Import aom library

This is the reference implementation for the Alliance for Open Media's av1 video code.

The commit used was 4d668d7feb1f8abd809d1bca0418570a7f142a36.

third_party/aom/aom_dsp/add_noise.c

@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+void aom_plane_add_noise_c(uint8_t *start, char *noise, char blackclamp[16],
+                           char whiteclamp[16], char bothclamp[16],
+                           unsigned int width, unsigned int height, int pitch) {
+  unsigned int i, j;
+
+  for (i = 0; i < height; ++i) {
+    uint8_t *pos = start + i * pitch;
+    char *ref = (char *)(noise + (rand() & 0xff));  // NOLINT
+
+    for (j = 0; j < width; ++j) {
+      int v = pos[j];
+
+      v = clamp(v - blackclamp[0], 0, 255);
+      v = clamp(v + bothclamp[0], 0, 255);
+      v = clamp(v - whiteclamp[0], 0, 255);
+
+      pos[j] = v + ref[j];
+    }
+  }
+}
+
+static double gaussian(double sigma, double mu, double x) {
+  return 1 / (sigma * sqrt(2.0 * 3.14159265)) *
+         (exp(-(x - mu) * (x - mu) / (2 * sigma * sigma)));
+}
+
+int aom_setup_noise(double sigma, int size, char *noise) {
+  char char_dist[256];
+  int next = 0, i, j;
+
+  // set up a 256 entry lookup that matches gaussian distribution
+  for (i = -32; i < 32; ++i) {
+    const int a_i = (int)(0.5 + 256 * gaussian(sigma, 0, i));
+    if (a_i) {
+      for (j = 0; j < a_i; ++j) {
+        char_dist[next + j] = (char)i;
+      }
+      next = next + j;
+    }
+  }
+
+  // Rounding error - might mean we have less than 256.
+  for (; next < 256; ++next) {
+    char_dist[next] = 0;
+  }
+
+  for (i = 0; i < size; ++i) {
+    noise[i] = char_dist[rand() & 0xff];  // NOLINT
+  }
+
+  // Returns the highest non 0 value used in distribution.
+  return -char_dist[0];
+}

third_party/aom/aom_dsp/ans.h

@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_ANS_H_
+#define AOM_DSP_ANS_H_
+// Constants, types and utilities for Asymmetric Numeral Systems
+// http://arxiv.org/abs/1311.2540v2
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+// Use windowed ANS, size is passed in at initialization
+#define ANS_MAX_SYMBOLS 1
+#define ANS_REVERSE 1
+
+typedef uint8_t AnsP8;
+#define ANS_P8_PRECISION 256u
+#define ANS_P8_SHIFT 8
+#define RANS_PROB_BITS 15
+#define RANS_PRECISION (1u << RANS_PROB_BITS)
+
+// L_BASE is the ANS base state. L_BASE % PRECISION must be 0.
+#define L_BASE (1u << 17)
+#define IO_BASE 256
+// Range I = { L_BASE, L_BASE + 1, ..., L_BASE * IO_BASE - 1 }
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_ANS_H_

third_party/aom/aom_dsp/ansreader.h

@@ -0,0 +1,214 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_ANSREADER_H_
+#define AOM_DSP_ANSREADER_H_
+// An implementation of Asymmetric Numeral Systems
+// http://arxiv.org/abs/1311.2540v2
+// Implements decoding of:
+// * rABS (range Asymmetric Binary Systems), a boolean coder
+// * rANS (range Asymmetric Numeral Systems), a multi-symbol coder
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/prob.h"
+#include "aom_dsp/ans.h"
+#include "aom_ports/mem_ops.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+struct AnsDecoder {
+  const uint8_t *buf;
+  int buf_offset;
+  uint32_t state;
+#if ANS_MAX_SYMBOLS
+  int symbols_left;
+  int window_size;
+#endif
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+};
+
+static INLINE int ans_read_reinit(struct AnsDecoder *const ans);
+
+static INLINE unsigned refill_state(struct AnsDecoder *const ans,
+                                    unsigned state) {
+#if ANS_REVERSE
+  while (state < L_BASE && ans->buf_offset < 0) {
+    state = state * IO_BASE + ans->buf[ans->buf_offset++];
+  }
+#else
+  while (state < L_BASE && ans->buf_offset > 0) {
+    state = state * IO_BASE + ans->buf[--ans->buf_offset];
+  }
+#endif
+  return state;
+}
+
+// Decode one rABS encoded boolean where the probability of the value being zero
+// is p0.
+static INLINE int rabs_read(struct AnsDecoder *ans, AnsP8 p0) {
+#if ANS_MAX_SYMBOLS
+  if (ans->symbols_left-- == 0) {
+    ans_read_reinit(ans);
+    ans->symbols_left--;
+  }
+#endif
+  unsigned state = refill_state(ans, ans->state);
+  const unsigned quotient = state / ANS_P8_PRECISION;
+  const unsigned remainder = state % ANS_P8_PRECISION;
+  const int value = remainder >= p0;
+  const unsigned qp0 = quotient * p0;
+  if (value)
+    state = state - qp0 - p0;
+  else
+    state = qp0 + remainder;
+  ans->state = state;
+  return value;
+}
+
+// Decode one rABS encoded boolean where the probability of the value being zero
+// is one half.
+static INLINE int rabs_read_bit(struct AnsDecoder *ans) {
+#if ANS_MAX_SYMBOLS
+  if (ans->symbols_left-- == 0) {
+    ans_read_reinit(ans);
+    ans->symbols_left--;
+  }
+#endif
+  unsigned state = refill_state(ans, ans->state);
+  const int value = !!(state & 0x80);
+  ans->state = ((state >> 1) & ~0x7F) | (state & 0x7F);
+  return value;
+}
+
+struct rans_dec_sym {
+  uint8_t val;
+  aom_cdf_prob prob;
+  aom_cdf_prob cum_prob;  // not-inclusive
+};
+
+static INLINE void fetch_sym(struct rans_dec_sym *out, const aom_cdf_prob *cdf,
+                             aom_cdf_prob rem) {
+  int i;
+  aom_cdf_prob cum_prob = 0, top_prob;
+  // TODO(skal): if critical, could be a binary search.
+  // Or, better, an O(1) alias-table.
+  for (i = 0; rem >= (top_prob = cdf[i]); ++i) {
+    cum_prob = top_prob;
+  }
+  out->val = i;
+  out->prob = top_prob - cum_prob;
+  out->cum_prob = cum_prob;
+}
+
+static INLINE int rans_read(struct AnsDecoder *ans, const aom_cdf_prob *tab) {
+  unsigned rem;
+  unsigned quo;
+  struct rans_dec_sym sym;
+#if ANS_MAX_SYMBOLS
+  if (ans->symbols_left-- == 0) {
+    ans_read_reinit(ans);
+    ans->symbols_left--;
+  }
+#endif
+  ans->state = refill_state(ans, ans->state);
+  quo = ans->state / RANS_PRECISION;
+  rem = ans->state % RANS_PRECISION;
+  fetch_sym(&sym, tab, rem);
+  ans->state = quo * sym.prob + rem - sym.cum_prob;
+  return sym.val;
+}
+
+static INLINE int ans_read_init(struct AnsDecoder *const ans,
+                                const uint8_t *const buf, int offset) {
+  unsigned x;
+  if (offset < 1) return 1;
+#if ANS_REVERSE
+  ans->buf = buf + offset;
+  ans->buf_offset = -offset;
+  x = buf[0];
+  if ((x & 0x80) == 0) {  // Marker is 0xxx xxxx
+    if (offset < 2) return 1;
+    ans->buf_offset += 2;
+    ans->state = mem_get_be16(buf) & 0x7FFF;
+#if L_BASE * IO_BASE > (1 << 23)
+  } else if ((x & 0xC0) == 0x80) {  // Marker is 10xx xxxx
+    if (offset < 3) return 1;
+    ans->buf_offset += 3;
+    ans->state = mem_get_be24(buf) & 0x3FFFFF;
+  } else {  // Marker is 11xx xxxx
+    if (offset < 4) return 1;
+    ans->buf_offset += 4;
+    ans->state = mem_get_be32(buf) & 0x3FFFFFFF;
+#else
+  } else {  // Marker is 1xxx xxxx
+    if (offset < 3) return 1;
+    ans->buf_offset += 3;
+    ans->state = mem_get_be24(buf) & 0x7FFFFF;
+#endif
+  }
+#else
+  ans->buf = buf;
+  x = buf[offset - 1];
+  if ((x & 0x80) == 0) {  // Marker is 0xxx xxxx
+    if (offset < 2) return 1;
+    ans->buf_offset = offset - 2;
+    ans->state = mem_get_le16(buf + offset - 2) & 0x7FFF;
+  } else if ((x & 0xC0) == 0x80) {  // Marker is 10xx xxxx
+    if (offset < 3) return 1;
+    ans->buf_offset = offset - 3;
+    ans->state = mem_get_le24(buf + offset - 3) & 0x3FFFFF;
+  } else if ((x & 0xE0) == 0xE0) {  // Marker is 111x xxxx
+    if (offset < 4) return 1;
+    ans->buf_offset = offset - 4;
+    ans->state = mem_get_le32(buf + offset - 4) & 0x1FFFFFFF;
+  } else {
+    // Marker 110x xxxx implies this byte is a superframe marker
+    return 1;
+  }
+#endif  // ANS_REVERSE
+#if CONFIG_ACCOUNTING
+  ans->accounting = NULL;
+#endif
+  ans->state += L_BASE;
+  if (ans->state >= L_BASE * IO_BASE) return 1;
+#if ANS_MAX_SYMBOLS
+  assert(ans->window_size > 1);
+  ans->symbols_left = ans->window_size;
+#endif
+  return 0;
+}
+
+#if ANS_REVERSE
+static INLINE int ans_read_reinit(struct AnsDecoder *const ans) {
+  return ans_read_init(ans, ans->buf + ans->buf_offset, -ans->buf_offset);
+}
+#endif
+
+static INLINE int ans_read_end(const struct AnsDecoder *const ans) {
+  return ans->buf_offset == 0 && ans->state < L_BASE;
+}
+
+static INLINE int ans_reader_has_error(const struct AnsDecoder *const ans) {
+  return ans->state < L_BASE / RANS_PRECISION;
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_ANSREADER_H_

third_party/aom/aom_dsp/answriter.h

@@ -0,0 +1,148 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_ANSWRITER_H_
+#define AOM_DSP_ANSWRITER_H_
+// An implementation of Asymmetric Numeral Systems
+// http://arxiv.org/abs/1311.2540v2
+// Implements encoding of:
+// * rABS (range Asymmetric Binary Systems), a boolean coder
+// * rANS (range Asymmetric Numeral Systems), a multi-symbol coder
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/ans.h"
+#include "aom_dsp/prob.h"
+#include "aom_ports/mem_ops.h"
+#include "av1/common/odintrin.h"
+
+#if RANS_PRECISION <= OD_DIVU_DMAX
+#define ANS_DIVREM(quotient, remainder, dividend, divisor) \
+  do {                                                     \
+    quotient = OD_DIVU_SMALL((dividend), (divisor));       \
+    remainder = (dividend) - (quotient) * (divisor);       \
+  } while (0)
+#else
+#define ANS_DIVREM(quotient, remainder, dividend, divisor) \
+  do {                                                     \
+    quotient = (dividend) / (divisor);                     \
+    remainder = (dividend) % (divisor);                    \
+  } while (0)
+#endif
+
+#define ANS_DIV8(dividend, divisor) OD_DIVU_SMALL((dividend), (divisor))
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+struct AnsCoder {
+  uint8_t *buf;
+  int buf_offset;
+  uint32_t state;
+};
+
+static INLINE void ans_write_init(struct AnsCoder *const ans,
+                                  uint8_t *const buf) {
+  ans->buf = buf;
+  ans->buf_offset = 0;
+  ans->state = L_BASE;
+}
+
+static INLINE int ans_write_end(struct AnsCoder *const ans) {
+  uint32_t state;
+  int ans_size;
+  assert(ans->state >= L_BASE);
+  assert(ans->state < L_BASE * IO_BASE);
+  state = ans->state - L_BASE;
+  if (state < (1u << 15)) {
+    mem_put_le16(ans->buf + ans->buf_offset, (0x00u << 15) + state);
+    ans_size = ans->buf_offset + 2;
+#if ANS_REVERSE
+#if L_BASE * IO_BASE > (1 << 23)
+  } else if (state < (1u << 22)) {
+    mem_put_le24(ans->buf + ans->buf_offset, (0x02u << 22) + state);
+    ans_size = ans->buf_offset + 3;
+  } else if (state < (1u << 30)) {
+    mem_put_le32(ans->buf + ans->buf_offset, (0x03u << 30) + state);
+    ans_size = ans->buf_offset + 4;
+#else
+  } else if (state < (1u << 23)) {
+    mem_put_le24(ans->buf + ans->buf_offset, (0x01u << 23) + state);
+    ans_size = ans->buf_offset + 3;
+#endif
+#else
+  } else if (state < (1u << 22)) {
+    mem_put_le24(ans->buf + ans->buf_offset, (0x02u << 22) + state);
+    ans_size = ans->buf_offset + 3;
+  } else if (state < (1u << 29)) {
+    mem_put_le32(ans->buf + ans->buf_offset, (0x07u << 29) + state);
+    ans_size = ans->buf_offset + 4;
+#endif
+  } else {
+    assert(0 && "State is too large to be serialized");
+    return ans->buf_offset;
+  }
+#if ANS_REVERSE
+  {
+    int i;
+    uint8_t tmp;
+    for (i = 0; i < (ans_size >> 1); i++) {
+      tmp = ans->buf[i];
+      ans->buf[i] = ans->buf[ans_size - 1 - i];
+      ans->buf[ans_size - 1 - i] = tmp;
+    }
+    ans->buf += ans_size;
+    ans->buf_offset = 0;
+    ans->state = L_BASE;
+  }
+#endif
+  return ans_size;
+}
+
+// Write one boolean using rABS where p0 is the probability of the value being
+// zero.
+static INLINE void rabs_write(struct AnsCoder *ans, int value, AnsP8 p0) {
+  const AnsP8 p = ANS_P8_PRECISION - p0;
+  const unsigned l_s = value ? p : p0;
+  unsigned state = ans->state;
+  while (state >= L_BASE / ANS_P8_PRECISION * IO_BASE * l_s) {
+    ans->buf[ans->buf_offset++] = state % IO_BASE;
+    state /= IO_BASE;
+  }
+  const unsigned quotient = ANS_DIV8(state, l_s);
+  const unsigned remainder = state - quotient * l_s;
+  ans->state = quotient * ANS_P8_PRECISION + remainder + (value ? p0 : 0);
+}
+
+// Encode one symbol using rANS.
+// cum_prob: The cumulative probability before this symbol (the offset of
+// the symbol in the symbol cycle)
+// prob: The probability of this symbol (l_s from the paper)
+// RANS_PRECISION takes the place of m from the paper.
+static INLINE void rans_write(struct AnsCoder *ans, aom_cdf_prob cum_prob,
+                              aom_cdf_prob prob) {
+  unsigned quotient, remainder;
+  while (ans->state >= L_BASE / RANS_PRECISION * IO_BASE * prob) {
+    ans->buf[ans->buf_offset++] = ans->state % IO_BASE;
+    ans->state /= IO_BASE;
+  }
+  ANS_DIVREM(quotient, remainder, ans->state, prob);
+  ans->state = quotient * RANS_PRECISION + remainder + cum_prob;
+}
+
+#undef ANS_DIV8
+#undef ANS_DIVREM
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_ANSWRITER_H_

third_party/aom/aom_dsp/aom_convolve.c

@@ -0,0 +1,854 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <string.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+
+static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const InterpKernel *x_filters, int x0_q4,
+                           int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const InterpKernel *x_filters, int x0_q4,
+                               int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(
+          dst[x] + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const InterpKernel *y_filters, int y0_q4,
+                          int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const InterpKernel *y_filters, int y0_q4,
+                              int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
+          dst[y * dst_stride] +
+              clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)),
+          1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                     ptrdiff_t dst_stride, const InterpKernel *const x_filters,
+                     int x0_q4, int x_step_q4,
+                     const InterpKernel *const y_filters, int y0_q4,
+                     int y_step_q4, int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint8_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp,
+                 MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
+                 intermediate_height);
+  convolve_vert(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), MAX_SB_SIZE, dst,
+                dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+}
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+  return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4, int w,
+                           int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+                 w, h);
+}
+
+void aom_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4, int w,
+                               int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                     x_step_q4, w, h);
+}
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const int16_t *filter_x, int x_step_q4,
+                          const int16_t *filter_y, int y_step_q4, int w,
+                          int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4,
+                w, h);
+}
+
+void aom_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4, int w,
+                              int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                    y_step_q4, w, h);
+}
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                     ptrdiff_t dst_stride, const int16_t *filter_x,
+                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+           filters_y, y0_q4, y_step_q4, w, h);
+}
+
+void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  /* Fixed size intermediate buffer places limits on parameters. */
+  DECLARE_ALIGNED(16, uint8_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  aom_convolve8_c(src, src_stride, temp, MAX_SB_SIZE, filter_x, x_step_q4,
+                  filter_y, y_step_q4, w, h);
+  aom_convolve_avg_c(temp, MAX_SB_SIZE, dst, dst_stride, NULL, 0, NULL, 0, w,
+                     h);
+}
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int filter_x_stride, const int16_t *filter_y,
+                         int filter_y_stride, int w, int h) {
+  int r;
+
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                        ptrdiff_t dst_stride, const int16_t *filter_x,
+                        int filter_x_stride, const int16_t *filter_y,
+                        int filter_y_stride, int w, int h) {
+  int x, y;
+
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                        ptrdiff_t dst_stride, const int16_t *filter_x,
+                        int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                        int w, int h) {
+  aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                        filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                       ptrdiff_t dst_stride, const int16_t *filter_x,
+                       int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                       int w, int h) {
+  aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                       filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                     ptrdiff_t dst_stride, const int16_t *filter_x,
+                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                  filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h) {
+  aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4, int w,
+                           int h) {
+  aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                      filter_y, y_step_q4, w, h);
+}
+
+#if CONFIG_LOOP_RESTORATION
+static void convolve_add_src_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *x_filters, int x0_q4,
+                                   int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
+                          src_x[SUBPEL_TAPS / 2 - 1]);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_add_src_vert(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const InterpKernel *y_filters, int y0_q4,
+                                  int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] =
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
+                     src_y[(SUBPEL_TAPS / 2 - 1) * src_stride]);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_add_src(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const InterpKernel *const x_filters, int x0_q4,
+                             int x_step_q4, const InterpKernel *const y_filters,
+                             int y0_q4, int y_step_q4, int w, int h) {
+  uint8_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_add_src_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride,
+                         temp, MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
+                         intermediate_height);
+  convolve_add_src_vert(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), MAX_SB_SIZE,
+                        dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_add_src_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                         x_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4, int w,
+                                  int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_add_src_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                        y_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4, int w,
+                             int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                   x_step_q4, filters_y, y0_q4, y_step_q4, w, h);
+}
+#endif  // CONFIG_LOOP_RESTORATION
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                  uint8_t *dst8, ptrdiff_t dst_stride,
+                                  const InterpKernel *x_filters, int x0_q4,
+                                  int x_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                      uint8_t *dst8, ptrdiff_t dst_stride,
+                                      const InterpKernel *x_filters, int x0_q4,
+                                      int x_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(
+          dst[x] + clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
+          1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                 uint8_t *dst8, ptrdiff_t dst_stride,
+                                 const InterpKernel *y_filters, int y0_q4,
+                                 int y_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                     uint8_t *dst8, ptrdiff_t dst_stride,
+                                     const InterpKernel *y_filters, int y0_q4,
+                                     int y_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
+          dst[y * dst_stride] +
+              clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
+          1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const InterpKernel *const x_filters, int x0_q4,
+                            int x_step_q4, const InterpKernel *const y_filters,
+                            int y0_q4, int y_step_q4, int w, int h, int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  highbd_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride,
+                        CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, x_filters, x0_q4,
+                        x_step_q4, w, intermediate_height, bd);
+  highbd_convolve_vert(
+      CONVERT_TO_BYTEPTR(temp) + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+      MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4, int w,
+                                  int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                        x_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const int16_t *filter_x, int x_step_q4,
+                                      const int16_t *filter_y, int y_step_q4,
+                                      int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                            x_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x, int x_step_q4,
+                                 const int16_t *filter_y, int y_step_q4, int w,
+                                 int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                       y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                           y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  highbd_convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+                  filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4, int w,
+                                int h, int bd) {
+  // Fixed size intermediate buffer places limits on parameters.
+  DECLARE_ALIGNED(16, uint16_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  aom_highbd_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE,
+                         filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
+  aom_highbd_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, dst,
+                            dst_stride, NULL, 0, NULL, 0, w, h, bd);
+}
+
+void aom_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
+                                uint8_t *dst8, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int filter_x_stride,
+                                const int16_t *filter_y, int filter_y_stride,
+                                int w, int h, int bd) {
+  int r;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w * sizeof(uint16_t));
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_highbd_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
+                               uint8_t *dst8, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int filter_x_stride,
+                               const int16_t *filter_y, int filter_y_stride,
+                               int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+#if CONFIG_LOOP_RESTORATION
+static void highbd_convolve_add_src_horiz(const uint8_t *src8,
+                                          ptrdiff_t src_stride, uint8_t *dst8,
+                                          ptrdiff_t dst_stride,
+                                          const InterpKernel *x_filters,
+                                          int x0_q4, int x_step_q4, int w,
+                                          int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_highbd(
+          ROUND_POWER_OF_TWO(sum, FILTER_BITS) + src_x[SUBPEL_TAPS / 2 - 1],
+          bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_add_src_vert(const uint8_t *src8,
+                                         ptrdiff_t src_stride, uint8_t *dst8,
+                                         ptrdiff_t dst_stride,
+                                         const InterpKernel *y_filters,
+                                         int y0_q4, int y_step_q4, int w, int h,
+                                         int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
+                                src_y[(SUBPEL_TAPS / 2 - 1) * src_stride],
+                            bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_add_src(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *const x_filters,
+                                    int x0_q4, int x_step_q4,
+                                    const InterpKernel *const y_filters,
+                                    int y0_q4, int y_step_q4, int w, int h,
+                                    int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  highbd_convolve_add_src_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                                src_stride, CONVERT_TO_BYTEPTR(temp),
+                                MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
+                                intermediate_height, bd);
+  highbd_convolve_add_src_vert(
+      CONVERT_TO_BYTEPTR(temp) + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+      MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_horiz_c(
+    const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+    ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,
+    const int16_t *filter_y, int y_step_q4, int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_add_src_horiz(src, src_stride, dst, dst_stride, filters_x,
+                                x0_q4, x_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_vert_c(const uint8_t *src,
+                                         ptrdiff_t src_stride, uint8_t *dst,
+                                         ptrdiff_t dst_stride,
+                                         const int16_t *filter_x, int x_step_q4,
+                                         const int16_t *filter_y, int y_step_q4,
+                                         int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_add_src_vert(src, src_stride, dst, dst_stride, filters_y,
+                               y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const int16_t *filter_x, int x_step_q4,
+                                    const int16_t *filter_y, int y_step_q4,
+                                    int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  highbd_convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                          x_step_q4, filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+#endif  // CONFIG_LOOP_RESTORATION
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/aom_dsp/aom_convolve.h

@@ -0,0 +1,57 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_DSP_AOM_CONVOLVE_H_
+#define AOM_DSP_AOM_CONVOLVE_H_
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Note: Fixed size intermediate buffers, place limits on parameters
+// of some functions. 2d filtering proceeds in 2 steps:
+//   (1) Interpolate horizontally into an intermediate buffer, temp.
+//   (2) Interpolate temp vertically to derive the sub-pixel result.
+// Deriving the maximum number of rows in the temp buffer (135):
+// --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+// --Largest block size is 64x64 pixels.
+// --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+//   original frame (in 1/16th pixel units).
+// --Must round-up because block may be located at sub-pixel position.
+// --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+// --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+#define MAX_EXT_SIZE 263
+#else
+#define MAX_EXT_SIZE 135
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+
+typedef void (*convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4, int w,
+                              int h);
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*highbd_convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_AOM_CONVOLVE_H_

third_party/aom/aom_dsp/aom_dsp.cmake

@@ -0,0 +1,509 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(AOM_DSP_COMMON_SOURCES
+    "${AOM_ROOT}/aom_dsp/aom_convolve.c"
+    "${AOM_ROOT}/aom_dsp/aom_convolve.h"
+    "${AOM_ROOT}/aom_dsp/aom_dsp_common.h"
+    "${AOM_ROOT}/aom_dsp/aom_filter.h"
+    "${AOM_ROOT}/aom_dsp/aom_simd.h"
+    "${AOM_ROOT}/aom_dsp/aom_simd_inline.h"
+    "${AOM_ROOT}/aom_dsp/blend.h"
+    "${AOM_ROOT}/aom_dsp/blend_a64_hmask.c"
+    "${AOM_ROOT}/aom_dsp/blend_a64_mask.c"
+    "${AOM_ROOT}/aom_dsp/blend_a64_vmask.c"
+    "${AOM_ROOT}/aom_dsp/intrapred.c"
+    "${AOM_ROOT}/aom_dsp/loopfilter.c"
+    "${AOM_ROOT}/aom_dsp/prob.c"
+    "${AOM_ROOT}/aom_dsp/prob.h"
+    "${AOM_ROOT}/aom_dsp/sad.c"
+    "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics.h"
+    "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics_c.h"
+    "${AOM_ROOT}/aom_dsp/simd/v256_intrinsics.h"
+    "${AOM_ROOT}/aom_dsp/simd/v256_intrinsics_c.h"
+    "${AOM_ROOT}/aom_dsp/simd/v64_intrinsics.h"
+    "${AOM_ROOT}/aom_dsp/simd/v64_intrinsics_c.h"
+    "${AOM_ROOT}/aom_dsp/subtract.c"
+    "${AOM_ROOT}/aom_dsp/txfm_common.h"
+    "${AOM_ROOT}/aom_dsp/x86/txfm_common_intrin.h")
+
+set(AOM_DSP_COMMON_ASM_SSE2
+    "${AOM_ROOT}/aom_dsp/x86/aom_convolve_copy_sse2.asm"
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_sse2.asm"
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm"
+    "${AOM_ROOT}/aom_dsp/x86/intrapred_sse2.asm")
+
+set(AOM_DSP_COMMON_INTRIN_SSE2
+    "${AOM_ROOT}/aom_dsp/x86/aom_asm_stubs.c"
+    "${AOM_ROOT}/aom_dsp/x86/convolve.h"
+    "${AOM_ROOT}/aom_dsp/x86/txfm_common_sse2.h"
+    "${AOM_ROOT}/aom_dsp/x86/loopfilter_sse2.c")
+
+set(AOM_DSP_COMMON_ASM_SSSE3
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_ssse3.asm"
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm"
+    "${AOM_ROOT}/aom_dsp/x86/intrapred_ssse3.asm")
+
+set(AOM_DSP_COMMON_INTRIN_SSSE3
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c"
+    "${AOM_ROOT}/aom_dsp/x86/inv_txfm_ssse3.c")
+
+set(AOM_DSP_COMMON_INTRIN_SSE4_1
+    "${AOM_ROOT}/aom_dsp/x86/blend_a64_hmask_sse4.c"
+    "${AOM_ROOT}/aom_dsp/x86/blend_a64_mask_sse4.c"
+    "${AOM_ROOT}/aom_dsp/x86/blend_a64_vmask_sse4.c")
+
+set(AOM_DSP_COMMON_INTRIN_AVX2
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c"
+    "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_avx2.c"
+    "${AOM_ROOT}/aom_dsp/x86/loopfilter_avx2.c")
+
+set(AOM_DSP_COMMON_ASM_NEON
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_avg_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve_avg_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve_copy_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct16x16_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct16x16_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct32x32_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct32x32_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct4x4_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct4x4_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct8x8_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct8x8_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/intrapred_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_16_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_4_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_8_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_mb_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/save_reg_neon.asm")
+
+set(AOM_DSP_COMMON_INTRIN_NEON
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/avg_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/fwd_txfm_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/hadamard_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/idct16x16_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/intrapred_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/sad4d_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/sad_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/subpel_variance_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/subtract_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/variance_neon.c")
+
+if ("${AOM_TARGET_CPU}" STREQUAL "arm64")
+  set(AOM_DSP_COMMON_INTRIN_NEON
+      ${AOM_DSP_COMMON_INTRIN_NEON}
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_avg_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve_avg_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve_copy_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct16x16_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct16x16_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct32x32_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct32x32_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct4x4_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct4x4_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct8x8_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct8x8_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/intrapred_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/loopfilter_16_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/loopfilter_4_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/loopfilter_8_neon.c")
+endif ()
+
+set(AOM_DSP_COMMON_INTRIN_DSPR2
+    "${AOM_ROOT}/aom_dsp/mips/common_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/common_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_avg_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_avg_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_vert_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_avg_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_avg_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_vert_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve_common_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred16_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred4_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred8_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/inv_txfm_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_filters_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_filters_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_macros_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_masks_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_mb_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_mb_vert_dspr2.c")
+
+set(AOM_DSP_COMMON_INTRIN_MSA
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_avg_horiz_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_avg_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_avg_vert_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_horiz_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_vert_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve_avg_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve_copy_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/fwd_dct32x32_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/fwd_txfm_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/fwd_txfm_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/idct16x16_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/idct32x32_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/idct4x4_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/idct8x8_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/inv_txfm_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_16_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_4_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_8_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/macros_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/txfm_macros_msa.h")
+
+if (CONFIG_HIGHBITDEPTH)
+  set(AOM_DSP_COMMON_ASM_SSE2
+      ${AOM_DSP_COMMON_ASM_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/highbd_intrapred_sse2.asm")
+
+  set(AOM_DSP_COMMON_INTRIN_SSE2
+      ${AOM_DSP_COMMON_INTRIN_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/highbd_loopfilter_sse2.c")
+
+  set(AOM_DSP_COMMON_INTRIN_AVX2
+      ${AOM_DSP_COMMON_INTRIN_AVX2}
+      "${AOM_ROOT}/aom_dsp/x86/highbd_convolve_avx2.c")
+else ()
+  set(AOM_DSP_COMMON_INTRIN_DSPR2
+      ${AOM_DSP_COMMON_INTRIN_DSPR2}
+      "${AOM_ROOT}/aom_dsp/mips/itrans16_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans32_cols_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans32_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans4_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans8_dspr2.c")
+endif ()
+
+if (CONFIG_ANS)
+  set(AOM_DSP_COMMON_SOURCES
+      ${AOM_DSP_COMMON_SOURCES}
+      "${AOM_ROOT}/aom_dsp/ans.h")
+elseif (CONFIG_DAALA_EC)
+  set(AOM_DSP_COMMON_SOURCES
+      ${AOM_DSP_COMMON_SOURCES}
+      "${AOM_ROOT}/aom_dsp/entcode.c"
+      "${AOM_ROOT}/aom_dsp/entcode.h")
+endif ()
+
+if (CONFIG_AV1)
+  set(AOM_DSP_COMMON_SOURCES
+      ${AOM_DSP_COMMON_SOURCES}
+      "${AOM_ROOT}/aom_dsp/inv_txfm.c"
+      "${AOM_ROOT}/aom_dsp/inv_txfm.h")
+
+  set(AOM_DSP_COMMON_ASM_SSE2
+      ${AOM_DSP_COMMON_ASM_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/inv_wht_sse2.asm")
+
+  set(AOM_DSP_COMMON_INTRIN_SSE2
+      ${AOM_DSP_COMMON_INTRIN_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/inv_txfm_sse2.c"
+      "${AOM_ROOT}/aom_dsp/x86/inv_txfm_sse2.h")
+endif ()
+
+if (CONFIG_DECODERS)
+  set(AOM_DSP_DECODER_SOURCES
+      "${AOM_ROOT}/aom_dsp/binary_codes_reader.c"
+      "${AOM_ROOT}/aom_dsp/binary_codes_reader.h"
+      "${AOM_ROOT}/aom_dsp/bitreader.h"
+      "${AOM_ROOT}/aom_dsp/bitreader_buffer.c"
+      "${AOM_ROOT}/aom_dsp/bitreader_buffer.h")
+
+  if (CONFIG_ANS)
+    set(AOM_DSP_DECODER_SOURCES
+        ${AOM_DSP_DECODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/ansreader.h")
+  elseif (CONFIG_DAALA_EC)
+    set(AOM_DSP_DECODER_SOURCES
+        ${AOM_DSP_DECODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/daalaboolreader.c"
+        "${AOM_ROOT}/aom_dsp/daalaboolreader.h"
+        "${AOM_ROOT}/aom_dsp/entdec.c"
+        "${AOM_ROOT}/aom_dsp/entdec.h")
+  else ()
+    set(AOM_DSP_DECODER_SOURCES
+        ${AOM_DSP_DECODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/dkboolreader.c"
+        "${AOM_ROOT}/aom_dsp/dkboolreader.h")
+  endif ()
+endif ()
+
+if (CONFIG_ENCODERS)
+  set(AOM_DSP_ENCODER_SOURCES
+      "${AOM_ROOT}/aom_dsp/binary_codes_writer.c"
+      "${AOM_ROOT}/aom_dsp/binary_codes_writer.h"
+      "${AOM_ROOT}/aom_dsp/bitwriter.h"
+      "${AOM_ROOT}/aom_dsp/bitwriter_buffer.c"
+      "${AOM_ROOT}/aom_dsp/bitwriter_buffer.h"
+      "${AOM_ROOT}/aom_dsp/psnr.c"
+      "${AOM_ROOT}/aom_dsp/psnr.h"
+      "${AOM_ROOT}/aom_dsp/variance.c"
+      "${AOM_ROOT}/aom_dsp/variance.h")
+
+  set(AOM_DSP_ENCODER_ASM_SSE2
+      ${AOM_DSP_ENCODER_ASM_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/halfpix_variance_impl_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/sad4d_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/sad_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/subtract_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/subpel_variance_sse2.asm")
+
+  set(AOM_DSP_ENCODER_INTRIN_SSE2
+      "${AOM_ROOT}/aom_dsp/x86/quantize_sse2.c")
+
+  set(AOM_DSP_ENCODER_ASM_SSSE3
+      "${AOM_ROOT}/aom_dsp/x86/sad_ssse3.asm")
+
+  set(AOM_DSP_ENCODER_ASM_SSSE3_X86_64
+      "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm"
+      "${AOM_ROOT}/aom_dsp/x86/ssim_opt_x86_64.asm")
+
+  set(AOM_DSP_ENCODER_INTRIN_SSE3 "${AOM_ROOT}/aom_dsp/x86/sad_sse3.asm")
+  set(AOM_DSP_ENCODER_ASM_SSE4_1 "${AOM_ROOT}/aom_dsp/x86/sad_sse4.asm")
+
+  set(AOM_DSP_ENCODER_INTRIN_AVX2
+      "${AOM_ROOT}/aom_dsp/x86/sad4d_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/sad_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/sad_impl_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/variance_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/variance_impl_avx2.c")
+
+  if (CONFIG_AV1_ENCODER)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/avg.c"
+        "${AOM_ROOT}/aom_dsp/fwd_txfm.c"
+        "${AOM_ROOT}/aom_dsp/fwd_txfm.h"
+        "${AOM_ROOT}/aom_dsp/quantize.c"
+        "${AOM_ROOT}/aom_dsp/quantize.h"
+        "${AOM_ROOT}/aom_dsp/sum_squares.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSE2
+        ${AOM_DSP_ENCODER_INTRIN_SSE2}
+        "${AOM_ROOT}/aom_dsp/x86/avg_intrin_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_dct32_8cols_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_dct32x32_impl_sse2.h"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_impl_sse2.h"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_sse2.h"
+        "${AOM_ROOT}/aom_dsp/x86/halfpix_variance_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/variance_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/sum_squares_sse2.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSSE3
+        ${AOM_DSP_ENCODER_INTRIN_SSSE3}
+        "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.c"
+        "${AOM_ROOT}/aom_dsp/x86/masked_variance_intrin_ssse3.c")
+
+    set(AOM_DSP_ENCODER_ASM_SSSE3_X86_64
+        ${AOM_DSP_ENCODER_ASM_SSSE3_X86_64}
+        "${AOM_ROOT}/aom_dsp/x86/avg_ssse3_x86_64.asm"
+        "${AOM_ROOT}/aom_dsp/x86/quantize_ssse3_x86_64.asm")
+
+    set(AOM_DSP_ENCODER_AVX_ASM_X86_64
+        ${AOM_DSP_ENCODER_AVX_ASM_X86_64}
+        "${AOM_ROOT}/aom_dsp/x86/quantize_avx_x86_64.asm")
+
+    set(AOM_DSP_ENCODER_INTRIN_MSA
+        "${AOM_ROOT}/aom_dsp/mips/avg_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/sad_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/subtract_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/variance_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/sub_pixel_variance_msa.c")
+
+    if (CONFIG_HIGHBITDEPTH)
+      set(AOM_DSP_ENCODER_INTRIN_SSE2
+          ${AOM_DSP_ENCODER_INTRIN_SSE2}
+          "${AOM_ROOT}/aom_dsp/x86/highbd_quantize_intrin_sse2.c"
+          "${AOM_ROOT}/aom_dsp/x86/highbd_subtract_sse2.c")
+    endif ()
+  endif ()
+
+  if (CONFIG_HIGHBITDEPTH)
+    set(AOM_DSP_ENCODER_ASM_SSE2
+        ${AOM_DSP_ENCODER_ASM_SSE2}
+        "${AOM_ROOT}/aom_dsp/x86/highbd_sad4d_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/highbd_sad_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/highbd_variance_impl_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSE2
+        ${AOM_DSP_ENCODER_INTRIN_SSE2}
+        "${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse2.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSE4_1
+        ${AOM_DSP_ENCODER_INTRIN_SSE4_1}
+        "${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse4.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_AVX2
+        ${AOM_DSP_ENCODER_INTRIN_AVX2}
+        "${AOM_ROOT}/aom_dsp/x86/sad_highbd_avx2.c")
+  endif ()
+
+  if (CONFIG_ANS)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/answriter.h"
+        "${AOM_ROOT}/aom_dsp/buf_ans.c"
+        "${AOM_ROOT}/aom_dsp/buf_ans.h")
+  elseif (CONFIG_DAALA_EC)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/daalaboolwriter.c"
+        "${AOM_ROOT}/aom_dsp/daalaboolwriter.h"
+        "${AOM_ROOT}/aom_dsp/entenc.c"
+        "${AOM_ROOT}/aom_dsp/entenc.h")
+  else ()
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/dkboolwriter.c"
+        "${AOM_ROOT}/aom_dsp/dkboolwriter.h")
+  endif ()
+
+  if (CONFIG_INTERNAL_STATS)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/fastssim.c"
+        "${AOM_ROOT}/aom_dsp/psnrhvs.c"
+        "${AOM_ROOT}/aom_dsp/ssim.c"
+        "${AOM_ROOT}/aom_dsp/ssim.h")
+  endif ()
+endif ()
+
+if (CONFIG_MOTION_VAR)
+  set(AOM_DSP_ENCODER_INTRIN_SSE4_1
+      ${AOM_DSP_ENCODER_INTRIN_SSE4_1}
+      "${AOM_ROOT}/aom_dsp/x86/obmc_sad_sse4.c"
+      "${AOM_ROOT}/aom_dsp/x86/obmc_variance_sse4.c")
+endif ()
+
+# Creates aom_dsp build targets. Must not be called until after libaom target
+# has been created.
+function (setup_aom_dsp_targets)
+  add_library(aom_dsp_common OBJECT ${AOM_DSP_COMMON_SOURCES})
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_common)
+  target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_dsp_common>)
+
+  if (CONFIG_DECODERS)
+    add_library(aom_dsp_decoder OBJECT ${AOM_DSP_DECODER_SOURCES})
+    set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_decoder)
+    target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_dsp_decoder>)
+  endif ()
+
+  if (CONFIG_ENCODERS)
+    add_library(aom_dsp_encoder OBJECT ${AOM_DSP_ENCODER_SOURCES})
+    set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_encoder)
+    target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_dsp_encoder>)
+  endif ()
+
+  if (HAVE_SSE2)
+    add_asm_library("aom_dsp_common_sse2" "AOM_DSP_COMMON_ASM_SSE2" "aom")
+    add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_SSE2")
+    if (CONFIG_ENCODERS)
+      add_asm_library("aom_dsp_encoder_sse2" "AOM_DSP_ENCODER_ASM_SSE2" "aom")
+      add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_SSE2")
+    endif()
+  endif ()
+
+  if (HAVE_SSE3 AND CONFIG_ENCODERS)
+    add_asm_library("aom_dsp_encoder_sse3" "AOM_DSP_ENCODER_INTRIN_SSE3" "aom")
+  endif ()
+
+  if (HAVE_SSSE3)
+    add_asm_library("aom_dsp_common_ssse3" "AOM_DSP_COMMON_ASM_SSSE3" "aom")
+    add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_SSSE3")
+
+    if (CONFIG_ENCODERS)
+      if ("${AOM_TARGET_CPU}" STREQUAL "x86_64")
+        list(APPEND AOM_DSP_ENCODER_ASM_SSSE3
+             ${AOM_DSP_ENCODER_ASM_SSSE3_X86_64})
+      endif ()
+      add_asm_library("aom_dsp_encoder_ssse3" "AOM_DSP_ENCODER_ASM_SSSE3" "aom")
+      add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_SSSE3")
+    endif ()
+  endif ()
+
+  if (HAVE_SSE4_1)
+    add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_SSE4_1")
+    if (CONFIG_ENCODERS)
+      if (AOM_DSP_ENCODER_INTRIN_SSE4_1)
+        add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_encoder"
+                                      "AOM_DSP_ENCODER_INTRIN_SSE4_1")
+      endif ()
+      add_asm_library("aom_dsp_encoder_sse4_1" "AOM_DSP_ENCODER_ASM_SSE4_1"
+                      "aom")
+    endif ()
+  endif ()
+
+  if (HAVE_AVX AND "${AOM_TARGET_CPU}" STREQUAL "x86_64")
+    add_asm_library("aom_dsp_encoder_avx" "AOM_DSP_ENCODER_AVX_ASM_X86_64"
+                    "aom")
+  endif ()
+
+  if (HAVE_AVX2)
+    add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_AVX2")
+    if (CONFIG_ENCODERS)
+      add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_AVX2")
+    endif ()
+  endif ()
+
+  if (HAVE_NEON_ASM)
+    if (AOM_ADS2GAS_REQUIRED)
+      add_gas_asm_library("aom_dsp_common_neon" "AOM_DSP_COMMON_ASM_NEON" "aom")
+    else ()
+      add_asm_library("aom_dsp_common_neon" "AOM_DSP_COMMON_ASM_NEON" "aom")
+    endif ()
+  endif ()
+
+  if (HAVE_NEON)
+    add_intrinsics_object_library("${AOM_NEON_INTRIN_FLAG}" "neon"
+                                  "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_NEON")
+  endif ()
+
+  if (HAVE_DSPR2)
+    add_intrinsics_object_library("" "dspr2" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_DSPR2")
+  endif ()
+
+  if (HAVE_MSA)
+    add_intrinsics_object_library("" "msa" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_MSA")
+    if (CONFIG_ENCODERS)
+      add_intrinsics_object_library("" "msa" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_MSA")
+    endif ()
+  endif ()
+
+  # Pass the new lib targets up to the parent scope instance of
+  # $AOM_LIB_TARGETS.
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE)
+endfunction ()

third_party/aom/aom_dsp/aom_dsp.mk

@@ -0,0 +1,428 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+DSP_SRCS-yes += aom_dsp.mk
+DSP_SRCS-yes += aom_dsp_common.h
+
+DSP_SRCS-$(HAVE_MSA)    += mips/macros_msa.h
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64)   += x86/synonyms.h
+
+# bit reader
+DSP_SRCS-yes += prob.h
+DSP_SRCS-yes += prob.c
+DSP_SRCS-$(CONFIG_ANS) += ans.h
+
+ifeq ($(CONFIG_ENCODERS),yes)
+ifeq ($(CONFIG_ANS),yes)
+DSP_SRCS-yes += answriter.h
+DSP_SRCS-yes += buf_ans.h
+DSP_SRCS-yes += buf_ans.c
+else ifeq ($(CONFIG_DAALA_EC),yes)
+DSP_SRCS-yes += entenc.c
+DSP_SRCS-yes += entenc.h
+DSP_SRCS-yes += daalaboolwriter.c
+DSP_SRCS-yes += daalaboolwriter.h
+else
+DSP_SRCS-yes += dkboolwriter.h
+DSP_SRCS-yes += dkboolwriter.c
+endif
+DSP_SRCS-yes += bitwriter.h
+DSP_SRCS-yes += bitwriter_buffer.c
+DSP_SRCS-yes += bitwriter_buffer.h
+DSP_SRCS-yes += binary_codes_writer.c
+DSP_SRCS-yes += binary_codes_writer.h
+DSP_SRCS-yes += psnr.c
+DSP_SRCS-yes += psnr.h
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += ssim.c
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += ssim.h
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += psnrhvs.c
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += fastssim.c
+endif
+
+ifeq ($(CONFIG_DECODERS),yes)
+ifeq ($(CONFIG_ANS),yes)
+DSP_SRCS-yes += ansreader.h
+else ifeq ($(CONFIG_DAALA_EC),yes)
+DSP_SRCS-yes += entdec.c
+DSP_SRCS-yes += entdec.h
+DSP_SRCS-yes += daalaboolreader.c
+DSP_SRCS-yes += daalaboolreader.h
+else
+DSP_SRCS-yes += dkboolreader.h
+DSP_SRCS-yes += dkboolreader.c
+endif
+DSP_SRCS-yes += bitreader.h
+DSP_SRCS-yes += bitreader_buffer.c
+DSP_SRCS-yes += bitreader_buffer.h
+DSP_SRCS-yes += binary_codes_reader.c
+DSP_SRCS-yes += binary_codes_reader.h
+endif
+
+# intra predictions
+DSP_SRCS-yes += intrapred.c
+
+ifeq ($(CONFIG_DAALA_EC),yes)
+DSP_SRCS-yes += entcode.c
+DSP_SRCS-yes += entcode.h
+endif
+
+DSP_SRCS-$(HAVE_SSE) += x86/intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/intrapred_ssse3.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_8t_ssse3.asm
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE)  += x86/highbd_intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_intrapred_sse2.asm
+endif  # CONFIG_HIGHBITDEPTH
+
+DSP_SRCS-$(HAVE_NEON_ASM) += arm/intrapred_neon_asm$(ASM)
+DSP_SRCS-$(HAVE_NEON) += arm/intrapred_neon.c
+DSP_SRCS-$(HAVE_MSA) += mips/intrapred_msa.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred4_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred16_dspr2.c
+
+DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.c
+
+# inter predictions
+DSP_SRCS-yes            += blend.h
+DSP_SRCS-yes            += blend_a64_mask.c
+DSP_SRCS-yes            += blend_a64_hmask.c
+DSP_SRCS-yes            += blend_a64_vmask.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_sse4.h
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_mask_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_hmask_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_vmask_sse4.c
+
+# interpolation filters
+DSP_SRCS-yes += aom_convolve.c
+DSP_SRCS-yes += aom_convolve.h
+DSP_SRCS-yes += aom_filter.h
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/convolve.h
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/aom_asm_stubs.c
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_subpixel_8t_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_subpixel_bilinear_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_8t_ssse3.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_bilinear_ssse3.asm
+DSP_SRCS-$(HAVE_AVX2)  += x86/aom_subpixel_8t_intrin_avx2.c
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_8t_intrin_ssse3.c
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_high_subpixel_8t_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_high_subpixel_bilinear_sse2.asm
+DSP_SRCS-$(HAVE_AVX2)  += x86/highbd_convolve_avx2.c
+endif
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_convolve_copy_sse2.asm
+
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes += arm/aom_convolve_copy_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve8_avg_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve8_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve_avg_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve_neon.c
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes += arm/aom_convolve_copy_neon.c
+DSP_SRCS-yes += arm/aom_convolve8_avg_neon.c
+DSP_SRCS-yes += arm/aom_convolve8_neon.c
+DSP_SRCS-yes += arm/aom_convolve_avg_neon.c
+DSP_SRCS-yes += arm/aom_convolve_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+
+# common (msa)
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_avg_horiz_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_avg_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_avg_vert_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_horiz_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_vert_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve_avg_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve_copy_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve_msa.h
+
+# common (dspr2)
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve_common_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_avg_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_avg_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_vert_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_avg_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_avg_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_vert_dspr2.c
+
+# loop filters
+DSP_SRCS-yes += loopfilter.c
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64)   += x86/loopfilter_sse2.c
+DSP_SRCS-$(HAVE_AVX2)                += x86/loopfilter_avx2.c
+
+DSP_SRCS-$(HAVE_NEON)   += arm/loopfilter_neon.c
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes  += arm/loopfilter_mb_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_16_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_8_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_4_neon$(ASM)
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes   += arm/loopfilter_16_neon.c
+DSP_SRCS-yes   += arm/loopfilter_8_neon.c
+DSP_SRCS-yes   += arm/loopfilter_4_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_msa.h
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_16_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_8_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_4_msa.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_filters_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_filters_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_macros_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_masks_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_vert_dspr2.c
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_loopfilter_sse2.c
+endif  # CONFIG_HIGHBITDEPTH
+
+DSP_SRCS-yes            += txfm_common.h
+DSP_SRCS-yes            += x86/txfm_common_intrin.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/txfm_common_sse2.h
+DSP_SRCS-$(HAVE_MSA)    += mips/txfm_macros_msa.h
+
+# forward transform
+ifneq ($(findstring yes,$(CONFIG_AV1)$(CONFIG_PVQ)),)
+DSP_SRCS-$(HAVE_AVX2)   += x86/txfm_common_avx2.h
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+DSP_SRCS-yes            += fwd_txfm.c
+DSP_SRCS-yes            += fwd_txfm.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_dct32_8cols_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_impl_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_dct32x32_impl_sse2.h
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/fwd_txfm_ssse3_x86_64.asm
+endif
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_txfm_avx2.h
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_txfm_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_dct32x32_impl_avx2.h
+DSP_SRCS-$(HAVE_NEON)   += arm/fwd_txfm_neon.c
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.h
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_dct32x32_msa.c
+endif  # CONFIG_AV1_ENCODER
+endif  # CONFIG_AV1
+
+# inverse transform
+ifeq ($(CONFIG_AV1), yes)
+DSP_SRCS-yes            += inv_txfm.h
+DSP_SRCS-yes            += inv_txfm.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_txfm_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_txfm_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_wht_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3)  += x86/inv_txfm_ssse3.c
+
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes  += arm/save_reg_neon$(ASM)
+DSP_SRCS-yes  += arm/idct4x4_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct4x4_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct8x8_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct8x8_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct16x16_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct16x16_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct32x32_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct32x32_add_neon$(ASM)
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes  += arm/idct4x4_1_add_neon.c
+DSP_SRCS-yes  += arm/idct4x4_add_neon.c
+DSP_SRCS-yes  += arm/idct8x8_1_add_neon.c
+DSP_SRCS-yes  += arm/idct8x8_add_neon.c
+DSP_SRCS-yes  += arm/idct16x16_1_add_neon.c
+DSP_SRCS-yes  += arm/idct16x16_add_neon.c
+DSP_SRCS-yes  += arm/idct32x32_1_add_neon.c
+DSP_SRCS-yes  += arm/idct32x32_add_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+DSP_SRCS-$(HAVE_NEON)  += arm/idct16x16_neon.c
+
+DSP_SRCS-$(HAVE_MSA)   += mips/inv_txfm_msa.h
+DSP_SRCS-$(HAVE_MSA)   += mips/idct4x4_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct8x8_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct16x16_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct32x32_msa.c
+
+ifneq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_DSPR2) += mips/inv_txfm_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans4_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans16_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_cols_dspr2.c
+endif  # CONFIG_HIGHBITDEPTH
+endif  # CONFIG_AV1
+
+# quantization
+ifneq ($(filter yes,$(CONFIG_AV1_ENCODER)),)
+DSP_SRCS-yes            += quantize.c
+DSP_SRCS-yes            += quantize.h
+
+DSP_SRCS-$(HAVE_SSE2)   += x86/quantize_sse2.c
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_quantize_intrin_sse2.c
+endif
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/quantize_ssse3_x86_64.asm
+DSP_SRCS-$(HAVE_AVX)    += x86/quantize_avx_x86_64.asm
+endif
+
+# avg
+DSP_SRCS-yes           += avg.c
+DSP_SRCS-$(HAVE_SSE2)  += x86/avg_intrin_sse2.c
+DSP_SRCS-$(HAVE_NEON)  += arm/avg_neon.c
+DSP_SRCS-$(HAVE_MSA)   += mips/avg_msa.c
+DSP_SRCS-$(HAVE_NEON)  += arm/hadamard_neon.c
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSSE3) += x86/avg_ssse3_x86_64.asm
+endif
+
+# high bit depth subtract
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)  += x86/highbd_subtract_sse2.c
+endif
+
+endif  # CONFIG_AV1_ENCODER
+
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+DSP_SRCS-yes            += sum_squares.c
+
+DSP_SRCS-$(HAVE_SSE2)   += x86/sum_squares_sse2.c
+endif # CONFIG_AV1_ENCODER
+
+ifeq ($(CONFIG_ENCODERS),yes)
+DSP_SRCS-yes            += sad.c
+DSP_SRCS-yes            += subtract.c
+
+DSP_SRCS-$(HAVE_MEDIA)  += arm/sad_media$(ASM)
+DSP_SRCS-$(HAVE_NEON)   += arm/sad4d_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/sad_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/subtract_neon.c
+
+DSP_SRCS-$(HAVE_MSA)    += mips/sad_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/subtract_msa.c
+
+DSP_SRCS-$(HAVE_SSE3)   += x86/sad_sse3.asm
+DSP_SRCS-$(HAVE_SSSE3)  += x86/sad_ssse3.asm
+DSP_SRCS-$(HAVE_SSE4_1) += x86/sad_sse4.asm
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad4d_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad_avx2.c
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad_highbd_avx2.c
+endif
+
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+ifeq ($(CONFIG_EXT_INTER),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/masked_sad_intrin_ssse3.c
+DSP_SRCS-$(HAVE_SSSE3)  += x86/masked_variance_intrin_ssse3.c
+endif  #CONFIG_EXT_INTER
+ifeq ($(CONFIG_MOTION_VAR),yes)
+DSP_SRCS-$(HAVE_SSE4_1) += x86/obmc_sad_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/obmc_variance_sse4.c
+endif  #CONFIG_MOTION_VAR
+ifeq ($(CONFIG_EXT_PARTITION),yes)
+DSP_SRCS-$(HAVE_AVX2) += x86/sad_impl_avx2.c
+endif
+endif  #CONFIG_AV1_ENCODER
+
+DSP_SRCS-$(HAVE_SSE)    += x86/sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE)    += x86/sad_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/sad_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/subtract_sse2.asm
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad_sse2.asm
+endif  # CONFIG_HIGHBITDEPTH
+
+endif  # CONFIG_ENCODERS
+
+ifneq ($(filter yes,$(CONFIG_ENCODERS)),)
+DSP_SRCS-yes            += variance.c
+DSP_SRCS-yes            += variance.h
+
+DSP_SRCS-$(HAVE_MEDIA)  += arm/bilinear_filter_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/subpel_variance_media.c
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_h_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_hv_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_v_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_media$(ASM)
+DSP_SRCS-$(HAVE_NEON)   += arm/subpel_variance_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/variance_neon.c
+
+DSP_SRCS-$(HAVE_MSA)    += mips/variance_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/sub_pixel_variance_msa.c
+
+DSP_SRCS-$(HAVE_SSE)    += x86/variance_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/variance_sse2.c  # Contains SSE2 and SSSE3
+DSP_SRCS-$(HAVE_SSE2)   += x86/halfpix_variance_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/halfpix_variance_impl_sse2.asm
+DSP_SRCS-$(HAVE_AVX2)   += x86/variance_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/variance_impl_avx2.c
+
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/ssim_opt_x86_64.asm
+endif  # ARCH_X86_64
+
+DSP_SRCS-$(HAVE_SSE)    += x86/subpel_variance_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/subpel_variance_sse2.asm  # Contains SSE2 and SSSE3
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_variance_sse2.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_variance_sse4.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_variance_impl_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_subpel_variance_impl_sse2.asm
+endif  # CONFIG_HIGHBITDEPTH
+endif  # CONFIG_ENCODERS
+
+DSP_SRCS-no += $(DSP_SRCS_REMOVE-yes)
+
+DSP_SRCS-yes += aom_dsp_rtcd.c
+DSP_SRCS-yes += aom_dsp_rtcd_defs.pl
+
+DSP_SRCS-yes += aom_simd.h
+DSP_SRCS-yes += aom_simd_inline.h
+DSP_SRCS-yes += simd/v64_intrinsics.h
+DSP_SRCS-yes += simd/v64_intrinsics_c.h
+DSP_SRCS-yes += simd/v128_intrinsics.h
+DSP_SRCS-yes += simd/v128_intrinsics_c.h
+DSP_SRCS-yes += simd/v256_intrinsics.h
+DSP_SRCS-yes += simd/v256_intrinsics_c.h
+DSP_SRCS-yes += simd/v256_intrinsics_v128.h
+DSP_SRCS-$(HAVE_SSE2) += simd/v64_intrinsics_x86.h
+DSP_SRCS-$(HAVE_SSE2) += simd/v128_intrinsics_x86.h
+DSP_SRCS-$(HAVE_SSE2) += simd/v256_intrinsics_x86.h
+DSP_SRCS-$(HAVE_NEON) += simd/v64_intrinsics_arm.h
+DSP_SRCS-$(HAVE_NEON) += simd/v128_intrinsics_arm.h
+DSP_SRCS-$(HAVE_NEON) += simd/v256_intrinsics_arm.h
+
+$(eval $(call rtcd_h_template,aom_dsp_rtcd,aom_dsp/aom_dsp_rtcd_defs.pl))

third_party/aom/aom_dsp/aom_dsp_common.h

@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_DSP_COMMON_H_
+#define AOM_DSP_AOM_DSP_COMMON_H_
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef MAX_SB_SIZE
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+#define MAX_SB_SIZE 128
+#else
+#define MAX_SB_SIZE 64
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+#endif  // ndef MAX_SB_SIZE
+
+#define AOMMIN(x, y) (((x) < (y)) ? (x) : (y))
+#define AOMMAX(x, y) (((x) > (y)) ? (x) : (y))
+
+#define IMPLIES(a, b) (!(a) || (b))  //  Logical 'a implies b' (or 'a -> b')
+
+#define IS_POWER_OF_TWO(x) (((x) & ((x)-1)) == 0)
+
+/* Left shifting a negative value became undefined behavior in C99 (downgraded
+   from merely implementation-defined in C89). This should still compile to the
+   correct thing on any two's-complement machine, but avoid ubsan warnings.*/
+#define AOM_SIGNED_SHL(x, shift) ((x) * (((x)*0 + 1) << (shift)))
+
+// These can be used to give a hint about branch outcomes.
+// This can have an effect, even if your target processor has a
+// good branch predictor, as these hints can affect basic block
+// ordering by the compiler.
+#ifdef __GNUC__
+#define LIKELY(v) __builtin_expect(v, 1)
+#define UNLIKELY(v) __builtin_expect(v, 0)
+#else
+#define LIKELY(v) (v)
+#define UNLIKELY(v) (v)
+#endif
+
+#define AOM_SWAP(type, a, b) \
+  do {                       \
+    type c = (b);            \
+    b = a;                   \
+    a = c;                   \
+  } while (0)
+
+#if CONFIG_AOM_QM
+typedef uint16_t qm_val_t;
+#define AOM_QM_BITS 6
+#endif
+#if CONFIG_HIGHBITDEPTH
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int64_t tran_high_t;
+typedef int32_t tran_low_t;
+#else
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int32_t tran_high_t;
+typedef int16_t tran_low_t;
+#endif  // CONFIG_HIGHBITDEPTH
+
+static INLINE uint8_t clip_pixel(int val) {
+  return (val > 255) ? 255 : (val < 0) ? 0 : val;
+}
+
+static INLINE int clamp(int value, int low, int high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+static INLINE double fclamp(double value, double low, double high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE uint16_t clip_pixel_highbd(int val, int bd) {
+  switch (bd) {
+    case 8:
+    default: return (uint16_t)clamp(val, 0, 255);
+    case 10: return (uint16_t)clamp(val, 0, 1023);
+    case 12: return (uint16_t)clamp(val, 0, 4095);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_AOM_DSP_COMMON_H_

third_party/aom/aom_dsp/aom_dsp_rtcd.c

@@ -0,0 +1,16 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "./aom_config.h"
+#define RTCD_C
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/aom_once.h"
+
+void aom_dsp_rtcd() { once(setup_rtcd_internal); }

third_party/aom/aom_dsp/aom_filter.h

@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_FILTER_H_
+#define AOM_DSP_AOM_FILTER_H_
+
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define FILTER_BITS 7
+
+#define SUBPEL_BITS 4
+#define SUBPEL_MASK ((1 << SUBPEL_BITS) - 1)
+#define SUBPEL_SHIFTS (1 << SUBPEL_BITS)
+#define SUBPEL_TAPS 8
+
+typedef int16_t InterpKernel[SUBPEL_TAPS];
+
+#define BIL_SUBPEL_BITS 3
+#define BIL_SUBPEL_SHIFTS (1 << BIL_SUBPEL_BITS)
+
+// 2 tap bilinear filters
+static const uint8_t bilinear_filters_2t[BIL_SUBPEL_SHIFTS][2] = {
+  { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 },
+  { 64, 64 }, { 48, 80 },  { 32, 96 }, { 16, 112 },
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_AOM_FILTER_H_

third_party/aom/aom_dsp/aom_simd.h

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_AOM_SIMD_H_
+#define AOM_DSP_AOM_AOM_SIMD_H_
+
+#include <stdint.h>
+
+#if defined(_WIN32)
+#include <intrin.h>
+#endif
+
+#include "./aom_config.h"
+#include "./aom_simd_inline.h"
+
+#define SIMD_CHECK 1  // Sanity checks in C equivalents
+
+#if HAVE_NEON
+#include "simd/v256_intrinsics_arm.h"
+// VS compiling for 32 bit targets does not support vector types in
+// structs as arguments, which makes the v256 type of the intrinsics
+// hard to support, so optimizations for this target are disabled.
+#elif HAVE_SSE2 && (defined(_WIN64) || !defined(_MSC_VER) || defined(__clang__))
+#include "simd/v256_intrinsics_x86.h"
+#else
+#include "simd/v256_intrinsics.h"
+#endif
+
+#endif  // AOM_DSP_AOM_AOM_SIMD_H_

third_party/aom/aom_dsp/aom_simd_inline.h

@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_SIMD_INLINE_H_
+#define AOM_DSP_AOM_SIMD_INLINE_H_
+
+#include "aom/aom_integer.h"
+
+#ifndef SIMD_INLINE
+#define SIMD_INLINE static AOM_FORCE_INLINE
+#endif
+
+#endif  // AOM_DSP_AOM_SIMD_INLINE_H_

third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon.c

@@ -0,0 +1,364 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(int16x4_t dsrc0, int16x4_t dsrc1,
+                                       int16x4_t dsrc2, int16x4_t dsrc3,
+                                       int16x4_t dsrc4, int16x4_t dsrc5,
+                                       int16x4_t dsrc6, int16x4_t dsrc7,
+                                       int16x8_t q0s16) {
+  int32x4_t qdst;
+  int16x4_t d0s16, d1s16;
+
+  d0s16 = vget_low_s16(q0s16);
+  d1s16 = vget_high_s16(q0s16);
+
+  qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+  qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+  return qdst;
+}
+
+void aom_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y,  // unused
+                                  int y_step_q4,            // unused
+                                  int w, int h) {
+  int width;
+  const uint8_t *s;
+  uint8_t *d;
+  uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+  uint32x2_t d2u32, d3u32, d6u32, d7u32, d28u32, d29u32, d30u32, d31u32;
+  uint8x16_t q1u8, q3u8, q12u8, q13u8, q14u8, q15u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+  uint16x8x2_t q0x2u16;
+  uint8x8x2_t d0x2u8, d1x2u8;
+  uint32x2x2_t d0x2u32;
+  uint16x4x2_t d0x2u16, d1x2u16;
+  uint32x4x2_t q0x2u32;
+
+  assert(x_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_y;
+
+  q0s16 = vld1q_s16(filter_x);
+
+  src -= 3;                // adjust for taps
+  for (; h > 0; h -= 4) {  // loop_horiz_v
+    s = src;
+    d24u8 = vld1_u8(s);
+    s += src_stride;
+    d25u8 = vld1_u8(s);
+    s += src_stride;
+    d26u8 = vld1_u8(s);
+    s += src_stride;
+    d27u8 = vld1_u8(s);
+
+    q12u8 = vcombine_u8(d24u8, d25u8);
+    q13u8 = vcombine_u8(d26u8, d27u8);
+
+    q0x2u16 =
+        vtrnq_u16(vreinterpretq_u16_u8(q12u8), vreinterpretq_u16_u8(q13u8));
+    d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+    d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+    d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+    d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+    d0x2u8 = vtrn_u8(d24u8, d25u8);
+    d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+    __builtin_prefetch(src + src_stride * 4);
+    __builtin_prefetch(src + src_stride * 5);
+
+    q8u16 = vmovl_u8(d0x2u8.val[0]);
+    q9u16 = vmovl_u8(d0x2u8.val[1]);
+    q10u16 = vmovl_u8(d1x2u8.val[0]);
+    q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+    src += 7;
+    d16u16 = vget_low_u16(q8u16);
+    d17u16 = vget_high_u16(q8u16);
+    d18u16 = vget_low_u16(q9u16);
+    d19u16 = vget_high_u16(q9u16);
+    q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+    q9u16 = vcombine_u16(d17u16, d19u16);
+
+    d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));         // vmov 23 21
+    for (width = w; width > 0; width -= 4, src += 4, dst += 4) {  // loop_horiz
+      s = src;
+      d28u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d29u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d31u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+      __builtin_prefetch(src + 64);
+
+      d0x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d28u32), vreinterpret_u16_u32(d31u32));
+      d1x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d29u32), vreinterpret_u16_u32(d30u32));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                       vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+      d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                       vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+      __builtin_prefetch(src + 64 + src_stride);
+
+      q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+      q0x2u32 =
+          vtrnq_u32(vreinterpretq_u32_u8(q14u8), vreinterpretq_u32_u8(q15u8));
+
+      d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+      d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+      q12u16 = vmovl_u8(d28u8);
+      q13u16 = vmovl_u8(d29u8);
+
+      __builtin_prefetch(src + 64 + src_stride * 2);
+
+      d = dst;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+      d += dst_stride;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+      d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+      d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16, d18s16, d19s16,
+                             d23s16, d24s16, q0s16);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16, d19s16, d23s16,
+                             d24s16, d26s16, q0s16);
+      q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16, d23s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      __builtin_prefetch(src + 64 + src_stride * 3);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8), vreinterpret_u16_u8(d3u8));
+      d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                         vreinterpret_u32_u16(d0x2u16.val[1]));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                       vreinterpret_u8_u32(d0x2u32.val[1]));
+
+      q1u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+      d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+      d = dst;
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+      q8u16 = q9u16;
+      d20s16 = d23s16;
+      q11u16 = q12u16;
+      q9u16 = q13u16;
+      d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    }
+    src += src_stride * 4 - w - 7;
+    dst += dst_stride * 4 - w;
+  }
+  return;
+}
+
+void aom_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x,  // unused
+                                 int x_step_q4,            // unused
+                                 const int16_t *filter_y, int y_step_q4, int w,
+                                 int h) {
+  int height;
+  const uint8_t *s;
+  uint8_t *d;
+  uint8x8_t d2u8, d3u8;
+  uint32x2_t d2u32, d3u32, d6u32, d7u32;
+  uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+  uint8x16_t q1u8, q3u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+  assert(y_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_x;
+
+  src -= src_stride * 3;
+  q0s16 = vld1q_s16(filter_y);
+  for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+    s = src;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+    s += src_stride;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+    s += src_stride;
+    d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+    s += src_stride;
+
+    q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
+    q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
+    q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+    q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+    d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+    d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d = dst;
+    for (height = h; height > 0; height -= 4) {  // loop_vert
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+      s += src_stride;
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+      s += src_stride;
+
+      q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+      q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+      d += dst_stride;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+      d -= dst_stride * 3;
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+      d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      __builtin_prefetch(s);
+      __builtin_prefetch(s + src_stride);
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16, d20s16, d21s16,
+                             d22s16, d24s16, q0s16);
+      __builtin_prefetch(s + src_stride * 2);
+      __builtin_prefetch(s + src_stride * 3);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16, d21s16, d22s16,
+                             d24s16, d26s16, q0s16);
+      __builtin_prefetch(d);
+      __builtin_prefetch(d + dst_stride);
+      q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16, d22s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      __builtin_prefetch(d + dst_stride * 2);
+      __builtin_prefetch(d + dst_stride * 3);
+      q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+      d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+      d += dst_stride;
+
+      q8u16 = q10u16;
+      d18s16 = d22s16;
+      d19s16 = d24s16;
+      q10u16 = q13u16;
+      d22s16 = d25s16;
+    }
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon_asm.asm

@@ -0,0 +1,295 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; AV1_FILTER_WEIGHT == 128
+    ; AV1_FILTER_SHIFT == 7
+
+    EXPORT  |aom_convolve8_avg_horiz_neon|
+    EXPORT  |aom_convolve8_avg_vert_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|aom_convolve8_avg_horiz_neon| PROC
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+aom_convolve8_avg_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+aom_convolve8_avg_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; slightly out of order load to match the existing data
+    vld1.u32        {d6[0]}, [r2], r3
+    vld1.u32        {d7[0]}, [r2], r3
+    vld1.u32        {d6[1]}, [r2], r3
+    vld1.u32        {d7[1]}, [r2], r3
+
+    sub             r2, r2, r3, lsl #2      ; reset for store
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_avg_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt aom_convolve8_avg_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|aom_convolve8_avg_vert_neon| PROC
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+aom_convolve8_avg_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+aom_convolve8_avg_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    vld1.u32        {d6[0]}, [r5@32], r3
+    vld1.u32        {d6[1]}, [r8@32], r3
+    vld1.u32        {d7[0]}, [r5@32], r3
+    vld1.u32        {d7[1]}, [r8@32], r3
+
+    pld             [r7]
+    pld             [r4]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    sub             r5, r5, r3, lsl #1      ; reset for store
+    sub             r8, r8, r3, lsl #1
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             aom_convolve8_avg_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_avg_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END

third_party/aom/aom_dsp/arm/aom_convolve8_neon.c

@@ -0,0 +1,331 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(int16x4_t dsrc0, int16x4_t dsrc1,
+                                       int16x4_t dsrc2, int16x4_t dsrc3,
+                                       int16x4_t dsrc4, int16x4_t dsrc5,
+                                       int16x4_t dsrc6, int16x4_t dsrc7,
+                                       int16x8_t q0s16) {
+  int32x4_t qdst;
+  int16x4_t d0s16, d1s16;
+
+  d0s16 = vget_low_s16(q0s16);
+  d1s16 = vget_high_s16(q0s16);
+
+  qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+  qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+  return qdst;
+}
+
+void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y,  // unused
+                              int y_step_q4,            // unused
+                              int w, int h) {
+  int width;
+  const uint8_t *s, *psrc;
+  uint8_t *d, *pdst;
+  uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+  uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
+  uint8x16_t q12u8, q13u8, q14u8, q15u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+  uint16x8x2_t q0x2u16;
+  uint8x8x2_t d0x2u8, d1x2u8;
+  uint32x2x2_t d0x2u32;
+  uint16x4x2_t d0x2u16, d1x2u16;
+  uint32x4x2_t q0x2u32;
+
+  assert(x_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_y;
+
+  q0s16 = vld1q_s16(filter_x);
+
+  src -= 3;  // adjust for taps
+  for (; h > 0; h -= 4, src += src_stride * 4,
+                dst += dst_stride * 4) {  // loop_horiz_v
+    s = src;
+    d24u8 = vld1_u8(s);
+    s += src_stride;
+    d25u8 = vld1_u8(s);
+    s += src_stride;
+    d26u8 = vld1_u8(s);
+    s += src_stride;
+    d27u8 = vld1_u8(s);
+
+    q12u8 = vcombine_u8(d24u8, d25u8);
+    q13u8 = vcombine_u8(d26u8, d27u8);
+
+    q0x2u16 =
+        vtrnq_u16(vreinterpretq_u16_u8(q12u8), vreinterpretq_u16_u8(q13u8));
+    d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+    d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+    d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+    d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+    d0x2u8 = vtrn_u8(d24u8, d25u8);
+    d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+    __builtin_prefetch(src + src_stride * 4);
+    __builtin_prefetch(src + src_stride * 5);
+    __builtin_prefetch(src + src_stride * 6);
+
+    q8u16 = vmovl_u8(d0x2u8.val[0]);
+    q9u16 = vmovl_u8(d0x2u8.val[1]);
+    q10u16 = vmovl_u8(d1x2u8.val[0]);
+    q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+    d16u16 = vget_low_u16(q8u16);
+    d17u16 = vget_high_u16(q8u16);
+    d18u16 = vget_low_u16(q9u16);
+    d19u16 = vget_high_u16(q9u16);
+    q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+    q9u16 = vcombine_u16(d17u16, d19u16);
+
+    d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));  // vmov 23 21
+    for (width = w, psrc = src + 7, pdst = dst; width > 0;
+         width -= 4, psrc += 4, pdst += 4) {  // loop_horiz
+      s = psrc;
+      d28u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d29u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d31u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+      __builtin_prefetch(psrc + 64);
+
+      d0x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d28u32), vreinterpret_u16_u32(d31u32));
+      d1x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d29u32), vreinterpret_u16_u32(d30u32));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                       vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+      d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                       vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+      __builtin_prefetch(psrc + 64 + src_stride);
+
+      q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+      q0x2u32 =
+          vtrnq_u32(vreinterpretq_u32_u8(q14u8), vreinterpretq_u32_u8(q15u8));
+
+      d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+      d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+      q12u16 = vmovl_u8(d28u8);
+      q13u16 = vmovl_u8(d29u8);
+
+      __builtin_prefetch(psrc + 64 + src_stride * 2);
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+      d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+      d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16, d18s16, d19s16,
+                             d23s16, d24s16, q0s16);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16, d19s16, d23s16,
+                             d24s16, d26s16, q0s16);
+      q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16, d23s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      __builtin_prefetch(psrc + 60 + src_stride * 3);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8), vreinterpret_u16_u8(d3u8));
+      d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                         vreinterpret_u32_u16(d0x2u16.val[1]));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                       vreinterpret_u8_u32(d0x2u32.val[1]));
+
+      d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
+      d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);
+
+      d = pdst;
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+      q8u16 = q9u16;
+      d20s16 = d23s16;
+      q11u16 = q12u16;
+      q9u16 = q13u16;
+      d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    }
+  }
+  return;
+}
+
+void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x,  // unused
+                             int x_step_q4,            // unused
+                             const int16_t *filter_y, int y_step_q4, int w,
+                             int h) {
+  int height;
+  const uint8_t *s;
+  uint8_t *d;
+  uint32x2_t d2u32, d3u32;
+  uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+  assert(y_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_x;
+
+  src -= src_stride * 3;
+  q0s16 = vld1q_s16(filter_y);
+  for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+    s = src;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+    s += src_stride;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+    s += src_stride;
+    d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+    s += src_stride;
+
+    q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
+    q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
+    q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+    q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+    d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+    d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d = dst;
+    for (height = h; height > 0; height -= 4) {  // loop_vert
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+      s += src_stride;
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+      s += src_stride;
+
+      q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+      q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+      d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      __builtin_prefetch(d);
+      __builtin_prefetch(d + dst_stride);
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16, d20s16, d21s16,
+                             d22s16, d24s16, q0s16);
+      __builtin_prefetch(d + dst_stride * 2);
+      __builtin_prefetch(d + dst_stride * 3);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16, d21s16, d22s16,
+                             d24s16, d26s16, q0s16);
+      __builtin_prefetch(s);
+      __builtin_prefetch(s + src_stride);
+      q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16, d22s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      __builtin_prefetch(s + src_stride * 2);
+      __builtin_prefetch(s + src_stride * 3);
+      q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
+      d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));
+
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+      d += dst_stride;
+
+      q8u16 = q10u16;
+      d18s16 = d22s16;
+      d19s16 = d24s16;
+      q10u16 = q13u16;
+      d22s16 = d25s16;
+    }
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/aom_convolve8_neon_asm.asm

@@ -0,0 +1,273 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; AV1_FILTER_WEIGHT == 128
+    ; AV1_FILTER_SHIFT == 7
+
+    EXPORT  |aom_convolve8_horiz_neon|
+    EXPORT  |aom_convolve8_vert_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|aom_convolve8_horiz_neon| PROC
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+aom_convolve8_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+aom_convolve8_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt aom_convolve8_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|aom_convolve8_vert_neon| PROC
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+aom_convolve8_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+aom_convolve8_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r7]
+    pld             [r4]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             aom_convolve8_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END

third_party/aom/aom_dsp/arm/aom_convolve_avg_neon.c

@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+void aom_convolve_avg_neon(const uint8_t *src,    // r0
+                           ptrdiff_t src_stride,  // r1
+                           uint8_t *dst,          // r2
+                           ptrdiff_t dst_stride,  // r3
+                           const int16_t *filter_x, int filter_x_stride,
+                           const int16_t *filter_y, int filter_y_stride, int w,
+                           int h) {
+  uint8_t *d;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  uint32x2_t d0u32, d2u32;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8, q8u8, q9u8, q10u8, q11u8;
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  d = dst;
+  if (w > 32) {  // avg64
+    for (; h > 0; h -= 1) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      q2u8 = vld1q_u8(src + 32);
+      q3u8 = vld1q_u8(src + 48);
+      src += src_stride;
+      q8u8 = vld1q_u8(d);
+      q9u8 = vld1q_u8(d + 16);
+      q10u8 = vld1q_u8(d + 32);
+      q11u8 = vld1q_u8(d + 48);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q8u8);
+      q1u8 = vrhaddq_u8(q1u8, q9u8);
+      q2u8 = vrhaddq_u8(q2u8, q10u8);
+      q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      vst1q_u8(dst + 32, q2u8);
+      vst1q_u8(dst + 48, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w == 32) {  // avg32
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q2u8 = vld1q_u8(src);
+      q3u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q8u8 = vld1q_u8(d);
+      q9u8 = vld1q_u8(d + 16);
+      d += dst_stride;
+      q10u8 = vld1q_u8(d);
+      q11u8 = vld1q_u8(d + 16);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q8u8);
+      q1u8 = vrhaddq_u8(q1u8, q9u8);
+      q2u8 = vrhaddq_u8(q2u8, q10u8);
+      q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q2u8);
+      vst1q_u8(dst + 16, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w > 8) {  // avg16
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      src += src_stride;
+      q1u8 = vld1q_u8(src);
+      src += src_stride;
+      q2u8 = vld1q_u8(d);
+      d += dst_stride;
+      q3u8 = vld1q_u8(d);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q2u8);
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      vst1q_u8(dst, q0u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q1u8);
+      dst += dst_stride;
+    }
+  } else if (w == 8) {  // avg8
+    for (; h > 0; h -= 2) {
+      d0u8 = vld1_u8(src);
+      src += src_stride;
+      d1u8 = vld1_u8(src);
+      src += src_stride;
+      d2u8 = vld1_u8(d);
+      d += dst_stride;
+      d3u8 = vld1_u8(d);
+      d += dst_stride;
+
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      q0u8 = vrhaddq_u8(q0u8, q1u8);
+
+      vst1_u8(dst, vget_low_u8(q0u8));
+      dst += dst_stride;
+      vst1_u8(dst, vget_high_u8(q0u8));
+      dst += dst_stride;
+    }
+  } else {  // avg4
+    for (; h > 0; h -= 2) {
+      d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 0);
+      src += src_stride;
+      d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 1);
+      src += src_stride;
+      d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+
+      d0u8 = vrhadd_u8(vreinterpret_u8_u32(d0u32), vreinterpret_u8_u32(d2u32));
+
+      d0u32 = vreinterpret_u32_u8(d0u8);
+      vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+      dst += dst_stride;
+      vst1_lane_u32((uint32_t *)dst, d0u32, 1);
+      dst += dst_stride;
+    }
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/aom_convolve_avg_neon_asm.asm

@@ -0,0 +1,119 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_convolve_avg_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|aom_convolve_avg_neon| PROC
+    push                {r4-r6, lr}
+    ldrd                r4, r5, [sp, #32]
+    mov                 r6, r2
+
+    cmp                 r4, #32
+    bgt                 avg64
+    beq                 avg32
+    cmp                 r4, #8
+    bgt                 avg16
+    beq                 avg8
+    b                   avg4
+
+avg64
+    sub                 lr, r1, #32
+    sub                 r4, r3, #32
+avg64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    pld                 [r2, r3]
+    vld1.8              {q8-q9},   [r6@128]!
+    vld1.8              {q10-q11}, [r6@128], r4
+    vrhadd.u8           q0, q0, q8
+    vrhadd.u8           q1, q1, q9
+    vrhadd.u8           q2, q2, q10
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r4
+    subs                r5, r5, #1
+    bgt                 avg64_h
+    pop                 {r4-r6, pc}
+
+avg32
+    vld1.8              {q0-q1}, [r0], r1
+    vld1.8              {q2-q3}, [r0], r1
+    vld1.8              {q8-q9},   [r6@128], r3
+    vld1.8              {q10-q11}, [r6@128], r3
+    pld                 [r0]
+    vrhadd.u8           q0, q0, q8
+    pld                 [r0, r1]
+    vrhadd.u8           q1, q1, q9
+    pld                 [r6]
+    vrhadd.u8           q2, q2, q10
+    pld                 [r6, r3]
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg32
+    pop                 {r4-r6, pc}
+
+avg16
+    vld1.8              {q0}, [r0], r1
+    vld1.8              {q1}, [r0], r1
+    vld1.8              {q2}, [r6@128], r3
+    vld1.8              {q3}, [r6@128], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q2
+    pld                 [r6]
+    pld                 [r6, r3]
+    vrhadd.u8           q1, q1, q3
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg16
+    pop                 {r4-r6, pc}
+
+avg8
+    vld1.8              {d0}, [r0], r1
+    vld1.8              {d1}, [r0], r1
+    vld1.8              {d2}, [r6@64], r3
+    vld1.8              {d3}, [r6@64], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q1
+    pld                 [r6]
+    pld                 [r6, r3]
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d1}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 avg8
+    pop                 {r4-r6, pc}
+
+avg4
+    vld1.32             {d0[0]}, [r0], r1
+    vld1.32             {d0[1]}, [r0], r1
+    vld1.32             {d2[0]}, [r6@32], r3
+    vld1.32             {d2[1]}, [r6@32], r3
+    vrhadd.u8           d0, d0, d2
+    vst1.32             {d0[0]}, [r2@32], r3
+    vst1.32             {d0[1]}, [r2@32], r3
+    subs                r5, r5, #2
+    bgt                 avg4
+    pop                 {r4-r6, pc}
+    ENDP
+
+    END

third_party/aom/aom_dsp/arm/aom_convolve_copy_neon.c

@@ -0,0 +1,93 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+void aom_convolve_copy_neon(const uint8_t *src,    // r0
+                            ptrdiff_t src_stride,  // r1
+                            uint8_t *dst,          // r2
+                            ptrdiff_t dst_stride,  // r3
+                            const int16_t *filter_x, int filter_x_stride,
+                            const int16_t *filter_y, int filter_y_stride, int w,
+                            int h) {
+  uint8x8_t d0u8, d2u8;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8;
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  if (w > 32) {  // copy64
+    for (; h > 0; h--) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      q2u8 = vld1q_u8(src + 32);
+      q3u8 = vld1q_u8(src + 48);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      vst1q_u8(dst + 32, q2u8);
+      vst1q_u8(dst + 48, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w == 32) {  // copy32
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q2u8 = vld1q_u8(src);
+      q3u8 = vld1q_u8(src + 16);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q2u8);
+      vst1q_u8(dst + 16, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w > 8) {  // copy16
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      src += src_stride;
+      q1u8 = vld1q_u8(src);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q1u8);
+      dst += dst_stride;
+    }
+  } else if (w == 8) {  // copy8
+    for (; h > 0; h -= 2) {
+      d0u8 = vld1_u8(src);
+      src += src_stride;
+      d2u8 = vld1_u8(src);
+      src += src_stride;
+
+      vst1_u8(dst, d0u8);
+      dst += dst_stride;
+      vst1_u8(dst, d2u8);
+      dst += dst_stride;
+    }
+  } else {  // copy4
+    for (; h > 0; h--) {
+      *(uint32_t *)dst = *(const uint32_t *)src;
+      src += src_stride;
+      dst += dst_stride;
+    }
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/aom_convolve_copy_neon_asm.asm

@@ -0,0 +1,87 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_convolve_copy_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|aom_convolve_copy_neon| PROC
+    push                {r4-r5, lr}
+    ldrd                r4, r5, [sp, #28]
+
+    cmp                 r4, #32
+    bgt                 copy64
+    beq                 copy32
+    cmp                 r4, #8
+    bgt                 copy16
+    beq                 copy8
+    b                   copy4
+
+copy64
+    sub                 lr, r1, #32
+    sub                 r3, r3, #32
+copy64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #1
+    bgt                 copy64_h
+    pop                 {r4-r5, pc}
+
+copy32
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q2-q3}, [r0], r1
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy32
+    pop                 {r4-r5, pc}
+
+copy16
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q1}, [r0], r1
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy16
+    pop                 {r4-r5, pc}
+
+copy8
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d2}, [r0], r1
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d2}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 copy8
+    pop                 {r4-r5, pc}
+
+copy4
+    ldr                 r12, [r0], r1
+    str                 r12, [r2], r3
+    subs                r5, r5, #1
+    bgt                 copy4
+    pop                 {r4-r5, pc}
+    ENDP
+
+    END

third_party/aom/aom_dsp/arm/aom_convolve_neon.c

@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+void aom_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                        ptrdiff_t dst_stride, const int16_t *filter_x,
+                        int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                        int w, int h) {
+  /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
+   * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
+   */
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+
+  // Account for the vertical phase needing 3 lines prior and 4 lines post
+  int intermediate_height = h + 7;
+
+  assert(y_step_q4 == 16);
+  assert(x_step_q4 == 16);
+
+  /* Filter starting 3 lines back. The neon implementation will ignore the
+   * given height and filter a multiple of 4 lines. Since this goes in to
+   * the temp buffer which has lots of extra room and is subsequently discarded
+   * this is safe if somewhat less than ideal.
+   */
+  aom_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, 64, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w,
+                           intermediate_height);
+
+  /* Step into the temp buffer 3 lines to get the actual frame data */
+  aom_convolve8_vert_neon(temp + 64 * 3, 64, dst, dst_stride, filter_x,
+                          x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void aom_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h) {
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+  int intermediate_height = h + 7;
+
+  assert(y_step_q4 == 16);
+  assert(x_step_q4 == 16);
+
+  /* This implementation has the same issues as above. In addition, we only want
+   * to average the values after both passes.
+   */
+  aom_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, 64, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w,
+                           intermediate_height);
+  aom_convolve8_avg_vert_neon(temp + 64 * 3, 64, dst, dst_stride, filter_x,
+                              x_step_q4, filter_y, y_step_q4, w, h);
+}

third_party/aom/aom_dsp/arm/avg_neon.c

@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+static INLINE unsigned int horizontal_add_u16x8(const uint16x8_t v_16x8) {
+  const uint32x4_t a = vpaddlq_u16(v_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+unsigned int aom_avg_4x4_neon(const uint8_t *s, int p) {
+  uint16x8_t v_sum;
+  uint32x2_t v_s0 = vdup_n_u32(0);
+  uint32x2_t v_s1 = vdup_n_u32(0);
+  v_s0 = vld1_lane_u32((const uint32_t *)s, v_s0, 0);
+  v_s0 = vld1_lane_u32((const uint32_t *)(s + p), v_s0, 1);
+  v_s1 = vld1_lane_u32((const uint32_t *)(s + 2 * p), v_s1, 0);
+  v_s1 = vld1_lane_u32((const uint32_t *)(s + 3 * p), v_s1, 1);
+  v_sum = vaddl_u8(vreinterpret_u8_u32(v_s0), vreinterpret_u8_u32(v_s1));
+  return (horizontal_add_u16x8(v_sum) + 8) >> 4;
+}
+
+unsigned int aom_avg_8x8_neon(const uint8_t *s, int p) {
+  uint8x8_t v_s0 = vld1_u8(s);
+  const uint8x8_t v_s1 = vld1_u8(s + p);
+  uint16x8_t v_sum = vaddl_u8(v_s0, v_s1);
+
+  v_s0 = vld1_u8(s + 2 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 3 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 4 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 5 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 6 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 7 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  return (horizontal_add_u16x8(v_sum) + 32) >> 6;
+}
+
+// coeff: 16 bits, dynamic range [-32640, 32640].
+// length: value range {16, 64, 256, 1024}.
+int aom_satd_neon(const int16_t *coeff, int length) {
+  const int16x4_t zero = vdup_n_s16(0);
+  int32x4_t accum = vdupq_n_s32(0);
+
+  do {
+    const int16x8_t src0 = vld1q_s16(coeff);
+    const int16x8_t src8 = vld1q_s16(coeff + 8);
+    accum = vabal_s16(accum, vget_low_s16(src0), zero);
+    accum = vabal_s16(accum, vget_high_s16(src0), zero);
+    accum = vabal_s16(accum, vget_low_s16(src8), zero);
+    accum = vabal_s16(accum, vget_high_s16(src8), zero);
+    length -= 16;
+    coeff += 16;
+  } while (length != 0);
+
+  {
+    // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
+    const int64x2_t s0 = vpaddlq_s32(accum);  // cascading summation of 'accum'.
+    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
+                                  vreinterpret_s32_s64(vget_high_s64(s0)));
+    const int satd = vget_lane_s32(s1, 0);
+    return satd;
+  }
+}
+
+void aom_int_pro_row_neon(int16_t hbuf[16], uint8_t const *ref, int ref_stride,
+                          int height) {
+  int i;
+  uint16x8_t vec_sum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_hi = vdupq_n_u16(0);
+  const int shift_factor = ((height >> 5) + 3) * -1;
+  const int16x8_t vec_shift = vdupq_n_s16(shift_factor);
+
+  for (i = 0; i < height; i += 8) {
+    const uint8x16_t vec_row1 = vld1q_u8(ref);
+    const uint8x16_t vec_row2 = vld1q_u8(ref + ref_stride);
+    const uint8x16_t vec_row3 = vld1q_u8(ref + ref_stride * 2);
+    const uint8x16_t vec_row4 = vld1q_u8(ref + ref_stride * 3);
+    const uint8x16_t vec_row5 = vld1q_u8(ref + ref_stride * 4);
+    const uint8x16_t vec_row6 = vld1q_u8(ref + ref_stride * 5);
+    const uint8x16_t vec_row7 = vld1q_u8(ref + ref_stride * 6);
+    const uint8x16_t vec_row8 = vld1q_u8(ref + ref_stride * 7);
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row1));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row1));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row2));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row2));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row3));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row3));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row4));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row4));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row5));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row5));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row6));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row6));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row7));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row7));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row8));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row8));
+
+    ref += ref_stride * 8;
+  }
+
+  vec_sum_lo = vshlq_u16(vec_sum_lo, vec_shift);
+  vec_sum_hi = vshlq_u16(vec_sum_hi, vec_shift);
+
+  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_lo));
+  hbuf += 8;
+  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_hi));
+}
+
+int16_t aom_int_pro_col_neon(uint8_t const *ref, const int width) {
+  int i;
+  uint16x8_t vec_sum = vdupq_n_u16(0);
+
+  for (i = 0; i < width; i += 16) {
+    const uint8x16_t vec_row = vld1q_u8(ref);
+    vec_sum = vaddw_u8(vec_sum, vget_low_u8(vec_row));
+    vec_sum = vaddw_u8(vec_sum, vget_high_u8(vec_row));
+    ref += 16;
+  }
+
+  return horizontal_add_u16x8(vec_sum);
+}
+
+// ref, src = [0, 510] - max diff = 16-bits
+// bwl = {2, 3, 4}, width = {16, 32, 64}
+int aom_vector_var_neon(int16_t const *ref, int16_t const *src, int bwl) {
+  int width = 4 << bwl;
+  int32x4_t sse = vdupq_n_s32(0);
+  int16x8_t total = vdupq_n_s16(0);
+
+  assert(width >= 8);
+  assert((width % 8) == 0);
+
+  do {
+    const int16x8_t r = vld1q_s16(ref);
+    const int16x8_t s = vld1q_s16(src);
+    const int16x8_t diff = vsubq_s16(r, s);  // [-510, 510], 10 bits.
+    const int16x4_t diff_lo = vget_low_s16(diff);
+    const int16x4_t diff_hi = vget_high_s16(diff);
+    sse = vmlal_s16(sse, diff_lo, diff_lo);  // dynamic range 26 bits.
+    sse = vmlal_s16(sse, diff_hi, diff_hi);
+    total = vaddq_s16(total, diff);  // dynamic range 16 bits.
+
+    ref += 8;
+    src += 8;
+    width -= 8;
+  } while (width != 0);
+
+  {
+    // Note: 'total''s pairwise addition could be implemented similarly to
+    // horizontal_add_u16x8(), but one less vpaddl with 'total' when paired
+    // with the summation of 'sse' performed better on a Cortex-A15.
+    const int32x4_t t0 = vpaddlq_s16(total);  // cascading summation of 'total'
+    const int32x2_t t1 = vadd_s32(vget_low_s32(t0), vget_high_s32(t0));
+    const int32x2_t t2 = vpadd_s32(t1, t1);
+    const int t = vget_lane_s32(t2, 0);
+    const int64x2_t s0 = vpaddlq_s32(sse);  // cascading summation of 'sse'.
+    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
+                                  vreinterpret_s32_s64(vget_high_s64(s0)));
+    const int s = vget_lane_s32(s1, 0);
+    const int shift_factor = bwl + 2;
+    return s - ((t * t) >> shift_factor);
+  }
+}
+
+void aom_minmax_8x8_neon(const uint8_t *a, int a_stride, const uint8_t *b,
+                         int b_stride, int *min, int *max) {
+  // Load and concatenate.
+  const uint8x16_t a01 = vcombine_u8(vld1_u8(a), vld1_u8(a + a_stride));
+  const uint8x16_t a23 =
+      vcombine_u8(vld1_u8(a + 2 * a_stride), vld1_u8(a + 3 * a_stride));
+  const uint8x16_t a45 =
+      vcombine_u8(vld1_u8(a + 4 * a_stride), vld1_u8(a + 5 * a_stride));
+  const uint8x16_t a67 =
+      vcombine_u8(vld1_u8(a + 6 * a_stride), vld1_u8(a + 7 * a_stride));
+
+  const uint8x16_t b01 = vcombine_u8(vld1_u8(b), vld1_u8(b + b_stride));
+  const uint8x16_t b23 =
+      vcombine_u8(vld1_u8(b + 2 * b_stride), vld1_u8(b + 3 * b_stride));
+  const uint8x16_t b45 =
+      vcombine_u8(vld1_u8(b + 4 * b_stride), vld1_u8(b + 5 * b_stride));
+  const uint8x16_t b67 =
+      vcombine_u8(vld1_u8(b + 6 * b_stride), vld1_u8(b + 7 * b_stride));
+
+  // Absolute difference.
+  const uint8x16_t ab01_diff = vabdq_u8(a01, b01);
+  const uint8x16_t ab23_diff = vabdq_u8(a23, b23);
+  const uint8x16_t ab45_diff = vabdq_u8(a45, b45);
+  const uint8x16_t ab67_diff = vabdq_u8(a67, b67);
+
+  // Max values between the Q vectors.
+  const uint8x16_t ab0123_max = vmaxq_u8(ab01_diff, ab23_diff);
+  const uint8x16_t ab4567_max = vmaxq_u8(ab45_diff, ab67_diff);
+  const uint8x16_t ab0123_min = vminq_u8(ab01_diff, ab23_diff);
+  const uint8x16_t ab4567_min = vminq_u8(ab45_diff, ab67_diff);
+
+  const uint8x16_t ab07_max = vmaxq_u8(ab0123_max, ab4567_max);
+  const uint8x16_t ab07_min = vminq_u8(ab0123_min, ab4567_min);
+
+  // Split to D and start doing pairwise.
+  uint8x8_t ab_max = vmax_u8(vget_high_u8(ab07_max), vget_low_u8(ab07_max));
+  uint8x8_t ab_min = vmin_u8(vget_high_u8(ab07_min), vget_low_u8(ab07_min));
+
+  // Enough runs of vpmax/min propogate the max/min values to every position.
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  *min = *max = 0;  // Clear high bits
+  // Store directly to avoid costly neon->gpr transfer.
+  vst1_lane_u8((uint8_t *)max, ab_max, 0);
+  vst1_lane_u8((uint8_t *)min, ab_min, 0);
+}

third_party/aom/aom_dsp/arm/bilinear_filter_media.asm

@@ -0,0 +1,240 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_filter_block2d_bil_first_pass_media|
+    EXPORT  |aom_filter_block2d_bil_second_pass_media|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+;-------------------------------------
+; r0    unsigned char  *src_ptr,
+; r1    unsigned short *dst_ptr,
+; r2    unsigned int    src_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *aom_filter
+;-------------------------------------
+; The output is transposed stroed in output array to make it easy for second pass filtering.
+|aom_filter_block2d_bil_first_pass_media| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; aom_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    mov     r12, r3                         ; outer-loop counter
+
+    add     r7, r2, r4                      ; preload next row
+    pld     [r0, r7]
+
+    sub     r2, r2, r4                      ; src increment for height loop
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+
+    mov     r3, r3, lsl #1                  ; height*2
+    add     r3, r3, #2                      ; plus 2 to make output buffer 4-bit aligned since height is actually (height+1)
+
+    mov     r11, r1                         ; save dst_ptr for each row
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_1st_filter
+
+|bil_height_loop_1st_v6|
+    ldrb    r6, [r0]                        ; load source data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    mov     lr, r4, lsr #2                  ; 4-in-parellel loop counter
+
+|bil_width_loop_1st_v6|
+    ldrb    r9, [r0, #3]
+    ldrb    r10, [r0, #4]
+
+    pkhbt   r6, r6, r7, lsl #16             ; src[1] | src[0]
+    pkhbt   r7, r7, r8, lsl #16             ; src[2] | src[1]
+
+    smuad   r6, r6, r5                      ; apply the filter
+    pkhbt   r8, r8, r9, lsl #16             ; src[3] | src[2]
+    smuad   r7, r7, r5
+    pkhbt   r9, r9, r10, lsl #16            ; src[4] | src[3]
+
+    smuad   r8, r8, r5
+    smuad   r9, r9, r5
+
+    add     r0, r0, #4
+    subs    lr, lr, #1
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #16, r6, asr #7
+    usat    r7, #16, r7, asr #7
+
+    strh    r6, [r1], r3                    ; result is transposed and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strh    r7, [r1], r3
+    add     r9, r9, #0x40
+    usat    r8, #16, r8, asr #7
+    usat    r9, #16, r9, asr #7
+
+    strh    r8, [r1], r3                    ; result is transposed and stored
+
+    ldrneb  r6, [r0]                        ; load source data
+    strh    r9, [r1], r3
+
+    ldrneb  r7, [r0, #1]
+    ldrneb  r8, [r0, #2]
+
+    bne     bil_width_loop_1st_v6
+
+    add     r0, r0, r2                      ; move to next input row
+    subs    r12, r12, #1
+
+    add     r9, r2, r4, lsl #1              ; adding back block width
+    pld     [r0, r9]                        ; preload next row
+
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_1st_v6
+
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_1st_filter|
+|bil_height_loop_null_1st|
+    mov     lr, r4, lsr #2                  ; loop counter
+
+|bil_width_loop_null_1st|
+    ldrb    r6, [r0]                        ; load data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    ldrb    r9, [r0, #3]
+
+    strh    r6, [r1], r3                    ; store it to immediate buffer
+    add     r0, r0, #4
+    strh    r7, [r1], r3
+    subs    lr, lr, #1
+    strh    r8, [r1], r3
+    strh    r9, [r1], r3
+
+    bne     bil_width_loop_null_1st
+
+    subs    r12, r12, #1
+    add     r0, r0, r2                      ; move to next input line
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_null_1st
+
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP  ; |aom_filter_block2d_bil_first_pass_media|
+
+
+;---------------------------------
+; r0    unsigned short *src_ptr,
+; r1    unsigned char  *dst_ptr,
+; r2    int             dst_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *aom_filter
+;---------------------------------
+|aom_filter_block2d_bil_second_pass_media| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; aom_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+    mov     r12, r4                         ; outer-loop counter = width, since we work on transposed data matrix
+    mov     r11, r1
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_2nd_filter
+
+|bil_height_loop_2nd|
+    ldr     r6, [r0]                        ; load the data
+    ldr     r8, [r0, #4]
+    ldrh    r10, [r0, #8]
+    mov     lr, r3, lsr #2                  ; loop counter
+
+|bil_width_loop_2nd|
+    pkhtb   r7, r6, r8                      ; src[1] | src[2]
+    pkhtb   r9, r8, r10                     ; src[3] | src[4]
+
+    smuad   r6, r6, r5                      ; apply filter
+    smuad   r8, r8, r5                      ; apply filter
+
+    subs    lr, lr, #1
+
+    smuadx  r7, r7, r5                      ; apply filter
+    smuadx  r9, r9, r5                      ; apply filter
+
+    add     r0, r0, #8
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #8, r6, asr #7
+    usat    r7, #8, r7, asr #7
+    strb    r6, [r1], r2                    ; the result is transposed back and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strb    r7, [r1], r2
+    add     r9, r9, #0x40
+    usat    r8, #8, r8, asr #7
+    usat    r9, #8, r9, asr #7
+    strb    r8, [r1], r2                    ; the result is transposed back and stored
+
+    ldrne   r6, [r0]                        ; load data
+    strb    r9, [r1], r2
+    ldrne   r8, [r0, #4]
+    ldrneh  r10, [r0, #8]
+
+    bne     bil_width_loop_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4                      ; update src for next row
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_2nd
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_2nd_filter|
+|bil_height_loop_null_2nd|
+    mov     lr, r3, lsr #2
+
+|bil_width_loop_null_2nd|
+    ldr     r6, [r0], #4                    ; load data
+    subs    lr, lr, #1
+    ldr     r8, [r0], #4
+
+    strb    r6, [r1], r2                    ; store data
+    mov     r7, r6, lsr #16
+    strb    r7, [r1], r2
+    mov     r9, r8, lsr #16
+    strb    r8, [r1], r2
+    strb    r9, [r1], r2
+
+    bne     bil_width_loop_null_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_null_2nd
+
+    ldmia   sp!, {r4 - r11, pc}
+    ENDP  ; |aom_filter_block2d_second_pass_media|
+
+    END

third_party/aom/aom_dsp/arm/fwd_txfm_neon.c

@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+
+void aom_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
+  int i;
+  // stage 1
+  int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
+  int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
+  int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
+  int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
+  int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
+  int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
+  int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
+  int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
+  for (i = 0; i < 2; ++i) {
+    int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
+    const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
+    const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
+    const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
+    const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
+    const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
+    const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
+    const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
+    const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
+    // fdct4(step, step);
+    int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
+    int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
+    int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
+    int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
+    // fdct4(step, step);
+    int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
+    int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
+    v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
+    v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
+    v_t0_lo = vmulq_n_s32(v_t0_lo, (int32_t)cospi_16_64);
+    v_t0_hi = vmulq_n_s32(v_t0_hi, (int32_t)cospi_16_64);
+    v_t1_lo = vmulq_n_s32(v_t1_lo, (int32_t)cospi_16_64);
+    v_t1_hi = vmulq_n_s32(v_t1_hi, (int32_t)cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_0 = vcombine_s16(a, c);  // 00 01 02 03 40 41 42 43
+      out_2 = vcombine_s16(e, g);  // 20 21 22 23 60 61 62 63
+      out_4 = vcombine_s16(b, d);  // 04 05 06 07 44 45 46 47
+      out_6 = vcombine_s16(f, h);  // 24 25 26 27 64 65 66 67
+    }
+    // Stage 2
+    v_x0 = vsubq_s16(v_s6, v_s5);
+    v_x1 = vaddq_s16(v_s6, v_s5);
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x8_t ab = vcombine_s16(a, b);
+      const int16x8_t cd = vcombine_s16(c, d);
+      // Stage 3
+      v_x0 = vaddq_s16(v_s4, ab);
+      v_x1 = vsubq_s16(v_s4, ab);
+      v_x2 = vsubq_s16(v_s7, cd);
+      v_x3 = vaddq_s16(v_s7, cd);
+    }
+    // Stage 4
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
+    v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
+    v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
+    v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
+    v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
+    v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_1 = vcombine_s16(a, c);  // 10 11 12 13 50 51 52 53
+      out_3 = vcombine_s16(e, g);  // 30 31 32 33 70 71 72 73
+      out_5 = vcombine_s16(b, d);  // 14 15 16 17 54 55 56 57
+      out_7 = vcombine_s16(f, h);  // 34 35 36 37 74 75 76 77
+    }
+    // transpose 8x8
+    {
+      // 00 01 02 03 40 41 42 43
+      // 10 11 12 13 50 51 52 53
+      // 20 21 22 23 60 61 62 63
+      // 30 31 32 33 70 71 72 73
+      // 04 05 06 07 44 45 46 47
+      // 14 15 16 17 54 55 56 57
+      // 24 25 26 27 64 65 66 67
+      // 34 35 36 37 74 75 76 77
+      const int32x4x2_t r02_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_0), vreinterpretq_s32_s16(out_2));
+      const int32x4x2_t r13_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_1), vreinterpretq_s32_s16(out_3));
+      const int32x4x2_t r46_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_4), vreinterpretq_s32_s16(out_6));
+      const int32x4x2_t r57_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_5), vreinterpretq_s32_s16(out_7));
+      const int16x8x2_t r01_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
+                    vreinterpretq_s16_s32(r13_s32.val[0]));
+      const int16x8x2_t r23_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
+                    vreinterpretq_s16_s32(r13_s32.val[1]));
+      const int16x8x2_t r45_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
+                    vreinterpretq_s16_s32(r57_s32.val[0]));
+      const int16x8x2_t r67_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
+                    vreinterpretq_s16_s32(r57_s32.val[1]));
+      input_0 = r01_s16.val[0];
+      input_1 = r01_s16.val[1];
+      input_2 = r23_s16.val[0];
+      input_3 = r23_s16.val[1];
+      input_4 = r45_s16.val[0];
+      input_5 = r45_s16.val[1];
+      input_6 = r67_s16.val[0];
+      input_7 = r67_s16.val[1];
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }  // for
+  {
+    // from aom_dct_sse2.c
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
+    const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
+    const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
+    const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
+    const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
+    const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
+    const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
+    const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
+    input_0 = vhsubq_s16(input_0, sign_in0);
+    input_1 = vhsubq_s16(input_1, sign_in1);
+    input_2 = vhsubq_s16(input_2, sign_in2);
+    input_3 = vhsubq_s16(input_3, sign_in3);
+    input_4 = vhsubq_s16(input_4, sign_in4);
+    input_5 = vhsubq_s16(input_5, sign_in5);
+    input_6 = vhsubq_s16(input_6, sign_in6);
+    input_7 = vhsubq_s16(input_7, sign_in7);
+    // store results
+    vst1q_s16(&final_output[0 * 8], input_0);
+    vst1q_s16(&final_output[1 * 8], input_1);
+    vst1q_s16(&final_output[2 * 8], input_2);
+    vst1q_s16(&final_output[3 * 8], input_3);
+    vst1q_s16(&final_output[4 * 8], input_4);
+    vst1q_s16(&final_output[5 * 8], input_5);
+    vst1q_s16(&final_output[6 * 8], input_6);
+    vst1q_s16(&final_output[7 * 8], input_7);
+  }
+}
+
+void aom_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride) {
+  int r;
+  int16x8_t sum = vld1q_s16(&input[0]);
+  for (r = 1; r < 8; ++r) {
+    const int16x8_t input_00 = vld1q_s16(&input[r * stride]);
+    sum = vaddq_s16(sum, input_00);
+  }
+  {
+    const int32x4_t a = vpaddlq_s16(sum);
+    const int64x2_t b = vpaddlq_s32(a);
+    const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                                 vreinterpret_s32_s64(vget_high_s64(b)));
+    output[0] = vget_lane_s16(vreinterpret_s16_s32(c), 0);
+    output[1] = 0;
+  }
+}

third_party/aom/aom_dsp/arm/hadamard_neon.c

@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+
+static void hadamard8x8_one_pass(int16x8_t *a0, int16x8_t *a1, int16x8_t *a2,
+                                 int16x8_t *a3, int16x8_t *a4, int16x8_t *a5,
+                                 int16x8_t *a6, int16x8_t *a7) {
+  const int16x8_t b0 = vaddq_s16(*a0, *a1);
+  const int16x8_t b1 = vsubq_s16(*a0, *a1);
+  const int16x8_t b2 = vaddq_s16(*a2, *a3);
+  const int16x8_t b3 = vsubq_s16(*a2, *a3);
+  const int16x8_t b4 = vaddq_s16(*a4, *a5);
+  const int16x8_t b5 = vsubq_s16(*a4, *a5);
+  const int16x8_t b6 = vaddq_s16(*a6, *a7);
+  const int16x8_t b7 = vsubq_s16(*a6, *a7);
+
+  const int16x8_t c0 = vaddq_s16(b0, b2);
+  const int16x8_t c1 = vaddq_s16(b1, b3);
+  const int16x8_t c2 = vsubq_s16(b0, b2);
+  const int16x8_t c3 = vsubq_s16(b1, b3);
+  const int16x8_t c4 = vaddq_s16(b4, b6);
+  const int16x8_t c5 = vaddq_s16(b5, b7);
+  const int16x8_t c6 = vsubq_s16(b4, b6);
+  const int16x8_t c7 = vsubq_s16(b5, b7);
+
+  *a0 = vaddq_s16(c0, c4);
+  *a1 = vsubq_s16(c2, c6);
+  *a2 = vsubq_s16(c0, c4);
+  *a3 = vaddq_s16(c2, c6);
+  *a4 = vaddq_s16(c3, c7);
+  *a5 = vsubq_s16(c3, c7);
+  *a6 = vsubq_s16(c1, c5);
+  *a7 = vaddq_s16(c1, c5);
+}
+
+// TODO(johannkoenig): Make a transpose library and dedup with idct. Consider
+// reversing transpose order which may make it easier for the compiler to
+// reconcile the vtrn.64 moves.
+static void transpose8x8(int16x8_t *a0, int16x8_t *a1, int16x8_t *a2,
+                         int16x8_t *a3, int16x8_t *a4, int16x8_t *a5,
+                         int16x8_t *a6, int16x8_t *a7) {
+  // Swap 64 bit elements. Goes from:
+  // a0: 00 01 02 03 04 05 06 07
+  // a1: 08 09 10 11 12 13 14 15
+  // a2: 16 17 18 19 20 21 22 23
+  // a3: 24 25 26 27 28 29 30 31
+  // a4: 32 33 34 35 36 37 38 39
+  // a5: 40 41 42 43 44 45 46 47
+  // a6: 48 49 50 51 52 53 54 55
+  // a7: 56 57 58 59 60 61 62 63
+  // to:
+  // a04_lo: 00 01 02 03 32 33 34 35
+  // a15_lo: 08 09 10 11 40 41 42 43
+  // a26_lo: 16 17 18 19 48 49 50 51
+  // a37_lo: 24 25 26 27 56 57 58 59
+  // a04_hi: 04 05 06 07 36 37 38 39
+  // a15_hi: 12 13 14 15 44 45 46 47
+  // a26_hi: 20 21 22 23 52 53 54 55
+  // a37_hi: 28 29 30 31 60 61 62 63
+  const int16x8_t a04_lo = vcombine_s16(vget_low_s16(*a0), vget_low_s16(*a4));
+  const int16x8_t a15_lo = vcombine_s16(vget_low_s16(*a1), vget_low_s16(*a5));
+  const int16x8_t a26_lo = vcombine_s16(vget_low_s16(*a2), vget_low_s16(*a6));
+  const int16x8_t a37_lo = vcombine_s16(vget_low_s16(*a3), vget_low_s16(*a7));
+  const int16x8_t a04_hi = vcombine_s16(vget_high_s16(*a0), vget_high_s16(*a4));
+  const int16x8_t a15_hi = vcombine_s16(vget_high_s16(*a1), vget_high_s16(*a5));
+  const int16x8_t a26_hi = vcombine_s16(vget_high_s16(*a2), vget_high_s16(*a6));
+  const int16x8_t a37_hi = vcombine_s16(vget_high_s16(*a3), vget_high_s16(*a7));
+
+  // Swap 32 bit elements resulting in:
+  // a0246_lo:
+  // 00 01 16 17 32 33 48 49
+  // 02 03 18 19 34 35 50 51
+  // a1357_lo:
+  // 08 09 24 25 40 41 56 57
+  // 10 11 26 27 42 43 58 59
+  // a0246_hi:
+  // 04 05 20 21 36 37 52 53
+  // 06 07 22 23 38 39 54 55
+  // a1657_hi:
+  // 12 13 28 29 44 45 60 61
+  // 14 15 30 31 46 47 62 63
+  const int32x4x2_t a0246_lo =
+      vtrnq_s32(vreinterpretq_s32_s16(a04_lo), vreinterpretq_s32_s16(a26_lo));
+  const int32x4x2_t a1357_lo =
+      vtrnq_s32(vreinterpretq_s32_s16(a15_lo), vreinterpretq_s32_s16(a37_lo));
+  const int32x4x2_t a0246_hi =
+      vtrnq_s32(vreinterpretq_s32_s16(a04_hi), vreinterpretq_s32_s16(a26_hi));
+  const int32x4x2_t a1357_hi =
+      vtrnq_s32(vreinterpretq_s32_s16(a15_hi), vreinterpretq_s32_s16(a37_hi));
+
+  // Swap 16 bit elements resulting in:
+  // b0:
+  // 00 08 16 24 32 40 48 56
+  // 01 09 17 25 33 41 49 57
+  // b1:
+  // 02 10 18 26 34 42 50 58
+  // 03 11 19 27 35 43 51 59
+  // b2:
+  // 04 12 20 28 36 44 52 60
+  // 05 13 21 29 37 45 53 61
+  // b3:
+  // 06 14 22 30 38 46 54 62
+  // 07 15 23 31 39 47 55 63
+  const int16x8x2_t b0 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[0]),
+                                   vreinterpretq_s16_s32(a1357_lo.val[0]));
+  const int16x8x2_t b1 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[1]),
+                                   vreinterpretq_s16_s32(a1357_lo.val[1]));
+  const int16x8x2_t b2 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[0]),
+                                   vreinterpretq_s16_s32(a1357_hi.val[0]));
+  const int16x8x2_t b3 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[1]),
+                                   vreinterpretq_s16_s32(a1357_hi.val[1]));
+
+  *a0 = b0.val[0];
+  *a1 = b0.val[1];
+  *a2 = b1.val[0];
+  *a3 = b1.val[1];
+  *a4 = b2.val[0];
+  *a5 = b2.val[1];
+  *a6 = b3.val[0];
+  *a7 = b3.val[1];
+}
+
+void aom_hadamard_8x8_neon(const int16_t *src_diff, int src_stride,
+                           int16_t *coeff) {
+  int16x8_t a0 = vld1q_s16(src_diff);
+  int16x8_t a1 = vld1q_s16(src_diff + src_stride);
+  int16x8_t a2 = vld1q_s16(src_diff + 2 * src_stride);
+  int16x8_t a3 = vld1q_s16(src_diff + 3 * src_stride);
+  int16x8_t a4 = vld1q_s16(src_diff + 4 * src_stride);
+  int16x8_t a5 = vld1q_s16(src_diff + 5 * src_stride);
+  int16x8_t a6 = vld1q_s16(src_diff + 6 * src_stride);
+  int16x8_t a7 = vld1q_s16(src_diff + 7 * src_stride);
+
+  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  transpose8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  // Skip the second transpose because it is not required.
+
+  vst1q_s16(coeff + 0, a0);
+  vst1q_s16(coeff + 8, a1);
+  vst1q_s16(coeff + 16, a2);
+  vst1q_s16(coeff + 24, a3);
+  vst1q_s16(coeff + 32, a4);
+  vst1q_s16(coeff + 40, a5);
+  vst1q_s16(coeff + 48, a6);
+  vst1q_s16(coeff + 56, a7);
+}
+
+void aom_hadamard_16x16_neon(const int16_t *src_diff, int src_stride,
+                             int16_t *coeff) {
+  int i;
+
+  /* Rearrange 16x16 to 8x32 and remove stride.
+   * Top left first. */
+  aom_hadamard_8x8_neon(src_diff + 0 + 0 * src_stride, src_stride, coeff + 0);
+  /* Top right. */
+  aom_hadamard_8x8_neon(src_diff + 8 + 0 * src_stride, src_stride, coeff + 64);
+  /* Bottom left. */
+  aom_hadamard_8x8_neon(src_diff + 0 + 8 * src_stride, src_stride, coeff + 128);
+  /* Bottom right. */
+  aom_hadamard_8x8_neon(src_diff + 8 + 8 * src_stride, src_stride, coeff + 192);
+
+  for (i = 0; i < 64; i += 8) {
+    const int16x8_t a0 = vld1q_s16(coeff + 0);
+    const int16x8_t a1 = vld1q_s16(coeff + 64);
+    const int16x8_t a2 = vld1q_s16(coeff + 128);
+    const int16x8_t a3 = vld1q_s16(coeff + 192);
+
+    const int16x8_t b0 = vhaddq_s16(a0, a1);
+    const int16x8_t b1 = vhsubq_s16(a0, a1);
+    const int16x8_t b2 = vhaddq_s16(a2, a3);
+    const int16x8_t b3 = vhsubq_s16(a2, a3);
+
+    const int16x8_t c0 = vaddq_s16(b0, b2);
+    const int16x8_t c1 = vaddq_s16(b1, b3);
+    const int16x8_t c2 = vsubq_s16(b0, b2);
+    const int16x8_t c3 = vsubq_s16(b1, b3);
+
+    vst1q_s16(coeff + 0, c0);
+    vst1q_s16(coeff + 64, c1);
+    vst1q_s16(coeff + 128, c2);
+    vst1q_s16(coeff + 192, c3);
+
+    coeff += 8;
+  }
+}

third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.asm

@@ -0,0 +1,201 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+
+    EXPORT  |aom_idct16x16_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_idct16x16_1_add_neon(int16_t *input, uint8_t *dest,
+;                                    int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct16x16_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asr              r0, r0, #6                ; >> 6
+
+    vdup.s16         q0, r0                    ; duplicate a1
+    mov              r0, #8
+    sub              r2, #8
+
+    ; load destination data row0 - row3
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row4 - row7
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row8 - row11
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row12 - row15
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |aom_idct16x16_1_add_neon|
+
+    END

third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.c

@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+void aom_idct16x16_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d2u8, d3u8, d30u8, d31u8;
+  uint64x1_t d2u64, d3u64, d4u64, d5u64;
+  uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+  int16x8_t q0s16;
+  uint8_t *d1, *d2;
+  int16_t i, j, a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  q0s16 = vdupq_n_s16(a1);
+  q0u16 = vreinterpretq_u16_s16(q0s16);
+
+  for (d1 = d2 = dest, i = 0; i < 4; i++) {
+    for (j = 0; j < 2; j++) {
+      d2u64 = vld1_u64((const uint64_t *)d1);
+      d3u64 = vld1_u64((const uint64_t *)(d1 + 8));
+      d1 += dest_stride;
+      d4u64 = vld1_u64((const uint64_t *)d1);
+      d5u64 = vld1_u64((const uint64_t *)(d1 + 8));
+      d1 += dest_stride;
+
+      q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+      q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+      q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+      q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+      d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+      d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+      d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+      d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+      vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d3u8));
+      d2 += dest_stride;
+      vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+      vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d31u8));
+      d2 += dest_stride;
+    }
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/idct16x16_neon.c

@@ -0,0 +1,152 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_dsp_common.h"
+
+void aom_idct16x16_256_add_neon_pass1(const int16_t *input, int16_t *output,
+                                      int output_stride);
+void aom_idct16x16_256_add_neon_pass2(const int16_t *src, int16_t *output,
+                                      int16_t *pass1Output, int16_t skip_adding,
+                                      uint8_t *dest, int dest_stride);
+void aom_idct16x16_10_add_neon_pass1(const int16_t *input, int16_t *output,
+                                     int output_stride);
+void aom_idct16x16_10_add_neon_pass2(const int16_t *src, int16_t *output,
+                                     int16_t *pass1Output, int16_t skip_adding,
+                                     uint8_t *dest, int dest_stride);
+
+#if HAVE_NEON_ASM
+/* For ARM NEON, d8-d15 are callee-saved registers, and need to be saved. */
+extern void aom_push_neon(int64_t *store);
+extern void aom_pop_neon(int64_t *store);
+#endif  // HAVE_NEON_ASM
+
+void aom_idct16x16_256_add_neon(const int16_t *input, uint8_t *dest,
+                                int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16 * 16] = { 0 };
+  int16_t row_idct_output[16 * 16] = { 0 };
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  aom_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  aom_idct16x16_256_add_neon_pass2(input + 1, row_idct_output, pass1_output, 0,
+                                   dest, dest_stride);
+
+  /* Parallel idct on the lower 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(input + 8 * 16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  aom_idct16x16_256_add_neon_pass2(input + 8 * 16 + 1, row_idct_output + 8,
+                                   pass1_output, 0, dest, dest_stride);
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 1, row_idct_output,
+                                   pass1_output, 1, dest, dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output + 8 * 16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 8 * 16 + 1,
+                                   row_idct_output + 8, pass1_output, 1,
+                                   dest + 8, dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  aom_pop_neon(store_reg);
+#endif
+
+  return;
+}
+
+void aom_idct16x16_10_add_neon(const int16_t *input, uint8_t *dest,
+                               int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16 * 16] = { 0 };
+  int16_t row_idct_output[16 * 16] = { 0 };
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  aom_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_10_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  aom_idct16x16_10_add_neon_pass2(input + 1, row_idct_output, pass1_output, 0,
+                                  dest, dest_stride);
+
+  /* Skip Parallel idct on the lower 8 rows as they are all 0s */
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 1, row_idct_output,
+                                   pass1_output, 1, dest, dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output + 8 * 16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 8 * 16 + 1,
+                                   row_idct_output + 8, pass1_output, 1,
+                                   dest + 8, dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  aom_pop_neon(store_reg);
+#endif
+
+  return;
+}

third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.asm

@@ -0,0 +1,147 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+    EXPORT  |aom_idct32x32_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ;TODO(hkuang): put the following macros in a seperate
+    ;file so other idct function could also use them.
+    MACRO
+    LD_16x8          $src, $stride
+    vld1.8           {q8}, [$src], $stride
+    vld1.8           {q9}, [$src], $stride
+    vld1.8           {q10}, [$src], $stride
+    vld1.8           {q11}, [$src], $stride
+    vld1.8           {q12}, [$src], $stride
+    vld1.8           {q13}, [$src], $stride
+    vld1.8           {q14}, [$src], $stride
+    vld1.8           {q15}, [$src], $stride
+    MEND
+
+    MACRO
+    ADD_DIFF_16x8    $diff
+    vqadd.u8         q8, q8, $diff
+    vqadd.u8         q9, q9, $diff
+    vqadd.u8         q10, q10, $diff
+    vqadd.u8         q11, q11, $diff
+    vqadd.u8         q12, q12, $diff
+    vqadd.u8         q13, q13, $diff
+    vqadd.u8         q14, q14, $diff
+    vqadd.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    SUB_DIFF_16x8    $diff
+    vqsub.u8         q8, q8, $diff
+    vqsub.u8         q9, q9, $diff
+    vqsub.u8         q10, q10, $diff
+    vqsub.u8         q11, q11, $diff
+    vqsub.u8         q12, q12, $diff
+    vqsub.u8         q13, q13, $diff
+    vqsub.u8         q14, q14, $diff
+    vqsub.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    ST_16x8          $dst, $stride
+    vst1.8           {q8}, [$dst], $stride
+    vst1.8           {q9}, [$dst], $stride
+    vst1.8           {q10},[$dst], $stride
+    vst1.8           {q11},[$dst], $stride
+    vst1.8           {q12},[$dst], $stride
+    vst1.8           {q13},[$dst], $stride
+    vst1.8           {q14},[$dst], $stride
+    vst1.8           {q15},[$dst], $stride
+    MEND
+
+;void aom_idct32x32_1_add_neon(int16_t *input, uint8_t *dest,
+;                              int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride
+
+|aom_idct32x32_1_add_neon| PROC
+    push             {lr}
+    pld              [r1]
+    add              r3, r1, #16               ; r3 dest + 16 for second loop
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asrs             r0, r0, #6                ; >> 6
+    bge              diff_positive_32_32
+
+diff_negative_32_32
+    neg              r0, r0
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_negative_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_negative_32_32_loop
+    pop              {pc}
+
+diff_positive_32_32
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_positive_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_positive_32_32_loop
+    pop              {pc}
+
+    ENDP             ; |aom_idct32x32_1_add_neon|
+    END

third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.c

@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+static INLINE void LD_16x8(uint8_t *d, int d_stride, uint8x16_t *q8u8,
+                           uint8x16_t *q9u8, uint8x16_t *q10u8,
+                           uint8x16_t *q11u8, uint8x16_t *q12u8,
+                           uint8x16_t *q13u8, uint8x16_t *q14u8,
+                           uint8x16_t *q15u8) {
+  *q8u8 = vld1q_u8(d);
+  d += d_stride;
+  *q9u8 = vld1q_u8(d);
+  d += d_stride;
+  *q10u8 = vld1q_u8(d);
+  d += d_stride;
+  *q11u8 = vld1q_u8(d);
+  d += d_stride;
+  *q12u8 = vld1q_u8(d);
+  d += d_stride;
+  *q13u8 = vld1q_u8(d);
+  d += d_stride;
+  *q14u8 = vld1q_u8(d);
+  d += d_stride;
+  *q15u8 = vld1q_u8(d);
+  return;
+}
+
+static INLINE void ADD_DIFF_16x8(uint8x16_t qdiffu8, uint8x16_t *q8u8,
+                                 uint8x16_t *q9u8, uint8x16_t *q10u8,
+                                 uint8x16_t *q11u8, uint8x16_t *q12u8,
+                                 uint8x16_t *q13u8, uint8x16_t *q14u8,
+                                 uint8x16_t *q15u8) {
+  *q8u8 = vqaddq_u8(*q8u8, qdiffu8);
+  *q9u8 = vqaddq_u8(*q9u8, qdiffu8);
+  *q10u8 = vqaddq_u8(*q10u8, qdiffu8);
+  *q11u8 = vqaddq_u8(*q11u8, qdiffu8);
+  *q12u8 = vqaddq_u8(*q12u8, qdiffu8);
+  *q13u8 = vqaddq_u8(*q13u8, qdiffu8);
+  *q14u8 = vqaddq_u8(*q14u8, qdiffu8);
+  *q15u8 = vqaddq_u8(*q15u8, qdiffu8);
+  return;
+}
+
+static INLINE void SUB_DIFF_16x8(uint8x16_t qdiffu8, uint8x16_t *q8u8,
+                                 uint8x16_t *q9u8, uint8x16_t *q10u8,
+                                 uint8x16_t *q11u8, uint8x16_t *q12u8,
+                                 uint8x16_t *q13u8, uint8x16_t *q14u8,
+                                 uint8x16_t *q15u8) {
+  *q8u8 = vqsubq_u8(*q8u8, qdiffu8);
+  *q9u8 = vqsubq_u8(*q9u8, qdiffu8);
+  *q10u8 = vqsubq_u8(*q10u8, qdiffu8);
+  *q11u8 = vqsubq_u8(*q11u8, qdiffu8);
+  *q12u8 = vqsubq_u8(*q12u8, qdiffu8);
+  *q13u8 = vqsubq_u8(*q13u8, qdiffu8);
+  *q14u8 = vqsubq_u8(*q14u8, qdiffu8);
+  *q15u8 = vqsubq_u8(*q15u8, qdiffu8);
+  return;
+}
+
+static INLINE void ST_16x8(uint8_t *d, int d_stride, uint8x16_t *q8u8,
+                           uint8x16_t *q9u8, uint8x16_t *q10u8,
+                           uint8x16_t *q11u8, uint8x16_t *q12u8,
+                           uint8x16_t *q13u8, uint8x16_t *q14u8,
+                           uint8x16_t *q15u8) {
+  vst1q_u8(d, *q8u8);
+  d += d_stride;
+  vst1q_u8(d, *q9u8);
+  d += d_stride;
+  vst1q_u8(d, *q10u8);
+  d += d_stride;
+  vst1q_u8(d, *q11u8);
+  d += d_stride;
+  vst1q_u8(d, *q12u8);
+  d += d_stride;
+  vst1q_u8(d, *q13u8);
+  d += d_stride;
+  vst1q_u8(d, *q14u8);
+  d += d_stride;
+  vst1q_u8(d, *q15u8);
+  return;
+}
+
+void aom_idct32x32_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x16_t q0u8, q8u8, q9u8, q10u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+  int i, j, dest_stride8;
+  uint8_t *d;
+  int16_t a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  dest_stride8 = dest_stride * 8;
+  if (a1 >= 0) {  // diff_positive_32_32
+    a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+    q0u8 = vdupq_n_u8(a1);
+    for (i = 0; i < 2; i++, dest += 16) {  // diff_positive_32_32_loop
+      d = dest;
+      for (j = 0; j < 4; j++) {
+        LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        ADD_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                      &q14u8, &q15u8);
+        ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        d += dest_stride8;
+      }
+    }
+  } else {  // diff_negative_32_32
+    a1 = -a1;
+    a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+    q0u8 = vdupq_n_u8(a1);
+    for (i = 0; i < 2; i++, dest += 16) {  // diff_negative_32_32_loop
+      d = dest;
+      for (j = 0; j < 4; j++) {
+        LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        SUB_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                      &q14u8, &q15u8);
+        ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        d += dest_stride8;
+      }
+    }
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/idct32x32_add_neon.c

@@ -0,0 +1,686 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+
+#define LOAD_FROM_TRANSPOSED(prev, first, second) \
+  q14s16 = vld1q_s16(trans_buf + first * 8);      \
+  q13s16 = vld1q_s16(trans_buf + second * 8);
+
+#define LOAD_FROM_OUTPUT(prev, first, second, qA, qB) \
+  qA = vld1q_s16(out + first * 32);                   \
+  qB = vld1q_s16(out + second * 32);
+
+#define STORE_IN_OUTPUT(prev, first, second, qA, qB) \
+  vst1q_s16(out + first * 32, qA);                   \
+  vst1q_s16(out + second * 32, qB);
+
+#define STORE_COMBINE_CENTER_RESULTS(r10, r9) \
+  __STORE_COMBINE_CENTER_RESULTS(r10, r9, stride, q6s16, q7s16, q8s16, q9s16);
+static INLINE void __STORE_COMBINE_CENTER_RESULTS(uint8_t *p1, uint8_t *p2,
+                                                  int stride, int16x8_t q6s16,
+                                                  int16x8_t q7s16,
+                                                  int16x8_t q8s16,
+                                                  int16x8_t q9s16) {
+  int16x4_t d8s16, d9s16, d10s16, d11s16;
+
+  d8s16 = vld1_s16((int16_t *)p1);
+  p1 += stride;
+  d11s16 = vld1_s16((int16_t *)p2);
+  p2 -= stride;
+  d9s16 = vld1_s16((int16_t *)p1);
+  d10s16 = vld1_s16((int16_t *)p2);
+
+  q7s16 = vrshrq_n_s16(q7s16, 6);
+  q8s16 = vrshrq_n_s16(q8s16, 6);
+  q9s16 = vrshrq_n_s16(q9s16, 6);
+  q6s16 = vrshrq_n_s16(q6s16, 6);
+
+  q7s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q7s16), vreinterpret_u8_s16(d9s16)));
+  q8s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_s16(d10s16)));
+  q9s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_s16(d11s16)));
+  q6s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q6s16), vreinterpret_u8_s16(d8s16)));
+
+  d9s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+  d10s16 = vreinterpret_s16_u8(vqmovun_s16(q8s16));
+  d11s16 = vreinterpret_s16_u8(vqmovun_s16(q9s16));
+  d8s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+
+  vst1_s16((int16_t *)p1, d9s16);
+  p1 -= stride;
+  vst1_s16((int16_t *)p2, d10s16);
+  p2 += stride;
+  vst1_s16((int16_t *)p1, d8s16);
+  vst1_s16((int16_t *)p2, d11s16);
+  return;
+}
+
+#define STORE_COMBINE_EXTREME_RESULTS(r7, r6) \
+  ;                                           \
+  __STORE_COMBINE_EXTREME_RESULTS(r7, r6, stride, q4s16, q5s16, q6s16, q7s16);
+static INLINE void __STORE_COMBINE_EXTREME_RESULTS(uint8_t *p1, uint8_t *p2,
+                                                   int stride, int16x8_t q4s16,
+                                                   int16x8_t q5s16,
+                                                   int16x8_t q6s16,
+                                                   int16x8_t q7s16) {
+  int16x4_t d4s16, d5s16, d6s16, d7s16;
+
+  d4s16 = vld1_s16((int16_t *)p1);
+  p1 += stride;
+  d7s16 = vld1_s16((int16_t *)p2);
+  p2 -= stride;
+  d5s16 = vld1_s16((int16_t *)p1);
+  d6s16 = vld1_s16((int16_t *)p2);
+
+  q5s16 = vrshrq_n_s16(q5s16, 6);
+  q6s16 = vrshrq_n_s16(q6s16, 6);
+  q7s16 = vrshrq_n_s16(q7s16, 6);
+  q4s16 = vrshrq_n_s16(q4s16, 6);
+
+  q5s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q5s16), vreinterpret_u8_s16(d5s16)));
+  q6s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q6s16), vreinterpret_u8_s16(d6s16)));
+  q7s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q7s16), vreinterpret_u8_s16(d7s16)));
+  q4s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q4s16), vreinterpret_u8_s16(d4s16)));
+
+  d5s16 = vreinterpret_s16_u8(vqmovun_s16(q5s16));
+  d6s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+  d7s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+  d4s16 = vreinterpret_s16_u8(vqmovun_s16(q4s16));
+
+  vst1_s16((int16_t *)p1, d5s16);
+  p1 -= stride;
+  vst1_s16((int16_t *)p2, d6s16);
+  p2 += stride;
+  vst1_s16((int16_t *)p2, d7s16);
+  vst1_s16((int16_t *)p1, d4s16);
+  return;
+}
+
+#define DO_BUTTERFLY_STD(const_1, const_2, qA, qB) \
+  DO_BUTTERFLY(q14s16, q13s16, const_1, const_2, qA, qB);
+static INLINE void DO_BUTTERFLY(int16x8_t q14s16, int16x8_t q13s16,
+                                int16_t first_const, int16_t second_const,
+                                int16x8_t *qAs16, int16x8_t *qBs16) {
+  int16x4_t d30s16, d31s16;
+  int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q15s32;
+  int16x4_t dCs16, dDs16, dAs16, dBs16;
+
+  dCs16 = vget_low_s16(q14s16);
+  dDs16 = vget_high_s16(q14s16);
+  dAs16 = vget_low_s16(q13s16);
+  dBs16 = vget_high_s16(q13s16);
+
+  d30s16 = vdup_n_s16(first_const);
+  d31s16 = vdup_n_s16(second_const);
+
+  q8s32 = vmull_s16(dCs16, d30s16);
+  q10s32 = vmull_s16(dAs16, d31s16);
+  q9s32 = vmull_s16(dDs16, d30s16);
+  q11s32 = vmull_s16(dBs16, d31s16);
+  q12s32 = vmull_s16(dCs16, d31s16);
+
+  q8s32 = vsubq_s32(q8s32, q10s32);
+  q9s32 = vsubq_s32(q9s32, q11s32);
+
+  q10s32 = vmull_s16(dDs16, d31s16);
+  q11s32 = vmull_s16(dAs16, d30s16);
+  q15s32 = vmull_s16(dBs16, d30s16);
+
+  q11s32 = vaddq_s32(q12s32, q11s32);
+  q10s32 = vaddq_s32(q10s32, q15s32);
+
+  *qAs16 = vcombine_s16(vqrshrn_n_s32(q8s32, 14), vqrshrn_n_s32(q9s32, 14));
+  *qBs16 = vcombine_s16(vqrshrn_n_s32(q11s32, 14), vqrshrn_n_s32(q10s32, 14));
+  return;
+}
+
+static INLINE void idct32_transpose_pair(int16_t *input, int16_t *t_buf) {
+  int16_t *in;
+  int i;
+  const int stride = 32;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+  int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+  for (i = 0; i < 4; i++, input += 8) {
+    in = input;
+    q8s16 = vld1q_s16(in);
+    in += stride;
+    q9s16 = vld1q_s16(in);
+    in += stride;
+    q10s16 = vld1q_s16(in);
+    in += stride;
+    q11s16 = vld1q_s16(in);
+    in += stride;
+    q12s16 = vld1q_s16(in);
+    in += stride;
+    q13s16 = vld1q_s16(in);
+    in += stride;
+    q14s16 = vld1q_s16(in);
+    in += stride;
+    q15s16 = vld1q_s16(in);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+    d30s16 = vget_low_s16(q15s16);
+    d31s16 = vget_high_s16(q15s16);
+
+    q8s16 = vcombine_s16(d16s16, d24s16);   // vswp d17, d24
+    q9s16 = vcombine_s16(d18s16, d26s16);   // vswp d19, d26
+    q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    q12s16 = vcombine_s16(d17s16, d25s16);
+    q13s16 = vcombine_s16(d19s16, d27s16);
+    q14s16 = vcombine_s16(d21s16, d29s16);
+    q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q10s16));
+    q1x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q9s16), vreinterpretq_s32_s16(q11s16));
+    q2x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q12s16), vreinterpretq_s32_s16(q14s16));
+    q3x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q13s16), vreinterpretq_s32_s16(q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    vst1q_s16(t_buf, q0x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q0x2s16.val[1]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q1x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q1x2s16.val[1]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q2x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q2x2s16.val[1]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q3x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q3x2s16.val[1]);
+    t_buf += 8;
+  }
+  return;
+}
+
+static INLINE void idct32_bands_end_1st_pass(int16_t *out, int16x8_t q2s16,
+                                             int16x8_t q3s16, int16x8_t q6s16,
+                                             int16x8_t q7s16, int16x8_t q8s16,
+                                             int16x8_t q9s16, int16x8_t q10s16,
+                                             int16x8_t q11s16, int16x8_t q12s16,
+                                             int16x8_t q13s16, int16x8_t q14s16,
+                                             int16x8_t q15s16) {
+  int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+  STORE_IN_OUTPUT(17, 16, 17, q6s16, q7s16);
+  STORE_IN_OUTPUT(17, 14, 15, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(15, 30, 31, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(31, 30, 31, q6s16, q7s16);
+  STORE_IN_OUTPUT(31, 0, 1, q4s16, q5s16);
+
+  LOAD_FROM_OUTPUT(1, 12, 13, q0s16, q1s16);
+  q2s16 = vaddq_s16(q10s16, q1s16);
+  q3s16 = vaddq_s16(q11s16, q0s16);
+  q4s16 = vsubq_s16(q11s16, q0s16);
+  q5s16 = vsubq_s16(q10s16, q1s16);
+
+  LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16);
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_IN_OUTPUT(19, 18, 19, q6s16, q7s16);
+  STORE_IN_OUTPUT(19, 12, 13, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(13, 28, 29, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(29, 28, 29, q6s16, q7s16);
+  STORE_IN_OUTPUT(29, 2, 3, q4s16, q5s16);
+
+  LOAD_FROM_OUTPUT(3, 10, 11, q0s16, q1s16);
+  q2s16 = vaddq_s16(q12s16, q1s16);
+  q3s16 = vaddq_s16(q13s16, q0s16);
+  q4s16 = vsubq_s16(q13s16, q0s16);
+  q5s16 = vsubq_s16(q12s16, q1s16);
+
+  LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16);
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_IN_OUTPUT(21, 20, 21, q6s16, q7s16);
+  STORE_IN_OUTPUT(21, 10, 11, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(11, 26, 27, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(27, 26, 27, q6s16, q7s16);
+  STORE_IN_OUTPUT(27, 4, 5, q4s16, q5s16);
+
+  LOAD_FROM_OUTPUT(5, 8, 9, q0s16, q1s16);
+  q2s16 = vaddq_s16(q14s16, q1s16);
+  q3s16 = vaddq_s16(q15s16, q0s16);
+  q4s16 = vsubq_s16(q15s16, q0s16);
+  q5s16 = vsubq_s16(q14s16, q1s16);
+
+  LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16);
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_IN_OUTPUT(23, 22, 23, q6s16, q7s16);
+  STORE_IN_OUTPUT(23, 8, 9, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(9, 24, 25, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(25, 24, 25, q6s16, q7s16);
+  STORE_IN_OUTPUT(25, 6, 7, q4s16, q5s16);
+  return;
+}
+
+static INLINE void idct32_bands_end_2nd_pass(
+    int16_t *out, uint8_t *dest, int stride, int16x8_t q2s16, int16x8_t q3s16,
+    int16x8_t q6s16, int16x8_t q7s16, int16x8_t q8s16, int16x8_t q9s16,
+    int16x8_t q10s16, int16x8_t q11s16, int16x8_t q12s16, int16x8_t q13s16,
+    int16x8_t q14s16, int16x8_t q15s16) {
+  uint8_t *r6 = dest + 31 * stride;
+  uint8_t *r7 = dest /* +  0 * stride*/;
+  uint8_t *r9 = dest + 15 * stride;
+  uint8_t *r10 = dest + 16 * stride;
+  int str2 = stride << 1;
+  int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+  r10 += str2;
+  r9 -= str2;
+
+  LOAD_FROM_OUTPUT(17, 30, 31, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  r7 += str2;
+  r6 -= str2;
+
+  LOAD_FROM_OUTPUT(31, 12, 13, q0s16, q1s16)
+  q2s16 = vaddq_s16(q10s16, q1s16);
+  q3s16 = vaddq_s16(q11s16, q0s16);
+  q4s16 = vsubq_s16(q11s16, q0s16);
+  q5s16 = vsubq_s16(q10s16, q1s16);
+
+  LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16)
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+  r10 += str2;
+  r9 -= str2;
+
+  LOAD_FROM_OUTPUT(19, 28, 29, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  r7 += str2;
+  r6 -= str2;
+
+  LOAD_FROM_OUTPUT(29, 10, 11, q0s16, q1s16)
+  q2s16 = vaddq_s16(q12s16, q1s16);
+  q3s16 = vaddq_s16(q13s16, q0s16);
+  q4s16 = vsubq_s16(q13s16, q0s16);
+  q5s16 = vsubq_s16(q12s16, q1s16);
+
+  LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16)
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+  r10 += str2;
+  r9 -= str2;
+
+  LOAD_FROM_OUTPUT(21, 26, 27, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  r7 += str2;
+  r6 -= str2;
+
+  LOAD_FROM_OUTPUT(27, 8, 9, q0s16, q1s16)
+  q2s16 = vaddq_s16(q14s16, q1s16);
+  q3s16 = vaddq_s16(q15s16, q0s16);
+  q4s16 = vsubq_s16(q15s16, q0s16);
+  q5s16 = vsubq_s16(q14s16, q1s16);
+
+  LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16)
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+
+  LOAD_FROM_OUTPUT(23, 24, 25, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  return;
+}
+
+void aom_idct32x32_1024_add_neon(int16_t *input, uint8_t *dest, int stride) {
+  int i, idct32_pass_loop;
+  int16_t trans_buf[32 * 8];
+  int16_t pass1[32 * 32];
+  int16_t pass2[32 * 32];
+  int16_t *out;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+
+  for (idct32_pass_loop = 0, out = pass1; idct32_pass_loop < 2;
+       idct32_pass_loop++,
+      input = pass1,  // the input of pass2 is the result of pass1
+       out = pass2) {
+    for (i = 0; i < 4; i++, input += 32 * 8, out += 8) {  // idct32_bands_loop
+      idct32_transpose_pair(input, trans_buf);
+
+      // -----------------------------------------
+      // BLOCK A: 16-19,28-31
+      // -----------------------------------------
+      // generate 16,17,30,31
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(0, 1, 31)
+      DO_BUTTERFLY_STD(cospi_31_64, cospi_1_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(31, 17, 15)
+      DO_BUTTERFLY_STD(cospi_15_64, cospi_17_64, &q1s16, &q3s16)
+      // part of stage 2
+      q4s16 = vaddq_s16(q0s16, q1s16);
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q6s16 = vaddq_s16(q2s16, q3s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q5s16, &q7s16)
+
+      // generate 18,19,28,29
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(15, 9, 23)
+      DO_BUTTERFLY_STD(cospi_23_64, cospi_9_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(23, 25, 7)
+      DO_BUTTERFLY_STD(cospi_7_64, cospi_25_64, &q1s16, &q3s16)
+      // part of stage 2
+      q13s16 = vsubq_s16(q3s16, q2s16);
+      q3s16 = vaddq_s16(q3s16, q2s16);
+      q14s16 = vsubq_s16(q1s16, q0s16);
+      q2s16 = vaddq_s16(q1s16, q0s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(-cospi_4_64, -cospi_28_64, &q1s16, &q0s16)
+      // part of stage 4
+      q8s16 = vaddq_s16(q4s16, q2s16);
+      q9s16 = vaddq_s16(q5s16, q0s16);
+      q10s16 = vaddq_s16(q7s16, q1s16);
+      q15s16 = vaddq_s16(q6s16, q3s16);
+      q13s16 = vsubq_s16(q5s16, q0s16);
+      q14s16 = vsubq_s16(q7s16, q1s16);
+      STORE_IN_OUTPUT(0, 16, 31, q8s16, q15s16)
+      STORE_IN_OUTPUT(31, 17, 30, q9s16, q10s16)
+      // part of stage 5
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q0s16, &q1s16)
+      STORE_IN_OUTPUT(30, 29, 18, q1s16, q0s16)
+      // part of stage 4
+      q13s16 = vsubq_s16(q4s16, q2s16);
+      q14s16 = vsubq_s16(q6s16, q3s16);
+      // part of stage 5
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q4s16, &q6s16)
+      STORE_IN_OUTPUT(18, 19, 28, q4s16, q6s16)
+
+      // -----------------------------------------
+      // BLOCK B: 20-23,24-27
+      // -----------------------------------------
+      // generate 20,21,26,27
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(7, 5, 27)
+      DO_BUTTERFLY_STD(cospi_27_64, cospi_5_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(27, 21, 11)
+      DO_BUTTERFLY_STD(cospi_11_64, cospi_21_64, &q1s16, &q3s16)
+      // part of stage 2
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q0s16 = vaddq_s16(q0s16, q1s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      q2s16 = vaddq_s16(q2s16, q3s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+
+      // generate 22,23,24,25
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(11, 13, 19)
+      DO_BUTTERFLY_STD(cospi_19_64, cospi_13_64, &q5s16, &q7s16)
+      LOAD_FROM_TRANSPOSED(19, 29, 3)
+      DO_BUTTERFLY_STD(cospi_3_64, cospi_29_64, &q4s16, &q6s16)
+      // part of stage 2
+      q14s16 = vsubq_s16(q4s16, q5s16);
+      q5s16 = vaddq_s16(q4s16, q5s16);
+      q13s16 = vsubq_s16(q6s16, q7s16);
+      q6s16 = vaddq_s16(q6s16, q7s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(-cospi_20_64, -cospi_12_64, &q4s16, &q7s16)
+      // part of stage 4
+      q10s16 = vaddq_s16(q7s16, q1s16);
+      q11s16 = vaddq_s16(q5s16, q0s16);
+      q12s16 = vaddq_s16(q6s16, q2s16);
+      q15s16 = vaddq_s16(q4s16, q3s16);
+      // part of stage 6
+      LOAD_FROM_OUTPUT(28, 16, 17, q14s16, q13s16)
+      q8s16 = vaddq_s16(q14s16, q11s16);
+      q9s16 = vaddq_s16(q13s16, q10s16);
+      q13s16 = vsubq_s16(q13s16, q10s16);
+      q11s16 = vsubq_s16(q14s16, q11s16);
+      STORE_IN_OUTPUT(17, 17, 16, q9s16, q8s16)
+      LOAD_FROM_OUTPUT(16, 30, 31, q14s16, q9s16)
+      q8s16 = vsubq_s16(q9s16, q12s16);
+      q10s16 = vaddq_s16(q14s16, q15s16);
+      q14s16 = vsubq_s16(q14s16, q15s16);
+      q12s16 = vaddq_s16(q9s16, q12s16);
+      STORE_IN_OUTPUT(31, 30, 31, q10s16, q12s16)
+      // part of stage 7
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+      STORE_IN_OUTPUT(31, 25, 22, q14s16, q13s16)
+      q13s16 = q11s16;
+      q14s16 = q8s16;
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+      STORE_IN_OUTPUT(22, 24, 23, q14s16, q13s16)
+      // part of stage 4
+      q14s16 = vsubq_s16(q5s16, q0s16);
+      q13s16 = vsubq_s16(q6s16, q2s16);
+      DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q5s16, &q6s16);
+      q14s16 = vsubq_s16(q7s16, q1s16);
+      q13s16 = vsubq_s16(q4s16, q3s16);
+      DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q0s16, &q1s16);
+      // part of stage 6
+      LOAD_FROM_OUTPUT(23, 18, 19, q14s16, q13s16)
+      q8s16 = vaddq_s16(q14s16, q1s16);
+      q9s16 = vaddq_s16(q13s16, q6s16);
+      q13s16 = vsubq_s16(q13s16, q6s16);
+      q1s16 = vsubq_s16(q14s16, q1s16);
+      STORE_IN_OUTPUT(19, 18, 19, q8s16, q9s16)
+      LOAD_FROM_OUTPUT(19, 28, 29, q8s16, q9s16)
+      q14s16 = vsubq_s16(q8s16, q5s16);
+      q10s16 = vaddq_s16(q8s16, q5s16);
+      q11s16 = vaddq_s16(q9s16, q0s16);
+      q0s16 = vsubq_s16(q9s16, q0s16);
+      STORE_IN_OUTPUT(29, 28, 29, q10s16, q11s16)
+      // part of stage 7
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+      STORE_IN_OUTPUT(29, 20, 27, q13s16, q14s16)
+      DO_BUTTERFLY(q0s16, q1s16, cospi_16_64, cospi_16_64, &q1s16, &q0s16);
+      STORE_IN_OUTPUT(27, 21, 26, q1s16, q0s16)
+
+      // -----------------------------------------
+      // BLOCK C: 8-10,11-15
+      // -----------------------------------------
+      // generate 8,9,14,15
+      // part of stage 2
+      LOAD_FROM_TRANSPOSED(3, 2, 30)
+      DO_BUTTERFLY_STD(cospi_30_64, cospi_2_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(30, 18, 14)
+      DO_BUTTERFLY_STD(cospi_14_64, cospi_18_64, &q1s16, &q3s16)
+      // part of stage 3
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q0s16 = vaddq_s16(q0s16, q1s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      q2s16 = vaddq_s16(q2s16, q3s16);
+      // part of stage 4
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q1s16, &q3s16)
+
+      // generate 10,11,12,13
+      // part of stage 2
+      LOAD_FROM_TRANSPOSED(14, 10, 22)
+      DO_BUTTERFLY_STD(cospi_22_64, cospi_10_64, &q5s16, &q7s16)
+      LOAD_FROM_TRANSPOSED(22, 26, 6)
+      DO_BUTTERFLY_STD(cospi_6_64, cospi_26_64, &q4s16, &q6s16)
+      // part of stage 3
+      q14s16 = vsubq_s16(q4s16, q5s16);
+      q5s16 = vaddq_s16(q4s16, q5s16);
+      q13s16 = vsubq_s16(q6s16, q7s16);
+      q6s16 = vaddq_s16(q6s16, q7s16);
+      // part of stage 4
+      DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q4s16, &q7s16)
+      // part of stage 5
+      q8s16 = vaddq_s16(q0s16, q5s16);
+      q9s16 = vaddq_s16(q1s16, q7s16);
+      q13s16 = vsubq_s16(q1s16, q7s16);
+      q14s16 = vsubq_s16(q3s16, q4s16);
+      q10s16 = vaddq_s16(q3s16, q4s16);
+      q15s16 = vaddq_s16(q2s16, q6s16);
+      STORE_IN_OUTPUT(26, 8, 15, q8s16, q15s16)
+      STORE_IN_OUTPUT(15, 9, 14, q9s16, q10s16)
+      // part of stage 6
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+      STORE_IN_OUTPUT(14, 13, 10, q3s16, q1s16)
+      q13s16 = vsubq_s16(q0s16, q5s16);
+      q14s16 = vsubq_s16(q2s16, q6s16);
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+      STORE_IN_OUTPUT(10, 11, 12, q1s16, q3s16)
+
+      // -----------------------------------------
+      // BLOCK D: 0-3,4-7
+      // -----------------------------------------
+      // generate 4,5,6,7
+      // part of stage 3
+      LOAD_FROM_TRANSPOSED(6, 4, 28)
+      DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(28, 20, 12)
+      DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+      // part of stage 4
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q0s16 = vaddq_s16(q0s16, q1s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      q2s16 = vaddq_s16(q2s16, q3s16);
+      // part of stage 5
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+
+      // generate 0,1,2,3
+      // part of stage 4
+      LOAD_FROM_TRANSPOSED(12, 0, 16)
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q5s16, &q7s16)
+      LOAD_FROM_TRANSPOSED(16, 8, 24)
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q14s16, &q6s16)
+      // part of stage 5
+      q4s16 = vaddq_s16(q7s16, q6s16);
+      q7s16 = vsubq_s16(q7s16, q6s16);
+      q6s16 = vsubq_s16(q5s16, q14s16);
+      q5s16 = vaddq_s16(q5s16, q14s16);
+      // part of stage 6
+      q8s16 = vaddq_s16(q4s16, q2s16);
+      q9s16 = vaddq_s16(q5s16, q3s16);
+      q10s16 = vaddq_s16(q6s16, q1s16);
+      q11s16 = vaddq_s16(q7s16, q0s16);
+      q12s16 = vsubq_s16(q7s16, q0s16);
+      q13s16 = vsubq_s16(q6s16, q1s16);
+      q14s16 = vsubq_s16(q5s16, q3s16);
+      q15s16 = vsubq_s16(q4s16, q2s16);
+      // part of stage 7
+      LOAD_FROM_OUTPUT(12, 14, 15, q0s16, q1s16)
+      q2s16 = vaddq_s16(q8s16, q1s16);
+      q3s16 = vaddq_s16(q9s16, q0s16);
+      q4s16 = vsubq_s16(q9s16, q0s16);
+      q5s16 = vsubq_s16(q8s16, q1s16);
+      LOAD_FROM_OUTPUT(15, 16, 17, q0s16, q1s16)
+      q8s16 = vaddq_s16(q4s16, q1s16);
+      q9s16 = vaddq_s16(q5s16, q0s16);
+      q6s16 = vsubq_s16(q5s16, q0s16);
+      q7s16 = vsubq_s16(q4s16, q1s16);
+
+      if (idct32_pass_loop == 0) {
+        idct32_bands_end_1st_pass(out, q2s16, q3s16, q6s16, q7s16, q8s16, q9s16,
+                                  q10s16, q11s16, q12s16, q13s16, q14s16,
+                                  q15s16);
+      } else {
+        idct32_bands_end_2nd_pass(out, dest, stride, q2s16, q3s16, q6s16, q7s16,
+                                  q8s16, q9s16, q10s16, q11s16, q12s16, q13s16,
+                                  q14s16, q15s16);
+        dest += 8;
+      }
+    }
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.asm

@@ -0,0 +1,71 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+
+    EXPORT  |aom_idct4x4_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_idct4x4_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct4x4_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 4)
+    add              r0, r0, #8                ; + (1 <<((4) - 1))
+    asr              r0, r0, #4                ; >> 4
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    vld1.32          {d2[0]}, [r1], r2
+    vld1.32          {d2[1]}, [r1], r2
+    vld1.32          {d4[0]}, [r1], r2
+    vld1.32          {d4[1]}, [r1]
+
+    vaddw.u8         q8, q0, d2                ; dest[x] + a1
+    vaddw.u8         q9, q0, d4
+
+    vqmovun.s16      d6, q8                    ; clip_pixel
+    vqmovun.s16      d7, q9
+
+    vst1.32          {d6[0]}, [r12], r2
+    vst1.32          {d6[1]}, [r12], r2
+    vst1.32          {d7[0]}, [r12], r2
+    vst1.32          {d7[1]}, [r12]
+
+    bx               lr
+    ENDP             ; |aom_idct4x4_1_add_neon|
+
+    END

third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.c

@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+void aom_idct4x4_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d6u8;
+  uint32x2_t d2u32 = vdup_n_u32(0);
+  uint16x8_t q8u16;
+  int16x8_t q0s16;
+  uint8_t *d1, *d2;
+  int16_t i, a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  q0s16 = vdupq_n_s16(a1);
+
+  // dc_only_idct_add
+  d1 = d2 = dest;
+  for (i = 0; i < 2; i++) {
+    d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0);
+    d1 += dest_stride;
+    d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16), vreinterpret_u8_u32(d2u32));
+    d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+
+    vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0);
+    d2 += dest_stride;
+    vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1);
+    d2 += dest_stride;
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/idct4x4_add_neon.asm

@@ -0,0 +1,193 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_idct4x4_16_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    AREA     Block, CODE, READONLY ; name this block of code
+;void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct4x4_16_add_neon| PROC
+
+    ; The 2D transform is done with two passes which are actually pretty
+    ; similar. We first transform the rows. This is done by transposing
+    ; the inputs, doing an SIMD column transform (the columns are the
+    ; transposed rows) and then transpose the results (so that it goes back
+    ; in normal/row positions). Then, we transform the columns by doing
+    ; another SIMD column transform.
+    ; So, two passes of a transpose followed by a column transform.
+
+    ; load the inputs into q8-q9, d16-d19
+    vld1.s16        {q8,q9}, [r0]!
+
+    ; generate scalar constants
+    ; cospi_8_64 = 15137 = 0x3b21
+    mov             r0, #0x3b00
+    add             r0, #0x21
+    ; cospi_16_64 = 11585 = 0x2d41
+    mov             r3, #0x2d00
+    add             r3, #0x41
+    ; cospi_24_64 = 6270 = 0x 187e
+    mov             r12, #0x1800
+    add             r12, #0x7e
+
+    ; transpose the input data
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+
+    ; generate constant vectors
+    vdup.16         d20, r0         ; replicate cospi_8_64
+    vdup.16         d21, r3         ; replicate cospi_16_64
+
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    vdup.16         d22, r12        ; replicate cospi_24_64
+
+    ; do the transform on transposed rows
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+    vswp     d18, d19
+
+    ; transpose the results
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    ; do the transform on columns
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+
+    ; The results are in two registers, one of them being swapped. This will
+    ; be taken care of by loading the 'dest' value in a swapped fashion and
+    ; also storing them in the same swapped fashion.
+    ; temp_out[0, 1] = d16, d17 = q8
+    ; temp_out[2, 3] = d19, d18 = q9 swapped
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4)
+    vrshr.s16 q8, q8, #4
+    vrshr.s16 q9, q9, #4
+
+    vld1.32 {d26[0]}, [r1], r2
+    vld1.32 {d26[1]}, [r1], r2
+    vld1.32 {d27[1]}, [r1], r2
+    vld1.32 {d27[0]}, [r1]  ; no post-increment
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i]
+    vaddw.u8 q8, q8, d26
+    vaddw.u8 q9, q9, d27
+
+    ; clip_pixel
+    vqmovun.s16 d26, q8
+    vqmovun.s16 d27, q9
+
+    ; do the stores in reverse order with negative post-increment, by changing
+    ; the sign of the stride
+    rsb r2, r2, #0
+    vst1.32 {d27[0]}, [r1], r2
+    vst1.32 {d27[1]}, [r1], r2
+    vst1.32 {d26[1]}, [r1], r2
+    vst1.32 {d26[0]}, [r1]  ; no post-increment
+    bx              lr
+    ENDP  ; |aom_idct4x4_16_add_neon|
+
+    END

third_party/aom/aom_dsp/arm/idct4x4_add_neon.c

@@ -0,0 +1,146 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/txfm_common.h"
+
+void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d26u8, d27u8;
+  uint32x2_t d26u32, d27u32;
+  uint16x8_t q8u16, q9u16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
+  int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
+  int16x8_t q8s16, q9s16, q13s16, q14s16;
+  int32x4_t q1s32, q13s32, q14s32, q15s32;
+  int16x4x2_t d0x2s16, d1x2s16;
+  int32x4x2_t q0x2s32;
+  uint8_t *d;
+
+  d26u32 = d27u32 = vdup_n_u32(0);
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+
+  d16s16 = vget_low_s16(q8s16);
+  d17s16 = vget_high_s16(q8s16);
+  d18s16 = vget_low_s16(q9s16);
+  d19s16 = vget_high_s16(q9s16);
+
+  d0x2s16 = vtrn_s16(d16s16, d17s16);
+  d1x2s16 = vtrn_s16(d18s16, d19s16);
+  q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+  q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+  d20s16 = vdup_n_s16((int16_t)cospi_8_64);
+  d21s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
+  d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+  d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+  d22s16 = vdup_n_s16((int16_t)cospi_24_64);
+
+  // stage 1
+  d23s16 = vadd_s16(d16s16, d18s16);
+  d24s16 = vsub_s16(d16s16, d18s16);
+
+  q15s32 = vmull_s16(d17s16, d22s16);
+  q1s32 = vmull_s16(d17s16, d20s16);
+  q13s32 = vmull_s16(d23s16, d21s16);
+  q14s32 = vmull_s16(d24s16, d21s16);
+
+  q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+  q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
+
+  d26s16 = vqrshrn_n_s32(q13s32, 14);
+  d27s16 = vqrshrn_n_s32(q14s32, 14);
+  d29s16 = vqrshrn_n_s32(q15s32, 14);
+  d28s16 = vqrshrn_n_s32(q1s32, 14);
+  q13s16 = vcombine_s16(d26s16, d27s16);
+  q14s16 = vcombine_s16(d28s16, d29s16);
+
+  // stage 2
+  q8s16 = vaddq_s16(q13s16, q14s16);
+  q9s16 = vsubq_s16(q13s16, q14s16);
+
+  d16s16 = vget_low_s16(q8s16);
+  d17s16 = vget_high_s16(q8s16);
+  d18s16 = vget_high_s16(q9s16);  // vswp d18 d19
+  d19s16 = vget_low_s16(q9s16);
+
+  d0x2s16 = vtrn_s16(d16s16, d17s16);
+  d1x2s16 = vtrn_s16(d18s16, d19s16);
+  q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+  q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
+  d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+  d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+  // do the transform on columns
+  // stage 1
+  d23s16 = vadd_s16(d16s16, d18s16);
+  d24s16 = vsub_s16(d16s16, d18s16);
+
+  q15s32 = vmull_s16(d17s16, d22s16);
+  q1s32 = vmull_s16(d17s16, d20s16);
+  q13s32 = vmull_s16(d23s16, d21s16);
+  q14s32 = vmull_s16(d24s16, d21s16);
+
+  q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+  q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
+
+  d26s16 = vqrshrn_n_s32(q13s32, 14);
+  d27s16 = vqrshrn_n_s32(q14s32, 14);
+  d29s16 = vqrshrn_n_s32(q15s32, 14);
+  d28s16 = vqrshrn_n_s32(q1s32, 14);
+  q13s16 = vcombine_s16(d26s16, d27s16);
+  q14s16 = vcombine_s16(d28s16, d29s16);
+
+  // stage 2
+  q8s16 = vaddq_s16(q13s16, q14s16);
+  q9s16 = vsubq_s16(q13s16, q14s16);
+
+  q8s16 = vrshrq_n_s16(q8s16, 4);
+  q9s16 = vrshrq_n_s16(q9s16, 4);
+
+  d = dest;
+  d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
+  d += dest_stride;
+  d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
+  d += dest_stride;
+  d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
+  d += dest_stride;
+  d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
+
+  d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+
+  d = dest;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
+  d += dest_stride;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
+  d += dest_stride;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
+  d += dest_stride;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
+  return;
+}

third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.asm

@@ -0,0 +1,91 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+
+    EXPORT  |aom_idct8x8_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_idct8x8_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct8x8_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 5)
+    add              r0, r0, #16               ; + (1 <<((5) - 1))
+    asr              r0, r0, #5                ; >> 5
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    ; load destination data
+    vld1.64          {d2}, [r1], r2
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r2
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r2
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r2
+    vld1.64          {d17}, [r1]
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |aom_idct8x8_1_add_neon|
+
+    END

third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.c

@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+void aom_idct8x8_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d2u8, d3u8, d30u8, d31u8;
+  uint64x1_t d2u64, d3u64, d4u64, d5u64;
+  uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+  int16x8_t q0s16;
+  uint8_t *d1, *d2;
+  int16_t i, a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+
+  q0s16 = vdupq_n_s16(a1);
+  q0u16 = vreinterpretq_u16_s16(q0s16);
+
+  d1 = d2 = dest;
+  for (i = 0; i < 2; i++) {
+    d2u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+    d4u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+    d5u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+
+    q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+    q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+    q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+    q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+    d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d31u8));
+    d2 += dest_stride;
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/idct8x8_add_neon.asm

@@ -0,0 +1,522 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_idct8x8_64_add_neon|
+    EXPORT  |aom_idct8x8_12_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Parallel 1D IDCT on all the columns of a 8x8 16bit data matrix which are
+    ; loaded in q8-q15. The output will be stored back into q8-q15 registers.
+    ; This macro will touch q0-q7 registers and use them as buffer during
+    ; calculation.
+    MACRO
+    IDCT8x8_1D
+    ; stage 1
+    vdup.16         d0, r3                    ; duplicate cospi_28_64
+    vdup.16         d1, r4                    ; duplicate cospi_4_64
+    vdup.16         d2, r5                    ; duplicate cospi_12_64
+    vdup.16         d3, r6                    ; duplicate cospi_20_64
+
+    ; input[1] * cospi_28_64
+    vmull.s16       q2, d18, d0
+    vmull.s16       q3, d19, d0
+
+    ; input[5] * cospi_12_64
+    vmull.s16       q5, d26, d2
+    vmull.s16       q6, d27, d2
+
+    ; input[1]*cospi_28_64-input[7]*cospi_4_64
+    vmlsl.s16       q2, d30, d1
+    vmlsl.s16       q3, d31, d1
+
+    ; input[5] * cospi_12_64 - input[3] * cospi_20_64
+    vmlsl.s16       q5, d22, d3
+    vmlsl.s16       q6, d23, d3
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d8, q2, #14               ; >> 14
+    vqrshrn.s32     d9, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q5, #14              ; >> 14
+    vqrshrn.s32     d11, q6, #14              ; >> 14
+
+    ; input[1] * cospi_4_64
+    vmull.s16       q2, d18, d1
+    vmull.s16       q3, d19, d1
+
+    ; input[5] * cospi_20_64
+    vmull.s16       q9, d26, d3
+    vmull.s16       q13, d27, d3
+
+    ; input[1]*cospi_4_64+input[7]*cospi_28_64
+    vmlal.s16       q2, d30, d0
+    vmlal.s16       q3, d31, d0
+
+    ; input[5] * cospi_20_64 + input[3] * cospi_12_64
+    vmlal.s16       q9, d22, d2
+    vmlal.s16       q13, d23, d2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d14, q2, #14              ; >> 14
+    vqrshrn.s32     d15, q3, #14              ; >> 14
+
+    ; stage 2 & stage 3 - even half
+    vdup.16         d0, r7                    ; duplicate cospi_16_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q13, #14              ; >> 14
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q2, d16, d0
+    vmull.s16       q3, d17, d0
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q13, d16, d0
+    vmull.s16       q15, d17, d0
+
+    ; (input[0] + input[2]) * cospi_16_64
+    vmlal.s16       q2,  d24, d0
+    vmlal.s16       q3, d25, d0
+
+    ; (input[0] - input[2]) * cospi_16_64
+    vmlsl.s16       q13, d24, d0
+    vmlsl.s16       q15, d25, d0
+
+    vdup.16         d0, r8                    ; duplicate cospi_24_64
+    vdup.16         d1, r9                    ; duplicate cospi_8_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d18, q2, #14              ; >> 14
+    vqrshrn.s32     d19, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d22, q13, #14              ; >> 14
+    vqrshrn.s32     d23, q15, #14              ; >> 14
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    ; input[1] * cospi_24_64
+    vmull.s16       q2, d20, d0
+    vmull.s16       q3, d21, d0
+
+    ; input[1] * cospi_8_64
+    vmull.s16       q8, d20, d1
+    vmull.s16       q12, d21, d1
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    vmlsl.s16       q2, d28, d1
+    vmlsl.s16       q3, d29, d1
+
+    ; input[1] * cospi_8_64 + input[3] * cospi_24_64
+    vmlal.s16       q8, d28, d0
+    vmlal.s16       q12, d29, d0
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d26, q2, #14              ; >> 14
+    vqrshrn.s32     d27, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d30, q8, #14              ; >> 14
+    vqrshrn.s32     d31, q12, #14              ; >> 14
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q11, q13              ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q11, q13              ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+    MEND
+
+    ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
+    MACRO
+    TRANSPOSE8X8
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vtrn.32         q8, q10
+    vtrn.32         q9, q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.16         q8, q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    MEND
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void aom_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct8x8_64_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    IDCT8x8_1D
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |aom_idct8x8_64_add_neon|
+
+;void aom_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct8x8_12_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    ; stage 1
+    ; The following instructions use vqrdmulh to do the
+    ; dct_const_round_shift(input[1] * cospi_28_64). vqrdmulh will do doubling
+    ; multiply and shift the result by 16 bits instead of 14 bits. So we need
+    ; to double the constants before multiplying to compensate this.
+    mov             r12, r3, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_28_64*2
+    mov             r12, r4, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_4_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_28_64)
+    vqrdmulh.s16    q4, q9, q0
+
+    mov             r12, r6, lsl #1
+    rsb             r12, #0
+    vdup.16         q0, r12                   ; duplicate -cospi_20_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_4_64)
+    vqrdmulh.s16    q7, q9, q1
+
+    mov             r12, r5, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_12_64*2
+
+    ; dct_const_round_shift(- input[3] * cospi_20_64)
+    vqrdmulh.s16    q5, q11, q0
+
+    mov             r12, r7, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_16_64*2
+
+    ; dct_const_round_shift(input[3] * cospi_12_64)
+    vqrdmulh.s16    q6, q11, q1
+
+    ; stage 2 & stage 3 - even half
+    mov             r12, r8, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_24_64*2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrdmulh.s16    q9, q8, q0
+
+    mov             r12, r9, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_8_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_24_64)
+    vqrdmulh.s16    q13, q10, q1
+
+    ; dct_const_round_shift(input[1] * cospi_8_64)
+    vqrdmulh.s16    q15, q10, q0
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q9, q13               ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q9, q13               ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |aom_idct8x8_12_add_neon|
+
+    END

third_party/aom/aom_dsp/arm/idct8x8_add_neon.c

@@ -0,0 +1,509 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+
+static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16,
+                                int16x8_t *q10s16, int16x8_t *q11s16,
+                                int16x8_t *q12s16, int16x8_t *q13s16,
+                                int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+  int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  *q8s16 = vcombine_s16(d16s16, d24s16);   // vswp d17, d24
+  *q9s16 = vcombine_s16(d18s16, d26s16);   // vswp d19, d26
+  *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+  *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+  *q12s16 = vcombine_s16(d17s16, d25s16);
+  *q13s16 = vcombine_s16(d19s16, d27s16);
+  *q14s16 = vcombine_s16(d21s16, d29s16);
+  *q15s16 = vcombine_s16(d23s16, d31s16);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q8s16), vreinterpretq_s32_s16(*q10s16));
+  q1x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q9s16), vreinterpretq_s32_s16(*q11s16));
+  q2x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q12s16), vreinterpretq_s32_s16(*q14s16));
+  q3x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q13s16), vreinterpretq_s32_s16(*q15s16));
+
+  q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                      vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+  q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                      vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+  q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                      vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+  q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                      vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+  *q8s16 = q0x2s16.val[0];
+  *q9s16 = q0x2s16.val[1];
+  *q10s16 = q1x2s16.val[0];
+  *q11s16 = q1x2s16.val[1];
+  *q12s16 = q2x2s16.val[0];
+  *q13s16 = q2x2s16.val[1];
+  *q14s16 = q3x2s16.val[0];
+  *q15s16 = q3x2s16.val[1];
+  return;
+}
+
+static INLINE void IDCT8x8_1D(int16x8_t *q8s16, int16x8_t *q9s16,
+                              int16x8_t *q10s16, int16x8_t *q11s16,
+                              int16x8_t *q12s16, int16x8_t *q13s16,
+                              int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d0s16, d1s16, d2s16, d3s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
+  int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+
+  d0s16 = vdup_n_s16((int16_t)cospi_28_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_4_64);
+  d2s16 = vdup_n_s16((int16_t)cospi_12_64);
+  d3s16 = vdup_n_s16((int16_t)cospi_20_64);
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  q2s32 = vmull_s16(d18s16, d0s16);
+  q3s32 = vmull_s16(d19s16, d0s16);
+  q5s32 = vmull_s16(d26s16, d2s16);
+  q6s32 = vmull_s16(d27s16, d2s16);
+
+  q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+  q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+  q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
+  q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
+
+  d8s16 = vqrshrn_n_s32(q2s32, 14);
+  d9s16 = vqrshrn_n_s32(q3s32, 14);
+  d10s16 = vqrshrn_n_s32(q5s32, 14);
+  d11s16 = vqrshrn_n_s32(q6s32, 14);
+  q4s16 = vcombine_s16(d8s16, d9s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  q2s32 = vmull_s16(d18s16, d1s16);
+  q3s32 = vmull_s16(d19s16, d1s16);
+  q9s32 = vmull_s16(d26s16, d3s16);
+  q13s32 = vmull_s16(d27s16, d3s16);
+
+  q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
+  q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
+  q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+  q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
+
+  d14s16 = vqrshrn_n_s32(q2s32, 14);
+  d15s16 = vqrshrn_n_s32(q3s32, 14);
+  d12s16 = vqrshrn_n_s32(q9s32, 14);
+  d13s16 = vqrshrn_n_s32(q13s32, 14);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+  q7s16 = vcombine_s16(d14s16, d15s16);
+
+  d0s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q2s32 = vmull_s16(d16s16, d0s16);
+  q3s32 = vmull_s16(d17s16, d0s16);
+  q13s32 = vmull_s16(d16s16, d0s16);
+  q15s32 = vmull_s16(d17s16, d0s16);
+
+  q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
+  q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
+  q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
+  q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
+
+  d0s16 = vdup_n_s16((int16_t)cospi_24_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_8_64);
+
+  d18s16 = vqrshrn_n_s32(q2s32, 14);
+  d19s16 = vqrshrn_n_s32(q3s32, 14);
+  d22s16 = vqrshrn_n_s32(q13s32, 14);
+  d23s16 = vqrshrn_n_s32(q15s32, 14);
+  *q9s16 = vcombine_s16(d18s16, d19s16);
+  *q11s16 = vcombine_s16(d22s16, d23s16);
+
+  q2s32 = vmull_s16(d20s16, d0s16);
+  q3s32 = vmull_s16(d21s16, d0s16);
+  q8s32 = vmull_s16(d20s16, d1s16);
+  q12s32 = vmull_s16(d21s16, d1s16);
+
+  q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
+  q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
+  q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
+  q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
+
+  d26s16 = vqrshrn_n_s32(q2s32, 14);
+  d27s16 = vqrshrn_n_s32(q3s32, 14);
+  d30s16 = vqrshrn_n_s32(q8s32, 14);
+  d31s16 = vqrshrn_n_s32(q12s32, 14);
+  *q13s16 = vcombine_s16(d26s16, d27s16);
+  *q15s16 = vcombine_s16(d30s16, d31s16);
+
+  q0s16 = vaddq_s16(*q9s16, *q15s16);
+  q1s16 = vaddq_s16(*q11s16, *q13s16);
+  q2s16 = vsubq_s16(*q11s16, *q13s16);
+  q3s16 = vsubq_s16(*q9s16, *q15s16);
+
+  *q13s16 = vsubq_s16(q4s16, q5s16);
+  q4s16 = vaddq_s16(q4s16, q5s16);
+  *q14s16 = vsubq_s16(q7s16, q6s16);
+  q7s16 = vaddq_s16(q7s16, q6s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+
+  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q9s32 = vmull_s16(d28s16, d16s16);
+  q10s32 = vmull_s16(d29s16, d16s16);
+  q11s32 = vmull_s16(d28s16, d16s16);
+  q12s32 = vmull_s16(d29s16, d16s16);
+
+  q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
+  q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+  q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+  q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+  d10s16 = vqrshrn_n_s32(q9s32, 14);
+  d11s16 = vqrshrn_n_s32(q10s32, 14);
+  d12s16 = vqrshrn_n_s32(q11s32, 14);
+  d13s16 = vqrshrn_n_s32(q12s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  *q8s16 = vaddq_s16(q0s16, q7s16);
+  *q9s16 = vaddq_s16(q1s16, q6s16);
+  *q10s16 = vaddq_s16(q2s16, q5s16);
+  *q11s16 = vaddq_s16(q3s16, q4s16);
+  *q12s16 = vsubq_s16(q3s16, q4s16);
+  *q13s16 = vsubq_s16(q2s16, q5s16);
+  *q14s16 = vsubq_s16(q1s16, q6s16);
+  *q15s16 = vsubq_s16(q0s16, q7s16);
+  return;
+}
+
+void aom_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8_t *d1, *d2;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  uint64x1_t d0u64, d1u64, d2u64, d3u64;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  uint16x8_t q8u16, q9u16, q10u16, q11u16;
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+  q10s16 = vld1q_s16(input + 16);
+  q11s16 = vld1q_s16(input + 24);
+  q12s16 = vld1q_s16(input + 32);
+  q13s16 = vld1q_s16(input + 40);
+  q14s16 = vld1q_s16(input + 48);
+  q15s16 = vld1q_s16(input + 56);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+             &q15s16);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+             &q15s16);
+
+  q8s16 = vrshrq_n_s16(q8s16, 5);
+  q9s16 = vrshrq_n_s16(q9s16, 5);
+  q10s16 = vrshrq_n_s16(q10s16, 5);
+  q11s16 = vrshrq_n_s16(q11s16, 5);
+  q12s16 = vrshrq_n_s16(q12s16, 5);
+  q13s16 = vrshrq_n_s16(q13s16, 5);
+  q14s16 = vrshrq_n_s16(q14s16, 5);
+  q15s16 = vrshrq_n_s16(q15s16, 5);
+
+  d1 = d2 = dest;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+
+  q8s16 = q12s16;
+  q9s16 = q13s16;
+  q10s16 = q14s16;
+  q11s16 = q15s16;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+  return;
+}
+
+void aom_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8_t *d1, *d2;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
+  int16x4_t d26s16, d27s16, d28s16, d29s16;
+  uint64x1_t d0u64, d1u64, d2u64, d3u64;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  uint16x8_t q8u16, q9u16, q10u16, q11u16;
+  int32x4_t q9s32, q10s32, q11s32, q12s32;
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+  q10s16 = vld1q_s16(input + 16);
+  q11s16 = vld1q_s16(input + 24);
+  q12s16 = vld1q_s16(input + 32);
+  q13s16 = vld1q_s16(input + 40);
+  q14s16 = vld1q_s16(input + 48);
+  q15s16 = vld1q_s16(input + 56);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  // First transform rows
+  // stage 1
+  q0s16 = vdupq_n_s16((int16_t)cospi_28_64 * 2);
+  q1s16 = vdupq_n_s16((int16_t)cospi_4_64 * 2);
+
+  q4s16 = vqrdmulhq_s16(q9s16, q0s16);
+
+  q0s16 = vdupq_n_s16(-(int16_t)cospi_20_64 * 2);
+
+  q7s16 = vqrdmulhq_s16(q9s16, q1s16);
+
+  q1s16 = vdupq_n_s16((int16_t)cospi_12_64 * 2);
+
+  q5s16 = vqrdmulhq_s16(q11s16, q0s16);
+
+  q0s16 = vdupq_n_s16((int16_t)cospi_16_64 * 2);
+
+  q6s16 = vqrdmulhq_s16(q11s16, q1s16);
+
+  // stage 2 & stage 3 - even half
+  q1s16 = vdupq_n_s16((int16_t)cospi_24_64 * 2);
+
+  q9s16 = vqrdmulhq_s16(q8s16, q0s16);
+
+  q0s16 = vdupq_n_s16((int16_t)cospi_8_64 * 2);
+
+  q13s16 = vqrdmulhq_s16(q10s16, q1s16);
+
+  q15s16 = vqrdmulhq_s16(q10s16, q0s16);
+
+  // stage 3 -odd half
+  q0s16 = vaddq_s16(q9s16, q15s16);
+  q1s16 = vaddq_s16(q9s16, q13s16);
+  q2s16 = vsubq_s16(q9s16, q13s16);
+  q3s16 = vsubq_s16(q9s16, q15s16);
+
+  // stage 2 - odd half
+  q13s16 = vsubq_s16(q4s16, q5s16);
+  q4s16 = vaddq_s16(q4s16, q5s16);
+  q14s16 = vsubq_s16(q7s16, q6s16);
+  q7s16 = vaddq_s16(q7s16, q6s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+  d28s16 = vget_low_s16(q14s16);
+  d29s16 = vget_high_s16(q14s16);
+
+  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
+  q9s32 = vmull_s16(d28s16, d16s16);
+  q10s32 = vmull_s16(d29s16, d16s16);
+  q11s32 = vmull_s16(d28s16, d16s16);
+  q12s32 = vmull_s16(d29s16, d16s16);
+
+  q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
+  q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+  q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+  q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+  d10s16 = vqrshrn_n_s32(q9s32, 14);
+  d11s16 = vqrshrn_n_s32(q10s32, 14);
+  d12s16 = vqrshrn_n_s32(q11s32, 14);
+  d13s16 = vqrshrn_n_s32(q12s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  // stage 4
+  q8s16 = vaddq_s16(q0s16, q7s16);
+  q9s16 = vaddq_s16(q1s16, q6s16);
+  q10s16 = vaddq_s16(q2s16, q5s16);
+  q11s16 = vaddq_s16(q3s16, q4s16);
+  q12s16 = vsubq_s16(q3s16, q4s16);
+  q13s16 = vsubq_s16(q2s16, q5s16);
+  q14s16 = vsubq_s16(q1s16, q6s16);
+  q15s16 = vsubq_s16(q0s16, q7s16);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+             &q15s16);
+
+  q8s16 = vrshrq_n_s16(q8s16, 5);
+  q9s16 = vrshrq_n_s16(q9s16, 5);
+  q10s16 = vrshrq_n_s16(q10s16, 5);
+  q11s16 = vrshrq_n_s16(q11s16, 5);
+  q12s16 = vrshrq_n_s16(q12s16, 5);
+  q13s16 = vrshrq_n_s16(q13s16, 5);
+  q14s16 = vrshrq_n_s16(q14s16, 5);
+  q15s16 = vrshrq_n_s16(q15s16, 5);
+
+  d1 = d2 = dest;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+
+  q8s16 = q12s16;
+  q9s16 = q13s16;
+  q10s16 = q14s16;
+  q11s16 = q15s16;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+  return;
+}

third_party/aom/aom_dsp/arm/intrapred_neon.c

@@ -0,0 +1,757 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+//------------------------------------------------------------------------------
+// DC 4x4
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_4x4(uint8_t *dst, ptrdiff_t stride, const uint8_t *above,
+                          const uint8_t *left, int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x8_t A = vld1_u8(above);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    sum_top = vcombine_u16(p1, p1);
+  }
+
+  if (do_left) {
+    const uint8x8_t L = vld1_u8(left);   // left border
+    const uint16x4_t p0 = vpaddl_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    sum_left = vcombine_u16(p1, p1);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 3);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 2);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 2);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 4; ++i) {
+      vst1_lane_u32((uint32_t *)(dst + i * stride), vreinterpret_u32_u8(dc), 0);
+    }
+  }
+}
+
+void aom_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  dc_4x4(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  dc_4x4(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  dc_4x4(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_4x4(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 8x8
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_8x8(uint8_t *dst, ptrdiff_t stride, const uint8_t *above,
+                          const uint8_t *left, int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x8_t A = vld1_u8(above);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_top = vcombine_u16(p2, p2);
+  }
+
+  if (do_left) {
+    const uint8x8_t L = vld1_u8(left);   // left border
+    const uint16x4_t p0 = vpaddl_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_left = vcombine_u16(p2, p2);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 4);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 3);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 3);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 8; ++i) {
+      vst1_u32((uint32_t *)(dst + i * stride), vreinterpret_u32_u8(dc));
+    }
+  }
+}
+
+void aom_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  dc_8x8(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  dc_8x8(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  dc_8x8(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_8x8(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 16x16
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_16x16(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A = vld1q_u8(above);  // top row
+    const uint16x8_t p0 = vpaddlq_u8(A);   // cascading summation of the top
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_top = vcombine_u16(p3, p3);
+  }
+
+  if (do_left) {
+    const uint8x16_t L = vld1q_u8(left);  // left row
+    const uint16x8_t p0 = vpaddlq_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_left = vcombine_u16(p3, p3);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 5);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 4);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 4);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 16; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+    }
+  }
+}
+
+void aom_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_16x16(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_16x16(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_16x16(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_16x16(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 32x32
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_32x32(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A0 = vld1q_u8(above);  // top row
+    const uint8x16_t A1 = vld1q_u8(above + 16);
+    const uint16x8_t p0 = vpaddlq_u8(A0);  // cascading summation of the top
+    const uint16x8_t p1 = vpaddlq_u8(A1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_top = vcombine_u16(p5, p5);
+  }
+
+  if (do_left) {
+    const uint8x16_t L0 = vld1q_u8(left);  // left row
+    const uint8x16_t L1 = vld1q_u8(left + 16);
+    const uint16x8_t p0 = vpaddlq_u8(L0);  // cascading summation of the left
+    const uint16x8_t p1 = vpaddlq_u8(L1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_left = vcombine_u16(p5, p5);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 6);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 5);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 5);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 32; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+      vst1q_u8(dst + i * stride + 16, dc);
+    }
+  }
+}
+
+void aom_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_32x32(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_32x32(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_32x32(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_32x32(dst, stride, NULL, NULL, 0, 0);
+}
+
+// -----------------------------------------------------------------------------
+
+void aom_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const uint8x8_t XABCD_u8 = vld1_u8(above - 1);
+  const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8);
+  const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32);
+  const uint32x2_t zero = vdup_n_u32(0);
+  const uint32x2_t IJKL = vld1_lane_u32((const uint32_t *)left, zero, 0);
+  const uint8x8_t IJKL_u8 = vreinterpret_u8_u32(IJKL);
+  const uint64x1_t LKJI____ = vreinterpret_u64_u8(vrev32_u8(IJKL_u8));
+  const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC);
+  const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8));
+  const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16));
+  const uint8_t D = vget_lane_u8(XABCD_u8, 4);
+  const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6);
+  const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC);
+  const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8);
+  const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_);
+  const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
+  const uint32x2_t r3 = vreinterpret_u32_u8(avg2);
+  const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
+  const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
+  const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
+  vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
+  vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
+  vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
+  vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
+}
+
+#if !HAVE_NEON_ASM
+
+void aom_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint32x2_t d0u32 = vdup_n_u32(0);
+  (void)left;
+
+  d0u32 = vld1_lane_u32((const uint32_t *)above, d0u32, 0);
+  for (i = 0; i < 4; i++, dst += stride)
+    vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+}
+
+void aom_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  (void)left;
+
+  d0u8 = vld1_u8(above);
+  for (i = 0; i < 8; i++, dst += stride) vst1_u8(dst, d0u8);
+}
+
+void aom_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  for (i = 0; i < 16; i++, dst += stride) vst1q_u8(dst, q0u8);
+}
+
+void aom_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  q1u8 = vld1q_u8(above + 16);
+  for (i = 0; i < 32; i++, dst += stride) {
+    vst1q_u8(dst, q0u8);
+    vst1q_u8(dst + 16, q1u8);
+  }
+}
+
+void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d1u32 = vdup_n_u32(0);
+  (void)above;
+
+  d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+}
+
+void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint64x1_t d1u64 = vdup_n_u64(0);
+  (void)above;
+
+  d1u64 = vld1_u64((const uint64_t *)left);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7);
+  vst1_u8(dst, d0u8);
+}
+
+void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  q1u8 = vld1q_u8(left);
+  d2u8 = vget_low_u8(q1u8);
+  for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+    q0u8 = vdupq_lane_u8(d2u8, 0);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 1);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 2);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 3);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 4);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 5);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 6);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 7);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+  }
+}
+
+void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  for (k = 0; k < 2; k++, left += 16) {
+    q1u8 = vld1q_u8(left);
+    d2u8 = vget_low_u8(q1u8);
+    for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+      q0u8 = vdupq_lane_u8(d2u8, 0);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 1);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 2);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 3);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 4);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 5);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 6);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 7);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+    }
+  }
+}
+
+void aom_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint16x8_t q1u16, q3u16;
+  int16x8_t q1s16;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d2u32 = vdup_n_u32(0);
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d2u32 = vld1_lane_u32((const uint32_t *)above, d2u32, 0);
+  q3u16 = vsubl_u8(vreinterpret_u8_u32(d2u32), d0u8);
+  for (i = 0; i < 4; i++, dst += stride) {
+    q1u16 = vdupq_n_u16((uint16_t)left[i]);
+    q1s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q1u16), vreinterpretq_s16_u16(q3u16));
+    d0u8 = vqmovun_s16(q1s16);
+    vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  }
+}
+
+void aom_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint16x8_t q0u16, q3u16, q10u16;
+  int16x8_t q0s16;
+  uint16x4_t d20u16;
+  uint8x8_t d0u8, d2u8, d30u8;
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d30u8 = vld1_u8(left);
+  d2u8 = vld1_u8(above);
+  q10u16 = vmovl_u8(d30u8);
+  q3u16 = vsubl_u8(d2u8, d0u8);
+  d20u16 = vget_low_u16(q10u16);
+  for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+    q0u16 = vdupq_lane_u16(d20u16, 0);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 1);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 2);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 3);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+  }
+}
+
+void aom_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q2u16, q3u16, q8u16, q10u16;
+  uint8x16_t q0u8, q1u8;
+  int16x8_t q0s16, q1s16, q8s16, q11s16;
+  uint16x4_t d20u16;
+  uint8x8_t d2u8, d3u8, d18u8, d22u8, d23u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q3u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  for (k = 0; k < 2; k++, left += 8) {
+    d18u8 = vld1_u8(left);
+    q10u16 = vmovl_u8(d18u8);
+    d20u16 = vget_low_u16(q10u16);
+    for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+      q0u16 = vdupq_lane_u16(d20u16, 0);
+      q8u16 = vdupq_lane_u16(d20u16, 1);
+      q1s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q2u16));
+      q0s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q3u16));
+      q11s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q2u16));
+      q8s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d20u16, 2);
+      q8u16 = vdupq_lane_u16(d20u16, 3);
+      q1s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q2u16));
+      q0s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q3u16));
+      q11s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q2u16));
+      q8s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+    }
+  }
+}
+
+void aom_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q3u16, q8u16, q9u16, q10u16, q11u16;
+  uint8x16_t q0u8, q1u8, q2u8;
+  int16x8_t q12s16, q13s16, q14s16, q15s16;
+  uint16x4_t d6u16;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d26u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u8 = vld1q_u8(above + 16);
+  q8u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q9u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  q10u16 = vsubl_u8(vget_low_u8(q2u8), vget_low_u8(q0u8));
+  q11u16 = vsubl_u8(vget_high_u8(q2u8), vget_high_u8(q0u8));
+  for (k = 0; k < 4; k++, left += 8) {
+    d26u8 = vld1_u8(left);
+    q3u16 = vmovl_u8(d26u8);
+    d6u16 = vget_low_u16(q3u16);
+    for (j = 0; j < 2; j++, d6u16 = vget_high_u16(q3u16)) {
+      q0u16 = vdupq_lane_u16(d6u16, 0);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 1);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 2);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 3);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+    }
+  }
+}
+#endif  // !HAVE_NEON_ASM

third_party/aom/aom_dsp/arm/intrapred_neon_asm.asm

@@ -0,0 +1,633 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_v_predictor_4x4_neon|
+    EXPORT  |aom_v_predictor_8x8_neon|
+    EXPORT  |aom_v_predictor_16x16_neon|
+    EXPORT  |aom_v_predictor_32x32_neon|
+    EXPORT  |aom_h_predictor_4x4_neon|
+    EXPORT  |aom_h_predictor_8x8_neon|
+    EXPORT  |aom_h_predictor_16x16_neon|
+    EXPORT  |aom_h_predictor_32x32_neon|
+    EXPORT  |aom_tm_predictor_4x4_neon|
+    EXPORT  |aom_tm_predictor_8x8_neon|
+    EXPORT  |aom_tm_predictor_16x16_neon|
+    EXPORT  |aom_tm_predictor_32x32_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_4x4_neon| PROC
+    vld1.32             {d0[0]}, [r2]
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_v_predictor_4x4_neon|
+
+;void aom_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_8x8_neon| PROC
+    vld1.8              {d0}, [r2]
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_v_predictor_8x8_neon|
+
+;void aom_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_16x16_neon| PROC
+    vld1.8              {q0}, [r2]
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_v_predictor_16x16_neon|
+
+;void aom_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_32x32_neon| PROC
+    vld1.8              {q0, q1}, [r2]
+    mov                 r2, #2
+loop_v
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_v
+    bx                  lr
+    ENDP                ; |aom_v_predictor_32x32_neon|
+
+;void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_4x4_neon| PROC
+    vld1.32             {d1[0]}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_h_predictor_4x4_neon|
+
+;void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_8x8_neon| PROC
+    vld1.64             {d1}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[4]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[5]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[6]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[7]
+    vst1.64             {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_h_predictor_8x8_neon|
+
+;void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_16x16_neon| PROC
+    vld1.8              {q1}, [r3]
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_h_predictor_16x16_neon|
+
+;void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_32x32_neon| PROC
+    sub                 r1, r1, #16
+    mov                 r2, #2
+loop_h
+    vld1.8              {q1}, [r3]!
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_h
+    bx                  lr
+    ENDP                ; |aom_h_predictor_32x32_neon|
+
+;void aom_tm_predictor_4x4_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_4x4_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.u8             {d0[]}, [r12]
+
+    ; Load above 4 pixels
+    vld1.32             {d2[0]}, [r2]
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]!
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+
+    ; 3rd row and 4th row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_4x4_neon|
+
+;void aom_tm_predictor_8x8_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_8x8_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; preload 8 left
+    vld1.8              {d30}, [r3]
+
+    ; Load above 8 pixels
+    vld1.64             {d2}, [r2]
+
+    vmovl.u8            q10, d30
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vdup.16             q0, d20[0]
+    vdup.16             q1, d20[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 3rd row and 4th row
+    vdup.16             q8, d20[2]
+    vdup.16             q9, d20[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    ; 5th row and 6th row
+    vdup.16             q0, d21[0]
+    vdup.16             q1, d21[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 7th row and 8th row
+    vdup.16             q8, d21[2]
+    vdup.16             q9, d21[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_8x8_neon|
+
+;void aom_tm_predictor_16x16_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_16x16_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 8 pixels
+    vld1.8              {q1}, [r2]
+
+    ; preload 8 left into r12
+    vld1.8              {d18}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q2, d2, d0
+    vsubl.u8            q3, d3, d0
+
+    vmovl.u8            q10, d18
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 2 times to process 16 rows.
+    mov                 r2, #2
+
+loop_16x16_neon
+    ; Process two rows.
+    vdup.16             q0, d20[0]
+    vdup.16             q8, d20[1]
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d20[2]                  ; proload next 2 rows data
+    vdup.16             q8, d20[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[0]                  ; proload next 2 rows data
+    vdup.16             q8, d21[1]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[2]                  ; proload next 2 rows data
+    vdup.16             q8, d21[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vld1.8              {d18}, [r3]!                  ; preload 8 left into r12
+    vmovl.u8            q10, d18
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_16x16_neon
+
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_16x16_neon|
+
+;void aom_tm_predictor_32x32_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                  const uint8_t *above,
+;                                  const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_32x32_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 32 pixels
+    vld1.8              {q1}, [r2]!
+    vld1.8              {q2}, [r2]
+
+    ; preload 8 left pixels
+    vld1.8              {d26}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q8, d2, d0
+    vsubl.u8            q9, d3, d0
+    vsubl.u8            q10, d4, d0
+    vsubl.u8            q11, d5, d0
+
+    vmovl.u8            q3, d26
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 4 times to process 32 rows.
+    mov                 r2, #4
+
+loop_32x32_neon
+    ; Process two rows.
+    vdup.16             q0, d6[0]
+    vdup.16             q2, d6[1]
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q1, d6[2]
+    vdup.16             q2, d6[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q1, q8
+    vadd.s16            q13, q1, q9
+    vadd.s16            q14, q1, q10
+    vadd.s16            q15, q1, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[0]
+    vdup.16             q2, d7[1]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[2]
+    vdup.16             q2, d7[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vld1.8              {d0}, [r3]!                   ; preload 8 left pixels
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vmovl.u8            q3, d0
+    vst1.64             {d24-d27}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_32x32_neon
+
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_32x32_neon|
+
+    END

third_party/aom/aom_dsp/arm/loopfilter_16_neon.asm

@@ -0,0 +1,202 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_4_dual_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_lpf_horizontal_4_dual_neon(uint8_t *s, int p,
+;                                    const uint8_t *blimit0,
+;                                    const uint8_t *limit0,
+;                                    const uint8_t *thresh0,
+;                                    const uint8_t *blimit1,
+;                                    const uint8_t *limit1,
+;                                    const uint8_t *thresh1)
+; r0    uint8_t *s,
+; r1    int p,
+; r2    const uint8_t *blimit0,
+; r3    const uint8_t *limit0,
+; sp    const uint8_t *thresh0,
+; sp+4  const uint8_t *blimit1,
+; sp+8  const uint8_t *limit1,
+; sp+12 const uint8_t *thresh1,
+
+|aom_lpf_horizontal_4_dual_neon| PROC
+    push        {lr}
+
+    ldr         r12, [sp, #4]              ; load thresh0
+    vld1.8      {d0}, [r2]                 ; load blimit0 to first half q
+    vld1.8      {d2}, [r3]                 ; load limit0 to first half q
+
+    add         r1, r1, r1                 ; double pitch
+    ldr         r2, [sp, #8]               ; load blimit1
+
+    vld1.8      {d4}, [r12]                ; load thresh0 to first half q
+
+    ldr         r3, [sp, #12]              ; load limit1
+    ldr         r12, [sp, #16]             ; load thresh1
+    vld1.8      {d1}, [r2]                 ; load blimit1 to 2nd half q
+
+    sub         r2, r0, r1, lsl #1         ; s[-4 * p]
+
+    vld1.8      {d3}, [r3]                 ; load limit1 to 2nd half q
+    vld1.8      {d5}, [r12]                ; load thresh1 to 2nd half q
+
+    vpush       {d8-d15}                   ; save neon registers
+
+    add         r3, r2, r1, lsr #1         ; s[-3 * p]
+
+    vld1.u8     {q3}, [r2@64], r1          ; p3
+    vld1.u8     {q4}, [r3@64], r1          ; p2
+    vld1.u8     {q5}, [r2@64], r1          ; p1
+    vld1.u8     {q6}, [r3@64], r1          ; p0
+    vld1.u8     {q7}, [r2@64], r1          ; q0
+    vld1.u8     {q8}, [r3@64], r1          ; q1
+    vld1.u8     {q9}, [r2@64]              ; q2
+    vld1.u8     {q10}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          aom_loop_filter_neon_16
+
+    vst1.u8     {q5}, [r2@64], r1          ; store op1
+    vst1.u8     {q6}, [r3@64], r1          ; store op0
+    vst1.u8     {q7}, [r2@64], r1          ; store oq0
+    vst1.u8     {q8}, [r3@64], r1          ; store oq1
+
+    vpop        {d8-d15}                   ; restore neon registers
+
+    pop         {pc}
+    ENDP        ; |aom_lpf_horizontal_4_dual_neon|
+
+; void aom_loop_filter_neon_16();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. This function uses
+; registers d8-d15, so the calling function must save those registers.
+;
+; r0-r3, r12 PRESERVE
+; q0    blimit
+; q1    limit
+; q2    thresh
+; q3    p3
+; q4    p2
+; q5    p1
+; q6    p0
+; q7    q0
+; q8    q1
+; q9    q2
+; q10   q3
+;
+; Outputs:
+; q5    op1
+; q6    op0
+; q7    oq0
+; q8    oq1
+|aom_loop_filter_neon_16| PROC
+
+    ; filter_mask
+    vabd.u8     q11, q3, q4                 ; m1 = abs(p3 - p2)
+    vabd.u8     q12, q4, q5                 ; m2 = abs(p2 - p1)
+    vabd.u8     q13, q5, q6                 ; m3 = abs(p1 - p0)
+    vabd.u8     q14, q8, q7                 ; m4 = abs(q1 - q0)
+    vabd.u8     q3, q9, q8                  ; m5 = abs(q2 - q1)
+    vabd.u8     q4, q10, q9                 ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     q11, q11, q12               ; m7 = max(m1, m2)
+    vmax.u8     q12, q13, q14               ; m8 = max(m3, m4)
+
+    vabd.u8     q9, q6, q7                  ; abs(p0 - q0)
+
+    vmax.u8     q3, q3, q4                  ; m9 = max(m5, m6)
+
+    vmov.u8     q10, #0x80
+
+    vmax.u8     q15, q11, q12               ; m10 = max(m7, m8)
+
+    vcgt.u8     q13, q13, q2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     q14, q14, q2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     q15, q15, q3                ; m11 = max(m10, m9)
+
+    vabd.u8     q2, q5, q8                  ; a = abs(p1 - q1)
+    vqadd.u8    q9, q9, q9                  ; b = abs(p0 - q0) * 2
+
+    veor        q7, q7, q10                 ; qs0
+
+    vcge.u8     q15, q1, q15                ; abs(m11) > limit
+
+    vshr.u8     q2, q2, #1                  ; a = a / 2
+    veor        q6, q6, q10                 ; ps0
+
+    veor        q5, q5, q10                 ; ps1
+    vqadd.u8    q9, q9, q2                  ; a = b + a
+
+    veor        q8, q8, q10                 ; qs1
+
+    vmov.u16    q4, #3
+
+    vsubl.s8    q2, d14, d12                ; ( qs0 - ps0)
+    vsubl.s8    q11, d15, d13
+
+    vcge.u8     q9, q0, q9                  ; a > blimit
+
+    vqsub.s8    q1, q5, q8                  ; filter = clamp(ps1-qs1)
+    vorr        q14, q13, q14               ; hev
+
+    vmul.i16    q2, q2, q4                  ; 3 * ( qs0 - ps0)
+    vmul.i16    q11, q11, q4
+
+    vand        q1, q1, q14                 ; filter &= hev
+    vand        q15, q15, q9                ; mask
+
+    vmov.u8     q4, #3
+
+    vaddw.s8    q2, q2, d2                  ; filter + 3 * (qs0 - ps0)
+    vaddw.s8    q11, q11, d3
+
+    vmov.u8     q9, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d2, q2
+    vqmovn.s16  d3, q11
+    vand        q1, q1, q15                 ; filter &= mask
+
+    vqadd.s8    q2, q1, q4                  ; filter2 = clamp(filter+3)
+    vqadd.s8    q1, q1, q9                  ; filter1 = clamp(filter+4)
+    vshr.s8     q2, q2, #3                  ; filter2 >>= 3
+    vshr.s8     q1, q1, #3                  ; filter1 >>= 3
+
+
+    vqadd.s8    q11, q6, q2                 ; u = clamp(ps0 + filter2)
+    vqsub.s8    q0, q7, q1                  ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    q1, q1, #1                  ; filter = ++filter1 >> 1
+
+    veor        q7, q0,  q10                ; *oq0 = u^0x80
+
+    vbic        q1, q1, q14                 ; filter &= ~hev
+
+    vqadd.s8    q13, q5, q1                 ; u = clamp(ps1 + filter)
+    vqsub.s8    q12, q8, q1                 ; u = clamp(qs1 - filter)
+
+    veor        q6, q11, q10                ; *op0 = u^0x80
+    veor        q5, q13, q10                ; *op1 = u^0x80
+    veor        q8, q12, q10                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |aom_loop_filter_neon_16|
+
+    END

third_party/aom/aom_dsp/arm/loopfilter_16_neon.c

@@ -0,0 +1,174 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+static INLINE void loop_filter_neon_16(uint8x16_t qblimit,  // blimit
+                                       uint8x16_t qlimit,   // limit
+                                       uint8x16_t qthresh,  // thresh
+                                       uint8x16_t q3,       // p3
+                                       uint8x16_t q4,       // p2
+                                       uint8x16_t q5,       // p1
+                                       uint8x16_t q6,       // p0
+                                       uint8x16_t q7,       // q0
+                                       uint8x16_t q8,       // q1
+                                       uint8x16_t q9,       // q2
+                                       uint8x16_t q10,      // q3
+                                       uint8x16_t *q5r,     // p1
+                                       uint8x16_t *q6r,     // p0
+                                       uint8x16_t *q7r,     // q0
+                                       uint8x16_t *q8r) {   // q1
+  uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+  int16x8_t q2s16, q11s16;
+  uint16x8_t q4u16;
+  int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
+  int8x8_t d2s8, d3s8;
+
+  q11u8 = vabdq_u8(q3, q4);
+  q12u8 = vabdq_u8(q4, q5);
+  q13u8 = vabdq_u8(q5, q6);
+  q14u8 = vabdq_u8(q8, q7);
+  q3 = vabdq_u8(q9, q8);
+  q4 = vabdq_u8(q10, q9);
+
+  q11u8 = vmaxq_u8(q11u8, q12u8);
+  q12u8 = vmaxq_u8(q13u8, q14u8);
+  q3 = vmaxq_u8(q3, q4);
+  q15u8 = vmaxq_u8(q11u8, q12u8);
+
+  q9 = vabdq_u8(q6, q7);
+
+  // aom_hevmask
+  q13u8 = vcgtq_u8(q13u8, qthresh);
+  q14u8 = vcgtq_u8(q14u8, qthresh);
+  q15u8 = vmaxq_u8(q15u8, q3);
+
+  q2u8 = vabdq_u8(q5, q8);
+  q9 = vqaddq_u8(q9, q9);
+
+  q15u8 = vcgeq_u8(qlimit, q15u8);
+
+  // aom_filter() function
+  // convert to signed
+  q10 = vdupq_n_u8(0x80);
+  q8 = veorq_u8(q8, q10);
+  q7 = veorq_u8(q7, q10);
+  q6 = veorq_u8(q6, q10);
+  q5 = veorq_u8(q5, q10);
+
+  q2u8 = vshrq_n_u8(q2u8, 1);
+  q9 = vqaddq_u8(q9, q2u8);
+
+  q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+                   vget_low_s8(vreinterpretq_s8_u8(q6)));
+  q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+                    vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+  q9 = vcgeq_u8(qblimit, q9);
+
+  q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5), vreinterpretq_s8_u8(q8));
+
+  q14u8 = vorrq_u8(q13u8, q14u8);
+
+  q4u16 = vdupq_n_u16(3);
+  q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
+  q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
+
+  q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
+  q15u8 = vandq_u8(q15u8, q9);
+
+  q1s8 = vreinterpretq_s8_u8(q1u8);
+  q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
+  q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
+
+  q4 = vdupq_n_u8(3);
+  q9 = vdupq_n_u8(4);
+  // aom_filter = clamp(aom_filter + 3 * ( qs0 - ps0))
+  d2s8 = vqmovn_s16(q2s16);
+  d3s8 = vqmovn_s16(q11s16);
+  q1s8 = vcombine_s8(d2s8, d3s8);
+  q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
+  q1s8 = vreinterpretq_s8_u8(q1u8);
+
+  q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
+  q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
+  q2s8 = vshrq_n_s8(q2s8, 3);
+  q1s8 = vshrq_n_s8(q1s8, 3);
+
+  q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
+  q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
+
+  q1s8 = vrshrq_n_s8(q1s8, 1);
+  q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+  q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
+  q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
+
+  *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
+  *q7r = veorq_u8(vreinterpretq_u8_s8(q0s8), q10);
+  *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
+  *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
+  return;
+}
+
+void aom_lpf_horizontal_4_dual_neon(
+    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
+    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
+    const uint8_t *limit1, const uint8_t *thresh1) {
+  uint8x8_t dblimit0, dlimit0, dthresh0, dblimit1, dlimit1, dthresh1;
+  uint8x16_t qblimit, qlimit, qthresh;
+  uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
+
+  dblimit0 = vld1_u8(blimit0);
+  dlimit0 = vld1_u8(limit0);
+  dthresh0 = vld1_u8(thresh0);
+  dblimit1 = vld1_u8(blimit1);
+  dlimit1 = vld1_u8(limit1);
+  dthresh1 = vld1_u8(thresh1);
+  qblimit = vcombine_u8(dblimit0, dblimit1);
+  qlimit = vcombine_u8(dlimit0, dlimit1);
+  qthresh = vcombine_u8(dthresh0, dthresh1);
+
+  s -= (p << 2);
+
+  q3u8 = vld1q_u8(s);
+  s += p;
+  q4u8 = vld1q_u8(s);
+  s += p;
+  q5u8 = vld1q_u8(s);
+  s += p;
+  q6u8 = vld1q_u8(s);
+  s += p;
+  q7u8 = vld1q_u8(s);
+  s += p;
+  q8u8 = vld1q_u8(s);
+  s += p;
+  q9u8 = vld1q_u8(s);
+  s += p;
+  q10u8 = vld1q_u8(s);
+
+  loop_filter_neon_16(qblimit, qlimit, qthresh, q3u8, q4u8, q5u8, q6u8, q7u8,
+                      q8u8, q9u8, q10u8, &q5u8, &q6u8, &q7u8, &q8u8);
+
+  s -= (p * 5);
+  vst1q_u8(s, q5u8);
+  s += p;
+  vst1q_u8(s, q6u8);
+  s += p;
+  vst1q_u8(s, q7u8);
+  s += p;
+  vst1q_u8(s, q8u8);
+  return;
+}

third_party/aom/aom_dsp/arm/loopfilter_4_neon.asm

@@ -0,0 +1,252 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_4_neon|
+    EXPORT  |aom_lpf_vertical_4_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently aom only works on iterations 8 at a time. The aom loop filter
+; works on 16 iterations at a time.
+;
+; void aom_lpf_horizontal_4_neon(uint8_t *s,
+;                                int p /* pitch */,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_horizontal_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r2, [sp, #4]               ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+    sub         r2, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r3, r2, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r2@64], r1          ; p3
+    vld1.u8     {d4}, [r3@64], r1          ; p2
+    vld1.u8     {d5}, [r2@64], r1          ; p1
+    vld1.u8     {d6}, [r3@64], r1          ; p0
+    vld1.u8     {d7}, [r2@64], r1          ; q0
+    vld1.u8     {d16}, [r3@64], r1         ; q1
+    vld1.u8     {d17}, [r2@64]             ; q2
+    vld1.u8     {d18}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          aom_loop_filter_neon
+
+    vst1.u8     {d4}, [r2@64], r1          ; store op1
+    vst1.u8     {d5}, [r3@64], r1          ; store op0
+    vst1.u8     {d6}, [r2@64], r1          ; store oq0
+    vst1.u8     {d7}, [r3@64], r1          ; store oq1
+
+    pop         {pc}
+    ENDP        ; |aom_lpf_horizontal_4_neon|
+
+; Currently aom only works on iterations 8 at a time. The aom loop filter
+; works on 16 iterations at a time.
+;
+; void aom_lpf_vertical_4_neon(uint8_t *s,
+;                              int p /* pitch */,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_vertical_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #4]              ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    bl          aom_loop_filter_neon
+
+    sub         r0, r0, #2
+
+    ;store op1, op0, oq0, oq1
+    vst4.8      {d4[0], d5[0], d6[0], d7[0]}, [r0], r1
+    vst4.8      {d4[1], d5[1], d6[1], d7[1]}, [r0], r1
+    vst4.8      {d4[2], d5[2], d6[2], d7[2]}, [r0], r1
+    vst4.8      {d4[3], d5[3], d6[3], d7[3]}, [r0], r1
+    vst4.8      {d4[4], d5[4], d6[4], d7[4]}, [r0], r1
+    vst4.8      {d4[5], d5[5], d6[5], d7[5]}, [r0], r1
+    vst4.8      {d4[6], d5[6], d6[6], d7[6]}, [r0], r1
+    vst4.8      {d4[7], d5[7], d6[7], d7[7]}, [r0]
+
+    pop         {pc}
+    ENDP        ; |aom_lpf_vertical_4_neon|
+
+; void aom_loop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d4    op1
+; d5    op0
+; d6    oq0
+; d7    oq1
+|aom_loop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                 ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                 ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                 ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7                ; m4 = abs(q1 - q0)
+    vabd.u8     d3, d17, d16                ; m5 = abs(q2 - q1)
+    vabd.u8     d4, d18, d17                ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20               ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22               ; m2 = max(m3, m4)
+
+    vabd.u8     d17, d6, d7                 ; abs(p0 - q0)
+
+    vmax.u8     d3, d3, d4                  ; m3 = max(m5, m6)
+
+    vmov.u8     d18, #0x80
+
+    vmax.u8     d23, d19, d20               ; m1 = max(m1, m2)
+
+    ; hevmask
+    vcgt.u8     d21, d21, d2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     d23, d23, d3                ; m1 = max(m1, m3)
+
+    vabd.u8     d28, d5, d16                ; a = abs(p1 - q1)
+    vqadd.u8    d17, d17, d17               ; b = abs(p0 - q0) * 2
+
+    veor        d7, d7, d18                 ; qs0
+
+    vcge.u8     d23, d1, d23                ; abs(m1) > limit
+
+    ; filter() function
+    ; convert to signed
+
+    vshr.u8     d28, d28, #1                ; a = a / 2
+    veor        d6, d6, d18                 ; ps0
+
+    veor        d5, d5, d18                 ; ps1
+    vqadd.u8    d17, d17, d28               ; a = b + a
+
+    veor        d16, d16, d18               ; qs1
+
+    vmov.u8     d19, #3
+
+    vsub.s8     d28, d7, d6                 ; ( qs0 - ps0)
+
+    vcge.u8     d17, d0, d17                ; a > blimit
+
+    vqsub.s8    d27, d5, d16                ; filter = clamp(ps1-qs1)
+    vorr        d22, d21, d22               ; hevmask
+
+    vmull.s8    q12, d28, d19               ; 3 * ( qs0 - ps0)
+
+    vand        d27, d27, d22               ; filter &= hev
+    vand        d23, d23, d17               ; filter_mask
+
+    vaddw.s8    q12, q12, d27               ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d17, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d27, q12
+
+    vand        d27, d27, d23               ; filter &= mask
+
+    vqadd.s8    d28, d27, d19               ; filter2 = clamp(filter+3)
+    vqadd.s8    d27, d27, d17               ; filter1 = clamp(filter+4)
+    vshr.s8     d28, d28, #3                ; filter2 >>= 3
+    vshr.s8     d27, d27, #3                ; filter1 >>= 3
+
+    vqadd.s8    d19, d6, d28                ; u = clamp(ps0 + filter2)
+    vqsub.s8    d26, d7, d27                ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    d27, d27, #1                ; filter = ++filter1 >> 1
+
+    veor        d6, d26, d18                ; *oq0 = u^0x80
+
+    vbic        d27, d27, d22               ; filter &= ~hev
+
+    vqadd.s8    d21, d5, d27                ; u = clamp(ps1 + filter)
+    vqsub.s8    d20, d16, d27               ; u = clamp(qs1 - filter)
+
+    veor        d5, d19, d18                ; *op0 = u^0x80
+    veor        d4, d21, d18                ; *op1 = u^0x80
+    veor        d7, d20, d18                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |aom_loop_filter_neon|
+
+    END

third_party/aom/aom_dsp/arm/loopfilter_4_neon.c

@@ -0,0 +1,250 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+
+static INLINE void loop_filter_neon(uint8x8_t dblimit,   // flimit
+                                    uint8x8_t dlimit,    // limit
+                                    uint8x8_t dthresh,   // thresh
+                                    uint8x8_t d3u8,      // p3
+                                    uint8x8_t d4u8,      // p2
+                                    uint8x8_t d5u8,      // p1
+                                    uint8x8_t d6u8,      // p0
+                                    uint8x8_t d7u8,      // q0
+                                    uint8x8_t d16u8,     // q1
+                                    uint8x8_t d17u8,     // q2
+                                    uint8x8_t d18u8,     // q3
+                                    uint8x8_t *d4ru8,    // p1
+                                    uint8x8_t *d5ru8,    // p0
+                                    uint8x8_t *d6ru8,    // q0
+                                    uint8x8_t *d7ru8) {  // q1
+  uint8x8_t d19u8, d20u8, d21u8, d22u8, d23u8, d27u8, d28u8;
+  int16x8_t q12s16;
+  int8x8_t d19s8, d20s8, d21s8, d26s8, d27s8, d28s8;
+
+  d19u8 = vabd_u8(d3u8, d4u8);
+  d20u8 = vabd_u8(d4u8, d5u8);
+  d21u8 = vabd_u8(d5u8, d6u8);
+  d22u8 = vabd_u8(d16u8, d7u8);
+  d3u8 = vabd_u8(d17u8, d16u8);
+  d4u8 = vabd_u8(d18u8, d17u8);
+
+  d19u8 = vmax_u8(d19u8, d20u8);
+  d20u8 = vmax_u8(d21u8, d22u8);
+  d3u8 = vmax_u8(d3u8, d4u8);
+  d23u8 = vmax_u8(d19u8, d20u8);
+
+  d17u8 = vabd_u8(d6u8, d7u8);
+
+  d21u8 = vcgt_u8(d21u8, dthresh);
+  d22u8 = vcgt_u8(d22u8, dthresh);
+  d23u8 = vmax_u8(d23u8, d3u8);
+
+  d28u8 = vabd_u8(d5u8, d16u8);
+  d17u8 = vqadd_u8(d17u8, d17u8);
+
+  d23u8 = vcge_u8(dlimit, d23u8);
+
+  d18u8 = vdup_n_u8(0x80);
+  d5u8 = veor_u8(d5u8, d18u8);
+  d6u8 = veor_u8(d6u8, d18u8);
+  d7u8 = veor_u8(d7u8, d18u8);
+  d16u8 = veor_u8(d16u8, d18u8);
+
+  d28u8 = vshr_n_u8(d28u8, 1);
+  d17u8 = vqadd_u8(d17u8, d28u8);
+
+  d19u8 = vdup_n_u8(3);
+
+  d28s8 = vsub_s8(vreinterpret_s8_u8(d7u8), vreinterpret_s8_u8(d6u8));
+
+  d17u8 = vcge_u8(dblimit, d17u8);
+
+  d27s8 = vqsub_s8(vreinterpret_s8_u8(d5u8), vreinterpret_s8_u8(d16u8));
+
+  d22u8 = vorr_u8(d21u8, d22u8);
+
+  q12s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+  d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d22u8);
+  d23u8 = vand_u8(d23u8, d17u8);
+
+  q12s16 = vaddw_s8(q12s16, vreinterpret_s8_u8(d27u8));
+
+  d17u8 = vdup_n_u8(4);
+
+  d27s8 = vqmovn_s16(q12s16);
+  d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d23u8);
+  d27s8 = vreinterpret_s8_u8(d27u8);
+
+  d28s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d19u8));
+  d27s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d17u8));
+  d28s8 = vshr_n_s8(d28s8, 3);
+  d27s8 = vshr_n_s8(d27s8, 3);
+
+  d19s8 = vqadd_s8(vreinterpret_s8_u8(d6u8), d28s8);
+  d26s8 = vqsub_s8(vreinterpret_s8_u8(d7u8), d27s8);
+
+  d27s8 = vrshr_n_s8(d27s8, 1);
+  d27s8 = vbic_s8(d27s8, vreinterpret_s8_u8(d22u8));
+
+  d21s8 = vqadd_s8(vreinterpret_s8_u8(d5u8), d27s8);
+  d20s8 = vqsub_s8(vreinterpret_s8_u8(d16u8), d27s8);
+
+  *d4ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d18u8);
+  *d5ru8 = veor_u8(vreinterpret_u8_s8(d19s8), d18u8);
+  *d6ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d18u8);
+  *d7ru8 = veor_u8(vreinterpret_u8_s8(d20s8), d18u8);
+  return;
+}
+
+void aom_lpf_horizontal_4_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+  uint8_t *s, *psrc;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  psrc = src - (pitch << 2);
+  for (i = 0; i < 1; i++) {
+    s = psrc + i * 8;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    loop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                     d16u8, d17u8, d18u8, &d4u8, &d5u8, &d6u8, &d7u8);
+
+    s -= (pitch * 5);
+    vst1_u8(s, d4u8);
+    s += pitch;
+    vst1_u8(s, d5u8);
+    s += pitch;
+    vst1_u8(s, d6u8);
+    s += pitch;
+    vst1_u8(s, d7u8);
+  }
+  return;
+}
+
+void aom_lpf_vertical_4_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  int i, pitch8;
+  uint8_t *s;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+  uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+  uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+  uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+  uint8x8x4_t d4Result;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  pitch8 = pitch * 8;
+  for (i = 0; i < 1; i++, src += pitch8) {
+    s = src - (i + 1) * 4;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8), vreinterpret_u32_u8(d7u8));
+    d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8), vreinterpret_u32_u8(d16u8));
+    d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8), vreinterpret_u32_u8(d17u8));
+    d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8), vreinterpret_u32_u8(d18u8));
+
+    d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                      vreinterpret_u16_u32(d2tmp2.val[0]));
+    d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                      vreinterpret_u16_u32(d2tmp3.val[0]));
+    d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                      vreinterpret_u16_u32(d2tmp2.val[1]));
+    d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                      vreinterpret_u16_u32(d2tmp3.val[1]));
+
+    d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                     vreinterpret_u8_u16(d2tmp5.val[0]));
+    d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                     vreinterpret_u8_u16(d2tmp5.val[1]));
+    d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                      vreinterpret_u8_u16(d2tmp7.val[0]));
+    d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                      vreinterpret_u8_u16(d2tmp7.val[1]));
+
+    d3u8 = d2tmp8.val[0];
+    d4u8 = d2tmp8.val[1];
+    d5u8 = d2tmp9.val[0];
+    d6u8 = d2tmp9.val[1];
+    d7u8 = d2tmp10.val[0];
+    d16u8 = d2tmp10.val[1];
+    d17u8 = d2tmp11.val[0];
+    d18u8 = d2tmp11.val[1];
+
+    loop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                     d16u8, d17u8, d18u8, &d4u8, &d5u8, &d6u8, &d7u8);
+
+    d4Result.val[0] = d4u8;
+    d4Result.val[1] = d5u8;
+    d4Result.val[2] = d6u8;
+    d4Result.val[3] = d7u8;
+
+    src -= 2;
+    vst4_lane_u8(src, d4Result, 0);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 1);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 2);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 3);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 4);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 5);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 6);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 7);
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/loopfilter_8_neon.asm

@@ -0,0 +1,428 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_8_neon|
+    EXPORT  |aom_lpf_vertical_8_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently aom only works on iterations 8 at a time. The aom loop filter
+; works on 16 iterations at a time.
+;
+; void aom_lpf_horizontal_8_neon(uint8_t *s, int p,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_horizontal_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r2, [sp, #12]              ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+    sub         r3, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r2, r3, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r3@64], r1          ; p3
+    vld1.u8     {d4}, [r2@64], r1          ; p2
+    vld1.u8     {d5}, [r3@64], r1          ; p1
+    vld1.u8     {d6}, [r2@64], r1          ; p0
+    vld1.u8     {d7}, [r3@64], r1          ; q0
+    vld1.u8     {d16}, [r2@64], r1         ; q1
+    vld1.u8     {d17}, [r3@64]             ; q2
+    vld1.u8     {d18}, [r2@64], r1         ; q3
+
+    sub         r3, r3, r1, lsl #1
+    sub         r2, r2, r1, lsl #2
+
+    bl          aom_mbloop_filter_neon
+
+    vst1.u8     {d0}, [r2@64], r1          ; store op2
+    vst1.u8     {d1}, [r3@64], r1          ; store op1
+    vst1.u8     {d2}, [r2@64], r1          ; store op0
+    vst1.u8     {d3}, [r3@64], r1          ; store oq0
+    vst1.u8     {d4}, [r2@64], r1          ; store oq1
+    vst1.u8     {d5}, [r3@64], r1          ; store oq2
+
+    pop         {r4-r5, pc}
+
+    ENDP        ; |aom_lpf_horizontal_8_neon|
+
+; void aom_lpf_vertical_8_neon(uint8_t *s,
+;                              int pitch,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_vertical_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #12]             ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    sub         r2, r0, #3
+    add         r3, r0, #1
+
+    bl          aom_mbloop_filter_neon
+
+    ;store op2, op1, op0, oq0
+    vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r2], r1
+    vst4.8      {d0[1], d1[1], d2[1], d3[1]}, [r2], r1
+    vst4.8      {d0[2], d1[2], d2[2], d3[2]}, [r2], r1
+    vst4.8      {d0[3], d1[3], d2[3], d3[3]}, [r2], r1
+    vst4.8      {d0[4], d1[4], d2[4], d3[4]}, [r2], r1
+    vst4.8      {d0[5], d1[5], d2[5], d3[5]}, [r2], r1
+    vst4.8      {d0[6], d1[6], d2[6], d3[6]}, [r2], r1
+    vst4.8      {d0[7], d1[7], d2[7], d3[7]}, [r2]
+
+    ;store oq1, oq2
+    vst2.8      {d4[0], d5[0]}, [r3], r1
+    vst2.8      {d4[1], d5[1]}, [r3], r1
+    vst2.8      {d4[2], d5[2]}, [r3], r1
+    vst2.8      {d4[3], d5[3]}, [r3], r1
+    vst2.8      {d4[4], d5[4]}, [r3], r1
+    vst2.8      {d4[5], d5[5]}, [r3], r1
+    vst2.8      {d4[6], d5[6]}, [r3], r1
+    vst2.8      {d4[7], d5[7]}, [r3]
+
+    pop         {r4-r5, pc}
+    ENDP        ; |aom_lpf_vertical_8_neon|
+
+; void aom_mbloop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d0    op2
+; d1    op1
+; d2    op0
+; d3    oq0
+; d4    oq1
+; d5    oq2
+|aom_mbloop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7               ; m4 = abs(q1 - q0)
+    vabd.u8     d23, d17, d16              ; m5 = abs(q2 - q1)
+    vabd.u8     d24, d18, d17              ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22              ; m2 = max(m3, m4)
+
+    vabd.u8     d25, d6, d4                ; m7 = abs(p0 - p2)
+
+    vmax.u8     d23, d23, d24              ; m3 = max(m5, m6)
+
+    vabd.u8     d26, d7, d17               ; m8 = abs(q0 - q2)
+
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d6, d7                ; m9 = abs(p0 - q0)
+    vabd.u8     d27, d3, d6                ; m10 = abs(p3 - p0)
+    vabd.u8     d28, d18, d7               ; m11 = abs(q3 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d5, d16               ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d1, d19
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; m4 = max(m7, m8)
+    vmax.u8     d26, d27, d28              ; m5 = max(m10, m11)
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+
+    vmax.u8     d25, d25, d26              ; m4 = max(m4, m5)
+
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmax.u8     d20, d20, d25              ; m2 = max(m2, m4)
+
+    vmov.u8     d23, #1
+    vcge.u8     d24, d0, d24               ; a > blimit
+
+    vcgt.u8     d21, d21, d2               ; (abs(p1 - p0) > thresh)*-1
+
+    vcge.u8     d20, d23, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    vcgt.u8     d23, d22, d2               ; (abs(q1 - q0) > thresh)*-1
+
+    vand        d20, d20, d19              ; flat & mask
+
+    vmov.u8     d22, #0x80
+
+    vorr        d23, d21, d23              ; hev
+
+    ; This instruction will truncate the "flat & mask" masks down to 4 bits
+    ; each to fit into one 32 bit arm register. The values are stored in
+    ; q10.64[0].
+    vshrn.u16   d30, q10, #4
+    vmov.u32    r4, d30[0]                 ; flat & mask 4bits
+
+    adds        r5, r4, #1                 ; Check for all 1's
+
+    ; If mask and flat are 1's for all vectors, then we only need to execute
+    ; the power branch for all vectors.
+    beq         power_branch_only
+
+    cmp         r4, #0                     ; Check for 0, set flag for later
+
+    ; mbfilter() function
+    ; filter() function
+    ; convert to signed
+    veor        d21, d7, d22               ; qs0
+    veor        d24, d6, d22               ; ps0
+    veor        d25, d5, d22               ; ps1
+    veor        d26, d16, d22              ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d21, d24              ; ( qs0 - ps0)
+
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+
+    vand        d29, d29, d23              ; filter &= hev
+
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d21, d21, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d23              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    ; If mask and flat are 0's for all vectors, then we only need to execute
+    ; the filter branch for all vectors.
+    beq         filter_branch_only
+
+    ; If mask and flat are mixed then we must perform both branches and
+    ; combine the data.
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d21, d21, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    ; At this point we have already executed the filter branch. The filter
+    ; branch does not set op2 or oq2, so use p2 and q2. Execute the power
+    ; branch and combine the data.
+    vmov.u8     d23, #2
+    vaddl.u8    q14, d6, d7                ; r_op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; r_op2 += p3 * 3
+    vmlal.u8    q14, d4, d23               ; r_op2 += p2 * 2
+
+    vbif        d0, d4, d20                ; op2 |= p2 & ~(flat & mask)
+
+    vaddw.u8    q14, d5                    ; r_op2 += p1
+
+    vbif        d1, d25, d20               ; op1 |= f_op1 & ~(flat & mask)
+
+    vqrshrn.u16 d30, q14, #3               ; r_op2
+
+    vsubw.u8    q14, d3                    ; r_op1 = r_op2 - p3
+    vsubw.u8    q14, d4                    ; r_op1 -= p2
+    vaddw.u8    q14, d5                    ; r_op1 += p1
+    vaddw.u8    q14, d16                   ; r_op1 += q1
+
+    vbif        d2, d24, d20               ; op0 |= f_op0 & ~(flat & mask)
+
+    vqrshrn.u16 d31, q14, #3               ; r_op1
+
+    vsubw.u8    q14, d3                    ; r_op0 = r_op1 - p3
+    vsubw.u8    q14, d5                    ; r_op0 -= p1
+    vaddw.u8    q14, d6                    ; r_op0 += p0
+    vaddw.u8    q14, d17                   ; r_op0 += q2
+
+    vbit        d0, d30, d20               ; op2 |= r_op2 & (flat & mask)
+
+    vqrshrn.u16 d23, q14, #3               ; r_op0
+
+    vsubw.u8    q14, d3                    ; r_oq0 = r_op0 - p3
+    vsubw.u8    q14, d6                    ; r_oq0 -= p0
+    vaddw.u8    q14, d7                    ; r_oq0 += q0
+
+    vbit        d1, d31, d20               ; op1 |= r_op1 & (flat & mask)
+
+    vaddw.u8    q14, d18                   ; oq0 += q3
+
+    vbit        d2, d23, d20               ; op0 |= r_op0 & (flat & mask)
+
+    vqrshrn.u16 d22, q14, #3               ; r_oq0
+
+    vsubw.u8    q14, d4                    ; r_oq1 = r_oq0 - p2
+    vsubw.u8    q14, d7                    ; r_oq1 -= q0
+    vaddw.u8    q14, d16                   ; r_oq1 += q1
+
+    vbif        d3, d21, d20               ; oq0 |= f_oq0 & ~(flat & mask)
+
+    vaddw.u8    q14, d18                   ; r_oq1 += q3
+
+    vbif        d4, d26, d20               ; oq1 |= f_oq1 & ~(flat & mask)
+
+    vqrshrn.u16 d6, q14, #3                ; r_oq1
+
+    vsubw.u8    q14, d5                    ; r_oq2 = r_oq1 - p1
+    vsubw.u8    q14, d16                   ; r_oq2 -= q1
+    vaddw.u8    q14, d17                   ; r_oq2 += q2
+    vaddw.u8    q14, d18                   ; r_oq2 += q3
+
+    vbif        d5, d17, d20               ; oq2 |= q2 & ~(flat & mask)
+
+    vqrshrn.u16 d7, q14, #3                ; r_oq2
+
+    vbit        d3, d22, d20               ; oq0 |= r_oq0 & (flat & mask)
+    vbit        d4, d6, d20                ; oq1 |= r_oq1 & (flat & mask)
+    vbit        d5, d7, d20                ; oq2 |= r_oq2 & (flat & mask)
+
+    bx          lr
+
+power_branch_only
+    vmov.u8     d27, #3
+    vmov.u8     d21, #2
+    vaddl.u8    q14, d6, d7                ; op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; op2 += p3 * 3
+    vmlal.u8    q14, d4, d21               ; op2 += p2 * 2
+    vaddw.u8    q14, d5                    ; op2 += p1
+    vqrshrn.u16 d0, q14, #3                ; op2
+
+    vsubw.u8    q14, d3                    ; op1 = op2 - p3
+    vsubw.u8    q14, d4                    ; op1 -= p2
+    vaddw.u8    q14, d5                    ; op1 += p1
+    vaddw.u8    q14, d16                   ; op1 += q1
+    vqrshrn.u16 d1, q14, #3                ; op1
+
+    vsubw.u8    q14, d3                    ; op0 = op1 - p3
+    vsubw.u8    q14, d5                    ; op0 -= p1
+    vaddw.u8    q14, d6                    ; op0 += p0
+    vaddw.u8    q14, d17                   ; op0 += q2
+    vqrshrn.u16 d2, q14, #3                ; op0
+
+    vsubw.u8    q14, d3                    ; oq0 = op0 - p3
+    vsubw.u8    q14, d6                    ; oq0 -= p0
+    vaddw.u8    q14, d7                    ; oq0 += q0
+    vaddw.u8    q14, d18                   ; oq0 += q3
+    vqrshrn.u16 d3, q14, #3                ; oq0
+
+    vsubw.u8    q14, d4                    ; oq1 = oq0 - p2
+    vsubw.u8    q14, d7                    ; oq1 -= q0
+    vaddw.u8    q14, d16                   ; oq1 += q1
+    vaddw.u8    q14, d18                   ; oq1 += q3
+    vqrshrn.u16 d4, q14, #3                ; oq1
+
+    vsubw.u8    q14, d5                    ; oq2 = oq1 - p1
+    vsubw.u8    q14, d16                   ; oq2 -= q1
+    vaddw.u8    q14, d17                   ; oq2 += q2
+    vaddw.u8    q14, d18                   ; oq2 += q3
+    vqrshrn.u16 d5, q14, #3                ; oq2
+
+    bx          lr
+
+filter_branch_only
+    ; TODO(fgalligan): See if we can rearange registers so we do not need to
+    ; do the 2 vswp.
+    vswp        d0, d4                      ; op2
+    vswp        d5, d17                     ; oq2
+    veor        d2, d24, d22                ; *op0 = u^0x80
+    veor        d3, d21, d22                ; *oq0 = u^0x80
+    veor        d1, d25, d22                ; *op1 = u^0x80
+    veor        d4, d26, d22                ; *oq1 = u^0x80
+
+    bx          lr
+
+    ENDP        ; |aom_mbloop_filter_neon|
+
+    END

third_party/aom/aom_dsp/arm/loopfilter_8_neon.c

@@ -0,0 +1,430 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+
+static INLINE void mbloop_filter_neon(uint8x8_t dblimit,   // mblimit
+                                      uint8x8_t dlimit,    // limit
+                                      uint8x8_t dthresh,   // thresh
+                                      uint8x8_t d3u8,      // p2
+                                      uint8x8_t d4u8,      // p2
+                                      uint8x8_t d5u8,      // p1
+                                      uint8x8_t d6u8,      // p0
+                                      uint8x8_t d7u8,      // q0
+                                      uint8x8_t d16u8,     // q1
+                                      uint8x8_t d17u8,     // q2
+                                      uint8x8_t d18u8,     // q3
+                                      uint8x8_t *d0ru8,    // p1
+                                      uint8x8_t *d1ru8,    // p1
+                                      uint8x8_t *d2ru8,    // p0
+                                      uint8x8_t *d3ru8,    // q0
+                                      uint8x8_t *d4ru8,    // q1
+                                      uint8x8_t *d5ru8) {  // q1
+  uint32_t flat;
+  uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
+  uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+  int16x8_t q15s16;
+  uint16x8_t q10u16, q14u16;
+  int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
+
+  d19u8 = vabd_u8(d3u8, d4u8);
+  d20u8 = vabd_u8(d4u8, d5u8);
+  d21u8 = vabd_u8(d5u8, d6u8);
+  d22u8 = vabd_u8(d16u8, d7u8);
+  d23u8 = vabd_u8(d17u8, d16u8);
+  d24u8 = vabd_u8(d18u8, d17u8);
+
+  d19u8 = vmax_u8(d19u8, d20u8);
+  d20u8 = vmax_u8(d21u8, d22u8);
+
+  d25u8 = vabd_u8(d6u8, d4u8);
+
+  d23u8 = vmax_u8(d23u8, d24u8);
+
+  d26u8 = vabd_u8(d7u8, d17u8);
+
+  d19u8 = vmax_u8(d19u8, d20u8);
+
+  d24u8 = vabd_u8(d6u8, d7u8);
+  d27u8 = vabd_u8(d3u8, d6u8);
+  d28u8 = vabd_u8(d18u8, d7u8);
+
+  d19u8 = vmax_u8(d19u8, d23u8);
+
+  d23u8 = vabd_u8(d5u8, d16u8);
+  d24u8 = vqadd_u8(d24u8, d24u8);
+
+  d19u8 = vcge_u8(dlimit, d19u8);
+
+  d25u8 = vmax_u8(d25u8, d26u8);
+  d26u8 = vmax_u8(d27u8, d28u8);
+
+  d23u8 = vshr_n_u8(d23u8, 1);
+
+  d25u8 = vmax_u8(d25u8, d26u8);
+
+  d24u8 = vqadd_u8(d24u8, d23u8);
+
+  d20u8 = vmax_u8(d20u8, d25u8);
+
+  d23u8 = vdup_n_u8(1);
+  d24u8 = vcge_u8(dblimit, d24u8);
+
+  d21u8 = vcgt_u8(d21u8, dthresh);
+
+  d20u8 = vcge_u8(d23u8, d20u8);
+
+  d19u8 = vand_u8(d19u8, d24u8);
+
+  d23u8 = vcgt_u8(d22u8, dthresh);
+
+  d20u8 = vand_u8(d20u8, d19u8);
+
+  d22u8 = vdup_n_u8(0x80);
+
+  d23u8 = vorr_u8(d21u8, d23u8);
+
+  q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8), vreinterpret_u16_u8(d21u8));
+
+  d30u8 = vshrn_n_u16(q10u16, 4);
+  flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
+
+  if (flat == 0xffffffff) {  // Check for all 1's, power_branch_only
+    d27u8 = vdup_n_u8(3);
+    d21u8 = vdup_n_u8(2);
+    q14u16 = vaddl_u8(d6u8, d7u8);
+    q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+    q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
+    q14u16 = vaddw_u8(q14u16, d5u8);
+    *d0ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vaddw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+    *d1ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+    *d2ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d7u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+    *d3ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vsubw_u8(q14u16, d7u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+    *d4ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vsubw_u8(q14u16, d16u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+    *d5ru8 = vqrshrn_n_u16(q14u16, 3);
+  } else {
+    d21u8 = veor_u8(d7u8, d22u8);
+    d24u8 = veor_u8(d6u8, d22u8);
+    d25u8 = veor_u8(d5u8, d22u8);
+    d26u8 = veor_u8(d16u8, d22u8);
+
+    d27u8 = vdup_n_u8(3);
+
+    d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
+    d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
+
+    q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
+
+    d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+    q15s16 = vaddw_s8(q15s16, d29s8);
+
+    d29u8 = vdup_n_u8(4);
+
+    d28s8 = vqmovn_s16(q15s16);
+
+    d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+    d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
+    d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
+    d30s8 = vshr_n_s8(d30s8, 3);
+    d29s8 = vshr_n_s8(d29s8, 3);
+
+    d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
+    d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
+
+    d29s8 = vrshr_n_s8(d29s8, 1);
+    d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+    d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
+    d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
+
+    if (flat == 0) {  // filter_branch_only
+      *d0ru8 = d4u8;
+      *d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+      *d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+      *d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+      *d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+      *d5ru8 = d17u8;
+      return;
+    }
+
+    d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+    d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+    d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+    d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+
+    d23u8 = vdup_n_u8(2);
+    q14u16 = vaddl_u8(d6u8, d7u8);
+    q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+    q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
+
+    d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
+
+    q14u16 = vaddw_u8(q14u16, d5u8);
+
+    d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
+
+    d30u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vaddw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+
+    d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
+
+    d31u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+
+    *d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
+
+    d23u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d7u8);
+
+    *d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
+
+    q14u16 = vaddw_u8(q14u16, d18u8);
+
+    *d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
+
+    d22u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vsubw_u8(q14u16, d7u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+
+    d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
+
+    q14u16 = vaddw_u8(q14u16, d18u8);
+
+    d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
+
+    d6u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vsubw_u8(q14u16, d16u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+
+    d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
+
+    d7u8 = vqrshrn_n_u16(q14u16, 3);
+
+    *d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
+    *d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
+    *d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
+  }
+  return;
+}
+
+void aom_lpf_horizontal_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+  uint8_t *s, *psrc;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+  uint8x8_t d16u8, d17u8, d18u8;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  psrc = src - (pitch << 2);
+  for (i = 0; i < 1; i++) {
+    s = psrc + i * 8;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                       d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
+                       &d5u8);
+
+    s -= (pitch * 6);
+    vst1_u8(s, d0u8);
+    s += pitch;
+    vst1_u8(s, d1u8);
+    s += pitch;
+    vst1_u8(s, d2u8);
+    s += pitch;
+    vst1_u8(s, d3u8);
+    s += pitch;
+    vst1_u8(s, d4u8);
+    s += pitch;
+    vst1_u8(s, d5u8);
+  }
+  return;
+}
+
+void aom_lpf_vertical_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+  uint8_t *s;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+  uint8x8_t d16u8, d17u8, d18u8;
+  uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+  uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+  uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+  uint8x8x4_t d4Result;
+  uint8x8x2_t d2Result;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  for (i = 0; i < 1; i++) {
+    s = src + (i * (pitch << 3)) - 4;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8), vreinterpret_u32_u8(d7u8));
+    d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8), vreinterpret_u32_u8(d16u8));
+    d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8), vreinterpret_u32_u8(d17u8));
+    d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8), vreinterpret_u32_u8(d18u8));
+
+    d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                      vreinterpret_u16_u32(d2tmp2.val[0]));
+    d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                      vreinterpret_u16_u32(d2tmp3.val[0]));
+    d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                      vreinterpret_u16_u32(d2tmp2.val[1]));
+    d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                      vreinterpret_u16_u32(d2tmp3.val[1]));
+
+    d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                     vreinterpret_u8_u16(d2tmp5.val[0]));
+    d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                     vreinterpret_u8_u16(d2tmp5.val[1]));
+    d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                      vreinterpret_u8_u16(d2tmp7.val[0]));
+    d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                      vreinterpret_u8_u16(d2tmp7.val[1]));
+
+    d3u8 = d2tmp8.val[0];
+    d4u8 = d2tmp8.val[1];
+    d5u8 = d2tmp9.val[0];
+    d6u8 = d2tmp9.val[1];
+    d7u8 = d2tmp10.val[0];
+    d16u8 = d2tmp10.val[1];
+    d17u8 = d2tmp11.val[0];
+    d18u8 = d2tmp11.val[1];
+
+    mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                       d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
+                       &d5u8);
+
+    d4Result.val[0] = d0u8;
+    d4Result.val[1] = d1u8;
+    d4Result.val[2] = d2u8;
+    d4Result.val[3] = d3u8;
+
+    d2Result.val[0] = d4u8;
+    d2Result.val[1] = d5u8;
+
+    s = src - 3;
+    vst4_lane_u8(s, d4Result, 0);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 1);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 2);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 3);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 4);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 5);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 6);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 7);
+
+    s = src + 1;
+    vst2_lane_u8(s, d2Result, 0);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 1);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 2);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 3);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 4);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 5);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 6);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 7);
+  }
+  return;
+}

third_party/aom/aom_dsp/arm/loopfilter_mb_neon.asm

@@ -0,0 +1,638 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_edge_8_neon|
+    EXPORT  |aom_lpf_horizontal_edge_16_neon|
+    EXPORT  |aom_lpf_vertical_16_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; void mb_lpf_horizontal_edge(uint8_t *s, int p,
+;                             const uint8_t *blimit,
+;                             const uint8_t *limit,
+;                             const uint8_t *thresh,
+;                             int count)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+; r12   int count
+|mb_lpf_horizontal_edge| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+
+h_count
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, r1, lsl #3         ; move src pointer down by 8 lines
+
+    vld1.u8     {d0}, [r8@64], r1          ; p7
+    vld1.u8     {d1}, [r8@64], r1          ; p6
+    vld1.u8     {d2}, [r8@64], r1          ; p5
+    vld1.u8     {d3}, [r8@64], r1          ; p4
+    vld1.u8     {d4}, [r8@64], r1          ; p3
+    vld1.u8     {d5}, [r8@64], r1          ; p2
+    vld1.u8     {d6}, [r8@64], r1          ; p1
+    vld1.u8     {d7}, [r8@64], r1          ; p0
+    vld1.u8     {d8}, [r8@64], r1          ; q0
+    vld1.u8     {d9}, [r8@64], r1          ; q1
+    vld1.u8     {d10}, [r8@64], r1         ; q2
+    vld1.u8     {d11}, [r8@64], r1         ; q3
+    vld1.u8     {d12}, [r8@64], r1         ; q4
+    vld1.u8     {d13}, [r8@64], r1         ; q5
+    vld1.u8     {d14}, [r8@64], r1         ; q6
+    vld1.u8     {d15}, [r8@64], r1         ; q7
+
+    bl          aom_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         h_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, r1, lsl #1
+
+    vst1.u8     {d25}, [r8@64], r1         ; store op1
+    vst1.u8     {d24}, [r8@64], r1         ; store op0
+    vst1.u8     {d23}, [r8@64], r1         ; store oq0
+    vst1.u8     {d26}, [r8@64], r1         ; store oq1
+
+    b           h_next
+
+h_mbfilter
+    tst         r7, #2
+    beq         h_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, r1, lsl #1
+    sub         r8, r8, r1
+
+    vst1.u8     {d18}, [r8@64], r1         ; store op2
+    vst1.u8     {d19}, [r8@64], r1         ; store op1
+    vst1.u8     {d20}, [r8@64], r1         ; store op0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq0
+    vst1.u8     {d22}, [r8@64], r1         ; store oq1
+    vst1.u8     {d23}, [r8@64], r1         ; store oq2
+
+    b           h_next
+
+h_wide_mbfilter
+    sub         r8, r0, r1, lsl #3
+    add         r8, r8, r1
+
+    vst1.u8     {d16}, [r8@64], r1         ; store op6
+    vst1.u8     {d24}, [r8@64], r1         ; store op5
+    vst1.u8     {d25}, [r8@64], r1         ; store op4
+    vst1.u8     {d26}, [r8@64], r1         ; store op3
+    vst1.u8     {d27}, [r8@64], r1         ; store op2
+    vst1.u8     {d18}, [r8@64], r1         ; store op1
+    vst1.u8     {d19}, [r8@64], r1         ; store op0
+    vst1.u8     {d20}, [r8@64], r1         ; store oq0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq1
+    vst1.u8     {d22}, [r8@64], r1         ; store oq2
+    vst1.u8     {d23}, [r8@64], r1         ; store oq3
+    vst1.u8     {d1}, [r8@64], r1          ; store oq4
+    vst1.u8     {d2}, [r8@64], r1          ; store oq5
+    vst1.u8     {d3}, [r8@64], r1          ; store oq6
+
+h_next
+    add         r0, r0, #8
+    subs        r12, r12, #1
+    bne         h_count
+
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |mb_lpf_horizontal_edge|
+
+; void aom_lpf_horizontal_edge_8_neon(uint8_t *s, int pitch,
+;                                     const uint8_t *blimit,
+;                                     const uint8_t *limit,
+;                                     const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh
+|aom_lpf_horizontal_edge_8_neon| PROC
+    mov r12, #1
+    b mb_lpf_horizontal_edge
+    ENDP        ; |aom_lpf_horizontal_edge_8_neon|
+
+; void aom_lpf_horizontal_edge_16_neon(uint8_t *s, int pitch,
+;                                      const uint8_t *blimit,
+;                                      const uint8_t *limit,
+;                                      const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh
+|aom_lpf_horizontal_edge_16_neon| PROC
+    mov r12, #2
+    b mb_lpf_horizontal_edge
+    ENDP        ; |aom_lpf_horizontal_edge_16_neon|
+
+; void aom_lpf_vertical_16_neon(uint8_t *s, int p,
+;                               const uint8_t *blimit,
+;                               const uint8_t *limit,
+;                               const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_vertical_16_neon| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, #8
+
+    vld1.8      {d0}, [r8@64], r1
+    vld1.8      {d8}, [r0@64], r1
+    vld1.8      {d1}, [r8@64], r1
+    vld1.8      {d9}, [r0@64], r1
+    vld1.8      {d2}, [r8@64], r1
+    vld1.8      {d10}, [r0@64], r1
+    vld1.8      {d3}, [r8@64], r1
+    vld1.8      {d11}, [r0@64], r1
+    vld1.8      {d4}, [r8@64], r1
+    vld1.8      {d12}, [r0@64], r1
+    vld1.8      {d5}, [r8@64], r1
+    vld1.8      {d13}, [r0@64], r1
+    vld1.8      {d6}, [r8@64], r1
+    vld1.8      {d14}, [r0@64], r1
+    vld1.8      {d7}, [r8@64], r1
+    vld1.8      {d15}, [r0@64], r1
+
+    sub         r0, r0, r1, lsl #3
+
+    vtrn.32     q0, q2
+    vtrn.32     q1, q3
+    vtrn.32     q4, q6
+    vtrn.32     q5, q7
+
+    vtrn.16     q0, q1
+    vtrn.16     q2, q3
+    vtrn.16     q4, q5
+    vtrn.16     q6, q7
+
+    vtrn.8      d0, d1
+    vtrn.8      d2, d3
+    vtrn.8      d4, d5
+    vtrn.8      d6, d7
+
+    vtrn.8      d8, d9
+    vtrn.8      d10, d11
+    vtrn.8      d12, d13
+    vtrn.8      d14, d15
+
+    bl          aom_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         v_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, #2
+
+    vswp        d23, d25
+
+    vst4.8      {d23[0], d24[0], d25[0], d26[0]}, [r8], r1
+    vst4.8      {d23[1], d24[1], d25[1], d26[1]}, [r8], r1
+    vst4.8      {d23[2], d24[2], d25[2], d26[2]}, [r8], r1
+    vst4.8      {d23[3], d24[3], d25[3], d26[3]}, [r8], r1
+    vst4.8      {d23[4], d24[4], d25[4], d26[4]}, [r8], r1
+    vst4.8      {d23[5], d24[5], d25[5], d26[5]}, [r8], r1
+    vst4.8      {d23[6], d24[6], d25[6], d26[6]}, [r8], r1
+    vst4.8      {d23[7], d24[7], d25[7], d26[7]}, [r8], r1
+
+    b           v_end
+
+v_mbfilter
+    tst         r7, #2
+    beq         v_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, #3
+
+    vst3.8      {d18[0], d19[0], d20[0]}, [r8], r1
+    vst3.8      {d21[0], d22[0], d23[0]}, [r0], r1
+    vst3.8      {d18[1], d19[1], d20[1]}, [r8], r1
+    vst3.8      {d21[1], d22[1], d23[1]}, [r0], r1
+    vst3.8      {d18[2], d19[2], d20[2]}, [r8], r1
+    vst3.8      {d21[2], d22[2], d23[2]}, [r0], r1
+    vst3.8      {d18[3], d19[3], d20[3]}, [r8], r1
+    vst3.8      {d21[3], d22[3], d23[3]}, [r0], r1
+    vst3.8      {d18[4], d19[4], d20[4]}, [r8], r1
+    vst3.8      {d21[4], d22[4], d23[4]}, [r0], r1
+    vst3.8      {d18[5], d19[5], d20[5]}, [r8], r1
+    vst3.8      {d21[5], d22[5], d23[5]}, [r0], r1
+    vst3.8      {d18[6], d19[6], d20[6]}, [r8], r1
+    vst3.8      {d21[6], d22[6], d23[6]}, [r0], r1
+    vst3.8      {d18[7], d19[7], d20[7]}, [r8], r1
+    vst3.8      {d21[7], d22[7], d23[7]}, [r0], r1
+
+    b           v_end
+
+v_wide_mbfilter
+    sub         r8, r0, #8
+
+    vtrn.32     d0,  d26
+    vtrn.32     d16, d27
+    vtrn.32     d24, d18
+    vtrn.32     d25, d19
+
+    vtrn.16     d0,  d24
+    vtrn.16     d16, d25
+    vtrn.16     d26, d18
+    vtrn.16     d27, d19
+
+    vtrn.8      d0,  d16
+    vtrn.8      d24, d25
+    vtrn.8      d26, d27
+    vtrn.8      d18, d19
+
+    vtrn.32     d20, d1
+    vtrn.32     d21, d2
+    vtrn.32     d22, d3
+    vtrn.32     d23, d15
+
+    vtrn.16     d20, d22
+    vtrn.16     d21, d23
+    vtrn.16     d1,  d3
+    vtrn.16     d2,  d15
+
+    vtrn.8      d20, d21
+    vtrn.8      d22, d23
+    vtrn.8      d1,  d2
+    vtrn.8      d3,  d15
+
+    vst1.8      {d0}, [r8@64], r1
+    vst1.8      {d20}, [r0@64], r1
+    vst1.8      {d16}, [r8@64], r1
+    vst1.8      {d21}, [r0@64], r1
+    vst1.8      {d24}, [r8@64], r1
+    vst1.8      {d22}, [r0@64], r1
+    vst1.8      {d25}, [r8@64], r1
+    vst1.8      {d23}, [r0@64], r1
+    vst1.8      {d26}, [r8@64], r1
+    vst1.8      {d1}, [r0@64], r1
+    vst1.8      {d27}, [r8@64], r1
+    vst1.8      {d2}, [r0@64], r1
+    vst1.8      {d18}, [r8@64], r1
+    vst1.8      {d3}, [r0@64], r1
+    vst1.8      {d19}, [r8@64], r1
+    vst1.8      {d15}, [r0@64], r1
+
+v_end
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |aom_lpf_vertical_16_neon|
+
+; void aom_wide_mbfilter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store.
+;
+; r0-r3 PRESERVE
+; d16    blimit
+; d17    limit
+; d18    thresh
+; d0    p7
+; d1    p6
+; d2    p5
+; d3    p4
+; d4    p3
+; d5    p2
+; d6    p1
+; d7    p0
+; d8    q0
+; d9    q1
+; d10   q2
+; d11   q3
+; d12   q4
+; d13   q5
+; d14   q6
+; d15   q7
+|aom_wide_mbfilter_neon| PROC
+    mov         r7, #0
+
+    ; filter_mask
+    vabd.u8     d19, d4, d5                ; abs(p3 - p2)
+    vabd.u8     d20, d5, d6                ; abs(p2 - p1)
+    vabd.u8     d21, d6, d7                ; abs(p1 - p0)
+    vabd.u8     d22, d9, d8                ; abs(q1 - q0)
+    vabd.u8     d23, d10, d9               ; abs(q2 - q1)
+    vabd.u8     d24, d11, d10              ; abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; max(abs(p3 - p2), abs(p2 - p1))
+    vmax.u8     d20, d21, d22              ; max(abs(p1 - p0), abs(q1 - q0))
+    vmax.u8     d23, d23, d24              ; max(abs(q2 - q1), abs(q3 - q2))
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d7, d8                ; abs(p0 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d6, d9                ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d17, d19
+
+    ; flatmask4
+    vabd.u8     d25, d7, d5                ; abs(p0 - p2)
+    vabd.u8     d26, d8, d10               ; abs(q0 - q2)
+    vabd.u8     d27, d4, d7                ; abs(p3 - p0)
+    vabd.u8     d28, d11, d8               ; abs(q3 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; max(abs(p0 - p2), abs(q0 - q2))
+    vmax.u8     d26, d27, d28              ; max(abs(p3 - p0), abs(q3 - q0))
+    vmax.u8     d25, d25, d26
+    vmax.u8     d20, d20, d25
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmov.u8     d30, #1
+    vcge.u8     d24, d16, d24              ; (a > blimit * 2 + limit) * -1
+
+    vcge.u8     d20, d30, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    ; hevmask
+    vcgt.u8     d21, d21, d18              ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d18              ; (abs(q1 - q0) > thresh)*-1
+    vorr        d21, d21, d22              ; hev
+
+    vand        d16, d20, d19              ; flat && mask
+    vmov        r5, r6, d16
+
+    ; flatmask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7)
+    vabd.u8     d22, d3, d7                ; abs(p4 - p0)
+    vabd.u8     d23, d12, d8               ; abs(q4 - q0)
+    vabd.u8     d24, d7, d2                ; abs(p0 - p5)
+    vabd.u8     d25, d8, d13               ; abs(q0 - q5)
+    vabd.u8     d26, d1, d7                ; abs(p6 - p0)
+    vabd.u8     d27, d14, d8               ; abs(q6 - q0)
+    vabd.u8     d28, d0, d7                ; abs(p7 - p0)
+    vabd.u8     d29, d15, d8               ; abs(q7 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d22, d22, d23              ; max(abs(p4 - p0), abs(q4 - q0))
+    vmax.u8     d23, d24, d25              ; max(abs(p0 - p5), abs(q0 - q5))
+    vmax.u8     d24, d26, d27              ; max(abs(p6 - p0), abs(q6 - q0))
+    vmax.u8     d25, d28, d29              ; max(abs(p7 - p0), abs(q7 - q0))
+
+    vmax.u8     d26, d22, d23
+    vmax.u8     d27, d24, d25
+    vmax.u8     d23, d26, d27
+
+    vcge.u8     d18, d30, d23              ; flat2
+
+    vmov.u8     d22, #0x80
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #1                 ; Only do filter branch
+
+    vand        d17, d18, d16              ; flat2 && flat && mask
+    vmov        r5, r6, d17
+
+    ; mbfilter() function
+
+    ; filter() function
+    ; convert to signed
+    veor        d23, d8, d22               ; qs0
+    veor        d24, d7, d22               ; ps0
+    veor        d25, d6, d22               ; ps1
+    veor        d26, d9, d22               ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d23, d24              ; ( qs0 - ps0)
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+    vand        d29, d29, d21              ; filter &= hev
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d23, d23, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d21              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d23, d23, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    tst         r7, #1
+    bxne        lr
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #2                 ; Only do mbfilter branch
+
+    ; mbfilter flat && mask branch
+    ; TODO(fgalligan): Can I decrease the cycles shifting to consective d's
+    ; and using vibt on the q's?
+    vmov.u8     d29, #2
+    vaddl.u8    q15, d7, d8                ; op2 = p0 + q0
+    vmlal.u8    q15, d4, d27               ; op2 = p0 + q0 + p3 * 3
+    vmlal.u8    q15, d5, d29               ; op2 = p0 + q0 + p3 * 3 + p2 * 2
+    vaddl.u8    q10, d4, d5
+    vaddw.u8    q15, d6                    ; op2=p1 + p0 + q0 + p3 * 3 + p2 *2
+    vaddl.u8    q14, d6, d9
+    vqrshrn.u16 d18, q15, #3               ; r_op2
+
+    vsub.i16    q15, q10
+    vaddl.u8    q10, d4, d6
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d7, d10
+    vqrshrn.u16 d19, q15, #3               ; r_op1
+
+    vsub.i16    q15, q10
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d8, d11
+    vqrshrn.u16 d20, q15, #3               ; r_op0
+
+    vsubw.u8    q15, d4                    ; oq0 = op0 - p3
+    vsubw.u8    q15, d7                    ; oq0 -= p0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d9, d11
+    vqrshrn.u16 d21, q15, #3               ; r_oq0
+
+    vsubw.u8    q15, d5                    ; oq1 = oq0 - p2
+    vsubw.u8    q15, d8                    ; oq1 -= q0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d10, d11
+    vqrshrn.u16 d22, q15, #3               ; r_oq1
+
+    vsubw.u8    q15, d6                    ; oq2 = oq0 - p1
+    vsubw.u8    q15, d9                    ; oq2 -= q1
+    vadd.i16    q15, q14
+    vqrshrn.u16 d27, q15, #3               ; r_oq2
+
+    ; Filter does not set op2 or oq2, so use p2 and q2.
+    vbif        d18, d5, d16               ; t_op2 |= p2 & ~(flat & mask)
+    vbif        d19, d25, d16              ; t_op1 |= f_op1 & ~(flat & mask)
+    vbif        d20, d24, d16              ; t_op0 |= f_op0 & ~(flat & mask)
+    vbif        d21, d23, d16              ; t_oq0 |= f_oq0 & ~(flat & mask)
+    vbif        d22, d26, d16              ; t_oq1 |= f_oq1 & ~(flat & mask)
+
+    vbit        d23, d27, d16              ; t_oq2 |= r_oq2 & (flat & mask)
+    vbif        d23, d10, d16              ; t_oq2 |= q2 & ~(flat & mask)
+
+    tst         r7, #2
+    bxne        lr
+
+    ; wide_mbfilter flat2 && flat && mask branch
+    vmov.u8     d16, #7
+    vaddl.u8    q15, d7, d8                ; op6 = p0 + q0
+    vaddl.u8    q12, d2, d3
+    vaddl.u8    q13, d4, d5
+    vaddl.u8    q14, d1, d6
+    vmlal.u8    q15, d0, d16               ; op6 += p7 * 3
+    vadd.i16    q12, q13
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vadd.i16    q15, q12
+    vaddl.u8    q12, d0, d1
+    vaddw.u8    q15, d1
+    vaddl.u8    q13, d0, d2
+    vadd.i16    q14, q15, q14
+    vqrshrn.u16 d16, q15, #4               ; w_op6
+
+    vsub.i16    q15, q14, q12
+    vaddl.u8    q14, d3, d10
+    vqrshrn.u16 d24, q15, #4               ; w_op5
+
+    vsub.i16    q15, q13
+    vaddl.u8    q13, d0, d3
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vqrshrn.u16 d25, q15, #4               ; w_op4
+
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d0, d4
+    vsub.i16    q15, q13
+    vsub.i16    q14, q15, q14
+    vqrshrn.u16 d26, q15, #4               ; w_op3
+
+    vaddw.u8    q15, q14, d5               ; op2 += p2
+    vaddl.u8    q14, d0, d5
+    vaddw.u8    q15, d12                   ; op2 += q4
+    vbif        d26, d4, d17               ; op3 |= p3 & ~(f2 & f & m)
+    vqrshrn.u16 d27, q15, #4               ; w_op2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d6
+    vaddw.u8    q15, d6                    ; op1 += p1
+    vaddw.u8    q15, d13                   ; op1 += q5
+    vbif        d27, d18, d17              ; op2 |= t_op2 & ~(f2 & f & m)
+    vqrshrn.u16 d18, q15, #4               ; w_op1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d7
+    vaddw.u8    q15, d7                    ; op0 += p0
+    vaddw.u8    q15, d14                   ; op0 += q6
+    vbif        d18, d19, d17              ; op1 |= t_op1 & ~(f2 & f & m)
+    vqrshrn.u16 d19, q15, #4               ; w_op0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d1, d8
+    vaddw.u8    q15, d8                    ; oq0 += q0
+    vaddw.u8    q15, d15                   ; oq0 += q7
+    vbif        d19, d20, d17              ; op0 |= t_op0 & ~(f2 & f & m)
+    vqrshrn.u16 d20, q15, #4               ; w_oq0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vaddw.u8    q15, d9                    ; oq1 += q1
+    vaddl.u8    q4, d10, d15
+    vaddw.u8    q15, d15                   ; oq1 += q7
+    vbif        d20, d21, d17              ; oq0 |= t_oq0 & ~(f2 & f & m)
+    vqrshrn.u16 d21, q15, #4               ; w_oq1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d3, d10
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d11, d15
+    vbif        d21, d22, d17              ; oq1 |= t_oq1 & ~(f2 & f & m)
+    vqrshrn.u16 d22, q15, #4               ; w_oq2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d12, d15
+    vbif        d22, d23, d17              ; oq2 |= t_oq2 & ~(f2 & f & m)
+    vqrshrn.u16 d23, q15, #4               ; w_oq3
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d5, d12
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d13, d15
+    vbif        d16, d1, d17               ; op6 |= p6 & ~(f2 & f & m)
+    vqrshrn.u16 d1, q15, #4                ; w_oq4
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d6, d13
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d14, d15
+    vbif        d24, d2, d17               ; op5 |= p5 & ~(f2 & f & m)
+    vqrshrn.u16 d2, q15, #4                ; w_oq5
+
+    vsub.i16    q15, q14
+    vbif        d25, d3, d17               ; op4 |= p4 & ~(f2 & f & m)
+    vadd.i16    q15, q4
+    vbif        d23, d11, d17              ; oq3 |= q3 & ~(f2 & f & m)
+    vqrshrn.u16 d3, q15, #4                ; w_oq6
+    vbif        d1, d12, d17               ; oq4 |= q4 & ~(f2 & f & m)
+    vbif        d2, d13, d17               ; oq5 |= q5 & ~(f2 & f & m)
+    vbif        d3, d14, d17               ; oq6 |= q6 & ~(f2 & f & m)
+
+    bx          lr
+    ENDP        ; |aom_wide_mbfilter_neon|
+
+    END

third_party/aom/aom_dsp/arm/loopfilter_neon.c

@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+void aom_lpf_vertical_4_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  aom_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0);
+  aom_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+#if HAVE_NEON_ASM
+void aom_lpf_horizontal_8_dual_neon(
+    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
+    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
+    const uint8_t *limit1, const uint8_t *thresh1) {
+  aom_lpf_horizontal_8_neon(s, p, blimit0, limit0, thresh0);
+  aom_lpf_horizontal_8_neon(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_8_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  aom_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0);
+  aom_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_16_dual_neon(uint8_t *s, int p, const uint8_t *blimit,
+                                   const uint8_t *limit,
+                                   const uint8_t *thresh) {
+  aom_lpf_vertical_16_neon(s, p, blimit, limit, thresh);
+  aom_lpf_vertical_16_neon(s + 8 * p, p, blimit, limit, thresh);
+}
+#endif  // HAVE_NEON_ASM

third_party/aom/aom_dsp/arm/sad4d_neon.c

@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo =
+      vaddl_u16(vget_low_u16(vec_lo), vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi =
+      vaddl_u16(vget_low_u16(vec_hi), vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+// Calculate the absolute difference of 64 bytes from vec_src_00, vec_src_16,
+// vec_src_32, vec_src_48 and ref. Accumulate partial sums in vec_sum_ref_lo
+// and vec_sum_ref_hi.
+static void sad_neon_64(const uint8x16_t vec_src_00,
+                        const uint8x16_t vec_src_16,
+                        const uint8x16_t vec_src_32,
+                        const uint8x16_t vec_src_48, const uint8_t *ref,
+                        uint16x8_t *vec_sum_ref_lo,
+                        uint16x8_t *vec_sum_ref_hi) {
+  const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+  const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+  const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+  const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
+                             vget_low_u8(vec_ref_00));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
+                             vget_high_u8(vec_ref_00));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
+                             vget_low_u8(vec_ref_16));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
+                             vget_high_u8(vec_ref_16));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_32),
+                             vget_low_u8(vec_ref_32));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_32),
+                             vget_high_u8(vec_ref_32));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_48),
+                             vget_low_u8(vec_ref_48));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_48),
+                             vget_high_u8(vec_ref_48));
+}
+
+// Calculate the absolute difference of 32 bytes from vec_src_00, vec_src_16,
+// and ref. Accumulate partial sums in vec_sum_ref_lo and vec_sum_ref_hi.
+static void sad_neon_32(const uint8x16_t vec_src_00,
+                        const uint8x16_t vec_src_16, const uint8_t *ref,
+                        uint16x8_t *vec_sum_ref_lo,
+                        uint16x8_t *vec_sum_ref_hi) {
+  const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+  const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
+                             vget_low_u8(vec_ref_00));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
+                             vget_high_u8(vec_ref_00));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
+                             vget_low_u8(vec_ref_16));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
+                             vget_high_u8(vec_ref_16));
+}
+
+void aom_sad64x64x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref0,
+                &vec_sum_ref0_lo, &vec_sum_ref0_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref1,
+                &vec_sum_ref1_lo, &vec_sum_ref1_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref2,
+                &vec_sum_ref2_lo, &vec_sum_ref2_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref3,
+                &vec_sum_ref3_lo, &vec_sum_ref3_hi);
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
+
+void aom_sad32x32x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+
+    sad_neon_32(vec_src_00, vec_src_16, ref0, &vec_sum_ref0_lo,
+                &vec_sum_ref0_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref1, &vec_sum_ref1_lo,
+                &vec_sum_ref1_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref2, &vec_sum_ref2_lo,
+                &vec_sum_ref2_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref3, &vec_sum_ref3_lo,
+                &vec_sum_ref3_hi);
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
+
+void aom_sad16x16x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref0 = vld1q_u8(ref0);
+    const uint8x16_t vec_ref1 = vld1q_u8(ref1);
+    const uint8x16_t vec_ref2 = vld1q_u8(ref2);
+    const uint8x16_t vec_ref3 = vld1q_u8(ref3);
+
+    vec_sum_ref0_lo =
+        vabal_u8(vec_sum_ref0_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref0));
+    vec_sum_ref0_hi = vabal_u8(vec_sum_ref0_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref0));
+    vec_sum_ref1_lo =
+        vabal_u8(vec_sum_ref1_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref1));
+    vec_sum_ref1_hi = vabal_u8(vec_sum_ref1_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref1));
+    vec_sum_ref2_lo =
+        vabal_u8(vec_sum_ref2_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref2));
+    vec_sum_ref2_hi = vabal_u8(vec_sum_ref2_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref2));
+    vec_sum_ref3_lo =
+        vabal_u8(vec_sum_ref3_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref3));
+    vec_sum_ref3_hi = vabal_u8(vec_sum_ref3_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref3));
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}

third_party/aom/aom_dsp/arm/sad_media.asm

@@ -0,0 +1,98 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_sad16x16_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    const unsigned char *src_ptr
+; r1    int  src_stride
+; r2    const unsigned char *ref_ptr
+; r3    int  ref_stride
+|aom_sad16x16_media| PROC
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    mov     r4, #0              ; sad = 0;
+    mov     r5, #8              ; loop count
+
+loop
+    ; 1st row
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels (1A)
+    ldr     r8, [r2, #0x0]      ; load 4 ref pixels (1A)
+    ldr     r7, [r0, #0x4]      ; load 4 src pixels (1A)
+    ldr     r9, [r2, #0x4]      ; load 4 ref pixels (1A)
+    ldr     r10, [r0, #0x8]     ; load 4 src pixels (1B)
+    ldr     r11, [r0, #0xC]     ; load 4 src pixels (1B)
+
+    usada8  r4, r8, r6, r4      ; calculate sad for 4 pixels
+    usad8   r8, r7, r9          ; calculate sad for 4 pixels
+
+    ldr     r12, [r2, #0x8]     ; load 4 ref pixels (1B)
+    ldr     lr, [r2, #0xC]      ; load 4 ref pixels (1B)
+
+    add     r0, r0, r1          ; set src pointer to next row
+    add     r2, r2, r3          ; set dst pointer to next row
+
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    usada8  r4, r10, r12, r4    ; calculate sad for 4 pixels
+    usada8  r8, r11, lr, r8     ; calculate sad for 4 pixels
+
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels (2A)
+    ldr     r7, [r0, #0x4]      ; load 4 src pixels (2A)
+    add     r4, r4, r8          ; add partial sad values
+
+    ; 2nd row
+    ldr     r8, [r2, #0x0]      ; load 4 ref pixels (2A)
+    ldr     r9, [r2, #0x4]      ; load 4 ref pixels (2A)
+    ldr     r10, [r0, #0x8]     ; load 4 src pixels (2B)
+    ldr     r11, [r0, #0xC]     ; load 4 src pixels (2B)
+
+    usada8  r4, r6, r8, r4      ; calculate sad for 4 pixels
+    usad8   r8, r7, r9          ; calculate sad for 4 pixels
+
+    ldr     r12, [r2, #0x8]     ; load 4 ref pixels (2B)
+    ldr     lr, [r2, #0xC]      ; load 4 ref pixels (2B)
+
+    add     r0, r0, r1          ; set src pointer to next row
+    add     r2, r2, r3          ; set dst pointer to next row
+
+    usada8  r4, r10, r12, r4    ; calculate sad for 4 pixels
+    usada8  r8, r11, lr, r8     ; calculate sad for 4 pixels
+
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    subs    r5, r5, #1          ; decrement loop counter
+    add     r4, r4, r8          ; add partial sad values
+
+    bne     loop
+
+    mov     r0, r4              ; return sad
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+    END
+

third_party/aom/aom_dsp/arm/sad_neon.c

@@ -0,0 +1,224 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+unsigned int aom_sad8x16_neon(unsigned char *src_ptr, int src_stride,
+                              unsigned char *ref_ptr, int ref_stride) {
+  uint8x8_t d0, d8;
+  uint16x8_t q12;
+  uint32x4_t q1;
+  uint64x2_t q3;
+  uint32x2_t d5;
+  int i;
+
+  d0 = vld1_u8(src_ptr);
+  src_ptr += src_stride;
+  d8 = vld1_u8(ref_ptr);
+  ref_ptr += ref_stride;
+  q12 = vabdl_u8(d0, d8);
+
+  for (i = 0; i < 15; i++) {
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabal_u8(q12, d0, d8);
+  }
+
+  q1 = vpaddlq_u16(q12);
+  q3 = vpaddlq_u32(q1);
+  d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                vreinterpret_u32_u64(vget_high_u64(q3)));
+
+  return vget_lane_u32(d5, 0);
+}
+
+unsigned int aom_sad4x4_neon(unsigned char *src_ptr, int src_stride,
+                             unsigned char *ref_ptr, int ref_stride) {
+  uint8x8_t d0, d8;
+  uint16x8_t q12;
+  uint32x2_t d1;
+  uint64x1_t d3;
+  int i;
+
+  d0 = vld1_u8(src_ptr);
+  src_ptr += src_stride;
+  d8 = vld1_u8(ref_ptr);
+  ref_ptr += ref_stride;
+  q12 = vabdl_u8(d0, d8);
+
+  for (i = 0; i < 3; i++) {
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabal_u8(q12, d0, d8);
+  }
+
+  d1 = vpaddl_u16(vget_low_u16(q12));
+  d3 = vpaddl_u32(d1);
+
+  return vget_lane_u32(vreinterpret_u32_u64(d3), 0);
+}
+
+unsigned int aom_sad16x8_neon(unsigned char *src_ptr, int src_stride,
+                              unsigned char *ref_ptr, int ref_stride) {
+  uint8x16_t q0, q4;
+  uint16x8_t q12, q13;
+  uint32x4_t q1;
+  uint64x2_t q3;
+  uint32x2_t d5;
+  int i;
+
+  q0 = vld1q_u8(src_ptr);
+  src_ptr += src_stride;
+  q4 = vld1q_u8(ref_ptr);
+  ref_ptr += ref_stride;
+  q12 = vabdl_u8(vget_low_u8(q0), vget_low_u8(q4));
+  q13 = vabdl_u8(vget_high_u8(q0), vget_high_u8(q4));
+
+  for (i = 0; i < 7; i++) {
+    q0 = vld1q_u8(src_ptr);
+    src_ptr += src_stride;
+    q4 = vld1q_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabal_u8(q12, vget_low_u8(q0), vget_low_u8(q4));
+    q13 = vabal_u8(q13, vget_high_u8(q0), vget_high_u8(q4));
+  }
+
+  q12 = vaddq_u16(q12, q13);
+  q1 = vpaddlq_u16(q12);
+  q3 = vpaddlq_u32(q1);
+  d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                vreinterpret_u32_u64(vget_high_u64(q3)));
+
+  return vget_lane_u32(d5, 0);
+}
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo =
+      vaddl_u16(vget_low_u16(vec_lo), vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi =
+      vaddl_u16(vget_low_u16(vec_hi), vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_16x8) {
+  const uint32x4_t a = vpaddlq_u16(vec_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+unsigned int aom_sad64x64_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+    const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_32),
+                            vget_low_u8(vec_ref_32));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_32),
+                            vget_high_u8(vec_ref_32));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_48),
+                            vget_low_u8(vec_ref_48));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_48),
+                            vget_high_u8(vec_ref_48));
+  }
+  return horizontal_long_add_16x8(vec_accum_lo, vec_accum_hi);
+}
+
+unsigned int aom_sad32x32_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int aom_sad16x16_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref = vld1q_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo =
+        vabal_u8(vec_accum_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref));
+    vec_accum_hi =
+        vabal_u8(vec_accum_hi, vget_high_u8(vec_src), vget_high_u8(vec_ref));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int aom_sad8x8_neon(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum = vdupq_n_u16(0);
+
+  for (i = 0; i < 8; ++i) {
+    const uint8x8_t vec_src = vld1_u8(src);
+    const uint8x8_t vec_ref = vld1_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum = vabal_u8(vec_accum, vec_src, vec_ref);
+  }
+  return horizontal_add_16x8(vec_accum);
+}

third_party/aom/aom_dsp/arm/save_reg_neon.asm

@@ -0,0 +1,39 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_push_neon|
+    EXPORT  |aom_pop_neon|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|aom_push_neon| PROC
+    vst1.i64            {d8, d9, d10, d11}, [r0]!
+    vst1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+|aom_pop_neon| PROC
+    vld1.i64            {d8, d9, d10, d11}, [r0]!
+    vld1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+    END
+

third_party/aom/aom_dsp/arm/subpel_variance_media.c

@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_MEDIA
+static const int16_t bilinear_filters_media[8][2] = { { 128, 0 }, { 112, 16 },
+                                                      { 96, 32 }, { 80, 48 },
+                                                      { 64, 64 }, { 48, 80 },
+                                                      { 32, 96 }, { 16, 112 } };
+
+extern void aom_filter_block2d_bil_first_pass_media(
+    const uint8_t *src_ptr, uint16_t *dst_ptr, uint32_t src_pitch,
+    uint32_t height, uint32_t width, const int16_t *filter);
+
+extern void aom_filter_block2d_bil_second_pass_media(
+    const uint16_t *src_ptr, uint8_t *dst_ptr, int32_t src_pitch,
+    uint32_t height, uint32_t width, const int16_t *filter);
+
+unsigned int aom_sub_pixel_variance8x8_media(
+    const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset,
+    const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse) {
+  uint16_t first_pass[10 * 8];
+  uint8_t second_pass[8 * 8];
+  const int16_t *HFilter, *VFilter;
+
+  HFilter = bilinear_filters_media[xoffset];
+  VFilter = bilinear_filters_media[yoffset];
+
+  aom_filter_block2d_bil_first_pass_media(src_ptr, first_pass,
+                                          src_pixels_per_line, 9, 8, HFilter);
+  aom_filter_block2d_bil_second_pass_media(first_pass, second_pass, 8, 8, 8,
+                                           VFilter);
+
+  return aom_variance8x8_media(second_pass, 8, dst_ptr, dst_pixels_per_line,
+                               sse);
+}
+
+unsigned int aom_sub_pixel_variance16x16_media(
+    const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset,
+    const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse) {
+  uint16_t first_pass[36 * 16];
+  uint8_t second_pass[20 * 16];
+  const int16_t *HFilter, *VFilter;
+  unsigned int var;
+
+  if (xoffset == 4 && yoffset == 0) {
+    var = aom_variance_halfpixvar16x16_h_media(
+        src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse);
+  } else if (xoffset == 0 && yoffset == 4) {
+    var = aom_variance_halfpixvar16x16_v_media(
+        src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse);
+  } else if (xoffset == 4 && yoffset == 4) {
+    var = aom_variance_halfpixvar16x16_hv_media(
+        src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse);
+  } else {
+    HFilter = bilinear_filters_media[xoffset];
+    VFilter = bilinear_filters_media[yoffset];
+
+    aom_filter_block2d_bil_first_pass_media(
+        src_ptr, first_pass, src_pixels_per_line, 17, 16, HFilter);
+    aom_filter_block2d_bil_second_pass_media(first_pass, second_pass, 16, 16,
+                                             16, VFilter);
+
+    var = aom_variance16x16_media(second_pass, 16, dst_ptr, dst_pixels_per_line,
+                                  sse);
+  }
+  return var;
+}
+#endif  // HAVE_MEDIA

third_party/aom/aom_dsp/arm/subpel_variance_neon.c

@@ -0,0 +1,134 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/variance.h"
+
+static const uint8_t bilinear_filters[8][2] = {
+  { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 },
+  { 64, 64 }, { 48, 80 },  { 32, 96 }, { 16, 112 },
+};
+
+static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
+                                      uint8_t *output_ptr,
+                                      unsigned int src_pixels_per_line,
+                                      int pixel_step,
+                                      unsigned int output_height,
+                                      unsigned int output_width,
+                                      const uint8_t *filter) {
+  const uint8x8_t f0 = vmov_n_u8(filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(filter[1]);
+  unsigned int i;
+  for (i = 0; i < output_height; ++i) {
+    const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
+    const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
+    const uint16x8_t a = vmull_u8(src_0, f0);
+    const uint16x8_t b = vmlal_u8(a, src_1, f1);
+    const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
+    vst1_u8(&output_ptr[0], out);
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+static void var_filter_block2d_bil_w16(const uint8_t *src_ptr,
+                                       uint8_t *output_ptr,
+                                       unsigned int src_pixels_per_line,
+                                       int pixel_step,
+                                       unsigned int output_height,
+                                       unsigned int output_width,
+                                       const uint8_t *filter) {
+  const uint8x8_t f0 = vmov_n_u8(filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(filter[1]);
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; j += 16) {
+      const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]);
+      const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]);
+      const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0);
+      const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1);
+      const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS);
+      const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0);
+      const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1);
+      const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS);
+      vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi));
+    }
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+unsigned int aom_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride,
+                                            int xoffset, int yoffset,
+                                            const uint8_t *dst, int dst_stride,
+                                            unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[8 * 8]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[9 * 8]);
+
+  var_filter_block2d_bil_w8(src, fdata3, src_stride, 1, 9, 8,
+                            bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w8(fdata3, temp2, 8, 8, 8, 8,
+                            bilinear_filters[yoffset]);
+  return aom_variance8x8_neon(temp2, 8, dst, dst_stride, sse);
+}
+
+unsigned int aom_sub_pixel_variance16x16_neon(const uint8_t *src,
+                                              int src_stride, int xoffset,
+                                              int yoffset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[16 * 16]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[17 * 16]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 17, 16,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 16, 16, 16, 16,
+                             bilinear_filters[yoffset]);
+  return aom_variance16x16_neon(temp2, 16, dst, dst_stride, sse);
+}
+
+unsigned int aom_sub_pixel_variance32x32_neon(const uint8_t *src,
+                                              int src_stride, int xoffset,
+                                              int yoffset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[32 * 32]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[33 * 32]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 33, 32,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 32, 32, 32, 32,
+                             bilinear_filters[yoffset]);
+  return aom_variance32x32_neon(temp2, 32, dst, dst_stride, sse);
+}
+
+unsigned int aom_sub_pixel_variance64x64_neon(const uint8_t *src,
+                                              int src_stride, int xoffset,
+                                              int yoffset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[64 * 64]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[65 * 64]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 65, 64,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 64, 64, 64, 64,
+                             bilinear_filters[yoffset]);
+  return aom_variance64x64_neon(temp2, 64, dst, dst_stride, sse);
+}

third_party/aom/aom_dsp/arm/subtract_neon.c

@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+void aom_subtract_block_neon(int rows, int cols, int16_t *diff,
+                             ptrdiff_t diff_stride, const uint8_t *src,
+                             ptrdiff_t src_stride, const uint8_t *pred,
+                             ptrdiff_t pred_stride) {
+  int r, c;
+
+  if (cols > 16) {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; c += 32) {
+        const uint8x16_t v_src_00 = vld1q_u8(&src[c + 0]);
+        const uint8x16_t v_src_16 = vld1q_u8(&src[c + 16]);
+        const uint8x16_t v_pred_00 = vld1q_u8(&pred[c + 0]);
+        const uint8x16_t v_pred_16 = vld1q_u8(&pred[c + 16]);
+        const uint16x8_t v_diff_lo_00 =
+            vsubl_u8(vget_low_u8(v_src_00), vget_low_u8(v_pred_00));
+        const uint16x8_t v_diff_hi_00 =
+            vsubl_u8(vget_high_u8(v_src_00), vget_high_u8(v_pred_00));
+        const uint16x8_t v_diff_lo_16 =
+            vsubl_u8(vget_low_u8(v_src_16), vget_low_u8(v_pred_16));
+        const uint16x8_t v_diff_hi_16 =
+            vsubl_u8(vget_high_u8(v_src_16), vget_high_u8(v_pred_16));
+        vst1q_s16(&diff[c + 0], vreinterpretq_s16_u16(v_diff_lo_00));
+        vst1q_s16(&diff[c + 8], vreinterpretq_s16_u16(v_diff_hi_00));
+        vst1q_s16(&diff[c + 16], vreinterpretq_s16_u16(v_diff_lo_16));
+        vst1q_s16(&diff[c + 24], vreinterpretq_s16_u16(v_diff_hi_16));
+      }
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  } else if (cols > 8) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x16_t v_src = vld1q_u8(&src[0]);
+      const uint8x16_t v_pred = vld1q_u8(&pred[0]);
+      const uint16x8_t v_diff_lo =
+          vsubl_u8(vget_low_u8(v_src), vget_low_u8(v_pred));
+      const uint16x8_t v_diff_hi =
+          vsubl_u8(vget_high_u8(v_src), vget_high_u8(v_pred));
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff_lo));
+      vst1q_s16(&diff[8], vreinterpretq_s16_u16(v_diff_hi));
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  } else if (cols > 4) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x8_t v_src = vld1_u8(&src[0]);
+      const uint8x8_t v_pred = vld1_u8(&pred[0]);
+      const uint16x8_t v_diff = vsubl_u8(v_src, v_pred);
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff));
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  } else {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; ++c) diff[c] = src[c] - pred[c];
+
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  }
+}

third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm

@@ -0,0 +1,185 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance_halfpixvar16x16_h_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance_halfpixvar16x16_h_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r6, [r0, #1]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r6, [r0, #5]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r6, [r0, #9]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r6, [r0, #13]       ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    subs    r12, r12, #1
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
+

third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm

@@ -0,0 +1,225 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance_halfpixvar16x16_hv_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance_halfpixvar16x16_hv_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    add     r9, r0, r1          ; pointer to pixels on the next row
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load source pixels a, row N
+    ldr     r6, [r0, #1]        ; load source pixels b, row N
+    ldr     r5, [r9, #0]        ; load source pixels c, row N+1
+    ldr     r7, [r9, #1]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load source pixels a, row N
+    ldr     r6, [r0, #5]        ; load source pixels b, row N
+    ldr     r5, [r9, #4]        ; load source pixels c, row N+1
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    ldr     r7, [r9, #5]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load source pixels a, row N
+    ldr     r6, [r0, #9]        ; load source pixels b, row N
+    ldr     r5, [r9, #8]        ; load source pixels c, row N+1
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    ldr     r7, [r9, #9]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load source pixels a, row N
+    ldr     r6, [r0, #13]       ; load source pixels b, row N
+    ldr     r5, [r9, #12]       ; load source pixels c, row N+1
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+    ldr     r7, [r9, #13]       ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    subs    r12, r12, #1
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END

third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm

@@ -0,0 +1,187 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance_halfpixvar16x16_v_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance_halfpixvar16x16_v_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    add     r9, r0, r1          ; set src pointer to next row
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r6, [r9, #0]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r6, [r9, #4]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r6, [r9, #8]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r6, [r9, #12]       ; load 4 src pixels from next row
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+
+    subs    r12, r12, #1
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
+

third_party/aom/aom_dsp/arm/variance_media.asm

@@ -0,0 +1,361 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance16x16_media|
+    EXPORT  |aom_variance8x8_media|
+    EXPORT  |aom_mse16x16_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance16x16_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+
+loop16x16
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+
+    subs    r12, r12, #1
+
+    bne     loop16x16
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance8x8_media| PROC
+
+    push    {r4-r10, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r12, #8             ; set loop counter to 8 (=block height)
+    mov     r4, #0              ; initialize sum = 0
+    mov     r5, #0              ; initialize sse = 0
+
+loop8x8
+    ; 1st 4 pixels
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels
+    ldr     r7, [r2, #0x0]      ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r8, r6, r7          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r10, r8, lr         ; select bytes with positive difference
+    usub8   r9, r7, r6          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r6, r10, lr         ; calculate sum of positive differences
+    usad8   r7, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r10         ; differences of all 4 pixels
+    ; calculate total sum
+    add    r4, r4, r6           ; add positive differences to sum
+    sub    r4, r4, r7           ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r7, r8              ; byte (two pixels) to halfwords
+    uxtb16  r10, r8, ror #8     ; another two pixels to halfwords
+    smlad   r5, r7, r7, r5      ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r6, [r0, #0x4]      ; load 4 src pixels
+    ldr     r7, [r2, #0x4]      ; load 4 ref pixels
+    smlad   r5, r10, r10, r5    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r6, r7          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r10, r8, lr         ; select bytes with positive difference
+    usub8   r9, r7, r6          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r6, r10, lr         ; calculate sum of positive differences
+    usad8   r7, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r10         ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r4, r4, r6          ; add positive differences to sum
+    sub     r4, r4, r7          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r7, r8              ; byte (two pixels) to halfwords
+    uxtb16  r10, r8, ror #8     ; another two pixels to halfwords
+    smlad   r5, r7, r7, r5      ; dual signed multiply, add and accumulate (1)
+    subs    r12, r12, #1        ; next row
+    smlad   r5, r10, r10, r5    ; dual signed multiply, add and accumulate (2)
+
+    bne     loop8x8
+
+    ; return stuff
+    ldr     r8, [sp, #32]       ; get address of sse
+    mul     r1, r4, r4          ; sum * sum
+    str     r5, [r8]            ; store sse
+    sub     r0, r5, r1, ASR #6  ; return (sse - ((sum * sum) >> 6))
+
+    pop     {r4-r10, pc}
+
+    ENDP
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+;
+;note: Based on aom_variance16x16_media. In this function, sum is never used.
+;      So, we can remove this part of calculation.
+
+|aom_mse16x16_media| PROC
+
+    push    {r4-r9, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     r4, #0              ; initialize sse = 0
+
+loopmse
+    ; 1st 4 pixels
+    ldr     r5, [r0, #0x0]      ; load 4 src pixels
+    ldr     r6, [r2, #0x0]      ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r8, r5, r6          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    ldr     r5, [r0, #0x4]      ; load 4 src pixels
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r6, [r2, #0x4]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+    ldr     r5, [r0, #0x8]      ; load 4 src pixels
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r6, [r2, #0x8]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    ldr     r5, [r0, #0xc]      ; load 4 src pixels
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r6, [r2, #0xc]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    subs    r12, r12, #1        ; next row
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    bne     loopmse
+
+    ; return stuff
+    ldr     r1, [sp, #28]       ; get address of sse
+    mov     r0, r4              ; return sse
+    str     r4, [r1]            ; store sse
+
+    pop     {r4-r9, pc}
+
+    ENDP
+
+    END

third_party/aom/aom_dsp/arm/variance_neon.c

@@ -0,0 +1,400 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {
+  const int32x4_t a = vpaddlq_s16(v_16x8);
+  const int64x2_t b = vpaddlq_s32(a);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+static INLINE int horizontal_add_s32x4(const int32x4_t v_32x4) {
+  const int64x2_t b = vpaddlq_s32(v_32x4);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+// w * h must be less than 2048 or local variable v_sum may overflow.
+static void variance_neon_w8(const uint8_t *a, int a_stride, const uint8_t *b,
+                             int b_stride, int w, int h, uint32_t *sse,
+                             int *sum) {
+  int i, j;
+  int16x8_t v_sum = vdupq_n_s16(0);
+  int32x4_t v_sse_lo = vdupq_n_s32(0);
+  int32x4_t v_sse_hi = vdupq_n_s32(0);
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; j += 8) {
+      const uint8x8_t v_a = vld1_u8(&a[j]);
+      const uint8x8_t v_b = vld1_u8(&b[j]);
+      const uint16x8_t v_diff = vsubl_u8(v_a, v_b);
+      const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff);
+      v_sum = vaddq_s16(v_sum, sv_diff);
+      v_sse_lo =
+          vmlal_s16(v_sse_lo, vget_low_s16(sv_diff), vget_low_s16(sv_diff));
+      v_sse_hi =
+          vmlal_s16(v_sse_hi, vget_high_s16(sv_diff), vget_high_s16(sv_diff));
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+
+  *sum = horizontal_add_s16x8(v_sum);
+  *sse = (unsigned int)horizontal_add_s32x4(vaddq_s32(v_sse_lo, v_sse_hi));
+}
+
+void aom_get8x8var_neon(const uint8_t *a, int a_stride, const uint8_t *b,
+                        int b_stride, unsigned int *sse, int *sum) {
+  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, sum);
+}
+
+void aom_get16x16var_neon(const uint8_t *a, int a_stride, const uint8_t *b,
+                          int b_stride, unsigned int *sse, int *sum) {
+  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, sum);
+}
+
+unsigned int aom_variance8x8_neon(const uint8_t *a, int a_stride,
+                                  const uint8_t *b, int b_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, &sum);
+  return *sse - ((sum * sum) >> 6);
+}
+
+unsigned int aom_variance16x16_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, &sum);
+  return *sse - (((unsigned int)((int64_t)sum * sum)) >> 8);
+}
+
+unsigned int aom_variance32x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, sse, &sum);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 10);
+}
+
+unsigned int aom_variance32x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, &sse1, &sum1);
+  variance_neon_w8(a + (32 * a_stride), a_stride, b + (32 * b_stride), b_stride,
+                   32, 32, &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 11);
+}
+
+unsigned int aom_variance64x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
+  variance_neon_w8(a + (16 * a_stride), a_stride, b + (16 * b_stride), b_stride,
+                   64, 16, &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 11);
+}
+
+unsigned int aom_variance64x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+
+  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
+  variance_neon_w8(a + (16 * a_stride), a_stride, b + (16 * b_stride), b_stride,
+                   64, 16, &sse2, &sum2);
+  sse1 += sse2;
+  sum1 += sum2;
+
+  variance_neon_w8(a + (16 * 2 * a_stride), a_stride, b + (16 * 2 * b_stride),
+                   b_stride, 64, 16, &sse2, &sum2);
+  sse1 += sse2;
+  sum1 += sum2;
+
+  variance_neon_w8(a + (16 * 3 * a_stride), a_stride, b + (16 * 3 * b_stride),
+                   b_stride, 64, 16, &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 12);
+}
+
+unsigned int aom_variance16x8_neon(const unsigned char *src_ptr,
+                                   int source_stride,
+                                   const unsigned char *ref_ptr,
+                                   int recon_stride, unsigned int *sse) {
+  int i;
+  int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+  uint32x2_t d0u32, d10u32;
+  int64x1_t d0s64, d1s64;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8;
+  uint16x8_t q11u16, q12u16, q13u16, q14u16;
+  int32x4_t q8s32, q9s32, q10s32;
+  int64x2_t q0s64, q1s64, q5s64;
+
+  q8s32 = vdupq_n_s32(0);
+  q9s32 = vdupq_n_s32(0);
+  q10s32 = vdupq_n_s32(0);
+
+  for (i = 0; i < 4; i++) {
+    q0u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    q1u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    __builtin_prefetch(src_ptr);
+
+    q2u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    q3u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    __builtin_prefetch(ref_ptr);
+
+    q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+    q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+    q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+    q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+    q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+    q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+    d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+    d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+    q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+    q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+    d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+    d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
+    q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
+    q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
+
+    d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+    d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
+    q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+    q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+  }
+
+  q10s32 = vaddq_s32(q10s32, q9s32);
+  q0s64 = vpaddlq_s32(q8s32);
+  q1s64 = vpaddlq_s32(q10s32);
+
+  d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+  d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64));
+  vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+  d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+  d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+  return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int aom_variance8x16_neon(const unsigned char *src_ptr,
+                                   int source_stride,
+                                   const unsigned char *ref_ptr,
+                                   int recon_stride, unsigned int *sse) {
+  int i;
+  uint8x8_t d0u8, d2u8, d4u8, d6u8;
+  int16x4_t d22s16, d23s16, d24s16, d25s16;
+  uint32x2_t d0u32, d10u32;
+  int64x1_t d0s64, d1s64;
+  uint16x8_t q11u16, q12u16;
+  int32x4_t q8s32, q9s32, q10s32;
+  int64x2_t q0s64, q1s64, q5s64;
+
+  q8s32 = vdupq_n_s32(0);
+  q9s32 = vdupq_n_s32(0);
+  q10s32 = vdupq_n_s32(0);
+
+  for (i = 0; i < 8; i++) {
+    d0u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    d2u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    __builtin_prefetch(src_ptr);
+
+    d4u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    d6u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    __builtin_prefetch(ref_ptr);
+
+    q11u16 = vsubl_u8(d0u8, d4u8);
+    q12u16 = vsubl_u8(d2u8, d6u8);
+
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+    q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+    q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+    d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+    d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+    q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+    q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+  }
+
+  q10s32 = vaddq_s32(q10s32, q9s32);
+  q0s64 = vpaddlq_s32(q8s32);
+  q1s64 = vpaddlq_s32(q10s32);
+
+  d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+  d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64));
+  vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+  d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+  d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+  return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int aom_mse16x16_neon(const unsigned char *src_ptr, int source_stride,
+                               const unsigned char *ref_ptr, int recon_stride,
+                               unsigned int *sse) {
+  int i;
+  int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+  int64x1_t d0s64;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8;
+  int32x4_t q7s32, q8s32, q9s32, q10s32;
+  uint16x8_t q11u16, q12u16, q13u16, q14u16;
+  int64x2_t q1s64;
+
+  q7s32 = vdupq_n_s32(0);
+  q8s32 = vdupq_n_s32(0);
+  q9s32 = vdupq_n_s32(0);
+  q10s32 = vdupq_n_s32(0);
+
+  for (i = 0; i < 8; i++) {  // mse16x16_neon_loop
+    q0u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    q1u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    q2u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    q3u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+
+    q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+    q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+    q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+    q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    q7s32 = vmlal_s16(q7s32, d22s16, d22s16);
+    q8s32 = vmlal_s16(q8s32, d23s16, d23s16);
+
+    d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+    d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+    q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+    q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+    d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+    d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+    q7s32 = vmlal_s16(q7s32, d26s16, d26s16);
+    q8s32 = vmlal_s16(q8s32, d27s16, d27s16);
+
+    d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+    d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+    q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+    q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+  }
+
+  q7s32 = vaddq_s32(q7s32, q8s32);
+  q9s32 = vaddq_s32(q9s32, q10s32);
+  q10s32 = vaddq_s32(q7s32, q9s32);
+
+  q1s64 = vpaddlq_s32(q10s32);
+  d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0);
+  return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
+}
+
+unsigned int aom_get4x4sse_cs_neon(const unsigned char *src_ptr,
+                                   int source_stride,
+                                   const unsigned char *ref_ptr,
+                                   int recon_stride) {
+  int16x4_t d22s16, d24s16, d26s16, d28s16;
+  int64x1_t d0s64;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+  int32x4_t q7s32, q8s32, q9s32, q10s32;
+  uint16x8_t q11u16, q12u16, q13u16, q14u16;
+  int64x2_t q1s64;
+
+  d0u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d4u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+  d1u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d5u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+  d2u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d6u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+  d3u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d7u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+
+  q11u16 = vsubl_u8(d0u8, d4u8);
+  q12u16 = vsubl_u8(d1u8, d5u8);
+  q13u16 = vsubl_u8(d2u8, d6u8);
+  q14u16 = vsubl_u8(d3u8, d7u8);
+
+  d22s16 = vget_low_s16(vreinterpretq_s16_u16(q11u16));
+  d24s16 = vget_low_s16(vreinterpretq_s16_u16(q12u16));
+  d26s16 = vget_low_s16(vreinterpretq_s16_u16(q13u16));
+  d28s16 = vget_low_s16(vreinterpretq_s16_u16(q14u16));
+
+  q7s32 = vmull_s16(d22s16, d22s16);
+  q8s32 = vmull_s16(d24s16, d24s16);
+  q9s32 = vmull_s16(d26s16, d26s16);
+  q10s32 = vmull_s16(d28s16, d28s16);
+
+  q7s32 = vaddq_s32(q7s32, q8s32);
+  q9s32 = vaddq_s32(q9s32, q10s32);
+  q9s32 = vaddq_s32(q7s32, q9s32);
+
+  q1s64 = vpaddlq_s32(q9s32);
+  d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
+}

third_party/aom/aom_dsp/avg.c

@@ -0,0 +1,232 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+
+unsigned int aom_avg_8x8_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 8; ++i, src += stride)
+    for (j = 0; j < 8; sum += src[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 6);
+}
+
+unsigned int aom_avg_4x4_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 4; ++i, src += stride)
+    for (j = 0; j < 4; sum += src[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 4);
+}
+
+// src_diff: first pass, 9 bit, dynamic range [-255, 255]
+//           second pass, 12 bit, dynamic range [-2040, 2040]
+static void hadamard_col8(const int16_t *src_diff, int src_stride,
+                          int16_t *coeff) {
+  int16_t b0 = src_diff[0 * src_stride] + src_diff[1 * src_stride];
+  int16_t b1 = src_diff[0 * src_stride] - src_diff[1 * src_stride];
+  int16_t b2 = src_diff[2 * src_stride] + src_diff[3 * src_stride];
+  int16_t b3 = src_diff[2 * src_stride] - src_diff[3 * src_stride];
+  int16_t b4 = src_diff[4 * src_stride] + src_diff[5 * src_stride];
+  int16_t b5 = src_diff[4 * src_stride] - src_diff[5 * src_stride];
+  int16_t b6 = src_diff[6 * src_stride] + src_diff[7 * src_stride];
+  int16_t b7 = src_diff[6 * src_stride] - src_diff[7 * src_stride];
+
+  int16_t c0 = b0 + b2;
+  int16_t c1 = b1 + b3;
+  int16_t c2 = b0 - b2;
+  int16_t c3 = b1 - b3;
+  int16_t c4 = b4 + b6;
+  int16_t c5 = b5 + b7;
+  int16_t c6 = b4 - b6;
+  int16_t c7 = b5 - b7;
+
+  coeff[0] = c0 + c4;
+  coeff[7] = c1 + c5;
+  coeff[3] = c2 + c6;
+  coeff[4] = c3 + c7;
+  coeff[2] = c0 - c4;
+  coeff[6] = c1 - c5;
+  coeff[1] = c2 - c6;
+  coeff[5] = c3 - c7;
+}
+
+// The order of the output coeff of the hadamard is not important. For
+// optimization purposes the final transpose may be skipped.
+void aom_hadamard_8x8_c(const int16_t *src_diff, int src_stride,
+                        int16_t *coeff) {
+  int idx;
+  int16_t buffer[64];
+  int16_t *tmp_buf = &buffer[0];
+  for (idx = 0; idx < 8; ++idx) {
+    hadamard_col8(src_diff, src_stride, tmp_buf);  // src_diff: 9 bit
+                                                   // dynamic range [-255, 255]
+    tmp_buf += 8;
+    ++src_diff;
+  }
+
+  tmp_buf = &buffer[0];
+  for (idx = 0; idx < 8; ++idx) {
+    hadamard_col8(tmp_buf, 8, coeff);  // tmp_buf: 12 bit
+                                       // dynamic range [-2040, 2040]
+    coeff += 8;                        // coeff: 15 bit
+                                       // dynamic range [-16320, 16320]
+    ++tmp_buf;
+  }
+}
+
+// In place 16x16 2D Hadamard transform
+void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride,
+                          int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    // src_diff: 9 bit, dynamic range [-255, 255]
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
+    aom_hadamard_8x8_c(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  // coeff: 15 bit, dynamic range [-16320, 16320]
+  for (idx = 0; idx < 64; ++idx) {
+    int16_t a0 = coeff[0];
+    int16_t a1 = coeff[64];
+    int16_t a2 = coeff[128];
+    int16_t a3 = coeff[192];
+
+    int16_t b0 = (a0 + a1) >> 1;  // (a0 + a1): 16 bit, [-32640, 32640]
+    int16_t b1 = (a0 - a1) >> 1;  // b0-b3: 15 bit, dynamic range
+    int16_t b2 = (a2 + a3) >> 1;  // [-16320, 16320]
+    int16_t b3 = (a2 - a3) >> 1;
+
+    coeff[0] = b0 + b2;  // 16 bit, [-32640, 32640]
+    coeff[64] = b1 + b3;
+    coeff[128] = b0 - b2;
+    coeff[192] = b1 - b3;
+
+    ++coeff;
+  }
+}
+
+// coeff: 16 bits, dynamic range [-32640, 32640].
+// length: value range {16, 64, 256, 1024}.
+int aom_satd_c(const int16_t *coeff, int length) {
+  int i;
+  int satd = 0;
+  for (i = 0; i < length; ++i) satd += abs(coeff[i]);
+
+  // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
+  return satd;
+}
+
+// Integer projection onto row vectors.
+// height: value range {16, 32, 64}.
+void aom_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref, int ref_stride,
+                       int height) {
+  int idx;
+  const int norm_factor = height >> 1;
+  for (idx = 0; idx < 16; ++idx) {
+    int i;
+    hbuf[idx] = 0;
+    // hbuf[idx]: 14 bit, dynamic range [0, 16320].
+    for (i = 0; i < height; ++i) hbuf[idx] += ref[i * ref_stride];
+    // hbuf[idx]: 9 bit, dynamic range [0, 510].
+    hbuf[idx] /= norm_factor;
+    ++ref;
+  }
+}
+
+// width: value range {16, 32, 64}.
+int16_t aom_int_pro_col_c(const uint8_t *ref, int width) {
+  int idx;
+  int16_t sum = 0;
+  // sum: 14 bit, dynamic range [0, 16320]
+  for (idx = 0; idx < width; ++idx) sum += ref[idx];
+  return sum;
+}
+
+// ref: [0 - 510]
+// src: [0 - 510]
+// bwl: {2, 3, 4}
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl) {
+  int i;
+  int width = 4 << bwl;
+  int sse = 0, mean = 0, var;
+
+  for (i = 0; i < width; ++i) {
+    int diff = ref[i] - src[i];  // diff: dynamic range [-510, 510], 10 bits.
+    mean += diff;                // mean: dynamic range 16 bits.
+    sse += diff * diff;          // sse:  dynamic range 26 bits.
+  }
+
+  // (mean * mean): dynamic range 31 bits.
+  var = sse - ((mean * mean) >> (bwl + 2));
+  return var;
+}
+
+void aom_minmax_8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref,
+                      int ref_stride, int *min, int *max) {
+  int i, j;
+  *min = 255;
+  *max = 0;
+  for (i = 0; i < 8; ++i, src += src_stride, ref += ref_stride) {
+    for (j = 0; j < 8; ++j) {
+      int diff = abs(src[j] - ref[j]);
+      *min = diff < *min ? diff : *min;
+      *max = diff > *max ? diff : *max;
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  const uint16_t *s = CONVERT_TO_SHORTPTR(src);
+  for (i = 0; i < 8; ++i, s += stride)
+    for (j = 0; j < 8; sum += s[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 6);
+}
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  const uint16_t *s = CONVERT_TO_SHORTPTR(src);
+  for (i = 0; i < 4; ++i, s += stride)
+    for (j = 0; j < 4; sum += s[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 4);
+}
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s8, int p, const uint8_t *d8,
+                             int dp, int *min, int *max) {
+  int i, j;
+  const uint16_t *s = CONVERT_TO_SHORTPTR(s8);
+  const uint16_t *d = CONVERT_TO_SHORTPTR(d8);
+  *min = 255;
+  *max = 0;
+  for (i = 0; i < 8; ++i, s += p, d += dp) {
+    for (j = 0; j < 8; ++j) {
+      int diff = abs(s[j] - d[j]);
+      *min = diff < *min ? diff : *min;
+      *max = diff > *max ? diff : *max;
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/aom_dsp/binary_codes_reader.c

@@ -0,0 +1,117 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/bitreader.h"
+
+#include "av1/common/common.h"
+
+// Inverse recenters a non-negative literal v around a reference r
+static uint16_t inv_recenter_nonneg(uint16_t r, uint16_t v) {
+  if (v > (r << 1))
+    return v;
+  else if ((v & 1) == 0)
+    return (v >> 1) + r;
+  else
+    return r - ((v + 1) >> 1);
+}
+
+// Inverse recenters a non-negative literal v in [0, n-1] around a
+// reference r also in [0, n-1]
+static uint16_t inv_recenter_finite_nonneg(uint16_t n, uint16_t r, uint16_t v) {
+  if ((r << 1) <= n) {
+    return inv_recenter_nonneg(r, v);
+  } else {
+    return n - 1 - inv_recenter_nonneg(n - 1 - r, v);
+  }
+}
+
+int16_t aom_read_primitive_symmetric(aom_reader *r, unsigned int mag_bits) {
+  if (aom_read_bit(r, NULL)) {
+    int s = aom_read_bit(r, NULL);
+    int16_t x = aom_read_literal(r, mag_bits, NULL) + 1;
+    return (s > 0 ? -x : x);
+  } else {
+    return 0;
+  }
+}
+
+uint16_t aom_read_primitive_quniform(aom_reader *r, uint16_t n) {
+  if (n <= 1) return 0;
+  const int l = get_msb(n - 1) + 1;
+  const int m = (1 << l) - n;
+  const int v = aom_read_literal(r, l - 1, NULL);
+  return v < m ? v : (v << 1) - m + aom_read_bit(r, NULL);
+}
+
+uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p,
+                                       uint16_t ref) {
+  if (n <= 1) return 0;
+  assert(p > 0 && p <= n);
+  assert(ref < n);
+  int lolimit = ref - p / 2;
+  const int hilimit = lolimit + p - 1;
+  if (lolimit < 0) {
+    lolimit = 0;
+  } else if (hilimit >= n) {
+    lolimit = n - p;
+  }
+  int v;
+  if (aom_read_bit(r, NULL)) {
+    v = aom_read_primitive_quniform(r, p) + lolimit;
+  } else {
+    v = aom_read_primitive_quniform(r, n - p);
+    if (v >= lolimit) v += p;
+  }
+  return v;
+}
+
+// Decode finite subexponential code that for a symbol v in [0, n-1] with
+// parameter k
+uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k) {
+  int i = 0;
+  int mk = 0;
+  uint16_t v;
+  while (1) {
+    int b = (i ? k + i - 1 : k);
+    int a = (1 << b);
+    if (n <= mk + 3 * a) {
+      v = aom_read_primitive_quniform(r, n - mk) + mk;
+      break;
+    } else {
+      if (aom_read_bit(r, NULL)) {
+        i = i + 1;
+        mk += a;
+      } else {
+        v = aom_read_literal(r, b, NULL) + mk;
+        break;
+      }
+    }
+  }
+  return v;
+}
+
+// Decode finite subexponential code that for a symbol v in [0, n-1] with
+// parameter k
+// based on a reference ref also in [0, n-1].
+uint16_t aom_read_primitive_refsubexpfin(aom_reader *r, uint16_t n, uint16_t k,
+                                         uint16_t ref) {
+  return inv_recenter_finite_nonneg(n, ref,
+                                    aom_read_primitive_subexpfin(r, n, k));
+}
+
+// Decode finite subexponential code that for a symbol v in [-(n-1), n-1] with
+// parameter k based on a reference ref also in [-(n-1), n-1].
+int16_t aom_read_signed_primitive_refsubexpfin(aom_reader *r, uint16_t n,
+                                               uint16_t k, int16_t ref) {
+  ref += n - 1;
+  const uint16_t scaled_n = (n << 1) - 1;
+  return aom_read_primitive_refsubexpfin(r, scaled_n, k, ref) - n + 1;
+}

third_party/aom/aom_dsp/binary_codes_reader.h

@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BINARY_CODES_READER_H_
+#define AOM_DSP_BINARY_CODES_READER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/bitreader.h"
+
+int16_t aom_read_primitive_symmetric(aom_reader *r, unsigned int mag_bits);
+
+uint16_t aom_read_primitive_quniform(aom_reader *r, uint16_t n);
+uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p,
+                                       uint16_t ref);
+uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k);
+uint16_t aom_read_primitive_refsubexpfin(aom_reader *r, uint16_t n, uint16_t k,
+                                         uint16_t ref);
+int16_t aom_read_signed_primitive_refsubexpfin(aom_reader *r, uint16_t n,
+                                               uint16_t k, int16_t ref);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BINARY_CODES_READER_H_

third_party/aom/aom_dsp/binary_codes_writer.c

@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/bitwriter.h"
+
+#include "av1/common/common.h"
+
+// Recenters a non-negative literal v around a reference r
+static uint16_t recenter_nonneg(uint16_t r, uint16_t v) {
+  if (v > (r << 1))
+    return v;
+  else if (v >= r)
+    return ((v - r) << 1);
+  else
+    return ((r - v) << 1) - 1;
+}
+
+// Recenters a non-negative literal v in [0, n-1] around a
+// reference r also in [0, n-1]
+static uint16_t recenter_finite_nonneg(uint16_t n, uint16_t r, uint16_t v) {
+  if ((r << 1) <= n) {
+    return recenter_nonneg(r, v);
+  } else {
+    return recenter_nonneg(n - 1 - r, n - 1 - v);
+  }
+}
+
+// Codes a symbol v in [-2^mag_bits, 2^mag_bits].
+// mag_bits is number of bits for magnitude. The alphabet is of size
+// 2 * 2^mag_bits + 1, symmetric around 0, where one bit is used to
+// indicate 0 or non-zero, mag_bits bits are used to indicate magnitide
+// and 1 more bit for the sign if non-zero.
+void aom_write_primitive_symmetric(aom_writer *w, int16_t v,
+                                   unsigned int abs_bits) {
+  if (v == 0) {
+    aom_write_bit(w, 0);
+  } else {
+    const int x = abs(v);
+    const int s = v < 0;
+    aom_write_bit(w, 1);
+    aom_write_bit(w, s);
+    aom_write_literal(w, x - 1, abs_bits);
+  }
+}
+
+int aom_count_primitive_symmetric(int16_t v, unsigned int abs_bits) {
+  return (v == 0 ? 1 : abs_bits + 2);
+}
+
+// Encodes a value v in [0, n-1] quasi-uniformly
+void aom_write_primitive_quniform(aom_writer *w, uint16_t n, uint16_t v) {
+  if (n <= 1) return;
+  const int l = get_msb(n - 1) + 1;
+  const int m = (1 << l) - n;
+  if (v < m) {
+    aom_write_literal(w, v, l - 1);
+  } else {
+    aom_write_literal(w, m + ((v - m) >> 1), l - 1);
+    aom_write_bit(w, (v - m) & 1);
+  }
+}
+
+int aom_count_primitive_quniform(uint16_t n, uint16_t v) {
+  if (n <= 1) return 0;
+  const int l = get_msb(n - 1) + 1;
+  const int m = (1 << l) - n;
+  return v < m ? l - 1 : l;
+}
+
+// Encodes a value v in [0, n-1] based on a reference ref also in [0, n-1]
+// The closest p values of v from ref are coded using a p-ary quasi-unoform
+// short code while the remaining n-p values are coded with a longer code.
+void aom_write_primitive_refbilevel(aom_writer *w, uint16_t n, uint16_t p,
+                                    uint16_t ref, uint16_t v) {
+  if (n <= 1) return;
+  assert(p > 0 && p <= n);
+  assert(ref < n);
+  int lolimit = ref - p / 2;
+  int hilimit = lolimit + p - 1;
+  if (lolimit < 0) {
+    lolimit = 0;
+    hilimit = p - 1;
+  } else if (hilimit >= n) {
+    hilimit = n - 1;
+    lolimit = n - p;
+  }
+  if (v >= lolimit && v <= hilimit) {
+    aom_write_bit(w, 1);
+    v = v - lolimit;
+    aom_write_primitive_quniform(w, p, v);
+  } else {
+    aom_write_bit(w, 0);
+    if (v > hilimit) v -= p;
+    aom_write_primitive_quniform(w, n - p, v);
+  }
+}
+
+int aom_count_primitive_refbilevel(uint16_t n, uint16_t p, uint16_t ref,
+                                   uint16_t v) {
+  if (n <= 1) return 0;
+  assert(p > 0 && p <= n);
+  assert(ref < n);
+  int lolimit = ref - p / 2;
+  int hilimit = lolimit + p - 1;
+  if (lolimit < 0) {
+    lolimit = 0;
+    hilimit = p - 1;
+  } else if (hilimit >= n) {
+    hilimit = n - 1;
+    lolimit = n - p;
+  }
+  int count = 0;
+  if (v >= lolimit && v <= hilimit) {
+    count++;
+    v = v - lolimit;
+    count += aom_count_primitive_quniform(p, v);
+  } else {
+    count++;
+    if (v > hilimit) v -= p;
+    count += aom_count_primitive_quniform(n - p, v);
+  }
+  return count;
+}
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+void aom_write_primitive_subexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                   uint16_t v) {
+  int i = 0;
+  int mk = 0;
+  while (1) {
+    int b = (i ? k + i - 1 : k);
+    int a = (1 << b);
+    if (n <= mk + 3 * a) {
+      aom_write_primitive_quniform(w, n - mk, v - mk);
+      break;
+    } else {
+      int t = (v >= mk + a);
+      aom_write_bit(w, t);
+      if (t) {
+        i = i + 1;
+        mk += a;
+      } else {
+        aom_write_literal(w, v - mk, b);
+        break;
+      }
+    }
+  }
+}
+
+int aom_count_primitive_subexpfin(uint16_t n, uint16_t k, uint16_t v) {
+  int count = 0;
+  int i = 0;
+  int mk = 0;
+  while (1) {
+    int b = (i ? k + i - 1 : k);
+    int a = (1 << b);
+    if (n <= mk + 3 * a) {
+      count += aom_count_primitive_quniform(n - mk, v - mk);
+      break;
+    } else {
+      int t = (v >= mk + a);
+      count++;
+      if (t) {
+        i = i + 1;
+        mk += a;
+      } else {
+        count += b;
+        break;
+      }
+    }
+  }
+  return count;
+}
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+// based on a reference ref also in [0, n-1].
+// Recenters symbol around r first and then uses a finite subexponential code.
+void aom_write_primitive_refsubexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                      int16_t ref, int16_t v) {
+  aom_write_primitive_subexpfin(w, n, k, recenter_finite_nonneg(n, ref, v));
+}
+
+void aom_write_signed_primitive_refsubexpfin(aom_writer *w, uint16_t n,
+                                             uint16_t k, uint16_t ref,
+                                             uint16_t v) {
+  ref += n - 1;
+  v += n - 1;
+  const uint16_t scaled_n = (n << 1) - 1;
+  aom_write_primitive_refsubexpfin(w, scaled_n, k, ref, v);
+}
+
+int aom_count_primitive_refsubexpfin(uint16_t n, uint16_t k, uint16_t ref,
+                                     uint16_t v) {
+  return aom_count_primitive_subexpfin(n, k, recenter_finite_nonneg(n, ref, v));
+}
+
+int aom_count_signed_primitive_refsubexpfin(uint16_t n, uint16_t k, int16_t ref,
+                                            int16_t v) {
+  ref += n - 1;
+  v += n - 1;
+  const uint16_t scaled_n = (n << 1) - 1;
+  return aom_count_primitive_refsubexpfin(scaled_n, k, ref, v);
+}

third_party/aom/aom_dsp/binary_codes_writer.h

@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BINARY_CODES_WRITER_H_
+#define AOM_DSP_BINARY_CODES_WRITER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/bitwriter.h"
+
+// Codes a symbol v in [-2^mag_bits, 2^mag_bits]
+// mag_bits is number of bits for magnitude. The alphabet is of size
+// 2 * 2^mag_bits + 1, symmetric around 0, where one bit is used to
+// indicate 0 or non-zero, mag_bits bits are used to indicate magnitide
+// and 1 more bit for the sign if non-zero.
+void aom_write_primitive_symmetric(aom_writer *w, int16_t v,
+                                   unsigned int mag_bits);
+
+// Encodes a value v in [0, n-1] quasi-uniformly
+void aom_write_primitive_quniform(aom_writer *w, uint16_t n, uint16_t v);
+
+// Encodes a value v in [0, n-1] based on a reference ref also in [0, n-1]
+// The closest p values of v from ref are coded using a p-ary quasi-unoform
+// short code while the remaining n-p values are coded with a longer code.
+void aom_write_primitive_refbilevel(aom_writer *w, uint16_t n, uint16_t p,
+                                    uint16_t ref, uint16_t v);
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+void aom_write_primitive_subexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                   uint16_t v);
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+// based on a reference ref also in [0, n-1].
+void aom_write_primitive_refsubexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                      uint16_t ref, uint16_t v);
+
+// Finite subexponential code that codes a symbol v in [-(n-1), n-1] with
+// parameter k based on a reference ref also in [-(n-1), n-1].
+void aom_write_signed_primitive_refsubexpfin(aom_writer *w, uint16_t n,
+                                             uint16_t k, int16_t ref,
+                                             int16_t v);
+
+// Functions that counts bits for the above primitives
+int aom_count_primitive_symmetric(int16_t v, unsigned int mag_bits);
+int aom_count_primitive_quniform(uint16_t n, uint16_t v);
+int aom_count_primitive_refbilevel(uint16_t n, uint16_t p, uint16_t ref,
+                                   uint16_t v);
+int aom_count_primitive_subexpfin(uint16_t n, uint16_t k, uint16_t v);
+int aom_count_primitive_refsubexpfin(uint16_t n, uint16_t k, uint16_t ref,
+                                     uint16_t v);
+int aom_count_signed_primitive_refsubexpfin(uint16_t n, uint16_t k, int16_t ref,
+                                            int16_t v);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BINARY_CODES_WRITER_H_

third_party/aom/aom_dsp/bitreader.h

@@ -0,0 +1,276 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITREADER_H_
+#define AOM_DSP_BITREADER_H_
+
+#include <assert.h>
+#include <limits.h>
+
+#include "./aom_config.h"
+#if CONFIG_EC_ADAPT && !CONFIG_EC_MULTISYMBOL
+#error "CONFIG_EC_ADAPT is enabled without enabling CONFIG_EC_MULTISYMBOL."
+#endif
+
+#include "aom/aomdx.h"
+#include "aom/aom_integer.h"
+#if CONFIG_ANS
+#include "aom_dsp/ansreader.h"
+#elif CONFIG_DAALA_EC
+#include "aom_dsp/daalaboolreader.h"
+#else
+#include "aom_dsp/dkboolreader.h"
+#endif
+#include "aom_dsp/prob.h"
+#include "av1/common/odintrin.h"
+
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#define ACCT_STR_NAME acct_str
+#define ACCT_STR_PARAM , const char *ACCT_STR_NAME
+#define ACCT_STR_ARG(s) , s
+#else
+#define ACCT_STR_PARAM
+#define ACCT_STR_ARG(s)
+#endif
+
+#define aom_read(r, prob, ACCT_STR_NAME) \
+  aom_read_(r, prob ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_bit(r, ACCT_STR_NAME) \
+  aom_read_bit_(r ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_tree(r, tree, probs, ACCT_STR_NAME) \
+  aom_read_tree_(r, tree, probs ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_literal(r, bits, ACCT_STR_NAME) \
+  aom_read_literal_(r, bits ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_cdf(r, cdf, nsymbs, ACCT_STR_NAME) \
+  aom_read_cdf_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_symbol(r, cdf, nsymbs, ACCT_STR_NAME) \
+  aom_read_symbol_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME))
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_ANS
+typedef struct AnsDecoder aom_reader;
+#elif CONFIG_DAALA_EC
+typedef struct daala_reader aom_reader;
+#else
+typedef struct aom_dk_reader aom_reader;
+#endif
+
+static INLINE int aom_reader_init(aom_reader *r, const uint8_t *buffer,
+                                  size_t size, aom_decrypt_cb decrypt_cb,
+                                  void *decrypt_state) {
+#if CONFIG_ANS
+  (void)decrypt_cb;
+  (void)decrypt_state;
+  if (size > INT_MAX) return 1;
+  return ans_read_init(r, buffer, (int)size);
+#elif CONFIG_DAALA_EC
+  (void)decrypt_cb;
+  (void)decrypt_state;
+  return aom_daala_reader_init(r, buffer, (int)size);
+#else
+  return aom_dk_reader_init(r, buffer, size, decrypt_cb, decrypt_state);
+#endif
+}
+
+static INLINE const uint8_t *aom_reader_find_end(aom_reader *r) {
+#if CONFIG_ANS
+  (void)r;
+  assert(0 && "Use the raw buffer size with ANS");
+  return NULL;
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_find_end(r);
+#else
+  return aom_dk_reader_find_end(r);
+#endif
+}
+
+static INLINE int aom_reader_has_error(aom_reader *r) {
+#if CONFIG_ANS
+  return ans_reader_has_error(r);
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_has_error(r);
+#else
+  return aom_dk_reader_has_error(r);
+#endif
+}
+
+// Returns the position in the bit reader in bits.
+static INLINE uint32_t aom_reader_tell(const aom_reader *r) {
+#if CONFIG_ANS
+  (void)r;
+  assert(0 && "aom_reader_tell() is unimplemented for ANS");
+  return 0;
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_tell(r);
+#else
+  return aom_dk_reader_tell(r);
+#endif
+}
+
+// Returns the position in the bit reader in 1/8th bits.
+static INLINE uint32_t aom_reader_tell_frac(const aom_reader *r) {
+#if CONFIG_ANS
+  (void)r;
+  assert(0 && "aom_reader_tell_frac() is unimplemented for ANS");
+  return 0;
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_tell_frac(r);
+#else
+  return aom_dk_reader_tell_frac(r);
+#endif
+}
+
+#if CONFIG_ACCOUNTING
+static INLINE void aom_process_accounting(const aom_reader *r ACCT_STR_PARAM) {
+  if (r->accounting != NULL) {
+    uint32_t tell_frac;
+    tell_frac = aom_reader_tell_frac(r);
+    aom_accounting_record(r->accounting, ACCT_STR_NAME,
+                          tell_frac - r->accounting->last_tell_frac);
+    r->accounting->last_tell_frac = tell_frac;
+  }
+}
+
+static INLINE void aom_update_symb_counts(const aom_reader *r, int is_binary) {
+  if (r->accounting != NULL) {
+    r->accounting->syms.num_multi_syms += !is_binary;
+    r->accounting->syms.num_binary_syms += !!is_binary;
+  }
+}
+#endif
+
+static INLINE int aom_read_(aom_reader *r, int prob ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_ANS
+  ret = rabs_read(r, prob);
+#elif CONFIG_DAALA_EC
+  ret = aom_daala_read(r, prob);
+#else
+  ret = aom_dk_read(r, prob);
+#endif
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+  aom_update_symb_counts(r, 1);
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_bit_(aom_reader *r ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_ANS
+  ret = rabs_read_bit(r);  // Non trivial optimization at half probability
+#elif CONFIG_DAALA_EC && CONFIG_RAWBITS
+  // Note this uses raw bits and is not the same as aom_daala_read(r, 128);
+  // Calls to this function are omitted from raw symbol accounting.
+  ret = aom_daala_read_bit(r);
+#else
+  ret = aom_read(r, 128, NULL);  // aom_prob_half
+#endif
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_literal_(aom_reader *r, int bits ACCT_STR_PARAM) {
+  int literal = 0, bit;
+
+  for (bit = bits - 1; bit >= 0; bit--) literal |= aom_read_bit(r, NULL) << bit;
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+#endif
+  return literal;
+}
+
+static INLINE int aom_read_tree_as_bits(aom_reader *r,
+                                        const aom_tree_index *tree,
+                                        const aom_prob *probs) {
+  aom_tree_index i = 0;
+
+  while ((i = tree[i + aom_read(r, probs[i >> 1], NULL)]) > 0) continue;
+  return -i;
+}
+
+#if CONFIG_EC_MULTISYMBOL
+static INLINE int aom_read_cdf_(aom_reader *r, const aom_cdf_prob *cdf,
+                                int nsymbs ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_ANS
+  (void)nsymbs;
+  ret = rans_read(r, cdf);
+#elif CONFIG_DAALA_EC
+  ret = daala_read_symbol(r, cdf, nsymbs);
+#else
+#error \
+    "CONFIG_EC_MULTISYMBOL is selected without a valid backing entropy " \
+  "coder. Enable daala_ec or ans for a valid configuration."
+#endif
+
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+  aom_update_symb_counts(r, (nsymbs == 2));
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_symbol_(aom_reader *r, aom_cdf_prob *cdf,
+                                   int nsymbs ACCT_STR_PARAM) {
+  int ret;
+  ret = aom_read_cdf(r, cdf, nsymbs, ACCT_STR_NAME);
+#if CONFIG_EC_ADAPT
+  update_cdf(cdf, ret, nsymbs);
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_tree_as_cdf(aom_reader *r,
+                                       const aom_tree_index *tree,
+                                       const aom_prob *probs) {
+  aom_tree_index i = 0;
+  do {
+    aom_cdf_prob cdf[16];
+    aom_tree_index index[16];
+    int path[16];
+    int dist[16];
+    int nsymbs;
+    int symb;
+    nsymbs = tree_to_cdf(tree, probs, i, cdf, index, path, dist);
+    symb = aom_read_cdf(r, cdf, nsymbs, NULL);
+    OD_ASSERT(symb >= 0 && symb < nsymbs);
+    i = index[symb];
+  } while (i > 0);
+  return -i;
+}
+#endif  // CONFIG_EC_MULTISYMBOL
+
+static INLINE int aom_read_tree_(aom_reader *r, const aom_tree_index *tree,
+                                 const aom_prob *probs ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_EC_MULTISYMBOL
+  ret = aom_read_tree_as_cdf(r, tree, probs);
+#else
+  ret = aom_read_tree_as_bits(r, tree, probs);
+#endif
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+#endif
+  return ret;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITREADER_H_

third_party/aom/aom_dsp/bitreader_buffer.c

@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "./aom_config.h"
+#include "./bitreader_buffer.h"
+
+size_t aom_rb_bytes_read(struct aom_read_bit_buffer *rb) {
+  return (rb->bit_offset + 7) >> 3;
+}
+
+int aom_rb_read_bit(struct aom_read_bit_buffer *rb) {
+  const uint32_t off = rb->bit_offset;
+  const uint32_t p = off >> 3;
+  const int q = 7 - (int)(off & 0x7);
+  if (rb->bit_buffer + p < rb->bit_buffer_end) {
+    const int bit = (rb->bit_buffer[p] >> q) & 1;
+    rb->bit_offset = off + 1;
+    return bit;
+  } else {
+    rb->error_handler(rb->error_handler_data);
+    return 0;
+  }
+}
+
+int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits) {
+  int value = 0, bit;
+  for (bit = bits - 1; bit >= 0; bit--) value |= aom_rb_read_bit(rb) << bit;
+  return value;
+}
+
+int aom_rb_read_signed_literal(struct aom_read_bit_buffer *rb, int bits) {
+  const int value = aom_rb_read_literal(rb, bits);
+  return aom_rb_read_bit(rb) ? -value : value;
+}
+
+int aom_rb_read_inv_signed_literal(struct aom_read_bit_buffer *rb, int bits) {
+  const int nbits = sizeof(unsigned) * 8 - bits - 1;
+  const unsigned value = (unsigned)aom_rb_read_literal(rb, bits + 1) << nbits;
+  return ((int)value) >> nbits;
+}

third_party/aom/aom_dsp/bitreader_buffer.h

@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITREADER_BUFFER_H_
+#define AOM_DSP_BITREADER_BUFFER_H_
+
+#include <limits.h>
+
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*aom_rb_error_handler)(void *data);
+
+struct aom_read_bit_buffer {
+  const uint8_t *bit_buffer;
+  const uint8_t *bit_buffer_end;
+  uint32_t bit_offset;
+
+  void *error_handler_data;
+  aom_rb_error_handler error_handler;
+};
+
+size_t aom_rb_bytes_read(struct aom_read_bit_buffer *rb);
+
+int aom_rb_read_bit(struct aom_read_bit_buffer *rb);
+
+int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits);
+
+int aom_rb_read_signed_literal(struct aom_read_bit_buffer *rb, int bits);
+
+int aom_rb_read_inv_signed_literal(struct aom_read_bit_buffer *rb, int bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITREADER_BUFFER_H_

third_party/aom/aom_dsp/bitwriter.h

@@ -0,0 +1,255 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITWRITER_H_
+#define AOM_DSP_BITWRITER_H_
+
+#include <assert.h>
+#include "./aom_config.h"
+#if CONFIG_EC_ADAPT && !CONFIG_EC_MULTISYMBOL
+#error "CONFIG_EC_ADAPT is enabled without enabling CONFIG_EC_MULTISYMBOL"
+#endif
+
+#if CONFIG_ANS
+#include "aom_dsp/buf_ans.h"
+#elif CONFIG_DAALA_EC
+#include "aom_dsp/daalaboolwriter.h"
+#else
+#include "aom_dsp/dkboolwriter.h"
+#endif
+#include "aom_dsp/prob.h"
+
+#if CONFIG_RD_DEBUG
+#include "av1/common/blockd.h"
+#include "av1/encoder/cost.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_ANS
+typedef struct BufAnsCoder aom_writer;
+#elif CONFIG_DAALA_EC
+typedef struct daala_writer aom_writer;
+#else
+typedef struct aom_dk_writer aom_writer;
+#endif
+
+typedef struct TOKEN_STATS {
+  int cost;
+#if CONFIG_VAR_TX
+#if CONFIG_RD_DEBUG
+  int txb_coeff_cost_map[TXB_COEFF_COST_MAP_SIZE][TXB_COEFF_COST_MAP_SIZE];
+#endif
+#endif
+} TOKEN_STATS;
+
+static INLINE void init_token_stats(TOKEN_STATS *token_stats) {
+#if CONFIG_VAR_TX
+#if CONFIG_RD_DEBUG
+  int r, c;
+  for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r) {
+    for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
+      token_stats->txb_coeff_cost_map[r][c] = 0;
+    }
+  }
+#endif
+#endif
+  token_stats->cost = 0;
+}
+
+static INLINE void aom_start_encode(aom_writer *bc, uint8_t *buffer) {
+#if CONFIG_ANS
+  (void)bc;
+  (void)buffer;
+  assert(0 && "buf_ans requires a more complicated startup procedure");
+#elif CONFIG_DAALA_EC
+  aom_daala_start_encode(bc, buffer);
+#else
+  aom_dk_start_encode(bc, buffer);
+#endif
+}
+
+static INLINE void aom_stop_encode(aom_writer *bc) {
+#if CONFIG_ANS
+  (void)bc;
+  assert(0 && "buf_ans requires a more complicated shutdown procedure");
+#elif CONFIG_DAALA_EC
+  aom_daala_stop_encode(bc);
+#else
+  aom_dk_stop_encode(bc);
+#endif
+}
+
+static INLINE void aom_write(aom_writer *br, int bit, int probability) {
+#if CONFIG_ANS
+  buf_rabs_write(br, bit, probability);
+#elif CONFIG_DAALA_EC
+  aom_daala_write(br, bit, probability);
+#else
+  aom_dk_write(br, bit, probability);
+#endif
+}
+
+static INLINE void aom_write_record(aom_writer *br, int bit, int probability,
+                                    TOKEN_STATS *token_stats) {
+  aom_write(br, bit, probability);
+#if CONFIG_RD_DEBUG
+  token_stats->cost += av1_cost_bit(probability, bit);
+#else
+  (void)token_stats;
+#endif
+}
+
+static INLINE void aom_write_bit(aom_writer *w, int bit) {
+#if CONFIG_ANS
+  buf_rabs_write_bit(w, bit);
+#elif CONFIG_DAALA_EC && CONFIG_RAWBITS
+  // Note this uses raw bits and is not the same as aom_daala_write(r, 128);
+  aom_daala_write_bit(w, bit);
+#else
+  aom_write(w, bit, 128);  // aom_prob_half
+#endif
+}
+
+static INLINE void aom_write_bit_record(aom_writer *w, int bit,
+                                        TOKEN_STATS *token_stats) {
+  aom_write_bit(w, bit);
+#if CONFIG_RD_DEBUG
+  token_stats->cost += av1_cost_bit(128, bit);  // aom_prob_half
+#else
+  (void)token_stats;
+#endif
+}
+
+static INLINE void aom_write_literal(aom_writer *w, int data, int bits) {
+  int bit;
+
+  for (bit = bits - 1; bit >= 0; bit--) aom_write_bit(w, 1 & (data >> bit));
+}
+
+static INLINE void aom_write_tree_as_bits(aom_writer *w,
+                                          const aom_tree_index *tr,
+                                          const aom_prob *probs, int bits,
+                                          int len, aom_tree_index i) {
+  do {
+    const int bit = (bits >> --len) & 1;
+    aom_write(w, bit, probs[i >> 1]);
+    i = tr[i + bit];
+  } while (len);
+}
+
+static INLINE void aom_write_tree_as_bits_record(
+    aom_writer *w, const aom_tree_index *tr, const aom_prob *probs, int bits,
+    int len, aom_tree_index i, TOKEN_STATS *token_stats) {
+  do {
+    const int bit = (bits >> --len) & 1;
+    aom_write_record(w, bit, probs[i >> 1], token_stats);
+    i = tr[i + bit];
+  } while (len);
+}
+
+#if CONFIG_EC_MULTISYMBOL
+static INLINE void aom_write_cdf(aom_writer *w, int symb,
+                                 const aom_cdf_prob *cdf, int nsymbs) {
+#if CONFIG_ANS
+  (void)nsymbs;
+  assert(cdf);
+  const aom_cdf_prob cum_prob = symb > 0 ? cdf[symb - 1] : 0;
+  const aom_cdf_prob prob = cdf[symb] - cum_prob;
+  buf_rans_write(w, cum_prob, prob);
+#elif CONFIG_DAALA_EC
+  daala_write_symbol(w, symb, cdf, nsymbs);
+#else
+#error \
+    "CONFIG_EC_MULTISYMBOL is selected without a valid backing entropy " \
+  "coder. Enable daala_ec or ans for a valid configuration."
+#endif
+}
+
+static INLINE void aom_write_symbol(aom_writer *w, int symb, aom_cdf_prob *cdf,
+                                    int nsymbs) {
+  aom_write_cdf(w, symb, cdf, nsymbs);
+#if CONFIG_EC_ADAPT
+  update_cdf(cdf, symb, nsymbs);
+#endif
+}
+
+static INLINE void aom_write_tree_as_cdf(aom_writer *w,
+                                         const aom_tree_index *tree,
+                                         const aom_prob *probs, int bits,
+                                         int len, aom_tree_index i) {
+  aom_tree_index root;
+  root = i;
+  do {
+    aom_cdf_prob cdf[16];
+    aom_tree_index index[16];
+    int path[16];
+    int dist[16];
+    int nsymbs;
+    int symb;
+    int j;
+    /* Compute the CDF of the binary tree using the given probabilities. */
+    nsymbs = tree_to_cdf(tree, probs, root, cdf, index, path, dist);
+    /* Find the symbol to code. */
+    symb = -1;
+    for (j = 0; j < nsymbs; j++) {
+      /* If this symbol codes a leaf node,  */
+      if (index[j] <= 0) {
+        if (len == dist[j] && path[j] == bits) {
+          symb = j;
+          break;
+        }
+      } else {
+        if (len > dist[j] && path[j] == bits >> (len - dist[j])) {
+          symb = j;
+          break;
+        }
+      }
+    }
+    OD_ASSERT(symb != -1);
+    aom_write_cdf(w, symb, cdf, nsymbs);
+    bits &= (1 << (len - dist[symb])) - 1;
+    len -= dist[symb];
+  } while (len);
+}
+
+#endif  // CONFIG_EC_MULTISYMBOL
+
+static INLINE void aom_write_tree(aom_writer *w, const aom_tree_index *tree,
+                                  const aom_prob *probs, int bits, int len,
+                                  aom_tree_index i) {
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_tree_as_cdf(w, tree, probs, bits, len, i);
+#else
+  aom_write_tree_as_bits(w, tree, probs, bits, len, i);
+#endif
+}
+
+static INLINE void aom_write_tree_record(aom_writer *w,
+                                         const aom_tree_index *tree,
+                                         const aom_prob *probs, int bits,
+                                         int len, aom_tree_index i,
+                                         TOKEN_STATS *token_stats) {
+#if CONFIG_EC_MULTISYMBOL
+  (void)token_stats;
+  aom_write_tree_as_cdf(w, tree, probs, bits, len, i);
+#else
+  aom_write_tree_as_bits_record(w, tree, probs, bits, len, i, token_stats);
+#endif
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITWRITER_H_

third_party/aom/aom_dsp/bitwriter_buffer.c

@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./bitwriter_buffer.h"
+
+uint32_t aom_wb_bytes_written(const struct aom_write_bit_buffer *wb) {
+  return wb->bit_offset / CHAR_BIT + (wb->bit_offset % CHAR_BIT > 0);
+}
+
+void aom_wb_write_bit(struct aom_write_bit_buffer *wb, int bit) {
+  const int off = (int)wb->bit_offset;
+  const int p = off / CHAR_BIT;
+  const int q = CHAR_BIT - 1 - off % CHAR_BIT;
+  if (q == CHAR_BIT - 1) {
+    // Zero next char and write bit
+    wb->bit_buffer[p] = bit << q;
+  } else {
+    wb->bit_buffer[p] &= ~(1 << q);
+    wb->bit_buffer[p] |= bit << q;
+  }
+  wb->bit_offset = off + 1;
+}
+
+void aom_wb_overwrite_bit(struct aom_write_bit_buffer *wb, int bit) {
+  // Do not zero bytes but overwrite exisiting values
+  const int off = (int)wb->bit_offset;
+  const int p = off / CHAR_BIT;
+  const int q = CHAR_BIT - 1 - off % CHAR_BIT;
+  wb->bit_buffer[p] &= ~(1 << q);
+  wb->bit_buffer[p] |= bit << q;
+  wb->bit_offset = off + 1;
+}
+
+void aom_wb_write_literal(struct aom_write_bit_buffer *wb, int data, int bits) {
+  int bit;
+  for (bit = bits - 1; bit >= 0; bit--) aom_wb_write_bit(wb, (data >> bit) & 1);
+}
+
+void aom_wb_overwrite_literal(struct aom_write_bit_buffer *wb, int data,
+                              int bits) {
+  int bit;
+  for (bit = bits - 1; bit >= 0; bit--)
+    aom_wb_overwrite_bit(wb, (data >> bit) & 1);
+}
+
+void aom_wb_write_inv_signed_literal(struct aom_write_bit_buffer *wb, int data,
+                                     int bits) {
+  aom_wb_write_literal(wb, data, bits + 1);
+}

third_party/aom/aom_dsp/bitwriter_buffer.h

@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITWRITER_BUFFER_H_
+#define AOM_DSP_BITWRITER_BUFFER_H_
+
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct aom_write_bit_buffer {
+  uint8_t *bit_buffer;
+  uint32_t bit_offset;
+};
+
+uint32_t aom_wb_bytes_written(const struct aom_write_bit_buffer *wb);
+
+void aom_wb_write_bit(struct aom_write_bit_buffer *wb, int bit);
+
+void aom_wb_overwrite_bit(struct aom_write_bit_buffer *wb, int bit);
+
+void aom_wb_write_literal(struct aom_write_bit_buffer *wb, int data, int bits);
+
+void aom_wb_overwrite_literal(struct aom_write_bit_buffer *wb, int data,
+                              int bits);
+
+void aom_wb_write_inv_signed_literal(struct aom_write_bit_buffer *wb, int data,
+                                     int bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITWRITER_BUFFER_H_

third_party/aom/aom_dsp/blend.h

@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BLEND_H_
+#define AOM_DSP_BLEND_H_
+
+#include "aom_ports/mem.h"
+
+// Various blending functions and macros.
+// See also the aom_blend_* functions in aom_dsp_rtcd.h
+
+// Alpha blending with alpha values from the range [0, 64], where 64
+// means use the first input and 0 means use the second input.
+
+#define AOM_BLEND_A64_ROUND_BITS 6
+#define AOM_BLEND_A64_MAX_ALPHA (1 << AOM_BLEND_A64_ROUND_BITS)  // 64
+
+#define AOM_BLEND_A64(a, v0, v1)                                          \
+  ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A64_MAX_ALPHA - (a)) * (v1), \
+                     AOM_BLEND_A64_ROUND_BITS)
+
+// Alpha blending with alpha values from the range [0, 256], where 256
+// means use the first input and 0 means use the second input.
+#define AOM_BLEND_A256_ROUND_BITS 8
+#define AOM_BLEND_A256_MAX_ALPHA (1 << AOM_BLEND_A256_ROUND_BITS)  // 256
+
+#define AOM_BLEND_A256(a, v0, v1)                                          \
+  ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A256_MAX_ALPHA - (a)) * (v1), \
+                     AOM_BLEND_A256_ROUND_BITS)
+
+// Blending by averaging.
+#define AOM_BLEND_AVG(v0, v1) ROUND_POWER_OF_TWO((v0) + (v1), 1)
+
+#endif  // AOM_DSP_BLEND_H_

third_party/aom/aom_dsp/blend_a64_hmask.c

@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "./aom_dsp_rtcd.h"
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(
+          mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]);
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                  const uint8_t *src0_8, uint32_t src0_stride,
+                                  const uint8_t *src1_8, uint32_t src1_stride,
+                                  const uint8_t *mask, int h, int w, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+  (void)bd;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(
+          mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]);
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/aom_dsp/blend_a64_mask.c

@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/blend.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+#include "./aom_dsp_rtcd.h"
+
+// Blending with alpha mask. Mask values come from the range [0, 64],
+// as described for AOM_BLEND_A64 in aom_dsp/blend.h. src0 or src1 can
+// be the same as dst, or dst can be different from both sources.
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride,
+                          const uint8_t *src0, uint32_t src0_stride,
+                          const uint8_t *src1, uint32_t src1_stride,
+                          const uint8_t *mask, uint32_t mask_stride, int h,
+                          int w, int subh, int subw) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (subw == 0 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = mask[i * mask_stride + j];
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 1) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = ROUND_POWER_OF_TWO(
+            mask[(2 * i) * mask_stride + (2 * j)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j)] +
+                mask[(2 * i) * mask_stride + (2 * j + 1)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j + 1)],
+            2);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)],
+                                    mask[i * mask_stride + (2 * j + 1)]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j],
+                                    mask[(2 * i + 1) * mask_stride + j]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_blend_a64_mask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                 const uint8_t *src0_8, uint32_t src0_stride,
+                                 const uint8_t *src1_8, uint32_t src1_stride,
+                                 const uint8_t *mask, uint32_t mask_stride,
+                                 int h, int w, int subh, int subw, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+  (void)bd;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  if (subw == 0 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = mask[i * mask_stride + j];
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 1) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = ROUND_POWER_OF_TWO(
+            mask[(2 * i) * mask_stride + (2 * j)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j)] +
+                mask[(2 * i) * mask_stride + (2 * j + 1)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j + 1)],
+            2);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)],
+                                    mask[i * mask_stride + (2 * j + 1)]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j],
+                                    mask[(2 * i + 1) * mask_stride + j]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/aom_dsp/blend_a64_vmask.c

@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "./aom_dsp_rtcd.h"
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  for (i = 0; i < h; ++i) {
+    const int m = mask[i];
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                              src1[i * src1_stride + j]);
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                  const uint8_t *src0_8, uint32_t src0_stride,
+                                  const uint8_t *src1_8, uint32_t src1_stride,
+                                  const uint8_t *mask, int h, int w, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+  (void)bd;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  for (i = 0; i < h; ++i) {
+    const int m = mask[i];
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                              src1[i * src1_stride + j]);
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/aom_dsp/buf_ans.c

@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <string.h>
+
+#include "aom_dsp/buf_ans.h"
+#include "aom_mem/aom_mem.h"
+#include "aom/internal/aom_codec_internal.h"
+
+void aom_buf_ans_alloc(struct BufAnsCoder *c,
+                       struct aom_internal_error_info *error, int size) {
+  c->error = error;
+  c->size = size;
+  assert(c->size > 1);
+  AOM_CHECK_MEM_ERROR(error, c->buf, aom_malloc(c->size * sizeof(*c->buf)));
+  // Initialize to overfull to trigger the assert in write.
+  c->offset = c->size + 1;
+}
+
+void aom_buf_ans_free(struct BufAnsCoder *c) {
+  aom_free(c->buf);
+  c->buf = NULL;
+  c->size = 0;
+}
+
+#if !ANS_MAX_SYMBOLS
+void aom_buf_ans_grow(struct BufAnsCoder *c) {
+  struct buffered_ans_symbol *new_buf = NULL;
+  int new_size = c->size * 2;
+  AOM_CHECK_MEM_ERROR(c->error, new_buf,
+                      aom_malloc(new_size * sizeof(*new_buf)));
+  memcpy(new_buf, c->buf, c->size * sizeof(*c->buf));
+  aom_free(c->buf);
+  c->buf = new_buf;
+  c->size = new_size;
+}
+#endif
+
+void aom_buf_ans_flush(struct BufAnsCoder *const c) {
+  int offset;
+#if ANS_MAX_SYMBOLS
+  if (c->offset == 0) return;
+#endif
+  assert(c->offset > 0);
+  offset = c->offset - 1;
+  // Code the first symbol such that it brings the state to the smallest normal
+  // state from an initial state that would have been a subnormal/refill state.
+  if (c->buf[offset].method == ANS_METHOD_RANS) {
+    c->ans.state += c->buf[offset].val_start;
+  } else {
+    c->ans.state += c->buf[offset].val_start ? c->buf[offset].prob : 0;
+  }
+  for (offset = offset - 1; offset >= 0; --offset) {
+    if (c->buf[offset].method == ANS_METHOD_RANS) {
+      rans_write(&c->ans, c->buf[offset].val_start, c->buf[offset].prob);
+    } else {
+      rabs_write(&c->ans, (uint8_t)c->buf[offset].val_start,
+                 (AnsP8)c->buf[offset].prob);
+    }
+  }
+  c->offset = 0;
+  c->output_bytes += ans_write_end(&c->ans);
+}

third_party/aom/aom_dsp/buf_ans.h

@@ -0,0 +1,133 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BUF_ANS_H_
+#define AOM_DSP_BUF_ANS_H_
+// Buffered forward ANS writer.
+// Symbols are written to the writer in forward (decode) order and serialized
+// backwards due to ANS's stack like behavior.
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/ans.h"
+#include "aom_dsp/answriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+#define ANS_METHOD_RABS 0
+#define ANS_METHOD_RANS 1
+
+struct buffered_ans_symbol {
+  unsigned int method : 1;  // one of ANS_METHOD_RABS or ANS_METHOD_RANS
+  // TODO(aconverse): Should be possible to write this in terms of start for ABS
+  unsigned int val_start : RANS_PROB_BITS;  // Boolean value for ABS
+                                            // start in symbol cycle for Rans
+  unsigned int prob : RANS_PROB_BITS;       // Probability of this symbol
+};
+
+struct BufAnsCoder {
+  struct aom_internal_error_info *error;
+  struct buffered_ans_symbol *buf;
+  struct AnsCoder ans;
+  int size;
+  int offset;
+  int output_bytes;
+#if ANS_MAX_SYMBOLS
+  int window_size;
+#endif
+};
+
+// Allocate a buffered ANS coder to store size symbols.
+// When ANS_MAX_SYMBOLS is turned on, the size is the fixed size of each ANS
+// partition.
+// When ANS_MAX_SYMBOLS is turned off, size is merely an initial hint and the
+// buffer will grow on demand
+void aom_buf_ans_alloc(struct BufAnsCoder *c,
+                       struct aom_internal_error_info *error, int hint);
+
+void aom_buf_ans_free(struct BufAnsCoder *c);
+
+#if !ANS_MAX_SYMBOLS
+void aom_buf_ans_grow(struct BufAnsCoder *c);
+#endif
+
+void aom_buf_ans_flush(struct BufAnsCoder *const c);
+
+static INLINE void buf_ans_write_init(struct BufAnsCoder *const c,
+                                      uint8_t *const output_buffer) {
+  c->offset = 0;
+  c->output_bytes = 0;
+  ans_write_init(&c->ans, output_buffer);
+}
+
+static INLINE void buf_rabs_write(struct BufAnsCoder *const c, uint8_t val,
+                                  AnsP8 prob) {
+  assert(c->offset <= c->size);
+#if !ANS_MAX_SYMBOLS
+  if (c->offset == c->size) {
+    aom_buf_ans_grow(c);
+  }
+#endif
+  c->buf[c->offset].method = ANS_METHOD_RABS;
+  c->buf[c->offset].val_start = val;
+  c->buf[c->offset].prob = prob;
+  ++c->offset;
+#if ANS_MAX_SYMBOLS
+  if (c->offset == c->size) aom_buf_ans_flush(c);
+#endif
+}
+
+// Buffer one symbol for encoding using rANS.
+// cum_prob: The cumulative probability before this symbol (the offset of
+// the symbol in the symbol cycle)
+// prob: The probability of this symbol (l_s from the paper)
+// RANS_PRECISION takes the place of m from the paper.
+static INLINE void buf_rans_write(struct BufAnsCoder *const c,
+                                  aom_cdf_prob cum_prob, aom_cdf_prob prob) {
+  assert(c->offset <= c->size);
+#if !ANS_MAX_SYMBOLS
+  if (c->offset == c->size) {
+    aom_buf_ans_grow(c);
+  }
+#endif
+  c->buf[c->offset].method = ANS_METHOD_RANS;
+  c->buf[c->offset].val_start = cum_prob;
+  c->buf[c->offset].prob = prob;
+  ++c->offset;
+#if ANS_MAX_SYMBOLS
+  if (c->offset == c->size) aom_buf_ans_flush(c);
+#endif
+}
+
+static INLINE void buf_rabs_write_bit(struct BufAnsCoder *c, int bit) {
+  buf_rabs_write(c, bit, 128);
+}
+
+static INLINE void buf_rabs_write_literal(struct BufAnsCoder *c, int literal,
+                                          int bits) {
+  int bit;
+
+  assert(bits < 31);
+  for (bit = bits - 1; bit >= 0; bit--)
+    buf_rabs_write_bit(c, 1 & (literal >> bit));
+}
+
+static INLINE int buf_ans_write_end(struct BufAnsCoder *const c) {
+  assert(c->offset == 0);
+  return c->output_bytes;
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_BUF_ANS_H_

third_party/aom/aom_dsp/daalaboolreader.c

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/daalaboolreader.h"
+
+int aom_daala_reader_init(daala_reader *r, const uint8_t *buffer, int size) {
+  if (size && !buffer) {
+    return 1;
+  }
+  r->buffer_end = buffer + size;
+  r->buffer = buffer;
+  od_ec_dec_init(&r->ec, buffer, size - 1);
+#if CONFIG_ACCOUNTING
+  r->accounting = NULL;
+#endif
+  return 0;
+}
+
+const uint8_t *aom_daala_reader_find_end(daala_reader *r) {
+  return r->buffer_end;
+}
+
+uint32_t aom_daala_reader_tell(const daala_reader *r) {
+  return od_ec_dec_tell(&r->ec);
+}
+
+uint32_t aom_daala_reader_tell_frac(const daala_reader *r) {
+  return od_ec_dec_tell_frac(&r->ec);
+}

third_party/aom/aom_dsp/daalaboolreader.h

@@ -0,0 +1,164 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DAALABOOLREADER_H_
+#define AOM_DSP_DAALABOOLREADER_H_
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/entdec.h"
+#include "aom_dsp/prob.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+#include <stdio.h>
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct daala_reader {
+  const uint8_t *buffer;
+  const uint8_t *buffer_end;
+  od_ec_dec ec;
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+};
+
+typedef struct daala_reader daala_reader;
+
+int aom_daala_reader_init(daala_reader *r, const uint8_t *buffer, int size);
+const uint8_t *aom_daala_reader_find_end(daala_reader *r);
+uint32_t aom_daala_reader_tell(const daala_reader *r);
+uint32_t aom_daala_reader_tell_frac(const daala_reader *r);
+
+static INLINE int aom_daala_read(daala_reader *r, int prob) {
+  int bit;
+#if CONFIG_EC_SMALLMUL
+  int p = (0x7FFFFF - (prob << 15) + prob) >> 8;
+#else
+  int p = ((prob << 15) + 256 - prob) >> 8;
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+/*{
+  const int queue_r = bitstream_queue_get_read();
+  const int frame_idx = bitstream_queue_get_frame_read();
+  if (frame_idx == 0 && queue_r == 0) {
+    fprintf(stderr, "\n *** bitstream queue at frame_idx_r %d queue_r %d\n",
+            frame_idx, queue_r);
+  }
+}*/
+#endif
+
+  bit = od_ec_decode_bool_q15(&r->ec, p);
+
+#if CONFIG_BITSTREAM_DEBUG
+  {
+    int i;
+    int ref_bit, ref_nsymbs;
+    aom_cdf_prob ref_cdf[16];
+    const int queue_r = bitstream_queue_get_read();
+    const int frame_idx = bitstream_queue_get_frame_read();
+    bitstream_queue_pop(&ref_bit, ref_cdf, &ref_nsymbs);
+    if (ref_nsymbs != 2) {
+      fprintf(stderr,
+              "\n *** [bit] nsymbs error, frame_idx_r %d nsymbs %d ref_nsymbs "
+              "%d queue_r %d\n",
+              frame_idx, 2, ref_nsymbs, queue_r);
+      assert(0);
+    }
+    if ((ref_nsymbs != 2) || (ref_cdf[0] != (aom_cdf_prob)p) ||
+        (ref_cdf[1] != 32767)) {
+      fprintf(stderr,
+              "\n *** [bit] cdf error, frame_idx_r %d cdf {%d, %d} ref_cdf {%d",
+              frame_idx, p, 32767, ref_cdf[0]);
+      for (i = 1; i < ref_nsymbs; ++i) fprintf(stderr, ", %d", ref_cdf[i]);
+      fprintf(stderr, "} queue_r %d\n", queue_r);
+      assert(0);
+    }
+    if (bit != ref_bit) {
+      fprintf(stderr,
+              "\n *** [bit] symb error, frame_idx_r %d symb %d ref_symb %d "
+              "queue_r %d\n",
+              frame_idx, bit, ref_bit, queue_r);
+      assert(0);
+    }
+  }
+#endif
+
+  return bit;
+}
+
+#if CONFIG_RAWBITS
+static INLINE int aom_daala_read_bit(daala_reader *r) {
+  return od_ec_dec_bits(&r->ec, 1, "aom_bits");
+}
+#endif
+
+static INLINE int aom_daala_reader_has_error(daala_reader *r) {
+  return r->ec.error;
+}
+
+static INLINE int daala_read_symbol(daala_reader *r, const aom_cdf_prob *cdf,
+                                    int nsymbs) {
+  int symb;
+  symb = od_ec_decode_cdf_q15(&r->ec, cdf, nsymbs);
+
+#if CONFIG_BITSTREAM_DEBUG
+  {
+    int i;
+    int cdf_error = 0;
+    int ref_symb, ref_nsymbs;
+    aom_cdf_prob ref_cdf[16];
+    const int queue_r = bitstream_queue_get_read();
+    const int frame_idx = bitstream_queue_get_frame_read();
+    bitstream_queue_pop(&ref_symb, ref_cdf, &ref_nsymbs);
+    if (nsymbs != ref_nsymbs) {
+      fprintf(stderr,
+              "\n *** nsymbs error, frame_idx_r %d nsymbs %d ref_nsymbs %d "
+              "queue_r %d\n",
+              frame_idx, nsymbs, ref_nsymbs, queue_r);
+      cdf_error = 0;
+      assert(0);
+    } else {
+      for (i = 0; i < nsymbs; ++i)
+        if (cdf[i] != ref_cdf[i]) cdf_error = 1;
+    }
+    if (cdf_error) {
+      fprintf(stderr, "\n *** cdf error, frame_idx_r %d cdf {%d", frame_idx,
+              cdf[0]);
+      for (i = 1; i < nsymbs; ++i) fprintf(stderr, ", %d", cdf[i]);
+      fprintf(stderr, "} ref_cdf {%d", ref_cdf[0]);
+      for (i = 1; i < ref_nsymbs; ++i) fprintf(stderr, ", %d", ref_cdf[i]);
+      fprintf(stderr, "} queue_r %d\n", queue_r);
+      assert(0);
+    }
+    if (symb != ref_symb) {
+      fprintf(
+          stderr,
+          "\n *** symb error, frame_idx_r %d symb %d ref_symb %d queue_r %d\n",
+          frame_idx, symb, ref_symb, queue_r);
+      assert(0);
+    }
+  }
+#endif
+
+  return symb;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif

third_party/aom/aom_dsp/daalaboolwriter.c

@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <string.h>
+#include "aom_dsp/daalaboolwriter.h"
+
+void aom_daala_start_encode(daala_writer *br, uint8_t *source) {
+  br->buffer = source;
+  br->pos = 0;
+  od_ec_enc_init(&br->ec, 62025);
+}
+
+void aom_daala_stop_encode(daala_writer *br) {
+  uint32_t daala_bytes;
+  unsigned char *daala_data;
+  daala_data = od_ec_enc_done(&br->ec, &daala_bytes);
+  memcpy(br->buffer, daala_data, daala_bytes);
+  br->pos = daala_bytes;
+  /* Prevent ec bitstream from being detected as a superframe marker.
+     Must always be added, so that rawbits knows the exact length of the
+      bitstream. */
+  br->buffer[br->pos++] = 0;
+  od_ec_enc_clear(&br->ec);
+}

third_party/aom/aom_dsp/daalaboolwriter.h

@@ -0,0 +1,87 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DAALABOOLWRITER_H_
+#define AOM_DSP_DAALABOOLWRITER_H_
+
+#include <stdio.h>
+
+#include "aom_dsp/entenc.h"
+#include "aom_dsp/prob.h"
+#if CONFIG_BITSTREAM_DEBUG
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct daala_writer {
+  unsigned int pos;
+  uint8_t *buffer;
+  od_ec_enc ec;
+};
+
+typedef struct daala_writer daala_writer;
+
+void aom_daala_start_encode(daala_writer *w, uint8_t *buffer);
+void aom_daala_stop_encode(daala_writer *w);
+
+static INLINE void aom_daala_write(daala_writer *w, int bit, int prob) {
+#if CONFIG_EC_SMALLMUL
+  int p = (0x7FFFFF - (prob << 15) + prob) >> 8;
+#else
+  int p = ((prob << 15) + 256 - prob) >> 8;
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+  aom_cdf_prob cdf[2] = { (aom_cdf_prob)p, 32767 };
+  /*int queue_r = 0;
+  int frame_idx_r = 0;
+  int queue_w = bitstream_queue_get_write();
+  int frame_idx_w = bitstream_queue_get_frame_write();
+  if (frame_idx_w == frame_idx_r && queue_w == queue_r) {
+    fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n",
+    frame_idx_w, queue_w);
+  }*/
+  bitstream_queue_push(bit, cdf, 2);
+#endif
+
+  od_ec_encode_bool_q15(&w->ec, bit, p);
+}
+
+#if CONFIG_RAWBITS
+static INLINE void aom_daala_write_bit(daala_writer *w, int bit) {
+  od_ec_enc_bits(&w->ec, bit, 1);
+}
+#endif
+
+static INLINE void daala_write_symbol(daala_writer *w, int symb,
+                                      const aom_cdf_prob *cdf, int nsymbs) {
+#if CONFIG_BITSTREAM_DEBUG
+  /*int queue_r = 0;
+  int frame_idx_r = 0;
+  int queue_w = bitstream_queue_get_write();
+  int frame_idx_w = bitstream_queue_get_frame_write();
+  if (frame_idx_w == frame_idx_r && queue_w == queue_r) {
+    fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n",
+    frame_idx_w, queue_w);
+  }*/
+  bitstream_queue_push(symb, cdf, nsymbs);
+#endif
+
+  od_ec_encode_cdf_q15(&w->ec, symb, cdf, nsymbs);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif

third_party/aom/aom_dsp/dkboolreader.c

@@ -0,0 +1,110 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_config.h"
+
+#include "aom_dsp/dkboolreader.h"
+#include "aom_dsp/prob.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_util/endian_inl.h"
+
+static INLINE int aom_dk_read_bit(struct aom_dk_reader *r) {
+  return aom_dk_read(r, 128);  // aom_prob_half
+}
+
+int aom_dk_reader_init(struct aom_dk_reader *r, const uint8_t *buffer,
+                       size_t size, aom_decrypt_cb decrypt_cb,
+                       void *decrypt_state) {
+  if (size && !buffer) {
+    return 1;
+  } else {
+    r->buffer_end = buffer + size;
+    r->buffer_start = r->buffer = buffer;
+    r->value = 0;
+    r->count = -8;
+    r->range = 255;
+    r->decrypt_cb = decrypt_cb;
+    r->decrypt_state = decrypt_state;
+    aom_dk_reader_fill(r);
+#if CONFIG_ACCOUNTING
+    r->accounting = NULL;
+#endif
+    return aom_dk_read_bit(r) != 0;  // marker bit
+  }
+}
+
+void aom_dk_reader_fill(struct aom_dk_reader *r) {
+  const uint8_t *const buffer_end = r->buffer_end;
+  const uint8_t *buffer = r->buffer;
+  const uint8_t *buffer_start = buffer;
+  BD_VALUE value = r->value;
+  int count = r->count;
+  const size_t bytes_left = buffer_end - buffer;
+  const size_t bits_left = bytes_left * CHAR_BIT;
+  int shift = BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
+
+  if (r->decrypt_cb) {
+    size_t n = AOMMIN(sizeof(r->clear_buffer), bytes_left);
+    r->decrypt_cb(r->decrypt_state, buffer, r->clear_buffer, (int)n);
+    buffer = r->clear_buffer;
+    buffer_start = r->clear_buffer;
+  }
+  if (bits_left > BD_VALUE_SIZE) {
+    const int bits = (shift & 0xfffffff8) + CHAR_BIT;
+    BD_VALUE nv;
+    BD_VALUE big_endian_values;
+    memcpy(&big_endian_values, buffer, sizeof(BD_VALUE));
+#if SIZE_MAX == 0xffffffffffffffffULL
+    big_endian_values = HToBE64(big_endian_values);
+#else
+    big_endian_values = HToBE32(big_endian_values);
+#endif
+    nv = big_endian_values >> (BD_VALUE_SIZE - bits);
+    count += bits;
+    buffer += (bits >> 3);
+    value = r->value | (nv << (shift & 0x7));
+  } else {
+    const int bits_over = (int)(shift + CHAR_BIT - (int)bits_left);
+    int loop_end = 0;
+    if (bits_over >= 0) {
+      count += LOTS_OF_BITS;
+      loop_end = bits_over;
+    }
+
+    if (bits_over < 0 || bits_left) {
+      while (shift >= loop_end) {
+        count += CHAR_BIT;
+        value |= (BD_VALUE)*buffer++ << shift;
+        shift -= CHAR_BIT;
+      }
+    }
+  }
+
+  // NOTE: Variable 'buffer' may not relate to 'r->buffer' after decryption,
+  // so we increase 'r->buffer' by the amount that 'buffer' moved, rather than
+  // assign 'buffer' to 'r->buffer'.
+  r->buffer += buffer - buffer_start;
+  r->value = value;
+  r->count = count;
+}
+
+const uint8_t *aom_dk_reader_find_end(struct aom_dk_reader *r) {
+  // Find the end of the coded buffer
+  while (r->count > CHAR_BIT && r->count < BD_VALUE_SIZE) {
+    r->count -= CHAR_BIT;
+    r->buffer--;
+  }
+  return r->buffer;
+}

third_party/aom/aom_dsp/dkboolreader.h

@@ -0,0 +1,181 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DKBOOLREADER_H_
+#define AOM_DSP_DKBOOLREADER_H_
+
+#include <assert.h>
+#include <stddef.h>
+#include <limits.h>
+
+#include "./aom_config.h"
+#if CONFIG_BITSTREAM_DEBUG
+#include <assert.h>
+#include <stdio.h>
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#include "aom_ports/mem.h"
+#include "aom/aomdx.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/prob.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef size_t BD_VALUE;
+
+#define BD_VALUE_SIZE ((int)sizeof(BD_VALUE) * CHAR_BIT)
+
+// This is meant to be a large, positive constant that can still be efficiently
+// loaded as an immediate (on platforms like ARM, for example).
+// Even relatively modest values like 100 would work fine.
+#define LOTS_OF_BITS 0x40000000
+
+struct aom_dk_reader {
+  // Be careful when reordering this struct, it may impact the cache negatively.
+  BD_VALUE value;
+  unsigned int range;
+  int count;
+  const uint8_t *buffer_start;
+  const uint8_t *buffer_end;
+  const uint8_t *buffer;
+  aom_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+  uint8_t clear_buffer[sizeof(BD_VALUE) + 1];
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+};
+
+int aom_dk_reader_init(struct aom_dk_reader *r, const uint8_t *buffer,
+                       size_t size, aom_decrypt_cb decrypt_cb,
+                       void *decrypt_state);
+
+void aom_dk_reader_fill(struct aom_dk_reader *r);
+
+const uint8_t *aom_dk_reader_find_end(struct aom_dk_reader *r);
+
+static INLINE uint32_t aom_dk_reader_tell(const struct aom_dk_reader *r) {
+  const uint32_t bits_read =
+      (uint32_t)((r->buffer - r->buffer_start) * CHAR_BIT);
+  const int count =
+      (r->count < LOTS_OF_BITS) ? r->count : r->count - LOTS_OF_BITS;
+  assert(r->buffer >= r->buffer_start);
+  return bits_read - (count + CHAR_BIT);
+}
+
+/*The resolution of fractional-precision bit usage measurements, i.e.,
+   3 => 1/8th bits.*/
+#define DK_BITRES (3)
+
+static INLINE uint32_t aom_dk_reader_tell_frac(const struct aom_dk_reader *r) {
+  uint32_t num_bits;
+  uint32_t range;
+  int l;
+  int i;
+  num_bits = aom_dk_reader_tell(r) << DK_BITRES;
+  range = r->range;
+  l = 0;
+  for (i = DK_BITRES; i-- > 0;) {
+    int b;
+    range = range * range >> 7;
+    b = (int)(range >> 8);
+    l = l << 1 | b;
+    range >>= b;
+  }
+  return num_bits - l;
+}
+
+static INLINE int aom_dk_reader_has_error(struct aom_dk_reader *r) {
+  // Check if we have reached the end of the buffer.
+  //
+  // Variable 'count' stores the number of bits in the 'value' buffer, minus
+  // 8. The top byte is part of the algorithm, and the remainder is buffered
+  // to be shifted into it. So if count == 8, the top 16 bits of 'value' are
+  // occupied, 8 for the algorithm and 8 in the buffer.
+  //
+  // When reading a byte from the user's buffer, count is filled with 8 and
+  // one byte is filled into the value buffer. When we reach the end of the
+  // data, count is additionally filled with LOTS_OF_BITS. So when
+  // count == LOTS_OF_BITS - 1, the user's data has been exhausted.
+  //
+  // 1 if we have tried to decode bits after the end of stream was encountered.
+  // 0 No error.
+  return r->count > BD_VALUE_SIZE && r->count < LOTS_OF_BITS;
+}
+
+static INLINE int aom_dk_read(struct aom_dk_reader *r, int prob) {
+  unsigned int bit = 0;
+  BD_VALUE value;
+  BD_VALUE bigsplit;
+  int count;
+  unsigned int range;
+  unsigned int split = (r->range * prob + (256 - prob)) >> CHAR_BIT;
+
+  if (r->count < 0) aom_dk_reader_fill(r);
+
+  value = r->value;
+  count = r->count;
+
+  bigsplit = (BD_VALUE)split << (BD_VALUE_SIZE - CHAR_BIT);
+
+  range = split;
+
+  if (value >= bigsplit) {
+    range = r->range - split;
+    value = value - bigsplit;
+    bit = 1;
+  }
+
+  {
+    register int shift = aom_norm[range];
+    range <<= shift;
+    value <<= shift;
+    count -= shift;
+  }
+  r->value = value;
+  r->count = count;
+  r->range = range;
+
+#if CONFIG_BITSTREAM_DEBUG
+  {
+    int ref_bit, ref_prob;
+    const int queue_r = bitstream_queue_get_read();
+    const int frame_idx = bitstream_queue_get_frame_read();
+    bitstream_queue_pop(&ref_bit, &ref_prob);
+    if (prob != ref_prob) {
+      fprintf(
+          stderr,
+          "\n *** prob error, frame_idx_r %d prob %d ref_prob %d queue_r %d\n",
+          frame_idx, prob, ref_prob, queue_r);
+      assert(0);
+    }
+    if ((int)bit != ref_bit) {
+      fprintf(stderr, "\n *** bit error, frame_idx_r %d bit %d ref_bit %d\n",
+              frame_idx, bit, ref_bit);
+      assert(0);
+    }
+  }
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  return bit;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_DKBOOLREADER_H_

third_party/aom/aom_dsp/dkboolwriter.c

@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./dkboolwriter.h"
+
+static INLINE void aom_dk_write_bit(aom_dk_writer *w, int bit) {
+  aom_dk_write(w, bit, 128);  // aom_prob_half
+}
+
+void aom_dk_start_encode(aom_dk_writer *br, uint8_t *source) {
+  br->lowvalue = 0;
+  br->range = 255;
+  br->count = -24;
+  br->buffer = source;
+  br->pos = 0;
+  aom_dk_write_bit(br, 0);
+}
+
+void aom_dk_stop_encode(aom_dk_writer *br) {
+  int i;
+
+#if CONFIG_BITSTREAM_DEBUG
+  bitstream_queue_set_skip_write(1);
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  for (i = 0; i < 32; i++) aom_dk_write_bit(br, 0);
+
+#if CONFIG_BITSTREAM_DEBUG
+  bitstream_queue_set_skip_write(0);
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  // Ensure there's no ambigous collision with any index marker bytes
+  if ((br->buffer[br->pos - 1] & 0xe0) == 0xc0) br->buffer[br->pos++] = 0;
+}

third_party/aom/aom_dsp/dkboolwriter.h

@@ -0,0 +1,104 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DKBOOLWRITER_H_
+#define AOM_DSP_DKBOOLWRITER_H_
+
+#include "./aom_config.h"
+
+#if CONFIG_BITSTREAM_DEBUG
+#include <stdio.h>
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#include "aom_dsp/prob.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct aom_dk_writer {
+  unsigned int lowvalue;
+  unsigned int range;
+  int count;
+  unsigned int pos;
+  uint8_t *buffer;
+} aom_dk_writer;
+
+void aom_dk_start_encode(aom_dk_writer *bc, uint8_t *buffer);
+void aom_dk_stop_encode(aom_dk_writer *bc);
+
+static INLINE void aom_dk_write(aom_dk_writer *br, int bit, int probability) {
+  unsigned int split;
+  int count = br->count;
+  unsigned int range = br->range;
+  unsigned int lowvalue = br->lowvalue;
+  register int shift;
+
+#if CONFIG_BITSTREAM_DEBUG
+  // int queue_r = 0;
+  // int frame_idx_r = 0;
+  // int queue_w = bitstream_queue_get_write();
+  // int frame_idx_w = bitstream_queue_get_frame_write();
+  // if (frame_idx_w == frame_idx_r && queue_w == queue_r) {
+  //   fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n",
+  //   frame_idx_w, queue_w);
+  // }
+  bitstream_queue_push(bit, probability);
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  split = 1 + (((range - 1) * probability) >> 8);
+
+  range = split;
+
+  if (bit) {
+    lowvalue += split;
+    range = br->range - split;
+  }
+
+  shift = aom_norm[range];
+
+  range <<= shift;
+  count += shift;
+
+  if (count >= 0) {
+    int offset = shift - count;
+
+    if ((lowvalue << (offset - 1)) & 0x80000000) {
+      int x = br->pos - 1;
+
+      while (x >= 0 && br->buffer[x] == 0xff) {
+        br->buffer[x] = 0;
+        x--;
+      }
+
+      br->buffer[x] += 1;
+    }
+
+    br->buffer[br->pos++] = (lowvalue >> (24 - offset));
+    lowvalue <<= offset;
+    shift = count;
+    lowvalue &= 0xffffff;
+    count -= 8;
+  }
+
+  lowvalue <<= shift;
+  br->count = count;
+  br->lowvalue = lowvalue;
+  br->range = range;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_DKBOOLWRITER_H_

third_party/aom/aom_dsp/entcode.c

@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "./config.h"
+#endif
+
+#include "aom_dsp/entcode.h"
+
+/*Given the current total integer number of bits used and the current value of
+   rng, computes the fraction number of bits used to OD_BITRES precision.
+  This is used by od_ec_enc_tell_frac() and od_ec_dec_tell_frac().
+  nbits_total: The number of whole bits currently used, i.e., the value
+                returned by od_ec_enc_tell() or od_ec_dec_tell().
+  rng: The current value of rng from either the encoder or decoder state.
+  Return: The number of bits scaled by 2**OD_BITRES.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+uint32_t od_ec_tell_frac(uint32_t nbits_total, uint32_t rng) {
+  uint32_t nbits;
+  int l;
+  int i;
+  /*To handle the non-integral number of bits still left in the encoder/decoder
+     state, we compute the worst-case number of bits of val that must be
+     encoded to ensure that the value is inside the range for any possible
+     subsequent bits.
+    The computation here is independent of val itself (the decoder does not
+     even track that value), even though the real number of bits used after
+     od_ec_enc_done() may be 1 smaller if rng is a power of two and the
+     corresponding trailing bits of val are all zeros.
+    If we did try to track that special case, then coding a value with a
+     probability of 1/(1 << n) might sometimes appear to use more than n bits.
+    This may help explain the surprising result that a newly initialized
+     encoder or decoder claims to have used 1 bit.*/
+  nbits = nbits_total << OD_BITRES;
+  l = 0;
+  for (i = OD_BITRES; i-- > 0;) {
+    int b;
+    rng = rng * rng >> 15;
+    b = (int)(rng >> 16);
+    l = l << 1 | b;
+    rng >>= b;
+  }
+  return nbits - l;
+}

third_party/aom/aom_dsp/entcode.h

@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_entcode_H)
+#define _entcode_H (1)
+#include <limits.h>
+#include <stddef.h>
+#include "av1/common/odintrin.h"
+
+/*OPT: od_ec_window must be at least 32 bits, but if you have fast arithmetic
+   on a larger type, you can speed up the decoder by using it here.*/
+typedef uint32_t od_ec_window;
+
+#define OD_EC_WINDOW_SIZE ((int)sizeof(od_ec_window) * CHAR_BIT)
+
+/*The number of bits to use for the range-coded part of unsigned integers.*/
+#define OD_EC_UINT_BITS (4)
+
+/*The resolution of fractional-precision bit usage measurements, i.e.,
+   3 => 1/8th bits.*/
+#define OD_BITRES (3)
+
+/*With CONFIG_EC_SMALLMUL, the value stored in a CDF is 32768 minus the actual
+   Q15 cumulative probability (an "inverse" CDF).
+  This function converts from one representation to the other (and is its own
+   inverse).*/
+#if CONFIG_EC_SMALLMUL
+#define OD_ICDF(x) (32768U - (x))
+#else
+#define OD_ICDF(x) (x)
+#endif
+
+/*See entcode.c for further documentation.*/
+
+OD_WARN_UNUSED_RESULT uint32_t od_ec_tell_frac(uint32_t nbits_total,
+                                               uint32_t rng);
+
+#endif

third_party/aom/aom_dsp/entdec.c

@@ -0,0 +1,300 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "./config.h"
+#endif
+
+#include "aom_dsp/entdec.h"
+
+/*A range decoder.
+  This is an entropy decoder based upon \cite{Mar79}, which is itself a
+   rediscovery of the FIFO arithmetic code introduced by \cite{Pas76}.
+  It is very similar to arithmetic encoding, except that encoding is done with
+   digits in any base, instead of with bits, and so it is faster when using
+   larger bases (i.e.: a byte).
+  The author claims an average waste of $\frac{1}{2}\log_b(2b)$ bits, where $b$
+   is the base, longer than the theoretical optimum, but to my knowledge there
+   is no published justification for this claim.
+  This only seems true when using near-infinite precision arithmetic so that
+   the process is carried out with no rounding errors.
+
+  An excellent description of implementation details is available at
+   http://www.arturocampos.com/ac_range.html
+  A recent work \cite{MNW98} which proposes several changes to arithmetic
+   encoding for efficiency actually re-discovers many of the principles
+   behind range encoding, and presents a good theoretical analysis of them.
+
+  End of stream is handled by writing out the smallest number of bits that
+   ensures that the stream will be correctly decoded regardless of the value of
+   any subsequent bits.
+  od_ec_dec_tell() can be used to determine how many bits were needed to decode
+   all the symbols thus far; other data can be packed in the remaining bits of
+   the input buffer.
+  @PHDTHESIS{Pas76,
+    author="Richard Clark Pasco",
+    title="Source coding algorithms for fast data compression",
+    school="Dept. of Electrical Engineering, Stanford University",
+    address="Stanford, CA",
+    month=May,
+    year=1976,
+    URL="http://www.richpasco.org/scaffdc.pdf"
+  }
+  @INPROCEEDINGS{Mar79,
+   author="Martin, G.N.N.",
+   title="Range encoding: an algorithm for removing redundancy from a digitised
+    message",
+   booktitle="Video & Data Recording Conference",
+   year=1979,
+   address="Southampton",
+   month=Jul,
+   URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz"
+  }
+  @ARTICLE{MNW98,
+   author="Alistair Moffat and Radford Neal and Ian H. Witten",
+   title="Arithmetic Coding Revisited",
+   journal="{ACM} Transactions on Information Systems",
+   year=1998,
+   volume=16,
+   number=3,
+   pages="256--294",
+   month=Jul,
+   URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf"
+  }*/
+
+/*This is meant to be a large, positive constant that can still be efficiently
+   loaded as an immediate (on platforms like ARM, for example).
+  Even relatively modest values like 100 would work fine.*/
+#define OD_EC_LOTS_OF_BITS (0x4000)
+
+static void od_ec_dec_refill(od_ec_dec *dec) {
+  int s;
+  od_ec_window dif;
+  int16_t cnt;
+  const unsigned char *bptr;
+  const unsigned char *end;
+  dif = dec->dif;
+  cnt = dec->cnt;
+  bptr = dec->bptr;
+  end = dec->end;
+  s = OD_EC_WINDOW_SIZE - 9 - (cnt + 15);
+  for (; s >= 0 && bptr < end; s -= 8, bptr++) {
+    OD_ASSERT(s <= OD_EC_WINDOW_SIZE - 8);
+    dif ^= (od_ec_window)bptr[0] << s;
+    cnt += 8;
+  }
+  if (bptr >= end) {
+    dec->tell_offs += OD_EC_LOTS_OF_BITS - cnt;
+    cnt = OD_EC_LOTS_OF_BITS;
+  }
+  dec->dif = dif;
+  dec->cnt = cnt;
+  dec->bptr = bptr;
+}
+
+/*Takes updated dif and range values, renormalizes them so that
+   32768 <= rng < 65536 (reading more bytes from the stream into dif if
+   necessary), and stores them back in the decoder context.
+  dif: The new value of dif.
+  rng: The new value of the range.
+  ret: The value to return.
+  Return: ret.
+          This allows the compiler to jump to this function via a tail-call.*/
+static int od_ec_dec_normalize(od_ec_dec *dec, od_ec_window dif, unsigned rng,
+                               int ret) {
+  int d;
+  OD_ASSERT(rng <= 65535U);
+  d = 16 - OD_ILOG_NZ(rng);
+  dec->cnt -= d;
+#if CONFIG_EC_SMALLMUL
+  /*This is equivalent to shifting in 1's instead of 0's.*/
+  dec->dif = ((dif + 1) << d) - 1;
+#else
+  dec->dif = dif << d;
+#endif
+  dec->rng = rng << d;
+  if (dec->cnt < 0) od_ec_dec_refill(dec);
+  return ret;
+}
+
+/*Initializes the decoder.
+  buf: The input buffer to use.
+  Return: 0 on success, or a negative value on error.*/
+void od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf,
+                    uint32_t storage) {
+  dec->buf = buf;
+  dec->eptr = buf + storage;
+  dec->end_window = 0;
+  dec->nend_bits = 0;
+  dec->tell_offs = 10 - (OD_EC_WINDOW_SIZE - 8);
+  dec->end = buf + storage;
+  dec->bptr = buf;
+#if CONFIG_EC_SMALLMUL
+  dec->dif = ((od_ec_window)1 << (OD_EC_WINDOW_SIZE - 1)) - 1;
+#else
+  dec->dif = 0;
+#endif
+  dec->rng = 0x8000;
+  dec->cnt = -15;
+  dec->error = 0;
+  od_ec_dec_refill(dec);
+}
+
+/*Decode a single binary value.
+  {EC_SMALLMUL} f: The probability that the bit is one, scaled by 32768.
+  {else} f: The probability that the bit is zero, scaled by 32768.
+  Return: The value decoded (0 or 1).*/
+int od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f) {
+  od_ec_window dif;
+  od_ec_window vw;
+  unsigned r;
+  unsigned r_new;
+  unsigned v;
+  int ret;
+  OD_ASSERT(0 < f);
+  OD_ASSERT(f < 32768U);
+  dif = dec->dif;
+  r = dec->rng;
+  OD_ASSERT(dif >> (OD_EC_WINDOW_SIZE - 16) < r);
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  v = (r >> 8) * (uint32_t)f >> 7;
+  vw = (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16);
+  ret = 1;
+  r_new = v;
+  if (dif >= vw) {
+    r_new = r - v;
+    dif -= vw;
+    ret = 0;
+  }
+#else
+  v = f * (uint32_t)r >> 15;
+  vw = (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16);
+  ret = 0;
+  r_new = v;
+  if (dif >= vw) {
+    r_new = r - v;
+    dif -= vw;
+    ret = 1;
+  }
+#endif
+  return od_ec_dec_normalize(dec, dif, r_new, ret);
+}
+
+/*Decodes a symbol given a cumulative distribution function (CDF) table in Q15.
+  cdf: The CDF, such that symbol s falls in the range
+        [s > 0 ? cdf[s - 1] : 0, cdf[s]).
+       The values must be monotonically non-increasing, and cdf[nsyms - 1]
+        must be 32768.
+       {EC_SMALLMUL}: The CDF contains 32768 minus those values.
+  nsyms: The number of symbols in the alphabet.
+         This should be at most 16.
+  Return: The decoded symbol s.*/
+int od_ec_decode_cdf_q15(od_ec_dec *dec, const uint16_t *cdf, int nsyms) {
+  od_ec_window dif;
+  unsigned r;
+  unsigned c;
+  unsigned u;
+  unsigned v;
+  int ret;
+  (void)nsyms;
+  dif = dec->dif;
+  r = dec->rng;
+  OD_ASSERT(dif >> (OD_EC_WINDOW_SIZE - 16) < r);
+  OD_ASSERT(cdf[nsyms - 1] == OD_ICDF(32768U));
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  c = (unsigned)(dif >> (OD_EC_WINDOW_SIZE - 16));
+  v = r;
+  ret = -1;
+  do {
+    u = v;
+    v = (r >> 8) * (uint32_t)cdf[++ret] >> 7;
+  } while (c < v);
+  OD_ASSERT(v < u);
+  OD_ASSERT(u <= r);
+  r = u - v;
+  dif -= (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16);
+#else
+  c = (unsigned)(dif >> (OD_EC_WINDOW_SIZE - 16));
+  v = 0;
+  ret = -1;
+  do {
+    u = v;
+    v = cdf[++ret] * (uint32_t)r >> 15;
+  } while (v <= c);
+  OD_ASSERT(u < v);
+  OD_ASSERT(v <= r);
+  r = v - u;
+  dif -= (od_ec_window)u << (OD_EC_WINDOW_SIZE - 16);
+#endif
+  return od_ec_dec_normalize(dec, dif, r, ret);
+}
+
+#if CONFIG_RAWBITS
+/*Extracts a sequence of raw bits from the stream.
+  The bits must have been encoded with od_ec_enc_bits().
+  ftb: The number of bits to extract.
+       This must be between 0 and 25, inclusive.
+  Return: The decoded bits.*/
+uint32_t od_ec_dec_bits_(od_ec_dec *dec, unsigned ftb) {
+  od_ec_window window;
+  int available;
+  uint32_t ret;
+  OD_ASSERT(ftb <= 25);
+  window = dec->end_window;
+  available = dec->nend_bits;
+  if ((unsigned)available < ftb) {
+    const unsigned char *buf;
+    const unsigned char *eptr;
+    buf = dec->buf;
+    eptr = dec->eptr;
+    OD_ASSERT(available <= OD_EC_WINDOW_SIZE - 8);
+    do {
+      if (eptr <= buf) {
+        dec->tell_offs += OD_EC_LOTS_OF_BITS - available;
+        available = OD_EC_LOTS_OF_BITS;
+        break;
+      }
+      window |= (od_ec_window) * --eptr << available;
+      available += 8;
+    } while (available <= OD_EC_WINDOW_SIZE - 8);
+    dec->eptr = eptr;
+  }
+  ret = (uint32_t)window & (((uint32_t)1 << ftb) - 1);
+  window >>= ftb;
+  available -= ftb;
+  dec->end_window = window;
+  dec->nend_bits = available;
+  return ret;
+}
+#endif
+
+/*Returns the number of bits "used" by the decoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Return: The number of bits.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+int od_ec_dec_tell(const od_ec_dec *dec) {
+  return (int)(((dec->end - dec->eptr) + (dec->bptr - dec->buf)) * 8 -
+               dec->cnt - dec->nend_bits + dec->tell_offs);
+}
+
+/*Returns the number of bits "used" by the decoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Return: The number of bits scaled by 2**OD_BITRES.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+uint32_t od_ec_dec_tell_frac(const od_ec_dec *dec) {
+  return od_ec_tell_frac(od_ec_dec_tell(dec), dec->rng);
+}

third_party/aom/aom_dsp/entdec.h

@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_entdec_H)
+#define _entdec_H (1)
+#include <limits.h>
+#include "aom_dsp/entcode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct od_ec_dec od_ec_dec;
+
+#if defined(OD_ACCOUNTING) && OD_ACCOUNTING
+#define OD_ACC_STR , char *acc_str
+#define od_ec_dec_bits(dec, ftb, str) od_ec_dec_bits_(dec, ftb, str)
+#else
+#define OD_ACC_STR
+#define od_ec_dec_bits(dec, ftb, str) od_ec_dec_bits_(dec, ftb)
+#endif
+
+/*The entropy decoder context.*/
+struct od_ec_dec {
+  /*The start of the current input buffer.*/
+  const unsigned char *buf;
+  /*The read pointer for the raw bits.*/
+  const unsigned char *eptr;
+  /*Bits that will be read from/written at the end.*/
+  od_ec_window end_window;
+  /*Number of valid bits in end_window.*/
+  int nend_bits;
+  /*An offset used to keep track of tell after reaching the end of the stream.
+    This is constant throughout most of the decoding process, but becomes
+     important once we hit the end of the buffer and stop incrementing pointers
+     (and instead pretend cnt/nend_bits have lots of bits).*/
+  int32_t tell_offs;
+  /*The end of the current input buffer.*/
+  const unsigned char *end;
+  /*The read pointer for the entropy-coded bits.*/
+  const unsigned char *bptr;
+  /*The difference between the coded value and the low end of the current
+     range.
+    {EC_SMALLMUL} The difference between the high end of the current range,
+     (low + rng), and the coded value, minus 1.
+    This stores up to OD_EC_WINDOW_SIZE bits of that difference, but the
+     decoder only uses the top 16 bits of the window to decode the next symbol.
+    As we shift up during renormalization, if we don't have enough bits left in
+     the window to fill the top 16, we'll read in more bits of the coded
+     value.*/
+  od_ec_window dif;
+  /*The number of values in the current range.*/
+  uint16_t rng;
+  /*The number of bits of data in the current value.*/
+  int16_t cnt;
+  /*Nonzero if an error occurred.*/
+  int error;
+};
+
+/*See entdec.c for further documentation.*/
+
+void od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf, uint32_t storage)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(2);
+
+OD_WARN_UNUSED_RESULT int od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT int od_ec_decode_cdf_q15(od_ec_dec *dec,
+                                               const uint16_t *cdf, int nsyms)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(2);
+
+OD_WARN_UNUSED_RESULT uint32_t od_ec_dec_bits_(od_ec_dec *dec, unsigned ftb)
+    OD_ARG_NONNULL(1);
+
+OD_WARN_UNUSED_RESULT int od_ec_dec_tell(const od_ec_dec *dec)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT uint32_t od_ec_dec_tell_frac(const od_ec_dec *dec)
+    OD_ARG_NONNULL(1);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif

third_party/aom/aom_dsp/entenc.c

@@ -0,0 +1,507 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "./config.h"
+#endif
+
+#include <stdlib.h>
+#include <string.h>
+#include "aom_dsp/entenc.h"
+
+/*A range encoder.
+  See entdec.c and the references for implementation details \cite{Mar79,MNW98}.
+
+  @INPROCEEDINGS{Mar79,
+   author="Martin, G.N.N.",
+   title="Range encoding: an algorithm for removing redundancy from a digitised
+    message",
+   booktitle="Video \& Data Recording Conference",
+   year=1979,
+   address="Southampton",
+   month=Jul,
+   URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz"
+  }
+  @ARTICLE{MNW98,
+   author="Alistair Moffat and Radford Neal and Ian H. Witten",
+   title="Arithmetic Coding Revisited",
+   journal="{ACM} Transactions on Information Systems",
+   year=1998,
+   volume=16,
+   number=3,
+   pages="256--294",
+   month=Jul,
+   URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf"
+  }*/
+
+/*Takes updated low and range values, renormalizes them so that
+   32768 <= rng < 65536 (flushing bytes from low to the pre-carry buffer if
+   necessary), and stores them back in the encoder context.
+  low: The new value of low.
+  rng: The new value of the range.*/
+static void od_ec_enc_normalize(od_ec_enc *enc, od_ec_window low,
+                                unsigned rng) {
+  int d;
+  int c;
+  int s;
+  c = enc->cnt;
+  OD_ASSERT(rng <= 65535U);
+  d = 16 - OD_ILOG_NZ(rng);
+  s = c + d;
+  /*TODO: Right now we flush every time we have at least one byte available.
+    Instead we should use an od_ec_window and flush right before we're about to
+     shift bits off the end of the window.
+    For a 32-bit window this is about the same amount of work, but for a 64-bit
+     window it should be a fair win.*/
+  if (s >= 0) {
+    uint16_t *buf;
+    uint32_t storage;
+    uint32_t offs;
+    unsigned m;
+    buf = enc->precarry_buf;
+    storage = enc->precarry_storage;
+    offs = enc->offs;
+    if (offs + 2 > storage) {
+      storage = 2 * storage + 2;
+      buf = (uint16_t *)realloc(buf, sizeof(*buf) * storage);
+      if (buf == NULL) {
+        enc->error = -1;
+        enc->offs = 0;
+        return;
+      }
+      enc->precarry_buf = buf;
+      enc->precarry_storage = storage;
+    }
+    c += 16;
+    m = (1 << c) - 1;
+    if (s >= 8) {
+      OD_ASSERT(offs < storage);
+      buf[offs++] = (uint16_t)(low >> c);
+      low &= m;
+      c -= 8;
+      m >>= 8;
+    }
+    OD_ASSERT(offs < storage);
+    buf[offs++] = (uint16_t)(low >> c);
+    s = c + d - 24;
+    low &= m;
+    enc->offs = offs;
+  }
+  enc->low = low << d;
+  enc->rng = rng << d;
+  enc->cnt = s;
+}
+
+/*Initializes the encoder.
+  size: The initial size of the buffer, in bytes.*/
+void od_ec_enc_init(od_ec_enc *enc, uint32_t size) {
+  od_ec_enc_reset(enc);
+  enc->buf = (unsigned char *)malloc(sizeof(*enc->buf) * size);
+  enc->storage = size;
+  if (size > 0 && enc->buf == NULL) {
+    enc->storage = 0;
+    enc->error = -1;
+  }
+  enc->precarry_buf = (uint16_t *)malloc(sizeof(*enc->precarry_buf) * size);
+  enc->precarry_storage = size;
+  if (size > 0 && enc->precarry_buf == NULL) {
+    enc->precarry_storage = 0;
+    enc->error = -1;
+  }
+}
+
+/*Reinitializes the encoder.*/
+void od_ec_enc_reset(od_ec_enc *enc) {
+  enc->end_offs = 0;
+  enc->end_window = 0;
+  enc->nend_bits = 0;
+  enc->offs = 0;
+  enc->low = 0;
+  enc->rng = 0x8000;
+  /*This is initialized to -9 so that it crosses zero after we've accumulated
+     one byte + one carry bit.*/
+  enc->cnt = -9;
+  enc->error = 0;
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy = 0;
+  enc->nb_symbols = 0;
+#endif
+}
+
+/*Frees the buffers used by the encoder.*/
+void od_ec_enc_clear(od_ec_enc *enc) {
+  free(enc->precarry_buf);
+  free(enc->buf);
+}
+
+/*Encodes a symbol given its frequency in Q15.
+  fl: The cumulative frequency of all symbols that come before the one to be
+       encoded.
+  fh: The cumulative frequency of all symbols up to and including the one to
+       be encoded.
+  {EC_SMALLMUL} Both values are 32768 minus that.*/
+static void od_ec_encode_q15(od_ec_enc *enc, unsigned fl, unsigned fh) {
+  od_ec_window l;
+  unsigned r;
+  unsigned u;
+  unsigned v;
+  l = enc->low;
+  r = enc->rng;
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  OD_ASSERT(fh < fl);
+  OD_ASSERT(fl <= 32768U);
+  if (fl < 32768U) {
+    u = (r >> 8) * (uint32_t)fl >> 7;
+    v = (r >> 8) * (uint32_t)fh >> 7;
+    l += r - u;
+    r = u - v;
+  } else {
+    r -= (r >> 8) * (uint32_t)fh >> 7;
+  }
+#else
+  OD_ASSERT(fl < fh);
+  OD_ASSERT(fh <= 32768U);
+  u = fl * (uint32_t)r >> 15;
+  v = fh * (uint32_t)r >> 15;
+  r = v - u;
+  l += u;
+#endif
+  od_ec_enc_normalize(enc, l, r);
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy -= OD_LOG2((double)(OD_ICDF(fh) - OD_ICDF(fl)) / 32768.);
+  enc->nb_symbols++;
+#endif
+}
+
+/*Encode a single binary value.
+  val: The value to encode (0 or 1).
+  {EC_SMALLMUL} f: The probability that the val is one, scaled by 32768.
+  {else} f: The probability that val is zero, scaled by 32768.*/
+void od_ec_encode_bool_q15(od_ec_enc *enc, int val, unsigned f) {
+  od_ec_window l;
+  unsigned r;
+  unsigned v;
+  OD_ASSERT(0 < f);
+  OD_ASSERT(f < 32768U);
+  l = enc->low;
+  r = enc->rng;
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  v = (r >> 8) * (uint32_t)f >> 7;
+  if (val) l += r - v;
+  r = val ? v : r - v;
+#else
+  v = f * (uint32_t)r >> 15;
+  if (val) l += v;
+  r = val ? r - v : v;
+#endif
+  od_ec_enc_normalize(enc, l, r);
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy -=
+      OD_LOG2((double)(val ? 32768 - OD_ICDF(f) : OD_ICDF(f)) / 32768.);
+  enc->nb_symbols++;
+#endif
+}
+
+/*Encodes a symbol given a cumulative distribution function (CDF) table in Q15.
+  s: The index of the symbol to encode.
+  cdf: The CDF, such that symbol s falls in the range
+        [s > 0 ? cdf[s - 1] : 0, cdf[s]).
+       The values must be monotonically non-decreasing, and the last value
+        must be exactly 32768.
+  nsyms: The number of symbols in the alphabet.
+         This should be at most 16.*/
+void od_ec_encode_cdf_q15(od_ec_enc *enc, int s, const uint16_t *cdf,
+                          int nsyms) {
+  (void)nsyms;
+  OD_ASSERT(s >= 0);
+  OD_ASSERT(s < nsyms);
+  OD_ASSERT(cdf[nsyms - 1] == OD_ICDF(32768U));
+  od_ec_encode_q15(enc, s > 0 ? cdf[s - 1] : OD_ICDF(0), cdf[s]);
+}
+
+#if CONFIG_RAWBITS
+/*Encodes a sequence of raw bits in the stream.
+  fl: The bits to encode.
+  ftb: The number of bits to encode.
+       This must be between 0 and 25, inclusive.*/
+void od_ec_enc_bits(od_ec_enc *enc, uint32_t fl, unsigned ftb) {
+  od_ec_window end_window;
+  int nend_bits;
+  OD_ASSERT(ftb <= 25);
+  OD_ASSERT(fl < (uint32_t)1 << ftb);
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy += ftb;
+#endif
+  end_window = enc->end_window;
+  nend_bits = enc->nend_bits;
+  if (nend_bits + ftb > OD_EC_WINDOW_SIZE) {
+    unsigned char *buf;
+    uint32_t storage;
+    uint32_t end_offs;
+    buf = enc->buf;
+    storage = enc->storage;
+    end_offs = enc->end_offs;
+    if (end_offs + (OD_EC_WINDOW_SIZE >> 3) >= storage) {
+      unsigned char *new_buf;
+      uint32_t new_storage;
+      new_storage = 2 * storage + (OD_EC_WINDOW_SIZE >> 3);
+      new_buf = (unsigned char *)malloc(sizeof(*new_buf) * new_storage);
+      if (new_buf == NULL) {
+        enc->error = -1;
+        enc->end_offs = 0;
+        return;
+      }
+      OD_COPY(new_buf + new_storage - end_offs, buf + storage - end_offs,
+              end_offs);
+      storage = new_storage;
+      free(buf);
+      enc->buf = buf = new_buf;
+      enc->storage = storage;
+    }
+    do {
+      OD_ASSERT(end_offs < storage);
+      buf[storage - ++end_offs] = (unsigned char)end_window;
+      end_window >>= 8;
+      nend_bits -= 8;
+    } while (nend_bits >= 8);
+    enc->end_offs = end_offs;
+  }
+  OD_ASSERT(nend_bits + ftb <= OD_EC_WINDOW_SIZE);
+  end_window |= (od_ec_window)fl << nend_bits;
+  nend_bits += ftb;
+  enc->end_window = end_window;
+  enc->nend_bits = nend_bits;
+}
+#endif
+
+/*Overwrites a few bits at the very start of an existing stream, after they
+   have already been encoded.
+  This makes it possible to have a few flags up front, where it is easy for
+   decoders to access them without parsing the whole stream, even if their
+   values are not determined until late in the encoding process, without having
+   to buffer all the intermediate symbols in the encoder.
+  In order for this to work, at least nbits bits must have already been encoded
+   using probabilities that are an exact power of two.
+  The encoder can verify the number of encoded bits is sufficient, but cannot
+   check this latter condition.
+  val: The bits to encode (in the least nbits significant bits).
+       They will be decoded in order from most-significant to least.
+  nbits: The number of bits to overwrite.
+         This must be no more than 8.*/
+void od_ec_enc_patch_initial_bits(od_ec_enc *enc, unsigned val, int nbits) {
+  int shift;
+  unsigned mask;
+  OD_ASSERT(nbits >= 0);
+  OD_ASSERT(nbits <= 8);
+  OD_ASSERT(val < 1U << nbits);
+  shift = 8 - nbits;
+  mask = ((1U << nbits) - 1) << shift;
+  if (enc->offs > 0) {
+    /*The first byte has been finalized.*/
+    enc->precarry_buf[0] =
+        (uint16_t)((enc->precarry_buf[0] & ~mask) | val << shift);
+  } else if (9 + enc->cnt + (enc->rng == 0x8000) > nbits) {
+    /*The first byte has yet to be output.*/
+    enc->low = (enc->low & ~((od_ec_window)mask << (16 + enc->cnt))) |
+               (od_ec_window)val << (16 + enc->cnt + shift);
+  } else {
+    /*The encoder hasn't even encoded _nbits of data yet.*/
+    enc->error = -1;
+  }
+}
+
+#if OD_MEASURE_EC_OVERHEAD
+#include <stdio.h>
+#endif
+
+/*Indicates that there are no more symbols to encode.
+  All remaining output bytes are flushed to the output buffer.
+  od_ec_enc_reset() should be called before using the encoder again.
+  bytes: Returns the size of the encoded data in the returned buffer.
+  Return: A pointer to the start of the final buffer, or NULL if there was an
+           encoding error.*/
+unsigned char *od_ec_enc_done(od_ec_enc *enc, uint32_t *nbytes) {
+  unsigned char *out;
+  uint32_t storage;
+  uint16_t *buf;
+  uint32_t offs;
+  uint32_t end_offs;
+  int nend_bits;
+  od_ec_window m;
+  od_ec_window e;
+  od_ec_window l;
+  unsigned r;
+  int c;
+  int s;
+  if (enc->error) return NULL;
+#if OD_MEASURE_EC_OVERHEAD
+  {
+    uint32_t tell;
+    /* Don't count the 1 bit we lose to raw bits as overhead. */
+    tell = od_ec_enc_tell(enc) - 1;
+    fprintf(stderr, "overhead: %f%%\n",
+            100 * (tell - enc->entropy) / enc->entropy);
+    fprintf(stderr, "efficiency: %f bits/symbol\n",
+            (double)tell / enc->nb_symbols);
+  }
+#endif
+  /*We output the minimum number of bits that ensures that the symbols encoded
+     thus far will be decoded correctly regardless of the bits that follow.*/
+  l = enc->low;
+  r = enc->rng;
+  c = enc->cnt;
+  s = 9;
+  m = 0x7FFF;
+  e = (l + m) & ~m;
+  while ((e | m) >= l + r) {
+    s++;
+    m >>= 1;
+    e = (l + m) & ~m;
+  }
+  s += c;
+  offs = enc->offs;
+  buf = enc->precarry_buf;
+  if (s > 0) {
+    unsigned n;
+    storage = enc->precarry_storage;
+    if (offs + ((s + 7) >> 3) > storage) {
+      storage = storage * 2 + ((s + 7) >> 3);
+      buf = (uint16_t *)realloc(buf, sizeof(*buf) * storage);
+      if (buf == NULL) {
+        enc->error = -1;
+        return NULL;
+      }
+      enc->precarry_buf = buf;
+      enc->precarry_storage = storage;
+    }
+    n = (1 << (c + 16)) - 1;
+    do {
+      OD_ASSERT(offs < storage);
+      buf[offs++] = (uint16_t)(e >> (c + 16));
+      e &= n;
+      s -= 8;
+      c -= 8;
+      n >>= 8;
+    } while (s > 0);
+  }
+  /*Make sure there's enough room for the entropy-coded bits and the raw
+     bits.*/
+  out = enc->buf;
+  storage = enc->storage;
+  end_offs = enc->end_offs;
+  e = enc->end_window;
+  nend_bits = enc->nend_bits;
+  s = -s;
+  c = OD_MAXI((nend_bits - s + 7) >> 3, 0);
+  if (offs + end_offs + c > storage) {
+    storage = offs + end_offs + c;
+    out = (unsigned char *)realloc(out, sizeof(*out) * storage);
+    if (out == NULL) {
+      enc->error = -1;
+      return NULL;
+    }
+    OD_MOVE(out + storage - end_offs, out + enc->storage - end_offs, end_offs);
+    enc->buf = out;
+    enc->storage = storage;
+  }
+  /*If we have buffered raw bits, flush them as well.*/
+  while (nend_bits > s) {
+    OD_ASSERT(end_offs < storage);
+    out[storage - ++end_offs] = (unsigned char)e;
+    e >>= 8;
+    nend_bits -= 8;
+  }
+  *nbytes = offs + end_offs;
+  /*Perform carry propagation.*/
+  OD_ASSERT(offs + end_offs <= storage);
+  out = out + storage - (offs + end_offs);
+  c = 0;
+  end_offs = offs;
+  while (offs > 0) {
+    offs--;
+    c = buf[offs] + c;
+    out[offs] = (unsigned char)c;
+    c >>= 8;
+  }
+  /*Add any remaining raw bits to the last byte.
+    There is guaranteed to be enough room, because nend_bits <= s.*/
+  OD_ASSERT(nend_bits <= 0 || end_offs > 0);
+  if (nend_bits > 0) out[end_offs - 1] |= (unsigned char)e;
+  /*Note: Unless there's an allocation error, if you keep encoding into the
+     current buffer and call this function again later, everything will work
+     just fine (you won't get a new packet out, but you will get a single
+     buffer with the new data appended to the old).
+    However, this function is O(N) where N is the amount of data coded so far,
+     so calling it more than once for a given packet is a bad idea.*/
+  return out;
+}
+
+/*Returns the number of bits "used" by the encoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Warning: The value returned by this function can decrease compared to an
+   earlier call, even after encoding more data, if there is an encoding error
+   (i.e., a failure to allocate enough space for the output buffer).
+  Return: The number of bits.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+int od_ec_enc_tell(const od_ec_enc *enc) {
+  /*The 10 here counteracts the offset of -9 baked into cnt, and adds 1 extra
+     bit, which we reserve for terminating the stream.*/
+  return (enc->offs + enc->end_offs) * 8 + enc->cnt + enc->nend_bits + 10;
+}
+
+/*Returns the number of bits "used" by the encoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Warning: The value returned by this function can decrease compared to an
+   earlier call, even after encoding more data, if there is an encoding error
+   (i.e., a failure to allocate enough space for the output buffer).
+  Return: The number of bits scaled by 2**OD_BITRES.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+uint32_t od_ec_enc_tell_frac(const od_ec_enc *enc) {
+  return od_ec_tell_frac(od_ec_enc_tell(enc), enc->rng);
+}
+
+/*Saves a entropy coder checkpoint to dst.
+  This allows an encoder to reverse a series of entropy coder
+   decisions if it decides that the information would have been
+   better coded some other way.*/
+void od_ec_enc_checkpoint(od_ec_enc *dst, const od_ec_enc *src) {
+  OD_COPY(dst, src, 1);
+}
+
+/*Restores an entropy coder checkpoint saved by od_ec_enc_checkpoint.
+  This can only be used to restore from checkpoints earlier in the target
+   state's history: you can not switch backwards and forwards or otherwise
+   switch to a state which isn't a casual ancestor of the current state.
+  Restore is also incompatible with patching the initial bits, as the
+   changes will remain in the restored version.*/
+void od_ec_enc_rollback(od_ec_enc *dst, const od_ec_enc *src) {
+  unsigned char *buf;
+  uint32_t storage;
+  uint16_t *precarry_buf;
+  uint32_t precarry_storage;
+  OD_ASSERT(dst->storage >= src->storage);
+  OD_ASSERT(dst->precarry_storage >= src->precarry_storage);
+  buf = dst->buf;
+  storage = dst->storage;
+  precarry_buf = dst->precarry_buf;
+  precarry_storage = dst->precarry_storage;
+  OD_COPY(dst, src, 1);
+  dst->buf = buf;
+  dst->storage = storage;
+  dst->precarry_buf = precarry_buf;
+  dst->precarry_storage = precarry_storage;
+}

third_party/aom/aom_dsp/entenc.h

@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_entenc_H)
+#define _entenc_H (1)
+#include <stddef.h>
+#include "aom_dsp/entcode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct od_ec_enc od_ec_enc;
+
+#define OD_MEASURE_EC_OVERHEAD (0)
+
+/*The entropy encoder context.*/
+struct od_ec_enc {
+  /*Buffered output.
+    This contains only the raw bits until the final call to od_ec_enc_done(),
+     where all the arithmetic-coded data gets prepended to it.*/
+  unsigned char *buf;
+  /*The size of the buffer.*/
+  uint32_t storage;
+  /*The offset at which the last byte containing raw bits was written.*/
+  uint32_t end_offs;
+  /*Bits that will be read from/written at the end.*/
+  od_ec_window end_window;
+  /*Number of valid bits in end_window.*/
+  int nend_bits;
+  /*A buffer for output bytes with their associated carry flags.*/
+  uint16_t *precarry_buf;
+  /*The size of the pre-carry buffer.*/
+  uint32_t precarry_storage;
+  /*The offset at which the next entropy-coded byte will be written.*/
+  uint32_t offs;
+  /*The low end of the current range.*/
+  od_ec_window low;
+  /*The number of values in the current range.*/
+  uint16_t rng;
+  /*The number of bits of data in the current value.*/
+  int16_t cnt;
+  /*Nonzero if an error occurred.*/
+  int error;
+#if OD_MEASURE_EC_OVERHEAD
+  double entropy;
+  int nb_symbols;
+#endif
+};
+
+/*See entenc.c for further documentation.*/
+
+void od_ec_enc_init(od_ec_enc *enc, uint32_t size) OD_ARG_NONNULL(1);
+void od_ec_enc_reset(od_ec_enc *enc) OD_ARG_NONNULL(1);
+void od_ec_enc_clear(od_ec_enc *enc) OD_ARG_NONNULL(1);
+
+void od_ec_encode_bool_q15(od_ec_enc *enc, int val, unsigned f_q15)
+    OD_ARG_NONNULL(1);
+void od_ec_encode_cdf_q15(od_ec_enc *enc, int s, const uint16_t *cdf, int nsyms)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(3);
+
+void od_ec_enc_bits(od_ec_enc *enc, uint32_t fl, unsigned ftb)
+    OD_ARG_NONNULL(1);
+
+void od_ec_enc_patch_initial_bits(od_ec_enc *enc, unsigned val, int nbits)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT unsigned char *od_ec_enc_done(od_ec_enc *enc,
+                                                    uint32_t *nbytes)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(2);
+
+OD_WARN_UNUSED_RESULT int od_ec_enc_tell(const od_ec_enc *enc)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT uint32_t od_ec_enc_tell_frac(const od_ec_enc *enc)
+    OD_ARG_NONNULL(1);
+
+void od_ec_enc_checkpoint(od_ec_enc *dst, const od_ec_enc *src);
+void od_ec_enc_rollback(od_ec_enc *dst, const od_ec_enc *src);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif

third_party/aom/aom_dsp/fastssim.c

@@ -0,0 +1,493 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ *
+ *  This code was originally written by: Nathan E. Egge, at the Daala
+ *  project.
+ */
+#include <assert.h>
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/ssim.h"
+#include "aom_ports/system_state.h"
+
+typedef struct fs_level fs_level;
+typedef struct fs_ctx fs_ctx;
+
+#define SSIM_C1 (255 * 255 * 0.01 * 0.01)
+#define SSIM_C2 (255 * 255 * 0.03 * 0.03)
+#if CONFIG_HIGHBITDEPTH
+#define SSIM_C1_10 (1023 * 1023 * 0.01 * 0.01)
+#define SSIM_C1_12 (4095 * 4095 * 0.01 * 0.01)
+#define SSIM_C2_10 (1023 * 1023 * 0.03 * 0.03)
+#define SSIM_C2_12 (4095 * 4095 * 0.03 * 0.03)
+#endif
+#define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
+#define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
+
+struct fs_level {
+  uint32_t *im1;
+  uint32_t *im2;
+  double *ssim;
+  int w;
+  int h;
+};
+
+struct fs_ctx {
+  fs_level *level;
+  int nlevels;
+  unsigned *col_buf;
+};
+
+static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) {
+  unsigned char *data;
+  size_t data_size;
+  int lw;
+  int lh;
+  int l;
+  lw = (_w + 1) >> 1;
+  lh = (_h + 1) >> 1;
+  data_size =
+      _nlevels * sizeof(fs_level) + 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf);
+  for (l = 0; l < _nlevels; l++) {
+    size_t im_size;
+    size_t level_size;
+    im_size = lw * (size_t)lh;
+    level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+    level_size += sizeof(*_ctx->level[l].ssim) - 1;
+    level_size /= sizeof(*_ctx->level[l].ssim);
+    level_size += im_size;
+    level_size *= sizeof(*_ctx->level[l].ssim);
+    data_size += level_size;
+    lw = (lw + 1) >> 1;
+    lh = (lh + 1) >> 1;
+  }
+  data = (unsigned char *)malloc(data_size);
+  _ctx->level = (fs_level *)data;
+  _ctx->nlevels = _nlevels;
+  data += _nlevels * sizeof(*_ctx->level);
+  lw = (_w + 1) >> 1;
+  lh = (_h + 1) >> 1;
+  for (l = 0; l < _nlevels; l++) {
+    size_t im_size;
+    size_t level_size;
+    _ctx->level[l].w = lw;
+    _ctx->level[l].h = lh;
+    im_size = lw * (size_t)lh;
+    level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+    level_size += sizeof(*_ctx->level[l].ssim) - 1;
+    level_size /= sizeof(*_ctx->level[l].ssim);
+    level_size *= sizeof(*_ctx->level[l].ssim);
+    _ctx->level[l].im1 = (uint32_t *)data;
+    _ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
+    data += level_size;
+    _ctx->level[l].ssim = (double *)data;
+    data += im_size * sizeof(*_ctx->level[l].ssim);
+    lw = (lw + 1) >> 1;
+    lh = (lh + 1) >> 1;
+  }
+  _ctx->col_buf = (unsigned *)data;
+}
+
+static void fs_ctx_clear(fs_ctx *_ctx) { free(_ctx->level); }
+
+static void fs_downsample_level(fs_ctx *_ctx, int _l) {
+  const uint32_t *src1;
+  const uint32_t *src2;
+  uint32_t *dst1;
+  uint32_t *dst2;
+  int w2;
+  int h2;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  dst1 = _ctx->level[_l].im1;
+  dst2 = _ctx->level[_l].im2;
+  w2 = _ctx->level[_l - 1].w;
+  h2 = _ctx->level[_l - 1].h;
+  src1 = _ctx->level[_l - 1].im1;
+  src2 = _ctx->level[_l - 1].im2;
+  for (j = 0; j < h; j++) {
+    int j0offs;
+    int j1offs;
+    j0offs = 2 * j * w2;
+    j1offs = FS_MINI(2 * j + 1, h2) * w2;
+    for (i = 0; i < w; i++) {
+      int i0;
+      int i1;
+      i0 = 2 * i;
+      i1 = FS_MINI(i0 + 1, w2);
+      dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1] +
+                        src1[j1offs + i0] + src1[j1offs + i1];
+      dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1] +
+                        src2[j1offs + i0] + src2[j1offs + i1];
+    }
+  }
+}
+
+static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
+                                 int _s1ystride, const uint8_t *_src2,
+                                 int _s2ystride, int _w, int _h, uint32_t bd,
+                                 uint32_t shift) {
+  uint32_t *dst1;
+  uint32_t *dst2;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[0].w;
+  h = _ctx->level[0].h;
+  dst1 = _ctx->level[0].im1;
+  dst2 = _ctx->level[0].im2;
+  for (j = 0; j < h; j++) {
+    int j0;
+    int j1;
+    j0 = 2 * j;
+    j1 = FS_MINI(j0 + 1, _h);
+    for (i = 0; i < w; i++) {
+      int i0;
+      int i1;
+      i0 = 2 * i;
+      i1 = FS_MINI(i0 + 1, _w);
+      if (bd == 8 && shift == 0) {
+        dst1[j * w + i] =
+            _src1[j0 * _s1ystride + i0] + _src1[j0 * _s1ystride + i1] +
+            _src1[j1 * _s1ystride + i0] + _src1[j1 * _s1ystride + i1];
+        dst2[j * w + i] =
+            _src2[j0 * _s2ystride + i0] + _src2[j0 * _s2ystride + i1] +
+            _src2[j1 * _s2ystride + i0] + _src2[j1 * _s2ystride + i1];
+      } else {
+        uint16_t *src1s = CONVERT_TO_SHORTPTR(_src1);
+        uint16_t *src2s = CONVERT_TO_SHORTPTR(_src2);
+        dst1[j * w + i] = (src1s[j0 * _s1ystride + i0] >> shift) +
+                          (src1s[j0 * _s1ystride + i1] >> shift) +
+                          (src1s[j1 * _s1ystride + i0] >> shift) +
+                          (src1s[j1 * _s1ystride + i1] >> shift);
+        dst2[j * w + i] = (src2s[j0 * _s2ystride + i0] >> shift) +
+                          (src2s[j0 * _s2ystride + i1] >> shift) +
+                          (src2s[j1 * _s2ystride + i0] >> shift) +
+                          (src2s[j1 * _s2ystride + i1] >> shift);
+      }
+    }
+  }
+}
+
+static void fs_apply_luminance(fs_ctx *_ctx, int _l, int bit_depth) {
+  unsigned *col_sums_x;
+  unsigned *col_sums_y;
+  uint32_t *im1;
+  uint32_t *im2;
+  double *ssim;
+  double c1;
+  int w;
+  int h;
+  int j0offs;
+  int j1offs;
+  int i;
+  int j;
+  double ssim_c1 = SSIM_C1;
+#if CONFIG_HIGHBITDEPTH
+  if (bit_depth == 10) ssim_c1 = SSIM_C1_10;
+  if (bit_depth == 12) ssim_c1 = SSIM_C1_12;
+#else
+  assert(bit_depth == 8);
+  (void)bit_depth;
+#endif
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  col_sums_x = _ctx->col_buf;
+  col_sums_y = col_sums_x + w;
+  im1 = _ctx->level[_l].im1;
+  im2 = _ctx->level[_l].im2;
+  for (i = 0; i < w; i++) col_sums_x[i] = 5 * im1[i];
+  for (i = 0; i < w; i++) col_sums_y[i] = 5 * im2[i];
+  for (j = 1; j < 4; j++) {
+    j1offs = FS_MINI(j, h - 1) * w;
+    for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
+    for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
+  }
+  ssim = _ctx->level[_l].ssim;
+  c1 = (double)(ssim_c1 * 4096 * (1 << 4 * _l));
+  for (j = 0; j < h; j++) {
+    unsigned mux;
+    unsigned muy;
+    int i0;
+    int i1;
+    mux = 5 * col_sums_x[0];
+    muy = 5 * col_sums_y[0];
+    for (i = 1; i < 4; i++) {
+      i1 = FS_MINI(i, w - 1);
+      mux += col_sums_x[i1];
+      muy += col_sums_y[i1];
+    }
+    for (i = 0; i < w; i++) {
+      ssim[j * w + i] *= (2 * mux * (double)muy + c1) /
+                         (mux * (double)mux + muy * (double)muy + c1);
+      if (i + 1 < w) {
+        i0 = FS_MAXI(0, i - 4);
+        i1 = FS_MINI(i + 4, w - 1);
+        mux += col_sums_x[i1] - col_sums_x[i0];
+        muy += col_sums_x[i1] - col_sums_x[i0];
+      }
+    }
+    if (j + 1 < h) {
+      j0offs = FS_MAXI(0, j - 4) * w;
+      for (i = 0; i < w; i++) col_sums_x[i] -= im1[j0offs + i];
+      for (i = 0; i < w; i++) col_sums_y[i] -= im2[j0offs + i];
+      j1offs = FS_MINI(j + 4, h - 1) * w;
+      for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
+      for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
+    }
+  }
+}
+
+#define FS_COL_SET(_col, _joffs, _ioffs)                       \
+  do {                                                         \
+    unsigned gx;                                               \
+    unsigned gy;                                               \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] = gx * (double)gx;                    \
+    col_sums_gy2[(_col)] = gy * (double)gy;                    \
+    col_sums_gxgy[(_col)] = gx * (double)gy;                   \
+  } while (0)
+
+#define FS_COL_ADD(_col, _joffs, _ioffs)                       \
+  do {                                                         \
+    unsigned gx;                                               \
+    unsigned gy;                                               \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] += gx * (double)gx;                   \
+    col_sums_gy2[(_col)] += gy * (double)gy;                   \
+    col_sums_gxgy[(_col)] += gx * (double)gy;                  \
+  } while (0)
+
+#define FS_COL_SUB(_col, _joffs, _ioffs)                       \
+  do {                                                         \
+    unsigned gx;                                               \
+    unsigned gy;                                               \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] -= gx * (double)gx;                   \
+    col_sums_gy2[(_col)] -= gy * (double)gy;                   \
+    col_sums_gxgy[(_col)] -= gx * (double)gy;                  \
+  } while (0)
+
+#define FS_COL_COPY(_col1, _col2)                    \
+  do {                                               \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)];   \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)];   \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
+  } while (0)
+
+#define FS_COL_HALVE(_col1, _col2)                         \
+  do {                                                     \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5;   \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5;   \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
+  } while (0)
+
+#define FS_COL_DOUBLE(_col1, _col2)                      \
+  do {                                                   \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2;   \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2;   \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
+  } while (0)
+
+static void fs_calc_structure(fs_ctx *_ctx, int _l, int bit_depth) {
+  uint32_t *im1;
+  uint32_t *im2;
+  unsigned *gx_buf;
+  unsigned *gy_buf;
+  double *ssim;
+  double col_sums_gx2[8];
+  double col_sums_gy2[8];
+  double col_sums_gxgy[8];
+  double c2;
+  int stride;
+  int w;
+  int h;
+  int i;
+  int j;
+  double ssim_c2 = SSIM_C2;
+#if CONFIG_HIGHBITDEPTH
+  if (bit_depth == 10) ssim_c2 = SSIM_C2_10;
+  if (bit_depth == 12) ssim_c2 = SSIM_C2_12;
+#else
+  assert(bit_depth == 8);
+  (void)bit_depth;
+#endif
+
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  im1 = _ctx->level[_l].im1;
+  im2 = _ctx->level[_l].im2;
+  ssim = _ctx->level[_l].ssim;
+  gx_buf = _ctx->col_buf;
+  stride = w + 8;
+  gy_buf = gx_buf + 8 * stride;
+  memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
+  c2 = ssim_c2 * (1 << 4 * _l) * 16 * 104;
+  for (j = 0; j < h + 4; j++) {
+    if (j < h - 1) {
+      for (i = 0; i < w - 1; i++) {
+        unsigned g1;
+        unsigned g2;
+        unsigned gx;
+        unsigned gy;
+        g1 = abs((int)im1[(j + 1) * w + i + 1] - (int)im1[j * w + i]);
+        g2 = abs((int)im1[(j + 1) * w + i] - (int)im1[j * w + i + 1]);
+        gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+        g1 = abs((int)im2[(j + 1) * w + i + 1] - (int)im2[j * w + i]);
+        g2 = abs((int)im2[(j + 1) * w + i] - (int)im2[j * w + i + 1]);
+        gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+        gx_buf[(j & 7) * stride + i + 4] = gx;
+        gy_buf[(j & 7) * stride + i + 4] = gy;
+      }
+    } else {
+      memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
+      memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
+    }
+    if (j >= 4) {
+      int k;
+      col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
+      col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
+      col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
+          col_sums_gxgy[0] = 0;
+      for (i = 4; i < 8; i++) {
+        FS_COL_SET(i, -1, 0);
+        FS_COL_ADD(i, 0, 0);
+        for (k = 1; k < 8 - i; k++) {
+          FS_COL_DOUBLE(i, i);
+          FS_COL_ADD(i, -k - 1, 0);
+          FS_COL_ADD(i, k, 0);
+        }
+      }
+      for (i = 0; i < w; i++) {
+        double mugx2;
+        double mugy2;
+        double mugxgy;
+        mugx2 = col_sums_gx2[0];
+        for (k = 1; k < 8; k++) mugx2 += col_sums_gx2[k];
+        mugy2 = col_sums_gy2[0];
+        for (k = 1; k < 8; k++) mugy2 += col_sums_gy2[k];
+        mugxgy = col_sums_gxgy[0];
+        for (k = 1; k < 8; k++) mugxgy += col_sums_gxgy[k];
+        ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
+        if (i + 1 < w) {
+          FS_COL_SET(0, -1, 1);
+          FS_COL_ADD(0, 0, 1);
+          FS_COL_SUB(2, -3, 2);
+          FS_COL_SUB(2, 2, 2);
+          FS_COL_HALVE(1, 2);
+          FS_COL_SUB(3, -4, 3);
+          FS_COL_SUB(3, 3, 3);
+          FS_COL_HALVE(2, 3);
+          FS_COL_COPY(3, 4);
+          FS_COL_DOUBLE(4, 5);
+          FS_COL_ADD(4, -4, 5);
+          FS_COL_ADD(4, 3, 5);
+          FS_COL_DOUBLE(5, 6);
+          FS_COL_ADD(5, -3, 6);
+          FS_COL_ADD(5, 2, 6);
+          FS_COL_DOUBLE(6, 7);
+          FS_COL_ADD(6, -2, 7);
+          FS_COL_ADD(6, 1, 7);
+          FS_COL_SET(7, -1, 8);
+          FS_COL_ADD(7, 0, 8);
+        }
+      }
+    }
+  }
+}
+
+#define FS_NLEVELS (4)
+
+/*These weights were derived from the default weights found in Wang's original
+ Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}.
+ We drop the finest scale and renormalize the rest to sum to 1.*/
+
+static const double FS_WEIGHTS[FS_NLEVELS] = {
+  0.2989654541015625, 0.3141326904296875, 0.2473602294921875, 0.1395416259765625
+};
+
+static double fs_average(fs_ctx *_ctx, int _l) {
+  double *ssim;
+  double ret;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  ssim = _ctx->level[_l].ssim;
+  ret = 0;
+  for (j = 0; j < h; j++)
+    for (i = 0; i < w; i++) ret += ssim[j * w + i];
+  return pow(ret / (w * h), FS_WEIGHTS[_l]);
+}
+
+static double convert_ssim_db(double _ssim, double _weight) {
+  assert(_weight >= _ssim);
+  if ((_weight - _ssim) < 1e-10) return MAX_SSIM_DB;
+  return 10 * (log10(_weight) - log10(_weight - _ssim));
+}
+
+static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst,
+                        int _dystride, int _w, int _h, uint32_t _bd,
+                        uint32_t _shift) {
+  fs_ctx ctx;
+  double ret;
+  int l;
+  ret = 1;
+  fs_ctx_init(&ctx, _w, _h, FS_NLEVELS);
+  fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h, _bd,
+                       _shift);
+  for (l = 0; l < FS_NLEVELS - 1; l++) {
+    fs_calc_structure(&ctx, l, _bd);
+    ret *= fs_average(&ctx, l);
+    fs_downsample_level(&ctx, l + 1);
+  }
+  fs_calc_structure(&ctx, l, _bd);
+  fs_apply_luminance(&ctx, l, _bd);
+  ret *= fs_average(&ctx, l);
+  fs_ctx_clear(&ctx);
+  return ret;
+}
+
+double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source,
+                         const YV12_BUFFER_CONFIG *dest, double *ssim_y,
+                         double *ssim_u, double *ssim_v, uint32_t bd,
+                         uint32_t in_bd) {
+  double ssimv;
+  uint32_t bd_shift = 0;
+  aom_clear_system_state();
+  assert(bd >= in_bd);
+
+  bd_shift = bd - in_bd;
+
+  *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
+                      dest->y_stride, source->y_crop_width,
+                      source->y_crop_height, in_bd, bd_shift);
+  *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height, in_bd, bd_shift);
+  *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height, in_bd, bd_shift);
+  ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
+  return convert_ssim_db(ssimv, 1.0);
+}

third_party/aom/aom_dsp/fwd_txfm.c

@@ -0,0 +1,809 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/fwd_txfm.h"
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  tran_low_t intermediate[4 * 4];
+  const tran_low_t *in_low = NULL;
+  tran_low_t *out = intermediate;
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t in_high[4];    // canbe16
+    tran_high_t step[4];       // canbe16
+    tran_high_t temp1, temp2;  // needs32
+    int i;
+    for (i = 0; i < 4; ++i) {
+      // Load inputs.
+      if (pass == 0) {
+        in_high[0] = input[0 * stride] * 16;
+        in_high[1] = input[1 * stride] * 16;
+        in_high[2] = input[2 * stride] * 16;
+        in_high[3] = input[3 * stride] * 16;
+        if (i == 0 && in_high[0]) {
+          ++in_high[0];
+        }
+      } else {
+        assert(in_low != NULL);
+        in_high[0] = in_low[0 * 4];
+        in_high[1] = in_low[1 * 4];
+        in_high[2] = in_low[2 * 4];
+        in_high[3] = in_low[3 * 4];
+        ++in_low;
+      }
+      // Transform.
+      step[0] = in_high[0] + in_high[3];
+      step[1] = in_high[1] + in_high[2];
+      step[2] = in_high[1] - in_high[2];
+      step[3] = in_high[0] - in_high[3];
+      temp1 = (step[0] + step[1]) * cospi_16_64;
+      temp2 = (step[0] - step[1]) * cospi_16_64;
+      out[0] = (tran_low_t)fdct_round_shift(temp1);
+      out[2] = (tran_low_t)fdct_round_shift(temp2);
+      temp1 = step[2] * cospi_24_64 + step[3] * cospi_8_64;
+      temp2 = -step[2] * cospi_8_64 + step[3] * cospi_24_64;
+      out[1] = (tran_low_t)fdct_round_shift(temp1);
+      out[3] = (tran_low_t)fdct_round_shift(temp2);
+      // Do next column (which is a transposed row in second/horizontal pass)
+      ++input;
+      out += 4;
+    }
+    // Setup in/out for next pass.
+    in_low = intermediate;
+    out = output;
+  }
+
+  {
+    int i, j;
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j) output[j + i * 4] = (output[j + i * 4] + 1) >> 2;
+    }
+  }
+}
+
+void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  tran_low_t sum = 0;
+  for (r = 0; r < 4; ++r)
+    for (c = 0; c < 4; ++c) sum += input[r * stride + c];
+
+  output[0] = sum << 1;
+}
+
+void aom_fdct8x8_c(const int16_t *input, tran_low_t *final_output, int stride) {
+  int i, j;
+  tran_low_t intermediate[64];
+  int pass;
+  tran_low_t *output = intermediate;
+  const tran_low_t *in = NULL;
+
+  // Transform columns
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+    tran_high_t t0, t1, t2, t3;                  // needs32
+    tran_high_t x0, x1, x2, x3;                  // canbe16
+
+    for (i = 0; i < 8; i++) {
+      // stage 1
+      if (pass == 0) {
+        s0 = (input[0 * stride] + input[7 * stride]) * 4;
+        s1 = (input[1 * stride] + input[6 * stride]) * 4;
+        s2 = (input[2 * stride] + input[5 * stride]) * 4;
+        s3 = (input[3 * stride] + input[4 * stride]) * 4;
+        s4 = (input[3 * stride] - input[4 * stride]) * 4;
+        s5 = (input[2 * stride] - input[5 * stride]) * 4;
+        s6 = (input[1 * stride] - input[6 * stride]) * 4;
+        s7 = (input[0 * stride] - input[7 * stride]) * 4;
+        ++input;
+      } else {
+        s0 = in[0 * 8] + in[7 * 8];
+        s1 = in[1 * 8] + in[6 * 8];
+        s2 = in[2 * 8] + in[5 * 8];
+        s3 = in[3 * 8] + in[4 * 8];
+        s4 = in[3 * 8] - in[4 * 8];
+        s5 = in[2 * 8] - in[5 * 8];
+        s6 = in[1 * 8] - in[6 * 8];
+        s7 = in[0 * 8] - in[7 * 8];
+        ++in;
+      }
+
+      // fdct4(step, step);
+      x0 = s0 + s3;
+      x1 = s1 + s2;
+      x2 = s1 - s2;
+      x3 = s0 - s3;
+      t0 = (x0 + x1) * cospi_16_64;
+      t1 = (x0 - x1) * cospi_16_64;
+      t2 = x2 * cospi_24_64 + x3 * cospi_8_64;
+      t3 = -x2 * cospi_8_64 + x3 * cospi_24_64;
+      output[0] = (tran_low_t)fdct_round_shift(t0);
+      output[2] = (tran_low_t)fdct_round_shift(t2);
+      output[4] = (tran_low_t)fdct_round_shift(t1);
+      output[6] = (tran_low_t)fdct_round_shift(t3);
+
+      // Stage 2
+      t0 = (s6 - s5) * cospi_16_64;
+      t1 = (s6 + s5) * cospi_16_64;
+      t2 = fdct_round_shift(t0);
+      t3 = fdct_round_shift(t1);
+
+      // Stage 3
+      x0 = s4 + t2;
+      x1 = s4 - t2;
+      x2 = s7 - t3;
+      x3 = s7 + t3;
+
+      // Stage 4
+      t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
+      t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
+      t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+      t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
+      output[1] = (tran_low_t)fdct_round_shift(t0);
+      output[3] = (tran_low_t)fdct_round_shift(t2);
+      output[5] = (tran_low_t)fdct_round_shift(t1);
+      output[7] = (tran_low_t)fdct_round_shift(t3);
+      output += 8;
+    }
+    in = intermediate;
+    output = final_output;
+  }
+
+  // Rows
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j) final_output[j + i * 8] /= 2;
+  }
+}
+
+void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  tran_low_t sum = 0;
+  for (r = 0; r < 8; ++r)
+    for (c = 0; c < 8; ++c) sum += input[r * stride + c];
+
+  output[0] = sum;
+}
+
+void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  tran_low_t intermediate[256];
+  const tran_low_t *in_low = NULL;
+  tran_low_t *out = intermediate;
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t step1[8];      // canbe16
+    tran_high_t step2[8];      // canbe16
+    tran_high_t step3[8];      // canbe16
+    tran_high_t in_high[8];    // canbe16
+    tran_high_t temp1, temp2;  // needs32
+    int i;
+    for (i = 0; i < 16; i++) {
+      if (0 == pass) {
+        // Calculate input for the first 8 results.
+        in_high[0] = (input[0 * stride] + input[15 * stride]) * 4;
+        in_high[1] = (input[1 * stride] + input[14 * stride]) * 4;
+        in_high[2] = (input[2 * stride] + input[13 * stride]) * 4;
+        in_high[3] = (input[3 * stride] + input[12 * stride]) * 4;
+        in_high[4] = (input[4 * stride] + input[11 * stride]) * 4;
+        in_high[5] = (input[5 * stride] + input[10 * stride]) * 4;
+        in_high[6] = (input[6 * stride] + input[9 * stride]) * 4;
+        in_high[7] = (input[7 * stride] + input[8 * stride]) * 4;
+        // Calculate input for the next 8 results.
+        step1[0] = (input[7 * stride] - input[8 * stride]) * 4;
+        step1[1] = (input[6 * stride] - input[9 * stride]) * 4;
+        step1[2] = (input[5 * stride] - input[10 * stride]) * 4;
+        step1[3] = (input[4 * stride] - input[11 * stride]) * 4;
+        step1[4] = (input[3 * stride] - input[12 * stride]) * 4;
+        step1[5] = (input[2 * stride] - input[13 * stride]) * 4;
+        step1[6] = (input[1 * stride] - input[14 * stride]) * 4;
+        step1[7] = (input[0 * stride] - input[15 * stride]) * 4;
+      } else {
+        // Calculate input for the first 8 results.
+        assert(in_low != NULL);
+        in_high[0] = ((in_low[0 * 16] + 1) >> 2) + ((in_low[15 * 16] + 1) >> 2);
+        in_high[1] = ((in_low[1 * 16] + 1) >> 2) + ((in_low[14 * 16] + 1) >> 2);
+        in_high[2] = ((in_low[2 * 16] + 1) >> 2) + ((in_low[13 * 16] + 1) >> 2);
+        in_high[3] = ((in_low[3 * 16] + 1) >> 2) + ((in_low[12 * 16] + 1) >> 2);
+        in_high[4] = ((in_low[4 * 16] + 1) >> 2) + ((in_low[11 * 16] + 1) >> 2);
+        in_high[5] = ((in_low[5 * 16] + 1) >> 2) + ((in_low[10 * 16] + 1) >> 2);
+        in_high[6] = ((in_low[6 * 16] + 1) >> 2) + ((in_low[9 * 16] + 1) >> 2);
+        in_high[7] = ((in_low[7 * 16] + 1) >> 2) + ((in_low[8 * 16] + 1) >> 2);
+        // Calculate input for the next 8 results.
+        step1[0] = ((in_low[7 * 16] + 1) >> 2) - ((in_low[8 * 16] + 1) >> 2);
+        step1[1] = ((in_low[6 * 16] + 1) >> 2) - ((in_low[9 * 16] + 1) >> 2);
+        step1[2] = ((in_low[5 * 16] + 1) >> 2) - ((in_low[10 * 16] + 1) >> 2);
+        step1[3] = ((in_low[4 * 16] + 1) >> 2) - ((in_low[11 * 16] + 1) >> 2);
+        step1[4] = ((in_low[3 * 16] + 1) >> 2) - ((in_low[12 * 16] + 1) >> 2);
+        step1[5] = ((in_low[2 * 16] + 1) >> 2) - ((in_low[13 * 16] + 1) >> 2);
+        step1[6] = ((in_low[1 * 16] + 1) >> 2) - ((in_low[14 * 16] + 1) >> 2);
+        step1[7] = ((in_low[0 * 16] + 1) >> 2) - ((in_low[15 * 16] + 1) >> 2);
+        in_low++;
+      }
+      // Work on the first eight values; fdct8(input, even_results);
+      {
+        tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+        tran_high_t t0, t1, t2, t3;                  // needs32
+        tran_high_t x0, x1, x2, x3;                  // canbe16
+
+        // stage 1
+        s0 = in_high[0] + in_high[7];
+        s1 = in_high[1] + in_high[6];
+        s2 = in_high[2] + in_high[5];
+        s3 = in_high[3] + in_high[4];
+        s4 = in_high[3] - in_high[4];
+        s5 = in_high[2] - in_high[5];
+        s6 = in_high[1] - in_high[6];
+        s7 = in_high[0] - in_high[7];
+
+        // fdct4(step, step);
+        x0 = s0 + s3;
+        x1 = s1 + s2;
+        x2 = s1 - s2;
+        x3 = s0 - s3;
+        t0 = (x0 + x1) * cospi_16_64;
+        t1 = (x0 - x1) * cospi_16_64;
+        t2 = x3 * cospi_8_64 + x2 * cospi_24_64;
+        t3 = x3 * cospi_24_64 - x2 * cospi_8_64;
+        out[0] = (tran_low_t)fdct_round_shift(t0);
+        out[4] = (tran_low_t)fdct_round_shift(t2);
+        out[8] = (tran_low_t)fdct_round_shift(t1);
+        out[12] = (tran_low_t)fdct_round_shift(t3);
+
+        // Stage 2
+        t0 = (s6 - s5) * cospi_16_64;
+        t1 = (s6 + s5) * cospi_16_64;
+        t2 = fdct_round_shift(t0);
+        t3 = fdct_round_shift(t1);
+
+        // Stage 3
+        x0 = s4 + t2;
+        x1 = s4 - t2;
+        x2 = s7 - t3;
+        x3 = s7 + t3;
+
+        // Stage 4
+        t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
+        t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
+        t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+        t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
+        out[2] = (tran_low_t)fdct_round_shift(t0);
+        out[6] = (tran_low_t)fdct_round_shift(t2);
+        out[10] = (tran_low_t)fdct_round_shift(t1);
+        out[14] = (tran_low_t)fdct_round_shift(t3);
+      }
+      // Work on the next eight values; step1 -> odd_results
+      {
+        // step 2
+        temp1 = (step1[5] - step1[2]) * cospi_16_64;
+        temp2 = (step1[4] - step1[3]) * cospi_16_64;
+        step2[2] = fdct_round_shift(temp1);
+        step2[3] = fdct_round_shift(temp2);
+        temp1 = (step1[4] + step1[3]) * cospi_16_64;
+        temp2 = (step1[5] + step1[2]) * cospi_16_64;
+        step2[4] = fdct_round_shift(temp1);
+        step2[5] = fdct_round_shift(temp2);
+        // step 3
+        step3[0] = step1[0] + step2[3];
+        step3[1] = step1[1] + step2[2];
+        step3[2] = step1[1] - step2[2];
+        step3[3] = step1[0] - step2[3];
+        step3[4] = step1[7] - step2[4];
+        step3[5] = step1[6] - step2[5];
+        step3[6] = step1[6] + step2[5];
+        step3[7] = step1[7] + step2[4];
+        // step 4
+        temp1 = step3[1] * -cospi_8_64 + step3[6] * cospi_24_64;
+        temp2 = step3[2] * cospi_24_64 + step3[5] * cospi_8_64;
+        step2[1] = fdct_round_shift(temp1);
+        step2[2] = fdct_round_shift(temp2);
+        temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
+        temp2 = step3[1] * cospi_24_64 + step3[6] * cospi_8_64;
+        step2[5] = fdct_round_shift(temp1);
+        step2[6] = fdct_round_shift(temp2);
+        // step 5
+        step1[0] = step3[0] + step2[1];
+        step1[1] = step3[0] - step2[1];
+        step1[2] = step3[3] + step2[2];
+        step1[3] = step3[3] - step2[2];
+        step1[4] = step3[4] - step2[5];
+        step1[5] = step3[4] + step2[5];
+        step1[6] = step3[7] - step2[6];
+        step1[7] = step3[7] + step2[6];
+        // step 6
+        temp1 = step1[0] * cospi_30_64 + step1[7] * cospi_2_64;
+        temp2 = step1[1] * cospi_14_64 + step1[6] * cospi_18_64;
+        out[1] = (tran_low_t)fdct_round_shift(temp1);
+        out[9] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[2] * cospi_22_64 + step1[5] * cospi_10_64;
+        temp2 = step1[3] * cospi_6_64 + step1[4] * cospi_26_64;
+        out[5] = (tran_low_t)fdct_round_shift(temp1);
+        out[13] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[3] * -cospi_26_64 + step1[4] * cospi_6_64;
+        temp2 = step1[2] * -cospi_10_64 + step1[5] * cospi_22_64;
+        out[3] = (tran_low_t)fdct_round_shift(temp1);
+        out[11] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[1] * -cospi_18_64 + step1[6] * cospi_14_64;
+        temp2 = step1[0] * -cospi_2_64 + step1[7] * cospi_30_64;
+        out[7] = (tran_low_t)fdct_round_shift(temp1);
+        out[15] = (tran_low_t)fdct_round_shift(temp2);
+      }
+      // Do next column (which is a transposed row in second/horizontal pass)
+      input++;
+      out += 16;
+    }
+    // Setup in/out for next pass.
+    in_low = intermediate;
+    out = output;
+  }
+}
+
+void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  int sum = 0;
+  for (r = 0; r < 16; ++r)
+    for (c = 0; c < 16; ++c) sum += input[r * stride + c];
+
+  output[0] = (tran_low_t)(sum >> 1);
+}
+
+static INLINE tran_high_t dct_32_round(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  // TODO(debargha, peter.derivaz): Find new bounds for this assert,
+  // and make the bounds consts.
+  // assert(-131072 <= rv && rv <= 131071);
+  return rv;
+}
+
+static INLINE tran_high_t half_round_shift(tran_high_t input) {
+  tran_high_t rv = (input + 1 + (input < 0)) >> 2;
+  return rv;
+}
+
+void aom_fdct32(const tran_high_t *input, tran_high_t *output, int round) {
+  tran_high_t step[32];
+  // Stage 1
+  step[0] = input[0] + input[(32 - 1)];
+  step[1] = input[1] + input[(32 - 2)];
+  step[2] = input[2] + input[(32 - 3)];
+  step[3] = input[3] + input[(32 - 4)];
+  step[4] = input[4] + input[(32 - 5)];
+  step[5] = input[5] + input[(32 - 6)];
+  step[6] = input[6] + input[(32 - 7)];
+  step[7] = input[7] + input[(32 - 8)];
+  step[8] = input[8] + input[(32 - 9)];
+  step[9] = input[9] + input[(32 - 10)];
+  step[10] = input[10] + input[(32 - 11)];
+  step[11] = input[11] + input[(32 - 12)];
+  step[12] = input[12] + input[(32 - 13)];
+  step[13] = input[13] + input[(32 - 14)];
+  step[14] = input[14] + input[(32 - 15)];
+  step[15] = input[15] + input[(32 - 16)];
+  step[16] = -input[16] + input[(32 - 17)];
+  step[17] = -input[17] + input[(32 - 18)];
+  step[18] = -input[18] + input[(32 - 19)];
+  step[19] = -input[19] + input[(32 - 20)];
+  step[20] = -input[20] + input[(32 - 21)];
+  step[21] = -input[21] + input[(32 - 22)];
+  step[22] = -input[22] + input[(32 - 23)];
+  step[23] = -input[23] + input[(32 - 24)];
+  step[24] = -input[24] + input[(32 - 25)];
+  step[25] = -input[25] + input[(32 - 26)];
+  step[26] = -input[26] + input[(32 - 27)];
+  step[27] = -input[27] + input[(32 - 28)];
+  step[28] = -input[28] + input[(32 - 29)];
+  step[29] = -input[29] + input[(32 - 30)];
+  step[30] = -input[30] + input[(32 - 31)];
+  step[31] = -input[31] + input[(32 - 32)];
+
+  // Stage 2
+  output[0] = step[0] + step[16 - 1];
+  output[1] = step[1] + step[16 - 2];
+  output[2] = step[2] + step[16 - 3];
+  output[3] = step[3] + step[16 - 4];
+  output[4] = step[4] + step[16 - 5];
+  output[5] = step[5] + step[16 - 6];
+  output[6] = step[6] + step[16 - 7];
+  output[7] = step[7] + step[16 - 8];
+  output[8] = -step[8] + step[16 - 9];
+  output[9] = -step[9] + step[16 - 10];
+  output[10] = -step[10] + step[16 - 11];
+  output[11] = -step[11] + step[16 - 12];
+  output[12] = -step[12] + step[16 - 13];
+  output[13] = -step[13] + step[16 - 14];
+  output[14] = -step[14] + step[16 - 15];
+  output[15] = -step[15] + step[16 - 16];
+
+  output[16] = step[16];
+  output[17] = step[17];
+  output[18] = step[18];
+  output[19] = step[19];
+
+  output[20] = dct_32_round((-step[20] + step[27]) * cospi_16_64);
+  output[21] = dct_32_round((-step[21] + step[26]) * cospi_16_64);
+  output[22] = dct_32_round((-step[22] + step[25]) * cospi_16_64);
+  output[23] = dct_32_round((-step[23] + step[24]) * cospi_16_64);
+
+  output[24] = dct_32_round((step[24] + step[23]) * cospi_16_64);
+  output[25] = dct_32_round((step[25] + step[22]) * cospi_16_64);
+  output[26] = dct_32_round((step[26] + step[21]) * cospi_16_64);
+  output[27] = dct_32_round((step[27] + step[20]) * cospi_16_64);
+
+  output[28] = step[28];
+  output[29] = step[29];
+  output[30] = step[30];
+  output[31] = step[31];
+
+  // dump the magnitude by 4, hence the intermediate values are within
+  // the range of 16 bits.
+  if (round) {
+    output[0] = half_round_shift(output[0]);
+    output[1] = half_round_shift(output[1]);
+    output[2] = half_round_shift(output[2]);
+    output[3] = half_round_shift(output[3]);
+    output[4] = half_round_shift(output[4]);
+    output[5] = half_round_shift(output[5]);
+    output[6] = half_round_shift(output[6]);
+    output[7] = half_round_shift(output[7]);
+    output[8] = half_round_shift(output[8]);
+    output[9] = half_round_shift(output[9]);
+    output[10] = half_round_shift(output[10]);
+    output[11] = half_round_shift(output[11]);
+    output[12] = half_round_shift(output[12]);
+    output[13] = half_round_shift(output[13]);
+    output[14] = half_round_shift(output[14]);
+    output[15] = half_round_shift(output[15]);
+
+    output[16] = half_round_shift(output[16]);
+    output[17] = half_round_shift(output[17]);
+    output[18] = half_round_shift(output[18]);
+    output[19] = half_round_shift(output[19]);
+    output[20] = half_round_shift(output[20]);
+    output[21] = half_round_shift(output[21]);
+    output[22] = half_round_shift(output[22]);
+    output[23] = half_round_shift(output[23]);
+    output[24] = half_round_shift(output[24]);
+    output[25] = half_round_shift(output[25]);
+    output[26] = half_round_shift(output[26]);
+    output[27] = half_round_shift(output[27]);
+    output[28] = half_round_shift(output[28]);
+    output[29] = half_round_shift(output[29]);
+    output[30] = half_round_shift(output[30]);
+    output[31] = half_round_shift(output[31]);
+  }
+
+  // Stage 3
+  step[0] = output[0] + output[(8 - 1)];
+  step[1] = output[1] + output[(8 - 2)];
+  step[2] = output[2] + output[(8 - 3)];
+  step[3] = output[3] + output[(8 - 4)];
+  step[4] = -output[4] + output[(8 - 5)];
+  step[5] = -output[5] + output[(8 - 6)];
+  step[6] = -output[6] + output[(8 - 7)];
+  step[7] = -output[7] + output[(8 - 8)];
+  step[8] = output[8];
+  step[9] = output[9];
+  step[10] = dct_32_round((-output[10] + output[13]) * cospi_16_64);
+  step[11] = dct_32_round((-output[11] + output[12]) * cospi_16_64);
+  step[12] = dct_32_round((output[12] + output[11]) * cospi_16_64);
+  step[13] = dct_32_round((output[13] + output[10]) * cospi_16_64);
+  step[14] = output[14];
+  step[15] = output[15];
+
+  step[16] = output[16] + output[23];
+  step[17] = output[17] + output[22];
+  step[18] = output[18] + output[21];
+  step[19] = output[19] + output[20];
+  step[20] = -output[20] + output[19];
+  step[21] = -output[21] + output[18];
+  step[22] = -output[22] + output[17];
+  step[23] = -output[23] + output[16];
+  step[24] = -output[24] + output[31];
+  step[25] = -output[25] + output[30];
+  step[26] = -output[26] + output[29];
+  step[27] = -output[27] + output[28];
+  step[28] = output[28] + output[27];
+  step[29] = output[29] + output[26];
+  step[30] = output[30] + output[25];
+  step[31] = output[31] + output[24];
+
+  // Stage 4
+  output[0] = step[0] + step[3];
+  output[1] = step[1] + step[2];
+  output[2] = -step[2] + step[1];
+  output[3] = -step[3] + step[0];
+  output[4] = step[4];
+  output[5] = dct_32_round((-step[5] + step[6]) * cospi_16_64);
+  output[6] = dct_32_round((step[6] + step[5]) * cospi_16_64);
+  output[7] = step[7];
+  output[8] = step[8] + step[11];
+  output[9] = step[9] + step[10];
+  output[10] = -step[10] + step[9];
+  output[11] = -step[11] + step[8];
+  output[12] = -step[12] + step[15];
+  output[13] = -step[13] + step[14];
+  output[14] = step[14] + step[13];
+  output[15] = step[15] + step[12];
+
+  output[16] = step[16];
+  output[17] = step[17];
+  output[18] = dct_32_round(step[18] * -cospi_8_64 + step[29] * cospi_24_64);
+  output[19] = dct_32_round(step[19] * -cospi_8_64 + step[28] * cospi_24_64);
+  output[20] = dct_32_round(step[20] * -cospi_24_64 + step[27] * -cospi_8_64);
+  output[21] = dct_32_round(step[21] * -cospi_24_64 + step[26] * -cospi_8_64);
+  output[22] = step[22];
+  output[23] = step[23];
+  output[24] = step[24];
+  output[25] = step[25];
+  output[26] = dct_32_round(step[26] * cospi_24_64 + step[21] * -cospi_8_64);
+  output[27] = dct_32_round(step[27] * cospi_24_64 + step[20] * -cospi_8_64);
+  output[28] = dct_32_round(step[28] * cospi_8_64 + step[19] * cospi_24_64);
+  output[29] = dct_32_round(step[29] * cospi_8_64 + step[18] * cospi_24_64);
+  output[30] = step[30];
+  output[31] = step[31];
+
+  // Stage 5
+  step[0] = dct_32_round((output[0] + output[1]) * cospi_16_64);
+  step[1] = dct_32_round((-output[1] + output[0]) * cospi_16_64);
+  step[2] = dct_32_round(output[2] * cospi_24_64 + output[3] * cospi_8_64);
+  step[3] = dct_32_round(output[3] * cospi_24_64 - output[2] * cospi_8_64);
+  step[4] = output[4] + output[5];
+  step[5] = -output[5] + output[4];
+  step[6] = -output[6] + output[7];
+  step[7] = output[7] + output[6];
+  step[8] = output[8];
+  step[9] = dct_32_round(output[9] * -cospi_8_64 + output[14] * cospi_24_64);
+  step[10] = dct_32_round(output[10] * -cospi_24_64 + output[13] * -cospi_8_64);
+  step[11] = output[11];
+  step[12] = output[12];
+  step[13] = dct_32_round(output[13] * cospi_24_64 + output[10] * -cospi_8_64);
+  step[14] = dct_32_round(output[14] * cospi_8_64 + output[9] * cospi_24_64);
+  step[15] = output[15];
+
+  step[16] = output[16] + output[19];
+  step[17] = output[17] + output[18];
+  step[18] = -output[18] + output[17];
+  step[19] = -output[19] + output[16];
+  step[20] = -output[20] + output[23];
+  step[21] = -output[21] + output[22];
+  step[22] = output[22] + output[21];
+  step[23] = output[23] + output[20];
+  step[24] = output[24] + output[27];
+  step[25] = output[25] + output[26];
+  step[26] = -output[26] + output[25];
+  step[27] = -output[27] + output[24];
+  step[28] = -output[28] + output[31];
+  step[29] = -output[29] + output[30];
+  step[30] = output[30] + output[29];
+  step[31] = output[31] + output[28];
+
+  // Stage 6
+  output[0] = step[0];
+  output[1] = step[1];
+  output[2] = step[2];
+  output[3] = step[3];
+  output[4] = dct_32_round(step[4] * cospi_28_64 + step[7] * cospi_4_64);
+  output[5] = dct_32_round(step[5] * cospi_12_64 + step[6] * cospi_20_64);
+  output[6] = dct_32_round(step[6] * cospi_12_64 + step[5] * -cospi_20_64);
+  output[7] = dct_32_round(step[7] * cospi_28_64 + step[4] * -cospi_4_64);
+  output[8] = step[8] + step[9];
+  output[9] = -step[9] + step[8];
+  output[10] = -step[10] + step[11];
+  output[11] = step[11] + step[10];
+  output[12] = step[12] + step[13];
+  output[13] = -step[13] + step[12];
+  output[14] = -step[14] + step[15];
+  output[15] = step[15] + step[14];
+
+  output[16] = step[16];
+  output[17] = dct_32_round(step[17] * -cospi_4_64 + step[30] * cospi_28_64);
+  output[18] = dct_32_round(step[18] * -cospi_28_64 + step[29] * -cospi_4_64);
+  output[19] = step[19];
+  output[20] = step[20];
+  output[21] = dct_32_round(step[21] * -cospi_20_64 + step[26] * cospi_12_64);
+  output[22] = dct_32_round(step[22] * -cospi_12_64 + step[25] * -cospi_20_64);
+  output[23] = step[23];
+  output[24] = step[24];
+  output[25] = dct_32_round(step[25] * cospi_12_64 + step[22] * -cospi_20_64);
+  output[26] = dct_32_round(step[26] * cospi_20_64 + step[21] * cospi_12_64);
+  output[27] = step[27];
+  output[28] = step[28];
+  output[29] = dct_32_round(step[29] * cospi_28_64 + step[18] * -cospi_4_64);
+  output[30] = dct_32_round(step[30] * cospi_4_64 + step[17] * cospi_28_64);
+  output[31] = step[31];
+
+  // Stage 7
+  step[0] = output[0];
+  step[1] = output[1];
+  step[2] = output[2];
+  step[3] = output[3];
+  step[4] = output[4];
+  step[5] = output[5];
+  step[6] = output[6];
+  step[7] = output[7];
+  step[8] = dct_32_round(output[8] * cospi_30_64 + output[15] * cospi_2_64);
+  step[9] = dct_32_round(output[9] * cospi_14_64 + output[14] * cospi_18_64);
+  step[10] = dct_32_round(output[10] * cospi_22_64 + output[13] * cospi_10_64);
+  step[11] = dct_32_round(output[11] * cospi_6_64 + output[12] * cospi_26_64);
+  step[12] = dct_32_round(output[12] * cospi_6_64 + output[11] * -cospi_26_64);
+  step[13] = dct_32_round(output[13] * cospi_22_64 + output[10] * -cospi_10_64);
+  step[14] = dct_32_round(output[14] * cospi_14_64 + output[9] * -cospi_18_64);
+  step[15] = dct_32_round(output[15] * cospi_30_64 + output[8] * -cospi_2_64);
+
+  step[16] = output[16] + output[17];
+  step[17] = -output[17] + output[16];
+  step[18] = -output[18] + output[19];
+  step[19] = output[19] + output[18];
+  step[20] = output[20] + output[21];
+  step[21] = -output[21] + output[20];
+  step[22] = -output[22] + output[23];
+  step[23] = output[23] + output[22];
+  step[24] = output[24] + output[25];
+  step[25] = -output[25] + output[24];
+  step[26] = -output[26] + output[27];
+  step[27] = output[27] + output[26];
+  step[28] = output[28] + output[29];
+  step[29] = -output[29] + output[28];
+  step[30] = -output[30] + output[31];
+  step[31] = output[31] + output[30];
+
+  // Final stage --- outputs indices are bit-reversed.
+  output[0] = step[0];
+  output[16] = step[1];
+  output[8] = step[2];
+  output[24] = step[3];
+  output[4] = step[4];
+  output[20] = step[5];
+  output[12] = step[6];
+  output[28] = step[7];
+  output[2] = step[8];
+  output[18] = step[9];
+  output[10] = step[10];
+  output[26] = step[11];
+  output[6] = step[12];
+  output[22] = step[13];
+  output[14] = step[14];
+  output[30] = step[15];
+
+  output[1] = dct_32_round(step[16] * cospi_31_64 + step[31] * cospi_1_64);
+  output[17] = dct_32_round(step[17] * cospi_15_64 + step[30] * cospi_17_64);
+  output[9] = dct_32_round(step[18] * cospi_23_64 + step[29] * cospi_9_64);
+  output[25] = dct_32_round(step[19] * cospi_7_64 + step[28] * cospi_25_64);
+  output[5] = dct_32_round(step[20] * cospi_27_64 + step[27] * cospi_5_64);
+  output[21] = dct_32_round(step[21] * cospi_11_64 + step[26] * cospi_21_64);
+  output[13] = dct_32_round(step[22] * cospi_19_64 + step[25] * cospi_13_64);
+  output[29] = dct_32_round(step[23] * cospi_3_64 + step[24] * cospi_29_64);
+  output[3] = dct_32_round(step[24] * cospi_3_64 + step[23] * -cospi_29_64);
+  output[19] = dct_32_round(step[25] * cospi_19_64 + step[22] * -cospi_13_64);
+  output[11] = dct_32_round(step[26] * cospi_11_64 + step[21] * -cospi_21_64);
+  output[27] = dct_32_round(step[27] * cospi_27_64 + step[20] * -cospi_5_64);
+  output[7] = dct_32_round(step[28] * cospi_7_64 + step[19] * -cospi_25_64);
+  output[23] = dct_32_round(step[29] * cospi_23_64 + step[18] * -cospi_9_64);
+  output[15] = dct_32_round(step[30] * cospi_15_64 + step[17] * -cospi_17_64);
+  output[31] = dct_32_round(step[31] * cospi_31_64 + step[16] * -cospi_1_64);
+}
+
+void aom_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
+  int i, j;
+  tran_high_t output[32 * 32];
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = input[j * stride + i] * 4;
+    aom_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = output[j + i * 32];
+    aom_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      out[j + i * 32] =
+          (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
+  }
+}
+
+// Note that although we use dct_32_round in dct32 computation flow,
+// this 2d fdct32x32 for rate-distortion optimization loop is operating
+// within 16 bits precision.
+void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *out, int stride) {
+  int i, j;
+  tran_high_t output[32 * 32];
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = input[j * stride + i] * 4;
+    aom_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      // TODO(cd): see quality impact of only doing
+      //           output[j * 32 + i] = (temp_out[j] + 1) >> 2;
+      //           PS: also change code in aom_dsp/x86/aom_dct_sse2.c
+      output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = output[j + i * 32];
+    aom_fdct32(temp_in, temp_out, 1);
+    for (j = 0; j < 32; ++j) out[j + i * 32] = (tran_low_t)temp_out[j];
+  }
+}
+
+void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  int sum = 0;
+  for (r = 0; r < 32; ++r)
+    for (c = 0; c < 32; ++c) sum += input[r * stride + c];
+
+  output[0] = (tran_low_t)(sum >> 3);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  aom_fdct4x4_c(input, output, stride);
+}
+
+void aom_highbd_fdct8x8_c(const int16_t *input, tran_low_t *final_output,
+                          int stride) {
+  aom_fdct8x8_c(input, final_output, stride);
+}
+
+void aom_highbd_fdct8x8_1_c(const int16_t *input, tran_low_t *final_output,
+                            int stride) {
+  aom_fdct8x8_1_c(input, final_output, stride);
+}
+
+void aom_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output,
+                            int stride) {
+  aom_fdct16x16_c(input, output, stride);
+}
+
+void aom_highbd_fdct16x16_1_c(const int16_t *input, tran_low_t *output,
+                              int stride) {
+  aom_fdct16x16_1_c(input, output, stride);
+}
+
+void aom_highbd_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
+  aom_fdct32x32_c(input, out, stride);
+}
+
+void aom_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *out,
+                               int stride) {
+  aom_fdct32x32_rd_c(input, out, stride);
+}
+
+void aom_highbd_fdct32x32_1_c(const int16_t *input, tran_low_t *out,
+                              int stride) {
+  aom_fdct32x32_1_c(input, out, stride);
+}
+#endif  // CONFIG_HIGHBITDEPTH

third_party/aom/aom_dsp/fwd_txfm.h

@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_FWD_TXFM_H_
+#define AOM_DSP_FWD_TXFM_H_
+
+#include "aom_dsp/txfm_common.h"
+
+static INLINE tran_high_t saturate_int16(tran_high_t value) {
+  tran_high_t result;
+  result = value > INT16_MAX ? INT16_MAX : value;
+  return result < INT16_MIN ? INT16_MIN : result;
+}
+
+static INLINE tran_high_t fdct_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return rv;
+}
+
+void aom_fdct32(const tran_high_t *input, tran_high_t *output, int round);
+#endif  // AOM_DSP_FWD_TXFM_H_