import FIREFOX_52_6_0esr_RELEASE from mozilla-esr52 hg repo

media/libvpx/vp9/encoder/arm/neon/vp9_dct_neon.c

@@ -0,0 +1,241 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_idct.h"
+
+void vp9_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;
+  }
+}
+
+void vp9_fdct8x8_quant_neon(const int16_t *input, int stride,
+                            int16_t* coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t* zbin_ptr,
+                            const int16_t* round_ptr, const int16_t* quant_ptr,
+                            const int16_t* quant_shift_ptr,
+                            int16_t* qcoeff_ptr, int16_t* dqcoeff_ptr,
+                            const int16_t* dequant_ptr, uint16_t* eob_ptr,
+                            const int16_t* scan_ptr,
+                            const int16_t* iscan_ptr) {
+  int16_t temp_buffer[64];
+  (void)coeff_ptr;
+
+  vp9_fdct8x8_neon(input, temp_buffer, stride);
+  vp9_quantize_fp_neon(temp_buffer, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                       quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                       dequant_ptr, eob_ptr, scan_ptr, iscan_ptr);
+}
+
+void vp9_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, cospi_16_64);
+    v_t0_hi = vmulq_n_s32(v_t0_hi, cospi_16_64);
+    v_t1_lo = vmulq_n_s32(v_t1_lo, cospi_16_64);
+    v_t1_hi = vmulq_n_s32(v_t1_hi, 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 vp9_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);
+  }
+}
+

media/libvpx/vp9/encoder/arm/neon/vp9_quantize_neon.c

@@ -0,0 +1,118 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include <math.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_rd.h"
+
+void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
+                          int skip_block, const int16_t *zbin_ptr,
+                          const int16_t *round_ptr, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                          uint16_t *eob_ptr,
+                          const int16_t *scan, const int16_t *iscan) {
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)scan;
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    int i;
+    const int16x8_t v_zero = vdupq_n_s16(0);
+    const int16x8_t v_one = vdupq_n_s16(1);
+    int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
+    int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
+    int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
+    int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    // adjust for dc
+    v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
+    v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
+    v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
+    // process dc and the first seven ac coeffs
+    {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[0]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
+                                           vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
+                                           vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
+                                            vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[0], v_dqcoeff);
+      v_round = vmovq_n_s16(round_ptr[1]);
+      v_quant = vmovq_n_s16(quant_ptr[1]);
+      v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    }
+    // now process the rest of the ac coeffs
+    for (i = 8; i < count; i += 8) {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
+                                           vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
+                                           vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
+                                            vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
+    }
+    {
+      const int16x4_t v_eobmax_3210 =
+          vmax_s16(vget_low_s16(v_eobmax_76543210),
+                   vget_high_s16(v_eobmax_76543210));
+      const int64x1_t v_eobmax_xx32 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+      const int16x4_t v_eobmax_tmp =
+          vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+      const int64x1_t v_eobmax_xxx3 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+      const int16x4_t v_eobmax_final =
+          vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+
+      *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
+    }
+  } else {
+    memset(qcoeff_ptr, 0, count * sizeof(int16_t));
+    memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
+    *eob_ptr = 0;
+  }
+}

media/libvpx/vp9/encoder/arm/neon/vp9_subtract_neon.c

@@ -0,0 +1,81 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+void vp9_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;
+    }
+  }
+}

media/libvpx/vp9/encoder/arm/neon/vp9_variance_neon.c

@@ -0,0 +1,153 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx_ports/mem.h"
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_filter.h"
+
+static 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 *vp9_filter) {
+  const uint8x8_t f0 = vmov_n_u8(vp9_filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(vp9_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 *vp9_filter) {
+  const uint8x8_t f0 = vmov_n_u8(vp9_filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(vp9_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 vp9_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 vpx_variance8x8_neon(temp2, 8, dst, dst_stride, sse);
+}
+
+unsigned int vp9_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 vpx_variance16x16_neon(temp2, 16, dst, dst_stride, sse);
+}
+
+unsigned int vp9_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 vpx_variance32x32_neon(temp2, 32, dst, dst_stride, sse);
+}
+
+unsigned int vp9_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 vpx_variance64x64_neon(temp2, 64, dst, dst_stride, sse);
+}

media/libvpx/vp9/encoder/arm/neon/vp9enc_avg_neon.c

@@ -0,0 +1,49 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_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 vp9_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;
+}

media/libvpx/vp9/encoder/vp9_aq_complexity.c

@@ -0,0 +1,163 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "vp9/encoder/vp9_aq_complexity.h"
+#include "vp9/encoder/vp9_aq_variance.h"
+#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+#define AQ_C_SEGMENTS  5
+#define DEFAULT_AQ2_SEG 3   // Neutral Q segment
+#define AQ_C_STRENGTHS 3
+static const double aq_c_q_adj_factor[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  { {1.75, 1.25, 1.05, 1.00, 0.90},
+    {2.00, 1.50, 1.15, 1.00, 0.85},
+    {2.50, 1.75, 1.25, 1.00, 0.80} };
+static const double aq_c_transitions[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  { {0.15, 0.30, 0.55, 2.00, 100.0},
+    {0.20, 0.40, 0.65, 2.00, 100.0},
+    {0.25, 0.50, 0.75, 2.00, 100.0} };
+static const double aq_c_var_thresholds[AQ_C_STRENGTHS][AQ_C_SEGMENTS] =
+  { {-4.0, -3.0, -2.0, 100.00, 100.0},
+    {-3.5, -2.5, -1.5, 100.00, 100.0},
+    {-3.0, -2.0, -1.0, 100.00, 100.0} };
+
+#define DEFAULT_COMPLEXITY 64
+
+
+static int get_aq_c_strength(int q_index, vpx_bit_depth_t bit_depth) {
+  // Approximate base quatizer (truncated to int)
+  const int base_quant = vp9_ac_quant(q_index, 0, bit_depth) / 4;
+  return (base_quant > 10) + (base_quant > 25);
+}
+
+void vp9_setup_in_frame_q_adj(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  struct segmentation *const seg = &cm->seg;
+
+  // Make SURE use of floating point in this function is safe.
+  vp9_clear_system_state();
+
+  if (cm->frame_type == KEY_FRAME ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    int segment;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    // Clear down the segment map.
+    memset(cpi->segmentation_map, DEFAULT_AQ2_SEG, cm->mi_rows * cm->mi_cols);
+
+    vp9_clearall_segfeatures(seg);
+
+    // Segmentation only makes sense if the target bits per SB is above a
+    // threshold. Below this the overheads will usually outweigh any benefit.
+    if (cpi->rc.sb64_target_rate < 256) {
+      vp9_disable_segmentation(seg);
+      return;
+    }
+
+    vp9_enable_segmentation(seg);
+
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Default segment "Q" feature is disabled so it defaults to the baseline Q.
+    vp9_disable_segfeature(seg, DEFAULT_AQ2_SEG, SEG_LVL_ALT_Q);
+
+    // Use some of the segments for in frame Q adjustment.
+    for (segment = 0; segment < AQ_C_SEGMENTS; ++segment) {
+      int qindex_delta;
+
+      if (segment == DEFAULT_AQ2_SEG)
+        continue;
+
+      qindex_delta =
+        vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
+                                   aq_c_q_adj_factor[aq_strength][segment],
+                                   cm->bit_depth);
+
+
+      // For AQ complexity mode, we dont allow Q0 in a segment if the base
+      // Q is not 0. Q0 (lossless) implies 4x4 only and in AQ mode 2 a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+      if ((cm->base_qindex + qindex_delta) > 0) {
+        vp9_enable_segfeature(seg, segment, SEG_LVL_ALT_Q);
+        vp9_set_segdata(seg, segment, SEG_LVL_ALT_Q, qindex_delta);
+      }
+    }
+  }
+}
+
+#define DEFAULT_LV_THRESH 10.0
+#define MIN_DEFAULT_LV_THRESH 8.0
+#define VAR_STRENGTH_STEP 0.25
+// Select a segment for the current block.
+// The choice of segment for a block depends on the ratio of the projected
+// bits for the block vs a target average and its spatial complexity.
+void vp9_caq_select_segment(VP9_COMP *cpi, MACROBLOCK *mb, BLOCK_SIZE bs,
+                            int mi_row, int mi_col, int projected_rate) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
+  const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
+  const int xmis = MIN(cm->mi_cols - mi_col, num_8x8_blocks_wide_lookup[bs]);
+  const int ymis = MIN(cm->mi_rows - mi_row, num_8x8_blocks_high_lookup[bs]);
+  int x, y;
+  int i;
+  unsigned char segment;
+
+  if (0) {
+    segment = DEFAULT_AQ2_SEG;
+  } else {
+    // Rate depends on fraction of a SB64 in frame (xmis * ymis / bw * bh).
+    // It is converted to bits * 256 units.
+    const int target_rate = (cpi->rc.sb64_target_rate * xmis * ymis * 256) /
+                            (bw * bh);
+    double logvar;
+    double low_var_thresh;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    vp9_clear_system_state();
+    low_var_thresh = (cpi->oxcf.pass == 2)
+      ? MAX(cpi->twopass.mb_av_energy, MIN_DEFAULT_LV_THRESH)
+      : DEFAULT_LV_THRESH;
+
+    vp9_setup_src_planes(mb, cpi->Source, mi_row, mi_col);
+    logvar = vp9_log_block_var(cpi, mb, bs);
+
+    segment = AQ_C_SEGMENTS - 1;    // Just in case no break out below.
+    for (i = 0; i < AQ_C_SEGMENTS; ++i) {
+      // Test rate against a threshold value and variance against a threshold.
+      // Increasing segment number (higher variance and complexity) = higher Q.
+      if ((projected_rate <
+           target_rate * aq_c_transitions[aq_strength][i]) &&
+          (logvar < (low_var_thresh + aq_c_var_thresholds[aq_strength][i]))) {
+        segment = i;
+        break;
+      }
+    }
+  }
+
+  // Fill in the entires in the segment map corresponding to this SB64.
+  for (y = 0; y < ymis; y++) {
+    for (x = 0; x < xmis; x++) {
+      cpi->segmentation_map[mi_offset + y * cm->mi_cols + x] = segment;
+    }
+  }
+}

media/libvpx/vp9/encoder/vp9_aq_complexity.h

@@ -0,0 +1,37 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_AQ_COMPLEXITY_H_
+#define VP9_ENCODER_VP9_AQ_COMPLEXITY_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/common/vp9_enums.h"
+
+struct VP9_COMP;
+struct macroblock;
+
+// Select a segment for the current Block.
+void vp9_caq_select_segment(struct VP9_COMP *cpi, struct macroblock *,
+                            BLOCK_SIZE bs,
+                            int mi_row, int mi_col, int projected_rate);
+
+// This function sets up a set of segments with delta Q values around
+// the baseline frame quantizer.
+void vp9_setup_in_frame_q_adj(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_AQ_COMPLEXITY_H_

media/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.c

@@ -0,0 +1,519 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+struct CYCLIC_REFRESH {
+  // Percentage of blocks per frame that are targeted as candidates
+  // for cyclic refresh.
+  int percent_refresh;
+  // Maximum q-delta as percentage of base q.
+  int max_qdelta_perc;
+  // Superblock starting index for cycling through the frame.
+  int sb_index;
+  // Controls how long block will need to wait to be refreshed again, in
+  // excess of the cycle time, i.e., in the case of all zero motion, block
+  // will be refreshed every (100/percent_refresh + time_for_refresh) frames.
+  int time_for_refresh;
+  // Target number of (8x8) blocks that are set for delta-q.
+  int target_num_seg_blocks;
+  // Actual number of (8x8) blocks that were applied delta-q.
+  int actual_num_seg1_blocks;
+  int actual_num_seg2_blocks;
+  // RD mult. parameters for segment 1.
+  int rdmult;
+  // Cyclic refresh map.
+  signed char *map;
+  // Thresholds applied to the projected rate/distortion of the coding block,
+  // when deciding whether block should be refreshed.
+  int64_t thresh_rate_sb;
+  int64_t thresh_dist_sb;
+  // Threshold applied to the motion vector (in units of 1/8 pel) of the
+  // coding block, when deciding whether block should be refreshed.
+  int16_t motion_thresh;
+  // Rate target ratio to set q delta.
+  double rate_ratio_qdelta;
+  double low_content_avg;
+  int qindex_delta_seg1;
+  int qindex_delta_seg2;
+};
+
+CYCLIC_REFRESH *vp9_cyclic_refresh_alloc(int mi_rows, int mi_cols) {
+  CYCLIC_REFRESH *const cr = vpx_calloc(1, sizeof(*cr));
+  if (cr == NULL)
+    return NULL;
+
+  cr->map = vpx_calloc(mi_rows * mi_cols, sizeof(*cr->map));
+  if (cr->map == NULL) {
+    vpx_free(cr);
+    return NULL;
+  }
+
+  return cr;
+}
+
+void vp9_cyclic_refresh_free(CYCLIC_REFRESH *cr) {
+  vpx_free(cr->map);
+  vpx_free(cr);
+}
+
+// Check if we should turn off cyclic refresh based on bitrate condition.
+static int apply_cyclic_refresh_bitrate(const VP9_COMMON *cm,
+                                        const RATE_CONTROL *rc) {
+  // Turn off cyclic refresh if bits available per frame is not sufficiently
+  // larger than bit cost of segmentation. Segment map bit cost should scale
+  // with number of seg blocks, so compare available bits to number of blocks.
+  // Average bits available per frame = avg_frame_bandwidth
+  // Number of (8x8) blocks in frame = mi_rows * mi_cols;
+  const float factor = 0.25;
+  const int number_blocks = cm->mi_rows  * cm->mi_cols;
+  // The condition below corresponds to turning off at target bitrates:
+  // (at 30fps), ~12kbps for CIF, 36kbps for VGA, 100kps for HD/720p.
+  // Also turn off at very small frame sizes, to avoid too large fraction of
+  // superblocks to be refreshed per frame. Threshold below is less than QCIF.
+  if (rc->avg_frame_bandwidth < factor * number_blocks ||
+      number_blocks / 64 < 5)
+    return 0;
+  else
+    return 1;
+}
+
+// Check if this coding block, of size bsize, should be considered for refresh
+// (lower-qp coding). Decision can be based on various factors, such as
+// size of the coding block (i.e., below min_block size rejected), coding
+// mode, and rate/distortion.
+static int candidate_refresh_aq(const CYCLIC_REFRESH *cr,
+                                const MB_MODE_INFO *mbmi,
+                                int64_t rate,
+                                int64_t dist,
+                                int bsize) {
+  MV mv = mbmi->mv[0].as_mv;
+  // Reject the block for lower-qp coding if projected distortion
+  // is above the threshold, and any of the following is true:
+  // 1) mode uses large mv
+  // 2) mode is an intra-mode
+  // Otherwise accept for refresh.
+  if (dist > cr->thresh_dist_sb &&
+      (mv.row > cr->motion_thresh || mv.row < -cr->motion_thresh ||
+       mv.col > cr->motion_thresh || mv.col < -cr->motion_thresh ||
+       !is_inter_block(mbmi)))
+    return CR_SEGMENT_ID_BASE;
+  else  if (bsize >= BLOCK_16X16 &&
+            rate < cr->thresh_rate_sb &&
+            is_inter_block(mbmi) &&
+            mbmi->mv[0].as_int == 0)
+    // More aggressive delta-q for bigger blocks with zero motion.
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BOOST1;
+}
+
+// Compute delta-q for the segment.
+static int compute_deltaq(const VP9_COMP *cpi, int q, double rate_factor) {
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int deltaq = vp9_compute_qdelta_by_rate(rc, cpi->common.frame_type,
+                                          q, rate_factor,
+                                          cpi->common.bit_depth);
+  if ((-deltaq) > cr->max_qdelta_perc * q / 100) {
+    deltaq = -cr->max_qdelta_perc * q / 100;
+  }
+  return deltaq;
+}
+
+// For the just encoded frame, estimate the bits, incorporating the delta-q
+// from non-base segment. For now ignore effect of multiple segments
+// (with different delta-q). Note this function is called in the postencode
+// (called from rc_update_rate_correction_factors()).
+int vp9_cyclic_refresh_estimate_bits_at_q(const VP9_COMP *cpi,
+                                          double correction_factor) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int estimated_bits;
+  int mbs = cm->MBs;
+  int num8x8bl = mbs << 2;
+  // Weight for non-base segments: use actual number of blocks refreshed in
+  // previous/just encoded frame. Note number of blocks here is in 8x8 units.
+  double weight_segment1 = (double)cr->actual_num_seg1_blocks / num8x8bl;
+  double weight_segment2 = (double)cr->actual_num_seg2_blocks / num8x8bl;
+  // Take segment weighted average for estimated bits.
+  estimated_bits = (int)((1.0 - weight_segment1 - weight_segment2) *
+      vp9_estimate_bits_at_q(cm->frame_type, cm->base_qindex, mbs,
+                             correction_factor, cm->bit_depth) +
+                             weight_segment1 *
+      vp9_estimate_bits_at_q(cm->frame_type,
+                             cm->base_qindex + cr->qindex_delta_seg1, mbs,
+                             correction_factor, cm->bit_depth) +
+                             weight_segment2 *
+      vp9_estimate_bits_at_q(cm->frame_type,
+                             cm->base_qindex + cr->qindex_delta_seg2, mbs,
+                             correction_factor, cm->bit_depth));
+  return estimated_bits;
+}
+
+// Prior to encoding the frame, estimate the bits per mb, for a given q = i and
+// a corresponding delta-q (for segment 1). This function is called in the
+// rc_regulate_q() to set the base qp index.
+// Note: the segment map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or
+// to 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock, prior to encoding.
+int vp9_cyclic_refresh_rc_bits_per_mb(const VP9_COMP *cpi, int i,
+                                      double correction_factor) {
+  const VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int bits_per_mb;
+  int num8x8bl = cm->MBs << 2;
+  // Weight for segment prior to encoding: take the average of the target
+  // number for the frame to be encoded and the actual from the previous frame.
+  double weight_segment = (double)((cr->target_num_seg_blocks +
+      cr->actual_num_seg1_blocks + cr->actual_num_seg2_blocks) >> 1) /
+      num8x8bl;
+  // Compute delta-q corresponding to qindex i.
+  int deltaq = compute_deltaq(cpi, i, cr->rate_ratio_qdelta);
+  // Take segment weighted average for bits per mb.
+  bits_per_mb = (int)((1.0 - weight_segment) *
+      vp9_rc_bits_per_mb(cm->frame_type, i, correction_factor, cm->bit_depth) +
+      weight_segment *
+      vp9_rc_bits_per_mb(cm->frame_type, i + deltaq, correction_factor,
+                         cm->bit_depth));
+  return bits_per_mb;
+}
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void vp9_cyclic_refresh_update_segment(VP9_COMP *const cpi,
+                                       MB_MODE_INFO *const mbmi,
+                                       int mi_row, int mi_col,
+                                       BLOCK_SIZE bsize,
+                                       int64_t rate,
+                                       int64_t dist,
+                                       int skip) {
+  const VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = MIN(cm->mi_cols - mi_col, bw);
+  const int ymis = MIN(cm->mi_rows - mi_row, bh);
+  const int block_index = mi_row * cm->mi_cols + mi_col;
+  const int refresh_this_block = candidate_refresh_aq(cr, mbmi, rate, dist,
+                                                      bsize);
+  // Default is to not update the refresh map.
+  int new_map_value = cr->map[block_index];
+  int x = 0; int y = 0;
+
+  // If this block is labeled for refresh, check if we should reset the
+  // segment_id.
+  if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
+    mbmi->segment_id = refresh_this_block;
+    // Reset segment_id if will be skipped.
+    if (skip)
+      mbmi->segment_id = CR_SEGMENT_ID_BASE;
+  }
+
+  // Update the cyclic refresh map, to be used for setting segmentation map
+  // for the next frame. If the block  will be refreshed this frame, mark it
+  // as clean. The magnitude of the -ve influences how long before we consider
+  // it for refresh again.
+  if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
+    new_map_value = -cr->time_for_refresh;
+  } else if (refresh_this_block) {
+    // Else if it is accepted as candidate for refresh, and has not already
+    // been refreshed (marked as 1) then mark it as a candidate for cleanup
+    // for future time (marked as 0), otherwise don't update it.
+    if (cr->map[block_index] == 1)
+      new_map_value = 0;
+  } else {
+    // Leave it marked as block that is not candidate for refresh.
+    new_map_value = 1;
+  }
+
+  // Update entries in the cyclic refresh map with new_map_value, and
+  // copy mbmi->segment_id into global segmentation map.
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++) {
+      cr->map[block_index + y * cm->mi_cols + x] = new_map_value;
+      cpi->segmentation_map[block_index + y * cm->mi_cols + x] =
+          mbmi->segment_id;
+    }
+}
+
+// Update the actual number of blocks that were applied the segment delta q.
+void vp9_cyclic_refresh_postencode(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int mi_row, mi_col;
+  cr->actual_num_seg1_blocks = 0;
+  cr->actual_num_seg2_blocks = 0;
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++)
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      if (cyclic_refresh_segment_id(
+          seg_map[mi_row * cm->mi_cols + mi_col]) == CR_SEGMENT_ID_BOOST1)
+        cr->actual_num_seg1_blocks++;
+      else if (cyclic_refresh_segment_id(
+          seg_map[mi_row * cm->mi_cols + mi_col]) == CR_SEGMENT_ID_BOOST2)
+        cr->actual_num_seg2_blocks++;
+    }
+}
+
+// Set golden frame update interval, for non-svc 1 pass CBR mode.
+void vp9_cyclic_refresh_set_golden_update(VP9_COMP *const cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  // Set minimum gf_interval for GF update to a multiple (== 2) of refresh
+  // period. Depending on past encoding stats, GF flag may be reset and update
+  // may not occur until next baseline_gf_interval.
+  if (cr->percent_refresh > 0)
+    rc->baseline_gf_interval = 4 * (100 / cr->percent_refresh);
+  else
+    rc->baseline_gf_interval = 40;
+}
+
+// Update some encoding stats (from the just encoded frame). If this frame's
+// background has high motion, refresh the golden frame. Otherwise, if the
+// golden reference is to be updated check if we should NOT update the golden
+// ref.
+void vp9_cyclic_refresh_check_golden_update(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int mi_row, mi_col;
+  double fraction_low = 0.0;
+  int low_content_frame = 0;
+
+  MODE_INFO **mi = cm->mi_grid_visible;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int rows = cm->mi_rows, cols = cm->mi_cols;
+  int cnt1 = 0, cnt2 = 0;
+  int force_gf_refresh = 0;
+
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      int16_t abs_mvr = mi[0]->mbmi.mv[0].as_mv.row >= 0 ?
+          mi[0]->mbmi.mv[0].as_mv.row : -1 * mi[0]->mbmi.mv[0].as_mv.row;
+      int16_t abs_mvc = mi[0]->mbmi.mv[0].as_mv.col >= 0 ?
+          mi[0]->mbmi.mv[0].as_mv.col : -1 * mi[0]->mbmi.mv[0].as_mv.col;
+
+      // Calculate the motion of the background.
+      if (abs_mvr <= 16 && abs_mvc <= 16) {
+        cnt1++;
+        if (abs_mvr == 0 && abs_mvc == 0)
+          cnt2++;
+      }
+      mi++;
+
+      // Accumulate low_content_frame.
+      if (cr->map[mi_row * cols + mi_col] < 1)
+        low_content_frame++;
+    }
+    mi += 8;
+  }
+
+  // For video conference clips, if the background has high motion in current
+  // frame because of the camera movement, set this frame as the golden frame.
+  // Use 70% and 5% as the thresholds for golden frame refreshing.
+  if (cnt1 * 10 > (70 * rows * cols) && cnt2 * 20 < cnt1) {
+    vp9_cyclic_refresh_set_golden_update(cpi);
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
+    if (rc->frames_till_gf_update_due > rc->frames_to_key)
+      rc->frames_till_gf_update_due = rc->frames_to_key;
+    cpi->refresh_golden_frame = 1;
+    force_gf_refresh = 1;
+  }
+
+  fraction_low =
+      (double)low_content_frame / (rows * cols);
+  // Update average.
+  cr->low_content_avg = (fraction_low + 3 * cr->low_content_avg) / 4;
+  if (!force_gf_refresh && cpi->refresh_golden_frame == 1) {
+    // Don't update golden reference if the amount of low_content for the
+    // current encoded frame is small, or if the recursive average of the
+    // low_content over the update interval window falls below threshold.
+    if (fraction_low < 0.8 || cr->low_content_avg < 0.7)
+      cpi->refresh_golden_frame = 0;
+    // Reset for next internal.
+    cr->low_content_avg = fraction_low;
+  }
+}
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+// The map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or to
+// 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock.
+// Blocks labeled as BOOST1 may later get set to BOOST2 (during the
+// encoding of the superblock).
+static void cyclic_refresh_update_map(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int i, block_count, bl_index, sb_rows, sb_cols, sbs_in_frame;
+  int xmis, ymis, x, y;
+  memset(seg_map, CR_SEGMENT_ID_BASE, cm->mi_rows * cm->mi_cols);
+  sb_cols = (cm->mi_cols + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
+  sb_rows = (cm->mi_rows + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
+  sbs_in_frame = sb_cols * sb_rows;
+  // Number of target blocks to get the q delta (segment 1).
+  block_count = cr->percent_refresh * cm->mi_rows * cm->mi_cols / 100;
+  // Set the segmentation map: cycle through the superblocks, starting at
+  // cr->mb_index, and stopping when either block_count blocks have been found
+  // to be refreshed, or we have passed through whole frame.
+  assert(cr->sb_index < sbs_in_frame);
+  i = cr->sb_index;
+  cr->target_num_seg_blocks = 0;
+  do {
+    int sum_map = 0;
+    // Get the mi_row/mi_col corresponding to superblock index i.
+    int sb_row_index = (i / sb_cols);
+    int sb_col_index = i - sb_row_index * sb_cols;
+    int mi_row = sb_row_index * MI_BLOCK_SIZE;
+    int mi_col = sb_col_index * MI_BLOCK_SIZE;
+    assert(mi_row >= 0 && mi_row < cm->mi_rows);
+    assert(mi_col >= 0 && mi_col < cm->mi_cols);
+    bl_index = mi_row * cm->mi_cols + mi_col;
+    // Loop through all 8x8 blocks in superblock and update map.
+    xmis = MIN(cm->mi_cols - mi_col,
+               num_8x8_blocks_wide_lookup[BLOCK_64X64]);
+    ymis = MIN(cm->mi_rows - mi_row,
+               num_8x8_blocks_high_lookup[BLOCK_64X64]);
+    for (y = 0; y < ymis; y++) {
+      for (x = 0; x < xmis; x++) {
+        const int bl_index2 = bl_index + y * cm->mi_cols + x;
+        // If the block is as a candidate for clean up then mark it
+        // for possible boost/refresh (segment 1). The segment id may get
+        // reset to 0 later if block gets coded anything other than ZEROMV.
+        if (cr->map[bl_index2] == 0) {
+          sum_map++;
+        } else if (cr->map[bl_index2] < 0) {
+          cr->map[bl_index2]++;
+        }
+      }
+    }
+    // Enforce constant segment over superblock.
+    // If segment is at least half of superblock, set to 1.
+    if (sum_map >= xmis * ymis / 2) {
+      for (y = 0; y < ymis; y++)
+        for (x = 0; x < xmis; x++) {
+          seg_map[bl_index + y * cm->mi_cols + x] = CR_SEGMENT_ID_BOOST1;
+        }
+      cr->target_num_seg_blocks += xmis * ymis;
+    }
+    i++;
+    if (i == sbs_in_frame) {
+      i = 0;
+    }
+  } while (cr->target_num_seg_blocks < block_count && i != cr->sb_index);
+  cr->sb_index = i;
+}
+
+// Set/update global/frame level cyclic refresh parameters.
+void vp9_cyclic_refresh_update_parameters(VP9_COMP *const cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  cr->percent_refresh = 10;
+  // Use larger delta-qp (increase rate_ratio_qdelta) for first few (~4)
+  // periods of the refresh cycle, after a key frame. This corresponds to ~40
+  // frames with cr->percent_refresh = 10.
+  if (rc->frames_since_key <  40)
+    cr->rate_ratio_qdelta = 3.0;
+  else
+    cr->rate_ratio_qdelta = 2.0;
+}
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void vp9_cyclic_refresh_setup(VP9_COMP *const cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  struct segmentation *const seg = &cm->seg;
+  const int apply_cyclic_refresh  = apply_cyclic_refresh_bitrate(cm, rc);
+  if (cm->current_video_frame == 0)
+    cr->low_content_avg = 0.0;
+  // Don't apply refresh on key frame or enhancement layer frames.
+  if (!apply_cyclic_refresh ||
+      (cm->frame_type == KEY_FRAME) ||
+      (cpi->svc.temporal_layer_id > 0) ||
+      (cpi->svc.spatial_layer_id > 0)) {
+    // Set segmentation map to 0 and disable.
+    unsigned char *const seg_map = cpi->segmentation_map;
+    memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
+    vp9_disable_segmentation(&cm->seg);
+    if (cm->frame_type == KEY_FRAME)
+      cr->sb_index = 0;
+    return;
+  } else {
+    int qindex_delta = 0;
+    int qindex2;
+    const double q = vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth);
+    vp9_clear_system_state();
+    cr->max_qdelta_perc = 50;
+    cr->time_for_refresh = 0;
+    // Set rate threshold to some multiple (set to 2 for now) of the target
+    // rate (target is given by sb64_target_rate and scaled by 256).
+    cr->thresh_rate_sb = ((int64_t)(rc->sb64_target_rate) << 8) << 2;
+    // Distortion threshold, quadratic in Q, scale factor to be adjusted.
+    // q will not exceed 457, so (q * q) is within 32bit; see:
+    // vp9_convert_qindex_to_q(), vp9_ac_quant(), ac_qlookup*[].
+    cr->thresh_dist_sb = ((int64_t)(q * q)) << 2;
+    cr->motion_thresh = 32;
+    // Set up segmentation.
+    // Clear down the segment map.
+    vp9_enable_segmentation(&cm->seg);
+    vp9_clearall_segfeatures(seg);
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Note: setting temporal_update has no effect, as the seg-map coding method
+    // (temporal or spatial) is determined in vp9_choose_segmap_coding_method(),
+    // based on the coding cost of each method. For error_resilient mode on the
+    // last_frame_seg_map is set to 0, so if temporal coding is used, it is
+    // relative to 0 previous map.
+    // seg->temporal_update = 0;
+
+    // Segment BASE "Q" feature is disabled so it defaults to the baseline Q.
+    vp9_disable_segfeature(seg, CR_SEGMENT_ID_BASE, SEG_LVL_ALT_Q);
+    // Use segment BOOST1 for in-frame Q adjustment.
+    vp9_enable_segfeature(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q);
+    // Use segment BOOST2 for more aggressive in-frame Q adjustment.
+    vp9_enable_segfeature(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q);
+
+    // Set the q delta for segment BOOST1.
+    qindex_delta = compute_deltaq(cpi, cm->base_qindex, cr->rate_ratio_qdelta);
+    cr->qindex_delta_seg1 = qindex_delta;
+
+    // Compute rd-mult for segment BOOST1.
+    qindex2 = clamp(cm->base_qindex + cm->y_dc_delta_q + qindex_delta, 0, MAXQ);
+
+    cr->rdmult = vp9_compute_rd_mult(cpi, qindex2);
+
+    vp9_set_segdata(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Set a more aggressive (higher) q delta for segment BOOST2.
+    qindex_delta = compute_deltaq(cpi, cm->base_qindex,
+                                  MIN(CR_MAX_RATE_TARGET_RATIO,
+                                      CR_BOOST2_FAC * cr->rate_ratio_qdelta));
+    cr->qindex_delta_seg2 = qindex_delta;
+    vp9_set_segdata(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Update the segmentation and refresh map.
+    cyclic_refresh_update_map(cpi);
+  }
+}
+
+int vp9_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr) {
+  return cr->rdmult;
+}

media/libvpx/vp9/encoder/vp9_aq_cyclicrefresh.h

@@ -0,0 +1,99 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_AQ_CYCLICREFRESH_H_
+#define VP9_ENCODER_VP9_AQ_CYCLICREFRESH_H_
+
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The segment ids used in cyclic refresh: from base (no boost) to increasing
+// boost (higher delta-qp).
+#define CR_SEGMENT_ID_BASE    0
+#define CR_SEGMENT_ID_BOOST1  1
+#define CR_SEGMENT_ID_BOOST2  2
+
+// Maximum rate target ratio for setting segment delta-qp.
+#define CR_MAX_RATE_TARGET_RATIO 4.0
+
+// Boost factor for rate target ratio, for segment CR_SEGMENT_ID_BOOST2.
+#define CR_BOOST2_FAC 1.7
+
+struct VP9_COMP;
+
+struct CYCLIC_REFRESH;
+typedef struct CYCLIC_REFRESH CYCLIC_REFRESH;
+
+CYCLIC_REFRESH *vp9_cyclic_refresh_alloc(int mi_rows, int mi_cols);
+
+void vp9_cyclic_refresh_free(CYCLIC_REFRESH *cr);
+
+// Estimate the bits, incorporating the delta-q from segment 1, after encoding
+// the frame.
+int vp9_cyclic_refresh_estimate_bits_at_q(const struct VP9_COMP *cpi,
+                                          double correction_factor);
+
+// Estimate the bits per mb, for a given q = i and a corresponding delta-q
+// (for segment 1), prior to encoding the frame.
+int vp9_cyclic_refresh_rc_bits_per_mb(const struct VP9_COMP *cpi, int i,
+                                      double correction_factor);
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void vp9_cyclic_refresh_update_segment(struct VP9_COMP *const cpi,
+                                       MB_MODE_INFO *const mbmi,
+                                       int mi_row, int mi_col, BLOCK_SIZE bsize,
+                                       int64_t rate, int64_t dist, int skip);
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+void vp9_cyclic_refresh_update__map(struct VP9_COMP *const cpi);
+
+// Update the actual number of blocks that were applied the segment delta q.
+void vp9_cyclic_refresh_postencode(struct VP9_COMP *const cpi);
+
+// Set golden frame update interval, for non-svc 1 pass CBR mode.
+void vp9_cyclic_refresh_set_golden_update(struct VP9_COMP *const cpi);
+
+// Check if we should not update golden reference, based on past refresh stats.
+void vp9_cyclic_refresh_check_golden_update(struct VP9_COMP *const cpi);
+
+// Set/update global/frame level refresh parameters.
+void vp9_cyclic_refresh_update_parameters(struct VP9_COMP *const cpi);
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void vp9_cyclic_refresh_setup(struct VP9_COMP *const cpi);
+
+int vp9_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr);
+
+static INLINE int cyclic_refresh_segment_id_boosted(int segment_id) {
+  return segment_id == CR_SEGMENT_ID_BOOST1 ||
+         segment_id == CR_SEGMENT_ID_BOOST2;
+}
+
+static INLINE int cyclic_refresh_segment_id(int segment_id) {
+  if (segment_id == CR_SEGMENT_ID_BOOST1)
+    return CR_SEGMENT_ID_BOOST1;
+  else if (segment_id == CR_SEGMENT_ID_BOOST2)
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BASE;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_AQ_CYCLICREFRESH_H_

media/libvpx/vp9/encoder/vp9_aq_variance.c

@@ -0,0 +1,207 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vpx_ports/mem.h"
+
+#include "vp9/encoder/vp9_aq_variance.h"
+
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+#define ENERGY_MIN (-4)
+#define ENERGY_MAX (1)
+#define ENERGY_SPAN (ENERGY_MAX - ENERGY_MIN +  1)
+#define ENERGY_IN_BOUNDS(energy)\
+  assert((energy) >= ENERGY_MIN && (energy) <= ENERGY_MAX)
+
+static const double rate_ratio[MAX_SEGMENTS] =
+  {2.5, 2.0, 1.5, 1.0, 0.75, 1.0, 1.0, 1.0};
+static const int segment_id[ENERGY_SPAN] = {0, 1, 1, 2, 3, 4};
+
+#define SEGMENT_ID(i) segment_id[(i) - ENERGY_MIN]
+
+DECLARE_ALIGNED(16, static const uint8_t, vp9_64_zeros[64]) = {0};
+#if CONFIG_VP9_HIGHBITDEPTH
+DECLARE_ALIGNED(16, static const uint16_t, vp9_highbd_64_zeros[64]) = {0};
+#endif
+
+unsigned int vp9_vaq_segment_id(int energy) {
+  ENERGY_IN_BOUNDS(energy);
+  return SEGMENT_ID(energy);
+}
+
+void vp9_vaq_frame_setup(VP9_COMP *cpi) {
+  VP9_COMMON *cm = &cpi->common;
+  struct segmentation *seg = &cm->seg;
+  int i;
+
+  if (cm->frame_type == KEY_FRAME ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    vp9_enable_segmentation(seg);
+    vp9_clearall_segfeatures(seg);
+
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    vp9_clear_system_state();
+
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      int qindex_delta =
+          vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
+                                     rate_ratio[i], cm->bit_depth);
+
+      // We don't allow qindex 0 in a segment if the base value is not 0.
+      // Q index 0 (lossless) implies 4x4 encoding only and in AQ mode a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+
+      // No need to enable SEG_LVL_ALT_Q for this segment.
+      if (rate_ratio[i] == 1.0) {
+        continue;
+      }
+
+      vp9_set_segdata(seg, i, SEG_LVL_ALT_Q, qindex_delta);
+      vp9_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
+    }
+  }
+}
+
+/* TODO(agrange, paulwilkins): The block_variance calls the unoptimized versions
+ * of variance() and highbd_8_variance(). It should not.
+ */
+static void aq_variance(const uint8_t *a, int  a_stride,
+                        const uint8_t *b, int  b_stride,
+                        int  w, int  h, unsigned int *sse, int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void aq_highbd_variance64(const uint8_t *a8, int  a_stride,
+                                 const uint8_t *b8, int  b_stride,
+                                 int w, int h, uint64_t *sse, uint64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+static void aq_highbd_8_variance(const uint8_t *a8, int  a_stride,
+                                 const uint8_t *b8, int  b_stride,
+                                 int w, int h, unsigned int *sse, int *sum) {
+  uint64_t sse_long = 0;
+  uint64_t sum_long = 0;
+  aq_highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (unsigned int)sse_long;
+  *sum = (int)sum_long;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static unsigned int block_variance(VP9_COMP *cpi, MACROBLOCK *x,
+                                   BLOCK_SIZE bs) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int var, sse;
+  int right_overflow = (xd->mb_to_right_edge < 0) ?
+      ((-xd->mb_to_right_edge) >> 3) : 0;
+  int bottom_overflow = (xd->mb_to_bottom_edge < 0) ?
+      ((-xd->mb_to_bottom_edge) >> 3) : 0;
+
+  if (right_overflow || bottom_overflow) {
+    const int bw = 8 * num_8x8_blocks_wide_lookup[bs] - right_overflow;
+    const int bh = 8 * num_8x8_blocks_high_lookup[bs] - bottom_overflow;
+    int avg;
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      aq_highbd_8_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                           CONVERT_TO_BYTEPTR(vp9_highbd_64_zeros), 0, bw, bh,
+                           &sse, &avg);
+      sse >>= 2 * (xd->bd - 8);
+      avg >>= (xd->bd - 8);
+    } else {
+      aq_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                  vp9_64_zeros, 0, bw, bh, &sse, &avg);
+    }
+#else
+    aq_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                vp9_64_zeros, 0, bw, bh, &sse, &avg);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    var = sse - (((int64_t)avg * avg) / (bw * bh));
+    return (256 * var) / (bw * bh);
+  } else {
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
+                               x->plane[0].src.stride,
+                               CONVERT_TO_BYTEPTR(vp9_highbd_64_zeros),
+                               0, &sse);
+    } else {
+      var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
+                               x->plane[0].src.stride,
+                               vp9_64_zeros, 0, &sse);
+    }
+#else
+    var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
+                             x->plane[0].src.stride,
+                             vp9_64_zeros, 0, &sse);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    return (256 * var) >> num_pels_log2_lookup[bs];
+  }
+}
+
+double vp9_log_block_var(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  unsigned int var = block_variance(cpi, x, bs);
+  vp9_clear_system_state();
+  return log(var + 1.0);
+}
+
+#define DEFAULT_E_MIDPOINT 10.0
+int vp9_block_energy(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  double energy;
+  double energy_midpoint;
+  vp9_clear_system_state();
+  energy_midpoint =
+    (cpi->oxcf.pass == 2) ? cpi->twopass.mb_av_energy : DEFAULT_E_MIDPOINT;
+  energy = vp9_log_block_var(cpi, x, bs) - energy_midpoint;
+  return clamp((int)round(energy), ENERGY_MIN, ENERGY_MAX);
+}

media/libvpx/vp9/encoder/vp9_aq_variance.h

@@ -0,0 +1,31 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_AQ_VARIANCE_H_
+#define VP9_ENCODER_VP9_AQ_VARIANCE_H_
+
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int vp9_vaq_segment_id(int energy);
+void vp9_vaq_frame_setup(VP9_COMP *cpi);
+
+int vp9_block_energy(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+double vp9_log_block_var(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_AQ_VARIANCE_H_

media/libvpx/vp9/encoder/vp9_avg.c

@@ -0,0 +1,211 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vpx_ports/mem.h"
+
+unsigned int vp9_avg_8x8_c(const uint8_t *s, int p) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 8; ++i, s+=p)
+    for (j = 0; j < 8; sum += s[j], ++j) {}
+
+  return (sum + 32) >> 6;
+}
+
+unsigned int vp9_avg_4x4_c(const uint8_t *s, int p) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 4; ++i, s+=p)
+    for (j = 0; j < 4; sum += s[j], ++j) {}
+
+  return (sum + 8) >> 4;
+}
+
+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;
+}
+
+void vp9_hadamard_8x8_c(int16_t const *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);
+    tmp_buf += 8;
+    ++src_diff;
+  }
+
+  tmp_buf = &buffer[0];
+  for (idx = 0; idx < 8; ++idx) {
+    hadamard_col8(tmp_buf, 8, coeff);
+    coeff += 8;
+    ++tmp_buf;
+  }
+}
+
+// In place 16x16 2D Hadamard transform
+void vp9_hadamard_16x16_c(int16_t const *src_diff, int src_stride,
+                          int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    int16_t const *src_ptr = src_diff + (idx >> 1) * 8 * src_stride
+                                + (idx & 0x01) * 8;
+    vp9_hadamard_8x8_c(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  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;
+    int16_t b1 = a0 - a1;
+    int16_t b2 = a2 + a3;
+    int16_t b3 = a2 - a3;
+
+    coeff[0]   = (b0 + b2) >> 1;
+    coeff[64]  = (b1 + b3) >> 1;
+    coeff[128] = (b0 - b2) >> 1;
+    coeff[192] = (b1 - b3) >> 1;
+
+    ++coeff;
+  }
+}
+
+int16_t vp9_satd_c(const int16_t *coeff, int length) {
+  int i;
+  int satd = 0;
+  for (i = 0; i < length; ++i)
+    satd += abs(coeff[i]);
+
+  return (int16_t)satd;
+}
+
+// Integer projection onto row vectors.
+void vp9_int_pro_row_c(int16_t *hbuf, uint8_t const *ref,
+                       const int ref_stride, const int height) {
+  int idx;
+  const int norm_factor = MAX(8, height >> 1);
+  for (idx = 0; idx < 16; ++idx) {
+    int i;
+    hbuf[idx] = 0;
+    for (i = 0; i < height; ++i)
+      hbuf[idx] += ref[i * ref_stride];
+    hbuf[idx] /= norm_factor;
+    ++ref;
+  }
+}
+
+int16_t vp9_int_pro_col_c(uint8_t const *ref, const int width) {
+  int idx;
+  int16_t sum = 0;
+  for (idx = 0; idx < width; ++idx)
+    sum += ref[idx];
+  return sum;
+}
+
+int vp9_vector_var_c(int16_t const *ref, int16_t const *src,
+                     const 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];
+    mean += diff;
+    sse += diff * diff;
+  }
+
+  var = sse - ((mean * mean) >> (bwl + 2));
+  return var;
+}
+
+void vp9_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp,
+                      int *min, int *max) {
+  int i, j;
+  *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;
+    }
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+unsigned int vp9_highbd_avg_8x8_c(const uint8_t *s8, int p) {
+  int i, j;
+  int sum = 0;
+  const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
+  for (i = 0; i < 8; ++i, s+=p)
+    for (j = 0; j < 8; sum += s[j], ++j) {}
+
+  return (sum + 32) >> 6;
+}
+
+unsigned int vp9_highbd_avg_4x4_c(const uint8_t *s8, int p) {
+  int i, j;
+  int sum = 0;
+  const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
+  for (i = 0; i < 4; ++i, s+=p)
+    for (j = 0; j < 4; sum += s[j], ++j) {}
+
+  return (sum + 8) >> 4;
+}
+
+void vp9_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_VP9_HIGHBITDEPTH
+
+

media/libvpx/vp9/encoder/vp9_bitstream.h

@@ -0,0 +1,37 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_BITSTREAM_H_
+#define VP9_ENCODER_VP9_BITSTREAM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/encoder/vp9_encoder.h"
+
+void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size);
+
+static INLINE int vp9_preserve_existing_gf(VP9_COMP *cpi) {
+  return !cpi->multi_arf_allowed && cpi->refresh_golden_frame &&
+         cpi->rc.is_src_frame_alt_ref &&
+         (!cpi->use_svc ||      // Add spatial svc base layer case here
+          (is_two_pass_svc(cpi) &&
+           cpi->svc.spatial_layer_id == 0 &&
+           cpi->svc.layer_context[0].gold_ref_idx >=0 &&
+           cpi->oxcf.ss_enable_auto_arf[0]));
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_BITSTREAM_H_

media/libvpx/vp9/encoder/vp9_block.h

@@ -0,0 +1,137 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_BLOCK_H_
+#define VP9_ENCODER_VP9_BLOCK_H_
+
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_entropy.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  unsigned int sse;
+  int sum;
+  unsigned int var;
+} diff;
+
+struct macroblock_plane {
+  DECLARE_ALIGNED(16, int16_t, src_diff[64 * 64]);
+  tran_low_t *qcoeff;
+  tran_low_t *coeff;
+  uint16_t *eobs;
+  struct buf_2d src;
+
+  // Quantizer setings
+  int16_t *quant_fp;
+  int16_t *round_fp;
+  int16_t *quant;
+  int16_t *quant_shift;
+  int16_t *zbin;
+  int16_t *round;
+
+  int64_t quant_thred[2];
+};
+
+/* The [2] dimension is for whether we skip the EOB node (i.e. if previous
+ * coefficient in this block was zero) or not. */
+typedef unsigned int vp9_coeff_cost[PLANE_TYPES][REF_TYPES][COEF_BANDS][2]
+                                   [COEFF_CONTEXTS][ENTROPY_TOKENS];
+
+typedef struct macroblock MACROBLOCK;
+struct macroblock {
+  struct macroblock_plane plane[MAX_MB_PLANE];
+
+  MACROBLOCKD e_mbd;
+  int skip_block;
+  int select_tx_size;
+  int skip_recode;
+  int skip_optimize;
+  int q_index;
+
+  int errorperbit;
+  int sadperbit16;
+  int sadperbit4;
+  int rddiv;
+  int rdmult;
+  int mb_energy;
+
+  // These are set to their default values at the beginning, and then adjusted
+  // further in the encoding process.
+  BLOCK_SIZE min_partition_size;
+  BLOCK_SIZE max_partition_size;
+
+  int mv_best_ref_index[MAX_REF_FRAMES];
+  unsigned int max_mv_context[MAX_REF_FRAMES];
+  unsigned int source_variance;
+  unsigned int pred_sse[MAX_REF_FRAMES];
+  int pred_mv_sad[MAX_REF_FRAMES];
+
+  int nmvjointcost[MV_JOINTS];
+  int *nmvcost[2];
+  int *nmvcost_hp[2];
+  int **mvcost;
+
+  int nmvjointsadcost[MV_JOINTS];
+  int *nmvsadcost[2];
+  int *nmvsadcost_hp[2];
+  int **mvsadcost;
+
+  // These define limits to motion vector components to prevent them
+  // from extending outside the UMV borders
+  int mv_col_min;
+  int mv_col_max;
+  int mv_row_min;
+  int mv_row_max;
+
+  uint8_t zcoeff_blk[TX_SIZES][256];
+  int skip;
+
+  int encode_breakout;
+
+  // note that token_costs is the cost when eob node is skipped
+  vp9_coeff_cost token_costs[TX_SIZES];
+
+  int optimize;
+
+  // indicate if it is in the rd search loop or encoding process
+  int use_lp32x32fdct;
+  int skip_encode;
+
+  // use fast quantization process
+  int quant_fp;
+
+  // skip forward transform and quantization
+  uint8_t skip_txfm[MAX_MB_PLANE << 2];
+
+  int64_t bsse[MAX_MB_PLANE << 2];
+
+  // Used to store sub partition's choices.
+  MV pred_mv[MAX_REF_FRAMES];
+
+  // Strong color activity detection. Used in RTC coding mode to enhance
+  // the visual quality at the boundary of moving color objects.
+  uint8_t color_sensitivity[2];
+
+  void (*fwd_txm4x4)(const int16_t *input, tran_low_t *output, int stride);
+  void (*itxm_add)(const tran_low_t *input, uint8_t *dest, int stride, int eob);
+#if CONFIG_VP9_HIGHBITDEPTH
+  void (*highbd_itxm_add)(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, int bd);
+#endif
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_BLOCK_H_

media/libvpx/vp9/encoder/vp9_blockiness.c

@@ -0,0 +1,138 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_convolve.h"
+#include "vp9/common/vp9_filter.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+static int horizontal_filter(const uint8_t *s) {
+  return (s[1] - s[-2]) * 2 + (s[-1] - s[0]) * 6;
+}
+
+static int vertical_filter(const uint8_t *s, int p) {
+  return (s[p] - s[-2 * p]) * 2 + (s[-p] - s[0]) * 6;
+}
+
+static int variance(int sum, int sum_squared, int size) {
+  return sum_squared / size - (sum / size) * (sum / size);
+}
+// Calculate a blockiness level for a vertical block edge.
+// This function returns a new blockiness metric that's defined as
+
+//              p0 p1 p2 p3
+//              q0 q1 q2 q3
+// block edge ->
+//              r0 r1 r2 r3
+//              s0 s1 s2 s3
+
+// blockiness =  p0*-2+q0*6+r0*-6+s0*2 +
+//               p1*-2+q1*6+r1*-6+s1*2 +
+//               p2*-2+q2*6+r2*-6+s2*2 +
+//               p3*-2+q3*6+r3*-6+s3*2 ;
+
+// reconstructed_blockiness = abs(blockiness from reconstructed buffer -
+//                                blockiness from source buffer,0)
+//
+// I make the assumption that flat blocks are much more visible than high
+// contrast blocks. As such, I scale the result of the blockiness calc
+// by dividing the blockiness by the variance of the pixels on either side
+// of the edge as follows:
+// var_0 = (q0^2+q1^2+q2^2+q3^2) - ((q0 + q1 + q2 + q3) / 4 )^2
+// var_1 = (r0^2+r1^2+r2^2+r3^2) - ((r0 + r1 + r2 + r3) / 4 )^2
+// The returned blockiness is the scaled value
+// Reconstructed blockiness / ( 1 + var_0 + var_1 ) ;
+int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r, int rp,
+                        int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, s += sp, r += rp) {
+    s_blockiness += horizontal_filter(s);
+    r_blockiness += horizontal_filter(r);
+    sum_0 += s[0];
+    sum_sq_0 += s[0]*s[0];
+    sum_1 += s[-1];
+    sum_sq_1 += s[-1]*s[-1];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// Calculate a blockiness level for a horizontal block edge
+// same as above.
+int blockiness_horizontal(const uint8_t *s, int sp, const uint8_t *r, int rp,
+                          int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, ++s, ++r) {
+    s_blockiness += vertical_filter(s, sp);
+    r_blockiness += vertical_filter(r, rp);
+    sum_0 += s[0];
+    sum_sq_0 += s[0] * s[0];
+    sum_1 += s[-sp];
+    sum_sq_1 += s[-sp] * s[-sp];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// This function returns the blockiness for the entire frame currently by
+// looking at all borders in steps of 4.
+double vp9_get_blockiness(const unsigned char *img1, int img1_pitch,
+                          const unsigned char *img2, int img2_pitch,
+                          int width, int height ) {
+  double blockiness = 0;
+  int i, j;
+  vp9_clear_system_state();
+  for (i = 0; i < height; i += 4, img1 += img1_pitch * 4,
+       img2 += img2_pitch * 4) {
+    for (j = 0; j < width; j += 4) {
+      if (i > 0 && i < height && j > 0 && j < width) {
+        blockiness += blockiness_vertical(img1 + j, img1_pitch,
+                                          img2 + j, img2_pitch, 4);
+        blockiness += blockiness_horizontal(img1 + j, img1_pitch,
+                                            img2 + j, img2_pitch, 4);
+      }
+    }
+  }
+  blockiness /= width * height / 16;
+  return blockiness;
+}

media/libvpx/vp9/encoder/vp9_context_tree.c

@@ -0,0 +1,158 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_encoder.h"
+
+static const BLOCK_SIZE square[] = {
+  BLOCK_8X8,
+  BLOCK_16X16,
+  BLOCK_32X32,
+  BLOCK_64X64,
+};
+
+static void alloc_mode_context(VP9_COMMON *cm, int num_4x4_blk,
+                               PICK_MODE_CONTEXT *ctx) {
+  const int num_blk = (num_4x4_blk < 4 ? 4 : num_4x4_blk);
+  const int num_pix = num_blk << 4;
+  int i, k;
+  ctx->num_4x4_blk = num_blk;
+
+  CHECK_MEM_ERROR(cm, ctx->zcoeff_blk,
+                  vpx_calloc(num_4x4_blk, sizeof(uint8_t)));
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (k = 0; k < 3; ++k) {
+      CHECK_MEM_ERROR(cm, ctx->coeff[i][k],
+                      vpx_memalign(16, num_pix * sizeof(*ctx->coeff[i][k])));
+      CHECK_MEM_ERROR(cm, ctx->qcoeff[i][k],
+                      vpx_memalign(16, num_pix * sizeof(*ctx->qcoeff[i][k])));
+      CHECK_MEM_ERROR(cm, ctx->dqcoeff[i][k],
+                      vpx_memalign(16, num_pix * sizeof(*ctx->dqcoeff[i][k])));
+      CHECK_MEM_ERROR(cm, ctx->eobs[i][k],
+                      vpx_memalign(16, num_pix * sizeof(*ctx->eobs[i][k])));
+      ctx->coeff_pbuf[i][k]   = ctx->coeff[i][k];
+      ctx->qcoeff_pbuf[i][k]  = ctx->qcoeff[i][k];
+      ctx->dqcoeff_pbuf[i][k] = ctx->dqcoeff[i][k];
+      ctx->eobs_pbuf[i][k]    = ctx->eobs[i][k];
+    }
+  }
+}
+
+static void free_mode_context(PICK_MODE_CONTEXT *ctx) {
+  int i, k;
+  vpx_free(ctx->zcoeff_blk);
+  ctx->zcoeff_blk = 0;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (k = 0; k < 3; ++k) {
+      vpx_free(ctx->coeff[i][k]);
+      ctx->coeff[i][k] = 0;
+      vpx_free(ctx->qcoeff[i][k]);
+      ctx->qcoeff[i][k] = 0;
+      vpx_free(ctx->dqcoeff[i][k]);
+      ctx->dqcoeff[i][k] = 0;
+      vpx_free(ctx->eobs[i][k]);
+      ctx->eobs[i][k] = 0;
+    }
+  }
+}
+
+static void alloc_tree_contexts(VP9_COMMON *cm, PC_TREE *tree,
+                                int num_4x4_blk) {
+  alloc_mode_context(cm, num_4x4_blk, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[0]);
+
+  /* TODO(Jbb): for 4x8 and 8x4 these allocated values are not used.
+   * Figure out a better way to do this. */
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->horizontal[1]);
+  alloc_mode_context(cm, num_4x4_blk/2, &tree->vertical[1]);
+}
+
+static void free_tree_contexts(PC_TREE *tree) {
+  free_mode_context(&tree->none);
+  free_mode_context(&tree->horizontal[0]);
+  free_mode_context(&tree->horizontal[1]);
+  free_mode_context(&tree->vertical[0]);
+  free_mode_context(&tree->vertical[1]);
+}
+
+// This function sets up a tree of contexts such that at each square
+// partition level. There are contexts for none, horizontal, vertical, and
+// split.  Along with a block_size value and a selected block_size which
+// represents the state of our search.
+void vp9_setup_pc_tree(VP9_COMMON *cm, ThreadData *td) {
+  int i, j;
+  const int leaf_nodes = 64;
+  const int tree_nodes = 64 + 16 + 4 + 1;
+  int pc_tree_index = 0;
+  PC_TREE *this_pc;
+  PICK_MODE_CONTEXT *this_leaf;
+  int square_index = 1;
+  int nodes;
+
+  vpx_free(td->leaf_tree);
+  CHECK_MEM_ERROR(cm, td->leaf_tree, vpx_calloc(leaf_nodes,
+                                                sizeof(*td->leaf_tree)));
+  vpx_free(td->pc_tree);
+  CHECK_MEM_ERROR(cm, td->pc_tree, vpx_calloc(tree_nodes,
+                                              sizeof(*td->pc_tree)));
+
+  this_pc = &td->pc_tree[0];
+  this_leaf = &td->leaf_tree[0];
+
+  // 4x4 blocks smaller than 8x8 but in the same 8x8 block share the same
+  // context so we only need to allocate 1 for each 8x8 block.
+  for (i = 0; i < leaf_nodes; ++i)
+    alloc_mode_context(cm, 1, &td->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (pc_tree_index = 0; pc_tree_index < leaf_nodes; ++pc_tree_index) {
+    PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+    tree->block_size = square[0];
+    alloc_tree_contexts(cm, tree, 4);
+    tree->leaf_split[0] = this_leaf++;
+    for (j = 1; j < 4; j++)
+      tree->leaf_split[j] = tree->leaf_split[0];
+  }
+
+  // Each node has 4 leaf nodes, fill each block_size level of the tree
+  // from leafs to the root.
+  for (nodes = 16; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i) {
+      PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+      alloc_tree_contexts(cm, tree, 4 << (2 * square_index));
+      tree->block_size = square[square_index];
+      for (j = 0; j < 4; j++)
+        tree->split[j] = this_pc++;
+      ++pc_tree_index;
+    }
+    ++square_index;
+  }
+  td->pc_root = &td->pc_tree[tree_nodes - 1];
+  td->pc_root[0].none.best_mode_index = 2;
+}
+
+void vp9_free_pc_tree(ThreadData *td) {
+  const int tree_nodes = 64 + 16 + 4 + 1;
+  int i;
+
+  // Set up all 4x4 mode contexts
+  for (i = 0; i < 64; ++i)
+    free_mode_context(&td->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (i = 0; i < tree_nodes; ++i)
+    free_tree_contexts(&td->pc_tree[i]);
+
+  vpx_free(td->pc_tree);
+  td->pc_tree = NULL;
+  vpx_free(td->leaf_tree);
+  td->leaf_tree = NULL;
+}

media/libvpx/vp9/encoder/vp9_context_tree.h

@@ -0,0 +1,86 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_CONTEXT_TREE_H_
+#define VP9_ENCODER_VP9_CONTEXT_TREE_H_
+
+#include "vp9/common/vp9_blockd.h"
+
+struct VP9_COMP;
+struct VP9Common;
+struct ThreadData;
+
+// Structure to hold snapshot of coding context during the mode picking process
+typedef struct {
+  MODE_INFO mic;
+  uint8_t *zcoeff_blk;
+  tran_low_t *coeff[MAX_MB_PLANE][3];
+  tran_low_t *qcoeff[MAX_MB_PLANE][3];
+  tran_low_t *dqcoeff[MAX_MB_PLANE][3];
+  uint16_t *eobs[MAX_MB_PLANE][3];
+
+  // dual buffer pointers, 0: in use, 1: best in store
+  tran_low_t *coeff_pbuf[MAX_MB_PLANE][3];
+  tran_low_t *qcoeff_pbuf[MAX_MB_PLANE][3];
+  tran_low_t *dqcoeff_pbuf[MAX_MB_PLANE][3];
+  uint16_t *eobs_pbuf[MAX_MB_PLANE][3];
+
+  int is_coded;
+  int num_4x4_blk;
+  int skip;
+  int pred_pixel_ready;
+  // For current partition, only if all Y, U, and V transform blocks'
+  // coefficients are quantized to 0, skippable is set to 0.
+  int skippable;
+  uint8_t skip_txfm[MAX_MB_PLANE << 2];
+  int best_mode_index;
+  int hybrid_pred_diff;
+  int comp_pred_diff;
+  int single_pred_diff;
+  int64_t tx_rd_diff[TX_MODES];
+  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+
+  // TODO(jingning) Use RD_COST struct here instead. This involves a boarder
+  // scope of refactoring.
+  int rate;
+  int64_t dist;
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  unsigned int newmv_sse;
+  unsigned int zeromv_sse;
+  PREDICTION_MODE best_sse_inter_mode;
+  int_mv best_sse_mv;
+  MV_REFERENCE_FRAME best_reference_frame;
+  MV_REFERENCE_FRAME best_zeromv_reference_frame;
+#endif
+
+  // motion vector cache for adaptive motion search control in partition
+  // search loop
+  MV pred_mv[MAX_REF_FRAMES];
+  INTERP_FILTER pred_interp_filter;
+} PICK_MODE_CONTEXT;
+
+typedef struct PC_TREE {
+  int index;
+  PARTITION_TYPE partitioning;
+  BLOCK_SIZE block_size;
+  PICK_MODE_CONTEXT none;
+  PICK_MODE_CONTEXT horizontal[2];
+  PICK_MODE_CONTEXT vertical[2];
+  union {
+    struct PC_TREE *split[4];
+    PICK_MODE_CONTEXT *leaf_split[4];
+  };
+} PC_TREE;
+
+void vp9_setup_pc_tree(struct VP9Common *cm, struct ThreadData *td);
+void vp9_free_pc_tree(struct ThreadData *td);
+
+#endif /* VP9_ENCODER_VP9_CONTEXT_TREE_H_ */

media/libvpx/vp9/encoder/vp9_cost.c

@@ -0,0 +1,62 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_cost.h"
+
+const unsigned int vp9_prob_cost[256] = {
+  2047, 2047, 1791, 1641, 1535, 1452, 1385, 1328, 1279, 1235, 1196, 1161,
+  1129, 1099, 1072, 1046, 1023, 1000, 979,  959,  940,  922,  905,  889,
+  873,  858,  843,  829,  816,  803,  790,  778,  767,  755,  744,  733,
+  723,  713,  703,  693,  684,  675,  666,  657,  649,  641,  633,  625,
+  617,  609,  602,  594,  587,  580,  573,  567,  560,  553,  547,  541,
+  534,  528,  522,  516,  511,  505,  499,  494,  488,  483,  477,  472,
+  467,  462,  457,  452,  447,  442,  437,  433,  428,  424,  419,  415,
+  410,  406,  401,  397,  393,  389,  385,  381,  377,  373,  369,  365,
+  361,  357,  353,  349,  346,  342,  338,  335,  331,  328,  324,  321,
+  317,  314,  311,  307,  304,  301,  297,  294,  291,  288,  285,  281,
+  278,  275,  272,  269,  266,  263,  260,  257,  255,  252,  249,  246,
+  243,  240,  238,  235,  232,  229,  227,  224,  221,  219,  216,  214,
+  211,  208,  206,  203,  201,  198,  196,  194,  191,  189,  186,  184,
+  181,  179,  177,  174,  172,  170,  168,  165,  163,  161,  159,  156,
+  154,  152,  150,  148,  145,  143,  141,  139,  137,  135,  133,  131,
+  129,  127,  125,  123,  121,  119,  117,  115,  113,  111,  109,  107,
+  105,  103,  101,  99,   97,   95,   93,   92,   90,   88,   86,   84,
+  82,   81,   79,   77,   75,   73,   72,   70,   68,   66,   65,   63,
+  61,   60,   58,   56,   55,   53,   51,   50,   48,   46,   45,   43,
+  41,   40,   38,   37,   35,   33,   32,   30,   29,   27,   25,   24,
+  22,   21,   19,   18,   16,   15,   13,   12,   10,   9,    7,    6,
+  4,    3,    1,    1};
+
+static void cost(int *costs, vp9_tree tree, const vp9_prob *probs,
+                 int i, int c) {
+  const vp9_prob prob = probs[i / 2];
+  int b;
+
+  for (b = 0; b <= 1; ++b) {
+    const int cc = c + vp9_cost_bit(prob, b);
+    const vp9_tree_index ii = tree[i + b];
+
+    if (ii <= 0)
+      costs[-ii] = cc;
+    else
+      cost(costs, tree, probs, ii, cc);
+  }
+}
+
+void vp9_cost_tokens(int *costs, const vp9_prob *probs, vp9_tree tree) {
+  cost(costs, tree, probs, 0, 0);
+}
+
+void vp9_cost_tokens_skip(int *costs, const vp9_prob *probs, vp9_tree tree) {
+  assert(tree[0] <= 0 && tree[1] > 0);
+
+  costs[-tree[0]] = vp9_cost_bit(probs[0], 0);
+  cost(costs, tree, probs, 2, 0);
+}

media/libvpx/vp9/encoder/vp9_cost.h

@@ -0,0 +1,55 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_COST_H_
+#define VP9_ENCODER_VP9_COST_H_
+
+#include "vp9/common/vp9_prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const unsigned int vp9_prob_cost[256];
+
+#define vp9_cost_zero(prob) (vp9_prob_cost[prob])
+
+#define vp9_cost_one(prob) vp9_cost_zero(vp9_complement(prob))
+
+#define vp9_cost_bit(prob, bit) vp9_cost_zero((bit) ? vp9_complement(prob) \
+                                                    : (prob))
+
+static INLINE unsigned int cost_branch256(const unsigned int ct[2],
+                                          vp9_prob p) {
+  return ct[0] * vp9_cost_zero(p) + ct[1] * vp9_cost_one(p);
+}
+
+static INLINE int treed_cost(vp9_tree tree, const vp9_prob *probs,
+                             int bits, int len) {
+  int cost = 0;
+  vp9_tree_index i = 0;
+
+  do {
+    const int bit = (bits >> --len) & 1;
+    cost += vp9_cost_bit(probs[i >> 1], bit);
+    i = tree[i + bit];
+  } while (len);
+
+  return cost;
+}
+
+void vp9_cost_tokens(int *costs, const vp9_prob *probs, vp9_tree tree);
+void vp9_cost_tokens_skip(int *costs, const vp9_prob *probs, vp9_tree tree);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_COST_H_

media/libvpx/vp9/encoder/vp9_dct.h

@@ -0,0 +1,61 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_DCT_H_
+#define VP9_ENCODER_VP9_DCT_H_
+
+#include "vp9/common/vp9_idct.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
+void vp9_highbd_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
+void vp9_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output,
+                            int stride);
+void vp9_highbd_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride);
+void vp9_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *out,
+                               int stride);
+
+void vp9_fdct4(const tran_low_t *input, tran_low_t *output);
+void vp9_fadst4(const tran_low_t *input, tran_low_t *output);
+void vp9_fdct8(const tran_low_t *input, tran_low_t *output);
+void vp9_fadst8(const tran_low_t *input, tran_low_t *output);
+void vp9_fdct16(const tran_low_t in[16], tran_low_t out[16]);
+void vp9_fadst16(const tran_low_t *input, tran_low_t *output);
+void vp9_fdct32(const tran_high_t *input, tran_high_t *output, int round);
+
+static const transform_2d FHT_4[] = {
+  { vp9_fdct4,  vp9_fdct4  },  // DCT_DCT  = 0
+  { vp9_fadst4, vp9_fdct4  },  // ADST_DCT = 1
+  { vp9_fdct4,  vp9_fadst4 },  // DCT_ADST = 2
+  { vp9_fadst4, vp9_fadst4 }   // ADST_ADST = 3
+};
+
+static const transform_2d FHT_8[] = {
+  { vp9_fdct8,  vp9_fdct8  },  // DCT_DCT  = 0
+  { vp9_fadst8, vp9_fdct8  },  // ADST_DCT = 1
+  { vp9_fdct8,  vp9_fadst8 },  // DCT_ADST = 2
+  { vp9_fadst8, vp9_fadst8 }   // ADST_ADST = 3
+};
+
+static const transform_2d FHT_16[] = {
+  { vp9_fdct16,  vp9_fdct16  },  // DCT_DCT  = 0
+  { vp9_fadst16, vp9_fdct16  },  // ADST_DCT = 1
+  { vp9_fdct16,  vp9_fadst16 },  // DCT_ADST = 2
+  { vp9_fadst16, vp9_fadst16 }   // ADST_ADST = 3
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_DCT_H_

media/libvpx/vp9/encoder/vp9_denoiser.c

@@ -0,0 +1,498 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+
+/* The VP9 denoiser is a work-in-progress. It currently is only designed to work
+ * with speed 6, though it (inexplicably) seems to also work with speed 5 (one
+ * would need to modify the source code in vp9_pickmode.c and vp9_encoder.c to
+ * make the calls to the vp9_denoiser_* functions when in speed 5).
+ *
+ * The implementation is very similar to that of the VP8 denoiser. While
+ * choosing the motion vectors / reference frames, the denoiser is run, and if
+ * it did not modify the signal to much, the denoised block is copied to the
+ * signal.
+ */
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv);
+#endif
+
+static int absdiff_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  return 3 + (increase_denoising ? 1 : 0);
+}
+
+static int delta_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  (void)increase_denoising;
+  return 4;
+}
+
+static int noise_motion_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  (void)bs;
+  (void)increase_denoising;
+  return 625;
+}
+
+static unsigned int sse_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 60 : 40);
+}
+
+static int sse_diff_thresh(BLOCK_SIZE bs, int increase_denoising,
+                           int motion_magnitude) {
+  if (motion_magnitude >
+      noise_motion_thresh(bs, increase_denoising)) {
+    return 0;
+  } else {
+    return (1 << num_pels_log2_lookup[bs]) * 20;
+  }
+}
+
+int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
+}
+
+static int total_adj_weak_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
+}
+
+// TODO(jackychen): If increase_denoising is enabled in the future,
+// we might need to update the code for calculating 'total_adj' in
+// case the C code is not bit-exact with corresponding sse2 code.
+int vp9_denoiser_filter_c(const uint8_t *sig, int sig_stride,
+                          const uint8_t *mc_avg,
+                          int mc_avg_stride,
+                          uint8_t *avg, int avg_stride,
+                          int increase_denoising,
+                          BLOCK_SIZE bs,
+                          int motion_magnitude) {
+  int r, c;
+  const uint8_t *sig_start = sig;
+  const uint8_t *mc_avg_start = mc_avg;
+  uint8_t *avg_start = avg;
+  int diff, adj, absdiff, delta;
+  int adj_val[] = {3, 4, 6};
+  int total_adj = 0;
+  int shift_inc = 1;
+
+  // If motion_magnitude is small, making the denoiser more aggressive by
+  // increasing the adjustment for each level. Add another increment for
+  // blocks that are labeled for increase denoising.
+  if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) {
+    if (increase_denoising) {
+      shift_inc = 2;
+    }
+    adj_val[0] += shift_inc;
+    adj_val[1] += shift_inc;
+    adj_val[2] += shift_inc;
+  }
+
+  // First attempt to apply a strong temporal denoising filter.
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
+      diff = mc_avg[c] - sig[c];
+      absdiff = abs(diff);
+
+      if (absdiff <= absdiff_thresh(bs, increase_denoising)) {
+        avg[c] = mc_avg[c];
+        total_adj += diff;
+      } else {
+        switch (absdiff) {
+          case 4: case 5: case 6: case 7:
+            adj = adj_val[0];
+            break;
+          case 8: case 9: case 10: case 11:
+          case 12: case 13: case 14: case 15:
+            adj = adj_val[1];
+            break;
+          default:
+            adj = adj_val[2];
+        }
+        if (diff > 0) {
+          avg[c] = MIN(UINT8_MAX, sig[c] + adj);
+          total_adj += adj;
+        } else {
+          avg[c] = MAX(0, sig[c] - adj);
+          total_adj -= adj;
+        }
+      }
+    }
+    sig += sig_stride;
+    avg += avg_stride;
+    mc_avg += mc_avg_stride;
+  }
+
+  // If the strong filter did not modify the signal too much, we're all set.
+  if (abs(total_adj) <= total_adj_strong_thresh(bs, increase_denoising)) {
+    return FILTER_BLOCK;
+  }
+
+  // Otherwise, we try to dampen the filter if the delta is not too high.
+  delta = ((abs(total_adj) - total_adj_strong_thresh(bs, increase_denoising))
+           >> num_pels_log2_lookup[bs]) + 1;
+
+  if (delta >= delta_thresh(bs, increase_denoising)) {
+    return COPY_BLOCK;
+  }
+
+  mc_avg =  mc_avg_start;
+  avg = avg_start;
+  sig = sig_start;
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
+      diff = mc_avg[c] - sig[c];
+      adj = abs(diff);
+      if (adj > delta) {
+        adj = delta;
+      }
+      if (diff > 0) {
+        // Diff positive means we made positive adjustment above
+        // (in first try/attempt), so now make negative adjustment to bring
+        // denoised signal down.
+        avg[c] = MAX(0, avg[c] - adj);
+        total_adj -= adj;
+      } else {
+        // Diff negative means we made negative adjustment above
+        // (in first try/attempt), so now make positive adjustment to bring
+        // denoised signal up.
+        avg[c] = MIN(UINT8_MAX, avg[c] + adj);
+        total_adj += adj;
+      }
+    }
+    sig += sig_stride;
+    avg += avg_stride;
+    mc_avg += mc_avg_stride;
+  }
+
+  // We can use the filter if it has been sufficiently dampened
+  if (abs(total_adj) <= total_adj_weak_thresh(bs, increase_denoising)) {
+    return FILTER_BLOCK;
+  }
+  return COPY_BLOCK;
+}
+
+static uint8_t *block_start(uint8_t *framebuf, int stride,
+                            int mi_row, int mi_col) {
+  return framebuf + (stride * mi_row * 8) + (mi_col * 8);
+}
+
+static VP9_DENOISER_DECISION perform_motion_compensation(VP9_DENOISER *denoiser,
+                                                         MACROBLOCK *mb,
+                                                         BLOCK_SIZE bs,
+                                                         int increase_denoising,
+                                                         int mi_row,
+                                                         int mi_col,
+                                                         PICK_MODE_CONTEXT *ctx,
+                                                         int *motion_magnitude
+                                                         ) {
+  int mv_col, mv_row;
+  int sse_diff = ctx->zeromv_sse - ctx->newmv_sse;
+  MV_REFERENCE_FRAME frame;
+  MACROBLOCKD *filter_mbd = &mb->e_mbd;
+  MB_MODE_INFO *mbmi = &filter_mbd->mi[0]->mbmi;
+  MB_MODE_INFO saved_mbmi;
+  int i, j;
+  struct buf_2d saved_dst[MAX_MB_PLANE];
+  struct buf_2d saved_pre[MAX_MB_PLANE][2];  // 2 pre buffers
+
+  mv_col = ctx->best_sse_mv.as_mv.col;
+  mv_row = ctx->best_sse_mv.as_mv.row;
+  *motion_magnitude = mv_row * mv_row + mv_col * mv_col;
+  frame = ctx->best_reference_frame;
+
+  saved_mbmi = *mbmi;
+
+  // If the best reference frame uses inter-prediction and there is enough of a
+  // difference in sum-squared-error, use it.
+  if (frame != INTRA_FRAME &&
+      sse_diff > sse_diff_thresh(bs, increase_denoising, *motion_magnitude)) {
+    mbmi->ref_frame[0] = ctx->best_reference_frame;
+    mbmi->mode = ctx->best_sse_inter_mode;
+    mbmi->mv[0] = ctx->best_sse_mv;
+  } else {
+    // Otherwise, use the zero reference frame.
+    frame = ctx->best_zeromv_reference_frame;
+
+    mbmi->ref_frame[0] = ctx->best_zeromv_reference_frame;
+    mbmi->mode = ZEROMV;
+    mbmi->mv[0].as_int = 0;
+
+    ctx->best_sse_inter_mode = ZEROMV;
+    ctx->best_sse_mv.as_int = 0;
+    ctx->newmv_sse = ctx->zeromv_sse;
+  }
+
+  if (ctx->newmv_sse > sse_thresh(bs, increase_denoising)) {
+    // Restore everything to its original state
+    *mbmi = saved_mbmi;
+    return COPY_BLOCK;
+  }
+  if (*motion_magnitude >
+     (noise_motion_thresh(bs, increase_denoising) << 3)) {
+    // Restore everything to its original state
+    *mbmi = saved_mbmi;
+    return COPY_BLOCK;
+  }
+
+  // We will restore these after motion compensation.
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (j = 0; j < 2; ++j) {
+      saved_pre[i][j] = filter_mbd->plane[i].pre[j];
+    }
+    saved_dst[i] = filter_mbd->plane[i].dst;
+  }
+
+  // Set the pointers in the MACROBLOCKD to point to the buffers in the denoiser
+  // struct.
+  for (j = 0; j < 2; ++j) {
+    filter_mbd->plane[0].pre[j].buf =
+        block_start(denoiser->running_avg_y[frame].y_buffer,
+                    denoiser->running_avg_y[frame].y_stride,
+                    mi_row, mi_col);
+    filter_mbd->plane[0].pre[j].stride =
+        denoiser->running_avg_y[frame].y_stride;
+    filter_mbd->plane[1].pre[j].buf =
+        block_start(denoiser->running_avg_y[frame].u_buffer,
+                    denoiser->running_avg_y[frame].uv_stride,
+                    mi_row, mi_col);
+    filter_mbd->plane[1].pre[j].stride =
+        denoiser->running_avg_y[frame].uv_stride;
+    filter_mbd->plane[2].pre[j].buf =
+        block_start(denoiser->running_avg_y[frame].v_buffer,
+                    denoiser->running_avg_y[frame].uv_stride,
+                    mi_row, mi_col);
+    filter_mbd->plane[2].pre[j].stride =
+        denoiser->running_avg_y[frame].uv_stride;
+  }
+  filter_mbd->plane[0].dst.buf =
+      block_start(denoiser->mc_running_avg_y.y_buffer,
+                  denoiser->mc_running_avg_y.y_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[0].dst.stride = denoiser->mc_running_avg_y.y_stride;
+  filter_mbd->plane[1].dst.buf =
+      block_start(denoiser->mc_running_avg_y.u_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[1].dst.stride = denoiser->mc_running_avg_y.uv_stride;
+  filter_mbd->plane[2].dst.buf =
+      block_start(denoiser->mc_running_avg_y.v_buffer,
+                  denoiser->mc_running_avg_y.uv_stride,
+                  mi_row, mi_col);
+  filter_mbd->plane[2].dst.stride = denoiser->mc_running_avg_y.uv_stride;
+
+  vp9_build_inter_predictors_sby(filter_mbd, mv_row, mv_col, bs);
+
+  // Restore everything to its original state
+  *mbmi = saved_mbmi;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (j = 0; j < 2; ++j) {
+      filter_mbd->plane[i].pre[j] = saved_pre[i][j];
+    }
+    filter_mbd->plane[i].dst = saved_dst[i];
+  }
+
+  mv_row = ctx->best_sse_mv.as_mv.row;
+  mv_col = ctx->best_sse_mv.as_mv.col;
+
+  return FILTER_BLOCK;
+}
+
+void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs,
+                          PICK_MODE_CONTEXT *ctx) {
+  int motion_magnitude = 0;
+  VP9_DENOISER_DECISION decision = FILTER_BLOCK;
+  YV12_BUFFER_CONFIG avg = denoiser->running_avg_y[INTRA_FRAME];
+  YV12_BUFFER_CONFIG mc_avg = denoiser->mc_running_avg_y;
+  uint8_t *avg_start = block_start(avg.y_buffer, avg.y_stride, mi_row, mi_col);
+  uint8_t *mc_avg_start = block_start(mc_avg.y_buffer, mc_avg.y_stride,
+                                          mi_row, mi_col);
+  struct buf_2d src = mb->plane[0].src;
+
+  decision = perform_motion_compensation(denoiser, mb, bs,
+                                         denoiser->increase_denoising,
+                                         mi_row, mi_col, ctx,
+                                         &motion_magnitude);
+
+  if (decision == FILTER_BLOCK) {
+    decision = vp9_denoiser_filter(src.buf, src.stride,
+                                 mc_avg_start, mc_avg.y_stride,
+                                 avg_start, avg.y_stride,
+                                 0, bs, motion_magnitude);
+  }
+
+  if (decision == FILTER_BLOCK) {
+    vp9_convolve_copy(avg_start, avg.y_stride, src.buf, src.stride,
+                      NULL, 0, NULL, 0,
+                      num_4x4_blocks_wide_lookup[bs] << 2,
+                      num_4x4_blocks_high_lookup[bs] << 2);
+  } else {  // COPY_BLOCK
+    vp9_convolve_copy(src.buf, src.stride, avg_start, avg.y_stride,
+                      NULL, 0, NULL, 0,
+                      num_4x4_blocks_wide_lookup[bs] << 2,
+                      num_4x4_blocks_high_lookup[bs] << 2);
+  }
+}
+
+static void copy_frame(YV12_BUFFER_CONFIG dest, const YV12_BUFFER_CONFIG src) {
+  int r;
+  const uint8_t *srcbuf = src.y_buffer;
+  uint8_t *destbuf = dest.y_buffer;
+
+  assert(dest.y_width == src.y_width);
+  assert(dest.y_height == src.y_height);
+
+  for (r = 0; r < dest.y_height; ++r) {
+    memcpy(destbuf, srcbuf, dest.y_width);
+    destbuf += dest.y_stride;
+    srcbuf += src.y_stride;
+  }
+}
+
+static void swap_frame_buffer(YV12_BUFFER_CONFIG *dest,
+                              YV12_BUFFER_CONFIG *src) {
+  uint8_t *tmp_buf = dest->y_buffer;
+  assert(dest->y_width == src->y_width);
+  assert(dest->y_height == src->y_height);
+  dest->y_buffer = src->y_buffer;
+  src->y_buffer = tmp_buf;
+}
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame) {
+  if (frame_type == KEY_FRAME) {
+    int i;
+    // Start at 1 so as not to overwrite the INTRA_FRAME
+    for (i = 1; i < MAX_REF_FRAMES; ++i)
+      copy_frame(denoiser->running_avg_y[i], src);
+    return;
+  }
+
+  /* For non key frames */
+  if (refresh_alt_ref_frame) {
+    swap_frame_buffer(&denoiser->running_avg_y[ALTREF_FRAME],
+                      &denoiser->running_avg_y[INTRA_FRAME]);
+  }
+  if (refresh_golden_frame) {
+    swap_frame_buffer(&denoiser->running_avg_y[GOLDEN_FRAME],
+                      &denoiser->running_avg_y[INTRA_FRAME]);
+  }
+  if (refresh_last_frame) {
+    swap_frame_buffer(&denoiser->running_avg_y[LAST_FRAME],
+                      &denoiser->running_avg_y[INTRA_FRAME]);
+  }
+}
+
+void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx) {
+  ctx->zeromv_sse = UINT_MAX;
+  ctx->newmv_sse = UINT_MAX;
+}
+
+void vp9_denoiser_update_frame_stats(MB_MODE_INFO *mbmi, unsigned int sse,
+                                     PREDICTION_MODE mode,
+                                     PICK_MODE_CONTEXT *ctx) {
+  // TODO(tkopp): Use both MVs if possible
+  if (mbmi->mv[0].as_int == 0 && sse < ctx->zeromv_sse) {
+    ctx->zeromv_sse = sse;
+    ctx->best_zeromv_reference_frame = mbmi->ref_frame[0];
+  }
+
+  if (mbmi->mv[0].as_int != 0 && sse < ctx->newmv_sse) {
+    ctx->newmv_sse = sse;
+    ctx->best_sse_inter_mode = mode;
+    ctx->best_sse_mv = mbmi->mv[0];
+    ctx->best_reference_frame = mbmi->ref_frame[0];
+  }
+}
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       int border) {
+  int i, fail;
+  const int legacy_byte_alignment = 0;
+  assert(denoiser != NULL);
+
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    fail = vp9_alloc_frame_buffer(&denoiser->running_avg_y[i], width, height,
+                                  ssx, ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                  use_highbitdepth,
+#endif
+                                  border, legacy_byte_alignment);
+    if (fail) {
+      vp9_denoiser_free(denoiser);
+      return 1;
+    }
+#ifdef OUTPUT_YUV_DENOISED
+    make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  }
+
+  fail = vp9_alloc_frame_buffer(&denoiser->mc_running_avg_y, width, height,
+                                ssx, ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                use_highbitdepth,
+#endif
+                                border, legacy_byte_alignment);
+  if (fail) {
+    vp9_denoiser_free(denoiser);
+    return 1;
+  }
+#ifdef OUTPUT_YUV_DENOISED
+  make_grayscale(&denoiser->running_avg_y[i]);
+#endif
+  denoiser->increase_denoising = 0;
+  denoiser->frame_buffer_initialized = 1;
+
+  return 0;
+}
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser) {
+  int i;
+  denoiser->frame_buffer_initialized = 0;
+  if (denoiser == NULL) {
+    return;
+  }
+  for (i = 0; i < MAX_REF_FRAMES; ++i) {
+    vp9_free_frame_buffer(&denoiser->running_avg_y[i]);
+  }
+  vp9_free_frame_buffer(&denoiser->mc_running_avg_y);
+}
+
+#ifdef OUTPUT_YUV_DENOISED
+static void make_grayscale(YV12_BUFFER_CONFIG *yuv) {
+  int r, c;
+  uint8_t *u = yuv->u_buffer;
+  uint8_t *v = yuv->v_buffer;
+
+  for (r = 0; r < yuv->uv_height; ++r) {
+    for (c = 0; c < yuv->uv_width; ++c) {
+      u[c] = UINT8_MAX / 2;
+      v[c] = UINT8_MAX / 2;
+    }
+    u += yuv->uv_stride;
+    v += yuv->uv_stride;
+  }
+}
+#endif

media/libvpx/vp9/encoder/vp9_denoiser.h

@@ -0,0 +1,69 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_DENOISER_H_
+#define VP9_ENCODER_DENOISER_H_
+
+#include "vp9/encoder/vp9_block.h"
+#include "vpx_scale/yv12config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MOTION_MAGNITUDE_THRESHOLD (8 * 3)
+
+typedef enum vp9_denoiser_decision {
+  COPY_BLOCK,
+  FILTER_BLOCK
+} VP9_DENOISER_DECISION;
+
+typedef struct vp9_denoiser {
+  YV12_BUFFER_CONFIG running_avg_y[MAX_REF_FRAMES];
+  YV12_BUFFER_CONFIG mc_running_avg_y;
+  int increase_denoising;
+  int frame_buffer_initialized;
+} VP9_DENOISER;
+
+void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
+                                    YV12_BUFFER_CONFIG src,
+                                    FRAME_TYPE frame_type,
+                                    int refresh_alt_ref_frame,
+                                    int refresh_golden_frame,
+                                    int refresh_last_frame);
+
+void vp9_denoiser_denoise(VP9_DENOISER *denoiser, MACROBLOCK *mb,
+                          int mi_row, int mi_col, BLOCK_SIZE bs,
+                          PICK_MODE_CONTEXT *ctx);
+
+void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx);
+
+void vp9_denoiser_update_frame_stats(MB_MODE_INFO *mbmi,
+                                     unsigned int sse, PREDICTION_MODE mode,
+                                     PICK_MODE_CONTEXT *ctx);
+
+int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
+                       int ssx, int ssy,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       int border);
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising);
+#endif
+
+void vp9_denoiser_free(VP9_DENOISER *denoiser);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_DENOISER_H_

media/libvpx/vp9/encoder/vp9_encodeframe.h

@@ -0,0 +1,49 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_ENCODEFRAME_H_
+#define VP9_ENCODER_VP9_ENCODEFRAME_H_
+
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct macroblock;
+struct yv12_buffer_config;
+struct VP9_COMP;
+struct ThreadData;
+
+// Constants used in SOURCE_VAR_BASED_PARTITION
+#define VAR_HIST_MAX_BG_VAR 1000
+#define VAR_HIST_FACTOR 10
+#define VAR_HIST_BINS (VAR_HIST_MAX_BG_VAR / VAR_HIST_FACTOR + 1)
+#define VAR_HIST_LARGE_CUT_OFF 75
+#define VAR_HIST_SMALL_CUT_OFF 45
+
+void vp9_setup_src_planes(struct macroblock *x,
+                          const struct yv12_buffer_config *src,
+                          int mi_row, int mi_col);
+
+void vp9_encode_frame(struct VP9_COMP *cpi);
+
+void vp9_init_tile_data(struct VP9_COMP *cpi);
+void vp9_encode_tile(struct VP9_COMP *cpi, struct ThreadData *td,
+                     int tile_row, int tile_col);
+
+void vp9_set_variance_partition_thresholds(struct VP9_COMP *cpi, int q);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODEFRAME_H_

media/libvpx/vp9/encoder/vp9_encodemb.h

@@ -0,0 +1,46 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_ENCODEMB_H_
+#define VP9_ENCODER_VP9_ENCODEMB_H_
+
+#include "./vpx_config.h"
+#include "vp9/encoder/vp9_block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct encode_b_args {
+  MACROBLOCK *x;
+  struct optimize_ctx *ctx;
+  int8_t *skip;
+};
+void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize);
+void vp9_encode_sby_pass1(MACROBLOCK *x, BLOCK_SIZE bsize);
+void vp9_xform_quant_fp(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+void vp9_xform_quant_dc(MACROBLOCK *x, int plane, int block,
+                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+void vp9_xform_quant(MACROBLOCK *x, int plane, int block,
+                     BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+
+void vp9_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
+                            TX_SIZE tx_size, void *arg);
+
+void vp9_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODEMB_H_

media/libvpx/vp9/encoder/vp9_encodemv.c

@@ -0,0 +1,266 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encodemv.h"
+
+static struct vp9_token mv_joint_encodings[MV_JOINTS];
+static struct vp9_token mv_class_encodings[MV_CLASSES];
+static struct vp9_token mv_fp_encodings[MV_FP_SIZE];
+static struct vp9_token mv_class0_encodings[CLASS0_SIZE];
+
+void vp9_entropy_mv_init(void) {
+  vp9_tokens_from_tree(mv_joint_encodings, vp9_mv_joint_tree);
+  vp9_tokens_from_tree(mv_class_encodings, vp9_mv_class_tree);
+  vp9_tokens_from_tree(mv_class0_encodings, vp9_mv_class0_tree);
+  vp9_tokens_from_tree(mv_fp_encodings, vp9_mv_fp_tree);
+}
+
+static void encode_mv_component(vp9_writer* w, int comp,
+                                const nmv_component* mvcomp, int usehp) {
+  int offset;
+  const int sign = comp < 0;
+  const int mag = sign ? -comp : comp;
+  const int mv_class = vp9_get_mv_class(mag - 1, &offset);
+  const int d = offset >> 3;                // int mv data
+  const int fr = (offset >> 1) & 3;         // fractional mv data
+  const int hp = offset & 1;                // high precision mv data
+
+  assert(comp != 0);
+
+  // Sign
+  vp9_write(w, sign, mvcomp->sign);
+
+  // Class
+  vp9_write_token(w, vp9_mv_class_tree, mvcomp->classes,
+                  &mv_class_encodings[mv_class]);
+
+  // Integer bits
+  if (mv_class == MV_CLASS_0) {
+    vp9_write_token(w, vp9_mv_class0_tree, mvcomp->class0,
+                    &mv_class0_encodings[d]);
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+    for (i = 0; i < n; ++i)
+      vp9_write(w, (d >> i) & 1, mvcomp->bits[i]);
+  }
+
+  // Fractional bits
+  vp9_write_token(w, vp9_mv_fp_tree,
+                  mv_class == MV_CLASS_0 ?  mvcomp->class0_fp[d] : mvcomp->fp,
+                  &mv_fp_encodings[fr]);
+
+  // High precision bit
+  if (usehp)
+    vp9_write(w, hp,
+              mv_class == MV_CLASS_0 ? mvcomp->class0_hp : mvcomp->hp);
+}
+
+
+static void build_nmv_component_cost_table(int *mvcost,
+                                           const nmv_component* const mvcomp,
+                                           int usehp) {
+  int i, v;
+  int sign_cost[2], class_cost[MV_CLASSES], class0_cost[CLASS0_SIZE];
+  int bits_cost[MV_OFFSET_BITS][2];
+  int class0_fp_cost[CLASS0_SIZE][MV_FP_SIZE], fp_cost[MV_FP_SIZE];
+  int class0_hp_cost[2], hp_cost[2];
+
+  sign_cost[0] = vp9_cost_zero(mvcomp->sign);
+  sign_cost[1] = vp9_cost_one(mvcomp->sign);
+  vp9_cost_tokens(class_cost, mvcomp->classes, vp9_mv_class_tree);
+  vp9_cost_tokens(class0_cost, mvcomp->class0, vp9_mv_class0_tree);
+  for (i = 0; i < MV_OFFSET_BITS; ++i) {
+    bits_cost[i][0] = vp9_cost_zero(mvcomp->bits[i]);
+    bits_cost[i][1] = vp9_cost_one(mvcomp->bits[i]);
+  }
+
+  for (i = 0; i < CLASS0_SIZE; ++i)
+    vp9_cost_tokens(class0_fp_cost[i], mvcomp->class0_fp[i], vp9_mv_fp_tree);
+  vp9_cost_tokens(fp_cost, mvcomp->fp, vp9_mv_fp_tree);
+
+  if (usehp) {
+    class0_hp_cost[0] = vp9_cost_zero(mvcomp->class0_hp);
+    class0_hp_cost[1] = vp9_cost_one(mvcomp->class0_hp);
+    hp_cost[0] = vp9_cost_zero(mvcomp->hp);
+    hp_cost[1] = vp9_cost_one(mvcomp->hp);
+  }
+  mvcost[0] = 0;
+  for (v = 1; v <= MV_MAX; ++v) {
+    int z, c, o, d, e, f, cost = 0;
+    z = v - 1;
+    c = vp9_get_mv_class(z, &o);
+    cost += class_cost[c];
+    d = (o >> 3);               /* int mv data */
+    f = (o >> 1) & 3;           /* fractional pel mv data */
+    e = (o & 1);                /* high precision mv data */
+    if (c == MV_CLASS_0) {
+      cost += class0_cost[d];
+    } else {
+      int i, b;
+      b = c + CLASS0_BITS - 1;  /* number of bits */
+      for (i = 0; i < b; ++i)
+        cost += bits_cost[i][((d >> i) & 1)];
+    }
+    if (c == MV_CLASS_0) {
+      cost += class0_fp_cost[d][f];
+    } else {
+      cost += fp_cost[f];
+    }
+    if (usehp) {
+      if (c == MV_CLASS_0) {
+        cost += class0_hp_cost[e];
+      } else {
+        cost += hp_cost[e];
+      }
+    }
+    mvcost[v] = cost + sign_cost[0];
+    mvcost[-v] = cost + sign_cost[1];
+  }
+}
+
+static int update_mv(vp9_writer *w, const unsigned int ct[2], vp9_prob *cur_p,
+                     vp9_prob upd_p) {
+  const vp9_prob new_p = get_binary_prob(ct[0], ct[1]) | 1;
+  const int update = cost_branch256(ct, *cur_p) + vp9_cost_zero(upd_p) >
+                     cost_branch256(ct, new_p) + vp9_cost_one(upd_p) + 7 * 256;
+  vp9_write(w, update, upd_p);
+  if (update) {
+    *cur_p = new_p;
+    vp9_write_literal(w, new_p >> 1, 7);
+  }
+  return update;
+}
+
+static void write_mv_update(const vp9_tree_index *tree,
+                            vp9_prob probs[/*n - 1*/],
+                            const unsigned int counts[/*n - 1*/],
+                            int n, vp9_writer *w) {
+  int i;
+  unsigned int branch_ct[32][2];
+
+  // Assuming max number of probabilities <= 32
+  assert(n <= 32);
+
+  vp9_tree_probs_from_distribution(tree, branch_ct, counts);
+  for (i = 0; i < n - 1; ++i)
+    update_mv(w, branch_ct[i], &probs[i], MV_UPDATE_PROB);
+}
+
+void vp9_write_nmv_probs(VP9_COMMON *cm, int usehp, vp9_writer *w,
+                         nmv_context_counts *const counts) {
+  int i, j;
+  nmv_context *const mvc = &cm->fc->nmvc;
+
+  write_mv_update(vp9_mv_joint_tree, mvc->joints, counts->joints, MV_JOINTS, w);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *comp = &mvc->comps[i];
+    nmv_component_counts *comp_counts = &counts->comps[i];
+
+    update_mv(w, comp_counts->sign, &comp->sign, MV_UPDATE_PROB);
+    write_mv_update(vp9_mv_class_tree, comp->classes, comp_counts->classes,
+                    MV_CLASSES, w);
+    write_mv_update(vp9_mv_class0_tree, comp->class0, comp_counts->class0,
+                    CLASS0_SIZE, w);
+    for (j = 0; j < MV_OFFSET_BITS; ++j)
+      update_mv(w, comp_counts->bits[j], &comp->bits[j], MV_UPDATE_PROB);
+  }
+
+  for (i = 0; i < 2; ++i) {
+    for (j = 0; j < CLASS0_SIZE; ++j)
+      write_mv_update(vp9_mv_fp_tree, mvc->comps[i].class0_fp[j],
+                      counts->comps[i].class0_fp[j], MV_FP_SIZE, w);
+
+    write_mv_update(vp9_mv_fp_tree, mvc->comps[i].fp, counts->comps[i].fp,
+                    MV_FP_SIZE, w);
+  }
+
+  if (usehp) {
+    for (i = 0; i < 2; ++i) {
+      update_mv(w, counts->comps[i].class0_hp, &mvc->comps[i].class0_hp,
+                MV_UPDATE_PROB);
+      update_mv(w, counts->comps[i].hp, &mvc->comps[i].hp, MV_UPDATE_PROB);
+    }
+  }
+}
+
+void vp9_encode_mv(VP9_COMP* cpi, vp9_writer* w,
+                   const MV* mv, const MV* ref,
+                   const nmv_context* mvctx, int usehp) {
+  const MV diff = {mv->row - ref->row,
+                   mv->col - ref->col};
+  const MV_JOINT_TYPE j = vp9_get_mv_joint(&diff);
+  usehp = usehp && vp9_use_mv_hp(ref);
+
+  vp9_write_token(w, vp9_mv_joint_tree, mvctx->joints, &mv_joint_encodings[j]);
+  if (mv_joint_vertical(j))
+    encode_mv_component(w, diff.row, &mvctx->comps[0], usehp);
+
+  if (mv_joint_horizontal(j))
+    encode_mv_component(w, diff.col, &mvctx->comps[1], usehp);
+
+  // If auto_mv_step_size is enabled then keep track of the largest
+  // motion vector component used.
+  if (cpi->sf.mv.auto_mv_step_size) {
+    unsigned int maxv = MAX(abs(mv->row), abs(mv->col)) >> 3;
+    cpi->max_mv_magnitude = MAX(maxv, cpi->max_mv_magnitude);
+  }
+}
+
+void vp9_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context* ctx, int usehp) {
+  vp9_cost_tokens(mvjoint, ctx->joints, vp9_mv_joint_tree);
+  build_nmv_component_cost_table(mvcost[0], &ctx->comps[0], usehp);
+  build_nmv_component_cost_table(mvcost[1], &ctx->comps[1], usehp);
+}
+
+static void inc_mvs(const MB_MODE_INFO *mbmi, const int_mv mvs[2],
+                    nmv_context_counts *counts) {
+  int i;
+
+  for (i = 0; i < 1 + has_second_ref(mbmi); ++i) {
+    const MV *ref = &mbmi->ref_mvs[mbmi->ref_frame[i]][0].as_mv;
+    const MV diff = {mvs[i].as_mv.row - ref->row,
+                     mvs[i].as_mv.col - ref->col};
+    vp9_inc_mv(&diff, counts);
+  }
+}
+
+void vp9_update_mv_count(ThreadData *td) {
+  const MACROBLOCKD *xd = &td->mb.e_mbd;
+  const MODE_INFO *mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+
+  if (mbmi->sb_type < BLOCK_8X8) {
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[mbmi->sb_type];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[mbmi->sb_type];
+    int idx, idy;
+
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int i = idy * 2 + idx;
+        if (mi->bmi[i].as_mode == NEWMV)
+          inc_mvs(mbmi, mi->bmi[i].as_mv, &td->counts->mv);
+      }
+    }
+  } else {
+    if (mbmi->mode == NEWMV)
+      inc_mvs(mbmi, mbmi->mv, &td->counts->mv);
+  }
+}
+

media/libvpx/vp9/encoder/vp9_encodemv.h

@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_ENCODEMV_H_
+#define VP9_ENCODER_VP9_ENCODEMV_H_
+
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_entropy_mv_init(void);
+
+void vp9_write_nmv_probs(VP9_COMMON *cm, int usehp, vp9_writer *w,
+                         nmv_context_counts *const counts);
+
+void vp9_encode_mv(VP9_COMP *cpi, vp9_writer* w, const MV* mv, const MV* ref,
+                   const nmv_context* mvctx, int usehp);
+
+void vp9_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context* mvctx, int usehp);
+
+void vp9_update_mv_count(ThreadData *td);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODEMV_H_

media/libvpx/vp9/encoder/vp9_encoder.h

@@ -0,0 +1,654 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_ENCODER_H_
+#define VP9_ENCODER_VP9_ENCODER_H_
+
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "vpx/vp8cx.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_ppflags.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_thread.h"
+
+#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_firstpass.h"
+#include "vp9/encoder/vp9_lookahead.h"
+#include "vp9/encoder/vp9_mbgraph.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#if CONFIG_INTERNAL_STATS
+#include "vp9/encoder/vp9_ssim.h"
+#endif
+#include "vp9/encoder/vp9_speed_features.h"
+#include "vp9/encoder/vp9_svc_layercontext.h"
+#include "vp9/encoder/vp9_tokenize.h"
+#include "vp9/encoder/vp9_variance.h"
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+#include "vp9/encoder/vp9_denoiser.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DEFAULT_GF_INTERVAL         10
+
+typedef struct {
+  int nmvjointcost[MV_JOINTS];
+  int nmvcosts[2][MV_VALS];
+  int nmvcosts_hp[2][MV_VALS];
+
+  vp9_prob segment_pred_probs[PREDICTION_PROBS];
+
+  unsigned char *last_frame_seg_map_copy;
+
+  // 0 = Intra, Last, GF, ARF
+  signed char last_ref_lf_deltas[MAX_REF_LF_DELTAS];
+  // 0 = ZERO_MV, MV
+  signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
+
+  FRAME_CONTEXT fc;
+} CODING_CONTEXT;
+
+
+typedef enum {
+  // encode_breakout is disabled.
+  ENCODE_BREAKOUT_DISABLED = 0,
+  // encode_breakout is enabled.
+  ENCODE_BREAKOUT_ENABLED = 1,
+  // encode_breakout is enabled with small max_thresh limit.
+  ENCODE_BREAKOUT_LIMITED = 2
+} ENCODE_BREAKOUT_TYPE;
+
+typedef enum {
+  NORMAL      = 0,
+  FOURFIVE    = 1,
+  THREEFIVE   = 2,
+  ONETWO      = 3
+} VPX_SCALING;
+
+typedef enum {
+  // Good Quality Fast Encoding. The encoder balances quality with the amount of
+  // time it takes to encode the output. Speed setting controls how fast.
+  GOOD,
+
+  // The encoder places priority on the quality of the output over encoding
+  // speed. The output is compressed at the highest possible quality. This
+  // option takes the longest amount of time to encode. Speed setting ignored.
+  BEST,
+
+  // Realtime/Live Encoding. This mode is optimized for realtime encoding (for
+  // example, capturing a television signal or feed from a live camera). Speed
+  // setting controls how fast.
+  REALTIME
+} MODE;
+
+typedef enum {
+  FRAMEFLAGS_KEY    = 1 << 0,
+  FRAMEFLAGS_GOLDEN = 1 << 1,
+  FRAMEFLAGS_ALTREF = 1 << 2,
+} FRAMETYPE_FLAGS;
+
+typedef enum {
+  NO_AQ = 0,
+  VARIANCE_AQ = 1,
+  COMPLEXITY_AQ = 2,
+  CYCLIC_REFRESH_AQ = 3,
+  AQ_MODE_COUNT  // This should always be the last member of the enum
+} AQ_MODE;
+
+typedef enum {
+  RESIZE_NONE = 0,    // No frame resizing allowed (except for SVC).
+  RESIZE_FIXED = 1,   // All frames are coded at the specified dimension.
+  RESIZE_DYNAMIC = 2  // Coded size of each frame is determined by the codec.
+} RESIZE_TYPE;
+
+typedef struct VP9EncoderConfig {
+  BITSTREAM_PROFILE profile;
+  vpx_bit_depth_t bit_depth;     // Codec bit-depth.
+  int width;  // width of data passed to the compressor
+  int height;  // height of data passed to the compressor
+  unsigned int input_bit_depth;  // Input bit depth.
+  double init_framerate;  // set to passed in framerate
+  int64_t target_bandwidth;  // bandwidth to be used in kilobits per second
+
+  int noise_sensitivity;  // pre processing blur: recommendation 0
+  int sharpness;  // sharpening output: recommendation 0:
+  int speed;
+  // maximum allowed bitrate for any intra frame in % of bitrate target.
+  unsigned int rc_max_intra_bitrate_pct;
+  // maximum allowed bitrate for any inter frame in % of bitrate target.
+  unsigned int rc_max_inter_bitrate_pct;
+  // percent of rate boost for golden frame in CBR mode.
+  unsigned int gf_cbr_boost_pct;
+
+  MODE mode;
+  int pass;
+
+  // Key Framing Operations
+  int auto_key;  // autodetect cut scenes and set the keyframes
+  int key_freq;  // maximum distance to key frame.
+
+  int lag_in_frames;  // how many frames lag before we start encoding
+
+  // ----------------------------------------------------------------
+  // DATARATE CONTROL OPTIONS
+
+  // vbr, cbr, constrained quality or constant quality
+  enum vpx_rc_mode rc_mode;
+
+  // buffer targeting aggressiveness
+  int under_shoot_pct;
+  int over_shoot_pct;
+
+  // buffering parameters
+  int64_t starting_buffer_level_ms;
+  int64_t optimal_buffer_level_ms;
+  int64_t maximum_buffer_size_ms;
+
+  // Frame drop threshold.
+  int drop_frames_water_mark;
+
+  // controlling quality
+  int fixed_q;
+  int worst_allowed_q;
+  int best_allowed_q;
+  int cq_level;
+  AQ_MODE aq_mode;  // Adaptive Quantization mode
+
+  // Internal frame size scaling.
+  RESIZE_TYPE resize_mode;
+  int scaled_frame_width;
+  int scaled_frame_height;
+
+  // Enable feature to reduce the frame quantization every x frames.
+  int frame_periodic_boost;
+
+  // two pass datarate control
+  int two_pass_vbrbias;        // two pass datarate control tweaks
+  int two_pass_vbrmin_section;
+  int two_pass_vbrmax_section;
+  // END DATARATE CONTROL OPTIONS
+  // ----------------------------------------------------------------
+
+  // Spatial and temporal scalability.
+  int ss_number_layers;  // Number of spatial layers.
+  int ts_number_layers;  // Number of temporal layers.
+  // Bitrate allocation for spatial layers.
+  int layer_target_bitrate[VPX_MAX_LAYERS];
+  int ss_target_bitrate[VPX_SS_MAX_LAYERS];
+  int ss_enable_auto_arf[VPX_SS_MAX_LAYERS];
+  // Bitrate allocation (CBR mode) and framerate factor, for temporal layers.
+  int ts_rate_decimator[VPX_TS_MAX_LAYERS];
+
+  int enable_auto_arf;
+
+  int encode_breakout;  // early breakout : for video conf recommend 800
+
+  /* Bitfield defining the error resiliency features to enable.
+   * Can provide decodable frames after losses in previous
+   * frames and decodable partitions after losses in the same frame.
+   */
+  unsigned int error_resilient_mode;
+
+  /* Bitfield defining the parallel decoding mode where the
+   * decoding in successive frames may be conducted in parallel
+   * just by decoding the frame headers.
+   */
+  unsigned int frame_parallel_decoding_mode;
+
+  int arnr_max_frames;
+  int arnr_strength;
+
+  int tile_columns;
+  int tile_rows;
+
+  int max_threads;
+
+  vpx_fixed_buf_t two_pass_stats_in;
+  struct vpx_codec_pkt_list *output_pkt_list;
+
+#if CONFIG_FP_MB_STATS
+  vpx_fixed_buf_t firstpass_mb_stats_in;
+#endif
+
+  vp8e_tuning tuning;
+  vp9e_tune_content content;
+#if CONFIG_VP9_HIGHBITDEPTH
+  int use_highbitdepth;
+#endif
+  vpx_color_space_t color_space;
+  VP9E_TEMPORAL_LAYERING_MODE temporal_layering_mode;
+} VP9EncoderConfig;
+
+static INLINE int is_lossless_requested(const VP9EncoderConfig *cfg) {
+  return cfg->best_allowed_q == 0 && cfg->worst_allowed_q == 0;
+}
+
+// TODO(jingning) All spatially adaptive variables should go to TileDataEnc.
+typedef struct TileDataEnc {
+  TileInfo tile_info;
+  int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
+  int mode_map[BLOCK_SIZES][MAX_MODES];
+} TileDataEnc;
+
+typedef struct RD_COUNTS {
+  vp9_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
+  int64_t comp_pred_diff[REFERENCE_MODES];
+  int64_t tx_select_diff[TX_MODES];
+  int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+} RD_COUNTS;
+
+typedef struct ThreadData {
+  MACROBLOCK mb;
+  RD_COUNTS rd_counts;
+  FRAME_COUNTS *counts;
+
+  PICK_MODE_CONTEXT *leaf_tree;
+  PC_TREE *pc_tree;
+  PC_TREE *pc_root;
+} ThreadData;
+
+struct EncWorkerData;
+
+typedef struct ActiveMap {
+  int enabled;
+  int update;
+  unsigned char *map;
+} ActiveMap;
+
+typedef enum {
+  Y,
+  U,
+  V,
+  ALL
+} STAT_TYPE;
+
+typedef struct IMAGE_STAT {
+  double stat[ALL+1];
+  double worst;
+} ImageStat;
+
+typedef struct VP9_COMP {
+  QUANTS quants;
+  ThreadData td;
+  DECLARE_ALIGNED(16, int16_t, y_dequant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_dequant[QINDEX_RANGE][8]);
+  VP9_COMMON common;
+  VP9EncoderConfig oxcf;
+  struct lookahead_ctx    *lookahead;
+  struct lookahead_entry  *alt_ref_source;
+
+  YV12_BUFFER_CONFIG *Source;
+  YV12_BUFFER_CONFIG *Last_Source;  // NULL for first frame and alt_ref frames
+  YV12_BUFFER_CONFIG *un_scaled_source;
+  YV12_BUFFER_CONFIG scaled_source;
+  YV12_BUFFER_CONFIG *unscaled_last_source;
+  YV12_BUFFER_CONFIG scaled_last_source;
+
+  TileDataEnc *tile_data;
+
+  // For a still frame, this flag is set to 1 to skip partition search.
+  int partition_search_skippable_frame;
+
+  int scaled_ref_idx[MAX_REF_FRAMES];
+  int lst_fb_idx;
+  int gld_fb_idx;
+  int alt_fb_idx;
+
+  int refresh_last_frame;
+  int refresh_golden_frame;
+  int refresh_alt_ref_frame;
+
+  int ext_refresh_frame_flags_pending;
+  int ext_refresh_last_frame;
+  int ext_refresh_golden_frame;
+  int ext_refresh_alt_ref_frame;
+
+  int ext_refresh_frame_context_pending;
+  int ext_refresh_frame_context;
+
+  YV12_BUFFER_CONFIG last_frame_uf;
+
+  TOKENEXTRA *tile_tok[4][1 << 6];
+  unsigned int tok_count[4][1 << 6];
+
+  // Ambient reconstruction err target for force key frames
+  int64_t ambient_err;
+
+  RD_OPT rd;
+
+  CODING_CONTEXT coding_context;
+
+  int *nmvcosts[2];
+  int *nmvcosts_hp[2];
+  int *nmvsadcosts[2];
+  int *nmvsadcosts_hp[2];
+
+  int64_t last_time_stamp_seen;
+  int64_t last_end_time_stamp_seen;
+  int64_t first_time_stamp_ever;
+
+  RATE_CONTROL rc;
+  double framerate;
+
+  int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE];
+
+  struct vpx_codec_pkt_list  *output_pkt_list;
+
+  MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
+  int mbgraph_n_frames;             // number of frames filled in the above
+  int static_mb_pct;                // % forced skip mbs by segmentation
+  int ref_frame_flags;
+
+  SPEED_FEATURES sf;
+
+  unsigned int max_mv_magnitude;
+  int mv_step_param;
+
+  int allow_comp_inter_inter;
+
+  // Default value is 1. From first pass stats, encode_breakout may be disabled.
+  ENCODE_BREAKOUT_TYPE allow_encode_breakout;
+
+  // Get threshold from external input. A suggested threshold is 800 for HD
+  // clips, and 300 for < HD clips.
+  int encode_breakout;
+
+  unsigned char *segmentation_map;
+
+  // segment threashold for encode breakout
+  int  segment_encode_breakout[MAX_SEGMENTS];
+
+  CYCLIC_REFRESH *cyclic_refresh;
+  ActiveMap active_map;
+
+  fractional_mv_step_fp *find_fractional_mv_step;
+  vp9_full_search_fn_t full_search_sad;
+  vp9_diamond_search_fn_t diamond_search_sad;
+  vp9_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
+  uint64_t time_receive_data;
+  uint64_t time_compress_data;
+  uint64_t time_pick_lpf;
+  uint64_t time_encode_sb_row;
+
+#if CONFIG_FP_MB_STATS
+  int use_fp_mb_stats;
+#endif
+
+  TWO_PASS twopass;
+
+  YV12_BUFFER_CONFIG alt_ref_buffer;
+
+
+#if CONFIG_INTERNAL_STATS
+  unsigned int mode_chosen_counts[MAX_MODES];
+
+  int    count;
+  uint64_t total_sq_error;
+  uint64_t total_samples;
+  ImageStat psnr;
+
+  uint64_t totalp_sq_error;
+  uint64_t totalp_samples;
+  ImageStat psnrp;
+
+  double total_blockiness;
+  double worst_blockiness;
+
+  int    bytes;
+  double summed_quality;
+  double summed_weights;
+  double summedp_quality;
+  double summedp_weights;
+  unsigned int tot_recode_hits;
+  double worst_ssim;
+
+  ImageStat ssimg;
+  ImageStat fastssim;
+  ImageStat psnrhvs;
+
+  int b_calculate_ssimg;
+  int b_calculate_blockiness;
+
+  int b_calculate_consistency;
+
+  double total_inconsistency;
+  double worst_consistency;
+  Ssimv *ssim_vars;
+  Metrics metrics;
+#endif
+  int b_calculate_psnr;
+
+  int droppable;
+
+  int initial_width;
+  int initial_height;
+  int initial_mbs;  // Number of MBs in the full-size frame; to be used to
+                    // normalize the firstpass stats. This will differ from the
+                    // number of MBs in the current frame when the frame is
+                    // scaled.
+
+  int use_svc;
+
+  SVC svc;
+
+  // Store frame variance info in SOURCE_VAR_BASED_PARTITION search type.
+  diff *source_diff_var;
+  // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
+  unsigned int source_var_thresh;
+  int frames_till_next_var_check;
+
+  int frame_flags;
+
+  search_site_config ss_cfg;
+
+  int mbmode_cost[INTRA_MODES];
+  unsigned int inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
+  int intra_uv_mode_cost[FRAME_TYPES][INTRA_MODES];
+  int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
+  int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
+  int partition_cost[PARTITION_CONTEXTS][PARTITION_TYPES];
+
+  int multi_arf_allowed;
+  int multi_arf_enabled;
+  int multi_arf_last_grp_enabled;
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+  VP9_DENOISER denoiser;
+#endif
+
+  int resize_pending;
+
+  // VAR_BASED_PARTITION thresholds
+  // 0 - threshold_64x64; 1 - threshold_32x32;
+  // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
+  int64_t vbp_thresholds[4];
+  int64_t vbp_threshold_minmax;
+  int64_t vbp_threshold_sad;
+  BLOCK_SIZE vbp_bsize_min;
+
+  // Multi-threading
+  int num_workers;
+  VP9Worker *workers;
+  struct EncWorkerData *tile_thr_data;
+  VP9LfSync lf_row_sync;
+} VP9_COMP;
+
+void vp9_initialize_enc(void);
+
+struct VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
+                                       BufferPool *const pool);
+void vp9_remove_compressor(VP9_COMP *cpi);
+
+void vp9_change_config(VP9_COMP *cpi, const VP9EncoderConfig *oxcf);
+
+  // receive a frames worth of data. caller can assume that a copy of this
+  // frame is made and not just a copy of the pointer..
+int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
+                          YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
+                          int64_t end_time_stamp);
+
+int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
+                            size_t *size, uint8_t *dest,
+                            int64_t *time_stamp, int64_t *time_end, int flush);
+
+int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
+                              vp9_ppflags_t *flags);
+
+int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags);
+
+void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags);
+
+int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
+                           YV12_BUFFER_CONFIG *sd);
+
+int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
+                          YV12_BUFFER_CONFIG *sd);
+
+int vp9_update_entropy(VP9_COMP *cpi, int update);
+
+int vp9_set_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int vp9_get_active_map(VP9_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int vp9_set_internal_size(VP9_COMP *cpi,
+                          VPX_SCALING horiz_mode, VPX_SCALING vert_mode);
+
+int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
+                         unsigned int height);
+
+void vp9_set_svc(VP9_COMP *cpi, int use_svc);
+
+int vp9_get_quantizer(struct VP9_COMP *cpi);
+
+static INLINE int frame_is_kf_gf_arf(const VP9_COMP *cpi) {
+  return frame_is_intra_only(&cpi->common) ||
+         cpi->refresh_alt_ref_frame ||
+         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref);
+}
+
+static INLINE int get_ref_frame_map_idx(const VP9_COMP *cpi,
+                                        MV_REFERENCE_FRAME ref_frame) {
+  if (ref_frame == LAST_FRAME) {
+    return cpi->lst_fb_idx;
+  } else if (ref_frame == GOLDEN_FRAME) {
+    return cpi->gld_fb_idx;
+  } else {
+    return cpi->alt_fb_idx;
+  }
+}
+
+static INLINE int get_ref_frame_buf_idx(const VP9_COMP *const cpi,
+                                        int ref_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
+  return (map_idx != INVALID_IDX) ? cm->ref_frame_map[map_idx] : INVALID_IDX;
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
+    VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+  return
+      buf_idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[buf_idx].buf : NULL;
+}
+
+static INLINE int get_token_alloc(int mb_rows, int mb_cols) {
+  // TODO(JBB): double check we can't exceed this token count if we have a
+  // 32x32 transform crossing a boundary at a multiple of 16.
+  // mb_rows, cols are in units of 16 pixels. We assume 3 planes all at full
+  // resolution. We assume up to 1 token per pixel, and then allow
+  // a head room of 4.
+  return mb_rows * mb_cols * (16 * 16 * 3 + 4);
+}
+
+// Get the allocated token size for a tile. It does the same calculation as in
+// the frame token allocation.
+static INLINE int allocated_tokens(TileInfo tile) {
+  int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 1) >> 1;
+  int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 1) >> 1;
+
+  return get_token_alloc(tile_mb_rows, tile_mb_cols);
+}
+
+int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b);
+#if CONFIG_VP9_HIGHBITDEPTH
+int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_alloc_compressor_data(VP9_COMP *cpi);
+
+void vp9_scale_references(VP9_COMP *cpi);
+
+void vp9_update_reference_frames(VP9_COMP *cpi);
+
+void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv);
+
+YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled);
+
+void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags);
+
+static INLINE int is_two_pass_svc(const struct VP9_COMP *const cpi) {
+  return cpi->use_svc && cpi->oxcf.pass != 0;
+}
+
+static INLINE int is_one_pass_cbr_svc(const struct VP9_COMP *const cpi) {
+  return (cpi->use_svc && cpi->oxcf.pass == 0);
+}
+
+static INLINE int is_altref_enabled(const VP9_COMP *const cpi) {
+  return cpi->oxcf.mode != REALTIME && cpi->oxcf.lag_in_frames > 0 &&
+         (cpi->oxcf.enable_auto_arf &&
+          (!is_two_pass_svc(cpi) ||
+           cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id]));
+}
+
+static INLINE void set_ref_ptrs(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                MV_REFERENCE_FRAME ref0,
+                                MV_REFERENCE_FRAME ref1) {
+  xd->block_refs[0] = &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME
+                                                         : 0];
+  xd->block_refs[1] = &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME
+                                                         : 0];
+}
+
+static INLINE int get_chessboard_index(const int frame_index) {
+  return frame_index & 0x1;
+}
+
+static INLINE int *cond_cost_list(const struct VP9_COMP *cpi, int *cost_list) {
+  return cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL;
+}
+
+void vp9_new_framerate(VP9_COMP *cpi, double framerate);
+
+#define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_ENCODER_H_

media/libvpx/vp9/encoder/vp9_ethread.c

@@ -0,0 +1,179 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_ethread.h"
+
+static void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) {
+  int i, j, k, l, m, n;
+
+  for (i = 0; i < REFERENCE_MODES; i++)
+    td->rd_counts.comp_pred_diff[i] += td_t->rd_counts.comp_pred_diff[i];
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+    td->rd_counts.filter_diff[i] += td_t->rd_counts.filter_diff[i];
+
+  for (i = 0; i < TX_MODES; i++)
+    td->rd_counts.tx_select_diff[i] += td_t->rd_counts.tx_select_diff[i];
+
+  for (i = 0; i < TX_SIZES; i++)
+    for (j = 0; j < PLANE_TYPES; j++)
+      for (k = 0; k < REF_TYPES; k++)
+        for (l = 0; l < COEF_BANDS; l++)
+          for (m = 0; m < COEFF_CONTEXTS; m++)
+            for (n = 0; n < ENTROPY_TOKENS; n++)
+              td->rd_counts.coef_counts[i][j][k][l][m][n] +=
+                  td_t->rd_counts.coef_counts[i][j][k][l][m][n];
+}
+
+static int enc_worker_hook(EncWorkerData *const thread_data, void *unused) {
+  VP9_COMP *const cpi = thread_data->cpi;
+  const VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  int t;
+
+  (void) unused;
+
+  for (t = thread_data->start; t < tile_rows * tile_cols;
+      t += cpi->num_workers) {
+    int tile_row = t / tile_cols;
+    int tile_col = t % tile_cols;
+
+    vp9_encode_tile(cpi, thread_data->td, tile_row, tile_col);
+  }
+
+  return 0;
+}
+
+void vp9_encode_tiles_mt(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+  const int num_workers = MIN(cpi->oxcf.max_threads, tile_cols);
+  int i;
+
+  vp9_init_tile_data(cpi);
+
+  // Only run once to create threads and allocate thread data.
+  if (cpi->num_workers == 0) {
+    CHECK_MEM_ERROR(cm, cpi->workers,
+                    vpx_malloc(num_workers * sizeof(*cpi->workers)));
+
+    CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
+                    vpx_calloc(num_workers, sizeof(*cpi->tile_thr_data)));
+
+    for (i = 0; i < num_workers; i++) {
+      VP9Worker *const worker = &cpi->workers[i];
+      EncWorkerData *thread_data = &cpi->tile_thr_data[i];
+
+      ++cpi->num_workers;
+      winterface->init(worker);
+
+      if (i < num_workers - 1) {
+        thread_data->cpi = cpi;
+
+        // Allocate thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td,
+                        vpx_memalign(32, sizeof(*thread_data->td)));
+        vp9_zero(*thread_data->td);
+
+        // Set up pc_tree.
+        thread_data->td->leaf_tree = NULL;
+        thread_data->td->pc_tree = NULL;
+        vp9_setup_pc_tree(cm, thread_data->td);
+
+        // Allocate frame counters in thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td->counts,
+                        vpx_calloc(1, sizeof(*thread_data->td->counts)));
+
+        // Create threads
+        if (!winterface->reset(worker))
+          vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                             "Tile encoder thread creation failed");
+      } else {
+        // Main thread acts as a worker and uses the thread data in cpi.
+        thread_data->cpi = cpi;
+        thread_data->td = &cpi->td;
+      }
+
+      winterface->sync(worker);
+    }
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    VP9Worker *const worker = &cpi->workers[i];
+    EncWorkerData *thread_data;
+
+    worker->hook = (VP9WorkerHook)enc_worker_hook;
+    worker->data1 = &cpi->tile_thr_data[i];
+    worker->data2 = NULL;
+    thread_data = (EncWorkerData*)worker->data1;
+
+    // Before encoding a frame, copy the thread data from cpi.
+    if (thread_data->td != &cpi->td) {
+      thread_data->td->mb = cpi->td.mb;
+      thread_data->td->rd_counts = cpi->td.rd_counts;
+    }
+    if (thread_data->td->counts != &cpi->common.counts) {
+      memcpy(thread_data->td->counts, &cpi->common.counts,
+             sizeof(cpi->common.counts));
+    }
+
+    // Handle use_nonrd_pick_mode case.
+    if (cpi->sf.use_nonrd_pick_mode) {
+      MACROBLOCK *const x = &thread_data->td->mb;
+      MACROBLOCKD *const xd = &x->e_mbd;
+      struct macroblock_plane *const p = x->plane;
+      struct macroblockd_plane *const pd = xd->plane;
+      PICK_MODE_CONTEXT *ctx = &thread_data->td->pc_root->none;
+      int j;
+
+      for (j = 0; j < MAX_MB_PLANE; ++j) {
+        p[j].coeff = ctx->coeff_pbuf[j][0];
+        p[j].qcoeff = ctx->qcoeff_pbuf[j][0];
+        pd[j].dqcoeff = ctx->dqcoeff_pbuf[j][0];
+        p[j].eobs = ctx->eobs_pbuf[j][0];
+      }
+    }
+  }
+
+  // Encode a frame
+  for (i = 0; i < num_workers; i++) {
+    VP9Worker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+
+    // Set the starting tile for each thread.
+    thread_data->start = i;
+
+    if (i == num_workers - 1)
+      winterface->execute(worker);
+    else
+      winterface->launch(worker);
+  }
+
+  // Encoding ends.
+  for (i = 0; i < num_workers; i++) {
+    VP9Worker *const worker = &cpi->workers[i];
+    winterface->sync(worker);
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    VP9Worker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+
+    // Accumulate counters.
+    if (i < num_workers - 1) {
+      vp9_accumulate_frame_counts(cm, thread_data->td->counts, 0);
+      accumulate_rd_opt(&cpi->td, thread_data->td);
+    }
+  }
+}

media/libvpx/vp9/encoder/vp9_ethread.h

@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_ETHREAD_H_
+#define VP9_ENCODER_VP9_ETHREAD_H_
+
+struct VP9_COMP;
+struct ThreadData;
+
+typedef struct EncWorkerData {
+  struct VP9_COMP *cpi;
+  struct ThreadData *td;
+  int start;
+} EncWorkerData;
+
+void vp9_encode_tiles_mt(struct VP9_COMP *cpi);
+
+#endif  // VP9_ENCODER_VP9_ETHREAD_H_

media/libvpx/vp9/encoder/vp9_extend.c

@@ -0,0 +1,198 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/encoder/vp9_extend.h"
+
+static void copy_and_extend_plane(const uint8_t *src, int src_pitch,
+                                  uint8_t *dst, int dst_pitch,
+                                  int w, int h,
+                                  int extend_top, int extend_left,
+                                  int extend_bottom, int extend_right) {
+  int i, linesize;
+
+  // copy the left and right most columns out
+  const uint8_t *src_ptr1 = src;
+  const uint8_t *src_ptr2 = src + w - 1;
+  uint8_t *dst_ptr1 = dst - extend_left;
+  uint8_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    memset(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w);
+    memset(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h) - extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize);
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize);
+    dst_ptr2 += dst_pitch;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_copy_and_extend_plane(const uint8_t *src8, int src_pitch,
+                                         uint8_t *dst8, int dst_pitch,
+                                         int w, int h,
+                                         int extend_top, int extend_left,
+                                         int extend_bottom, int extend_right) {
+  int i, linesize;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+
+  // copy the left and right most columns out
+  const uint16_t *src_ptr1 = src;
+  const uint16_t *src_ptr2 = src + w - 1;
+  uint16_t *dst_ptr1 = dst - extend_left;
+  uint16_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    vpx_memset16(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w * sizeof(uint16_t));
+    vpx_memset16(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h) - extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize * sizeof(uint16_t));
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize * sizeof(uint16_t));
+    dst_ptr2 += dst_pitch;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst) {
+  // Extend src frame in buffer
+  // Altref filtering assumes 16 pixel extension
+  const int et_y = 16;
+  const int el_y = 16;
+  // Motion estimation may use src block variance with the block size up
+  // to 64x64, so the right and bottom need to be extended to 64 multiple
+  // or up to 16, whichever is greater.
+  const int er_y = MAX(src->y_width + 16, ALIGN_POWER_OF_TWO(src->y_width, 6))
+      - src->y_crop_width;
+  const int eb_y = MAX(src->y_height + 16, ALIGN_POWER_OF_TWO(src->y_height, 6))
+      - src->y_crop_height;
+  const int uv_width_subsampling = (src->uv_width != src->y_width);
+  const int uv_height_subsampling = (src->uv_height != src->y_height);
+  const int et_uv = et_y >> uv_height_subsampling;
+  const int el_uv = el_y >> uv_width_subsampling;
+  const int eb_uv = eb_y >> uv_height_subsampling;
+  const int er_uv = er_y >> uv_width_subsampling;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+    highbd_copy_and_extend_plane(src->y_buffer, src->y_stride,
+                                 dst->y_buffer, dst->y_stride,
+                                 src->y_crop_width, src->y_crop_height,
+                                 et_y, el_y, eb_y, er_y);
+
+    highbd_copy_and_extend_plane(src->u_buffer, src->uv_stride,
+                                 dst->u_buffer, dst->uv_stride,
+                                 src->uv_crop_width, src->uv_crop_height,
+                                 et_uv, el_uv, eb_uv, er_uv);
+
+    highbd_copy_and_extend_plane(src->v_buffer, src->uv_stride,
+                                 dst->v_buffer, dst->uv_stride,
+                                 src->uv_crop_width, src->uv_crop_height,
+                                 et_uv, el_uv, eb_uv, er_uv);
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  copy_and_extend_plane(src->y_buffer, src->y_stride,
+                        dst->y_buffer, dst->y_stride,
+                        src->y_crop_width, src->y_crop_height,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer, src->uv_stride,
+                        dst->u_buffer, dst->uv_stride,
+                        src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer, src->uv_stride,
+                        dst->v_buffer, dst->uv_stride,
+                        src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+}
+
+void vp9_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst,
+                                         int srcy, int srcx,
+                                         int srch, int srcw) {
+  // If the side is not touching the bounder then don't extend.
+  const int et_y = srcy ? 0 : dst->border;
+  const int el_y = srcx ? 0 : dst->border;
+  const int eb_y = srcy + srch != src->y_height ? 0 :
+                      dst->border + dst->y_height - src->y_height;
+  const int er_y = srcx + srcw != src->y_width ? 0 :
+                      dst->border + dst->y_width - src->y_width;
+  const int src_y_offset = srcy * src->y_stride + srcx;
+  const int dst_y_offset = srcy * dst->y_stride + srcx;
+
+  const int et_uv = ROUND_POWER_OF_TWO(et_y, 1);
+  const int el_uv = ROUND_POWER_OF_TWO(el_y, 1);
+  const int eb_uv = ROUND_POWER_OF_TWO(eb_y, 1);
+  const int er_uv = ROUND_POWER_OF_TWO(er_y, 1);
+  const int src_uv_offset = ((srcy * src->uv_stride) >> 1) + (srcx >> 1);
+  const int dst_uv_offset = ((srcy * dst->uv_stride) >> 1) + (srcx >> 1);
+  const int srch_uv = ROUND_POWER_OF_TWO(srch, 1);
+  const int srcw_uv = ROUND_POWER_OF_TWO(srcw, 1);
+
+  copy_and_extend_plane(src->y_buffer + src_y_offset, src->y_stride,
+                        dst->y_buffer + dst_y_offset, dst->y_stride,
+                        srcw, srch,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer + src_uv_offset, src->uv_stride,
+                        dst->u_buffer + dst_uv_offset, dst->uv_stride,
+                        srcw_uv, srch_uv,
+                        et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer + src_uv_offset, src->uv_stride,
+                        dst->v_buffer + dst_uv_offset, dst->uv_stride,
+                        srcw_uv, srch_uv,
+                        et_uv, el_uv, eb_uv, er_uv);
+}

media/libvpx/vp9/encoder/vp9_extend.h

@@ -0,0 +1,33 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_EXTEND_H_
+#define VP9_ENCODER_VP9_EXTEND_H_
+
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+void vp9_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst);
+
+void vp9_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst,
+                                         int srcy, int srcx,
+                                         int srch, int srcw);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_EXTEND_H_

media/libvpx/vp9/encoder/vp9_fastssim.c

@@ -0,0 +1,465 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ *
+ *  This code was originally written by: Nathan E. Egge, at the Daala
+ *  project.
+ */
+#include <math.h>
+#include <string.h>
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/encoder/vp9_ssim.h"
+/* TODO(jbb): High bit depth version of this code needed */
+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)
+
+#define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
+#define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
+
+struct fs_level {
+  uint16_t *im1;
+  uint16_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 = (uint16_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 uint16_t *src1;
+  const uint16_t *src2;
+  uint16_t *dst1;
+  uint16_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 unsigned char *_src1,
+                                 int _s1ystride, const unsigned char *_src2,
+                                 int _s2ystride, int _w, int _h) {
+  uint16_t *dst1;
+  uint16_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);
+      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];
+    }
+  }
+}
+
+static void fs_apply_luminance(fs_ctx *_ctx, int _l) {
+  unsigned *col_sums_x;
+  unsigned *col_sums_y;
+  uint16_t *im1;
+  uint16_t *im2;
+  double *ssim;
+  double c1;
+  int w;
+  int h;
+  int j0offs;
+  int j1offs;
+  int i;
+  int j;
+  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) {
+  uint16_t *im1;
+  uint16_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;
+  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(im1[(j + 1) * w + i + 1] - im1[j * w + i]);
+        g2 = abs(im1[(j + 1) * w + i] - im1[j * w + i + 1]);
+        gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+        g1 = abs(im2[(j + 1) * w + i + 1] - im2[j * w + i]);
+        g2 = abs(im2[(j + 1) * w + i] - 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 calc_ssim(const unsigned char *_src, int _systride,
+                 const unsigned char *_dst, int _dystride, int _w, int _h) {
+  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);
+  for (l = 0; l < FS_NLEVELS - 1; l++) {
+    fs_calc_structure(&ctx, l);
+    ret *= fs_average(&ctx, l);
+    fs_downsample_level(&ctx, l + 1);
+  }
+  fs_calc_structure(&ctx, l);
+  fs_apply_luminance(&ctx, l);
+  ret *= fs_average(&ctx, l);
+  fs_ctx_clear(&ctx);
+  return ret;
+}
+
+static double convert_ssim_db(double _ssim, double _weight) {
+  return 10 * (log10(_weight) - log10(_weight - _ssim));
+}
+
+double vp9_calc_fastssim(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+                         double *ssim_y, double *ssim_u, double *ssim_v) {
+  double ssimv;
+  vp9_clear_system_state();
+
+  *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
+                      dest->y_stride, source->y_crop_width,
+                      source->y_crop_height);
+
+  *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height);
+
+  *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height);
+  ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
+
+  return convert_ssim_db(ssimv, 1.0);
+}

media/libvpx/vp9/encoder/vp9_firstpass.h

@@ -0,0 +1,154 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_FIRSTPASS_H_
+#define VP9_ENCODER_VP9_FIRSTPASS_H_
+
+#include "vp9/encoder/vp9_lookahead.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_FP_MB_STATS
+
+#define FPMB_DCINTRA_MASK 0x01
+
+#define FPMB_MOTION_ZERO_MASK 0x02
+#define FPMB_MOTION_LEFT_MASK 0x04
+#define FPMB_MOTION_RIGHT_MASK 0x08
+#define FPMB_MOTION_UP_MASK 0x10
+#define FPMB_MOTION_DOWN_MASK 0x20
+
+#define FPMB_ERROR_SMALL_MASK 0x40
+#define FPMB_ERROR_LARGE_MASK 0x80
+#define FPMB_ERROR_SMALL_TH 2000
+#define FPMB_ERROR_LARGE_TH 48000
+
+typedef struct {
+  uint8_t *mb_stats_start;
+  uint8_t *mb_stats_end;
+} FIRSTPASS_MB_STATS;
+#endif
+
+#define VLOW_MOTION_THRESHOLD 950
+
+typedef struct {
+  double frame;
+  double weight;
+  double intra_error;
+  double coded_error;
+  double sr_coded_error;
+  double pcnt_inter;
+  double pcnt_motion;
+  double pcnt_second_ref;
+  double pcnt_neutral;
+  double MVr;
+  double mvr_abs;
+  double MVc;
+  double mvc_abs;
+  double MVrv;
+  double MVcv;
+  double mv_in_out_count;
+  double new_mv_count;
+  double duration;
+  double count;
+  int64_t spatial_layer_id;
+} FIRSTPASS_STATS;
+
+typedef enum {
+  KF_UPDATE = 0,
+  LF_UPDATE = 1,
+  GF_UPDATE = 2,
+  ARF_UPDATE = 3,
+  OVERLAY_UPDATE = 4,
+  FRAME_UPDATE_TYPES = 5
+} FRAME_UPDATE_TYPE;
+
+typedef struct {
+  unsigned char index;
+  RATE_FACTOR_LEVEL rf_level[(MAX_LAG_BUFFERS * 2) + 1];
+  FRAME_UPDATE_TYPE update_type[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_src_offset[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_update_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_ref_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  int bit_allocation[(MAX_LAG_BUFFERS * 2) + 1];
+} GF_GROUP;
+
+typedef struct {
+  unsigned int section_intra_rating;
+  FIRSTPASS_STATS total_stats;
+  FIRSTPASS_STATS this_frame_stats;
+  const FIRSTPASS_STATS *stats_in;
+  const FIRSTPASS_STATS *stats_in_start;
+  const FIRSTPASS_STATS *stats_in_end;
+  FIRSTPASS_STATS total_left_stats;
+  int first_pass_done;
+  int64_t bits_left;
+  double modified_error_min;
+  double modified_error_max;
+  double modified_error_left;
+  double mb_av_energy;
+
+#if CONFIG_FP_MB_STATS
+  uint8_t *frame_mb_stats_buf;
+  uint8_t *this_frame_mb_stats;
+  FIRSTPASS_MB_STATS firstpass_mb_stats;
+#endif
+
+  // Projected total bits available for a key frame group of frames
+  int64_t kf_group_bits;
+
+  // Error score of frames still to be coded in kf group
+  int64_t kf_group_error_left;
+
+  // The fraction for a kf groups total bits allocated to the inter frames
+  double kfgroup_inter_fraction;
+
+  int sr_update_lag;
+
+  int kf_zeromotion_pct;
+  int last_kfgroup_zeromotion_pct;
+  int gf_zeromotion_pct;
+  int active_worst_quality;
+  int baseline_active_worst_quality;
+  int extend_minq;
+  int extend_maxq;
+  int extend_minq_fast;
+
+  GF_GROUP gf_group;
+} TWO_PASS;
+
+struct VP9_COMP;
+
+void vp9_init_first_pass(struct VP9_COMP *cpi);
+void vp9_rc_get_first_pass_params(struct VP9_COMP *cpi);
+void vp9_first_pass(struct VP9_COMP *cpi, const struct lookahead_entry *source);
+void vp9_end_first_pass(struct VP9_COMP *cpi);
+
+void vp9_init_second_pass(struct VP9_COMP *cpi);
+void vp9_rc_get_second_pass_params(struct VP9_COMP *cpi);
+void vp9_twopass_postencode_update(struct VP9_COMP *cpi);
+
+// Post encode update of the rate control parameters for 2-pass
+void vp9_twopass_postencode_update(struct VP9_COMP *cpi);
+
+void vp9_init_subsampling(struct VP9_COMP *cpi);
+
+void calculate_coded_size(struct VP9_COMP *cpi,
+                          int *scaled_frame_width,
+                          int *scaled_frame_height);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_FIRSTPASS_H_

media/libvpx/vp9/encoder/vp9_lookahead.c

@@ -0,0 +1,245 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include <assert.h>
+#include <stdlib.h>
+
+#include "./vpx_config.h"
+
+#include "vp9/common/vp9_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_extend.h"
+#include "vp9/encoder/vp9_lookahead.h"
+
+/* Return the buffer at the given absolute index and increment the index */
+static struct lookahead_entry *pop(struct lookahead_ctx *ctx,
+                                   unsigned int *idx) {
+  unsigned int index = *idx;
+  struct lookahead_entry *buf = ctx->buf + index;
+
+  assert(index < ctx->max_sz);
+  if (++index >= ctx->max_sz)
+    index -= ctx->max_sz;
+  *idx = index;
+  return buf;
+}
+
+
+void vp9_lookahead_destroy(struct lookahead_ctx *ctx) {
+  if (ctx) {
+    if (ctx->buf) {
+      unsigned int i;
+
+      for (i = 0; i < ctx->max_sz; i++)
+        vp9_free_frame_buffer(&ctx->buf[i].img);
+      free(ctx->buf);
+    }
+    free(ctx);
+  }
+}
+
+
+struct lookahead_ctx *vp9_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth) {
+  struct lookahead_ctx *ctx = NULL;
+
+  // Clamp the lookahead queue depth
+  depth = clamp(depth, 1, MAX_LAG_BUFFERS);
+
+  // Allocate memory to keep previous source frames available.
+  depth += MAX_PRE_FRAMES;
+
+  // Allocate the lookahead structures
+  ctx = calloc(1, sizeof(*ctx));
+  if (ctx) {
+    const int legacy_byte_alignment = 0;
+    unsigned int i;
+    ctx->max_sz = depth;
+    ctx->buf = calloc(depth, sizeof(*ctx->buf));
+    if (!ctx->buf)
+      goto bail;
+    for (i = 0; i < depth; i++)
+      if (vp9_alloc_frame_buffer(&ctx->buf[i].img,
+                                 width, height, subsampling_x, subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 legacy_byte_alignment))
+        goto bail;
+  }
+  return ctx;
+ bail:
+  vp9_lookahead_destroy(ctx);
+  return NULL;
+}
+
+#define USE_PARTIAL_COPY 0
+
+int vp9_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG   *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       unsigned int flags) {
+  struct lookahead_entry *buf;
+#if USE_PARTIAL_COPY
+  int row, col, active_end;
+  int mb_rows = (src->y_height + 15) >> 4;
+  int mb_cols = (src->y_width + 15) >> 4;
+#endif
+  int width = src->y_crop_width;
+  int height = src->y_crop_height;
+  int uv_width = src->uv_crop_width;
+  int uv_height = src->uv_crop_height;
+  int subsampling_x = src->subsampling_x;
+  int subsampling_y = src->subsampling_y;
+  int larger_dimensions, new_dimensions;
+
+  if (ctx->sz + 1  + MAX_PRE_FRAMES > ctx->max_sz)
+    return 1;
+  ctx->sz++;
+  buf = pop(ctx, &ctx->write_idx);
+
+  new_dimensions = width != buf->img.y_crop_width ||
+                   height != buf->img.y_crop_height ||
+                   uv_width != buf->img.uv_crop_width ||
+                   uv_height != buf->img.uv_crop_height;
+  larger_dimensions = width > buf->img.y_width ||
+                      height > buf->img.y_height ||
+                      uv_width > buf->img.uv_width ||
+                      uv_height > buf->img.uv_height;
+  assert(!larger_dimensions || new_dimensions);
+
+#if USE_PARTIAL_COPY
+  // TODO(jkoleszar): This is disabled for now, as
+  // vp9_copy_and_extend_frame_with_rect is not subsampling/alpha aware.
+
+  // Only do this partial copy if the following conditions are all met:
+  // 1. Lookahead queue has has size of 1.
+  // 2. Active map is provided.
+  // 3. This is not a key frame, golden nor altref frame.
+  if (!new_dimensions && ctx->max_sz == 1 && active_map && !flags) {
+    for (row = 0; row < mb_rows; ++row) {
+      col = 0;
+
+      while (1) {
+        // Find the first active macroblock in this row.
+        for (; col < mb_cols; ++col) {
+          if (active_map[col])
+            break;
+        }
+
+        // No more active macroblock in this row.
+        if (col == mb_cols)
+          break;
+
+        // Find the end of active region in this row.
+        active_end = col;
+
+        for (; active_end < mb_cols; ++active_end) {
+          if (!active_map[active_end])
+            break;
+        }
+
+        // Only copy this active region.
+        vp9_copy_and_extend_frame_with_rect(src, &buf->img,
+                                            row << 4,
+                                            col << 4, 16,
+                                            (active_end - col) << 4);
+
+        // Start again from the end of this active region.
+        col = active_end;
+      }
+
+      active_map += mb_cols;
+    }
+  } else {
+#endif
+    if (larger_dimensions) {
+      YV12_BUFFER_CONFIG new_img;
+      memset(&new_img, 0, sizeof(new_img));
+      if (vp9_alloc_frame_buffer(&new_img,
+                                 width, height, subsampling_x, subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 0))
+          return 1;
+      vp9_free_frame_buffer(&buf->img);
+      buf->img = new_img;
+    } else if (new_dimensions) {
+      buf->img.y_crop_width = src->y_crop_width;
+      buf->img.y_crop_height = src->y_crop_height;
+      buf->img.uv_crop_width = src->uv_crop_width;
+      buf->img.uv_crop_height = src->uv_crop_height;
+      buf->img.subsampling_x = src->subsampling_x;
+      buf->img.subsampling_y = src->subsampling_y;
+    }
+    // Partial copy not implemented yet
+    vp9_copy_and_extend_frame(src, &buf->img);
+#if USE_PARTIAL_COPY
+  }
+#endif
+
+  buf->ts_start = ts_start;
+  buf->ts_end = ts_end;
+  buf->flags = flags;
+  return 0;
+}
+
+
+struct lookahead_entry *vp9_lookahead_pop(struct lookahead_ctx *ctx,
+                                          int drain) {
+  struct lookahead_entry *buf = NULL;
+
+  if (ctx && ctx->sz && (drain || ctx->sz == ctx->max_sz - MAX_PRE_FRAMES)) {
+    buf = pop(ctx, &ctx->read_idx);
+    ctx->sz--;
+  }
+  return buf;
+}
+
+
+struct lookahead_entry *vp9_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index) {
+  struct lookahead_entry *buf = NULL;
+
+  if (index >= 0) {
+    // Forward peek
+    if (index < (int)ctx->sz) {
+      index += ctx->read_idx;
+      if (index >= (int)ctx->max_sz)
+        index -= ctx->max_sz;
+      buf = ctx->buf + index;
+    }
+  } else if (index < 0) {
+    // Backward peek
+    if (-index <= MAX_PRE_FRAMES) {
+      index += ctx->read_idx;
+      if (index < 0)
+        index += ctx->max_sz;
+      buf = ctx->buf + index;
+    }
+  }
+
+  return buf;
+}
+
+unsigned int vp9_lookahead_depth(struct lookahead_ctx *ctx) {
+  return ctx->sz;
+}

media/libvpx/vp9/encoder/vp9_lookahead.h

@@ -0,0 +1,124 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_LOOKAHEAD_H_
+#define VP9_ENCODER_VP9_LOOKAHEAD_H_
+
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+
+#if CONFIG_SPATIAL_SVC
+#include "vpx/vp8cx.h"
+#include "vpx/vpx_encoder.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LAG_BUFFERS 25
+
+struct lookahead_entry {
+  YV12_BUFFER_CONFIG  img;
+  int64_t             ts_start;
+  int64_t             ts_end;
+  unsigned int        flags;
+};
+
+// The max of past frames we want to keep in the queue.
+#define MAX_PRE_FRAMES 1
+
+struct lookahead_ctx {
+  unsigned int max_sz;         /* Absolute size of the queue */
+  unsigned int sz;             /* Number of buffers currently in the queue */
+  unsigned int read_idx;       /* Read index */
+  unsigned int write_idx;      /* Write index */
+  struct lookahead_entry *buf; /* Buffer list */
+};
+
+/**\brief Initializes the lookahead stage
+ *
+ * The lookahead stage is a queue of frame buffers on which some analysis
+ * may be done when buffers are enqueued.
+ */
+struct lookahead_ctx *vp9_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth);
+
+
+/**\brief Destroys the lookahead stage
+ */
+void vp9_lookahead_destroy(struct lookahead_ctx *ctx);
+
+
+/**\brief Enqueue a source buffer
+ *
+ * This function will copy the source image into a new framebuffer with
+ * the expected stride/border.
+ *
+ * If active_map is non-NULL and there is only one frame in the queue, then copy
+ * only active macroblocks.
+ *
+ * \param[in] ctx         Pointer to the lookahead context
+ * \param[in] src         Pointer to the image to enqueue
+ * \param[in] ts_start    Timestamp for the start of this frame
+ * \param[in] ts_end      Timestamp for the end of this frame
+ * \param[in] flags       Flags set on this frame
+ * \param[in] active_map  Map that specifies which macroblock is active
+ */
+int vp9_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_VP9_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       unsigned int flags);
+
+
+/**\brief Get the next source buffer to encode
+ *
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] drain     Flag indicating the buffer should be drained
+ *                      (return a buffer regardless of the current queue depth)
+ *
+ * \retval NULL, if drain set and queue is empty
+ * \retval NULL, if drain not set and queue not of the configured depth
+ */
+struct lookahead_entry *vp9_lookahead_pop(struct lookahead_ctx *ctx,
+                                          int drain);
+
+
+/**\brief Get a future source buffer to encode
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] index     Index of the frame to be returned, 0 == next frame
+ *
+ * \retval NULL, if no buffer exists at the specified index
+ */
+struct lookahead_entry *vp9_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index);
+
+
+/**\brief Get the number of frames currently in the lookahead queue
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ */
+unsigned int vp9_lookahead_depth(struct lookahead_ctx *ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_LOOKAHEAD_H_

media/libvpx/vp9/encoder/vp9_mbgraph.c

@@ -0,0 +1,416 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+
+static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
+                                              const MV *ref_mv,
+                                              MV *dst_mv,
+                                              int mb_row,
+                                              int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
+
+  const int tmp_col_min = x->mv_col_min;
+  const int tmp_col_max = x->mv_col_max;
+  const int tmp_row_min = x->mv_row_min;
+  const int tmp_row_max = x->mv_row_max;
+  MV ref_full;
+  int cost_list[5];
+
+  // Further step/diamond searches as necessary
+  int step_param = mv_sf->reduce_first_step_size;
+  step_param = MIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  vp9_set_mv_search_range(x, ref_mv);
+
+  ref_full.col = ref_mv->col >> 3;
+  ref_full.row = ref_mv->row >> 3;
+
+  /*cpi->sf.search_method == HEX*/
+  vp9_hex_search(x, &ref_full, step_param, x->errorperbit, 0,
+                 cond_cost_list(cpi, cost_list),
+                 &v_fn_ptr, 0, ref_mv, dst_mv);
+
+  // Try sub-pixel MC
+  // if (bestsme > error_thresh && bestsme < INT_MAX)
+  {
+    int distortion;
+    unsigned int sse;
+    cpi->find_fractional_mv_step(
+        x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
+        &v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
+        cond_cost_list(cpi, cost_list),
+        NULL, NULL,
+        &distortion, &sse, NULL, 0, 0);
+  }
+
+  xd->mi[0]->mbmi.mode = NEWMV;
+  xd->mi[0]->mbmi.mv[0].as_mv = *dst_mv;
+
+  vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
+
+  /* restore UMV window */
+  x->mv_col_min = tmp_col_min;
+  x->mv_col_max = tmp_col_max;
+  x->mv_row_min = tmp_row_min;
+  x->mv_row_max = tmp_row_max;
+
+  return vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                      xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+}
+
+static int do_16x16_motion_search(VP9_COMP *cpi, const MV *ref_mv,
+                                  int_mv *dst_mv, int mb_row, int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err, tmp_err;
+  MV tmp_mv;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+  dst_mv->as_int = 0;
+
+  // Test last reference frame using the previous best mv as the
+  // starting point (best reference) for the search
+  tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
+  if (tmp_err < err) {
+    err = tmp_err;
+    dst_mv->as_mv = tmp_mv;
+  }
+
+  // If the current best reference mv is not centered on 0,0 then do a 0,0
+  // based search as well.
+  if (ref_mv->row != 0 || ref_mv->col != 0) {
+    unsigned int tmp_err;
+    MV zero_ref_mv = {0, 0}, tmp_mv;
+
+    tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv,
+                                        mb_row, mb_col);
+    if (tmp_err < err) {
+      dst_mv->as_mv = tmp_mv;
+      err = tmp_err;
+    }
+  }
+
+  return err;
+}
+
+static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+
+  dst_mv->as_int = 0;
+
+  return err;
+}
+static int find_best_16x16_intra(VP9_COMP *cpi, PREDICTION_MODE *pbest_mode) {
+  MACROBLOCK   *const x  = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  PREDICTION_MODE best_mode = -1, mode;
+  unsigned int best_err = INT_MAX;
+
+  // calculate SATD for each intra prediction mode;
+  // we're intentionally not doing 4x4, we just want a rough estimate
+  for (mode = DC_PRED; mode <= TM_PRED; mode++) {
+    unsigned int err;
+
+    xd->mi[0]->mbmi.mode = mode;
+    vp9_predict_intra_block(xd, 0, 2, TX_16X16, mode,
+                            x->plane[0].src.buf, x->plane[0].src.stride,
+                            xd->plane[0].dst.buf, xd->plane[0].dst.stride,
+                            0, 0, 0);
+    err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                       xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+
+    // find best
+    if (err < best_err) {
+      best_err  = err;
+      best_mode = mode;
+    }
+  }
+
+  if (pbest_mode)
+    *pbest_mode = best_mode;
+
+  return best_err;
+}
+
+static void update_mbgraph_mb_stats
+(
+  VP9_COMP *cpi,
+  MBGRAPH_MB_STATS *stats,
+  YV12_BUFFER_CONFIG *buf,
+  int mb_y_offset,
+  YV12_BUFFER_CONFIG *golden_ref,
+  const MV *prev_golden_ref_mv,
+  YV12_BUFFER_CONFIG *alt_ref,
+  int mb_row,
+  int mb_col
+) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int intra_error;
+  VP9_COMMON *cm = &cpi->common;
+
+  // FIXME in practice we're completely ignoring chroma here
+  x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
+  x->plane[0].src.stride = buf->y_stride;
+
+  xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
+  xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
+
+  // do intra 16x16 prediction
+  intra_error = find_best_16x16_intra(cpi,
+                                      &stats->ref[INTRA_FRAME].m.mode);
+  if (intra_error <= 0)
+    intra_error = 1;
+  stats->ref[INTRA_FRAME].err = intra_error;
+
+  // Golden frame MV search, if it exists and is different than last frame
+  if (golden_ref) {
+    int g_motion_error;
+    xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = golden_ref->y_stride;
+    g_motion_error = do_16x16_motion_search(cpi,
+                                            prev_golden_ref_mv,
+                                            &stats->ref[GOLDEN_FRAME].m.mv,
+                                            mb_row, mb_col);
+    stats->ref[GOLDEN_FRAME].err = g_motion_error;
+  } else {
+    stats->ref[GOLDEN_FRAME].err = INT_MAX;
+    stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
+  }
+
+  // Do an Alt-ref frame MV search, if it exists and is different than
+  // last/golden frame.
+  if (alt_ref) {
+    int a_motion_error;
+    xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = alt_ref->y_stride;
+    a_motion_error = do_16x16_zerozero_search(cpi,
+                                              &stats->ref[ALTREF_FRAME].m.mv);
+
+    stats->ref[ALTREF_FRAME].err = a_motion_error;
+  } else {
+    stats->ref[ALTREF_FRAME].err = INT_MAX;
+    stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
+  }
+}
+
+static void update_mbgraph_frame_stats(VP9_COMP *cpi,
+                                       MBGRAPH_FRAME_STATS *stats,
+                                       YV12_BUFFER_CONFIG *buf,
+                                       YV12_BUFFER_CONFIG *golden_ref,
+                                       YV12_BUFFER_CONFIG *alt_ref) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  VP9_COMMON *const cm = &cpi->common;
+
+  int mb_col, mb_row, offset = 0;
+  int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
+  MV gld_top_mv = {0, 0};
+  MODE_INFO mi_local;
+
+  vp9_zero(mi_local);
+  // Set up limit values for motion vectors to prevent them extending outside
+  // the UMV borders.
+  x->mv_row_min     = -BORDER_MV_PIXELS_B16;
+  x->mv_row_max     = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
+  xd->up_available  = 0;
+  xd->plane[0].dst.stride  = buf->y_stride;
+  xd->plane[0].pre[0].stride  = buf->y_stride;
+  xd->plane[1].dst.stride = buf->uv_stride;
+  xd->mi[0] = &mi_local;
+  mi_local.mbmi.sb_type = BLOCK_16X16;
+  mi_local.mbmi.ref_frame[0] = LAST_FRAME;
+  mi_local.mbmi.ref_frame[1] = NONE;
+
+  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
+    MV gld_left_mv = gld_top_mv;
+    int mb_y_in_offset  = mb_y_offset;
+    int arf_y_in_offset = arf_y_offset;
+    int gld_y_in_offset = gld_y_offset;
+
+    // Set up limit values for motion vectors to prevent them extending outside
+    // the UMV borders.
+    x->mv_col_min      = -BORDER_MV_PIXELS_B16;
+    x->mv_col_max      = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
+    xd->left_available = 0;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+      MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
+
+      update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
+                              golden_ref, &gld_left_mv, alt_ref,
+                              mb_row, mb_col);
+      gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
+      if (mb_col == 0) {
+        gld_top_mv = gld_left_mv;
+      }
+      xd->left_available = 1;
+      mb_y_in_offset    += 16;
+      gld_y_in_offset   += 16;
+      arf_y_in_offset   += 16;
+      x->mv_col_min     -= 16;
+      x->mv_col_max     -= 16;
+    }
+    xd->up_available = 1;
+    mb_y_offset     += buf->y_stride * 16;
+    gld_y_offset    += golden_ref->y_stride * 16;
+    if (alt_ref)
+      arf_y_offset    += alt_ref->y_stride * 16;
+    x->mv_row_min   -= 16;
+    x->mv_row_max   -= 16;
+    offset          += cm->mb_cols;
+  }
+}
+
+// void separate_arf_mbs_byzz
+static void separate_arf_mbs(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int mb_col, mb_row, offset, i;
+  int mi_row, mi_col;
+  int ncnt[4] = { 0 };
+  int n_frames = cpi->mbgraph_n_frames;
+
+  int *arf_not_zz;
+
+  CHECK_MEM_ERROR(cm, arf_not_zz,
+                  vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz),
+                             1));
+
+  // We are not interested in results beyond the alt ref itself.
+  if (n_frames > cpi->rc.frames_till_gf_update_due)
+    n_frames = cpi->rc.frames_till_gf_update_due;
+
+  // defer cost to reference frames
+  for (i = n_frames - 1; i >= 0; i--) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+
+    for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
+         offset += cm->mb_cols, mb_row++) {
+      for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+        MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
+
+        int altref_err = mb_stats->ref[ALTREF_FRAME].err;
+        int intra_err  = mb_stats->ref[INTRA_FRAME ].err;
+        int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
+
+        // Test for altref vs intra and gf and that its mv was 0,0.
+        if (altref_err > 1000 ||
+            altref_err > intra_err ||
+            altref_err > golden_err) {
+          arf_not_zz[offset + mb_col]++;
+        }
+      }
+    }
+  }
+
+  // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
+  // of bound access in segmentation_map
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      // If any of the blocks in the sequence failed then the MB
+      // goes in segment 0
+      if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
+        ncnt[0]++;
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
+      } else {
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
+        ncnt[1]++;
+      }
+    }
+  }
+
+  // Only bother with segmentation if over 10% of the MBs in static segment
+  // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
+  if (1) {
+    // Note % of blocks that are marked as static
+    if (cm->MBs)
+      cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
+
+    // This error case should not be reachable as this function should
+    // never be called with the common data structure uninitialized.
+    else
+      cpi->static_mb_pct = 0;
+
+    vp9_enable_segmentation(&cm->seg);
+  } else {
+    cpi->static_mb_pct = 0;
+    vp9_disable_segmentation(&cm->seg);
+  }
+
+  // Free localy allocated storage
+  vpx_free(arf_not_zz);
+}
+
+void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
+  YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+
+  assert(golden_ref != NULL);
+
+  // we need to look ahead beyond where the ARF transitions into
+  // being a GF - so exit if we don't look ahead beyond that
+  if (n_frames <= cpi->rc.frames_till_gf_update_due)
+    return;
+
+  if (n_frames > MAX_LAG_BUFFERS)
+    n_frames = MAX_LAG_BUFFERS;
+
+  cpi->mbgraph_n_frames = n_frames;
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    memset(frame_stats->mb_stats, 0,
+           cm->mb_rows * cm->mb_cols * sizeof(*cpi->mbgraph_stats[i].mb_stats));
+  }
+
+  // do motion search to find contribution of each reference to data
+  // later on in this GF group
+  // FIXME really, the GF/last MC search should be done forward, and
+  // the ARF MC search backwards, to get optimal results for MV caching
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
+
+    assert(q_cur != NULL);
+
+    update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img,
+                               golden_ref, cpi->Source);
+  }
+
+  vp9_clear_system_state();
+
+  separate_arf_mbs(cpi);
+}

media/libvpx/vp9/encoder/vp9_mbgraph.h

@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_MBGRAPH_H_
+#define VP9_ENCODER_VP9_MBGRAPH_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  struct {
+    int err;
+    union {
+      int_mv mv;
+      PREDICTION_MODE mode;
+    } m;
+  } ref[MAX_REF_FRAMES];
+} MBGRAPH_MB_STATS;
+
+typedef struct {
+  MBGRAPH_MB_STATS *mb_stats;
+} MBGRAPH_FRAME_STATS;
+
+struct VP9_COMP;
+
+void vp9_update_mbgraph_stats(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_MBGRAPH_H_

media/libvpx/vp9/encoder/vp9_mcomp.h

@@ -0,0 +1,165 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_MCOMP_H_
+#define VP9_ENCODER_VP9_MCOMP_H_
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_variance.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The maximum number of steps in a step search given the largest
+// allowed initial step
+#define MAX_MVSEARCH_STEPS 11
+// Max full pel mv specified in the unit of full pixel
+// Enable the use of motion vector in range [-1023, 1023].
+#define MAX_FULL_PEL_VAL ((1 << (MAX_MVSEARCH_STEPS - 1)) - 1)
+// Maximum size of the first step in full pel units
+#define MAX_FIRST_STEP (1 << (MAX_MVSEARCH_STEPS-1))
+// Allowed motion vector pixel distance outside image border
+// for Block_16x16
+#define BORDER_MV_PIXELS_B16 (16 + VP9_INTERP_EXTEND)
+
+// motion search site
+typedef struct search_site {
+  MV mv;
+  int offset;
+} search_site;
+
+typedef struct search_site_config {
+  search_site ss[8 * MAX_MVSEARCH_STEPS + 1];
+  int ss_count;
+  int searches_per_step;
+} search_site_config;
+
+void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride);
+void vp9_init3smotion_compensation(search_site_config *cfg,  int stride);
+
+void vp9_set_mv_search_range(MACROBLOCK *x, const MV *mv);
+int vp9_mv_bit_cost(const MV *mv, const MV *ref,
+                    const int *mvjcost, int *mvcost[2], int weight);
+
+// Utility to compute variance + MV rate cost for a given MV
+int vp9_get_mvpred_var(const MACROBLOCK *x,
+                       const MV *best_mv, const MV *center_mv,
+                       const vp9_variance_fn_ptr_t *vfp,
+                       int use_mvcost);
+int vp9_get_mvpred_av_var(const MACROBLOCK *x,
+                          const MV *best_mv, const MV *center_mv,
+                          const uint8_t *second_pred,
+                          const vp9_variance_fn_ptr_t *vfp,
+                          int use_mvcost);
+
+struct VP9_COMP;
+struct SPEED_FEATURES;
+
+int vp9_init_search_range(int size);
+
+int vp9_refining_search_sad(const struct macroblock *x,
+                            struct mv *ref_mv,
+                            int sad_per_bit, int distance,
+                            const struct vp9_variance_vtable *fn_ptr,
+                            const struct mv *center_mv);
+
+// Runs sequence of diamond searches in smaller steps for RD.
+int vp9_full_pixel_diamond(const struct VP9_COMP *cpi, MACROBLOCK *x,
+                           MV *mvp_full, int step_param,
+                           int sadpb, int further_steps, int do_refine,
+                           int *cost_list,
+                           const vp9_variance_fn_ptr_t *fn_ptr,
+                           const MV *ref_mv, MV *dst_mv);
+
+// Perform integral projection based motion estimation.
+unsigned int vp9_int_pro_motion_estimation(const struct VP9_COMP *cpi,
+                                           MACROBLOCK *x,
+                                           BLOCK_SIZE bsize,
+                                           int mi_row, int mi_col);
+
+typedef int (integer_mv_pattern_search_fn) (
+    const MACROBLOCK *x,
+    MV *ref_mv,
+    int search_param,
+    int error_per_bit,
+    int do_init_search,
+    int *cost_list,
+    const vp9_variance_fn_ptr_t *vf,
+    int use_mvcost,
+    const MV *center_mv,
+    MV *best_mv);
+
+integer_mv_pattern_search_fn vp9_hex_search;
+integer_mv_pattern_search_fn vp9_bigdia_search;
+integer_mv_pattern_search_fn vp9_square_search;
+integer_mv_pattern_search_fn vp9_fast_hex_search;
+integer_mv_pattern_search_fn vp9_fast_dia_search;
+
+typedef int (fractional_mv_step_fp) (
+    const MACROBLOCK *x,
+    MV *bestmv, const MV *ref_mv,
+    int allow_hp,
+    int error_per_bit,
+    const vp9_variance_fn_ptr_t *vfp,
+    int forced_stop,  // 0 - full, 1 - qtr only, 2 - half only
+    int iters_per_step,
+    int *cost_list,
+    int *mvjcost, int *mvcost[2],
+    int *distortion, unsigned int *sse1,
+    const uint8_t *second_pred,
+    int w, int h);
+
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree;
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree_pruned;
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree_pruned_more;
+extern fractional_mv_step_fp vp9_find_best_sub_pixel_tree_pruned_evenmore;
+
+typedef int (*vp9_full_search_fn_t)(const MACROBLOCK *x,
+                                    const MV *ref_mv, int sad_per_bit,
+                                    int distance,
+                                    const vp9_variance_fn_ptr_t *fn_ptr,
+                                    const MV *center_mv, MV *best_mv);
+
+typedef int (*vp9_refining_search_fn_t)(const MACROBLOCK *x,
+                                        MV *ref_mv, int sad_per_bit,
+                                        int distance,
+                                        const vp9_variance_fn_ptr_t *fn_ptr,
+                                        const MV *center_mv);
+
+typedef int (*vp9_diamond_search_fn_t)(const MACROBLOCK *x,
+                                       const search_site_config *cfg,
+                                       MV *ref_mv, MV *best_mv,
+                                       int search_param, int sad_per_bit,
+                                       int *num00,
+                                       const vp9_variance_fn_ptr_t *fn_ptr,
+                                       const MV *center_mv);
+
+int vp9_refining_search_8p_c(const MACROBLOCK *x,
+                             MV *ref_mv, int error_per_bit,
+                             int search_range,
+                             const vp9_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv, const uint8_t *second_pred);
+
+struct VP9_COMP;
+
+int vp9_full_pixel_search(struct VP9_COMP *cpi, MACROBLOCK *x,
+                          BLOCK_SIZE bsize, MV *mvp_full,
+                          int step_param, int error_per_bit,
+                          int *cost_list,
+                          const MV *ref_mv, MV *tmp_mv,
+                          int var_max, int rd);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_MCOMP_H_

media/libvpx/vp9/encoder/vp9_picklpf.c

@@ -0,0 +1,192 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+
+#include "./vpx_scale_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_quant_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_picklpf.h"
+#include "vp9/encoder/vp9_quantize.h"
+
+static int get_max_filter_level(const VP9_COMP *cpi) {
+  if (cpi->oxcf.pass == 2) {
+    return cpi->twopass.section_intra_rating > 8 ? MAX_LOOP_FILTER * 3 / 4
+                                                 : MAX_LOOP_FILTER;
+  } else {
+    return MAX_LOOP_FILTER;
+  }
+}
+
+
+static int64_t try_filter_frame(const YV12_BUFFER_CONFIG *sd,
+                                VP9_COMP *const cpi,
+                                int filt_level, int partial_frame) {
+  VP9_COMMON *const cm = &cpi->common;
+  int64_t filt_err;
+
+  if (cpi->num_workers > 1)
+    vp9_loop_filter_frame_mt(cm->frame_to_show, cm, cpi->td.mb.e_mbd.plane,
+                             filt_level, 1, partial_frame,
+                             cpi->workers, cpi->num_workers, &cpi->lf_row_sync);
+  else
+    vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level,
+                          1, partial_frame);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    filt_err = vp9_highbd_get_y_sse(sd, cm->frame_to_show);
+  } else {
+    filt_err = vp9_get_y_sse(sd, cm->frame_to_show);
+  }
+#else
+  filt_err = vp9_get_y_sse(sd, cm->frame_to_show);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+  // Re-instate the unfiltered frame
+  vpx_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
+
+  return filt_err;
+}
+
+static int search_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
+                               int partial_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const struct loopfilter *const lf = &cm->lf;
+  const int min_filter_level = 0;
+  const int max_filter_level = get_max_filter_level(cpi);
+  int filt_direction = 0;
+  int64_t best_err;
+  int filt_best;
+
+  // Start the search at the previous frame filter level unless it is now out of
+  // range.
+  int filt_mid = clamp(lf->filter_level, min_filter_level, max_filter_level);
+  int filter_step = filt_mid < 16 ? 4 : filt_mid / 4;
+  // Sum squared error at each filter level
+  int64_t ss_err[MAX_LOOP_FILTER + 1];
+
+  // Set each entry to -1
+  memset(ss_err, 0xFF, sizeof(ss_err));
+
+  //  Make a copy of the unfiltered / processed recon buffer
+  vpx_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
+
+  best_err = try_filter_frame(sd, cpi, filt_mid, partial_frame);
+  filt_best = filt_mid;
+  ss_err[filt_mid] = best_err;
+
+  while (filter_step > 0) {
+    const int filt_high = MIN(filt_mid + filter_step, max_filter_level);
+    const int filt_low = MAX(filt_mid - filter_step, min_filter_level);
+
+    // Bias against raising loop filter in favor of lowering it.
+    int64_t bias = (best_err >> (15 - (filt_mid / 8))) * filter_step;
+
+    if ((cpi->oxcf.pass == 2) && (cpi->twopass.section_intra_rating < 20))
+      bias = (bias * cpi->twopass.section_intra_rating) / 20;
+
+    // yx, bias less for large block size
+    if (cm->tx_mode != ONLY_4X4)
+      bias >>= 1;
+
+    if (filt_direction <= 0 && filt_low != filt_mid) {
+      // Get Low filter error score
+      if (ss_err[filt_low] < 0) {
+        ss_err[filt_low] = try_filter_frame(sd, cpi, filt_low, partial_frame);
+      }
+      // If value is close to the best so far then bias towards a lower loop
+      // filter value.
+      if ((ss_err[filt_low] - bias) < best_err) {
+        // Was it actually better than the previous best?
+        if (ss_err[filt_low] < best_err)
+          best_err = ss_err[filt_low];
+
+        filt_best = filt_low;
+      }
+    }
+
+    // Now look at filt_high
+    if (filt_direction >= 0 && filt_high != filt_mid) {
+      if (ss_err[filt_high] < 0) {
+        ss_err[filt_high] = try_filter_frame(sd, cpi, filt_high, partial_frame);
+      }
+      // Was it better than the previous best?
+      if (ss_err[filt_high] < (best_err - bias)) {
+        best_err = ss_err[filt_high];
+        filt_best = filt_high;
+      }
+    }
+
+    // Half the step distance if the best filter value was the same as last time
+    if (filt_best == filt_mid) {
+      filter_step /= 2;
+      filt_direction = 0;
+    } else {
+      filt_direction = (filt_best < filt_mid) ? -1 : 1;
+      filt_mid = filt_best;
+    }
+  }
+
+  return filt_best;
+}
+
+void vp9_pick_filter_level(const YV12_BUFFER_CONFIG *sd, VP9_COMP *cpi,
+                           LPF_PICK_METHOD method) {
+  VP9_COMMON *const cm = &cpi->common;
+  struct loopfilter *const lf = &cm->lf;
+
+  lf->sharpness_level = cm->frame_type == KEY_FRAME ? 0
+                                                    : cpi->oxcf.sharpness;
+
+  if (method == LPF_PICK_MINIMAL_LPF && lf->filter_level) {
+      lf->filter_level = 0;
+  } else if (method >= LPF_PICK_FROM_Q) {
+    const int min_filter_level = 0;
+    const int max_filter_level = get_max_filter_level(cpi);
+    const int q = vp9_ac_quant(cm->base_qindex, 0, cm->bit_depth);
+    // These values were determined by linear fitting the result of the
+    // searched level, filt_guess = q * 0.316206 + 3.87252
+#if CONFIG_VP9_HIGHBITDEPTH
+    int filt_guess;
+    switch (cm->bit_depth) {
+      case VPX_BITS_8:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+        break;
+      case VPX_BITS_10:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 4060632, 20);
+        break;
+      case VPX_BITS_12:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 16242526, 22);
+        break;
+      default:
+        assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 "
+                    "or VPX_BITS_12");
+        return;
+    }
+#else
+    int filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    if (cm->frame_type == KEY_FRAME)
+      filt_guess -= 4;
+    lf->filter_level = clamp(filt_guess, min_filter_level, max_filter_level);
+  } else {
+    lf->filter_level = search_filter_level(sd, cpi,
+                                           method == LPF_PICK_FROM_SUBIMAGE);
+  }
+}

media/libvpx/vp9/encoder/vp9_picklpf.h

@@ -0,0 +1,30 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_PICKLPF_H_
+#define VP9_ENCODER_VP9_PICKLPF_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/encoder/vp9_encoder.h"
+
+struct yv12_buffer_config;
+struct VP9_COMP;
+
+void vp9_pick_filter_level(const struct yv12_buffer_config *sd,
+                           struct VP9_COMP *cpi, LPF_PICK_METHOD method);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_PICKLPF_H_

media/libvpx/vp9/encoder/vp9_pickmode.h

@@ -0,0 +1,39 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_PICKMODE_H_
+#define VP9_ENCODER_VP9_PICKMODE_H_
+
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
+                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);
+
+void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
+                         TileDataEnc *tile_data,
+                         int mi_row, int mi_col, RD_COST *rd_cost,
+                         BLOCK_SIZE bsize,
+                         PICK_MODE_CONTEXT *ctx);
+
+void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
+                                TileDataEnc *tile_data,
+                                int mi_row, int mi_col, RD_COST *rd_cost,
+                                BLOCK_SIZE bsize,
+                                PICK_MODE_CONTEXT *ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_PICKMODE_H_

media/libvpx/vp9/encoder/vp9_psnrhvs.c

@@ -0,0 +1,223 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ *
+ *  This code was originally written by: Gregory Maxwell, at the Daala
+ *  project.
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <math.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/encoder/vp9_ssim.h"
+
+#if !defined(M_PI)
+# define M_PI (3.141592653589793238462643)
+#endif
+#include <string.h>
+
+void od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x, int xstride) {
+  (void) xstride;
+  vp9_fdct8x8_c(x, y, ystride);
+}
+
+/* Normalized inverse quantization matrix for 8x8 DCT at the point of
+ * transparency. This is not the JPEG based matrix from the paper,
+ this one gives a slightly higher MOS agreement.*/
+float csf_y[8][8] = {{1.6193873005, 2.2901594831, 2.08509755623, 1.48366094411,
+    1.00227514334, 0.678296995242, 0.466224900598, 0.3265091542}, {2.2901594831,
+    1.94321815382, 2.04793073064, 1.68731108984, 1.2305666963, 0.868920337363,
+    0.61280991668, 0.436405793551}, {2.08509755623, 2.04793073064,
+    1.34329019223, 1.09205635862, 0.875748795257, 0.670882927016,
+    0.501731932449, 0.372504254596}, {1.48366094411, 1.68731108984,
+    1.09205635862, 0.772819797575, 0.605636379554, 0.48309405692,
+    0.380429446972, 0.295774038565}, {1.00227514334, 1.2305666963,
+    0.875748795257, 0.605636379554, 0.448996256676, 0.352889268808,
+    0.283006984131, 0.226951348204}, {0.678296995242, 0.868920337363,
+    0.670882927016, 0.48309405692, 0.352889268808, 0.27032073436,
+    0.215017739696, 0.17408067321}, {0.466224900598, 0.61280991668,
+    0.501731932449, 0.380429446972, 0.283006984131, 0.215017739696,
+    0.168869545842, 0.136153931001}, {0.3265091542, 0.436405793551,
+    0.372504254596, 0.295774038565, 0.226951348204, 0.17408067321,
+    0.136153931001, 0.109083846276}};
+float csf_cb420[8][8] = {
+    {1.91113096927, 2.46074210438, 1.18284184739, 1.14982565193, 1.05017074788,
+        0.898018824055, 0.74725392039, 0.615105596242}, {2.46074210438,
+        1.58529308355, 1.21363250036, 1.38190029285, 1.33100189972,
+        1.17428548929, 0.996404342439, 0.830890433625}, {1.18284184739,
+        1.21363250036, 0.978712413627, 1.02624506078, 1.03145147362,
+        0.960060382087, 0.849823426169, 0.731221236837}, {1.14982565193,
+        1.38190029285, 1.02624506078, 0.861317501629, 0.801821139099,
+        0.751437590932, 0.685398513368, 0.608694761374}, {1.05017074788,
+        1.33100189972, 1.03145147362, 0.801821139099, 0.676555426187,
+        0.605503172737, 0.55002013668, 0.495804539034}, {0.898018824055,
+        1.17428548929, 0.960060382087, 0.751437590932, 0.605503172737,
+        0.514674450957, 0.454353482512, 0.407050308965}, {0.74725392039,
+        0.996404342439, 0.849823426169, 0.685398513368, 0.55002013668,
+        0.454353482512, 0.389234902883, 0.342353999733}, {0.615105596242,
+        0.830890433625, 0.731221236837, 0.608694761374, 0.495804539034,
+        0.407050308965, 0.342353999733, 0.295530605237}};
+float csf_cr420[8][8] = {
+    {2.03871978502, 2.62502345193, 1.26180942886, 1.11019789803, 1.01397751469,
+        0.867069376285, 0.721500455585, 0.593906509971}, {2.62502345193,
+        1.69112867013, 1.17180569821, 1.3342742857, 1.28513006198,
+        1.13381474809, 0.962064122248, 0.802254508198}, {1.26180942886,
+        1.17180569821, 0.944981930573, 0.990876405848, 0.995903384143,
+        0.926972725286, 0.820534991409, 0.706020324706}, {1.11019789803,
+        1.3342742857, 0.990876405848, 0.831632933426, 0.77418706195,
+        0.725539939514, 0.661776842059, 0.587716619023}, {1.01397751469,
+        1.28513006198, 0.995903384143, 0.77418706195, 0.653238524286,
+        0.584635025748, 0.531064164893, 0.478717061273}, {0.867069376285,
+        1.13381474809, 0.926972725286, 0.725539939514, 0.584635025748,
+        0.496936637883, 0.438694579826, 0.393021669543}, {0.721500455585,
+        0.962064122248, 0.820534991409, 0.661776842059, 0.531064164893,
+        0.438694579826, 0.375820256136, 0.330555063063}, {0.593906509971,
+        0.802254508198, 0.706020324706, 0.587716619023, 0.478717061273,
+        0.393021669543, 0.330555063063, 0.285345396658}};
+
+static double convert_score_db(double _score, double _weight) {
+  return 10 * (log10(255 * 255) - log10(_weight * _score));
+}
+
+static double calc_psnrhvs(const unsigned char *_src, int _systride,
+                           const unsigned char *_dst, int _dystride,
+                           double _par, int _w, int _h, int _step,
+                           float _csf[8][8]) {
+  float ret;
+  int16_t dct_s[8 * 8], dct_d[8 * 8];
+  tran_low_t dct_s_coef[8 * 8], dct_d_coef[8 * 8];
+  float mask[8][8];
+  int pixels;
+  int x;
+  int y;
+  (void) _par;
+  ret = pixels = 0;
+  /*In the PSNR-HVS-M paper[1] the authors describe the construction of
+   their masking table as "we have used the quantization table for the
+   color component Y of JPEG [6] that has been also obtained on the
+   basis of CSF. Note that the values in quantization table JPEG have
+   been normalized and then squared." Their CSF matrix (from PSNR-HVS)
+   was also constructed from the JPEG matrices. I can not find any obvious
+   scheme of normalizing to produce their table, but if I multiply their
+   CSF by 0.38857 and square the result I get their masking table.
+   I have no idea where this constant comes from, but deviating from it
+   too greatly hurts MOS agreement.
+
+   [1] Nikolay Ponomarenko, Flavia Silvestri, Karen Egiazarian, Marco Carli,
+   Jaakko Astola, Vladimir Lukin, "On between-coefficient contrast masking
+   of DCT basis functions", CD-ROM Proceedings of the Third
+   International Workshop on Video Processing and Quality Metrics for Consumer
+   Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p.*/
+  for (x = 0; x < 8; x++)
+    for (y = 0; y < 8; y++)
+      mask[x][y] = (_csf[x][y] * 0.3885746225901003)
+          * (_csf[x][y] * 0.3885746225901003);
+  for (y = 0; y < _h - 7; y += _step) {
+    for (x = 0; x < _w - 7; x += _step) {
+      int i;
+      int j;
+      float s_means[4];
+      float d_means[4];
+      float s_vars[4];
+      float d_vars[4];
+      float s_gmean = 0;
+      float d_gmean = 0;
+      float s_gvar = 0;
+      float d_gvar = 0;
+      float s_mask = 0;
+      float d_mask = 0;
+      for (i = 0; i < 4; i++)
+        s_means[i] = d_means[i] = s_vars[i] = d_vars[i] = 0;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+          dct_s[i * 8 + j] = _src[(y + i) * _systride + (j + x)];
+          dct_d[i * 8 + j] = _dst[(y + i) * _dystride + (j + x)];
+          s_gmean += dct_s[i * 8 + j];
+          d_gmean += dct_d[i * 8 + j];
+          s_means[sub] += dct_s[i * 8 + j];
+          d_means[sub] += dct_d[i * 8 + j];
+        }
+      }
+      s_gmean /= 64.f;
+      d_gmean /= 64.f;
+      for (i = 0; i < 4; i++)
+        s_means[i] /= 16.f;
+      for (i = 0; i < 4; i++)
+        d_means[i] /= 16.f;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+          s_gvar += (dct_s[i * 8 + j] - s_gmean) * (dct_s[i * 8 + j] - s_gmean);
+          d_gvar += (dct_d[i * 8 + j] - d_gmean) * (dct_d[i * 8 + j] - d_gmean);
+          s_vars[sub] += (dct_s[i * 8 + j] - s_means[sub])
+              * (dct_s[i * 8 + j] - s_means[sub]);
+          d_vars[sub] += (dct_d[i * 8 + j] - d_means[sub])
+              * (dct_d[i * 8 + j] - d_means[sub]);
+        }
+      }
+      s_gvar *= 1 / 63.f * 64;
+      d_gvar *= 1 / 63.f * 64;
+      for (i = 0; i < 4; i++)
+        s_vars[i] *= 1 / 15.f * 16;
+      for (i = 0; i < 4; i++)
+        d_vars[i] *= 1 / 15.f * 16;
+      if (s_gvar > 0)
+        s_gvar = (s_vars[0] + s_vars[1] + s_vars[2] + s_vars[3]) / s_gvar;
+      if (d_gvar > 0)
+        d_gvar = (d_vars[0] + d_vars[1] + d_vars[2] + d_vars[3]) / d_gvar;
+      od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8);
+      od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8);
+      for (i = 0; i < 8; i++)
+        for (j = (i == 0); j < 8; j++)
+          s_mask += dct_s_coef[i * 8 + j] * dct_s_coef[i * 8 + j] * mask[i][j];
+      for (i = 0; i < 8; i++)
+        for (j = (i == 0); j < 8; j++)
+          d_mask += dct_d_coef[i * 8 + j] * dct_d_coef[i * 8 + j] * mask[i][j];
+      s_mask = sqrt(s_mask * s_gvar) / 32.f;
+      d_mask = sqrt(d_mask * d_gvar) / 32.f;
+      if (d_mask > s_mask)
+        s_mask = d_mask;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          float err;
+          err = fabs(dct_s_coef[i * 8 + j] - dct_d_coef[i * 8 + j]);
+          if (i != 0 || j != 0)
+            err = err < s_mask / mask[i][j] ? 0 : err - s_mask / mask[i][j];
+          ret += (err * _csf[i][j]) * (err * _csf[i][j]);
+          pixels++;
+        }
+      }
+    }
+  }
+  ret /= pixels;
+  return ret;
+}
+double vp9_psnrhvs(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+                   double *y_psnrhvs, double *u_psnrhvs, double *v_psnrhvs) {
+  double psnrhvs;
+  double par = 1.0;
+  int step = 7;
+  vp9_clear_system_state();
+  *y_psnrhvs = calc_psnrhvs(source->y_buffer, source->y_stride, dest->y_buffer,
+                            dest->y_stride, par, source->y_crop_width,
+                            source->y_crop_height, step, csf_y);
+
+  *u_psnrhvs = calc_psnrhvs(source->u_buffer, source->uv_stride, dest->u_buffer,
+                            dest->uv_stride, par, source->uv_crop_width,
+                            source->uv_crop_height, step, csf_cb420);
+
+  *v_psnrhvs = calc_psnrhvs(source->v_buffer, source->uv_stride, dest->v_buffer,
+                            dest->uv_stride, par, source->uv_crop_width,
+                            source->uv_crop_height, step, csf_cr420);
+  psnrhvs = (*y_psnrhvs) * .8 + .1 * ((*u_psnrhvs) + (*v_psnrhvs));
+
+  return convert_score_db(psnrhvs, 1.0);
+}

media/libvpx/vp9/encoder/vp9_quantize.c

@@ -0,0 +1,728 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_rd.h"
+
+void vp9_quantize_dc(const tran_low_t *coeff_ptr,
+                     int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 16;
+    qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_dc(const tran_low_t *coeff_ptr,
+                            int n_coeffs, int skip_block,
+                            const int16_t *round_ptr, const int16_t quant,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int rc = 0;
+    const int coeff = coeff_ptr[rc];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+    const int64_t tmp =
+        (clamp(abs_coeff + round_ptr[rc != 0], INT32_MIN, INT32_MAX) *
+         quant) >> 16;
+    qcoeff_ptr[rc] = (tran_low_t)((tmp ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int n_coeffs = 1024;
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+
+    tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 15;
+    qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
+                                  int skip_block,
+                                  const int16_t *round_ptr,
+                                  const int16_t quant,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr,
+                                  uint16_t *eob_ptr) {
+  const int n_coeffs = 1024;
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int rc = 0;
+    const int coeff = coeff_ptr[rc];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+    const int64_t tmp =
+        (clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+               INT32_MIN, INT32_MAX) * quant) >> 15;
+    qcoeff_ptr[rc] = (tran_low_t)((tmp ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
+    if (tmp)
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                       int skip_block,
+                       const int16_t *zbin_ptr, const int16_t *round_ptr,
+                       const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
+                       tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                       const int16_t *dequant_ptr,
+                       uint16_t *eob_ptr,
+                       const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+      tmp = (tmp * quant_ptr[rc != 0]) >> 16;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_fp_c(const tran_low_t *coeff_ptr,
+                              intptr_t count,
+                              int skip_block,
+                              const int16_t *zbin_ptr,
+                              const int16_t *round_ptr,
+                              const int16_t *quant_ptr,
+                              const int16_t *quant_shift_ptr,
+                              tran_low_t *qcoeff_ptr,
+                              tran_low_t *dqcoeff_ptr,
+                              const int16_t *dequant_ptr,
+                              uint16_t *eob_ptr,
+                              const int16_t *scan,
+                              const int16_t *iscan) {
+  int i;
+  int eob = -1;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      const int64_t tmp =
+          (clamp(abs_coeff + round_ptr[rc != 0], INT32_MIN, INT32_MAX) *
+           quant_ptr[rc != 0]) >> 16;
+
+      qcoeff_ptr[rc] = (tran_low_t)((tmp ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+// TODO(jingning) Refactor this file and combine functions with similar
+// operations.
+void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block,
+                             const int16_t *zbin_ptr, const int16_t *round_ptr,
+                             const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr,
+                             uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int tmp = 0;
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= (dequant_ptr[rc != 0] >> 2)) {
+        abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+        abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+        tmp = (abs_coeff * quant_ptr[rc != 0]) >> 15;
+        qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      }
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_fp_32x32_c(const tran_low_t *coeff_ptr,
+                                    intptr_t n_coeffs, int skip_block,
+                                    const int16_t *zbin_ptr,
+                                    const int16_t *round_ptr,
+                                    const int16_t *quant_ptr,
+                                    const int16_t *quant_shift_ptr,
+                                    tran_low_t *qcoeff_ptr,
+                                    tran_low_t *dqcoeff_ptr,
+                                    const int16_t *dequant_ptr,
+                                    uint16_t *eob_ptr,
+                                    const int16_t *scan, const int16_t *iscan) {
+  int i, eob = -1;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int64_t tmp = 0;
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= (dequant_ptr[rc != 0] >> 2)) {
+        tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                    INT32_MIN, INT32_MAX);
+        tmp = (tmp * quant_ptr[rc != 0]) >> 15;
+        qcoeff_ptr[rc] = (tran_low_t)((tmp ^ coeff_sign) - coeff_sign);
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      }
+
+      if (tmp)
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                      int skip_block,
+                      const int16_t *zbin_ptr, const int16_t *round_ptr,
+                      const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
+                      tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                      const int16_t *dequant_ptr,
+                      uint16_t *eob_ptr,
+                      const int16_t *scan, const int16_t *iscan) {
+  int i, non_zero_count = (int)n_coeffs, eob = -1;
+  const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = (int)n_coeffs - 1; i >= 0; i--) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
+        non_zero_count--;
+      else
+        break;
+    }
+
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < non_zero_count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= zbins[rc != 0]) {
+        int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+        tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+                  quant_shift_ptr[rc != 0]) >> 16;  // quantization
+        qcoeff_ptr[rc]  = (tmp ^ coeff_sign) - coeff_sign;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+        if (tmp)
+          eob = i;
+      }
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *zbin_ptr,
+                             const int16_t *round_ptr, const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr,
+                             uint16_t *eob_ptr, const int16_t *scan,
+                             const int16_t *iscan) {
+  int i, non_zero_count = (int)n_coeffs, eob = -1;
+  const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = (int)n_coeffs - 1; i >= 0; i--) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
+        non_zero_count--;
+      else
+        break;
+    }
+
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < non_zero_count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= zbins[rc != 0]) {
+        int64_t tmp = clamp(abs_coeff + round_ptr[rc != 0],
+                            INT32_MIN, INT32_MAX);
+        tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+                  quant_shift_ptr[rc != 0]) >> 16;  // quantization
+        qcoeff_ptr[rc]  = (tran_low_t)((tmp ^ coeff_sign) - coeff_sign);
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+
+        if (tmp)
+          eob = i;
+      }
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block,
+                            const int16_t *zbin_ptr, const int16_t *round_ptr,
+                            const int16_t *quant_ptr,
+                            const int16_t *quant_shift_ptr,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t *dequant_ptr,
+                            uint16_t *eob_ptr,
+                            const int16_t *scan, const int16_t *iscan) {
+  const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
+                        ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int tmp;
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+      abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+      tmp = ((((abs_coeff * quant_ptr[rc != 0]) >> 16) + abs_coeff) *
+               quant_shift_ptr[rc != 0]) >> 15;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+
+      if (tmp)
+        eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, int skip_block,
+                                   const int16_t *zbin_ptr,
+                                   const int16_t *round_ptr,
+                                   const int16_t *quant_ptr,
+                                   const int16_t *quant_shift_ptr,
+                                   tran_low_t *qcoeff_ptr,
+                                   tran_low_t *dqcoeff_ptr,
+                                   const int16_t *dequant_ptr,
+                                   uint16_t *eob_ptr,
+                                   const int16_t *scan, const int16_t *iscan) {
+  const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
+                        ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
+  const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
+
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      int64_t tmp = clamp(abs_coeff +
+                          ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                          INT32_MIN, INT32_MAX);
+      tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+               quant_shift_ptr[rc != 0]) >> 15;
+
+      qcoeff_ptr[rc] = (tran_low_t)((tmp ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+
+      if (tmp)
+        eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
+
+void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
+                                const int16_t *scan, const int16_t *iscan) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_highbd_quantize_b(BLOCK_OFFSET(p->coeff, block),
+                          16, x->skip_block,
+                          p->zbin, p->round, p->quant, p->quant_shift,
+                          BLOCK_OFFSET(p->qcoeff, block),
+                          BLOCK_OFFSET(pd->dqcoeff, block),
+                          pd->dequant, &p->eobs[block],
+                          scan, iscan);
+    return;
+  }
+#endif
+  vp9_quantize_b(BLOCK_OFFSET(p->coeff, block),
+                 16, x->skip_block,
+                 p->zbin, p->round, p->quant, p->quant_shift,
+                 BLOCK_OFFSET(p->qcoeff, block),
+                 BLOCK_OFFSET(pd->dqcoeff, block),
+                 pd->dequant, &p->eobs[block], scan, iscan);
+}
+
+static void invert_quant(int16_t *quant, int16_t *shift, int d) {
+  unsigned t;
+  int l;
+  t = d;
+  for (l = 0; t > 1; l++)
+    t >>= 1;
+  t = 1 + (1 << (16 + l)) / d;
+  *quant = (int16_t)(t - (1 << 16));
+  *shift = 1 << (16 - l);
+}
+
+static int get_qzbin_factor(int q, vpx_bit_depth_t bit_depth) {
+  const int quant = vp9_dc_quant(q, 0, bit_depth);
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return q == 0 ? 64 : (quant < 148 ? 84 : 80);
+    case VPX_BITS_10:
+      return q == 0 ? 64 : (quant < 592 ? 84 : 80);
+    case VPX_BITS_12:
+      return q == 0 ? 64 : (quant < 2368 ? 84 : 80);
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  return q == 0 ? 64 : (quant < 148 ? 84 : 80);
+#endif
+}
+
+void vp9_init_quantizer(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  QUANTS *const quants = &cpi->quants;
+  int i, q, quant;
+
+  for (q = 0; q < QINDEX_RANGE; q++) {
+    const int qzbin_factor = get_qzbin_factor(q, cm->bit_depth);
+    const int qrounding_factor = q == 0 ? 64 : 48;
+
+    for (i = 0; i < 2; ++i) {
+      int qrounding_factor_fp = i == 0 ? 48 : 42;
+      if (q == 0)
+        qrounding_factor_fp = 64;
+
+      // y
+      quant = i == 0 ? vp9_dc_quant(q, cm->y_dc_delta_q, cm->bit_depth)
+                     : vp9_ac_quant(q, 0, cm->bit_depth);
+      invert_quant(&quants->y_quant[q][i], &quants->y_quant_shift[q][i], quant);
+      quants->y_quant_fp[q][i] = (1 << 16) / quant;
+      quants->y_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
+      quants->y_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
+      quants->y_round[q][i] = (qrounding_factor * quant) >> 7;
+      cpi->y_dequant[q][i] = quant;
+
+      // uv
+      quant = i == 0 ? vp9_dc_quant(q, cm->uv_dc_delta_q, cm->bit_depth)
+                     : vp9_ac_quant(q, cm->uv_ac_delta_q, cm->bit_depth);
+      invert_quant(&quants->uv_quant[q][i],
+                   &quants->uv_quant_shift[q][i], quant);
+      quants->uv_quant_fp[q][i] = (1 << 16) / quant;
+      quants->uv_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
+      quants->uv_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
+      quants->uv_round[q][i] = (qrounding_factor * quant) >> 7;
+      cpi->uv_dequant[q][i] = quant;
+    }
+
+    for (i = 2; i < 8; i++) {
+      quants->y_quant[q][i] = quants->y_quant[q][1];
+      quants->y_quant_fp[q][i] = quants->y_quant_fp[q][1];
+      quants->y_round_fp[q][i] = quants->y_round_fp[q][1];
+      quants->y_quant_shift[q][i] = quants->y_quant_shift[q][1];
+      quants->y_zbin[q][i] = quants->y_zbin[q][1];
+      quants->y_round[q][i] = quants->y_round[q][1];
+      cpi->y_dequant[q][i] = cpi->y_dequant[q][1];
+
+      quants->uv_quant[q][i] = quants->uv_quant[q][1];
+      quants->uv_quant_fp[q][i] = quants->uv_quant_fp[q][1];
+      quants->uv_round_fp[q][i] = quants->uv_round_fp[q][1];
+      quants->uv_quant_shift[q][i] = quants->uv_quant_shift[q][1];
+      quants->uv_zbin[q][i] = quants->uv_zbin[q][1];
+      quants->uv_round[q][i] = quants->uv_round[q][1];
+      cpi->uv_dequant[q][i] = cpi->uv_dequant[q][1];
+    }
+  }
+}
+
+void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
+  const VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  QUANTS *const quants = &cpi->quants;
+  const int segment_id = xd->mi[0]->mbmi.segment_id;
+  const int qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+  const int rdmult = vp9_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q);
+  int i;
+
+  // Y
+  x->plane[0].quant = quants->y_quant[qindex];
+  x->plane[0].quant_fp = quants->y_quant_fp[qindex];
+  x->plane[0].round_fp = quants->y_round_fp[qindex];
+  x->plane[0].quant_shift = quants->y_quant_shift[qindex];
+  x->plane[0].zbin = quants->y_zbin[qindex];
+  x->plane[0].round = quants->y_round[qindex];
+  xd->plane[0].dequant = cpi->y_dequant[qindex];
+
+  x->plane[0].quant_thred[0] = x->plane[0].zbin[0] * x->plane[0].zbin[0];
+  x->plane[0].quant_thred[1] = x->plane[0].zbin[1] * x->plane[0].zbin[1];
+
+  // UV
+  for (i = 1; i < 3; i++) {
+    x->plane[i].quant = quants->uv_quant[qindex];
+    x->plane[i].quant_fp = quants->uv_quant_fp[qindex];
+    x->plane[i].round_fp = quants->uv_round_fp[qindex];
+    x->plane[i].quant_shift = quants->uv_quant_shift[qindex];
+    x->plane[i].zbin = quants->uv_zbin[qindex];
+    x->plane[i].round = quants->uv_round[qindex];
+    xd->plane[i].dequant = cpi->uv_dequant[qindex];
+
+    x->plane[i].quant_thred[0] = x->plane[i].zbin[0] * x->plane[i].zbin[0];
+    x->plane[i].quant_thred[1] = x->plane[i].zbin[1] * x->plane[i].zbin[1];
+  }
+
+  x->skip_block = vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
+  x->q_index = qindex;
+
+  x->errorperbit = rdmult >> 6;
+  x->errorperbit += (x->errorperbit == 0);
+
+  vp9_initialize_me_consts(cpi, x, x->q_index);
+}
+
+void vp9_frame_init_quantizer(VP9_COMP *cpi) {
+  vp9_init_plane_quantizers(cpi, &cpi->td.mb);
+}
+
+void vp9_set_quantizer(VP9_COMMON *cm, int q) {
+  // quantizer has to be reinitialized with vp9_init_quantizer() if any
+  // delta_q changes.
+  cm->base_qindex = q;
+  cm->y_dc_delta_q = 0;
+  cm->uv_dc_delta_q = 0;
+  cm->uv_ac_delta_q = 0;
+}
+
+// Table that converts 0-63 Q-range values passed in outside to the Qindex
+// range used internally.
+static const int quantizer_to_qindex[] = {
+  0,    4,   8,  12,  16,  20,  24,  28,
+  32,   36,  40,  44,  48,  52,  56,  60,
+  64,   68,  72,  76,  80,  84,  88,  92,
+  96,  100, 104, 108, 112, 116, 120, 124,
+  128, 132, 136, 140, 144, 148, 152, 156,
+  160, 164, 168, 172, 176, 180, 184, 188,
+  192, 196, 200, 204, 208, 212, 216, 220,
+  224, 228, 232, 236, 240, 244, 249, 255,
+};
+
+int vp9_quantizer_to_qindex(int quantizer) {
+  return quantizer_to_qindex[quantizer];
+}
+
+int vp9_qindex_to_quantizer(int qindex) {
+  int quantizer;
+
+  for (quantizer = 0; quantizer < 64; ++quantizer)
+    if (quantizer_to_qindex[quantizer] >= qindex)
+      return quantizer;
+
+  return 63;
+}

media/libvpx/vp9/encoder/vp9_quantize.h

@@ -0,0 +1,87 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_QUANTIZE_H_
+#define VP9_ENCODER_VP9_QUANTIZE_H_
+
+#include "./vpx_config.h"
+#include "vp9/encoder/vp9_block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  DECLARE_ALIGNED(16, int16_t, y_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round[QINDEX_RANGE][8]);
+
+  // TODO(jingning): in progress of re-working the quantization. will decide
+  // if we want to deprecate the current use of y_quant.
+  DECLARE_ALIGNED(16, int16_t, y_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round_fp[QINDEX_RANGE][8]);
+
+  DECLARE_ALIGNED(16, int16_t, uv_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round[QINDEX_RANGE][8]);
+} QUANTS;
+
+void vp9_quantize_dc(const tran_low_t *coeff_ptr,
+                     int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant_ptr,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vp9_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant_ptr,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
+                                const int16_t *scan, const int16_t *iscan);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_quantize_dc(const tran_low_t *coeff_ptr,
+                            int n_coeffs, int skip_block,
+                            const int16_t *round_ptr, const int16_t quant_ptr,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t dequant_ptr, uint16_t *eob_ptr);
+void vp9_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
+                                  int skip_block,
+                                  const int16_t *round_ptr,
+                                  const int16_t quant_ptr,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr,
+                                  uint16_t *eob_ptr);
+#endif
+
+struct VP9_COMP;
+struct VP9Common;
+
+void vp9_frame_init_quantizer(struct VP9_COMP *cpi);
+
+void vp9_init_plane_quantizers(struct VP9_COMP *cpi, MACROBLOCK *x);
+
+void vp9_init_quantizer(struct VP9_COMP *cpi);
+
+void vp9_set_quantizer(struct VP9Common *cm, int q);
+
+int vp9_quantizer_to_qindex(int quantizer);
+
+int vp9_qindex_to_quantizer(int qindex);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_QUANTIZE_H_

media/libvpx/vp9/encoder/vp9_ratectrl.h

@@ -0,0 +1,252 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_RATECTRL_H_
+#define VP9_ENCODER_VP9_RATECTRL_H_
+
+#include "vpx/vpx_codec.h"
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Bits Per MB at different Q (Multiplied by 512)
+#define BPER_MB_NORMBITS    9
+
+typedef enum {
+  INTER_NORMAL = 0,
+  INTER_HIGH = 1,
+  GF_ARF_LOW = 2,
+  GF_ARF_STD = 3,
+  KF_STD = 4,
+  RATE_FACTOR_LEVELS = 5
+} RATE_FACTOR_LEVEL;
+
+// Internal frame scaling level.
+typedef enum {
+  UNSCALED = 0,     // Frame is unscaled.
+  SCALE_STEP1 = 1,  // First-level down-scaling.
+  FRAME_SCALE_STEPS
+} FRAME_SCALE_LEVEL;
+
+// Frame dimensions multiplier wrt the native frame size, in 1/16ths,
+// specified for the scale-up case.
+// e.g. 24 => 16/24 = 2/3 of native size. The restriction to 1/16th is
+// intended to match the capabilities of the normative scaling filters,
+// giving precedence to the up-scaling accuracy.
+static const int frame_scale_factor[FRAME_SCALE_STEPS] = {16, 24};
+
+// Multiplier of the target rate to be used as threshold for triggering scaling.
+static const double rate_thresh_mult[FRAME_SCALE_STEPS] = {1.0, 2.0};
+
+// Scale dependent Rate Correction Factor multipliers. Compensates for the
+// greater number of bits per pixel generated in down-scaled frames.
+static const double rcf_mult[FRAME_SCALE_STEPS] = {1.0, 2.0};
+
+typedef struct {
+  // Rate targetting variables
+  int base_frame_target;           // A baseline frame target before adjustment
+                                   // for previous under or over shoot.
+  int this_frame_target;           // Actual frame target after rc adjustment.
+  int projected_frame_size;
+  int sb64_target_rate;
+  int last_q[FRAME_TYPES];         // Separate values for Intra/Inter
+  int last_boosted_qindex;         // Last boosted GF/KF/ARF q
+  int last_kf_qindex;              // Q index of the last key frame coded.
+
+  int gfu_boost;
+  int last_boost;
+  int kf_boost;
+
+  double rate_correction_factors[RATE_FACTOR_LEVELS];
+
+  int frames_since_golden;
+  int frames_till_gf_update_due;
+  int min_gf_interval;
+  int max_gf_interval;
+  int static_scene_max_gf_interval;
+  int baseline_gf_interval;
+  int constrained_gf_group;
+  int frames_to_key;
+  int frames_since_key;
+  int this_key_frame_forced;
+  int next_key_frame_forced;
+  int source_alt_ref_pending;
+  int source_alt_ref_active;
+  int is_src_frame_alt_ref;
+
+  int avg_frame_bandwidth;  // Average frame size target for clip
+  int min_frame_bandwidth;  // Minimum allocation used for any frame
+  int max_frame_bandwidth;  // Maximum burst rate allowed for a frame.
+
+  int ni_av_qi;
+  int ni_tot_qi;
+  int ni_frames;
+  int avg_frame_qindex[FRAME_TYPES];
+  double tot_q;
+  double avg_q;
+
+  int64_t buffer_level;
+  int64_t bits_off_target;
+  int64_t vbr_bits_off_target;
+  int64_t vbr_bits_off_target_fast;
+
+  int decimation_factor;
+  int decimation_count;
+
+  int rolling_target_bits;
+  int rolling_actual_bits;
+
+  int long_rolling_target_bits;
+  int long_rolling_actual_bits;
+
+  int rate_error_estimate;
+
+  int64_t total_actual_bits;
+  int64_t total_target_bits;
+  int64_t total_target_vs_actual;
+
+  int worst_quality;
+  int best_quality;
+
+  int64_t starting_buffer_level;
+  int64_t optimal_buffer_level;
+  int64_t maximum_buffer_size;
+
+  // rate control history for last frame(1) and the frame before(2).
+  // -1: undershot
+  //  1: overshoot
+  //  0: not initialized.
+  int rc_1_frame;
+  int rc_2_frame;
+  int q_1_frame;
+  int q_2_frame;
+
+  // Auto frame-scaling variables.
+  FRAME_SCALE_LEVEL frame_size_selector;
+  FRAME_SCALE_LEVEL next_frame_size_selector;
+  int frame_width[FRAME_SCALE_STEPS];
+  int frame_height[FRAME_SCALE_STEPS];
+  int rf_level_maxq[RATE_FACTOR_LEVELS];
+} RATE_CONTROL;
+
+struct VP9_COMP;
+struct VP9EncoderConfig;
+
+void vp9_rc_init(const struct VP9EncoderConfig *oxcf, int pass,
+                 RATE_CONTROL *rc);
+
+int vp9_estimate_bits_at_q(FRAME_TYPE frame_kind, int q, int mbs,
+                           double correction_factor,
+                           vpx_bit_depth_t bit_depth);
+
+double vp9_convert_qindex_to_q(int qindex, vpx_bit_depth_t bit_depth);
+
+void vp9_rc_init_minq_luts(void);
+
+// Generally at the high level, the following flow is expected
+// to be enforced for rate control:
+// First call per frame, one of:
+//   vp9_rc_get_one_pass_vbr_params()
+//   vp9_rc_get_one_pass_cbr_params()
+//   vp9_rc_get_svc_params()
+//   vp9_rc_get_first_pass_params()
+//   vp9_rc_get_second_pass_params()
+// depending on the usage to set the rate control encode parameters desired.
+//
+// Then, call encode_frame_to_data_rate() to perform the
+// actual encode. This function will in turn call encode_frame()
+// one or more times, followed by one of:
+//   vp9_rc_postencode_update()
+//   vp9_rc_postencode_update_drop_frame()
+//
+// The majority of rate control parameters are only expected
+// to be set in the vp9_rc_get_..._params() functions and
+// updated during the vp9_rc_postencode_update...() functions.
+// The only exceptions are vp9_rc_drop_frame() and
+// vp9_rc_update_rate_correction_factors() functions.
+
+// Functions to set parameters for encoding before the actual
+// encode_frame_to_data_rate() function.
+void vp9_rc_get_one_pass_vbr_params(struct VP9_COMP *cpi);
+void vp9_rc_get_one_pass_cbr_params(struct VP9_COMP *cpi);
+void vp9_rc_get_svc_params(struct VP9_COMP *cpi);
+
+// Post encode update of the rate control parameters based
+// on bytes used
+void vp9_rc_postencode_update(struct VP9_COMP *cpi, uint64_t bytes_used);
+// Post encode update of the rate control parameters for dropped frames
+void vp9_rc_postencode_update_drop_frame(struct VP9_COMP *cpi);
+
+// Updates rate correction factors
+// Changes only the rate correction factors in the rate control structure.
+void vp9_rc_update_rate_correction_factors(struct VP9_COMP *cpi);
+
+// Decide if we should drop this frame: For 1-pass CBR.
+// Changes only the decimation count in the rate control structure
+int vp9_rc_drop_frame(struct VP9_COMP *cpi);
+
+// Computes frame size bounds.
+void vp9_rc_compute_frame_size_bounds(const struct VP9_COMP *cpi,
+                                      int this_frame_target,
+                                      int *frame_under_shoot_limit,
+                                      int *frame_over_shoot_limit);
+
+// Picks q and q bounds given the target for bits
+int vp9_rc_pick_q_and_bounds(const struct VP9_COMP *cpi,
+                             int *bottom_index,
+                             int *top_index);
+
+// Estimates q to achieve a target bits per frame
+int vp9_rc_regulate_q(const struct VP9_COMP *cpi, int target_bits_per_frame,
+                      int active_best_quality, int active_worst_quality);
+
+// Estimates bits per mb for a given qindex and correction factor.
+int vp9_rc_bits_per_mb(FRAME_TYPE frame_type, int qindex,
+                       double correction_factor, vpx_bit_depth_t bit_depth);
+
+// Clamping utilities for bitrate targets for iframes and pframes.
+int vp9_rc_clamp_iframe_target_size(const struct VP9_COMP *const cpi,
+                                    int target);
+int vp9_rc_clamp_pframe_target_size(const struct VP9_COMP *const cpi,
+                                    int target);
+// Utility to set frame_target into the RATE_CONTROL structure
+// This function is called only from the vp9_rc_get_..._params() functions.
+void vp9_rc_set_frame_target(struct VP9_COMP *cpi, int target);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a target q value
+int vp9_compute_qdelta(const RATE_CONTROL *rc, double qstart, double qtarget,
+                       vpx_bit_depth_t bit_depth);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a value that should equate to the given rate ratio.
+int vp9_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
+                               int qindex, double rate_target_ratio,
+                               vpx_bit_depth_t bit_depth);
+
+int vp9_frame_type_qdelta(const struct VP9_COMP *cpi, int rf_level, int q);
+
+void vp9_rc_update_framerate(struct VP9_COMP *cpi);
+
+void vp9_rc_set_gf_interval_range(const struct VP9_COMP *const cpi,
+                                  RATE_CONTROL *const rc);
+
+void vp9_set_target_rate(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_RATECTRL_H_

media/libvpx/vp9/encoder/vp9_rd.c

@@ -0,0 +1,664 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./vp9_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encodemb.h"
+#include "vp9/encoder/vp9_encodemv.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rd.h"
+#include "vp9/encoder/vp9_tokenize.h"
+#include "vp9/encoder/vp9_variance.h"
+
+#define RD_THRESH_POW      1.25
+#define RD_MULT_EPB_RATIO  64
+
+// Factor to weigh the rate for switchable interp filters.
+#define SWITCHABLE_INTERP_RATE_FACTOR 1
+
+void vp9_rd_cost_reset(RD_COST *rd_cost) {
+  rd_cost->rate = INT_MAX;
+  rd_cost->dist = INT64_MAX;
+  rd_cost->rdcost = INT64_MAX;
+}
+
+void vp9_rd_cost_init(RD_COST *rd_cost) {
+  rd_cost->rate = 0;
+  rd_cost->dist = 0;
+  rd_cost->rdcost = 0;
+}
+
+// The baseline rd thresholds for breaking out of the rd loop for
+// certain modes are assumed to be based on 8x8 blocks.
+// This table is used to correct for block size.
+// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
+static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
+  2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32
+};
+
+static void fill_mode_costs(VP9_COMP *cpi) {
+  const FRAME_CONTEXT *const fc = cpi->common.fc;
+  int i, j;
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    for (j = 0; j < INTRA_MODES; ++j)
+      vp9_cost_tokens(cpi->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j],
+                      vp9_intra_mode_tree);
+
+  vp9_cost_tokens(cpi->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree);
+  vp9_cost_tokens(cpi->intra_uv_mode_cost[KEY_FRAME],
+                  vp9_kf_uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
+  vp9_cost_tokens(cpi->intra_uv_mode_cost[INTER_FRAME],
+                  fc->uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    vp9_cost_tokens(cpi->switchable_interp_costs[i],
+                    fc->switchable_interp_prob[i], vp9_switchable_interp_tree);
+}
+
+static void fill_token_costs(vp9_coeff_cost *c,
+                             vp9_coeff_probs_model (*p)[PLANE_TYPES]) {
+  int i, j, k, l;
+  TX_SIZE t;
+  for (t = TX_4X4; t <= TX_32X32; ++t)
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+            vp9_prob probs[ENTROPY_NODES];
+            vp9_model_to_full_probs(p[t][i][j][k][l], probs);
+            vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
+                            vp9_coef_tree);
+            vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
+                                 vp9_coef_tree);
+            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
+                   c[t][i][j][k][1][l][EOB_TOKEN]);
+          }
+}
+
+// Values are now correlated to quantizer.
+static int sad_per_bit16lut_8[QINDEX_RANGE];
+static int sad_per_bit4lut_8[QINDEX_RANGE];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static int sad_per_bit16lut_10[QINDEX_RANGE];
+static int sad_per_bit4lut_10[QINDEX_RANGE];
+static int sad_per_bit16lut_12[QINDEX_RANGE];
+static int sad_per_bit4lut_12[QINDEX_RANGE];
+#endif
+
+static void init_me_luts_bd(int *bit16lut, int *bit4lut, int range,
+                            vpx_bit_depth_t bit_depth) {
+  int i;
+  // Initialize the sad lut tables using a formulaic calculation for now.
+  // This is to make it easier to resolve the impact of experimental changes
+  // to the quantizer tables.
+  for (i = 0; i < range; i++) {
+    const double q = vp9_convert_qindex_to_q(i, bit_depth);
+    bit16lut[i] = (int)(0.0418 * q + 2.4107);
+    bit4lut[i] = (int)(0.063 * q + 2.742);
+  }
+}
+
+void vp9_init_me_luts(void) {
+  init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE,
+                  VPX_BITS_8);
+#if CONFIG_VP9_HIGHBITDEPTH
+  init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE,
+                  VPX_BITS_10);
+  init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE,
+                  VPX_BITS_12);
+#endif
+}
+
+static const int rd_boost_factor[16] = {
+  64, 32, 32, 32, 24, 16, 12, 12,
+  8, 8, 4, 4, 2, 2, 1, 0
+};
+static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = {
+  128, 144, 128, 128, 144
+};
+
+int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
+  const int64_t q = vp9_dc_quant(qindex, 0, cpi->common.bit_depth);
+#if CONFIG_VP9_HIGHBITDEPTH
+  int64_t rdmult = 0;
+  switch (cpi->common.bit_depth) {
+    case VPX_BITS_8:
+      rdmult = 88 * q * q / 24;
+      break;
+    case VPX_BITS_10:
+      rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 4);
+      break;
+    case VPX_BITS_12:
+      rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 8);
+      break;
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  int64_t rdmult = 88 * q * q / 24;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
+    const int boost_index = MIN(15, (cpi->rc.gfu_boost / 100));
+
+    rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
+    rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
+  }
+  return (int)rdmult;
+}
+
+static int compute_rd_thresh_factor(int qindex, vpx_bit_depth_t bit_depth) {
+  double q;
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
+      break;
+    case VPX_BITS_10:
+      q = vp9_dc_quant(qindex, 0, VPX_BITS_10) / 16.0;
+      break;
+    case VPX_BITS_12:
+      q = vp9_dc_quant(qindex, 0, VPX_BITS_12) / 64.0;
+      break;
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  // TODO(debargha): Adjust the function below.
+  return MAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
+}
+
+void vp9_initialize_me_consts(VP9_COMP *cpi, MACROBLOCK *x, int qindex) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (cpi->common.bit_depth) {
+    case VPX_BITS_8:
+      x->sadperbit16 = sad_per_bit16lut_8[qindex];
+      x->sadperbit4 = sad_per_bit4lut_8[qindex];
+      break;
+    case VPX_BITS_10:
+      x->sadperbit16 = sad_per_bit16lut_10[qindex];
+      x->sadperbit4 = sad_per_bit4lut_10[qindex];
+      break;
+    case VPX_BITS_12:
+      x->sadperbit16 = sad_per_bit16lut_12[qindex];
+      x->sadperbit4 = sad_per_bit4lut_12[qindex];
+      break;
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+  }
+#else
+  (void)cpi;
+  x->sadperbit16 = sad_per_bit16lut_8[qindex];
+  x->sadperbit4 = sad_per_bit4lut_8[qindex];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) {
+  int i, bsize, segment_id;
+
+  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
+    const int qindex =
+        clamp(vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
+              cm->y_dc_delta_q, 0, MAXQ);
+    const int q = compute_rd_thresh_factor(qindex, cm->bit_depth);
+
+    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
+      // Threshold here seems unnecessarily harsh but fine given actual
+      // range of values used for cpi->sf.thresh_mult[].
+      const int t = q * rd_thresh_block_size_factor[bsize];
+      const int thresh_max = INT_MAX / t;
+
+      if (bsize >= BLOCK_8X8) {
+        for (i = 0; i < MAX_MODES; ++i)
+          rd->threshes[segment_id][bsize][i] =
+              rd->thresh_mult[i] < thresh_max
+                  ? rd->thresh_mult[i] * t / 4
+                  : INT_MAX;
+      } else {
+        for (i = 0; i < MAX_REFS; ++i)
+          rd->threshes[segment_id][bsize][i] =
+              rd->thresh_mult_sub8x8[i] < thresh_max
+                  ? rd->thresh_mult_sub8x8[i] * t / 4
+                  : INT_MAX;
+      }
+    }
+  }
+}
+
+void vp9_initialize_rd_consts(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->td.mb;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+  vp9_clear_system_state();
+
+  rd->RDDIV = RDDIV_BITS;  // In bits (to multiply D by 128).
+  rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
+
+  x->errorperbit = rd->RDMULT / RD_MULT_EPB_RATIO;
+  x->errorperbit += (x->errorperbit == 0);
+
+  x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
+                       cm->frame_type != KEY_FRAME) ? 0 : 1;
+
+  set_block_thresholds(cm, rd);
+
+  if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME)
+    fill_token_costs(x->token_costs, cm->fc->coef_probs);
+
+  if (cpi->sf.partition_search_type != VAR_BASED_PARTITION ||
+      cm->frame_type == KEY_FRAME) {
+    for (i = 0; i < PARTITION_CONTEXTS; ++i)
+      vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(cm, i),
+                      vp9_partition_tree);
+  }
+
+  if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 ||
+      cm->frame_type == KEY_FRAME) {
+    fill_mode_costs(cpi);
+
+    if (!frame_is_intra_only(cm)) {
+      vp9_build_nmv_cost_table(x->nmvjointcost,
+                               cm->allow_high_precision_mv ? x->nmvcost_hp
+                                                           : x->nmvcost,
+                               &cm->fc->nmvc, cm->allow_high_precision_mv);
+
+      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+        vp9_cost_tokens((int *)cpi->inter_mode_cost[i],
+                        cm->fc->inter_mode_probs[i], vp9_inter_mode_tree);
+    }
+  }
+}
+
+static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
+  // NOTE: The tables below must be of the same size.
+
+  // The functions described below are sampled at the four most significant
+  // bits of x^2 + 8 / 256.
+
+  // Normalized rate:
+  // This table models the rate for a Laplacian source with given variance
+  // when quantized with a uniform quantizer with given stepsize. The
+  // closed form expression is:
+  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
+  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
+  // and H(x) is the binary entropy function.
+  static const int rate_tab_q10[] = {
+    65536,  6086,  5574,  5275,  5063,  4899,  4764,  4651,
+     4553,  4389,  4255,  4142,  4044,  3958,  3881,  3811,
+     3748,  3635,  3538,  3453,  3376,  3307,  3244,  3186,
+     3133,  3037,  2952,  2877,  2809,  2747,  2690,  2638,
+     2589,  2501,  2423,  2353,  2290,  2232,  2179,  2130,
+     2084,  2001,  1928,  1862,  1802,  1748,  1698,  1651,
+     1608,  1530,  1460,  1398,  1342,  1290,  1243,  1199,
+     1159,  1086,  1021,   963,   911,   864,   821,   781,
+      745,   680,   623,   574,   530,   490,   455,   424,
+      395,   345,   304,   269,   239,   213,   190,   171,
+      154,   126,   104,    87,    73,    61,    52,    44,
+       38,    28,    21,    16,    12,    10,     8,     6,
+        5,     3,     2,     1,     1,     1,     0,     0,
+  };
+  // Normalized distortion:
+  // This table models the normalized distortion for a Laplacian source
+  // with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expression is:
+  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
+  // where x = qpstep / sqrt(variance).
+  // Note the actual distortion is Dn * variance.
+  static const int dist_tab_q10[] = {
+       0,     0,     1,     1,     1,     2,     2,     2,
+       3,     3,     4,     5,     5,     6,     7,     7,
+       8,     9,    11,    12,    13,    15,    16,    17,
+      18,    21,    24,    26,    29,    31,    34,    36,
+      39,    44,    49,    54,    59,    64,    69,    73,
+      78,    88,    97,   106,   115,   124,   133,   142,
+     151,   167,   184,   200,   215,   231,   245,   260,
+     274,   301,   327,   351,   375,   397,   418,   439,
+     458,   495,   528,   559,   587,   613,   637,   659,
+     680,   717,   749,   777,   801,   823,   842,   859,
+     874,   899,   919,   936,   949,   960,   969,   977,
+     983,   994,  1001,  1006,  1010,  1013,  1015,  1017,
+    1018,  1020,  1022,  1022,  1023,  1023,  1023,  1024,
+  };
+  static const int xsq_iq_q10[] = {
+         0,      4,      8,     12,     16,     20,     24,     28,
+        32,     40,     48,     56,     64,     72,     80,     88,
+        96,    112,    128,    144,    160,    176,    192,    208,
+       224,    256,    288,    320,    352,    384,    416,    448,
+       480,    544,    608,    672,    736,    800,    864,    928,
+       992,   1120,   1248,   1376,   1504,   1632,   1760,   1888,
+      2016,   2272,   2528,   2784,   3040,   3296,   3552,   3808,
+      4064,   4576,   5088,   5600,   6112,   6624,   7136,   7648,
+      8160,   9184,  10208,  11232,  12256,  13280,  14304,  15328,
+     16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,
+     32736,  36832,  40928,  45024,  49120,  53216,  57312,  61408,
+     65504,  73696,  81888,  90080,  98272, 106464, 114656, 122848,
+    131040, 147424, 163808, 180192, 196576, 212960, 229344, 245728,
+  };
+  const int tmp = (xsq_q10 >> 2) + 8;
+  const int k = get_msb(tmp) - 3;
+  const int xq = (k << 3) + ((tmp >> k) & 0x7);
+  const int one_q10 = 1 << 10;
+  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
+  const int b_q10 = one_q10 - a_q10;
+  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
+  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
+}
+
+void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n_log2,
+                                  unsigned int qstep, int *rate,
+                                  int64_t *dist) {
+  // This function models the rate and distortion for a Laplacian
+  // source with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expressions are in:
+  // Hang and Chen, "Source Model for transform video coder and its
+  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
+  // Sys. for Video Tech., April 1997.
+  if (var == 0) {
+    *rate = 0;
+    *dist = 0;
+  } else {
+    int d_q10, r_q10;
+    static const uint32_t MAX_XSQ_Q10 = 245727;
+    const uint64_t xsq_q10_64 =
+        (((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var;
+    const int xsq_q10 = (int)MIN(xsq_q10_64, MAX_XSQ_Q10);
+    model_rd_norm(xsq_q10, &r_q10, &d_q10);
+    *rate = ((r_q10 << n_log2) + 2) >> 2;
+    *dist = (var * (int64_t)d_q10 + 512) >> 10;
+  }
+}
+
+void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[16],
+                              ENTROPY_CONTEXT t_left[16]) {
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+  const ENTROPY_CONTEXT *const above = pd->above_context;
+  const ENTROPY_CONTEXT *const left = pd->left_context;
+
+  int i;
+  switch (tx_size) {
+    case TX_4X4:
+      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
+      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
+      break;
+    case TX_8X8:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_16X16:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_32X32:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+    default:
+      assert(0 && "Invalid transform size.");
+      break;
+  }
+}
+
+void vp9_mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride,
+                 int ref_frame, BLOCK_SIZE block_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int i;
+  int zero_seen = 0;
+  int best_index = 0;
+  int best_sad = INT_MAX;
+  int this_sad = INT_MAX;
+  int max_mv = 0;
+  int near_same_nearest;
+  uint8_t *src_y_ptr = x->plane[0].src.buf;
+  uint8_t *ref_y_ptr;
+  const int num_mv_refs = MAX_MV_REF_CANDIDATES +
+                    (cpi->sf.adaptive_motion_search &&
+                     block_size < x->max_partition_size);
+
+  MV pred_mv[3];
+  pred_mv[0] = mbmi->ref_mvs[ref_frame][0].as_mv;
+  pred_mv[1] = mbmi->ref_mvs[ref_frame][1].as_mv;
+  pred_mv[2] = x->pred_mv[ref_frame];
+  assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0])));
+
+  near_same_nearest =
+      mbmi->ref_mvs[ref_frame][0].as_int == mbmi->ref_mvs[ref_frame][1].as_int;
+  // Get the sad for each candidate reference mv.
+  for (i = 0; i < num_mv_refs; ++i) {
+    const MV *this_mv = &pred_mv[i];
+    int fp_row, fp_col;
+
+    if (i == 1 && near_same_nearest)
+      continue;
+    fp_row = (this_mv->row + 3 + (this_mv->row >= 0)) >> 3;
+    fp_col = (this_mv->col + 3 + (this_mv->col >= 0)) >> 3;
+    max_mv = MAX(max_mv, MAX(abs(this_mv->row), abs(this_mv->col)) >> 3);
+
+    if (fp_row ==0 && fp_col == 0 && zero_seen)
+      continue;
+    zero_seen |= (fp_row ==0 && fp_col == 0);
+
+    ref_y_ptr =&ref_y_buffer[ref_y_stride * fp_row + fp_col];
+    // Find sad for current vector.
+    this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
+                                           ref_y_ptr, ref_y_stride);
+    // Note if it is the best so far.
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      best_index = i;
+    }
+  }
+
+  // Note the index of the mv that worked best in the reference list.
+  x->mv_best_ref_index[ref_frame] = best_index;
+  x->max_mv_context[ref_frame] = max_mv;
+  x->pred_mv_sad[ref_frame] = best_sad;
+}
+
+void vp9_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv) {
+  int i;
+
+  dst[0].buf = src->y_buffer;
+  dst[0].stride = src->y_stride;
+  dst[1].buf = src->u_buffer;
+  dst[2].buf = src->v_buffer;
+  dst[1].stride = dst[2].stride = src->uv_stride;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col,
+                     i ? scale_uv : scale,
+                     xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
+  }
+}
+
+int vp9_raster_block_offset(BLOCK_SIZE plane_bsize,
+                            int raster_block, int stride) {
+  const int bw = b_width_log2_lookup[plane_bsize];
+  const int y = 4 * (raster_block >> bw);
+  const int x = 4 * (raster_block & ((1 << bw) - 1));
+  return y * stride + x;
+}
+
+int16_t* vp9_raster_block_offset_int16(BLOCK_SIZE plane_bsize,
+                                       int raster_block, int16_t *base) {
+  const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
+  return base + vp9_raster_block_offset(plane_bsize, raster_block, stride);
+}
+
+YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
+                                             int ref_frame) {
+  const VP9_COMMON *const cm = &cpi->common;
+  const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
+  const int ref_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+  return
+      (scaled_idx != ref_idx && scaled_idx != INVALID_IDX) ?
+          &cm->buffer_pool->frame_bufs[scaled_idx].buf : NULL;
+}
+
+int vp9_get_switchable_rate(const VP9_COMP *cpi, const MACROBLOCKD *const xd) {
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx = vp9_get_pred_context_switchable_interp(xd);
+  return SWITCHABLE_INTERP_RATE_FACTOR *
+             cpi->switchable_interp_costs[ctx][mbmi->interp_filter];
+}
+
+void vp9_set_rd_speed_thresholds(VP9_COMP *cpi) {
+  int i;
+  RD_OPT *const rd = &cpi->rd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+
+  // Set baseline threshold values.
+  for (i = 0; i < MAX_MODES; ++i)
+    rd->thresh_mult[i] = cpi->oxcf.mode == BEST ? -500 : 0;
+
+  if (sf->adaptive_rd_thresh) {
+    rd->thresh_mult[THR_NEARESTMV] = 300;
+    rd->thresh_mult[THR_NEARESTG] = 300;
+    rd->thresh_mult[THR_NEARESTA] = 300;
+  } else {
+    rd->thresh_mult[THR_NEARESTMV] = 0;
+    rd->thresh_mult[THR_NEARESTG] = 0;
+    rd->thresh_mult[THR_NEARESTA] = 0;
+  }
+
+  rd->thresh_mult[THR_DC] += 1000;
+
+  rd->thresh_mult[THR_NEWMV] += 1000;
+  rd->thresh_mult[THR_NEWA] += 1000;
+  rd->thresh_mult[THR_NEWG] += 1000;
+
+  rd->thresh_mult[THR_NEARMV] += 1000;
+  rd->thresh_mult[THR_NEARA] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTLA] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTGA] += 1000;
+
+  rd->thresh_mult[THR_TM] += 1000;
+
+  rd->thresh_mult[THR_COMP_NEARLA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWLA] += 2000;
+  rd->thresh_mult[THR_NEARG] += 1000;
+  rd->thresh_mult[THR_COMP_NEARGA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWGA] += 2000;
+
+  rd->thresh_mult[THR_ZEROMV] += 2000;
+  rd->thresh_mult[THR_ZEROG] += 2000;
+  rd->thresh_mult[THR_ZEROA] += 2000;
+  rd->thresh_mult[THR_COMP_ZEROLA] += 2500;
+  rd->thresh_mult[THR_COMP_ZEROGA] += 2500;
+
+  rd->thresh_mult[THR_H_PRED] += 2000;
+  rd->thresh_mult[THR_V_PRED] += 2000;
+  rd->thresh_mult[THR_D45_PRED ] += 2500;
+  rd->thresh_mult[THR_D135_PRED] += 2500;
+  rd->thresh_mult[THR_D117_PRED] += 2500;
+  rd->thresh_mult[THR_D153_PRED] += 2500;
+  rd->thresh_mult[THR_D207_PRED] += 2500;
+  rd->thresh_mult[THR_D63_PRED] += 2500;
+}
+
+void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) {
+  static const int thresh_mult[2][MAX_REFS] =
+      {{2500, 2500, 2500, 4500, 4500, 2500},
+       {2000, 2000, 2000, 4000, 4000, 2000}};
+  RD_OPT *const rd = &cpi->rd;
+  const int idx = cpi->oxcf.mode == BEST;
+  memcpy(rd->thresh_mult_sub8x8, thresh_mult[idx], sizeof(thresh_mult[idx]));
+}
+
+void vp9_update_rd_thresh_fact(int (*factor_buf)[MAX_MODES], int rd_thresh,
+                               int bsize, int best_mode_index) {
+  if (rd_thresh > 0) {
+    const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
+    int mode;
+    for (mode = 0; mode < top_mode; ++mode) {
+      const BLOCK_SIZE min_size = MAX(bsize - 1, BLOCK_4X4);
+      const BLOCK_SIZE max_size = MIN(bsize + 2, BLOCK_64X64);
+      BLOCK_SIZE bs;
+      for (bs = min_size; bs <= max_size; ++bs) {
+        int *const fact = &factor_buf[bs][mode];
+        if (mode == best_mode_index) {
+          *fact -= (*fact >> 4);
+        } else {
+          *fact = MIN(*fact + RD_THRESH_INC,
+                      rd_thresh * RD_THRESH_MAX_FACT);
+        }
+      }
+    }
+  }
+}
+
+int vp9_get_intra_cost_penalty(int qindex, int qdelta,
+                               vpx_bit_depth_t bit_depth) {
+  const int q = vp9_dc_quant(qindex, qdelta, bit_depth);
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return 20 * q;
+    case VPX_BITS_10:
+      return 5 * q;
+    case VPX_BITS_12:
+      return ROUND_POWER_OF_TWO(5 * q, 2);
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  return 20 * q;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+

media/libvpx/vp9/encoder/vp9_rd.h

@@ -0,0 +1,190 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_RD_H_
+#define VP9_ENCODER_VP9_RD_H_
+
+#include <limits.h>
+
+#include "vp9/common/vp9_blockd.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_context_tree.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RDDIV_BITS          7
+
+#define RDCOST(RM, DM, R, D) \
+  (((128 + ((int64_t)R) * (RM)) >> 8) + (D << DM))
+#define QIDX_SKIP_THRESH     115
+
+#define MV_COST_WEIGHT      108
+#define MV_COST_WEIGHT_SUB  120
+
+#define INVALID_MV 0x80008000
+
+#define MAX_MODES 30
+#define MAX_REFS  6
+
+#define RD_THRESH_MAX_FACT 64
+#define RD_THRESH_INC      1
+
+// This enumerator type needs to be kept aligned with the mode order in
+// const MODE_DEFINITION vp9_mode_order[MAX_MODES] used in the rd code.
+typedef enum {
+  THR_NEARESTMV,
+  THR_NEARESTA,
+  THR_NEARESTG,
+
+  THR_DC,
+
+  THR_NEWMV,
+  THR_NEWA,
+  THR_NEWG,
+
+  THR_NEARMV,
+  THR_NEARA,
+  THR_NEARG,
+
+  THR_ZEROMV,
+  THR_ZEROG,
+  THR_ZEROA,
+
+  THR_COMP_NEARESTLA,
+  THR_COMP_NEARESTGA,
+
+  THR_TM,
+
+  THR_COMP_NEARLA,
+  THR_COMP_NEWLA,
+  THR_COMP_NEARGA,
+  THR_COMP_NEWGA,
+
+  THR_COMP_ZEROLA,
+  THR_COMP_ZEROGA,
+
+  THR_H_PRED,
+  THR_V_PRED,
+  THR_D135_PRED,
+  THR_D207_PRED,
+  THR_D153_PRED,
+  THR_D63_PRED,
+  THR_D117_PRED,
+  THR_D45_PRED,
+} THR_MODES;
+
+typedef enum {
+  THR_LAST,
+  THR_GOLD,
+  THR_ALTR,
+  THR_COMP_LA,
+  THR_COMP_GA,
+  THR_INTRA,
+} THR_MODES_SUB8X8;
+
+typedef struct RD_OPT {
+  // Thresh_mult is used to set a threshold for the rd score. A higher value
+  // means that we will accept the best mode so far more often. This number
+  // is used in combination with the current block size, and thresh_freq_fact
+  // to pick a threshold.
+  int thresh_mult[MAX_MODES];
+  int thresh_mult_sub8x8[MAX_REFS];
+
+  int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
+
+  int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
+  // TODO(agrange): can this overflow?
+  int tx_select_threshes[MAX_REF_FRAMES][TX_MODES];
+
+  int64_t filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
+
+  int RDMULT;
+  int RDDIV;
+} RD_OPT;
+
+typedef struct RD_COST {
+  int rate;
+  int64_t dist;
+  int64_t rdcost;
+} RD_COST;
+
+// Reset the rate distortion cost values to maximum (invalid) value.
+void vp9_rd_cost_reset(RD_COST *rd_cost);
+// Initialize the rate distortion cost values to zero.
+void vp9_rd_cost_init(RD_COST *rd_cost);
+
+struct TileInfo;
+struct TileDataEnc;
+struct VP9_COMP;
+struct macroblock;
+
+int vp9_compute_rd_mult(const struct VP9_COMP *cpi, int qindex);
+
+void vp9_initialize_rd_consts(struct VP9_COMP *cpi);
+
+void vp9_initialize_me_consts(struct VP9_COMP *cpi, MACROBLOCK *x, int qindex);
+
+void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
+                                  unsigned int qstep, int *rate,
+                                  int64_t *dist);
+
+int vp9_get_switchable_rate(const struct VP9_COMP *cpi,
+                            const MACROBLOCKD *const xd);
+
+int vp9_raster_block_offset(BLOCK_SIZE plane_bsize,
+                            int raster_block, int stride);
+
+int16_t* vp9_raster_block_offset_int16(BLOCK_SIZE plane_bsize,
+                                       int raster_block, int16_t *base);
+
+YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const struct VP9_COMP *cpi,
+                                             int ref_frame);
+
+void vp9_init_me_luts(void);
+
+void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[16],
+                              ENTROPY_CONTEXT t_left[16]);
+
+void vp9_set_rd_speed_thresholds(struct VP9_COMP *cpi);
+
+void vp9_set_rd_speed_thresholds_sub8x8(struct VP9_COMP *cpi);
+
+void vp9_update_rd_thresh_fact(int (*fact)[MAX_MODES], int rd_thresh,
+                               int bsize, int best_mode_index);
+
+static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
+                                      int thresh_fact) {
+    return best_rd < ((int64_t)thresh * thresh_fact >> 5) || thresh == INT_MAX;
+}
+
+void vp9_mv_pred(struct VP9_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride,
+                 int ref_frame, BLOCK_SIZE block_size);
+
+void vp9_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv);
+
+int vp9_get_intra_cost_penalty(int qindex, int qdelta,
+                               vpx_bit_depth_t bit_depth);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_RD_H_

media/libvpx/vp9/encoder/vp9_rdopt.h

@@ -0,0 +1,68 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_RDOPT_H_
+#define VP9_ENCODER_VP9_RDOPT_H_
+
+#include "vp9/common/vp9_blockd.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_context_tree.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct TileInfo;
+struct VP9_COMP;
+struct macroblock;
+struct RD_COST;
+
+void vp9_rd_pick_intra_mode_sb(struct VP9_COMP *cpi, struct macroblock *x,
+                               struct RD_COST *rd_cost, BLOCK_SIZE bsize,
+                               PICK_MODE_CONTEXT *ctx, int64_t best_rd);
+
+unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
+                                           const struct buf_2d *ref,
+                                           BLOCK_SIZE bs);
+#if CONFIG_VP9_HIGHBITDEPTH
+unsigned int vp9_high_get_sby_perpixel_variance(VP9_COMP *cpi,
+                                                const struct buf_2d *ref,
+                                                BLOCK_SIZE bs, int bd);
+#endif
+
+void vp9_rd_pick_inter_mode_sb(struct VP9_COMP *cpi,
+                               struct TileDataEnc *tile_data,
+                               struct macroblock *x,
+                               int mi_row, int mi_col,
+                               struct RD_COST *rd_cost,
+                               BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                               int64_t best_rd_so_far);
+
+void vp9_rd_pick_inter_mode_sb_seg_skip(struct VP9_COMP *cpi,
+                                        struct TileDataEnc *tile_data,
+                                        struct macroblock *x,
+                                        struct RD_COST *rd_cost,
+                                        BLOCK_SIZE bsize,
+                                        PICK_MODE_CONTEXT *ctx,
+                                        int64_t best_rd_so_far);
+
+void vp9_rd_pick_inter_mode_sub8x8(struct VP9_COMP *cpi,
+                                   struct TileDataEnc *tile_data,
+                                   struct macroblock *x,
+                                   int mi_row, int mi_col,
+                                   struct RD_COST *rd_cost,
+                                   BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                                   int64_t best_rd_so_far);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_RDOPT_H_

media/libvpx/vp9/encoder/vp9_resize.c

@@ -0,0 +1,925 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/encoder/vp9_resize.h"
+
+#define FILTER_BITS               7
+
+#define INTERP_TAPS               8
+#define SUBPEL_BITS               5
+#define SUBPEL_MASK               ((1 << SUBPEL_BITS) - 1)
+#define INTERP_PRECISION_BITS     32
+
+typedef int16_t interp_kernel[INTERP_TAPS];
+
+// Filters for interpolation (0.5-band) - note this also filters integer pels.
+const interp_kernel vp9_filteredinterp_filters500[(1 << SUBPEL_BITS)] = {
+  {-3,  0, 35, 64, 35,  0, -3, 0},
+  {-3, -1, 34, 64, 36,  1, -3, 0},
+  {-3, -1, 32, 64, 38,  1, -3, 0},
+  {-2, -2, 31, 63, 39,  2, -3, 0},
+  {-2, -2, 29, 63, 41,  2, -3, 0},
+  {-2, -2, 28, 63, 42,  3, -4, 0},
+  {-2, -3, 27, 63, 43,  4, -4, 0},
+  {-2, -3, 25, 62, 45,  5, -4, 0},
+  {-2, -3, 24, 62, 46,  5, -4, 0},
+  {-2, -3, 23, 61, 47,  6, -4, 0},
+  {-2, -3, 21, 60, 49,  7, -4, 0},
+  {-1, -4, 20, 60, 50,  8, -4, -1},
+  {-1, -4, 19, 59, 51,  9, -4, -1},
+  {-1, -4, 17, 58, 52, 10, -4, 0},
+  {-1, -4, 16, 57, 53, 12, -4, -1},
+  {-1, -4, 15, 56, 54, 13, -4, -1},
+  {-1, -4, 14, 55, 55, 14, -4, -1},
+  {-1, -4, 13, 54, 56, 15, -4, -1},
+  {-1, -4, 12, 53, 57, 16, -4, -1},
+  {0, -4, 10, 52, 58, 17, -4, -1},
+  {-1, -4,  9, 51, 59, 19, -4, -1},
+  {-1, -4,  8, 50, 60, 20, -4, -1},
+  {0, -4,  7, 49, 60, 21, -3, -2},
+  {0, -4,  6, 47, 61, 23, -3, -2},
+  {0, -4,  5, 46, 62, 24, -3, -2},
+  {0, -4,  5, 45, 62, 25, -3, -2},
+  {0, -4,  4, 43, 63, 27, -3, -2},
+  {0, -4,  3, 42, 63, 28, -2, -2},
+  {0, -3,  2, 41, 63, 29, -2, -2},
+  {0, -3,  2, 39, 63, 31, -2, -2},
+  {0, -3,  1, 38, 64, 32, -1, -3},
+  {0, -3,  1, 36, 64, 34, -1, -3}
+};
+
+// Filters for interpolation (0.625-band) - note this also filters integer pels.
+const interp_kernel vp9_filteredinterp_filters625[(1 << SUBPEL_BITS)] = {
+  {-1, -8, 33, 80, 33, -8, -1, 0},
+  {-1, -8, 30, 80, 35, -8, -1, 1},
+  {-1, -8, 28, 80, 37, -7, -2, 1},
+  {0, -8, 26, 79, 39, -7, -2, 1},
+  {0, -8, 24, 79, 41, -7, -2, 1},
+  {0, -8, 22, 78, 43, -6, -2, 1},
+  {0, -8, 20, 78, 45, -5, -3, 1},
+  {0, -8, 18, 77, 48, -5, -3, 1},
+  {0, -8, 16, 76, 50, -4, -3, 1},
+  {0, -8, 15, 75, 52, -3, -4, 1},
+  {0, -7, 13, 74, 54, -3, -4, 1},
+  {0, -7, 11, 73, 56, -2, -4, 1},
+  {0, -7, 10, 71, 58, -1, -4, 1},
+  {1, -7,  8, 70, 60,  0, -5, 1},
+  {1, -6,  6, 68, 62,  1, -5, 1},
+  {1, -6,  5, 67, 63,  2, -5, 1},
+  {1, -6,  4, 65, 65,  4, -6, 1},
+  {1, -5,  2, 63, 67,  5, -6, 1},
+  {1, -5,  1, 62, 68,  6, -6, 1},
+  {1, -5,  0, 60, 70,  8, -7, 1},
+  {1, -4, -1, 58, 71, 10, -7, 0},
+  {1, -4, -2, 56, 73, 11, -7, 0},
+  {1, -4, -3, 54, 74, 13, -7, 0},
+  {1, -4, -3, 52, 75, 15, -8, 0},
+  {1, -3, -4, 50, 76, 16, -8, 0},
+  {1, -3, -5, 48, 77, 18, -8, 0},
+  {1, -3, -5, 45, 78, 20, -8, 0},
+  {1, -2, -6, 43, 78, 22, -8, 0},
+  {1, -2, -7, 41, 79, 24, -8, 0},
+  {1, -2, -7, 39, 79, 26, -8, 0},
+  {1, -2, -7, 37, 80, 28, -8, -1},
+  {1, -1, -8, 35, 80, 30, -8, -1},
+};
+
+// Filters for interpolation (0.75-band) - note this also filters integer pels.
+const interp_kernel vp9_filteredinterp_filters750[(1 << SUBPEL_BITS)] = {
+  {2, -11,  25,  96,  25, -11,   2, 0},
+  {2, -11,  22,  96,  28, -11,   2, 0},
+  {2, -10,  19,  95,  31, -11,   2, 0},
+  {2, -10,  17,  95,  34, -12,   2, 0},
+  {2,  -9,  14,  94,  37, -12,   2, 0},
+  {2,  -8,  12,  93,  40, -12,   1, 0},
+  {2,  -8,   9,  92,  43, -12,   1, 1},
+  {2,  -7,   7,  91,  46, -12,   1, 0},
+  {2,  -7,   5,  90,  49, -12,   1, 0},
+  {2,  -6,   3,  88,  52, -12,   0, 1},
+  {2,  -5,   1,  86,  55, -12,   0, 1},
+  {2,  -5,  -1,  84,  58, -11,   0, 1},
+  {2,  -4,  -2,  82,  61, -11,  -1, 1},
+  {2,  -4,  -4,  80,  64, -10,  -1, 1},
+  {1, -3, -5, 77, 67, -9, -1, 1},
+  {1, -3, -6, 75, 70, -8, -2, 1},
+  {1, -2, -7, 72, 72, -7, -2, 1},
+  {1, -2, -8, 70, 75, -6, -3, 1},
+  {1, -1, -9, 67, 77, -5, -3, 1},
+  {1,  -1, -10,  64,  80,  -4,  -4, 2},
+  {1,  -1, -11,  61,  82,  -2,  -4, 2},
+  {1,   0, -11,  58,  84,  -1,  -5, 2},
+  {1,   0, -12,  55,  86,   1,  -5, 2},
+  {1,   0, -12,  52,  88,   3,  -6, 2},
+  {0,   1, -12,  49,  90,   5,  -7, 2},
+  {0,   1, -12,  46,  91,   7,  -7, 2},
+  {1,   1, -12,  43,  92,   9,  -8, 2},
+  {0,   1, -12,  40,  93,  12,  -8, 2},
+  {0,   2, -12,  37,  94,  14,  -9, 2},
+  {0,   2, -12,  34,  95,  17, -10, 2},
+  {0,   2, -11,  31,  95,  19, -10, 2},
+  {0,   2, -11,  28,  96,  22, -11, 2}
+};
+
+// Filters for interpolation (0.875-band) - note this also filters integer pels.
+const interp_kernel vp9_filteredinterp_filters875[(1 << SUBPEL_BITS)] = {
+  {3,  -8,  13, 112,  13,  -8,   3, 0},
+  {3,  -7,  10, 112,  17,  -9,   3, -1},
+  {2,  -6,   7, 111,  21,  -9,   3, -1},
+  {2,  -5,   4, 111,  24, -10,   3, -1},
+  {2,  -4,   1, 110,  28, -11,   3, -1},
+  {1,  -3,  -1, 108,  32, -12,   4, -1},
+  {1,  -2,  -3, 106,  36, -13,   4, -1},
+  {1,  -1,  -6, 105,  40, -14,   4, -1},
+  {1,  -1,  -7, 102,  44, -14,   4, -1},
+  {1,   0,  -9, 100,  48, -15,   4, -1},
+  {1,   1, -11,  97,  53, -16,   4, -1},
+  {0,   1, -12,  95,  57, -16,   4, -1},
+  {0,   2, -13,  91,  61, -16,   4, -1},
+  {0,   2, -14,  88,  65, -16,   4, -1},
+  {0,   3, -15,  84,  69, -17,   4, 0},
+  {0,   3, -16,  81,  73, -16,   3, 0},
+  {0,   3, -16,  77,  77, -16,   3, 0},
+  {0,   3, -16,  73,  81, -16,   3, 0},
+  {0,   4, -17,  69,  84, -15,   3, 0},
+  {-1,   4, -16,  65,  88, -14,   2, 0},
+  {-1,   4, -16,  61,  91, -13,   2, 0},
+  {-1,   4, -16,  57,  95, -12,   1, 0},
+  {-1,   4, -16,  53,  97, -11,   1, 1},
+  {-1,   4, -15,  48, 100,  -9,   0, 1},
+  {-1,   4, -14,  44, 102,  -7,  -1, 1},
+  {-1,   4, -14,  40, 105,  -6,  -1, 1},
+  {-1,   4, -13,  36, 106,  -3,  -2, 1},
+  {-1,   4, -12,  32, 108,  -1,  -3, 1},
+  {-1,   3, -11,  28, 110,   1,  -4, 2},
+  {-1,   3, -10,  24, 111,   4,  -5, 2},
+  {-1,   3,  -9,  21, 111,   7,  -6, 2},
+  {-1,   3,  -9,  17, 112,  10,  -7, 3}
+};
+
+// Filters for interpolation (full-band) - no filtering for integer pixels
+const interp_kernel vp9_filteredinterp_filters1000[(1 << SUBPEL_BITS)] = {
+  {0,   0,   0, 128,   0,   0,   0, 0},
+  {0,   1,  -3, 128,   3,  -1,   0, 0},
+  {-1,   2,  -6, 127,   7,  -2,   1, 0},
+  {-1,   3,  -9, 126,  12,  -4,   1, 0},
+  {-1,   4, -12, 125,  16,  -5,   1, 0},
+  {-1,   4, -14, 123,  20,  -6,   2, 0},
+  {-1,   5, -15, 120,  25,  -8,   2, 0},
+  {-1,   5, -17, 118,  30,  -9,   3, -1},
+  {-1,   6, -18, 114,  35, -10,   3, -1},
+  {-1,   6, -19, 111,  41, -12,   3, -1},
+  {-1,   6, -20, 107,  46, -13,   4, -1},
+  {-1,   6, -21, 103,  52, -14,   4, -1},
+  {-1,   6, -21,  99,  57, -16,   5, -1},
+  {-1,   6, -21,  94,  63, -17,   5, -1},
+  {-1,   6, -20,  89,  68, -18,   5, -1},
+  {-1,   6, -20,  84,  73, -19,   6, -1},
+  {-1,   6, -20,  79,  79, -20,   6, -1},
+  {-1,   6, -19,  73,  84, -20,   6, -1},
+  {-1,   5, -18,  68,  89, -20,   6, -1},
+  {-1,   5, -17,  63,  94, -21,   6, -1},
+  {-1,   5, -16,  57,  99, -21,   6, -1},
+  {-1,   4, -14,  52, 103, -21,   6, -1},
+  {-1,   4, -13,  46, 107, -20,   6, -1},
+  {-1,   3, -12,  41, 111, -19,   6, -1},
+  {-1,   3, -10,  35, 114, -18,   6, -1},
+  {-1,   3,  -9,  30, 118, -17,   5, -1},
+  {0,   2,  -8,  25, 120, -15,   5, -1},
+  {0,   2,  -6,  20, 123, -14,   4, -1},
+  {0,   1,  -5,  16, 125, -12,   4, -1},
+  {0,   1,  -4,  12, 126,  -9,   3, -1},
+  {0,   1,  -2,   7, 127,  -6,   2, -1},
+  {0,   0,  -1,   3, 128,  -3,   1, 0}
+};
+
+// Filters for factor of 2 downsampling.
+static const int16_t vp9_down2_symeven_half_filter[] = {56, 12, -3, -1};
+static const int16_t vp9_down2_symodd_half_filter[] = {64, 35, 0, -3};
+
+static const interp_kernel *choose_interp_filter(int inlength, int outlength) {
+  int outlength16 = outlength * 16;
+  if (outlength16 >= inlength * 16)
+    return vp9_filteredinterp_filters1000;
+  else if (outlength16 >= inlength * 13)
+    return vp9_filteredinterp_filters875;
+  else if (outlength16 >= inlength * 11)
+    return vp9_filteredinterp_filters750;
+  else if (outlength16 >= inlength * 9)
+    return vp9_filteredinterp_filters625;
+  else
+    return vp9_filteredinterp_filters500;
+}
+
+static void interpolate(const uint8_t *const input, int inlength,
+                        uint8_t *output, int outlength) {
+  const int64_t delta = (((uint64_t)inlength << 32) + outlength / 2) /
+      outlength;
+  const int64_t offset = inlength > outlength ?
+      (((int64_t)(inlength - outlength) << 31) + outlength / 2) / outlength :
+      -(((int64_t)(outlength - inlength) << 31) + outlength / 2) / outlength;
+  uint8_t *optr = output;
+  int x, x1, x2, sum, k, int_pel, sub_pel;
+  int64_t y;
+
+  const interp_kernel *interp_filters =
+      choose_interp_filter(inlength, outlength);
+
+  x = 0;
+  y = offset;
+  while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
+    x++;
+    y += delta;
+  }
+  x1 = x;
+  x = outlength - 1;
+  y = delta * x + offset;
+  while ((y >> INTERP_PRECISION_BITS) +
+         (int64_t)(INTERP_TAPS / 2) >= inlength) {
+    x--;
+    y -= delta;
+  }
+  x2 = x;
+  if (x1 > x2) {
+    for (x = 0, y = offset; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k) {
+        const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
+        sum += filter[k] * input[(pk < 0 ? 0 :
+                                  (pk >= inlength ? inlength - 1 : pk))];
+      }
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+  } else {
+    // Initial part.
+    for (x = 0, y = offset; x < x1; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0 ?
+                                  0 :
+                                  int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    // Middle part.
+    for (; x <= x2; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    // End part.
+    for (; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >=
+                                  inlength ?  inlength - 1 :
+                                  int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+  }
+}
+
+static void down2_symeven(const uint8_t *const input, int length,
+                          uint8_t *output) {
+  // Actual filter len = 2 * filter_len_half.
+  const int16_t *filter = vp9_down2_symeven_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symeven_half_filter) / 2;
+  int i, j;
+  uint8_t *optr = output;
+  int l1 = filter_len_half;
+  int l2 = (length - filter_len_half);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  }
+}
+
+static void down2_symodd(const uint8_t *const input, int length,
+                         uint8_t *output) {
+  // Actual filter len = 2 * filter_len_half - 1.
+  const int16_t *filter = vp9_down2_symodd_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symodd_half_filter) / 2;
+  int i, j;
+  uint8_t *optr = output;
+  int l1 = filter_len_half - 1;
+  int l2 = (length - filter_len_half + 1);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  }
+}
+
+static int get_down2_length(int length, int steps) {
+  int s;
+  for (s = 0; s < steps; ++s)
+    length = (length + 1) >> 1;
+  return length;
+}
+
+static int get_down2_steps(int in_length, int out_length) {
+  int steps = 0;
+  int proj_in_length;
+  while ((proj_in_length = get_down2_length(in_length, 1)) >= out_length) {
+    ++steps;
+    in_length = proj_in_length;
+  }
+  return steps;
+}
+
+static void resize_multistep(const uint8_t *const input,
+                             int length,
+                             uint8_t *output,
+                             int olength,
+                             uint8_t *buf) {
+  int steps;
+  if (length == olength) {
+    memcpy(output, input, sizeof(uint8_t) * length);
+    return;
+  }
+  steps = get_down2_steps(length, olength);
+
+  if (steps > 0) {
+    int s;
+    uint8_t *out = NULL;
+    uint8_t *tmpbuf = NULL;
+    uint8_t *otmp, *otmp2;
+    int filteredlength = length;
+    if (!tmpbuf) {
+      tmpbuf = (uint8_t *)malloc(sizeof(uint8_t) * length);
+      otmp = tmpbuf;
+    } else {
+      otmp = buf;
+    }
+    otmp2 = otmp + get_down2_length(length, 1);
+    for (s = 0; s < steps; ++s) {
+      const int proj_filteredlength = get_down2_length(filteredlength, 1);
+      const uint8_t *const in = (s == 0 ? input : out);
+      if (s == steps - 1 && proj_filteredlength == olength)
+        out = output;
+      else
+        out = (s & 1 ? otmp2 : otmp);
+      if (filteredlength & 1)
+        down2_symodd(in, filteredlength, out);
+      else
+        down2_symeven(in, filteredlength, out);
+      filteredlength = proj_filteredlength;
+    }
+    if (filteredlength != olength) {
+      interpolate(out, filteredlength, output, olength);
+    }
+    if (tmpbuf)
+      free(tmpbuf);
+  } else {
+    interpolate(input, length, output, olength);
+  }
+}
+
+static void fill_col_to_arr(uint8_t *img, int stride, int len, uint8_t *arr) {
+  int i;
+  uint8_t *iptr = img;
+  uint8_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *aptr++ = *iptr;
+  }
+}
+
+static void fill_arr_to_col(uint8_t *img, int stride, int len, uint8_t *arr) {
+  int i;
+  uint8_t *iptr = img;
+  uint8_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *iptr = *aptr++;
+  }
+}
+
+void vp9_resize_plane(const uint8_t *const input,
+                      int height,
+                      int width,
+                      int in_stride,
+                      uint8_t *output,
+                      int height2,
+                      int width2,
+                      int out_stride) {
+  int i;
+  uint8_t *intbuf = (uint8_t *)malloc(sizeof(uint8_t) * width2 * height);
+  uint8_t *tmpbuf = (uint8_t *)malloc(sizeof(uint8_t) *
+                                      (width < height ? height : width));
+  uint8_t *arrbuf = (uint8_t *)malloc(sizeof(uint8_t) * (height + height2));
+  assert(width > 0);
+  assert(height > 0);
+  assert(width2 > 0);
+  assert(height2 > 0);
+  for (i = 0; i < height; ++i)
+    resize_multistep(input + in_stride * i, width,
+                        intbuf + width2 * i, width2, tmpbuf);
+  for (i = 0; i < width2; ++i) {
+    fill_col_to_arr(intbuf + i, width2, height, arrbuf);
+    resize_multistep(arrbuf, height, arrbuf + height, height2, tmpbuf);
+    fill_arr_to_col(output + i, out_stride, height2, arrbuf + height);
+  }
+  free(intbuf);
+  free(tmpbuf);
+  free(arrbuf);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_interpolate(const uint16_t *const input, int inlength,
+                               uint16_t *output, int outlength, int bd) {
+  const int64_t delta =
+      (((uint64_t)inlength << 32) + outlength / 2) / outlength;
+  const int64_t offset = inlength > outlength ?
+      (((int64_t)(inlength - outlength) << 31) + outlength / 2) / outlength :
+      -(((int64_t)(outlength - inlength) << 31) + outlength / 2) / outlength;
+  uint16_t *optr = output;
+  int x, x1, x2, sum, k, int_pel, sub_pel;
+  int64_t y;
+
+  const interp_kernel *interp_filters =
+      choose_interp_filter(inlength, outlength);
+
+  x = 0;
+  y = offset;
+  while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
+    x++;
+    y += delta;
+  }
+  x1 = x;
+  x = outlength - 1;
+  y = delta * x + offset;
+  while ((y >> INTERP_PRECISION_BITS) +
+         (int64_t)(INTERP_TAPS / 2) >= inlength) {
+    x--;
+    y -= delta;
+  }
+  x2 = x;
+  if (x1 > x2) {
+    for (x = 0, y = offset; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k) {
+        const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
+        sum += filter[k] *
+            input[(pk < 0 ? 0 : (pk >= inlength ? inlength - 1 : pk))];
+      }
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+  } else {
+    // Initial part.
+    for (x = 0, y = offset; x < x1; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] *
+            input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0 ?
+                   0 : int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+    // Middle part.
+    for (; x <= x2; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+    // End part.
+    for (; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >=
+                                  inlength ?  inlength - 1 :
+                                  int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+  }
+}
+
+static void highbd_down2_symeven(const uint16_t *const input, int length,
+                                 uint16_t *output, int bd) {
+  // Actual filter len = 2 * filter_len_half.
+  static const int16_t *filter = vp9_down2_symeven_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symeven_half_filter) / 2;
+  int i, j;
+  uint16_t *optr = output;
+  int l1 = filter_len_half;
+  int l2 = (length - filter_len_half);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  }
+}
+
+static void highbd_down2_symodd(const uint16_t *const input, int length,
+                              uint16_t *output, int bd) {
+  // Actual filter len = 2 * filter_len_half - 1.
+  static const int16_t *filter = vp9_down2_symodd_half_filter;
+  const int filter_len_half = sizeof(vp9_down2_symodd_half_filter) / 2;
+  int i, j;
+  uint16_t *optr = output;
+  int l1 = filter_len_half - 1;
+  int l2 = (length - filter_len_half + 1);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
+            filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  }
+}
+
+static void highbd_resize_multistep(const uint16_t *const input,
+                                    int length,
+                                    uint16_t *output,
+                                    int olength,
+                                    uint16_t *buf,
+                                    int bd) {
+  int steps;
+  if (length == olength) {
+    memcpy(output, input, sizeof(uint16_t) * length);
+    return;
+  }
+  steps = get_down2_steps(length, olength);
+
+  if (steps > 0) {
+    int s;
+    uint16_t *out = NULL;
+    uint16_t *tmpbuf = NULL;
+    uint16_t *otmp, *otmp2;
+    int filteredlength = length;
+    if (!tmpbuf) {
+      tmpbuf = (uint16_t *)malloc(sizeof(uint16_t) * length);
+      otmp = tmpbuf;
+    } else {
+      otmp = buf;
+    }
+    otmp2 = otmp + get_down2_length(length, 1);
+    for (s = 0; s < steps; ++s) {
+      const int proj_filteredlength = get_down2_length(filteredlength, 1);
+      const uint16_t *const in = (s == 0 ? input : out);
+      if (s == steps - 1 && proj_filteredlength == olength)
+        out = output;
+      else
+        out = (s & 1 ? otmp2 : otmp);
+      if (filteredlength & 1)
+        highbd_down2_symodd(in, filteredlength, out, bd);
+      else
+        highbd_down2_symeven(in, filteredlength, out, bd);
+      filteredlength = proj_filteredlength;
+    }
+    if (filteredlength != olength) {
+      highbd_interpolate(out, filteredlength, output, olength, bd);
+    }
+    if (tmpbuf)
+      free(tmpbuf);
+  } else {
+    highbd_interpolate(input, length, output, olength, bd);
+  }
+}
+
+static void highbd_fill_col_to_arr(uint16_t *img, int stride, int len,
+                                   uint16_t *arr) {
+  int i;
+  uint16_t *iptr = img;
+  uint16_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *aptr++ = *iptr;
+  }
+}
+
+static void highbd_fill_arr_to_col(uint16_t *img, int stride, int len,
+                                   uint16_t *arr) {
+  int i;
+  uint16_t *iptr = img;
+  uint16_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *iptr = *aptr++;
+  }
+}
+
+void vp9_highbd_resize_plane(const uint8_t *const input,
+                             int height,
+                             int width,
+                             int in_stride,
+                             uint8_t *output,
+                             int height2,
+                             int width2,
+                             int out_stride,
+                             int bd) {
+  int i;
+  uint16_t *intbuf = (uint16_t *)malloc(sizeof(uint16_t) * width2 * height);
+  uint16_t *tmpbuf = (uint16_t *)malloc(sizeof(uint16_t) *
+                                        (width < height ? height : width));
+  uint16_t *arrbuf = (uint16_t *)malloc(sizeof(uint16_t) * (height + height2));
+  for (i = 0; i < height; ++i) {
+    highbd_resize_multistep(CONVERT_TO_SHORTPTR(input + in_stride * i), width,
+                            intbuf + width2 * i, width2, tmpbuf, bd);
+  }
+  for (i = 0; i < width2; ++i) {
+    highbd_fill_col_to_arr(intbuf + i, width2, height, arrbuf);
+    highbd_resize_multistep(arrbuf, height, arrbuf + height, height2, tmpbuf,
+                            bd);
+    highbd_fill_arr_to_col(CONVERT_TO_SHORTPTR(output + i), out_stride, height2,
+                           arrbuf + height);
+  }
+  free(intbuf);
+  free(tmpbuf);
+  free(arrbuf);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_resize_frame420(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride,
+                         int height, int width,
+                         uint8_t *oy, int oy_stride,
+                         uint8_t *ou, uint8_t *ov, int ouv_stride,
+                         int oheight, int owidth) {
+  vp9_resize_plane(y, height, width, y_stride,
+                   oy, oheight, owidth, oy_stride);
+  vp9_resize_plane(u, height / 2, width / 2, uv_stride,
+                   ou, oheight / 2, owidth / 2, ouv_stride);
+  vp9_resize_plane(v, height / 2, width / 2, uv_stride,
+                   ov, oheight / 2, owidth / 2, ouv_stride);
+}
+
+void vp9_resize_frame422(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride,
+                         int height, int width,
+                         uint8_t *oy, int oy_stride,
+                         uint8_t *ou, uint8_t *ov, int ouv_stride,
+                         int oheight, int owidth) {
+  vp9_resize_plane(y, height, width, y_stride,
+                   oy, oheight, owidth, oy_stride);
+  vp9_resize_plane(u, height, width / 2, uv_stride,
+                   ou, oheight, owidth / 2, ouv_stride);
+  vp9_resize_plane(v, height, width / 2, uv_stride,
+                   ov, oheight, owidth / 2, ouv_stride);
+}
+
+void vp9_resize_frame444(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride,
+                         int height, int width,
+                         uint8_t *oy, int oy_stride,
+                         uint8_t *ou, uint8_t *ov, int ouv_stride,
+                         int oheight, int owidth) {
+  vp9_resize_plane(y, height, width, y_stride,
+                   oy, oheight, owidth, oy_stride);
+  vp9_resize_plane(u, height, width, uv_stride,
+                   ou, oheight, owidth, ouv_stride);
+  vp9_resize_plane(v, height, width, uv_stride,
+                   ov, oheight, owidth, ouv_stride);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_resize_frame420(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride,
+                                int height, int width,
+                                uint8_t *oy, int oy_stride,
+                                uint8_t *ou, uint8_t *ov, int ouv_stride,
+                                int oheight, int owidth, int bd) {
+  vp9_highbd_resize_plane(y, height, width, y_stride,
+                          oy, oheight, owidth, oy_stride, bd);
+  vp9_highbd_resize_plane(u, height / 2, width / 2, uv_stride,
+                          ou, oheight / 2, owidth / 2, ouv_stride, bd);
+  vp9_highbd_resize_plane(v, height / 2, width / 2, uv_stride,
+                          ov, oheight / 2, owidth / 2, ouv_stride, bd);
+}
+
+void vp9_highbd_resize_frame422(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride,
+                                int height, int width,
+                                uint8_t *oy, int oy_stride,
+                                uint8_t *ou, uint8_t *ov, int ouv_stride,
+                                int oheight, int owidth, int bd) {
+  vp9_highbd_resize_plane(y, height, width, y_stride,
+                          oy, oheight, owidth, oy_stride, bd);
+  vp9_highbd_resize_plane(u, height, width / 2, uv_stride,
+                          ou, oheight, owidth / 2, ouv_stride, bd);
+  vp9_highbd_resize_plane(v, height, width / 2, uv_stride,
+                          ov, oheight, owidth / 2, ouv_stride, bd);
+}
+
+void vp9_highbd_resize_frame444(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride,
+                                int height, int width,
+                                uint8_t *oy, int oy_stride,
+                                uint8_t *ou, uint8_t *ov, int ouv_stride,
+                                int oheight, int owidth, int bd) {
+  vp9_highbd_resize_plane(y, height, width, y_stride,
+                          oy, oheight, owidth, oy_stride, bd);
+  vp9_highbd_resize_plane(u, height, width, uv_stride,
+                          ou, oheight, owidth, ouv_stride, bd);
+  vp9_highbd_resize_plane(v, height, width, uv_stride,
+                          ov, oheight, owidth, ouv_stride, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH

media/libvpx/vp9/encoder/vp9_resize.h

@@ -0,0 +1,124 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_RESIZE_H_
+#define VP9_ENCODER_VP9_RESIZE_H_
+
+#include <stdio.h>
+#include "vpx/vpx_integer.h"
+
+void vp9_resize_plane(const uint8_t *const input,
+                      int height,
+                      int width,
+                      int in_stride,
+                      uint8_t *output,
+                      int height2,
+                      int width2,
+                      int out_stride);
+void vp9_resize_frame420(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u,
+                         const uint8_t *const v,
+                         int uv_stride,
+                         int height,
+                         int width,
+                         uint8_t *oy,
+                         int oy_stride,
+                         uint8_t *ou,
+                         uint8_t *ov,
+                         int ouv_stride,
+                         int oheight,
+                         int owidth);
+void vp9_resize_frame422(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u,
+                         const uint8_t *const v,
+                         int uv_stride,
+                         int height,
+                         int width,
+                         uint8_t *oy,
+                         int oy_stride,
+                         uint8_t *ou,
+                         uint8_t *ov,
+                         int ouv_stride,
+                         int oheight,
+                         int owidth);
+void vp9_resize_frame444(const uint8_t *const y,
+                         int y_stride,
+                         const uint8_t *const u,
+                         const uint8_t *const v,
+                         int uv_stride,
+                         int height,
+                         int width,
+                         uint8_t *oy,
+                         int oy_stride,
+                         uint8_t *ou,
+                         uint8_t *ov,
+                         int ouv_stride,
+                         int oheight,
+                         int owidth);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_resize_plane(const uint8_t *const input,
+                             int height,
+                             int width,
+                             int in_stride,
+                             uint8_t *output,
+                             int height2,
+                             int width2,
+                             int out_stride,
+                             int bd);
+void vp9_highbd_resize_frame420(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u,
+                                const uint8_t *const v,
+                                int uv_stride,
+                                int height,
+                                int width,
+                                uint8_t *oy,
+                                int oy_stride,
+                                uint8_t *ou,
+                                uint8_t *ov,
+                                int ouv_stride,
+                                int oheight,
+                                int owidth,
+                                int bd);
+void vp9_highbd_resize_frame422(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u,
+                                const uint8_t *const v,
+                                int uv_stride,
+                                int height,
+                                int width,
+                                uint8_t *oy,
+                                int oy_stride,
+                                uint8_t *ou,
+                                uint8_t *ov,
+                                int ouv_stride,
+                                int oheight,
+                                int owidth,
+                                int bd);
+void vp9_highbd_resize_frame444(const uint8_t *const y,
+                                int y_stride,
+                                const uint8_t *const u,
+                                const uint8_t *const v,
+                                int uv_stride,
+                                int height,
+                                int width,
+                                uint8_t *oy,
+                                int oy_stride,
+                                uint8_t *ou,
+                                uint8_t *ov,
+                                int ouv_stride,
+                                int oheight,
+                                int owidth,
+                                int bd);
+#endif    // CONFIG_VP9_HIGHBITDEPTH
+#endif    // VP9_ENCODER_VP9_RESIZE_H_

media/libvpx/vp9/encoder/vp9_segmentation.c

@@ -0,0 +1,281 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include <limits.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_segmentation.h"
+
+void vp9_enable_segmentation(struct segmentation *seg) {
+  seg->enabled = 1;
+  seg->update_map = 1;
+  seg->update_data = 1;
+}
+
+void vp9_disable_segmentation(struct segmentation *seg) {
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+}
+
+void vp9_set_segment_data(struct segmentation *seg,
+                          signed char *feature_data,
+                          unsigned char abs_delta) {
+  seg->abs_delta = abs_delta;
+
+  memcpy(seg->feature_data, feature_data, sizeof(seg->feature_data));
+}
+void vp9_disable_segfeature(struct segmentation *seg, int segment_id,
+                            SEG_LVL_FEATURES feature_id) {
+  seg->feature_mask[segment_id] &= ~(1 << feature_id);
+}
+
+void vp9_clear_segdata(struct segmentation *seg, int segment_id,
+                       SEG_LVL_FEATURES feature_id) {
+  seg->feature_data[segment_id][feature_id] = 0;
+}
+
+// Based on set of segment counts calculate a probability tree
+static void calc_segtree_probs(int *segcounts, vp9_prob *segment_tree_probs) {
+  // Work out probabilities of each segment
+  const int c01 = segcounts[0] + segcounts[1];
+  const int c23 = segcounts[2] + segcounts[3];
+  const int c45 = segcounts[4] + segcounts[5];
+  const int c67 = segcounts[6] + segcounts[7];
+
+  segment_tree_probs[0] = get_binary_prob(c01 + c23, c45 + c67);
+  segment_tree_probs[1] = get_binary_prob(c01, c23);
+  segment_tree_probs[2] = get_binary_prob(c45, c67);
+  segment_tree_probs[3] = get_binary_prob(segcounts[0], segcounts[1]);
+  segment_tree_probs[4] = get_binary_prob(segcounts[2], segcounts[3]);
+  segment_tree_probs[5] = get_binary_prob(segcounts[4], segcounts[5]);
+  segment_tree_probs[6] = get_binary_prob(segcounts[6], segcounts[7]);
+}
+
+// Based on set of segment counts and probabilities calculate a cost estimate
+static int cost_segmap(int *segcounts, vp9_prob *probs) {
+  const int c01 = segcounts[0] + segcounts[1];
+  const int c23 = segcounts[2] + segcounts[3];
+  const int c45 = segcounts[4] + segcounts[5];
+  const int c67 = segcounts[6] + segcounts[7];
+  const int c0123 = c01 + c23;
+  const int c4567 = c45 + c67;
+
+  // Cost the top node of the tree
+  int cost = c0123 * vp9_cost_zero(probs[0]) +
+             c4567 * vp9_cost_one(probs[0]);
+
+  // Cost subsequent levels
+  if (c0123 > 0) {
+    cost += c01 * vp9_cost_zero(probs[1]) +
+            c23 * vp9_cost_one(probs[1]);
+
+    if (c01 > 0)
+      cost += segcounts[0] * vp9_cost_zero(probs[3]) +
+              segcounts[1] * vp9_cost_one(probs[3]);
+    if (c23 > 0)
+      cost += segcounts[2] * vp9_cost_zero(probs[4]) +
+              segcounts[3] * vp9_cost_one(probs[4]);
+  }
+
+  if (c4567 > 0) {
+    cost += c45 * vp9_cost_zero(probs[2]) +
+            c67 * vp9_cost_one(probs[2]);
+
+    if (c45 > 0)
+      cost += segcounts[4] * vp9_cost_zero(probs[5]) +
+              segcounts[5] * vp9_cost_one(probs[5]);
+    if (c67 > 0)
+      cost += segcounts[6] * vp9_cost_zero(probs[6]) +
+              segcounts[7] * vp9_cost_one(probs[6]);
+  }
+
+  return cost;
+}
+
+static void count_segs(const VP9_COMMON *cm, MACROBLOCKD *xd,
+                       const TileInfo *tile, MODE_INFO **mi,
+                       int *no_pred_segcounts,
+                       int (*temporal_predictor_count)[2],
+                       int *t_unpred_seg_counts,
+                       int bw, int bh, int mi_row, int mi_col) {
+  int segment_id;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  xd->mi = mi;
+  segment_id = xd->mi[0]->mbmi.segment_id;
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
+
+  // Count the number of hits on each segment with no prediction
+  no_pred_segcounts[segment_id]++;
+
+  // Temporal prediction not allowed on key frames
+  if (cm->frame_type != KEY_FRAME) {
+    const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+    // Test to see if the segment id matches the predicted value.
+    const int pred_segment_id = vp9_get_segment_id(cm, cm->last_frame_seg_map,
+                                                   bsize, mi_row, mi_col);
+    const int pred_flag = pred_segment_id == segment_id;
+    const int pred_context = vp9_get_pred_context_seg_id(xd);
+
+    // Store the prediction status for this mb and update counts
+    // as appropriate
+    xd->mi[0]->mbmi.seg_id_predicted = pred_flag;
+    temporal_predictor_count[pred_context][pred_flag]++;
+
+    // Update the "unpredicted" segment count
+    if (!pred_flag)
+      t_unpred_seg_counts[segment_id]++;
+  }
+}
+
+static void count_segs_sb(const VP9_COMMON *cm, MACROBLOCKD *xd,
+                          const TileInfo *tile, MODE_INFO **mi,
+                          int *no_pred_segcounts,
+                          int (*temporal_predictor_count)[2],
+                          int *t_unpred_seg_counts,
+                          int mi_row, int mi_col,
+                          BLOCK_SIZE bsize) {
+  const int mis = cm->mi_stride;
+  int bw, bh;
+  const int bs = num_8x8_blocks_wide_lookup[bsize], hbs = bs / 2;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  bw = num_8x8_blocks_wide_lookup[mi[0]->mbmi.sb_type];
+  bh = num_8x8_blocks_high_lookup[mi[0]->mbmi.sb_type];
+
+  if (bw == bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, bs, mi_row, mi_col);
+  } else if (bw == bs && bh < bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+               temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+               mi_row + hbs, mi_col);
+  } else if (bw < bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs,
+               no_pred_segcounts, temporal_predictor_count, t_unpred_seg_counts,
+               hbs, bs, mi_row, mi_col + hbs);
+  } else {
+    const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
+    int n;
+
+    assert(bw < bs && bh < bs);
+
+    for (n = 0; n < 4; n++) {
+      const int mi_dc = hbs * (n & 1);
+      const int mi_dr = hbs * (n >> 1);
+
+      count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc],
+                    no_pred_segcounts, temporal_predictor_count,
+                    t_unpred_seg_counts,
+                    mi_row + mi_dr, mi_col + mi_dc, subsize);
+    }
+  }
+}
+
+void vp9_choose_segmap_coding_method(VP9_COMMON *cm, MACROBLOCKD *xd) {
+  struct segmentation *seg = &cm->seg;
+
+  int no_pred_cost;
+  int t_pred_cost = INT_MAX;
+
+  int i, tile_col, mi_row, mi_col;
+
+  int temporal_predictor_count[PREDICTION_PROBS][2] = { { 0 } };
+  int no_pred_segcounts[MAX_SEGMENTS] = { 0 };
+  int t_unpred_seg_counts[MAX_SEGMENTS] = { 0 };
+
+  vp9_prob no_pred_tree[SEG_TREE_PROBS];
+  vp9_prob t_pred_tree[SEG_TREE_PROBS];
+  vp9_prob t_nopred_prob[PREDICTION_PROBS];
+
+  // Set default state for the segment tree probabilities and the
+  // temporal coding probabilities
+  memset(seg->tree_probs, 255, sizeof(seg->tree_probs));
+  memset(seg->pred_probs, 255, sizeof(seg->pred_probs));
+
+  // First of all generate stats regarding how well the last segment map
+  // predicts this one
+  for (tile_col = 0; tile_col < 1 << cm->log2_tile_cols; tile_col++) {
+    TileInfo tile;
+    MODE_INFO **mi_ptr;
+    vp9_tile_init(&tile, cm, 0, tile_col);
+
+    mi_ptr = cm->mi_grid_visible + tile.mi_col_start;
+    for (mi_row = 0; mi_row < cm->mi_rows;
+         mi_row += 8, mi_ptr += 8 * cm->mi_stride) {
+      MODE_INFO **mi = mi_ptr;
+      for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
+           mi_col += 8, mi += 8)
+        count_segs_sb(cm, xd, &tile, mi, no_pred_segcounts,
+                      temporal_predictor_count, t_unpred_seg_counts,
+                      mi_row, mi_col, BLOCK_64X64);
+    }
+  }
+
+  // Work out probability tree for coding segments without prediction
+  // and the cost.
+  calc_segtree_probs(no_pred_segcounts, no_pred_tree);
+  no_pred_cost = cost_segmap(no_pred_segcounts, no_pred_tree);
+
+  // Key frames cannot use temporal prediction
+  if (!frame_is_intra_only(cm)) {
+    // Work out probability tree for coding those segments not
+    // predicted using the temporal method and the cost.
+    calc_segtree_probs(t_unpred_seg_counts, t_pred_tree);
+    t_pred_cost = cost_segmap(t_unpred_seg_counts, t_pred_tree);
+
+    // Add in the cost of the signaling for each prediction context.
+    for (i = 0; i < PREDICTION_PROBS; i++) {
+      const int count0 = temporal_predictor_count[i][0];
+      const int count1 = temporal_predictor_count[i][1];
+
+      t_nopred_prob[i] = get_binary_prob(count0, count1);
+
+      // Add in the predictor signaling cost
+      t_pred_cost += count0 * vp9_cost_zero(t_nopred_prob[i]) +
+                     count1 * vp9_cost_one(t_nopred_prob[i]);
+    }
+  }
+
+  // Now choose which coding method to use.
+  if (t_pred_cost < no_pred_cost) {
+    seg->temporal_update = 1;
+    memcpy(seg->tree_probs, t_pred_tree, sizeof(t_pred_tree));
+    memcpy(seg->pred_probs, t_nopred_prob, sizeof(t_nopred_prob));
+  } else {
+    seg->temporal_update = 0;
+    memcpy(seg->tree_probs, no_pred_tree, sizeof(no_pred_tree));
+  }
+}
+
+void vp9_reset_segment_features(struct segmentation *seg) {
+  // Set up default state for MB feature flags
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+  memset(seg->tree_probs, 255, sizeof(seg->tree_probs));
+  vp9_clearall_segfeatures(seg);
+}

media/libvpx/vp9/encoder/vp9_segmentation.h

@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_SEGMENTATION_H_
+#define VP9_ENCODER_VP9_SEGMENTATION_H_
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/encoder/vp9_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_enable_segmentation(struct segmentation *seg);
+void vp9_disable_segmentation(struct segmentation *seg);
+
+void vp9_disable_segfeature(struct segmentation *seg,
+                            int segment_id,
+                            SEG_LVL_FEATURES feature_id);
+void vp9_clear_segdata(struct segmentation *seg,
+                       int segment_id,
+                       SEG_LVL_FEATURES feature_id);
+
+// The values given for each segment can be either deltas (from the default
+// value chosen for the frame) or absolute values.
+//
+// Valid range for abs values is (0-127 for MB_LVL_ALT_Q), (0-63 for
+// SEGMENT_ALT_LF)
+// Valid range for delta values are (+/-127 for MB_LVL_ALT_Q), (+/-63 for
+// SEGMENT_ALT_LF)
+//
+// abs_delta = SEGMENT_DELTADATA (deltas) abs_delta = SEGMENT_ABSDATA (use
+// the absolute values given).
+void vp9_set_segment_data(struct segmentation *seg, signed char *feature_data,
+                          unsigned char abs_delta);
+
+void vp9_choose_segmap_coding_method(VP9_COMMON *cm, MACROBLOCKD *xd);
+
+void vp9_reset_segment_features(struct segmentation *seg);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SEGMENTATION_H_

media/libvpx/vp9/encoder/vp9_skin_detection.c

@@ -0,0 +1,104 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_skin_detection.h"
+
+// Fixed-point skin color model parameters.
+static const int skin_mean[2] = {7463, 9614};                 // q6
+static const int skin_inv_cov[4] = {4107, 1663, 1663, 2157};  // q16
+static const int skin_threshold = 1570636;                    // q18
+
+// Thresholds on luminance.
+static const int y_low = 20;
+static const int y_high = 220;
+
+// Evaluates the Mahalanobis distance measure for the input CbCr values.
+static int evaluate_skin_color_difference(int cb, int cr) {
+  const int cb_q6 = cb << 6;
+  const int cr_q6 = cr << 6;
+  const int cb_diff_q12 = (cb_q6 - skin_mean[0]) * (cb_q6 - skin_mean[0]);
+  const int cbcr_diff_q12 = (cb_q6 - skin_mean[0]) * (cr_q6 - skin_mean[1]);
+  const int cr_diff_q12 = (cr_q6 - skin_mean[1]) * (cr_q6 - skin_mean[1]);
+  const int cb_diff_q2 = (cb_diff_q12 + (1 << 9)) >> 10;
+  const int cbcr_diff_q2 = (cbcr_diff_q12 + (1 << 9)) >> 10;
+  const int cr_diff_q2 = (cr_diff_q12 + (1 << 9)) >> 10;
+  const int skin_diff = skin_inv_cov[0] * cb_diff_q2 +
+      skin_inv_cov[1] * cbcr_diff_q2 +
+      skin_inv_cov[2] * cbcr_diff_q2 +
+      skin_inv_cov[3] * cr_diff_q2;
+  return skin_diff;
+}
+
+int vp9_skin_pixel(const uint8_t y, const uint8_t cb, const uint8_t cr) {
+  if (y < y_low || y > y_high)
+    return 0;
+  else
+    return (evaluate_skin_color_difference(cb, cr) < skin_threshold);
+}
+
+#ifdef OUTPUT_YUV_SKINMAP
+// For viewing skin map on input source.
+void vp9_compute_skin_map(VP9_COMP *const cpi, FILE *yuv_skinmap_file) {
+  int i, j, mi_row, mi_col;
+  VP9_COMMON *const cm = &cpi->common;
+  uint8_t *y;
+  const uint8_t *src_y = cpi->Source->y_buffer;
+  const uint8_t *src_u = cpi->Source->u_buffer;
+  const uint8_t *src_v = cpi->Source->v_buffer;
+  const int src_ystride = cpi->Source->y_stride;
+  const int src_uvstride = cpi->Source->uv_stride;
+  YV12_BUFFER_CONFIG skinmap;
+  memset(&skinmap, 0, sizeof(YV12_BUFFER_CONFIG));
+  if (vp9_alloc_frame_buffer(&skinmap, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment)) {
+      vp9_free_frame_buffer(&skinmap);
+      return;
+  }
+  memset(skinmap.buffer_alloc, 128, skinmap.frame_size);
+  y = skinmap.y_buffer;
+  // Loop through 8x8 blocks and set skin map based on center pixel of block.
+  // Set y to white for skin block, otherwise set to source with gray scale.
+  // Ignore rightmost/bottom boundary blocks.
+  for (mi_row = 0; mi_row < cm->mi_rows - 1; ++mi_row) {
+    for (mi_col = 0; mi_col < cm->mi_cols - 1; ++mi_col) {
+      // Use middle pixel for each 8x8 block for skin detection.
+      // If middle pixel is skin, assign whole 8x8 block to skin.
+      const uint8_t ysource = src_y[4 * src_ystride + 4];
+      const uint8_t usource = src_u[2 * src_uvstride + 2];
+      const uint8_t vsource = src_v[2 * src_uvstride + 2];
+      const int is_skin = vp9_skin_pixel(ysource, usource, vsource);
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          if (is_skin)
+            y[i * src_ystride + j] = 255;
+          else
+            y[i * src_ystride + j] = src_y[i * src_ystride + j];
+        }
+      }
+      y += 8;
+      src_y += 8;
+      src_u += 4;
+      src_v += 4;
+    }
+    y += (src_ystride << 3) - ((cm->mi_cols - 1) << 3);
+    src_y += (src_ystride << 3) - ((cm->mi_cols - 1) << 3);
+    src_u += (src_uvstride << 2) - ((cm->mi_cols - 1) << 2);
+    src_v += (src_uvstride << 2) - ((cm->mi_cols - 1) << 2);
+  }
+  vp9_write_yuv_frame_420(&skinmap, yuv_skinmap_file);
+  vp9_free_frame_buffer(&skinmap);
+}
+#endif

media/libvpx/vp9/encoder/vp9_skin_detection.h

@@ -0,0 +1,35 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_SKIN_MAP_H_
+#define VP9_ENCODER_VP9_SKIN_MAP_H_
+
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9_COMP;
+
+// #define OUTPUT_YUV_SKINMAP
+
+int vp9_skin_pixel(const uint8_t y, const uint8_t cb, const uint8_t cr);
+
+#ifdef OUTPUT_YUV_SKINMAP
+// For viewing skin map on input source.
+void vp9_compute_skin_map(VP9_COMP *const cpi, FILE *yuv_skinmap_file);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SKIN_MAP_H_

media/libvpx/vp9/encoder/vp9_speed_features.c

@@ -0,0 +1,531 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_speed_features.h"
+
+// Intra only frames, golden frames (except alt ref overlays) and
+// alt ref frames tend to be coded at a higher than ambient quality
+static int frame_is_boosted(const VP9_COMP *cpi) {
+  return frame_is_kf_gf_arf(cpi) || vp9_is_upper_layer_key_frame(cpi);
+}
+
+static void set_good_speed_feature_framesize_dependent(VP9_COMMON *cm,
+                                                       SPEED_FEATURES *sf,
+                                                       int speed) {
+  if (speed >= 1) {
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+    } else {
+      sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 21);
+    }
+  }
+
+  if (speed >= 2) {
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+      sf->adaptive_pred_interp_filter = 0;
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+      sf->partition_search_breakout_rate_thr = 120;
+    } else {
+      sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
+      sf->partition_search_breakout_dist_thr = (1 << 22);
+      sf->partition_search_breakout_rate_thr = 100;
+    }
+  }
+
+  if (speed >= 3) {
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = DISABLE_ALL_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
+      sf->partition_search_breakout_dist_thr = (1 << 25);
+      sf->partition_search_breakout_rate_thr = 200;
+    } else {
+      sf->max_intra_bsize = BLOCK_32X32;
+      sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0;
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+      sf->partition_search_breakout_rate_thr = 120;
+    }
+  }
+
+  if (speed >= 4) {
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->partition_search_breakout_dist_thr = (1 << 26);
+    } else {
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+    }
+    sf->disable_split_mask = DISABLE_ALL_SPLIT;
+  }
+}
+
+// Sets a partition size down to which the auto partition code will always
+// search (can go lower), based on the image dimensions. The logic here
+// is that the extent to which ringing artefacts are offensive, depends
+// partly on the screen area that over which they propogate. Propogation is
+// limited by transform block size but the screen area take up by a given block
+// size will be larger for a small image format stretched to full screen.
+static BLOCK_SIZE set_partition_min_limit(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  unsigned int screen_area = (cm->width * cm->height);
+
+  // Select block size based on image format size.
+  if (screen_area < 1280 * 720) {
+    // Formats smaller in area than 720P
+    return BLOCK_4X4;
+  } else if (screen_area < 1920 * 1080) {
+    // Format >= 720P and < 1080P
+    return BLOCK_8X8;
+  } else {
+    // Formats 1080P and up
+    return BLOCK_16X16;
+  }
+}
+
+static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
+                                   SPEED_FEATURES *sf, int speed) {
+  const int boosted = frame_is_boosted(cpi);
+
+  sf->adaptive_rd_thresh = 1;
+  sf->allow_skip_recode = 1;
+
+  if (speed >= 1) {
+    sf->use_square_partition_only = !frame_is_intra_only(cm);
+    sf->less_rectangular_check  = 1;
+
+    sf->use_rd_breakout = 1;
+    sf->adaptive_motion_search = 1;
+    sf->mv.auto_mv_step_size = 1;
+    sf->adaptive_rd_thresh = 2;
+    sf->mv.subpel_iters_per_step = 1;
+    sf->mode_skip_start = 10;
+    sf->adaptive_pred_interp_filter = 1;
+
+    sf->recode_loop = ALLOW_RECODE_KFARFGF;
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
+
+    sf->tx_size_search_breakout = 1;
+    sf->partition_search_breakout_rate_thr = 80;
+  }
+
+  if (speed >= 2) {
+    sf->tx_size_search_method = frame_is_boosted(cpi) ? USE_FULL_RD
+                                                      : USE_LARGESTALL;
+
+    // Reference masking is not supported in dynamic scaling mode.
+    sf->reference_masking = cpi->oxcf.resize_mode != RESIZE_DYNAMIC ? 1 : 0;
+
+    sf->mode_search_skip_flags = (cm->frame_type == KEY_FRAME) ? 0 :
+                                 FLAG_SKIP_INTRA_DIRMISMATCH |
+                                 FLAG_SKIP_INTRA_BESTINTER |
+                                 FLAG_SKIP_COMP_BESTINTRA |
+                                 FLAG_SKIP_INTRA_LOWVAR;
+    sf->disable_filter_search_var_thresh = 100;
+    sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
+    sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
+    sf->rd_auto_partition_min_limit = set_partition_min_limit(cpi);
+    sf->allow_partition_search_skip = 1;
+  }
+
+  if (speed >= 3) {
+    sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
+                                                        : USE_LARGESTALL;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
+    sf->adaptive_pred_interp_filter = 0;
+    sf->adaptive_mode_search = 1;
+    sf->cb_partition_search = !boosted;
+    sf->cb_pred_filter_search = 1;
+    sf->alt_ref_search_fp = 1;
+    sf->recode_loop = ALLOW_RECODE_KFMAXBW;
+    sf->adaptive_rd_thresh = 3;
+    sf->mode_skip_start = 6;
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
+    sf->adaptive_interp_filter_search = 1;
+  }
+
+  if (speed >= 4) {
+    sf->use_square_partition_only = 1;
+    sf->tx_size_search_method = USE_LARGESTALL;
+    sf->mv.search_method = BIGDIA;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
+    sf->adaptive_rd_thresh = 4;
+    if (cm->frame_type != KEY_FRAME)
+      sf->mode_search_skip_flags |= FLAG_EARLY_TERMINATE;
+    sf->disable_filter_search_var_thresh = 200;
+    sf->use_lp32x32fdct = 1;
+    sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
+    sf->use_fast_coef_costing = 1;
+    sf->motion_field_mode_search = !boosted;
+    sf->partition_search_breakout_rate_thr = 300;
+  }
+
+  if (speed >= 5) {
+    int i;
+    sf->optimize_coefficients = 0;
+    sf->mv.search_method = HEX;
+    sf->disable_filter_search_var_thresh = 500;
+    for (i = 0; i < TX_SIZES; ++i) {
+      sf->intra_y_mode_mask[i] = INTRA_DC;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
+    }
+    sf->partition_search_breakout_rate_thr = 500;
+    sf->mv.reduce_first_step_size = 1;
+  }
+}
+
+static void set_rt_speed_feature_framesize_dependent(VP9_COMP *cpi,
+    SPEED_FEATURES *sf, int speed) {
+  VP9_COMMON *const cm = &cpi->common;
+
+  if (speed >= 1) {
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+    } else {
+      sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+    }
+  }
+
+  if (speed >= 2) {
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+                                              : DISABLE_ALL_INTER_SPLIT;
+    } else {
+      sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
+    }
+  }
+
+  if (speed >= 5) {
+    if (MIN(cm->width, cm->height) >= 720) {
+      sf->partition_search_breakout_dist_thr = (1 << 25);
+    } else {
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+    }
+  }
+
+  if (speed >= 7) {
+    sf->encode_breakout_thresh = (MIN(cm->width, cm->height) >= 720) ?
+        800 : 300;
+  }
+}
+
+static void set_rt_speed_feature(VP9_COMP *cpi, SPEED_FEATURES *sf,
+                                 int speed, vp9e_tune_content content) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int is_keyframe = cm->frame_type == KEY_FRAME;
+  const int frames_since_key = is_keyframe ? 0 : cpi->rc.frames_since_key;
+  sf->static_segmentation = 0;
+  sf->adaptive_rd_thresh = 1;
+  sf->use_fast_coef_costing = 1;
+
+  if (speed >= 1) {
+    sf->use_square_partition_only = !frame_is_intra_only(cm);
+    sf->less_rectangular_check = 1;
+    sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
+                                                        : USE_LARGESTALL;
+
+    sf->use_rd_breakout = 1;
+
+    sf->adaptive_motion_search = 1;
+    sf->adaptive_pred_interp_filter = 1;
+    sf->mv.auto_mv_step_size = 1;
+    sf->adaptive_rd_thresh = 2;
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
+  }
+
+  if (speed >= 2) {
+    sf->mode_search_skip_flags = (cm->frame_type == KEY_FRAME) ? 0 :
+                                 FLAG_SKIP_INTRA_DIRMISMATCH |
+                                 FLAG_SKIP_INTRA_BESTINTER |
+                                 FLAG_SKIP_COMP_BESTINTRA |
+                                 FLAG_SKIP_INTRA_LOWVAR;
+    sf->adaptive_pred_interp_filter = 2;
+
+    // Disable reference masking if using spatial scaling since
+    // pred_mv_sad will not be set (since vp9_mv_pred will not
+    // be called).
+    // TODO(marpan/agrange): Fix this condition.
+    sf->reference_masking = (cpi->oxcf.resize_mode != RESIZE_DYNAMIC &&
+                             cpi->svc.number_spatial_layers == 1) ? 1 : 0;
+
+    sf->disable_filter_search_var_thresh = 50;
+    sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
+    sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
+    sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
+    sf->adjust_partitioning_from_last_frame = 1;
+    sf->last_partitioning_redo_frequency = 3;
+    sf->use_lp32x32fdct = 1;
+    sf->mode_skip_start = 11;
+    sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
+  }
+
+  if (speed >= 3) {
+    sf->use_square_partition_only = 1;
+    sf->disable_filter_search_var_thresh = 100;
+    sf->use_uv_intra_rd_estimate = 1;
+    sf->skip_encode_sb = 1;
+    sf->mv.subpel_iters_per_step = 1;
+    sf->adaptive_rd_thresh = 4;
+    sf->mode_skip_start = 6;
+    sf->allow_skip_recode = 0;
+    sf->optimize_coefficients = 0;
+    sf->disable_split_mask = DISABLE_ALL_SPLIT;
+    sf->lpf_pick = LPF_PICK_FROM_Q;
+  }
+
+  if (speed >= 4) {
+    int i;
+    sf->last_partitioning_redo_frequency = 4;
+    sf->adaptive_rd_thresh = 5;
+    sf->use_fast_coef_costing = 0;
+    sf->auto_min_max_partition_size = STRICT_NEIGHBORING_MIN_MAX;
+    sf->adjust_partitioning_from_last_frame =
+        cm->last_frame_type != cm->frame_type || (0 ==
+        (frames_since_key + 1) % sf->last_partitioning_redo_frequency);
+    sf->mv.subpel_force_stop = 1;
+    for (i = 0; i < TX_SIZES; i++) {
+      sf->intra_y_mode_mask[i] = INTRA_DC_H_V;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
+    }
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->frame_parameter_update = 0;
+    sf->mv.search_method = FAST_HEX;
+
+    sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEAR_NEW;
+    sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST;
+    sf->max_intra_bsize = BLOCK_32X32;
+    sf->allow_skip_recode = 1;
+  }
+
+  if (speed >= 5) {
+    sf->use_quant_fp = !is_keyframe;
+    sf->auto_min_max_partition_size = is_keyframe ? RELAXED_NEIGHBORING_MIN_MAX
+                                                  : STRICT_NEIGHBORING_MIN_MAX;
+    sf->default_max_partition_size = BLOCK_32X32;
+    sf->default_min_partition_size = BLOCK_8X8;
+    sf->force_frame_boost = is_keyframe ||
+        (frames_since_key % (sf->last_partitioning_redo_frequency << 1) == 1);
+    sf->max_delta_qindex = is_keyframe ? 20 : 15;
+    sf->partition_search_type = REFERENCE_PARTITION;
+    sf->use_nonrd_pick_mode = 1;
+    sf->allow_skip_recode = 0;
+    sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEW_ZERO;
+    sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST_NEW_ZERO;
+    sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST_NEW_ZERO;
+    sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST_NEW_ZERO;
+    sf->adaptive_rd_thresh = 2;
+    // This feature is only enabled when partition search is disabled.
+    sf->reuse_inter_pred_sby = 1;
+    sf->partition_search_breakout_rate_thr = 200;
+    sf->coeff_prob_appx_step = 4;
+    sf->use_fast_coef_updates = is_keyframe ? TWO_LOOP : ONE_LOOP_REDUCED;
+    sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH;
+    sf->tx_size_search_method = is_keyframe ? USE_LARGESTALL : USE_TX_8X8;
+
+    if (!is_keyframe) {
+      int i;
+      if (content == VP9E_CONTENT_SCREEN) {
+        for (i = 0; i < BLOCK_SIZES; ++i)
+          sf->intra_y_mode_bsize_mask[i] = INTRA_DC_TM_H_V;
+      } else {
+        for (i = 0; i < BLOCK_SIZES; ++i)
+          if (i >= BLOCK_16X16)
+            sf->intra_y_mode_bsize_mask[i] = INTRA_DC;
+          else
+            // Use H and V intra mode for block sizes <= 16X16.
+            sf->intra_y_mode_bsize_mask[i] = INTRA_DC_H_V;
+      }
+    }
+  }
+
+  if (speed >= 6) {
+    // Adaptively switch between SOURCE_VAR_BASED_PARTITION and FIXED_PARTITION.
+    sf->partition_search_type = VAR_BASED_PARTITION;
+    // Turn on this to use non-RD key frame coding mode.
+    sf->use_nonrd_pick_mode = 1;
+    sf->mv.search_method = NSTEP;
+    sf->mv.reduce_first_step_size = 1;
+    sf->skip_encode_sb = 0;
+  }
+
+  if (speed >= 7) {
+    sf->adaptive_rd_thresh = 3;
+    sf->mv.search_method = FAST_DIAMOND;
+    sf->mv.fullpel_search_step_param = 10;
+  }
+  if (speed >= 8) {
+    sf->adaptive_rd_thresh = 4;
+    sf->mv.subpel_force_stop = 2;
+    sf->lpf_pick = LPF_PICK_MINIMAL_LPF;
+  }
+}
+
+void vp9_set_speed_features_framesize_dependent(VP9_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  VP9_COMMON *const cm = &cpi->common;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+  if (oxcf->mode == REALTIME) {
+    set_rt_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
+  } else if (oxcf->mode == GOOD) {
+    set_good_speed_feature_framesize_dependent(cm, sf, oxcf->speed);
+  }
+
+  if (sf->disable_split_mask == DISABLE_ALL_SPLIT) {
+    sf->adaptive_pred_interp_filter = 0;
+  }
+
+  if (cpi->encode_breakout && oxcf->mode == REALTIME &&
+      sf->encode_breakout_thresh > cpi->encode_breakout) {
+    cpi->encode_breakout = sf->encode_breakout_thresh;
+  }
+
+  // Check for masked out split cases.
+  for (i = 0; i < MAX_REFS; ++i) {
+    if (sf->disable_split_mask & (1 << i)) {
+      rd->thresh_mult_sub8x8[i] = INT_MAX;
+    }
+  }
+}
+
+void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->td.mb;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  int i;
+
+  // best quality defaults
+  sf->frame_parameter_update = 1;
+  sf->mv.search_method = NSTEP;
+  sf->recode_loop = ALLOW_RECODE;
+  sf->mv.subpel_search_method = SUBPEL_TREE;
+  sf->mv.subpel_iters_per_step = 2;
+  sf->mv.subpel_force_stop = 0;
+  sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
+  sf->mv.reduce_first_step_size = 0;
+  sf->coeff_prob_appx_step = 1;
+  sf->mv.auto_mv_step_size = 0;
+  sf->mv.fullpel_search_step_param = 6;
+  sf->comp_inter_joint_search_thresh = BLOCK_4X4;
+  sf->adaptive_rd_thresh = 0;
+  sf->tx_size_search_method = USE_FULL_RD;
+  sf->use_lp32x32fdct = 0;
+  sf->adaptive_motion_search = 0;
+  sf->adaptive_pred_interp_filter = 0;
+  sf->adaptive_mode_search = 0;
+  sf->cb_pred_filter_search = 0;
+  sf->cb_partition_search = 0;
+  sf->motion_field_mode_search = 0;
+  sf->alt_ref_search_fp = 0;
+  sf->use_quant_fp = 0;
+  sf->reference_masking = 0;
+  sf->partition_search_type = SEARCH_PARTITION;
+  sf->less_rectangular_check = 0;
+  sf->use_square_partition_only = 0;
+  sf->auto_min_max_partition_size = NOT_IN_USE;
+  sf->rd_auto_partition_min_limit = BLOCK_4X4;
+  sf->default_max_partition_size = BLOCK_64X64;
+  sf->default_min_partition_size = BLOCK_4X4;
+  sf->adjust_partitioning_from_last_frame = 0;
+  sf->last_partitioning_redo_frequency = 4;
+  sf->disable_split_mask = 0;
+  sf->mode_search_skip_flags = 0;
+  sf->force_frame_boost = 0;
+  sf->max_delta_qindex = 0;
+  sf->disable_filter_search_var_thresh = 0;
+  sf->adaptive_interp_filter_search = 0;
+  sf->allow_partition_search_skip = 0;
+
+  for (i = 0; i < TX_SIZES; i++) {
+    sf->intra_y_mode_mask[i] = INTRA_ALL;
+    sf->intra_uv_mode_mask[i] = INTRA_ALL;
+  }
+  sf->use_rd_breakout = 0;
+  sf->skip_encode_sb = 0;
+  sf->use_uv_intra_rd_estimate = 0;
+  sf->allow_skip_recode = 0;
+  sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE;
+  sf->use_fast_coef_updates = TWO_LOOP;
+  sf->use_fast_coef_costing = 0;
+  sf->mode_skip_start = MAX_MODES;  // Mode index at which mode skip mask set
+  sf->schedule_mode_search = 0;
+  sf->use_nonrd_pick_mode = 0;
+  for (i = 0; i < BLOCK_SIZES; ++i)
+    sf->inter_mode_mask[i] = INTER_ALL;
+  sf->max_intra_bsize = BLOCK_64X64;
+  sf->reuse_inter_pred_sby = 0;
+  // This setting only takes effect when partition_search_type is set
+  // to FIXED_PARTITION.
+  sf->always_this_block_size = BLOCK_16X16;
+  sf->search_type_check_frequency = 50;
+  sf->encode_breakout_thresh = 0;
+  // Recode loop tolerance %.
+  sf->recode_tolerance = 25;
+  sf->default_interp_filter = SWITCHABLE;
+  sf->tx_size_search_breakout = 0;
+  sf->partition_search_breakout_dist_thr = 0;
+  sf->partition_search_breakout_rate_thr = 0;
+
+  if (oxcf->mode == REALTIME)
+    set_rt_speed_feature(cpi, sf, oxcf->speed, oxcf->content);
+  else if (oxcf->mode == GOOD)
+    set_good_speed_feature(cpi, cm, sf, oxcf->speed);
+
+  cpi->full_search_sad = vp9_full_search_sad;
+  cpi->diamond_search_sad = oxcf->mode == BEST ? vp9_full_range_search
+                                               : vp9_diamond_search_sad;
+
+  // Slow quant, dct and trellis not worthwhile for first pass
+  // so make sure they are always turned off.
+  if (oxcf->pass == 1)
+    sf->optimize_coefficients = 0;
+
+  // No recode for 1 pass.
+  if (oxcf->pass == 0) {
+    sf->recode_loop = DISALLOW_RECODE;
+    sf->optimize_coefficients = 0;
+  }
+
+  if (sf->mv.subpel_search_method == SUBPEL_TREE) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_MORE) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_more;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_EVENMORE) {
+    cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_evenmore;
+  }
+
+  x->optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1;
+
+  x->min_partition_size = sf->default_min_partition_size;
+  x->max_partition_size = sf->default_max_partition_size;
+
+  if (!cpi->oxcf.frame_periodic_boost) {
+    sf->max_delta_qindex = 0;
+  }
+}

media/libvpx/vp9/encoder/vp9_speed_features.h

@@ -0,0 +1,429 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_SPEED_FEATURES_H_
+#define VP9_ENCODER_VP9_SPEED_FEATURES_H_
+
+#include "vp9/common/vp9_enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  INTRA_ALL       = (1 << DC_PRED) |
+                    (1 << V_PRED) | (1 << H_PRED) |
+                    (1 << D45_PRED) | (1 << D135_PRED) |
+                    (1 << D117_PRED) | (1 << D153_PRED) |
+                    (1 << D207_PRED) | (1 << D63_PRED) |
+                    (1 << TM_PRED),
+  INTRA_DC        = (1 << DC_PRED),
+  INTRA_DC_TM     = (1 << DC_PRED) | (1 << TM_PRED),
+  INTRA_DC_H_V    = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
+  INTRA_DC_TM_H_V = (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) |
+                    (1 << H_PRED)
+};
+
+enum {
+  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST = (1 << NEARESTMV),
+  INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
+  INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV),
+  INTER_NEAREST_NEW_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV),
+};
+
+enum {
+  DISABLE_ALL_INTER_SPLIT   = (1 << THR_COMP_GA) |
+                              (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) |
+                              (1 << THR_GOLD) |
+                              (1 << THR_LAST),
+
+  DISABLE_ALL_SPLIT         = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
+
+  DISABLE_COMPOUND_SPLIT    = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
+
+  LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) |
+                              (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) |
+                              (1 << THR_GOLD)
+};
+
+typedef enum {
+  DIAMOND = 0,
+  NSTEP = 1,
+  HEX = 2,
+  BIGDIA = 3,
+  SQUARE = 4,
+  FAST_HEX = 5,
+  FAST_DIAMOND = 6
+} SEARCH_METHODS;
+
+typedef enum {
+  // No recode.
+  DISALLOW_RECODE = 0,
+  // Allow recode for KF and exceeding maximum frame bandwidth.
+  ALLOW_RECODE_KFMAXBW = 1,
+  // Allow recode only for KF/ARF/GF frames.
+  ALLOW_RECODE_KFARFGF = 2,
+  // Allow recode for all frames based on bitrate constraints.
+  ALLOW_RECODE = 3,
+} RECODE_LOOP_TYPE;
+
+typedef enum {
+  SUBPEL_TREE = 0,
+  SUBPEL_TREE_PRUNED = 1,           // Prunes 1/2-pel searches
+  SUBPEL_TREE_PRUNED_MORE = 2,      // Prunes 1/2-pel searches more aggressively
+  SUBPEL_TREE_PRUNED_EVENMORE = 3,  // Prunes 1/2- and 1/4-pel searches
+  // Other methods to come
+} SUBPEL_SEARCH_METHODS;
+
+typedef enum {
+  NO_MOTION_THRESHOLD = 0,
+  LOW_MOTION_THRESHOLD = 7
+} MOTION_THRESHOLD;
+
+typedef enum {
+  USE_FULL_RD = 0,
+  USE_LARGESTALL,
+  USE_TX_8X8
+} TX_SIZE_SEARCH_METHOD;
+
+typedef enum {
+  NOT_IN_USE = 0,
+  RELAXED_NEIGHBORING_MIN_MAX = 1,
+  STRICT_NEIGHBORING_MIN_MAX = 2
+} AUTO_MIN_MAX_MODE;
+
+typedef enum {
+  // Try the full image with different values.
+  LPF_PICK_FROM_FULL_IMAGE,
+  // Try a small portion of the image with different values.
+  LPF_PICK_FROM_SUBIMAGE,
+  // Estimate the level based on quantizer and frame type
+  LPF_PICK_FROM_Q,
+  // Pick 0 to disable LPF if LPF was enabled last frame
+  LPF_PICK_MINIMAL_LPF
+} LPF_PICK_METHOD;
+
+typedef enum {
+  // Terminate search early based on distortion so far compared to
+  // qp step, distortion in the neighborhood of the frame, etc.
+  FLAG_EARLY_TERMINATE = 1 << 0,
+
+  // Skips comp inter modes if the best so far is an intra mode.
+  FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
+
+  // Skips oblique intra modes if the best so far is an inter mode.
+  FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
+
+  // Skips oblique intra modes  at angles 27, 63, 117, 153 if the best
+  // intra so far is not one of the neighboring directions.
+  FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
+
+  // Skips intra modes other than DC_PRED if the source variance is small
+  FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
+} MODE_SEARCH_SKIP_LOGIC;
+
+typedef enum {
+  FLAG_SKIP_EIGHTTAP = 1 << EIGHTTAP,
+  FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
+  FLAG_SKIP_EIGHTTAP_SHARP = 1 << EIGHTTAP_SHARP,
+} INTERP_FILTER_MASK;
+
+typedef enum {
+  // Search partitions using RD/NONRD criterion
+  SEARCH_PARTITION,
+
+  // Always use a fixed size partition
+  FIXED_PARTITION,
+
+  REFERENCE_PARTITION,
+
+  // Use an arbitrary partitioning scheme based on source variance within
+  // a 64X64 SB
+  VAR_BASED_PARTITION,
+
+  // Use non-fixed partitions based on source variance
+  SOURCE_VAR_BASED_PARTITION
+} PARTITION_SEARCH_TYPE;
+
+typedef enum {
+  // Does a dry run to see if any of the contexts need to be updated or not,
+  // before the final run.
+  TWO_LOOP = 0,
+
+  // No dry run, also only half the coef contexts and bands are updated.
+  // The rest are not updated at all.
+  ONE_LOOP_REDUCED = 1
+} FAST_COEFF_UPDATE;
+
+typedef struct MV_SPEED_FEATURES {
+  // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
+  SEARCH_METHODS search_method;
+
+  // This parameter controls which step in the n-step process we start at.
+  // It's changed adaptively based on circumstances.
+  int reduce_first_step_size;
+
+  // If this is set to 1, we limit the motion search range to 2 times the
+  // largest motion vector found in the last frame.
+  int auto_mv_step_size;
+
+  // Subpel_search_method can only be subpel_tree which does a subpixel
+  // logarithmic search that keeps stepping at 1/2 pixel units until
+  // you stop getting a gain, and then goes on to 1/4 and repeats
+  // the same process. Along the way it skips many diagonals.
+  SUBPEL_SEARCH_METHODS subpel_search_method;
+
+  // Maximum number of steps in logarithmic subpel search before giving up.
+  int subpel_iters_per_step;
+
+  // Control when to stop subpel search
+  int subpel_force_stop;
+
+  // This variable sets the step_param used in full pel motion search.
+  int fullpel_search_step_param;
+} MV_SPEED_FEATURES;
+
+typedef struct SPEED_FEATURES {
+  MV_SPEED_FEATURES mv;
+
+  // Frame level coding parameter update
+  int frame_parameter_update;
+
+  RECODE_LOOP_TYPE recode_loop;
+
+  // Trellis (dynamic programming) optimization of quantized values (+1, 0).
+  int optimize_coefficients;
+
+  // Always set to 0. If on it enables 0 cost background transmission
+  // (except for the initial transmission of the segmentation). The feature is
+  // disabled because the addition of very large block sizes make the
+  // backgrounds very to cheap to encode, and the segmentation we have
+  // adds overhead.
+  int static_segmentation;
+
+  // If 1 we iterate finding a best reference for 2 ref frames together - via
+  // a log search that iterates 4 times (check around mv for last for best
+  // error of combined predictor then check around mv for alt). If 0 we
+  // we just use the best motion vector found for each frame by itself.
+  BLOCK_SIZE comp_inter_joint_search_thresh;
+
+  // This variable is used to cap the maximum number of times we skip testing a
+  // mode to be evaluated. A high value means we will be faster.
+  int adaptive_rd_thresh;
+
+  // Enables skipping the reconstruction step (idct, recon) in the
+  // intermediate steps assuming the last frame didn't have too many intra
+  // blocks and the q is less than a threshold.
+  int skip_encode_sb;
+  int skip_encode_frame;
+  // Speed feature to allow or disallow skipping of recode at block
+  // level within a frame.
+  int allow_skip_recode;
+
+  // Coefficient probability model approximation step size
+  int coeff_prob_appx_step;
+
+  // The threshold is to determine how slow the motino is, it is used when
+  // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
+  MOTION_THRESHOLD lf_motion_threshold;
+
+  // Determine which method we use to determine transform size. We can choose
+  // between options like full rd, largest for prediction size, largest
+  // for intra and model coefs for the rest.
+  TX_SIZE_SEARCH_METHOD tx_size_search_method;
+
+  // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
+  // precise but significantly faster than the non lp version.
+  int use_lp32x32fdct;
+
+  // After looking at the first set of modes (set by index here), skip
+  // checking modes for reference frames that don't match the reference frame
+  // of the best so far.
+  int mode_skip_start;
+
+  // TODO(JBB): Remove this.
+  int reference_masking;
+
+  PARTITION_SEARCH_TYPE partition_search_type;
+
+  // Used if partition_search_type = FIXED_SIZE_PARTITION
+  BLOCK_SIZE always_this_block_size;
+
+  // Skip rectangular partition test when partition type none gives better
+  // rd than partition type split.
+  int less_rectangular_check;
+
+  // Disable testing non square partitions. (eg 16x32)
+  int use_square_partition_only;
+
+  // Sets min and max partition sizes for this 64x64 region based on the
+  // same 64x64 in last encoded frame, and the left and above neighbor.
+  AUTO_MIN_MAX_MODE auto_min_max_partition_size;
+  // Ensures the rd based auto partition search will always
+  // go down at least to the specified level.
+  BLOCK_SIZE rd_auto_partition_min_limit;
+
+  // Min and max partition size we enable (block_size) as per auto
+  // min max, but also used by adjust partitioning, and pick_partitioning.
+  BLOCK_SIZE default_min_partition_size;
+  BLOCK_SIZE default_max_partition_size;
+
+  // Whether or not we allow partitions one smaller or one greater than the last
+  // frame's partitioning. Only used if use_lastframe_partitioning is set.
+  int adjust_partitioning_from_last_frame;
+
+  // How frequently we re do the partitioning from scratch. Only used if
+  // use_lastframe_partitioning is set.
+  int last_partitioning_redo_frequency;
+
+  // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
+  // it always, to allow it for only Last frame and Intra, disable it for all
+  // inter modes or to enable it always.
+  int disable_split_mask;
+
+  // TODO(jingning): combine the related motion search speed features
+  // This allows us to use motion search at other sizes as a starting
+  // point for this motion search and limits the search range around it.
+  int adaptive_motion_search;
+
+  int schedule_mode_search;
+
+  // Allows sub 8x8 modes to use the prediction filter that was determined
+  // best for 8x8 mode. If set to 0 we always re check all the filters for
+  // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
+  // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
+  int adaptive_pred_interp_filter;
+
+  // Adaptive prediction mode search
+  int adaptive_mode_search;
+
+  // Chessboard pattern prediction filter type search
+  int cb_pred_filter_search;
+
+  int cb_partition_search;
+
+  int motion_field_mode_search;
+
+  int alt_ref_search_fp;
+
+  // Fast quantization process path
+  int use_quant_fp;
+
+  // Use finer quantizer in every other few frames that run variable block
+  // partition type search.
+  int force_frame_boost;
+
+  // Maximally allowed base quantization index fluctuation.
+  int max_delta_qindex;
+
+  // Implements various heuristics to skip searching modes
+  // The heuristics selected are based on  flags
+  // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
+  unsigned int mode_search_skip_flags;
+
+  // A source variance threshold below which filter search is disabled
+  // Choose a very large value (UINT_MAX) to use 8-tap always
+  unsigned int disable_filter_search_var_thresh;
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // transform size separately.
+  int intra_y_mode_mask[TX_SIZES];
+  int intra_uv_mode_mask[TX_SIZES];
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // prediction block size separately.
+  int intra_y_mode_bsize_mask[BLOCK_SIZES];
+
+  // This variable enables an early break out of mode testing if the model for
+  // rd built from the prediction signal indicates a value that's much
+  // higher than the best rd we've seen so far.
+  int use_rd_breakout;
+
+  // This enables us to use an estimate for intra rd based on dc mode rather
+  // than choosing an actual uv mode in the stage of encoding before the actual
+  // final encode.
+  int use_uv_intra_rd_estimate;
+
+  // This feature controls how the loop filter level is determined.
+  LPF_PICK_METHOD lpf_pick;
+
+  // This feature limits the number of coefficients updates we actually do
+  // by only looking at counts from 1/2 the bands.
+  FAST_COEFF_UPDATE use_fast_coef_updates;
+
+  // This flag controls the use of non-RD mode decision.
+  int use_nonrd_pick_mode;
+
+  // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
+  // modes are used in order from LSB to MSB for each BLOCK_SIZE.
+  int inter_mode_mask[BLOCK_SIZES];
+
+  // This feature controls whether we do the expensive context update and
+  // calculation in the rd coefficient costing loop.
+  int use_fast_coef_costing;
+
+  // This feature controls the tolerence vs target used in deciding whether to
+  // recode a frame. It has no meaning if recode is disabled.
+  int recode_tolerance;
+
+  // This variable controls the maximum block size where intra blocks can be
+  // used in inter frames.
+  // TODO(aconverse): Fold this into one of the other many mode skips
+  BLOCK_SIZE max_intra_bsize;
+
+  // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
+  // FIXED_PARTITION search type should be used.
+  int search_type_check_frequency;
+
+  // When partition is pre-set, the inter prediction result from pick_inter_mode
+  // can be reused in final block encoding process. It is enabled only for real-
+  // time mode speed 6.
+  int reuse_inter_pred_sby;
+
+  // This variable sets the encode_breakout threshold. Currently, it is only
+  // enabled in real time mode.
+  int encode_breakout_thresh;
+
+  // default interp filter choice
+  INTERP_FILTER default_interp_filter;
+
+  // Early termination in transform size search, which only applies while
+  // tx_size_search_method is USE_FULL_RD.
+  int tx_size_search_breakout;
+
+  // adaptive interp_filter search to allow skip of certain filter types.
+  int adaptive_interp_filter_search;
+
+  // mask for skip evaluation of certain interp_filter type.
+  INTERP_FILTER_MASK interp_filter_search_mask;
+
+  // Partition search early breakout thresholds.
+  int64_t partition_search_breakout_dist_thr;
+  int partition_search_breakout_rate_thr;
+
+  // Allow skipping partition search for still image frame
+  int allow_partition_search_skip;
+} SPEED_FEATURES;
+
+struct VP9_COMP;
+
+void vp9_set_speed_features_framesize_independent(struct VP9_COMP *cpi);
+void vp9_set_speed_features_framesize_dependent(struct VP9_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SPEED_FEATURES_H_

media/libvpx/vp9/encoder/vp9_ssim.h

@@ -0,0 +1,96 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_SSIM_H_
+#define VP9_ENCODER_VP9_SSIM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vpx_scale/yv12config.h"
+
+// metrics used for calculating ssim, ssim2, dssim, and ssimc
+typedef struct {
+  // source sum ( over 8x8 region )
+  uint64_t sum_s;
+
+  // reference sum (over 8x8 region )
+  uint64_t sum_r;
+
+  // source sum squared ( over 8x8 region )
+  uint64_t sum_sq_s;
+
+  // reference sum squared (over 8x8 region )
+  uint64_t sum_sq_r;
+
+  // sum of source times reference (over 8x8 region)
+  uint64_t sum_sxr;
+
+  // calculated ssim score between source and reference
+  double ssim;
+} Ssimv;
+
+// metrics collected on a frame basis
+typedef struct {
+  // ssim consistency error metric ( see code for explanation )
+  double ssimc;
+
+  // standard ssim
+  double ssim;
+
+  // revised ssim ( see code for explanation)
+  double ssim2;
+
+  // ssim restated as an error metric like sse
+  double dssim;
+
+  // dssim converted to decibels
+  double dssimd;
+
+  // ssimc converted to decibels
+  double ssimcd;
+} Metrics;
+
+double vp9_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2,
+                      int img2_pitch, int width, int height, Ssimv *sv2,
+                      Metrics *m, int do_inconsistency);
+
+double vp9_calc_ssim(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+                     double *weight);
+
+double vp9_calc_ssimg(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+                      double *ssim_y, double *ssim_u, double *ssim_v);
+
+double vp9_calc_fastssim(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+                         double *ssim_y, double *ssim_u, double *ssim_v);
+
+double vp9_psnrhvs(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest,
+                   double *ssim_y, double *ssim_u, double *ssim_v);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+double vp9_highbd_calc_ssim(YV12_BUFFER_CONFIG *source,
+                            YV12_BUFFER_CONFIG *dest,
+                            double *weight,
+                            unsigned int bd);
+
+double vp9_highbd_calc_ssimg(YV12_BUFFER_CONFIG *source,
+                             YV12_BUFFER_CONFIG *dest,
+                             double *ssim_y,
+                             double *ssim_u,
+                             double *ssim_v,
+                             unsigned int bd);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SSIM_H_

media/libvpx/vp9/encoder/vp9_subexp.c

@@ -0,0 +1,214 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_subexp.h"
+#include "vp9/encoder/vp9_writer.h"
+
+#define vp9_cost_upd256  ((int)(vp9_cost_one(upd) - vp9_cost_zero(upd)))
+
+static const int update_bits[255] = {
+   5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,
+   6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,  0,
+};
+
+static int recenter_nonneg(int v, int m) {
+  if (v > (m << 1))
+    return v;
+  else if (v >= m)
+    return ((v - m) << 1);
+  else
+    return ((m - v) << 1) - 1;
+}
+
+static int remap_prob(int v, int m) {
+  int i;
+  static const int map_table[MAX_PROB - 1] = {
+    // generated by:
+    //   map_table[j] = split_index(j, MAX_PROB - 1, MODULUS_PARAM);
+     20,  21,  22,  23,  24,  25,   0,  26,  27,  28,  29,  30,  31,  32,  33,
+     34,  35,  36,  37,   1,  38,  39,  40,  41,  42,  43,  44,  45,  46,  47,
+     48,  49,   2,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,
+      3,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,   4,  74,
+     75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,   5,  86,  87,  88,
+     89,  90,  91,  92,  93,  94,  95,  96,  97,   6,  98,  99, 100, 101, 102,
+    103, 104, 105, 106, 107, 108, 109,   7, 110, 111, 112, 113, 114, 115, 116,
+    117, 118, 119, 120, 121,   8, 122, 123, 124, 125, 126, 127, 128, 129, 130,
+    131, 132, 133,   9, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144,
+    145,  10, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157,  11,
+    158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169,  12, 170, 171,
+    172, 173, 174, 175, 176, 177, 178, 179, 180, 181,  13, 182, 183, 184, 185,
+    186, 187, 188, 189, 190, 191, 192, 193,  14, 194, 195, 196, 197, 198, 199,
+    200, 201, 202, 203, 204, 205,  15, 206, 207, 208, 209, 210, 211, 212, 213,
+    214, 215, 216, 217,  16, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227,
+    228, 229,  17, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241,
+     18, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253,  19,
+  };
+  v--;
+  m--;
+  if ((m << 1) <= MAX_PROB)
+    i = recenter_nonneg(v, m) - 1;
+  else
+    i = recenter_nonneg(MAX_PROB - 1 - v, MAX_PROB - 1 - m) - 1;
+
+  i = map_table[i];
+  return i;
+}
+
+static int prob_diff_update_cost(vp9_prob newp, vp9_prob oldp) {
+  int delp = remap_prob(newp, oldp);
+  return update_bits[delp] * 256;
+}
+
+static void encode_uniform(vp9_writer *w, int v) {
+  const int l = 8;
+  const int m = (1 << l) - 191;
+  if (v < m) {
+    vp9_write_literal(w, v, l - 1);
+  } else {
+    vp9_write_literal(w, m + ((v - m) >> 1), l - 1);
+    vp9_write_literal(w, (v - m) & 1, 1);
+  }
+}
+
+static INLINE int write_bit_gte(vp9_writer *w, int word, int test) {
+  vp9_write_literal(w, word >= test, 1);
+  return word >= test;
+}
+
+static void encode_term_subexp(vp9_writer *w, int word) {
+  if (!write_bit_gte(w, word, 16)) {
+    vp9_write_literal(w, word, 4);
+  } else if (!write_bit_gte(w, word, 32)) {
+    vp9_write_literal(w, word - 16, 4);
+  } else if (!write_bit_gte(w, word, 64)) {
+    vp9_write_literal(w, word - 32, 5);
+  } else {
+    encode_uniform(w, word - 64);
+  }
+}
+
+void vp9_write_prob_diff_update(vp9_writer *w, vp9_prob newp, vp9_prob oldp) {
+  const int delp = remap_prob(newp, oldp);
+  encode_term_subexp(w, delp);
+}
+
+int vp9_prob_diff_update_savings_search(const unsigned int *ct,
+                                        vp9_prob oldp, vp9_prob *bestp,
+                                        vp9_prob upd) {
+  const int old_b = cost_branch256(ct, oldp);
+  int bestsavings = 0;
+  vp9_prob newp, bestnewp = oldp;
+  const int step = *bestp > oldp ? -1 : 1;
+
+  for (newp = *bestp; newp != oldp; newp += step) {
+    const int new_b = cost_branch256(ct, newp);
+    const int update_b = prob_diff_update_cost(newp, oldp) + vp9_cost_upd256;
+    const int savings = old_b - new_b - update_b;
+    if (savings > bestsavings) {
+      bestsavings = savings;
+      bestnewp = newp;
+    }
+  }
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const vp9_prob *oldp,
+                                              vp9_prob *bestp,
+                                              vp9_prob upd,
+                                              int stepsize) {
+  int i, old_b, new_b, update_b, savings, bestsavings, step;
+  int newp;
+  vp9_prob bestnewp, newplist[ENTROPY_NODES], oldplist[ENTROPY_NODES];
+  vp9_model_to_full_probs(oldp, oldplist);
+  memcpy(newplist, oldp, sizeof(vp9_prob) * UNCONSTRAINED_NODES);
+  for (i = UNCONSTRAINED_NODES, old_b = 0; i < ENTROPY_NODES; ++i)
+    old_b += cost_branch256(ct + 2 * i, oldplist[i]);
+  old_b += cost_branch256(ct + 2 * PIVOT_NODE, oldplist[PIVOT_NODE]);
+
+  bestsavings = 0;
+  bestnewp = oldp[PIVOT_NODE];
+
+  if (*bestp > oldp[PIVOT_NODE]) {
+    step = -stepsize;
+    for (newp = *bestp; newp > oldp[PIVOT_NODE]; newp += step) {
+      if (newp < 1 || newp > 255)
+        continue;
+      newplist[PIVOT_NODE] = newp;
+      vp9_model_to_full_probs(newplist, newplist);
+      for (i = UNCONSTRAINED_NODES, new_b = 0; i < ENTROPY_NODES; ++i)
+        new_b += cost_branch256(ct + 2 * i, newplist[i]);
+      new_b += cost_branch256(ct + 2 * PIVOT_NODE, newplist[PIVOT_NODE]);
+      update_b = prob_diff_update_cost(newp, oldp[PIVOT_NODE]) +
+          vp9_cost_upd256;
+      savings = old_b - new_b - update_b;
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  } else {
+    step = stepsize;
+    for (newp = *bestp; newp < oldp[PIVOT_NODE]; newp += step) {
+      if (newp < 1 || newp > 255)
+        continue;
+      newplist[PIVOT_NODE] = newp;
+      vp9_model_to_full_probs(newplist, newplist);
+      for (i = UNCONSTRAINED_NODES, new_b = 0; i < ENTROPY_NODES; ++i)
+        new_b += cost_branch256(ct + 2 * i, newplist[i]);
+      new_b += cost_branch256(ct + 2 * PIVOT_NODE, newplist[PIVOT_NODE]);
+      update_b = prob_diff_update_cost(newp, oldp[PIVOT_NODE]) +
+          vp9_cost_upd256;
+      savings = old_b - new_b - update_b;
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+void vp9_cond_prob_diff_update(vp9_writer *w, vp9_prob *oldp,
+                               const unsigned int ct[2]) {
+  const vp9_prob upd = DIFF_UPDATE_PROB;
+  vp9_prob newp = get_binary_prob(ct[0], ct[1]);
+  const int savings = vp9_prob_diff_update_savings_search(ct, *oldp, &newp,
+                                                          upd);
+  assert(newp >= 1);
+  if (savings > 0) {
+    vp9_write(w, 1, upd);
+    vp9_write_prob_diff_update(w, newp, *oldp);
+    *oldp = newp;
+  } else {
+    vp9_write(w, 0, upd);
+  }
+}

media/libvpx/vp9/encoder/vp9_subexp.h

@@ -0,0 +1,44 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_ENCODER_VP9_SUBEXP_H_
+#define VP9_ENCODER_VP9_SUBEXP_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "vp9/common/vp9_prob.h"
+
+struct vp9_writer;
+
+void vp9_write_prob_diff_update(struct vp9_writer *w,
+                                vp9_prob newp, vp9_prob oldp);
+
+void vp9_cond_prob_diff_update(struct vp9_writer *w, vp9_prob *oldp,
+                               const unsigned int ct[2]);
+
+int vp9_prob_diff_update_savings_search(const unsigned int *ct,
+                                        vp9_prob oldp, vp9_prob *bestp,
+                                        vp9_prob upd);
+
+
+int vp9_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const vp9_prob *oldp,
+                                              vp9_prob *bestp,
+                                              vp9_prob upd,
+                                              int stepsize);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SUBEXP_H_

media/libvpx/vp9/encoder/vp9_svc_layercontext.c

@@ -0,0 +1,646 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_svc_layercontext.h"
+#include "vp9/encoder/vp9_extend.h"
+
+#define SMALL_FRAME_FB_IDX 7
+#define SMALL_FRAME_WIDTH  16
+#define SMALL_FRAME_HEIGHT 16
+
+void vp9_init_layer_context(VP9_COMP *const cpi) {
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  int sl, tl;
+  int alt_ref_idx = svc->number_spatial_layers;
+
+  svc->spatial_layer_id = 0;
+  svc->temporal_layer_id = 0;
+
+  if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2) {
+    if (vp9_realloc_frame_buffer(&cpi->svc.empty_frame.img,
+                                 SMALL_FRAME_WIDTH, SMALL_FRAME_HEIGHT,
+                                 cpi->common.subsampling_x,
+                                 cpi->common.subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 cpi->common.use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 cpi->common.byte_alignment,
+                                 NULL, NULL, NULL))
+      vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                         "Failed to allocate empty frame for multiple frame "
+                         "contexts");
+
+    memset(cpi->svc.empty_frame.img.buffer_alloc, 0x80,
+           cpi->svc.empty_frame.img.buffer_alloc_sz);
+  }
+
+  for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
+    for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+      int layer = LAYER_IDS_TO_IDX(sl, tl, oxcf->ts_number_layers);
+      LAYER_CONTEXT *const lc = &svc->layer_context[layer];
+      RATE_CONTROL *const lrc = &lc->rc;
+      int i;
+      lc->current_video_frame_in_layer = 0;
+      lc->layer_size = 0;
+      lc->frames_from_key_frame = 0;
+      lc->last_frame_type = FRAME_TYPES;
+      lrc->ni_av_qi = oxcf->worst_allowed_q;
+      lrc->total_actual_bits = 0;
+      lrc->total_target_vs_actual = 0;
+      lrc->ni_tot_qi = 0;
+      lrc->tot_q = 0.0;
+      lrc->avg_q = 0.0;
+      lrc->ni_frames = 0;
+      lrc->decimation_count = 0;
+      lrc->decimation_factor = 0;
+
+      for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
+        lrc->rate_correction_factors[i] = 1.0;
+      }
+
+      if (cpi->oxcf.rc_mode == VPX_CBR) {
+        lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
+        lrc->last_q[INTER_FRAME] = oxcf->worst_allowed_q;
+        lrc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
+        lrc->avg_frame_qindex[KEY_FRAME] = oxcf->worst_allowed_q;
+      } else {
+        lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
+        lrc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
+        lrc->last_q[INTER_FRAME] = oxcf->best_allowed_q;
+        lrc->avg_frame_qindex[KEY_FRAME] = (oxcf->worst_allowed_q +
+                                            oxcf->best_allowed_q) / 2;
+        lrc->avg_frame_qindex[INTER_FRAME] = (oxcf->worst_allowed_q +
+                                              oxcf->best_allowed_q) / 2;
+        if (oxcf->ss_enable_auto_arf[sl])
+          lc->alt_ref_idx = alt_ref_idx++;
+        else
+          lc->alt_ref_idx = INVALID_IDX;
+        lc->gold_ref_idx = INVALID_IDX;
+      }
+
+      lrc->buffer_level = oxcf->starting_buffer_level_ms *
+                              lc->target_bandwidth / 1000;
+      lrc->bits_off_target = lrc->buffer_level;
+    }
+  }
+
+  // Still have extra buffer for base layer golden frame
+  if (!(svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR)
+      && alt_ref_idx < REF_FRAMES)
+    svc->layer_context[0].gold_ref_idx = alt_ref_idx;
+}
+
+// Update the layer context from a change_config() call.
+void vp9_update_layer_context_change_config(VP9_COMP *const cpi,
+                                            const int target_bandwidth) {
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int sl, tl, layer = 0, spatial_layer_target;
+  float bitrate_alloc = 1.0;
+
+  if (svc->temporal_layering_mode != VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING) {
+    for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
+      spatial_layer_target = 0;
+
+      for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+        layer = LAYER_IDS_TO_IDX(sl, tl, oxcf->ts_number_layers);
+        svc->layer_context[layer].target_bandwidth =
+            oxcf->layer_target_bitrate[layer];
+      }
+
+      layer = LAYER_IDS_TO_IDX(sl, ((oxcf->ts_number_layers - 1) < 0 ?
+          0 : (oxcf->ts_number_layers - 1)), oxcf->ts_number_layers);
+      spatial_layer_target =
+          svc->layer_context[layer].target_bandwidth =
+              oxcf->layer_target_bitrate[layer];
+
+      for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
+        LAYER_CONTEXT *const lc =
+            &svc->layer_context[sl * oxcf->ts_number_layers + tl];
+        RATE_CONTROL *const lrc = &lc->rc;
+
+        lc->spatial_layer_target_bandwidth = spatial_layer_target;
+        bitrate_alloc = (float)lc->target_bandwidth / spatial_layer_target;
+        lrc->starting_buffer_level =
+            (int64_t)(rc->starting_buffer_level * bitrate_alloc);
+        lrc->optimal_buffer_level =
+            (int64_t)(rc->optimal_buffer_level * bitrate_alloc);
+        lrc->maximum_buffer_size =
+            (int64_t)(rc->maximum_buffer_size * bitrate_alloc);
+        lrc->bits_off_target =
+            MIN(lrc->bits_off_target, lrc->maximum_buffer_size);
+        lrc->buffer_level = MIN(lrc->buffer_level, lrc->maximum_buffer_size);
+        lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[tl];
+        lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+        lrc->max_frame_bandwidth = rc->max_frame_bandwidth;
+        lrc->worst_quality = rc->worst_quality;
+        lrc->best_quality = rc->best_quality;
+      }
+    }
+  } else {
+    int layer_end;
+    float bitrate_alloc = 1.0;
+
+    if (svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) {
+      layer_end = svc->number_temporal_layers;
+    } else {
+      layer_end = svc->number_spatial_layers;
+    }
+
+    for (layer = 0; layer < layer_end; ++layer) {
+      LAYER_CONTEXT *const lc = &svc->layer_context[layer];
+      RATE_CONTROL *const lrc = &lc->rc;
+
+      lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
+
+      bitrate_alloc = (float)lc->target_bandwidth / target_bandwidth;
+      // Update buffer-related quantities.
+      lrc->starting_buffer_level =
+          (int64_t)(rc->starting_buffer_level * bitrate_alloc);
+      lrc->optimal_buffer_level =
+          (int64_t)(rc->optimal_buffer_level * bitrate_alloc);
+      lrc->maximum_buffer_size =
+          (int64_t)(rc->maximum_buffer_size * bitrate_alloc);
+      lrc->bits_off_target = MIN(lrc->bits_off_target,
+                                 lrc->maximum_buffer_size);
+      lrc->buffer_level = MIN(lrc->buffer_level, lrc->maximum_buffer_size);
+      // Update framerate-related quantities.
+      if (svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) {
+        lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[layer];
+      } else {
+        lc->framerate = cpi->framerate;
+      }
+      lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+      lrc->max_frame_bandwidth = rc->max_frame_bandwidth;
+      // Update qp-related quantities.
+      lrc->worst_quality = rc->worst_quality;
+      lrc->best_quality = rc->best_quality;
+    }
+  }
+}
+
+static LAYER_CONTEXT *get_layer_context(VP9_COMP *const cpi) {
+  if (is_one_pass_cbr_svc(cpi))
+    return &cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+        cpi->svc.number_temporal_layers + cpi->svc.temporal_layer_id];
+  else
+    return (cpi->svc.number_temporal_layers > 1 &&
+            cpi->oxcf.rc_mode == VPX_CBR) ?
+             &cpi->svc.layer_context[cpi->svc.temporal_layer_id] :
+             &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
+}
+
+void vp9_update_temporal_layer_framerate(VP9_COMP *const cpi) {
+  SVC *const svc = &cpi->svc;
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+  RATE_CONTROL *const lrc = &lc->rc;
+  // Index into spatial+temporal arrays.
+  const int st_idx = svc->spatial_layer_id * svc->number_temporal_layers +
+      svc->temporal_layer_id;
+  const int tl = svc->temporal_layer_id;
+
+  lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[tl];
+  lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+  lrc->max_frame_bandwidth = cpi->rc.max_frame_bandwidth;
+  // Update the average layer frame size (non-cumulative per-frame-bw).
+  if (tl == 0) {
+    lc->avg_frame_size = lrc->avg_frame_bandwidth;
+  } else {
+    const double prev_layer_framerate =
+        cpi->framerate / oxcf->ts_rate_decimator[tl - 1];
+    const int prev_layer_target_bandwidth =
+        oxcf->layer_target_bitrate[st_idx - 1];
+    lc->avg_frame_size =
+        (int)((lc->target_bandwidth - prev_layer_target_bandwidth) /
+              (lc->framerate - prev_layer_framerate));
+  }
+}
+
+void vp9_update_spatial_layer_framerate(VP9_COMP *const cpi, double framerate) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+  RATE_CONTROL *const lrc = &lc->rc;
+
+  lc->framerate = framerate;
+  lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
+  lrc->min_frame_bandwidth = (int)(lrc->avg_frame_bandwidth *
+                                   oxcf->two_pass_vbrmin_section / 100);
+  lrc->max_frame_bandwidth = (int)(((int64_t)lrc->avg_frame_bandwidth *
+                                   oxcf->two_pass_vbrmax_section) / 100);
+  vp9_rc_set_gf_interval_range(cpi, lrc);
+}
+
+void vp9_restore_layer_context(VP9_COMP *const cpi) {
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+  const int old_frame_since_key = cpi->rc.frames_since_key;
+  const int old_frame_to_key = cpi->rc.frames_to_key;
+
+  cpi->rc = lc->rc;
+  cpi->twopass = lc->twopass;
+  cpi->oxcf.target_bandwidth = lc->target_bandwidth;
+  cpi->alt_ref_source = lc->alt_ref_source;
+  // Reset the frames_since_key and frames_to_key counters to their values
+  // before the layer restore. Keep these defined for the stream (not layer).
+  if (cpi->svc.number_temporal_layers > 1) {
+    cpi->rc.frames_since_key = old_frame_since_key;
+    cpi->rc.frames_to_key = old_frame_to_key;
+  }
+}
+
+void vp9_save_layer_context(VP9_COMP *const cpi) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  LAYER_CONTEXT *const lc = get_layer_context(cpi);
+
+  lc->rc = cpi->rc;
+  lc->twopass = cpi->twopass;
+  lc->target_bandwidth = (int)oxcf->target_bandwidth;
+  lc->alt_ref_source = cpi->alt_ref_source;
+}
+
+void vp9_init_second_pass_spatial_svc(VP9_COMP *cpi) {
+  SVC *const svc = &cpi->svc;
+  int i;
+
+  for (i = 0; i < svc->number_spatial_layers; ++i) {
+    TWO_PASS *const twopass = &svc->layer_context[i].twopass;
+
+    svc->spatial_layer_id = i;
+    vp9_init_second_pass(cpi);
+
+    twopass->total_stats.spatial_layer_id = i;
+    twopass->total_left_stats.spatial_layer_id = i;
+  }
+  svc->spatial_layer_id = 0;
+}
+
+void vp9_inc_frame_in_layer(VP9_COMP *const cpi) {
+  LAYER_CONTEXT *const lc =
+      &cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+                              cpi->svc.number_temporal_layers];
+  ++lc->current_video_frame_in_layer;
+  ++lc->frames_from_key_frame;
+}
+
+int vp9_is_upper_layer_key_frame(const VP9_COMP *const cpi) {
+  return is_two_pass_svc(cpi) &&
+         cpi->svc.spatial_layer_id > 0 &&
+         cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+                                cpi->svc.number_temporal_layers +
+                                cpi->svc.temporal_layer_id].is_key_frame;
+}
+
+static void get_layer_resolution(const int width_org, const int height_org,
+                                 const int num, const int den,
+                                 int *width_out, int *height_out) {
+  int w, h;
+
+  if (width_out == NULL || height_out == NULL || den == 0)
+    return;
+
+  w = width_org * num / den;
+  h = height_org * num / den;
+
+  // make height and width even to make chrome player happy
+  w += w % 2;
+  h += h % 2;
+
+  *width_out = w;
+  *height_out = h;
+}
+
+// The function sets proper ref_frame_flags, buffer indices, and buffer update
+// variables for temporal layering mode 3 - that does 0-2-1-2 temporal layering
+// scheme.
+static void set_flags_and_fb_idx_for_temporal_mode3(VP9_COMP *const cpi) {
+  int frame_num_within_temporal_struct = 0;
+  int spatial_id, temporal_id;
+  spatial_id = cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  frame_num_within_temporal_struct =
+      cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+      cpi->svc.number_temporal_layers].current_video_frame_in_layer % 4;
+  temporal_id = cpi->svc.temporal_layer_id =
+      (frame_num_within_temporal_struct & 1) ? 2 :
+      (frame_num_within_temporal_struct >> 1);
+  cpi->ext_refresh_last_frame = cpi->ext_refresh_golden_frame =
+      cpi->ext_refresh_alt_ref_frame = 0;
+  if (!temporal_id) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_last_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else if (cpi->svc.layer_context[temporal_id].is_key_frame) {
+      // base layer is a key frame.
+      cpi->ref_frame_flags = VP9_GOLD_FLAG;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  } else if (temporal_id == 1) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_alt_ref_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  } else {
+    if (frame_num_within_temporal_struct == 1) {
+      // the first tl2 picture
+      if (!spatial_id) {
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ext_refresh_alt_ref_frame = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG;
+      } else if (spatial_id < cpi->svc.number_spatial_layers - 1) {
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ext_refresh_alt_ref_frame = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+      } else {  // Top layer
+        cpi->ext_refresh_frame_flags_pending = 0;
+        cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+      }
+    } else {
+      //  The second tl2 picture
+      if (!spatial_id) {
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG;
+        cpi->ext_refresh_last_frame = 1;
+      } else if (spatial_id < cpi->svc.number_spatial_layers - 1) {
+        cpi->ext_refresh_frame_flags_pending = 1;
+        cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+        cpi->ext_refresh_last_frame = 1;
+      } else {  // top layer
+        cpi->ext_refresh_frame_flags_pending = 0;
+        cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+      }
+    }
+  }
+  if (temporal_id == 0) {
+    cpi->lst_fb_idx = spatial_id;
+    if (spatial_id)
+      cpi->gld_fb_idx = spatial_id - 1;
+    else
+      cpi->gld_fb_idx = 0;
+    cpi->alt_fb_idx = 0;
+  } else if (temporal_id == 1) {
+    cpi->lst_fb_idx = spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+  } else if (frame_num_within_temporal_struct == 1) {
+    cpi->lst_fb_idx = spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+  } else {
+    cpi->lst_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = 0;
+  }
+}
+
+// The function sets proper ref_frame_flags, buffer indices, and buffer update
+// variables for temporal layering mode 2 - that does 0-1-0-1 temporal layering
+// scheme.
+static void set_flags_and_fb_idx_for_temporal_mode2(VP9_COMP *const cpi) {
+  int spatial_id, temporal_id;
+  spatial_id = cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  temporal_id = cpi->svc.temporal_layer_id =
+      cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+      cpi->svc.number_temporal_layers].current_video_frame_in_layer & 1;
+  cpi->ext_refresh_last_frame = cpi->ext_refresh_golden_frame =
+                                cpi->ext_refresh_alt_ref_frame = 0;
+  if (!temporal_id) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_last_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else if (cpi->svc.layer_context[temporal_id].is_key_frame) {
+      // base layer is a key frame.
+      cpi->ref_frame_flags = VP9_GOLD_FLAG;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  } else if (temporal_id == 1) {
+    cpi->ext_refresh_frame_flags_pending = 1;
+    cpi->ext_refresh_alt_ref_frame = 1;
+    if (!spatial_id) {
+      cpi->ref_frame_flags = VP9_LAST_FLAG;
+    } else {
+      cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+    }
+  }
+
+  if (temporal_id == 0) {
+    cpi->lst_fb_idx = spatial_id;
+    if (spatial_id)
+      cpi->gld_fb_idx = spatial_id - 1;
+    else
+      cpi->gld_fb_idx = 0;
+    cpi->alt_fb_idx = 0;
+  } else if (temporal_id == 1) {
+    cpi->lst_fb_idx = spatial_id;
+    cpi->gld_fb_idx = cpi->svc.number_spatial_layers + spatial_id - 1;
+    cpi->alt_fb_idx = cpi->svc.number_spatial_layers + spatial_id;
+  }
+}
+
+// The function sets proper ref_frame_flags, buffer indices, and buffer update
+// variables for temporal layering mode 0 - that has no temporal layering.
+static void set_flags_and_fb_idx_for_temporal_mode_noLayering(
+    VP9_COMP *const cpi) {
+  int spatial_id;
+  spatial_id = cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  cpi->ext_refresh_last_frame =
+      cpi->ext_refresh_golden_frame = cpi->ext_refresh_alt_ref_frame = 0;
+  cpi->ext_refresh_frame_flags_pending = 1;
+  cpi->ext_refresh_last_frame = 1;
+  if (!spatial_id) {
+    cpi->ref_frame_flags = VP9_LAST_FLAG;
+  } else if (cpi->svc.layer_context[0].is_key_frame) {
+    cpi->ref_frame_flags = VP9_GOLD_FLAG;
+  } else {
+    cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
+  }
+  cpi->lst_fb_idx = spatial_id;
+  if (spatial_id)
+    cpi->gld_fb_idx = spatial_id - 1;
+  else
+    cpi->gld_fb_idx = 0;
+}
+
+int vp9_one_pass_cbr_svc_start_layer(VP9_COMP *const cpi) {
+  int width = 0, height = 0;
+  LAYER_CONTEXT *lc = NULL;
+
+  if (cpi->svc.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_0212) {
+    set_flags_and_fb_idx_for_temporal_mode3(cpi);
+  } else if (cpi->svc.temporal_layering_mode ==
+           VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING) {
+    set_flags_and_fb_idx_for_temporal_mode_noLayering(cpi);
+  } else if (cpi->svc.temporal_layering_mode ==
+           VP9E_TEMPORAL_LAYERING_MODE_0101) {
+    set_flags_and_fb_idx_for_temporal_mode2(cpi);
+  } else if (cpi->svc.temporal_layering_mode ==
+      VP9E_TEMPORAL_LAYERING_MODE_BYPASS) {
+    // VP9E_TEMPORAL_LAYERING_MODE_BYPASS :
+    // if the code goes here, it means the encoder will be relying on the
+    // flags from outside for layering.
+    // However, since when spatial+temporal layering is used, the buffer indices
+    // cannot be derived automatically, the bypass mode will only work when the
+    // number of spatial layers equals 1.
+    assert(cpi->svc.number_spatial_layers == 1);
+  }
+
+  lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id *
+                               cpi->svc.number_temporal_layers +
+                               cpi->svc.temporal_layer_id];
+
+  get_layer_resolution(cpi->oxcf.width, cpi->oxcf.height,
+                       lc->scaling_factor_num, lc->scaling_factor_den,
+                       &width, &height);
+
+  if (vp9_set_size_literal(cpi, width, height) != 0)
+    return VPX_CODEC_INVALID_PARAM;
+
+  return 0;
+}
+
+#if CONFIG_SPATIAL_SVC
+int vp9_svc_start_frame(VP9_COMP *const cpi) {
+  int width = 0, height = 0;
+  LAYER_CONTEXT *lc;
+  struct lookahead_entry *buf;
+  int count = 1 << (cpi->svc.number_temporal_layers - 1);
+
+  cpi->svc.spatial_layer_id = cpi->svc.spatial_layer_to_encode;
+  lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
+
+  cpi->svc.temporal_layer_id = 0;
+  while ((lc->current_video_frame_in_layer % count) != 0) {
+    ++cpi->svc.temporal_layer_id;
+    count >>= 1;
+  }
+
+  cpi->ref_frame_flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
+
+  cpi->lst_fb_idx = cpi->svc.spatial_layer_id;
+
+  if (cpi->svc.spatial_layer_id == 0)
+    cpi->gld_fb_idx = (lc->gold_ref_idx >= 0) ?
+                      lc->gold_ref_idx : cpi->lst_fb_idx;
+  else
+    cpi->gld_fb_idx = cpi->svc.spatial_layer_id - 1;
+
+  if (lc->current_video_frame_in_layer == 0) {
+    if (cpi->svc.spatial_layer_id >= 2) {
+      cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
+    } else {
+      cpi->alt_fb_idx = cpi->lst_fb_idx;
+      cpi->ref_frame_flags &= (~VP9_LAST_FLAG & ~VP9_ALT_FLAG);
+    }
+  } else {
+    if (cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id]) {
+      cpi->alt_fb_idx = lc->alt_ref_idx;
+      if (!lc->has_alt_frame)
+        cpi->ref_frame_flags &= (~VP9_ALT_FLAG);
+    } else {
+      // Find a proper alt_fb_idx for layers that don't have alt ref frame
+      if (cpi->svc.spatial_layer_id == 0) {
+        cpi->alt_fb_idx = cpi->lst_fb_idx;
+      } else {
+        LAYER_CONTEXT *lc_lower =
+            &cpi->svc.layer_context[cpi->svc.spatial_layer_id - 1];
+
+        if (cpi->oxcf.ss_enable_auto_arf[cpi->svc.spatial_layer_id - 1] &&
+            lc_lower->alt_ref_source != NULL)
+          cpi->alt_fb_idx = lc_lower->alt_ref_idx;
+        else if (cpi->svc.spatial_layer_id >= 2)
+          cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
+        else
+          cpi->alt_fb_idx = cpi->lst_fb_idx;
+      }
+    }
+  }
+
+  get_layer_resolution(cpi->oxcf.width, cpi->oxcf.height,
+                       lc->scaling_factor_num, lc->scaling_factor_den,
+                       &width, &height);
+
+  // Workaround for multiple frame contexts. In some frames we can't use prev_mi
+  // since its previous frame could be changed during decoding time. The idea is
+  // we put a empty invisible frame in front of them, then we will not use
+  // prev_mi when encoding these frames.
+
+  buf = vp9_lookahead_peek(cpi->lookahead, 0);
+  if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2 &&
+      cpi->svc.encode_empty_frame_state == NEED_TO_ENCODE &&
+      lc->rc.frames_to_key != 0 &&
+      !(buf != NULL && (buf->flags & VPX_EFLAG_FORCE_KF))) {
+    if ((cpi->svc.number_temporal_layers > 1 &&
+         cpi->svc.temporal_layer_id < cpi->svc.number_temporal_layers - 1) ||
+        (cpi->svc.number_spatial_layers > 1 &&
+         cpi->svc.spatial_layer_id == 0)) {
+      struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead, 0);
+
+      if (buf != NULL) {
+        cpi->svc.empty_frame.ts_start = buf->ts_start;
+        cpi->svc.empty_frame.ts_end = buf->ts_end;
+        cpi->svc.encode_empty_frame_state = ENCODING;
+        cpi->common.show_frame = 0;
+        cpi->ref_frame_flags = 0;
+        cpi->common.frame_type = INTER_FRAME;
+        cpi->lst_fb_idx =
+            cpi->gld_fb_idx = cpi->alt_fb_idx = SMALL_FRAME_FB_IDX;
+
+        if (cpi->svc.encode_intra_empty_frame != 0)
+          cpi->common.intra_only = 1;
+
+        width = SMALL_FRAME_WIDTH;
+        height = SMALL_FRAME_HEIGHT;
+      }
+    }
+  }
+
+  cpi->oxcf.worst_allowed_q = vp9_quantizer_to_qindex(lc->max_q);
+  cpi->oxcf.best_allowed_q = vp9_quantizer_to_qindex(lc->min_q);
+
+  vp9_change_config(cpi, &cpi->oxcf);
+
+  if (vp9_set_size_literal(cpi, width, height) != 0)
+    return VPX_CODEC_INVALID_PARAM;
+
+  vp9_set_high_precision_mv(cpi, 1);
+
+  cpi->alt_ref_source = get_layer_context(cpi)->alt_ref_source;
+
+  return 0;
+}
+
+#endif
+
+struct lookahead_entry *vp9_svc_lookahead_pop(VP9_COMP *const cpi,
+                                              struct lookahead_ctx *ctx,
+                                              int drain) {
+  struct lookahead_entry *buf = NULL;
+  if (ctx->sz && (drain || ctx->sz == ctx->max_sz - MAX_PRE_FRAMES)) {
+    buf = vp9_lookahead_peek(ctx, 0);
+    if (buf != NULL) {
+      // Only remove the buffer when pop the highest layer.
+      if (cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1) {
+        vp9_lookahead_pop(ctx, drain);
+      }
+    }
+  }
+  return buf;
+}

media/libvpx/vp9/encoder/vp9_svc_layercontext.h

@@ -0,0 +1,122 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_SVC_LAYERCONTEXT_H_
+#define VP9_ENCODER_VP9_SVC_LAYERCONTEXT_H_
+
+#include "vpx/vpx_encoder.h"
+
+#include "vp9/encoder/vp9_ratectrl.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  RATE_CONTROL rc;
+  int target_bandwidth;
+  int spatial_layer_target_bandwidth;  // Target for the spatial layer.
+  double framerate;
+  int avg_frame_size;
+  int max_q;
+  int min_q;
+  int scaling_factor_num;
+  int scaling_factor_den;
+  TWO_PASS twopass;
+  vpx_fixed_buf_t rc_twopass_stats_in;
+  unsigned int current_video_frame_in_layer;
+  int is_key_frame;
+  int frames_from_key_frame;
+  FRAME_TYPE last_frame_type;
+  struct lookahead_entry  *alt_ref_source;
+  int alt_ref_idx;
+  int gold_ref_idx;
+  int has_alt_frame;
+  size_t layer_size;
+  struct vpx_psnr_pkt psnr_pkt;
+} LAYER_CONTEXT;
+
+typedef struct {
+  int spatial_layer_id;
+  int temporal_layer_id;
+  int number_spatial_layers;
+  int number_temporal_layers;
+
+  int spatial_layer_to_encode;
+
+  // Workaround for multiple frame contexts
+  enum {
+    ENCODED = 0,
+    ENCODING,
+    NEED_TO_ENCODE
+  }encode_empty_frame_state;
+  struct lookahead_entry empty_frame;
+  int encode_intra_empty_frame;
+
+  // Store scaled source frames to be used for temporal filter to generate
+  // a alt ref frame.
+  YV12_BUFFER_CONFIG scaled_frames[MAX_LAG_BUFFERS];
+
+  // Layer context used for rate control in one pass temporal CBR mode or
+  // two pass spatial mode.
+  LAYER_CONTEXT layer_context[VPX_MAX_LAYERS];
+  // Indicates what sort of temporal layering is used.
+  // Currently, this only works for CBR mode.
+  VP9E_TEMPORAL_LAYERING_MODE temporal_layering_mode;
+} SVC;
+
+struct VP9_COMP;
+
+// Initialize layer context data from init_config().
+void vp9_init_layer_context(struct VP9_COMP *const cpi);
+
+// Update the layer context from a change_config() call.
+void vp9_update_layer_context_change_config(struct VP9_COMP *const cpi,
+                                            const int target_bandwidth);
+
+// Prior to encoding the frame, update framerate-related quantities
+// for the current temporal layer.
+void vp9_update_temporal_layer_framerate(struct VP9_COMP *const cpi);
+
+// Update framerate-related quantities for the current spatial layer.
+void vp9_update_spatial_layer_framerate(struct VP9_COMP *const cpi,
+                                        double framerate);
+
+// Prior to encoding the frame, set the layer context, for the current layer
+// to be encoded, to the cpi struct.
+void vp9_restore_layer_context(struct VP9_COMP *const cpi);
+
+// Save the layer context after encoding the frame.
+void vp9_save_layer_context(struct VP9_COMP *const cpi);
+
+// Initialize second pass rc for spatial svc.
+void vp9_init_second_pass_spatial_svc(struct VP9_COMP *cpi);
+
+// Increment number of video frames in layer
+void vp9_inc_frame_in_layer(struct VP9_COMP *const cpi);
+
+// Check if current layer is key frame in spatial upper layer
+int vp9_is_upper_layer_key_frame(const struct VP9_COMP *const cpi);
+
+// Get the next source buffer to encode
+struct lookahead_entry *vp9_svc_lookahead_pop(struct VP9_COMP *const cpi,
+                                              struct lookahead_ctx *ctx,
+                                              int drain);
+
+// Start a frame and initialize svc parameters
+int vp9_svc_start_frame(struct VP9_COMP *const cpi);
+
+int vp9_one_pass_cbr_svc_start_layer(struct VP9_COMP *const cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_SVC_LAYERCONTEXT_

media/libvpx/vp9/encoder/vp9_temporal_filter.c

@@ -0,0 +1,750 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <limits.h>
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_systemdependent.h"
+#include "vp9/encoder/vp9_extend.h"
+#include "vp9/encoder/vp9_firstpass.h"
+#include "vp9/encoder/vp9_mcomp.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_quantize.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_segmentation.h"
+#include "vp9/encoder/vp9_temporal_filter.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/vpx_timer.h"
+#include "vpx_scale/vpx_scale.h"
+
+static int fixed_divide[512];
+
+static void temporal_filter_predictors_mb_c(MACROBLOCKD *xd,
+                                            uint8_t *y_mb_ptr,
+                                            uint8_t *u_mb_ptr,
+                                            uint8_t *v_mb_ptr,
+                                            int stride,
+                                            int uv_block_width,
+                                            int uv_block_height,
+                                            int mv_row,
+                                            int mv_col,
+                                            uint8_t *pred,
+                                            struct scale_factors *scale,
+                                            int x, int y) {
+  const int which_mv = 0;
+  const MV mv = { mv_row, mv_col };
+  const InterpKernel *const kernel =
+    vp9_get_interp_kernel(xd->mi[0]->mbmi.interp_filter);
+
+  enum mv_precision mv_precision_uv;
+  int uv_stride;
+  if (uv_block_width == 8) {
+    uv_stride = (stride + 1) >> 1;
+    mv_precision_uv = MV_PRECISION_Q4;
+  } else {
+    uv_stride = stride;
+    mv_precision_uv = MV_PRECISION_Q3;
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_highbd_build_inter_predictor(y_mb_ptr, stride,
+                                     &pred[0], 16,
+                                     &mv,
+                                     scale,
+                                     16, 16,
+                                     which_mv,
+                                     kernel, MV_PRECISION_Q3, x, y, xd->bd);
+
+    vp9_highbd_build_inter_predictor(u_mb_ptr, uv_stride,
+                                     &pred[256], uv_block_width,
+                                     &mv,
+                                     scale,
+                                     uv_block_width, uv_block_height,
+                                     which_mv,
+                                     kernel, mv_precision_uv, x, y, xd->bd);
+
+    vp9_highbd_build_inter_predictor(v_mb_ptr, uv_stride,
+                                     &pred[512], uv_block_width,
+                                     &mv,
+                                     scale,
+                                     uv_block_width, uv_block_height,
+                                     which_mv,
+                                     kernel, mv_precision_uv, x, y, xd->bd);
+    return;
+  }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  vp9_build_inter_predictor(y_mb_ptr, stride,
+                            &pred[0], 16,
+                            &mv,
+                            scale,
+                            16, 16,
+                            which_mv,
+                            kernel, MV_PRECISION_Q3, x, y);
+
+  vp9_build_inter_predictor(u_mb_ptr, uv_stride,
+                            &pred[256], uv_block_width,
+                            &mv,
+                            scale,
+                            uv_block_width, uv_block_height,
+                            which_mv,
+                            kernel, mv_precision_uv, x, y);
+
+  vp9_build_inter_predictor(v_mb_ptr, uv_stride,
+                            &pred[512], uv_block_width,
+                            &mv,
+                            scale,
+                            uv_block_width, uv_block_height,
+                            which_mv,
+                            kernel, mv_precision_uv, x, y);
+}
+
+void vp9_temporal_filter_init(void) {
+  int i;
+
+  fixed_divide[0] = 0;
+  for (i = 1; i < 512; ++i)
+    fixed_divide[i] = 0x80000 / i;
+}
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1,
+                                 unsigned int stride,
+                                 uint8_t *frame2,
+                                 unsigned int block_width,
+                                 unsigned int block_height,
+                                 int strength,
+                                 int filter_weight,
+                                 unsigned int *accumulator,
+                                 uint16_t *count) {
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int src_byte = frame1[byte];
+      int pixel_value = *frame2++;
+
+      modifier   = src_byte - pixel_value;
+      // This is an integer approximation of:
+      // float coeff = (3.0 * modifer * modifier) / pow(2, strength);
+      // modifier =  (int)roundf(coeff > 16 ? 0 : 16-coeff);
+      modifier  *= modifier;
+      modifier  *= 3;
+      modifier  += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16)
+        modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_temporal_filter_apply_c(uint8_t *frame1_8,
+                                        unsigned int stride,
+                                        uint8_t *frame2_8,
+                                        unsigned int block_width,
+                                        unsigned int block_height,
+                                        int strength,
+                                        int filter_weight,
+                                        unsigned int *accumulator,
+                                        uint16_t *count) {
+  uint16_t *frame1 = CONVERT_TO_SHORTPTR(frame1_8);
+  uint16_t *frame2 = CONVERT_TO_SHORTPTR(frame2_8);
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int src_byte = frame1[byte];
+      int pixel_value = *frame2++;
+
+      modifier   = src_byte - pixel_value;
+      // This is an integer approximation of:
+      // float coeff = (3.0 * modifer * modifier) / pow(2, strength);
+      // modifier =  (int)roundf(coeff > 16 ? 0 : 16-coeff);
+      modifier *= modifier;
+      modifier *= 3;
+      modifier += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16)
+        modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static int temporal_filter_find_matching_mb_c(VP9_COMP *cpi,
+                                              uint8_t *arf_frame_buf,
+                                              uint8_t *frame_ptr_buf,
+                                              int stride) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  int step_param;
+  int sadpb = x->sadperbit16;
+  int bestsme = INT_MAX;
+  int distortion;
+  unsigned int sse;
+  int cost_list[5];
+
+  MV best_ref_mv1 = {0, 0};
+  MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
+  MV *ref_mv = &x->e_mbd.mi[0]->bmi[0].as_mv[0].as_mv;
+
+  // Save input state
+  struct buf_2d src = x->plane[0].src;
+  struct buf_2d pre = xd->plane[0].pre[0];
+
+  best_ref_mv1_full.col = best_ref_mv1.col >> 3;
+  best_ref_mv1_full.row = best_ref_mv1.row >> 3;
+
+  // Setup frame pointers
+  x->plane[0].src.buf = arf_frame_buf;
+  x->plane[0].src.stride = stride;
+  xd->plane[0].pre[0].buf = frame_ptr_buf;
+  xd->plane[0].pre[0].stride = stride;
+
+  step_param = mv_sf->reduce_first_step_size;
+  step_param = MIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  // Ignore mv costing by sending NULL pointer instead of cost arrays
+  vp9_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1,
+                 cond_cost_list(cpi, cost_list),
+                 &cpi->fn_ptr[BLOCK_16X16], 0, &best_ref_mv1, ref_mv);
+
+  // Ignore mv costing by sending NULL pointer instead of cost array
+  bestsme = cpi->find_fractional_mv_step(x, ref_mv,
+                                         &best_ref_mv1,
+                                         cpi->common.allow_high_precision_mv,
+                                         x->errorperbit,
+                                         &cpi->fn_ptr[BLOCK_16X16],
+                                         0, mv_sf->subpel_iters_per_step,
+                                         cond_cost_list(cpi, cost_list),
+                                         NULL, NULL,
+                                         &distortion, &sse, NULL, 0, 0);
+
+  // Restore input state
+  x->plane[0].src = src;
+  xd->plane[0].pre[0] = pre;
+
+  return bestsme;
+}
+
+static void temporal_filter_iterate_c(VP9_COMP *cpi,
+                                      YV12_BUFFER_CONFIG **frames,
+                                      int frame_count,
+                                      int alt_ref_index,
+                                      int strength,
+                                      struct scale_factors *scale) {
+  int byte;
+  int frame;
+  int mb_col, mb_row;
+  unsigned int filter_weight;
+  int mb_cols = (frames[alt_ref_index]->y_crop_width + 15) >> 4;
+  int mb_rows = (frames[alt_ref_index]->y_crop_height + 15) >> 4;
+  int mb_y_offset = 0;
+  int mb_uv_offset = 0;
+  DECLARE_ALIGNED(16, unsigned int, accumulator[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint16_t, count[16 * 16 * 3]);
+  MACROBLOCKD *mbd = &cpi->td.mb.e_mbd;
+  YV12_BUFFER_CONFIG *f = frames[alt_ref_index];
+  uint8_t *dst1, *dst2;
+#if CONFIG_VP9_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t,  predictor16[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint8_t,  predictor8[16 * 16 * 3]);
+  uint8_t *predictor;
+#else
+  DECLARE_ALIGNED(16, uint8_t,  predictor[16 * 16 * 3]);
+#endif
+  const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y;
+  const int mb_uv_width  = 16 >> mbd->plane[1].subsampling_x;
+
+  // Save input state
+  uint8_t* input_buffer[MAX_MB_PLANE];
+  int i;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    predictor = CONVERT_TO_BYTEPTR(predictor16);
+  } else {
+    predictor = predictor8;
+  }
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; i++)
+    input_buffer[i] = mbd->plane[i].pre[0].buf;
+
+  for (mb_row = 0; mb_row < mb_rows; mb_row++) {
+    // Source frames are extended to 16 pixels. This is different than
+    //  L/A/G reference frames that have a border of 32 (VP9ENCBORDERINPIXELS)
+    // A 6/8 tap filter is used for motion search.  This requires 2 pixels
+    //  before and 3 pixels after.  So the largest Y mv on a border would
+    //  then be 16 - VP9_INTERP_EXTEND. The UV blocks are half the size of the
+    //  Y and therefore only extended by 8.  The largest mv that a UV block
+    //  can support is 8 - VP9_INTERP_EXTEND.  A UV mv is half of a Y mv.
+    //  (16 - VP9_INTERP_EXTEND) >> 1 which is greater than
+    //  8 - VP9_INTERP_EXTEND.
+    // To keep the mv in play for both Y and UV planes the max that it
+    //  can be on a border is therefore 16 - (2*VP9_INTERP_EXTEND+1).
+    cpi->td.mb.mv_row_min = -((mb_row * 16) + (17 - 2 * VP9_INTERP_EXTEND));
+    cpi->td.mb.mv_row_max = ((mb_rows - 1 - mb_row) * 16)
+                         + (17 - 2 * VP9_INTERP_EXTEND);
+
+    for (mb_col = 0; mb_col < mb_cols; mb_col++) {
+      int i, j, k;
+      int stride;
+
+      memset(accumulator, 0, 16 * 16 * 3 * sizeof(accumulator[0]));
+      memset(count, 0, 16 * 16 * 3 * sizeof(count[0]));
+
+      cpi->td.mb.mv_col_min = -((mb_col * 16) + (17 - 2 * VP9_INTERP_EXTEND));
+      cpi->td.mb.mv_col_max = ((mb_cols - 1 - mb_col) * 16)
+                           + (17 - 2 * VP9_INTERP_EXTEND);
+
+      for (frame = 0; frame < frame_count; frame++) {
+        const int thresh_low  = 10000;
+        const int thresh_high = 20000;
+
+        if (frames[frame] == NULL)
+          continue;
+
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.row = 0;
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.col = 0;
+
+        if (frame == alt_ref_index) {
+          filter_weight = 2;
+        } else {
+          // Find best match in this frame by MC
+          int err = temporal_filter_find_matching_mb_c(cpi,
+              frames[alt_ref_index]->y_buffer + mb_y_offset,
+              frames[frame]->y_buffer + mb_y_offset,
+              frames[frame]->y_stride);
+
+          // Assign higher weight to matching MB if it's error
+          // score is lower. If not applying MC default behavior
+          // is to weight all MBs equal.
+          filter_weight = err < thresh_low
+                          ? 2 : err < thresh_high ? 1 : 0;
+        }
+
+        if (filter_weight != 0) {
+          // Construct the predictors
+          temporal_filter_predictors_mb_c(mbd,
+              frames[frame]->y_buffer + mb_y_offset,
+              frames[frame]->u_buffer + mb_uv_offset,
+              frames[frame]->v_buffer + mb_uv_offset,
+              frames[frame]->y_stride,
+              mb_uv_width, mb_uv_height,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.row,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.col,
+              predictor, scale,
+              mb_col * 16, mb_row * 16);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+          if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            int adj_strength = strength + 2 * (mbd->bd - 8);
+            // Apply the filter (YUV)
+            vp9_highbd_temporal_filter_apply(f->y_buffer + mb_y_offset,
+                                             f->y_stride,
+                                             predictor, 16, 16, adj_strength,
+                                             filter_weight,
+                                             accumulator, count);
+            vp9_highbd_temporal_filter_apply(f->u_buffer + mb_uv_offset,
+                                             f->uv_stride, predictor + 256,
+                                             mb_uv_width, mb_uv_height,
+                                             adj_strength,
+                                             filter_weight, accumulator + 256,
+                                             count + 256);
+            vp9_highbd_temporal_filter_apply(f->v_buffer + mb_uv_offset,
+                                             f->uv_stride, predictor + 512,
+                                             mb_uv_width, mb_uv_height,
+                                             adj_strength, filter_weight,
+                                             accumulator + 512, count + 512);
+          } else {
+            // Apply the filter (YUV)
+            vp9_temporal_filter_apply(f->y_buffer + mb_y_offset, f->y_stride,
+                                      predictor, 16, 16,
+                                      strength, filter_weight,
+                                      accumulator, count);
+            vp9_temporal_filter_apply(f->u_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 256,
+                                      mb_uv_width, mb_uv_height, strength,
+                                      filter_weight, accumulator + 256,
+                                      count + 256);
+            vp9_temporal_filter_apply(f->v_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 512,
+                                      mb_uv_width, mb_uv_height, strength,
+                                      filter_weight, accumulator + 512,
+                                      count + 512);
+          }
+#else
+          // Apply the filter (YUV)
+          vp9_temporal_filter_apply(f->y_buffer + mb_y_offset, f->y_stride,
+                                    predictor, 16, 16,
+                                    strength, filter_weight,
+                                    accumulator, count);
+          vp9_temporal_filter_apply(f->u_buffer + mb_uv_offset, f->uv_stride,
+                                    predictor + 256,
+                                    mb_uv_width, mb_uv_height, strength,
+                                    filter_weight, accumulator + 256,
+                                    count + 256);
+          vp9_temporal_filter_apply(f->v_buffer + mb_uv_offset, f->uv_stride,
+                                    predictor + 512,
+                                    mb_uv_width, mb_uv_height, strength,
+                                    filter_weight, accumulator + 512,
+                                    count + 512);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+        }
+      }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        uint16_t *dst1_16;
+        uint16_t *dst2_16;
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+
+            dst1_16[byte] = (uint16_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        dst2_16 = CONVERT_TO_SHORTPTR(dst2);
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+            dst1_16[byte] = (uint16_t)pval;
+
+            // V
+            pval = accumulator[m] + (count[m] >> 1);
+            pval *= fixed_divide[count[m]];
+            pval >>= 19;
+            dst2_16[byte] = (uint16_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - mb_uv_width;
+        }
+      } else {
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+
+            dst1[byte] = (uint8_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            unsigned int pval = accumulator[k] + (count[k] >> 1);
+            pval *= fixed_divide[count[k]];
+            pval >>= 19;
+            dst1[byte] = (uint8_t)pval;
+
+            // V
+            pval = accumulator[m] + (count[m] >> 1);
+            pval *= fixed_divide[count[m]];
+            pval >>= 19;
+            dst2[byte] = (uint8_t)pval;
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - mb_uv_width;
+        }
+      }
+#else
+      // Normalize filter output to produce AltRef frame
+      dst1 = cpi->alt_ref_buffer.y_buffer;
+      stride = cpi->alt_ref_buffer.y_stride;
+      byte = mb_y_offset;
+      for (i = 0, k = 0; i < 16; i++) {
+        for (j = 0; j < 16; j++, k++) {
+          unsigned int pval = accumulator[k] + (count[k] >> 1);
+          pval *= fixed_divide[count[k]];
+          pval >>= 19;
+
+          dst1[byte] = (uint8_t)pval;
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - 16;
+      }
+
+      dst1 = cpi->alt_ref_buffer.u_buffer;
+      dst2 = cpi->alt_ref_buffer.v_buffer;
+      stride = cpi->alt_ref_buffer.uv_stride;
+      byte = mb_uv_offset;
+      for (i = 0, k = 256; i < mb_uv_height; i++) {
+        for (j = 0; j < mb_uv_width; j++, k++) {
+          int m = k + 256;
+
+          // U
+          unsigned int pval = accumulator[k] + (count[k] >> 1);
+          pval *= fixed_divide[count[k]];
+          pval >>= 19;
+          dst1[byte] = (uint8_t)pval;
+
+          // V
+          pval = accumulator[m] + (count[m] >> 1);
+          pval *= fixed_divide[count[m]];
+          pval >>= 19;
+          dst2[byte] = (uint8_t)pval;
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - mb_uv_width;
+      }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+      mb_y_offset += 16;
+      mb_uv_offset += mb_uv_width;
+    }
+    mb_y_offset += 16 * (f->y_stride - mb_cols);
+    mb_uv_offset += mb_uv_height * f->uv_stride - mb_uv_width * mb_cols;
+  }
+
+  // Restore input state
+  for (i = 0; i < MAX_MB_PLANE; i++)
+    mbd->plane[i].pre[0].buf = input_buffer[i];
+}
+
+// Apply buffer limits and context specific adjustments to arnr filter.
+static void adjust_arnr_filter(VP9_COMP *cpi,
+                               int distance, int group_boost,
+                               int *arnr_frames, int *arnr_strength) {
+  const VP9EncoderConfig *const oxcf = &cpi->oxcf;
+  const int frames_after_arf =
+      vp9_lookahead_depth(cpi->lookahead) - distance - 1;
+  int frames_fwd = (cpi->oxcf.arnr_max_frames - 1) >> 1;
+  int frames_bwd;
+  int q, frames, strength;
+
+  // Define the forward and backwards filter limits for this arnr group.
+  if (frames_fwd > frames_after_arf)
+    frames_fwd = frames_after_arf;
+  if (frames_fwd > distance)
+    frames_fwd = distance;
+
+  frames_bwd = frames_fwd;
+
+  // For even length filter there is one more frame backward
+  // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
+  if (frames_bwd < distance)
+    frames_bwd += (oxcf->arnr_max_frames + 1) & 0x1;
+
+  // Set the baseline active filter size.
+  frames = frames_bwd + 1 + frames_fwd;
+
+  // Adjust the strength based on active max q.
+  if (cpi->common.current_video_frame > 1)
+    q = ((int)vp9_convert_qindex_to_q(
+        cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth));
+  else
+    q = ((int)vp9_convert_qindex_to_q(
+        cpi->rc.avg_frame_qindex[KEY_FRAME], cpi->common.bit_depth));
+  if (q > 16) {
+    strength = oxcf->arnr_strength;
+  } else {
+    strength = oxcf->arnr_strength - ((16 - q) / 2);
+    if (strength < 0)
+      strength = 0;
+  }
+
+  // Adjust number of frames in filter and strength based on gf boost level.
+  if (frames > group_boost / 150) {
+    frames = group_boost / 150;
+    frames += !(frames & 1);
+  }
+
+  if (strength > group_boost / 300) {
+    strength = group_boost / 300;
+  }
+
+  // Adjustments for second level arf in multi arf case.
+  if (cpi->oxcf.pass == 2 && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    if (gf_group->rf_level[gf_group->index] != GF_ARF_STD) {
+      strength >>= 1;
+    }
+  }
+
+  *arnr_frames = frames;
+  *arnr_strength = strength;
+}
+
+void vp9_temporal_filter(VP9_COMP *cpi, int distance) {
+  VP9_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  int frame;
+  int frames_to_blur;
+  int start_frame;
+  int strength;
+  int frames_to_blur_backward;
+  int frames_to_blur_forward;
+  struct scale_factors sf;
+  YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS] = {NULL};
+
+  // Apply context specific adjustments to the arnr filter parameters.
+  adjust_arnr_filter(cpi, distance, rc->gfu_boost, &frames_to_blur, &strength);
+  frames_to_blur_backward = (frames_to_blur / 2);
+  frames_to_blur_forward = ((frames_to_blur - 1) / 2);
+  start_frame = distance + frames_to_blur_forward;
+
+  // Setup frame pointers, NULL indicates frame not included in filter.
+  for (frame = 0; frame < frames_to_blur; ++frame) {
+    const int which_buffer = start_frame - frame;
+    struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead,
+                                                     which_buffer);
+    frames[frames_to_blur - 1 - frame] = &buf->img;
+  }
+
+  if (frames_to_blur > 0) {
+    // Setup scaling factors. Scaling on each of the arnr frames is not
+    // supported.
+    if (cpi->use_svc) {
+      // In spatial svc the scaling factors might be less then 1/2.
+      // So we will use non-normative scaling.
+      int frame_used = 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+      vp9_setup_scale_factors_for_frame(
+          &sf,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height,
+          cm->use_highbitdepth);
+#else
+      vp9_setup_scale_factors_for_frame(
+          &sf,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height,
+          get_frame_new_buffer(cm)->y_crop_width,
+          get_frame_new_buffer(cm)->y_crop_height);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+      for (frame = 0; frame < frames_to_blur; ++frame) {
+        if (cm->mi_cols * MI_SIZE != frames[frame]->y_width ||
+            cm->mi_rows * MI_SIZE != frames[frame]->y_height) {
+          if (vp9_realloc_frame_buffer(&cpi->svc.scaled_frames[frame_used],
+                                       cm->width, cm->height,
+                                       cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                       cm->use_highbitdepth,
+#endif
+                                       VP9_ENC_BORDER_IN_PIXELS,
+                                       cm->byte_alignment,
+                                       NULL, NULL, NULL)) {
+            vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                               "Failed to reallocate alt_ref_buffer");
+          }
+          frames[frame] = vp9_scale_if_required(
+              cm, frames[frame], &cpi->svc.scaled_frames[frame_used]);
+          ++frame_used;
+        }
+      }
+      cm->mi = cm->mip + cm->mi_stride + 1;
+      xd->mi = cm->mi_grid_visible;
+      xd->mi[0] = cm->mi;
+    } else {
+      // ARF is produced at the native frame size and resized when coded.
+#if CONFIG_VP9_HIGHBITDEPTH
+      vp9_setup_scale_factors_for_frame(&sf,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height,
+                                        cm->use_highbitdepth);
+#else
+      vp9_setup_scale_factors_for_frame(&sf,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height,
+                                        frames[0]->y_crop_width,
+                                        frames[0]->y_crop_height);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    }
+  }
+
+  temporal_filter_iterate_c(cpi, frames, frames_to_blur,
+                            frames_to_blur_backward, strength, &sf);
+}

media/libvpx/vp9/encoder/vp9_temporal_filter.h

@@ -0,0 +1,25 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_TEMPORAL_FILTER_H_
+#define VP9_ENCODER_VP9_TEMPORAL_FILTER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_temporal_filter_init(void);
+void vp9_temporal_filter(VP9_COMP *cpi, int distance);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_TEMPORAL_FILTER_H_

media/libvpx/vp9/encoder/vp9_tokenize.c

@@ -0,0 +1,634 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_scan.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/encoder/vp9_cost.h"
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_tokenize.h"
+
+static const TOKENVALUE dct_cat_lt_10_value_tokens[] = {
+  {9, 63}, {9, 61}, {9, 59}, {9, 57}, {9, 55}, {9, 53}, {9, 51}, {9, 49},
+  {9, 47}, {9, 45}, {9, 43}, {9, 41}, {9, 39}, {9, 37}, {9, 35}, {9, 33},
+  {9, 31}, {9, 29}, {9, 27}, {9, 25}, {9, 23}, {9, 21}, {9, 19}, {9, 17},
+  {9, 15}, {9, 13}, {9, 11}, {9, 9}, {9, 7}, {9, 5}, {9, 3}, {9, 1},
+  {8, 31}, {8, 29}, {8, 27}, {8, 25}, {8, 23}, {8, 21},
+  {8, 19}, {8, 17}, {8, 15}, {8, 13}, {8, 11}, {8, 9},
+  {8, 7}, {8, 5}, {8, 3}, {8, 1},
+  {7, 15}, {7, 13}, {7, 11}, {7, 9}, {7, 7}, {7, 5}, {7, 3}, {7, 1},
+  {6, 7}, {6, 5}, {6, 3}, {6, 1}, {5, 3}, {5, 1},
+  {4, 1}, {3, 1}, {2, 1}, {1, 1}, {0, 0},
+  {1, 0},  {2, 0}, {3, 0}, {4, 0},
+  {5, 0}, {5, 2}, {6, 0}, {6, 2}, {6, 4}, {6, 6},
+  {7, 0}, {7, 2}, {7, 4}, {7, 6}, {7, 8}, {7, 10}, {7, 12}, {7, 14},
+  {8, 0}, {8, 2}, {8, 4}, {8, 6}, {8, 8}, {8, 10}, {8, 12},
+  {8, 14}, {8, 16}, {8, 18}, {8, 20}, {8, 22}, {8, 24},
+  {8, 26}, {8, 28}, {8, 30}, {9, 0}, {9, 2},
+  {9, 4}, {9, 6}, {9, 8}, {9, 10}, {9, 12}, {9, 14}, {9, 16},
+  {9, 18}, {9, 20}, {9, 22}, {9, 24}, {9, 26}, {9, 28},
+  {9, 30}, {9, 32}, {9, 34}, {9, 36}, {9, 38}, {9, 40},
+  {9, 42}, {9, 44}, {9, 46}, {9, 48}, {9, 50}, {9, 52},
+  {9, 54}, {9, 56}, {9, 58}, {9, 60}, {9, 62}
+};
+const TOKENVALUE *vp9_dct_cat_lt_10_value_tokens = dct_cat_lt_10_value_tokens +
+    (sizeof(dct_cat_lt_10_value_tokens) / sizeof(*dct_cat_lt_10_value_tokens))
+    / 2;
+
+// Array indices are identical to previously-existing CONTEXT_NODE indices
+const vp9_tree_index vp9_coef_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
+  -EOB_TOKEN, 2,                       // 0  = EOB
+  -ZERO_TOKEN, 4,                      // 1  = ZERO
+  -ONE_TOKEN, 6,                       // 2  = ONE
+  8, 12,                               // 3  = LOW_VAL
+  -TWO_TOKEN, 10,                      // 4  = TWO
+  -THREE_TOKEN, -FOUR_TOKEN,           // 5  = THREE
+  14, 16,                              // 6  = HIGH_LOW
+  -CATEGORY1_TOKEN, -CATEGORY2_TOKEN,  // 7  = CAT_ONE
+  18, 20,                              // 8  = CAT_THREEFOUR
+  -CATEGORY3_TOKEN, -CATEGORY4_TOKEN,  // 9  = CAT_THREE
+  -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 10 = CAT_FIVE
+};
+
+static const vp9_tree_index cat1[2] = {0, 0};
+static const vp9_tree_index cat2[4] = {2, 2, 0, 0};
+static const vp9_tree_index cat3[6] = {2, 2, 4, 4, 0, 0};
+static const vp9_tree_index cat4[8] = {2, 2, 4, 4, 6, 6, 0, 0};
+static const vp9_tree_index cat5[10] = {2, 2, 4, 4, 6, 6, 8, 8, 0, 0};
+static const vp9_tree_index cat6[28] = {2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12,
+    14, 14, 16, 16, 18, 18, 20, 20, 22, 22, 24, 24, 26, 26, 0, 0};
+
+static const int16_t zero_cost[] = {0};
+static const int16_t one_cost[] = {255, 257};
+static const int16_t two_cost[] = {255, 257};
+static const int16_t three_cost[] = {255, 257};
+static const int16_t four_cost[] = {255, 257};
+static const int16_t cat1_cost[] = {429, 431, 616, 618};
+static const int16_t cat2_cost[] = {624, 626, 727, 729, 848, 850, 951, 953};
+static const int16_t cat3_cost[] = {
+  820, 822, 893, 895, 940, 942, 1013, 1015, 1096, 1098, 1169, 1171, 1216, 1218,
+  1289, 1291
+};
+static const int16_t cat4_cost[] = {
+  1032, 1034, 1075, 1077, 1105, 1107, 1148, 1150, 1194, 1196, 1237, 1239,
+  1267, 1269, 1310, 1312, 1328, 1330, 1371, 1373, 1401, 1403, 1444, 1446,
+  1490, 1492, 1533, 1535, 1563, 1565, 1606, 1608
+};
+static const int16_t cat5_cost[] = {
+  1269, 1271, 1283, 1285, 1306, 1308, 1320,
+  1322, 1347, 1349, 1361, 1363, 1384, 1386, 1398, 1400, 1443, 1445, 1457,
+  1459, 1480, 1482, 1494, 1496, 1521, 1523, 1535, 1537, 1558, 1560, 1572,
+  1574, 1592, 1594, 1606, 1608, 1629, 1631, 1643, 1645, 1670, 1672, 1684,
+  1686, 1707, 1709, 1721, 1723, 1766, 1768, 1780, 1782, 1803, 1805, 1817,
+  1819, 1844, 1846, 1858, 1860, 1881, 1883, 1895, 1897
+};
+const int16_t vp9_cat6_low_cost[256] = {
+  1638, 1640, 1646, 1648, 1652, 1654, 1660, 1662,
+  1670, 1672, 1678, 1680, 1684, 1686, 1692, 1694, 1711, 1713, 1719, 1721,
+  1725, 1727, 1733, 1735, 1743, 1745, 1751, 1753, 1757, 1759, 1765, 1767,
+  1787, 1789, 1795, 1797, 1801, 1803, 1809, 1811, 1819, 1821, 1827, 1829,
+  1833, 1835, 1841, 1843, 1860, 1862, 1868, 1870, 1874, 1876, 1882, 1884,
+  1892, 1894, 1900, 1902, 1906, 1908, 1914, 1916, 1940, 1942, 1948, 1950,
+  1954, 1956, 1962, 1964, 1972, 1974, 1980, 1982, 1986, 1988, 1994, 1996,
+  2013, 2015, 2021, 2023, 2027, 2029, 2035, 2037, 2045, 2047, 2053, 2055,
+  2059, 2061, 2067, 2069, 2089, 2091, 2097, 2099, 2103, 2105, 2111, 2113,
+  2121, 2123, 2129, 2131, 2135, 2137, 2143, 2145, 2162, 2164, 2170, 2172,
+  2176, 2178, 2184, 2186, 2194, 2196, 2202, 2204, 2208, 2210, 2216, 2218,
+  2082, 2084, 2090, 2092, 2096, 2098, 2104, 2106, 2114, 2116, 2122, 2124,
+  2128, 2130, 2136, 2138, 2155, 2157, 2163, 2165, 2169, 2171, 2177, 2179,
+  2187, 2189, 2195, 2197, 2201, 2203, 2209, 2211, 2231, 2233, 2239, 2241,
+  2245, 2247, 2253, 2255, 2263, 2265, 2271, 2273, 2277, 2279, 2285, 2287,
+  2304, 2306, 2312, 2314, 2318, 2320, 2326, 2328, 2336, 2338, 2344, 2346,
+  2350, 2352, 2358, 2360, 2384, 2386, 2392, 2394, 2398, 2400, 2406, 2408,
+  2416, 2418, 2424, 2426, 2430, 2432, 2438, 2440, 2457, 2459, 2465, 2467,
+  2471, 2473, 2479, 2481, 2489, 2491, 2497, 2499, 2503, 2505, 2511, 2513,
+  2533, 2535, 2541, 2543, 2547, 2549, 2555, 2557, 2565, 2567, 2573, 2575,
+  2579, 2581, 2587, 2589, 2606, 2608, 2614, 2616, 2620, 2622, 2628, 2630,
+  2638, 2640, 2646, 2648, 2652, 2654, 2660, 2662
+};
+const int16_t vp9_cat6_high_cost[128] = {
+  72, 892, 1183, 2003, 1448, 2268, 2559, 3379,
+  1709, 2529, 2820, 3640, 3085, 3905, 4196, 5016, 2118, 2938, 3229, 4049,
+  3494, 4314, 4605, 5425, 3755, 4575, 4866, 5686, 5131, 5951, 6242, 7062,
+  2118, 2938, 3229, 4049, 3494, 4314, 4605, 5425, 3755, 4575, 4866, 5686,
+  5131, 5951, 6242, 7062, 4164, 4984, 5275, 6095, 5540, 6360, 6651, 7471,
+  5801, 6621, 6912, 7732, 7177, 7997, 8288, 9108, 2118, 2938, 3229, 4049,
+  3494, 4314, 4605, 5425, 3755, 4575, 4866, 5686, 5131, 5951, 6242, 7062,
+  4164, 4984, 5275, 6095, 5540, 6360, 6651, 7471, 5801, 6621, 6912, 7732,
+  7177, 7997, 8288, 9108, 4164, 4984, 5275, 6095, 5540, 6360, 6651, 7471,
+  5801, 6621, 6912, 7732, 7177, 7997, 8288, 9108, 6210, 7030, 7321, 8141,
+  7586, 8406, 8697, 9517, 7847, 8667, 8958, 9778, 9223, 10043, 10334, 11154
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+const int16_t vp9_cat6_high10_high_cost[512] = {
+  74, 894, 1185, 2005, 1450, 2270, 2561,
+  3381, 1711, 2531, 2822, 3642, 3087, 3907, 4198, 5018, 2120, 2940, 3231,
+  4051, 3496, 4316, 4607, 5427, 3757, 4577, 4868, 5688, 5133, 5953, 6244,
+  7064, 2120, 2940, 3231, 4051, 3496, 4316, 4607, 5427, 3757, 4577, 4868,
+  5688, 5133, 5953, 6244, 7064, 4166, 4986, 5277, 6097, 5542, 6362, 6653,
+  7473, 5803, 6623, 6914, 7734, 7179, 7999, 8290, 9110, 2120, 2940, 3231,
+  4051, 3496, 4316, 4607, 5427, 3757, 4577, 4868, 5688, 5133, 5953, 6244,
+  7064, 4166, 4986, 5277, 6097, 5542, 6362, 6653, 7473, 5803, 6623, 6914,
+  7734, 7179, 7999, 8290, 9110, 4166, 4986, 5277, 6097, 5542, 6362, 6653,
+  7473, 5803, 6623, 6914, 7734, 7179, 7999, 8290, 9110, 6212, 7032, 7323,
+  8143, 7588, 8408, 8699, 9519, 7849, 8669, 8960, 9780, 9225, 10045, 10336,
+  11156, 2120, 2940, 3231, 4051, 3496, 4316, 4607, 5427, 3757, 4577, 4868,
+  5688, 5133, 5953, 6244, 7064, 4166, 4986, 5277, 6097, 5542, 6362, 6653,
+  7473, 5803, 6623, 6914, 7734, 7179, 7999, 8290, 9110, 4166, 4986, 5277,
+  6097, 5542, 6362, 6653, 7473, 5803, 6623, 6914, 7734, 7179, 7999, 8290,
+  9110, 6212, 7032, 7323, 8143, 7588, 8408, 8699, 9519, 7849, 8669, 8960,
+  9780, 9225, 10045, 10336, 11156, 4166, 4986, 5277, 6097, 5542, 6362, 6653,
+  7473, 5803, 6623, 6914, 7734, 7179, 7999, 8290, 9110, 6212, 7032, 7323,
+  8143, 7588, 8408, 8699, 9519, 7849, 8669, 8960, 9780, 9225, 10045, 10336,
+  11156, 6212, 7032, 7323, 8143, 7588, 8408, 8699, 9519, 7849, 8669, 8960,
+  9780, 9225, 10045, 10336, 11156, 8258, 9078, 9369, 10189, 9634, 10454,
+  10745, 11565, 9895, 10715, 11006, 11826, 11271, 12091, 12382, 13202, 2120,
+  2940, 3231, 4051, 3496, 4316, 4607, 5427, 3757, 4577, 4868, 5688, 5133,
+  5953, 6244, 7064, 4166, 4986, 5277, 6097, 5542, 6362, 6653, 7473, 5803,
+  6623, 6914, 7734, 7179, 7999, 8290, 9110, 4166, 4986, 5277, 6097, 5542,
+  6362, 6653, 7473, 5803, 6623, 6914, 7734, 7179, 7999, 8290, 9110, 6212,
+  7032, 7323, 8143, 7588, 8408, 8699, 9519, 7849, 8669, 8960, 9780, 9225,
+  10045, 10336, 11156, 4166, 4986, 5277, 6097, 5542, 6362, 6653, 7473, 5803,
+  6623, 6914, 7734, 7179, 7999, 8290, 9110, 6212, 7032, 7323, 8143, 7588,
+  8408, 8699, 9519, 7849, 8669, 8960, 9780, 9225, 10045, 10336, 11156, 6212,
+  7032, 7323, 8143, 7588, 8408, 8699, 9519, 7849, 8669, 8960, 9780, 9225,
+  10045, 10336, 11156, 8258, 9078, 9369, 10189, 9634, 10454, 10745, 11565,
+  9895, 10715, 11006, 11826, 11271, 12091, 12382, 13202, 4166, 4986, 5277,
+  6097, 5542, 6362, 6653, 7473, 5803, 6623, 6914, 7734, 7179, 7999, 8290,
+  9110, 6212, 7032, 7323, 8143, 7588, 8408, 8699, 9519, 7849, 8669, 8960,
+  9780, 9225, 10045, 10336, 11156, 6212, 7032, 7323, 8143, 7588, 8408, 8699,
+  9519, 7849, 8669, 8960, 9780, 9225, 10045, 10336, 11156, 8258, 9078, 9369,
+  10189, 9634, 10454, 10745, 11565, 9895, 10715, 11006, 11826, 11271, 12091,
+  12382, 13202, 6212, 7032, 7323, 8143, 7588, 8408, 8699, 9519, 7849, 8669,
+  8960, 9780, 9225, 10045, 10336, 11156, 8258, 9078, 9369, 10189, 9634, 10454,
+  10745, 11565, 9895, 10715, 11006, 11826, 11271, 12091, 12382, 13202, 8258,
+  9078, 9369, 10189, 9634, 10454, 10745, 11565, 9895, 10715, 11006, 11826,
+  11271, 12091, 12382, 13202, 10304, 11124, 11415, 12235, 11680, 12500, 12791,
+  13611, 11941, 12761, 13052, 13872, 13317, 14137, 14428, 15248,
+};
+const int16_t vp9_cat6_high12_high_cost[2048] = {
+  76, 896, 1187, 2007, 1452, 2272, 2563,
+  3383, 1713, 2533, 2824, 3644, 3089, 3909, 4200, 5020, 2122, 2942, 3233,
+  4053, 3498, 4318, 4609, 5429, 3759, 4579, 4870, 5690, 5135, 5955, 6246,
+  7066, 2122, 2942, 3233, 4053, 3498, 4318, 4609, 5429, 3759, 4579, 4870,
+  5690, 5135, 5955, 6246, 7066, 4168, 4988, 5279, 6099, 5544, 6364, 6655,
+  7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112, 2122, 2942, 3233,
+  4053, 3498, 4318, 4609, 5429, 3759, 4579, 4870, 5690, 5135, 5955, 6246,
+  7066, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916,
+  7736, 7181, 8001, 8292, 9112, 4168, 4988, 5279, 6099, 5544, 6364, 6655,
+  7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325,
+  8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338,
+  11158, 2122, 2942, 3233, 4053, 3498, 4318, 4609, 5429, 3759, 4579, 4870,
+  5690, 5135, 5955, 6246, 7066, 4168, 4988, 5279, 6099, 5544, 6364, 6655,
+  7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112, 4168, 4988, 5279,
+  6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292,
+  9112, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962,
+  9782, 9227, 10047, 10338, 11158, 4168, 4988, 5279, 6099, 5544, 6364, 6655,
+  7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325,
+  8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338,
+  11158, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962,
+  9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456,
+  10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 2122,
+  2942, 3233, 4053, 3498, 4318, 4609, 5429, 3759, 4579, 4870, 5690, 5135,
+  5955, 6246, 7066, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805,
+  6625, 6916, 7736, 7181, 8001, 8292, 9112, 4168, 4988, 5279, 6099, 5544,
+  6364, 6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112, 6214,
+  7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227,
+  10047, 10338, 11158, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805,
+  6625, 6916, 7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325, 8145, 7590,
+  8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 6214,
+  7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227,
+  10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567,
+  9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 4168, 4988, 5279,
+  6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292,
+  9112, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962,
+  9782, 9227, 10047, 10338, 11158, 6214, 7034, 7325, 8145, 7590, 8410, 8701,
+  9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371,
+  10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093,
+  12384, 13204, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671,
+  8962, 9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456,
+  10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 8260,
+  9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828,
+  11273, 12093, 12384, 13204, 10306, 11126, 11417, 12237, 11682, 12502, 12793,
+  13613, 11943, 12763, 13054, 13874, 13319, 14139, 14430, 15250, 2122, 2942,
+  3233, 4053, 3498, 4318, 4609, 5429, 3759, 4579, 4870, 5690, 5135, 5955,
+  6246, 7066, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805, 6625,
+  6916, 7736, 7181, 8001, 8292, 9112, 4168, 4988, 5279, 6099, 5544, 6364,
+  6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112, 6214, 7034,
+  7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047,
+  10338, 11158, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805, 6625,
+  6916, 7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325, 8145, 7590, 8410,
+  8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 6214, 7034,
+  7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047,
+  10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897,
+  10717, 11008, 11828, 11273, 12093, 12384, 13204, 4168, 4988, 5279, 6099,
+  5544, 6364, 6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112,
+  6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782,
+  9227, 10047, 10338, 11158, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521,
+  7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371, 10191,
+  9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384,
+  13204, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962,
+  9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456,
+  10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 8260,
+  9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828,
+  11273, 12093, 12384, 13204, 10306, 11126, 11417, 12237, 11682, 12502, 12793,
+  13613, 11943, 12763, 13054, 13874, 13319, 14139, 14430, 15250, 4168, 4988,
+  5279, 6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001,
+  8292, 9112, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671,
+  8962, 9782, 9227, 10047, 10338, 11158, 6214, 7034, 7325, 8145, 7590, 8410,
+  8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260, 9080,
+  9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273,
+  12093, 12384, 13204, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851,
+  8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636,
+  10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204,
+  8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008,
+  11828, 11273, 12093, 12384, 13204, 10306, 11126, 11417, 12237, 11682, 12502,
+  12793, 13613, 11943, 12763, 13054, 13874, 13319, 14139, 14430, 15250, 6214,
+  7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227,
+  10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567,
+  9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 8260, 9080, 9371,
+  10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093,
+  12384, 13204, 10306, 11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943,
+  12763, 13054, 13874, 13319, 14139, 14430, 15250, 8260, 9080, 9371, 10191,
+  9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384,
+  13204, 10306, 11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763,
+  13054, 13874, 13319, 14139, 14430, 15250, 10306, 11126, 11417, 12237, 11682,
+  12502, 12793, 13613, 11943, 12763, 13054, 13874, 13319, 14139, 14430, 15250,
+  12352, 13172, 13463, 14283, 13728, 14548, 14839, 15659, 13989, 14809, 15100,
+  15920, 15365, 16185, 16476, 17296, 2122, 2942, 3233, 4053, 3498, 4318, 4609,
+  5429, 3759, 4579, 4870, 5690, 5135, 5955, 6246, 7066, 4168, 4988, 5279,
+  6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292,
+  9112, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916,
+  7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325, 8145, 7590, 8410, 8701,
+  9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 4168, 4988, 5279,
+  6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916, 7736, 7181, 8001, 8292,
+  9112, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962,
+  9782, 9227, 10047, 10338, 11158, 6214, 7034, 7325, 8145, 7590, 8410, 8701,
+  9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371,
+  10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093,
+  12384, 13204, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805, 6625,
+  6916, 7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325, 8145, 7590, 8410,
+  8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 6214, 7034,
+  7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047,
+  10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897,
+  10717, 11008, 11828, 11273, 12093, 12384, 13204, 6214, 7034, 7325, 8145,
+  7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158,
+  8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008,
+  11828, 11273, 12093, 12384, 13204, 8260, 9080, 9371, 10191, 9636, 10456,
+  10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 10306,
+  11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054, 13874,
+  13319, 14139, 14430, 15250, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475,
+  5805, 6625, 6916, 7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325, 8145,
+  7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158,
+  6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782,
+  9227, 10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747,
+  11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 6214, 7034,
+  7325, 8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047,
+  10338, 11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897,
+  10717, 11008, 11828, 11273, 12093, 12384, 13204, 8260, 9080, 9371, 10191,
+  9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384,
+  13204, 10306, 11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763,
+  13054, 13874, 13319, 14139, 14430, 15250, 6214, 7034, 7325, 8145, 7590,
+  8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260,
+  9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828,
+  11273, 12093, 12384, 13204, 8260, 9080, 9371, 10191, 9636, 10456, 10747,
+  11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 10306, 11126,
+  11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054, 13874, 13319,
+  14139, 14430, 15250, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567,
+  9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 10306, 11126, 11417,
+  12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054, 13874, 13319, 14139,
+  14430, 15250, 10306, 11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943,
+  12763, 13054, 13874, 13319, 14139, 14430, 15250, 12352, 13172, 13463, 14283,
+  13728, 14548, 14839, 15659, 13989, 14809, 15100, 15920, 15365, 16185, 16476,
+  17296, 4168, 4988, 5279, 6099, 5544, 6364, 6655, 7475, 5805, 6625, 6916,
+  7736, 7181, 8001, 8292, 9112, 6214, 7034, 7325, 8145, 7590, 8410, 8701,
+  9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 6214, 7034, 7325,
+  8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338,
+  11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717,
+  11008, 11828, 11273, 12093, 12384, 13204, 6214, 7034, 7325, 8145, 7590,
+  8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260,
+  9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828,
+  11273, 12093, 12384, 13204, 8260, 9080, 9371, 10191, 9636, 10456, 10747,
+  11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 10306, 11126,
+  11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054, 13874, 13319,
+  14139, 14430, 15250, 6214, 7034, 7325, 8145, 7590, 8410, 8701, 9521, 7851,
+  8671, 8962, 9782, 9227, 10047, 10338, 11158, 8260, 9080, 9371, 10191, 9636,
+  10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204,
+  8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008,
+  11828, 11273, 12093, 12384, 13204, 10306, 11126, 11417, 12237, 11682, 12502,
+  12793, 13613, 11943, 12763, 13054, 13874, 13319, 14139, 14430, 15250, 8260,
+  9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717, 11008, 11828,
+  11273, 12093, 12384, 13204, 10306, 11126, 11417, 12237, 11682, 12502, 12793,
+  13613, 11943, 12763, 13054, 13874, 13319, 14139, 14430, 15250, 10306, 11126,
+  11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054, 13874, 13319,
+  14139, 14430, 15250, 12352, 13172, 13463, 14283, 13728, 14548, 14839, 15659,
+  13989, 14809, 15100, 15920, 15365, 16185, 16476, 17296, 6214, 7034, 7325,
+  8145, 7590, 8410, 8701, 9521, 7851, 8671, 8962, 9782, 9227, 10047, 10338,
+  11158, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567, 9897, 10717,
+  11008, 11828, 11273, 12093, 12384, 13204, 8260, 9080, 9371, 10191, 9636,
+  10456, 10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204,
+  10306, 11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054,
+  13874, 13319, 14139, 14430, 15250, 8260, 9080, 9371, 10191, 9636, 10456,
+  10747, 11567, 9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 10306,
+  11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054, 13874,
+  13319, 14139, 14430, 15250, 10306, 11126, 11417, 12237, 11682, 12502, 12793,
+  13613, 11943, 12763, 13054, 13874, 13319, 14139, 14430, 15250, 12352, 13172,
+  13463, 14283, 13728, 14548, 14839, 15659, 13989, 14809, 15100, 15920, 15365,
+  16185, 16476, 17296, 8260, 9080, 9371, 10191, 9636, 10456, 10747, 11567,
+  9897, 10717, 11008, 11828, 11273, 12093, 12384, 13204, 10306, 11126, 11417,
+  12237, 11682, 12502, 12793, 13613, 11943, 12763, 13054, 13874, 13319, 14139,
+  14430, 15250, 10306, 11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943,
+  12763, 13054, 13874, 13319, 14139, 14430, 15250, 12352, 13172, 13463, 14283,
+  13728, 14548, 14839, 15659, 13989, 14809, 15100, 15920, 15365, 16185, 16476,
+  17296, 10306, 11126, 11417, 12237, 11682, 12502, 12793, 13613, 11943, 12763,
+  13054, 13874, 13319, 14139, 14430, 15250, 12352, 13172, 13463, 14283, 13728,
+  14548, 14839, 15659, 13989, 14809, 15100, 15920, 15365, 16185, 16476, 17296,
+  12352, 13172, 13463, 14283, 13728, 14548, 14839, 15659, 13989, 14809, 15100,
+  15920, 15365, 16185, 16476, 17296, 14398, 15218, 15509, 16329, 15774, 16594,
+  16885, 17705, 16035, 16855, 17146, 17966, 17411, 18231, 18522, 19342
+};
+#endif
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const vp9_tree_index cat1_high10[2] = {0, 0};
+static const vp9_tree_index cat2_high10[4] = {2, 2, 0, 0};
+static const vp9_tree_index cat3_high10[6] = {2, 2, 4, 4, 0, 0};
+static const vp9_tree_index cat4_high10[8] = {2, 2, 4, 4, 6, 6, 0, 0};
+static const vp9_tree_index cat5_high10[10] = {2, 2, 4, 4, 6, 6, 8, 8, 0, 0};
+static const vp9_tree_index cat6_high10[32] = {2, 2, 4, 4, 6, 6, 8, 8, 10, 10,
+  12, 12, 14, 14, 16, 16, 18, 18, 20, 20, 22, 22, 24, 24, 26, 26, 28, 28,
+  30, 30, 0, 0};
+static const vp9_tree_index cat1_high12[2] = {0, 0};
+static const vp9_tree_index cat2_high12[4] = {2, 2, 0, 0};
+static const vp9_tree_index cat3_high12[6] = {2, 2, 4, 4, 0, 0};
+static const vp9_tree_index cat4_high12[8] = {2, 2, 4, 4, 6, 6, 0, 0};
+static const vp9_tree_index cat5_high12[10] = {2, 2, 4, 4, 6, 6, 8, 8, 0, 0};
+static const vp9_tree_index cat6_high12[36] = {2, 2, 4, 4, 6, 6, 8, 8, 10, 10,
+  12, 12, 14, 14, 16, 16, 18, 18, 20, 20, 22, 22, 24, 24, 26, 26, 28, 28,
+  30, 30, 32, 32, 34, 34, 0, 0};
+#endif
+
+const vp9_extra_bit vp9_extra_bits[ENTROPY_TOKENS] = {
+  {0, 0, 0, 0, zero_cost},                             // ZERO_TOKEN
+  {0, 0, 0, 1, one_cost},                              // ONE_TOKEN
+  {0, 0, 0, 2, two_cost},                              // TWO_TOKEN
+  {0, 0, 0, 3, three_cost},                            // THREE_TOKEN
+  {0, 0, 0, 4, four_cost},                             // FOUR_TOKEN
+  {cat1, vp9_cat1_prob, 1,  CAT1_MIN_VAL, cat1_cost},  // CATEGORY1_TOKEN
+  {cat2, vp9_cat2_prob, 2,  CAT2_MIN_VAL, cat2_cost},  // CATEGORY2_TOKEN
+  {cat3, vp9_cat3_prob, 3,  CAT3_MIN_VAL, cat3_cost},  // CATEGORY3_TOKEN
+  {cat4, vp9_cat4_prob, 4,  CAT4_MIN_VAL, cat4_cost},  // CATEGORY4_TOKEN
+  {cat5, vp9_cat5_prob, 5,  CAT5_MIN_VAL, cat5_cost},  // CATEGORY5_TOKEN
+  {cat6, vp9_cat6_prob, 14, CAT6_MIN_VAL, 0},          // CATEGORY6_TOKEN
+  {0, 0, 0, 0, zero_cost}                              // EOB_TOKEN
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+const vp9_extra_bit vp9_extra_bits_high10[ENTROPY_TOKENS] = {
+  {0, 0, 0, 0, zero_cost},                                           // ZERO
+  {0, 0, 0, 1, one_cost},                                            // ONE
+  {0, 0, 0, 2, two_cost},                                            // TWO
+  {0, 0, 0, 3, three_cost},                                          // THREE
+  {0, 0, 0, 4, four_cost},                                           // FOUR
+  {cat1_high10, vp9_cat1_prob_high10, 1,  CAT1_MIN_VAL, cat1_cost},  // CAT1
+  {cat2_high10, vp9_cat2_prob_high10, 2,  CAT2_MIN_VAL, cat2_cost},  // CAT2
+  {cat3_high10, vp9_cat3_prob_high10, 3,  CAT3_MIN_VAL, cat3_cost},  // CAT3
+  {cat4_high10, vp9_cat4_prob_high10, 4,  CAT4_MIN_VAL, cat4_cost},  // CAT4
+  {cat5_high10, vp9_cat5_prob_high10, 5,  CAT5_MIN_VAL, cat5_cost},  // CAT5
+  {cat6_high10, vp9_cat6_prob_high10, 16, CAT6_MIN_VAL, 0},          // CAT6
+  {0, 0, 0, 0, zero_cost}                                            // EOB
+};
+const vp9_extra_bit vp9_extra_bits_high12[ENTROPY_TOKENS] = {
+  {0, 0, 0, 0, zero_cost},                                           // ZERO
+  {0, 0, 0, 1, one_cost},                                            // ONE
+  {0, 0, 0, 2, two_cost},                                            // TWO
+  {0, 0, 0, 3, three_cost},                                          // THREE
+  {0, 0, 0, 4, four_cost},                                           // FOUR
+  {cat1_high12, vp9_cat1_prob_high12, 1,  CAT1_MIN_VAL, cat1_cost},  // CAT1
+  {cat2_high12, vp9_cat2_prob_high12, 2,  CAT2_MIN_VAL, cat2_cost},  // CAT2
+  {cat3_high12, vp9_cat3_prob_high12, 3,  CAT3_MIN_VAL, cat3_cost},  // CAT3
+  {cat4_high12, vp9_cat4_prob_high12, 4,  CAT4_MIN_VAL, cat4_cost},  // CAT4
+  {cat5_high12, vp9_cat5_prob_high12, 5,  CAT5_MIN_VAL, cat5_cost},  // CAT5
+  {cat6_high12, vp9_cat6_prob_high12, 18, CAT6_MIN_VAL, 0},          // CAT6
+  {0, 0, 0, 0, zero_cost}                                            // EOB
+};
+#endif
+
+const struct vp9_token vp9_coef_encodings[ENTROPY_TOKENS] = {
+  {2, 2}, {6, 3}, {28, 5}, {58, 6}, {59, 6}, {60, 6}, {61, 6}, {124, 7},
+  {125, 7}, {126, 7}, {127, 7}, {0, 1}
+};
+
+
+struct tokenize_b_args {
+  VP9_COMP *cpi;
+  ThreadData *td;
+  TOKENEXTRA **tp;
+};
+
+static void set_entropy_context_b(int plane, int block, BLOCK_SIZE plane_bsize,
+                                  TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args* const args = arg;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  int aoff, loff;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
+  vp9_set_contexts(xd, pd, plane_bsize, tx_size, p->eobs[block] > 0,
+                   aoff, loff);
+}
+
+static INLINE void add_token(TOKENEXTRA **t, const vp9_prob *context_tree,
+                             int32_t extra, uint8_t token,
+                             uint8_t skip_eob_node,
+                             unsigned int *counts) {
+  (*t)->token = token;
+  (*t)->extra = extra;
+  (*t)->context_tree = context_tree;
+  (*t)->skip_eob_node = skip_eob_node;
+  (*t)++;
+  ++counts[token];
+}
+
+static INLINE void add_token_no_extra(TOKENEXTRA **t,
+                                      const vp9_prob *context_tree,
+                                      uint8_t token,
+                                      uint8_t skip_eob_node,
+                                      unsigned int *counts) {
+  (*t)->token = token;
+  (*t)->context_tree = context_tree;
+  (*t)->skip_eob_node = skip_eob_node;
+  (*t)++;
+  ++counts[token];
+}
+
+static INLINE int get_tx_eob(const struct segmentation *seg, int segment_id,
+                             TX_SIZE tx_size) {
+  const int eob_max = 16 << (tx_size << 1);
+  return vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP) ? 0 : eob_max;
+}
+
+static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
+                       TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args* const args = arg;
+  VP9_COMP *cpi = args->cpi;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  TOKENEXTRA **tp = args->tp;
+  uint8_t token_cache[32 * 32];
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int pt; /* near block/prev token context index */
+  int c;
+  TOKENEXTRA *t = *tp;        /* store tokens starting here */
+  int eob = p->eobs[block];
+  const PLANE_TYPE type = pd->plane_type;
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  const int segment_id = mbmi->segment_id;
+  const int16_t *scan, *nb;
+  const scan_order *so;
+  const int ref = is_inter_block(mbmi);
+  unsigned int (*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      td->rd_counts.coef_counts[tx_size][type][ref];
+  vp9_prob (*const coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      cpi->common.fc->coef_probs[tx_size][type][ref];
+  unsigned int (*const eob_branch)[COEFF_CONTEXTS] =
+      td->counts->eob_branch[tx_size][type][ref];
+  const uint8_t *const band = get_band_translate(tx_size);
+  const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
+  int16_t token;
+  EXTRABIT extra;
+  int aoff, loff;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
+
+  pt = get_entropy_context(tx_size, pd->above_context + aoff,
+                           pd->left_context + loff);
+  so = get_scan(xd, tx_size, type, block);
+  scan = so->scan;
+  nb = so->neighbors;
+  c = 0;
+
+  while (c < eob) {
+    int v = 0;
+    int skip_eob = 0;
+    v = qcoeff[scan[c]];
+
+    while (!v) {
+      add_token_no_extra(&t, coef_probs[band[c]][pt], ZERO_TOKEN, skip_eob,
+                         counts[band[c]][pt]);
+      eob_branch[band[c]][pt] += !skip_eob;
+
+      skip_eob = 1;
+      token_cache[scan[c]] = 0;
+      ++c;
+      pt = get_coef_context(nb, token_cache, c);
+      v = qcoeff[scan[c]];
+    }
+
+    vp9_get_token_extra(v, &token, &extra);
+
+    add_token(&t, coef_probs[band[c]][pt], extra, (uint8_t)token,
+              (uint8_t)skip_eob, counts[band[c]][pt]);
+    eob_branch[band[c]][pt] += !skip_eob;
+
+    token_cache[scan[c]] = vp9_pt_energy_class[token];
+    ++c;
+    pt = get_coef_context(nb, token_cache, c);
+  }
+  if (c < seg_eob) {
+    add_token_no_extra(&t, coef_probs[band[c]][pt], EOB_TOKEN, 0,
+                       counts[band[c]][pt]);
+    ++eob_branch[band[c]][pt];
+  }
+
+  *tp = t;
+
+  vp9_set_contexts(xd, pd, plane_bsize, tx_size, c > 0, aoff, loff);
+}
+
+struct is_skippable_args {
+  MACROBLOCK *x;
+  int *skippable;
+};
+static void is_skippable(int plane, int block,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                         void *argv) {
+  struct is_skippable_args *args = argv;
+  (void)plane_bsize;
+  (void)tx_size;
+  args->skippable[0] &= (!args->x->plane[plane].eobs[block]);
+}
+
+// TODO(yaowu): rewrite and optimize this function to remove the usage of
+//              vp9_foreach_transform_block() and simplify is_skippable().
+int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  int result = 1;
+  struct is_skippable_args args = {x, &result};
+  vp9_foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane, is_skippable,
+                                         &args);
+  return result;
+}
+
+static void has_high_freq_coeff(int plane, int block,
+                                BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                                void *argv) {
+  struct is_skippable_args *args = argv;
+  int eobs = (tx_size == TX_4X4) ? 3 : 10;
+  (void) plane_bsize;
+
+  *(args->skippable) |= (args->x->plane[plane].eobs[block] > eobs);
+}
+
+int vp9_has_high_freq_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  int result = 0;
+  struct is_skippable_args args = {x, &result};
+  vp9_foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane,
+                                         has_high_freq_coeff, &args);
+  return result;
+}
+
+void vp9_tokenize_sb(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
+                     int dry_run, BLOCK_SIZE bsize) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx = vp9_get_skip_context(xd);
+  const int skip_inc = !vp9_segfeature_active(&cm->seg, mbmi->segment_id,
+                                              SEG_LVL_SKIP);
+  struct tokenize_b_args arg = {cpi, td, t};
+  if (mbmi->skip) {
+    if (!dry_run)
+      td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    return;
+  }
+
+  if (!dry_run) {
+    td->counts->skip[ctx][0] += skip_inc;
+    vp9_foreach_transformed_block(xd, bsize, tokenize_b, &arg);
+  } else {
+    vp9_foreach_transformed_block(xd, bsize, set_entropy_context_b, &arg);
+  }
+}

media/libvpx/vp9/encoder/vp9_tokenize.h

@@ -0,0 +1,112 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_TOKENIZE_H_
+#define VP9_ENCODER_VP9_TOKENIZE_H_
+
+#include "vp9/common/vp9_entropy.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_treewriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define EOSB_TOKEN 127     // Not signalled, encoder only
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  typedef int32_t EXTRABIT;
+#else
+  typedef int16_t EXTRABIT;
+#endif
+
+
+typedef struct {
+  int16_t token;
+  EXTRABIT extra;
+} TOKENVALUE;
+
+typedef struct {
+  const vp9_prob *context_tree;
+  EXTRABIT extra;
+  uint8_t token;
+  uint8_t skip_eob_node;
+} TOKENEXTRA;
+
+extern const vp9_tree_index vp9_coef_tree[];
+extern const vp9_tree_index vp9_coef_con_tree[];
+extern const struct vp9_token vp9_coef_encodings[];
+
+int vp9_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+int vp9_has_high_freq_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+struct VP9_COMP;
+struct ThreadData;
+
+void vp9_tokenize_sb(struct VP9_COMP *cpi, struct ThreadData *td,
+                     TOKENEXTRA **t, int dry_run, BLOCK_SIZE bsize);
+
+extern const int16_t *vp9_dct_value_cost_ptr;
+/* TODO: The Token field should be broken out into a separate char array to
+ *  improve cache locality, since it's needed for costing when the rest of the
+ *  fields are not.
+ */
+extern const TOKENVALUE *vp9_dct_value_tokens_ptr;
+extern const TOKENVALUE *vp9_dct_cat_lt_10_value_tokens;
+extern const int16_t vp9_cat6_low_cost[256];
+extern const int16_t vp9_cat6_high_cost[128];
+extern const int16_t vp9_cat6_high10_high_cost[512];
+extern const int16_t vp9_cat6_high12_high_cost[2048];
+static INLINE int16_t vp9_get_cost(int16_t token, EXTRABIT extrabits,
+                                   const int16_t *cat6_high_table) {
+  if (token != CATEGORY6_TOKEN)
+    return vp9_extra_bits[token].cost[extrabits];
+  return vp9_cat6_low_cost[extrabits & 0xff]
+      + cat6_high_table[extrabits >> 8];
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE const int16_t* vp9_get_high_cost_table(int bit_depth) {
+  return bit_depth == 8 ? vp9_cat6_high_cost
+      : (bit_depth == 10 ? vp9_cat6_high10_high_cost :
+         vp9_cat6_high12_high_cost);
+}
+#else
+static INLINE const int16_t* vp9_get_high_cost_table(int bit_depth) {
+  (void) bit_depth;
+  return vp9_cat6_high_cost;
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static INLINE void vp9_get_token_extra(int v, int16_t *token, EXTRABIT *extra) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    *token = CATEGORY6_TOKEN;
+    if (v >= CAT6_MIN_VAL)
+      *extra = 2 * v - 2 * CAT6_MIN_VAL;
+    else
+      *extra = -2 * v - 2 * CAT6_MIN_VAL + 1;
+    return;
+  }
+  *token = vp9_dct_cat_lt_10_value_tokens[v].token;
+  *extra = vp9_dct_cat_lt_10_value_tokens[v].extra;
+}
+static INLINE int16_t vp9_get_token(int v) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL)
+    return 10;
+  return vp9_dct_cat_lt_10_value_tokens[v].token;
+}
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_TOKENIZE_H_

media/libvpx/vp9/encoder/vp9_treewriter.c

@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/encoder/vp9_treewriter.h"
+
+static void tree2tok(struct vp9_token *tokens, const vp9_tree_index *tree,
+                     int i, int v, int l) {
+  v += v;
+  ++l;
+
+  do {
+    const vp9_tree_index j = tree[i++];
+    if (j <= 0) {
+      tokens[-j].value = v;
+      tokens[-j].len = l;
+    } else {
+      tree2tok(tokens, tree, j, v, l);
+    }
+  } while (++v & 1);
+}
+
+void vp9_tokens_from_tree(struct vp9_token *tokens,
+                          const vp9_tree_index *tree) {
+  tree2tok(tokens, tree, 0, 0, 0);
+}
+
+static unsigned int convert_distribution(unsigned int i, vp9_tree tree,
+                                         unsigned int branch_ct[][2],
+                                         const unsigned int num_events[]) {
+  unsigned int left, right;
+
+  if (tree[i] <= 0)
+    left = num_events[-tree[i]];
+  else
+    left = convert_distribution(tree[i], tree, branch_ct, num_events);
+
+  if (tree[i + 1] <= 0)
+    right = num_events[-tree[i + 1]];
+  else
+    right = convert_distribution(tree[i + 1], tree, branch_ct, num_events);
+
+  branch_ct[i >> 1][0] = left;
+  branch_ct[i >> 1][1] = right;
+  return left + right;
+}
+
+void vp9_tree_probs_from_distribution(vp9_tree tree,
+                                      unsigned int branch_ct[/* n-1 */][2],
+                                      const unsigned int num_events[/* n */]) {
+  convert_distribution(0, tree, branch_ct, num_events);
+}

media/libvpx/vp9/encoder/vp9_treewriter.h

@@ -0,0 +1,51 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_TREEWRITER_H_
+#define VP9_ENCODER_VP9_TREEWRITER_H_
+
+#include "vp9/encoder/vp9_writer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_tree_probs_from_distribution(vp9_tree tree,
+                                      unsigned int branch_ct[ /* n - 1 */ ][2],
+                                      const unsigned int num_events[ /* n */ ]);
+
+struct vp9_token {
+  int value;
+  int len;
+};
+
+void vp9_tokens_from_tree(struct vp9_token*, const vp9_tree_index *);
+
+static INLINE void vp9_write_tree(vp9_writer *w, const vp9_tree_index *tree,
+                                  const vp9_prob *probs, int bits, int len,
+                                  vp9_tree_index i) {
+  do {
+    const int bit = (bits >> --len) & 1;
+    vp9_write(w, bit, probs[i >> 1]);
+    i = tree[i + bit];
+  } while (len);
+}
+
+static INLINE void vp9_write_token(vp9_writer *w, const vp9_tree_index *tree,
+                                   const vp9_prob *probs,
+                                   const struct vp9_token *token) {
+  vp9_write_tree(w, tree, probs, token->value, token->len, 0);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_TREEWRITER_H_

media/libvpx/vp9/encoder/vp9_variance.c

@@ -0,0 +1,380 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+
+#include "vpx_ports/mem.h"
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_filter.h"
+
+#include "vp9/encoder/vp9_variance.h"
+
+static uint8_t bilinear_filters[8][2] = {
+  { 128,   0, },
+  { 112,  16, },
+  {  96,  32, },
+  {  80,  48, },
+  {  64,  64, },
+  {  48,  80, },
+  {  32,  96, },
+  {  16, 112, },
+};
+
+// Applies a 1-D 2-tap bi-linear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// first-pass of 2-D separable filter.
+//
+// Produces int32_t output to retain precision for next pass. Two filter taps
+// should sum to VP9_FILTER_WEIGHT. pixel_step defines whether the filter is
+// applied horizontally (pixel_step=1) or vertically (pixel_step=stride). It
+// defines the offset required to move from one input to the next.
+static void var_filter_block2d_bil_first_pass(const uint8_t *src_ptr,
+                                              uint16_t *output_ptr,
+                                              unsigned int src_pixels_per_line,
+                                              int pixel_step,
+                                              unsigned int output_height,
+                                              unsigned int output_width,
+                                              const uint8_t *vp9_filter) {
+  unsigned int i, j;
+
+  for (i = 0; i < output_height; i++) {
+    for (j = 0; j < output_width; j++) {
+      output_ptr[j] = ROUND_POWER_OF_TWO((int)src_ptr[0] * vp9_filter[0] +
+                          (int)src_ptr[pixel_step] * vp9_filter[1],
+                          FILTER_BITS);
+
+      src_ptr++;
+    }
+
+    // Next row...
+    src_ptr    += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+// Applies a 1-D 2-tap bi-linear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// second-pass of 2-D separable filter.
+//
+// Requires 32-bit input as produced by filter_block2d_bil_first_pass. Two
+// filter taps should sum to VP9_FILTER_WEIGHT. pixel_step defines whether the
+// filter is applied horizontally (pixel_step=1) or vertically (pixel_step=
+// stride). It defines the offset required to move from one input to the next.
+static void var_filter_block2d_bil_second_pass(const uint16_t *src_ptr,
+                                               uint8_t *output_ptr,
+                                               unsigned int src_pixels_per_line,
+                                               unsigned int pixel_step,
+                                               unsigned int output_height,
+                                               unsigned int output_width,
+                                               const uint8_t *vp9_filter) {
+  unsigned int  i, j;
+
+  for (i = 0; i < output_height; i++) {
+    for (j = 0; j < output_width; j++) {
+      output_ptr[j] = ROUND_POWER_OF_TWO((int)src_ptr[0] * vp9_filter[0] +
+                          (int)src_ptr[pixel_step] * vp9_filter[1],
+                          FILTER_BITS);
+      src_ptr++;
+    }
+
+    src_ptr += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+#define SUBPIX_VAR(W, H) \
+unsigned int vp9_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint8_t temp2[H * W]; \
+\
+  var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, W, \
+                                    bilinear_filters[xoffset]); \
+  var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                     bilinear_filters[yoffset]); \
+\
+  return vpx_variance##W##x##H##_c(temp2, W, dst, dst_stride, sse); \
+}
+
+#define SUBPIX_AVG_VAR(W, H) \
+unsigned int vp9_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint8_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint8_t, temp3[H * W]); \
+\
+  var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, W, \
+                                    bilinear_filters[xoffset]); \
+  var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                     bilinear_filters[yoffset]); \
+\
+  vpx_comp_avg_pred(temp3, second_pred, W, H, temp2, W); \
+\
+  return vpx_variance##W##x##H##_c(temp3, W, dst, dst_stride, sse); \
+}
+
+SUBPIX_VAR(4, 4)
+SUBPIX_AVG_VAR(4, 4)
+
+SUBPIX_VAR(4, 8)
+SUBPIX_AVG_VAR(4, 8)
+
+SUBPIX_VAR(8, 4)
+SUBPIX_AVG_VAR(8, 4)
+
+SUBPIX_VAR(8, 8)
+SUBPIX_AVG_VAR(8, 8)
+
+SUBPIX_VAR(8, 16)
+SUBPIX_AVG_VAR(8, 16)
+
+SUBPIX_VAR(16, 8)
+SUBPIX_AVG_VAR(16, 8)
+
+SUBPIX_VAR(16, 16)
+SUBPIX_AVG_VAR(16, 16)
+
+SUBPIX_VAR(16, 32)
+SUBPIX_AVG_VAR(16, 32)
+
+SUBPIX_VAR(32, 16)
+SUBPIX_AVG_VAR(32, 16)
+
+SUBPIX_VAR(32, 32)
+SUBPIX_AVG_VAR(32, 32)
+
+SUBPIX_VAR(32, 64)
+SUBPIX_AVG_VAR(32, 64)
+
+SUBPIX_VAR(64, 32)
+SUBPIX_AVG_VAR(64, 32)
+
+SUBPIX_VAR(64, 64)
+SUBPIX_AVG_VAR(64, 64)
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_var_filter_block2d_bil_first_pass(
+    const uint8_t *src_ptr8,
+    uint16_t *output_ptr,
+    unsigned int src_pixels_per_line,
+    int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const uint8_t *vp9_filter) {
+  unsigned int i, j;
+  uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src_ptr8);
+  for (i = 0; i < output_height; i++) {
+    for (j = 0; j < output_width; j++) {
+      output_ptr[j] =
+          ROUND_POWER_OF_TWO((int)src_ptr[0] * vp9_filter[0] +
+                             (int)src_ptr[pixel_step] * vp9_filter[1],
+                             FILTER_BITS);
+
+      src_ptr++;
+    }
+
+    // Next row...
+    src_ptr += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+static void highbd_var_filter_block2d_bil_second_pass(
+    const uint16_t *src_ptr,
+    uint16_t *output_ptr,
+    unsigned int src_pixels_per_line,
+    unsigned int pixel_step,
+    unsigned int output_height,
+    unsigned int output_width,
+    const uint8_t *vp9_filter) {
+  unsigned int  i, j;
+
+  for (i = 0; i < output_height; i++) {
+    for (j = 0; j < output_width; j++) {
+      output_ptr[j] =
+          ROUND_POWER_OF_TWO((int)src_ptr[0] * vp9_filter[0] +
+                             (int)src_ptr[pixel_step] * vp9_filter[1],
+                             FILTER_BITS);
+      src_ptr++;
+    }
+
+    src_ptr += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+#define HIGHBD_SUBPIX_VAR(W, H) \
+unsigned int vp9_highbd_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  return vpx_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, dst, \
+                                          dst_stride, sse); \
+} \
+\
+unsigned int vp9_highbd_10_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  return vpx_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \
+                                             W, dst, dst_stride, sse); \
+} \
+\
+unsigned int vp9_highbd_12_sub_pixel_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  return vpx_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \
+                                             W, dst, dst_stride, sse); \
+}
+
+#define HIGHBD_SUBPIX_AVG_VAR(W, H) \
+unsigned int vp9_highbd_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  vpx_highbd_comp_avg_pred(temp3, second_pred, W, H, \
+                           CONVERT_TO_BYTEPTR(temp2), W); \
+\
+  return vpx_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, dst, \
+                                          dst_stride, sse); \
+} \
+\
+unsigned int vp9_highbd_10_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  vpx_highbd_comp_avg_pred(temp3, second_pred, W, H, \
+                           CONVERT_TO_BYTEPTR(temp2), W); \
+\
+  return vpx_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), \
+                                             W, dst, dst_stride, sse); \
+} \
+\
+unsigned int vp9_highbd_12_sub_pixel_avg_variance##W##x##H##_c( \
+  const uint8_t *src, int  src_stride, \
+  int xoffset, int  yoffset, \
+  const uint8_t *dst, int dst_stride, \
+  unsigned int *sse, \
+  const uint8_t *second_pred) { \
+  uint16_t fdata3[(H + 1) * W]; \
+  uint16_t temp2[H * W]; \
+  DECLARE_ALIGNED(16, uint16_t, temp3[H * W]); \
+\
+  highbd_var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, \
+                                           W, bilinear_filters[xoffset]); \
+  highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W, \
+                                            bilinear_filters[yoffset]); \
+\
+  vpx_highbd_comp_avg_pred(temp3, second_pred, W, H, \
+                           CONVERT_TO_BYTEPTR(temp2), W); \
+\
+  return vpx_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), \
+                                             W, dst, dst_stride, sse); \
+}
+
+HIGHBD_SUBPIX_VAR(4, 4)
+HIGHBD_SUBPIX_AVG_VAR(4, 4)
+
+HIGHBD_SUBPIX_VAR(4, 8)
+HIGHBD_SUBPIX_AVG_VAR(4, 8)
+
+HIGHBD_SUBPIX_VAR(8, 4)
+HIGHBD_SUBPIX_AVG_VAR(8, 4)
+
+HIGHBD_SUBPIX_VAR(8, 8)
+HIGHBD_SUBPIX_AVG_VAR(8, 8)
+
+HIGHBD_SUBPIX_VAR(8, 16)
+HIGHBD_SUBPIX_AVG_VAR(8, 16)
+
+HIGHBD_SUBPIX_VAR(16, 8)
+HIGHBD_SUBPIX_AVG_VAR(16, 8)
+
+HIGHBD_SUBPIX_VAR(16, 16)
+HIGHBD_SUBPIX_AVG_VAR(16, 16)
+
+HIGHBD_SUBPIX_VAR(16, 32)
+HIGHBD_SUBPIX_AVG_VAR(16, 32)
+
+HIGHBD_SUBPIX_VAR(32, 16)
+HIGHBD_SUBPIX_AVG_VAR(32, 16)
+
+HIGHBD_SUBPIX_VAR(32, 32)
+HIGHBD_SUBPIX_AVG_VAR(32, 32)
+
+HIGHBD_SUBPIX_VAR(32, 64)
+HIGHBD_SUBPIX_AVG_VAR(32, 64)
+
+HIGHBD_SUBPIX_VAR(64, 32)
+HIGHBD_SUBPIX_AVG_VAR(64, 32)
+
+HIGHBD_SUBPIX_VAR(64, 64)
+HIGHBD_SUBPIX_AVG_VAR(64, 64)
+#endif  // CONFIG_VP9_HIGHBITDEPTH

media/libvpx/vp9/encoder/vp9_variance.h

@@ -0,0 +1,81 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_VARIANCE_H_
+#define VP9_ENCODER_VP9_VARIANCE_H_
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef unsigned int(*vp9_sad_fn_t)(const uint8_t *src_ptr,
+                                    int source_stride,
+                                    const uint8_t *ref_ptr,
+                                    int ref_stride);
+
+typedef unsigned int(*vp9_sad_avg_fn_t)(const uint8_t *src_ptr,
+                                        int source_stride,
+                                        const uint8_t *ref_ptr,
+                                        int ref_stride,
+                                        const uint8_t *second_pred);
+
+typedef void (*vp9_sad_multi_fn_t)(const uint8_t *src_ptr,
+                                   int source_stride,
+                                   const uint8_t *ref_ptr,
+                                   int  ref_stride,
+                                   unsigned int *sad_array);
+
+typedef void (*vp9_sad_multi_d_fn_t)(const uint8_t *src_ptr,
+                                     int source_stride,
+                                     const uint8_t* const ref_ptr[],
+                                     int  ref_stride, unsigned int *sad_array);
+
+typedef unsigned int (*vp9_variance_fn_t)(const uint8_t *src_ptr,
+                                          int source_stride,
+                                          const uint8_t *ref_ptr,
+                                          int ref_stride,
+                                          unsigned int *sse);
+
+typedef unsigned int (*vp9_subpixvariance_fn_t)(const uint8_t *src_ptr,
+                                                int source_stride,
+                                                int xoffset,
+                                                int yoffset,
+                                                const uint8_t *ref_ptr,
+                                                int Refstride,
+                                                unsigned int *sse);
+
+typedef unsigned int (*vp9_subp_avg_variance_fn_t)(const uint8_t *src_ptr,
+                                                   int source_stride,
+                                                   int xoffset,
+                                                   int yoffset,
+                                                   const uint8_t *ref_ptr,
+                                                   int Refstride,
+                                                   unsigned int *sse,
+                                                   const uint8_t *second_pred);
+
+typedef struct vp9_variance_vtable {
+  vp9_sad_fn_t               sdf;
+  vp9_sad_avg_fn_t           sdaf;
+  vp9_variance_fn_t          vf;
+  vp9_subpixvariance_fn_t    svf;
+  vp9_subp_avg_variance_fn_t svaf;
+  vp9_sad_multi_fn_t         sdx3f;
+  vp9_sad_multi_fn_t         sdx8f;
+  vp9_sad_multi_d_fn_t       sdx4df;
+} vp9_variance_fn_ptr_t;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_VARIANCE_H_

media/libvpx/vp9/encoder/vp9_write_bit_buffer.c

@@ -0,0 +1,35 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+#include "vp9/encoder/vp9_write_bit_buffer.h"
+
+size_t vp9_wb_bytes_written(const struct vp9_write_bit_buffer *wb) {
+  return wb->bit_offset / CHAR_BIT + (wb->bit_offset % CHAR_BIT > 0);
+}
+
+void vp9_wb_write_bit(struct vp9_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) {
+    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 vp9_wb_write_literal(struct vp9_write_bit_buffer *wb, int data, int bits) {
+  int bit;
+  for (bit = bits - 1; bit >= 0; bit--)
+    vp9_wb_write_bit(wb, (data >> bit) & 1);
+}

media/libvpx/vp9/encoder/vp9_write_bit_buffer.h

@@ -0,0 +1,36 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_WRITE_BIT_BUFFER_H_
+#define VP9_ENCODER_VP9_WRITE_BIT_BUFFER_H_
+
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct vp9_write_bit_buffer {
+  uint8_t *bit_buffer;
+  size_t bit_offset;
+};
+
+size_t vp9_wb_bytes_written(const struct vp9_write_bit_buffer *wb);
+
+void vp9_wb_write_bit(struct vp9_write_bit_buffer *wb, int bit);
+
+void vp9_wb_write_literal(struct vp9_write_bit_buffer *wb, int data, int bits);
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_WRITE_BIT_BUFFER_H_

media/libvpx/vp9/encoder/vp9_writer.c

@@ -0,0 +1,34 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include "vp9/encoder/vp9_writer.h"
+#include "vp9/common/vp9_entropy.h"
+
+void vp9_start_encode(vp9_writer *br, uint8_t *source) {
+  br->lowvalue = 0;
+  br->range    = 255;
+  br->count    = -24;
+  br->buffer   = source;
+  br->pos      = 0;
+  vp9_write_bit(br, 0);
+}
+
+void vp9_stop_encode(vp9_writer *br) {
+  int i;
+
+  for (i = 0; i < 32; i++)
+    vp9_write_bit(br, 0);
+
+  // Ensure there's no ambigous collision with any index marker bytes
+  if ((br->buffer[br->pos - 1] & 0xe0) == 0xc0)
+    br->buffer[br->pos++] = 0;
+}
+

media/libvpx/vp9/encoder/vp9_writer.h

@@ -0,0 +1,98 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_VP9_WRITER_H_
+#define VP9_ENCODER_VP9_WRITER_H_
+
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct vp9_writer {
+  unsigned int lowvalue;
+  unsigned int range;
+  int count;
+  unsigned int pos;
+  uint8_t *buffer;
+} vp9_writer;
+
+void vp9_start_encode(vp9_writer *bc, uint8_t *buffer);
+void vp9_stop_encode(vp9_writer *bc);
+
+static INLINE void vp9_write(vp9_writer *br, int bit, int probability) {
+  unsigned int split;
+  int count = br->count;
+  unsigned int range = br->range;
+  unsigned int lowvalue = br->lowvalue;
+  register unsigned int shift;
+
+  split = 1 + (((range - 1) * probability) >> 8);
+
+  range = split;
+
+  if (bit) {
+    lowvalue += split;
+    range = br->range - split;
+  }
+
+  shift = vp9_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;
+}
+
+static INLINE void vp9_write_bit(vp9_writer *w, int bit) {
+  vp9_write(w, bit, 128);  // vp9_prob_half
+}
+
+static INLINE void vp9_write_literal(vp9_writer *w, int data, int bits) {
+  int bit;
+
+  for (bit = bits - 1; bit >= 0; bit--)
+    vp9_write_bit(w, 1 & (data >> bit));
+}
+
+#define vp9_write_prob(w, v) vp9_write_literal((w), (v), 8)
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_VP9_WRITER_H_

media/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c

@@ -0,0 +1,423 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "./vp9_rtcd.h"
+#include "vpx_ports/mem.h"
+
+void vp9_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp,
+                         int *min, int *max) {
+  __m128i u0, s0, d0, diff, maxabsdiff, minabsdiff, negdiff, absdiff0, absdiff;
+  u0  = _mm_setzero_si128();
+  // Row 0
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff0 = _mm_max_epi16(diff, negdiff);
+  // Row 1
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(absdiff0, absdiff);
+  minabsdiff = _mm_min_epi16(absdiff0, absdiff);
+  // Row 2
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 3
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 4
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 5
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 5 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 6
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 6 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 7
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 7 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_si128(maxabsdiff, 8));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 32));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 16));
+  *max = _mm_extract_epi16(maxabsdiff, 0);
+
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_si128(minabsdiff, 8));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 32));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 16));
+  *min = _mm_extract_epi16(minabsdiff, 0);
+}
+
+unsigned int vp9_avg_8x8_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0  = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 8));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 32));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 32) >> 6;
+}
+
+unsigned int vp9_avg_4x4_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0  = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 4));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 8) >> 4;
+}
+
+static void hadamard_col8_sse2(__m128i *in, int iter) {
+  __m128i a0 = in[0];
+  __m128i a1 = in[1];
+  __m128i a2 = in[2];
+  __m128i a3 = in[3];
+  __m128i a4 = in[4];
+  __m128i a5 = in[5];
+  __m128i a6 = in[6];
+  __m128i a7 = in[7];
+
+  __m128i b0 = _mm_add_epi16(a0, a1);
+  __m128i b1 = _mm_sub_epi16(a0, a1);
+  __m128i b2 = _mm_add_epi16(a2, a3);
+  __m128i b3 = _mm_sub_epi16(a2, a3);
+  __m128i b4 = _mm_add_epi16(a4, a5);
+  __m128i b5 = _mm_sub_epi16(a4, a5);
+  __m128i b6 = _mm_add_epi16(a6, a7);
+  __m128i b7 = _mm_sub_epi16(a6, a7);
+
+  a0 = _mm_add_epi16(b0, b2);
+  a1 = _mm_add_epi16(b1, b3);
+  a2 = _mm_sub_epi16(b0, b2);
+  a3 = _mm_sub_epi16(b1, b3);
+  a4 = _mm_add_epi16(b4, b6);
+  a5 = _mm_add_epi16(b5, b7);
+  a6 = _mm_sub_epi16(b4, b6);
+  a7 = _mm_sub_epi16(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm_add_epi16(a0, a4);
+    b7 = _mm_add_epi16(a1, a5);
+    b3 = _mm_add_epi16(a2, a6);
+    b4 = _mm_add_epi16(a3, a7);
+    b2 = _mm_sub_epi16(a0, a4);
+    b6 = _mm_sub_epi16(a1, a5);
+    b1 = _mm_sub_epi16(a2, a6);
+    b5 = _mm_sub_epi16(a3, a7);
+
+    a0 = _mm_unpacklo_epi16(b0, b1);
+    a1 = _mm_unpacklo_epi16(b2, b3);
+    a2 = _mm_unpackhi_epi16(b0, b1);
+    a3 = _mm_unpackhi_epi16(b2, b3);
+    a4 = _mm_unpacklo_epi16(b4, b5);
+    a5 = _mm_unpacklo_epi16(b6, b7);
+    a6 = _mm_unpackhi_epi16(b4, b5);
+    a7 = _mm_unpackhi_epi16(b6, b7);
+
+    b0 = _mm_unpacklo_epi32(a0, a1);
+    b1 = _mm_unpacklo_epi32(a4, a5);
+    b2 = _mm_unpackhi_epi32(a0, a1);
+    b3 = _mm_unpackhi_epi32(a4, a5);
+    b4 = _mm_unpacklo_epi32(a2, a3);
+    b5 = _mm_unpacklo_epi32(a6, a7);
+    b6 = _mm_unpackhi_epi32(a2, a3);
+    b7 = _mm_unpackhi_epi32(a6, a7);
+
+    in[0] = _mm_unpacklo_epi64(b0, b1);
+    in[1] = _mm_unpackhi_epi64(b0, b1);
+    in[2] = _mm_unpacklo_epi64(b2, b3);
+    in[3] = _mm_unpackhi_epi64(b2, b3);
+    in[4] = _mm_unpacklo_epi64(b4, b5);
+    in[5] = _mm_unpackhi_epi64(b4, b5);
+    in[6] = _mm_unpacklo_epi64(b6, b7);
+    in[7] = _mm_unpackhi_epi64(b6, b7);
+  } else {
+    in[0] = _mm_add_epi16(a0, a4);
+    in[7] = _mm_add_epi16(a1, a5);
+    in[3] = _mm_add_epi16(a2, a6);
+    in[4] = _mm_add_epi16(a3, a7);
+    in[2] = _mm_sub_epi16(a0, a4);
+    in[6] = _mm_sub_epi16(a1, a5);
+    in[1] = _mm_sub_epi16(a2, a6);
+    in[5] = _mm_sub_epi16(a3, a7);
+  }
+}
+
+void vp9_hadamard_8x8_sse2(int16_t const *src_diff, int src_stride,
+                           int16_t *coeff) {
+  __m128i src[8];
+  src[0] = _mm_load_si128((const __m128i *)src_diff);
+  src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[7] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+
+  hadamard_col8_sse2(src, 0);
+  hadamard_col8_sse2(src, 1);
+
+  _mm_store_si128((__m128i *)coeff, src[0]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[1]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[2]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[3]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[4]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[5]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[6]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[7]);
+}
+
+void vp9_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride,
+                             int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    int16_t const *src_ptr = src_diff + (idx >> 1) * 8 * src_stride
+                                + (idx & 0x01) * 8;
+    vp9_hadamard_8x8_sse2(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  for (idx = 0; idx < 64; idx += 8) {
+    __m128i coeff0 = _mm_load_si128((const __m128i *)coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(coeff + 64));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(coeff + 128));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(coeff + 192));
+
+    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
+    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
+    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
+    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    coeff0 = _mm_srai_epi16(coeff0, 1);
+    coeff1 = _mm_srai_epi16(coeff1, 1);
+    _mm_store_si128((__m128i *)coeff, coeff0);
+    _mm_store_si128((__m128i *)(coeff + 64), coeff1);
+
+    coeff2 = _mm_sub_epi16(b0, b2);
+    coeff3 = _mm_sub_epi16(b1, b3);
+    coeff2 = _mm_srai_epi16(coeff2, 1);
+    coeff3 = _mm_srai_epi16(coeff3, 1);
+    _mm_store_si128((__m128i *)(coeff + 128), coeff2);
+    _mm_store_si128((__m128i *)(coeff + 192), coeff3);
+
+    coeff += 8;
+  }
+}
+
+int16_t vp9_satd_sse2(const int16_t *coeff, int length) {
+  int i;
+  __m128i sum = _mm_load_si128((const __m128i *)coeff);
+  __m128i sign = _mm_srai_epi16(sum, 15);
+  __m128i val = _mm_xor_si128(sum, sign);
+  sum = _mm_sub_epi16(val, sign);
+  coeff += 8;
+
+  for (i = 8; i < length; i += 8) {
+    __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    sign = _mm_srai_epi16(src_line, 15);
+    val = _mm_xor_si128(src_line, sign);
+    val = _mm_sub_epi16(val, sign);
+    sum = _mm_add_epi16(sum, val);
+    coeff += 8;
+  }
+
+  val = _mm_srli_si128(sum, 8);
+  sum = _mm_add_epi16(sum, val);
+  val = _mm_srli_epi64(sum, 32);
+  sum = _mm_add_epi16(sum, val);
+  val = _mm_srli_epi32(sum, 16);
+  sum = _mm_add_epi16(sum, val);
+
+  return _mm_extract_epi16(sum, 0);
+}
+
+void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
+                          const int ref_stride, const int height) {
+  int idx;
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_loadu_si128((const __m128i *)ref);
+  __m128i s0 = _mm_unpacklo_epi8(src_line, zero);
+  __m128i s1 = _mm_unpackhi_epi8(src_line, zero);
+  __m128i t0, t1;
+  int height_1 = height - 1;
+  ref += ref_stride;
+
+  for (idx = 1; idx < height_1; idx += 2) {
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+  }
+
+  src_line = _mm_loadu_si128((const __m128i *)ref);
+  t0 = _mm_unpacklo_epi8(src_line, zero);
+  t1 = _mm_unpackhi_epi8(src_line, zero);
+  s0 = _mm_adds_epu16(s0, t0);
+  s1 = _mm_adds_epu16(s1, t1);
+
+  if (height == 64) {
+    s0 = _mm_srai_epi16(s0, 5);
+    s1 = _mm_srai_epi16(s1, 5);
+  } else if (height == 32) {
+    s0 = _mm_srai_epi16(s0, 4);
+    s1 = _mm_srai_epi16(s1, 4);
+  } else {
+    s0 = _mm_srai_epi16(s0, 3);
+    s1 = _mm_srai_epi16(s1, 3);
+  }
+
+  _mm_storeu_si128((__m128i *)hbuf, s0);
+  hbuf += 8;
+  _mm_storeu_si128((__m128i *)hbuf, s1);
+}
+
+int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width) {
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_load_si128((const __m128i *)ref);
+  __m128i s0 = _mm_sad_epu8(src_line, zero);
+  __m128i s1;
+  int i;
+
+  for (i = 16; i < width; i += 16) {
+    ref += 16;
+    src_line = _mm_load_si128((const __m128i *)ref);
+    s1 = _mm_sad_epu8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, s1);
+  }
+
+  s1 = _mm_srli_si128(s0, 8);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  return _mm_extract_epi16(s0, 0);
+}
+
+int vp9_vector_var_sse2(int16_t const *ref, int16_t const *src,
+                        const int bwl) {
+  int idx;
+  int width = 4 << bwl;
+  int16_t mean;
+  __m128i v0 = _mm_loadu_si128((const __m128i *)ref);
+  __m128i v1 = _mm_load_si128((const __m128i *)src);
+  __m128i diff = _mm_subs_epi16(v0, v1);
+  __m128i sum = diff;
+  __m128i sse = _mm_madd_epi16(diff, diff);
+
+  ref += 8;
+  src += 8;
+
+  for (idx = 8; idx < width; idx += 8) {
+    v0 = _mm_loadu_si128((const __m128i *)ref);
+    v1 = _mm_load_si128((const __m128i *)src);
+    diff = _mm_subs_epi16(v0, v1);
+
+    sum = _mm_add_epi16(sum, diff);
+    v0  = _mm_madd_epi16(diff, diff);
+    sse = _mm_add_epi32(sse, v0);
+
+    ref += 8;
+    src += 8;
+  }
+
+  v0  = _mm_srli_si128(sum, 8);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi64(sum, 32);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi32(sum, 16);
+  sum = _mm_add_epi16(sum, v0);
+
+  v1  = _mm_srli_si128(sse, 8);
+  sse = _mm_add_epi32(sse, v1);
+  v1  = _mm_srli_epi64(sse, 32);
+  sse = _mm_add_epi32(sse, v1);
+
+  mean = _mm_extract_epi16(sum, 0);
+
+  return _mm_cvtsi128_si32(sse) - ((mean * mean) >> (bwl + 2));
+}

media/libvpx/vp9/encoder/x86/vp9_dct_avx2.c

@@ -0,0 +1,26 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+#include "vp9/common/vp9_idct.h"  // for cospi constants
+#include "vpx_ports/mem.h"
+
+
+#define FDCT32x32_2D_AVX2 vp9_fdct32x32_rd_avx2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "vp9/encoder/x86/vp9_dct32x32_avx2_impl.h"
+#undef  FDCT32x32_2D_AVX2
+#undef  FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D_AVX2 vp9_fdct32x32_avx2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "vp9/encoder/x86/vp9_dct32x32_avx2_impl.h" // NOLINT
+#undef  FDCT32x32_2D_AVX2
+#undef  FDCT32x32_HIGH_PRECISION

media/libvpx/vp9/encoder/x86/vp9_dct_mmx.asm

@@ -0,0 +1,101 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro TRANSFORM_COLS 0
+  paddw           m0,        m1
+  movq            m4,        m0
+  psubw           m3,        m2
+  psubw           m4,        m3
+  psraw           m4,        1
+  movq            m5,        m4
+  psubw           m5,        m1 ;b1
+  psubw           m4,        m2 ;c1
+  psubw           m0,        m4
+  paddw           m3,        m5
+                                ; m0 a0
+  SWAP            1,         4  ; m1 c1
+  SWAP            2,         3  ; m2 d1
+  SWAP            3,         5  ; m3 b1
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+  movq            m4,        m0
+  movq            m5,        m2
+  punpcklwd       m4,        m1
+  punpckhwd       m0,        m1
+  punpcklwd       m5,        m3
+  punpckhwd       m2,        m3
+  movq            m1,        m4
+  movq            m3,        m0
+  punpckldq       m1,        m5
+  punpckhdq       m4,        m5
+  punpckldq       m3,        m2
+  punpckhdq       m0,        m2
+  SWAP            2, 3, 0, 1, 4
+%endmacro
+
+INIT_MMX mmx
+cglobal fwht4x4, 3, 4, 8, input, output, stride
+  lea             r3q,       [inputq + strideq*4]
+  movq            m0,        [inputq] ;a1
+  movq            m1,        [inputq + strideq*2] ;b1
+  movq            m2,        [r3q] ;c1
+  movq            m3,        [r3q + strideq*2] ;d1
+
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+
+  psllw           m0,        2
+  psllw           m1,        2
+  psllw           m2,        2
+  psllw           m3,        2
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  pxor            m4,             m4
+  pxor            m5,             m5
+  pcmpgtw         m4,             m0
+  pcmpgtw         m5,             m1
+  movq            m6,             m0
+  movq            m7,             m1
+  punpcklwd       m0,             m4
+  punpcklwd       m1,             m5
+  punpckhwd       m6,             m4
+  punpckhwd       m7,             m5
+  movq            [outputq],      m0
+  movq            [outputq + 8],  m6
+  movq            [outputq + 16], m1
+  movq            [outputq + 24], m7
+  pxor            m4,             m4
+  pxor            m5,             m5
+  pcmpgtw         m4,             m2
+  pcmpgtw         m5,             m3
+  movq            m6,             m2
+  movq            m7,             m3
+  punpcklwd       m2,             m4
+  punpcklwd       m3,             m5
+  punpckhwd       m6,             m4
+  punpckhwd       m7,             m5
+  movq            [outputq + 32], m2
+  movq            [outputq + 40], m6
+  movq            [outputq + 48], m3
+  movq            [outputq + 56], m7
+%else
+  movq            [outputq],      m0
+  movq            [outputq + 8],  m1
+  movq            [outputq + 16], m2
+  movq            [outputq + 24], m3
+%endif
+
+  RET

media/libvpx/vp9/encoder/x86/vp9_dct_sse2.h

@@ -0,0 +1,464 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_X86_VP9_DCT_SSE2_H_
+#define VP9_ENCODER_X86_VP9_DCT_SSE2_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define pair_set_epi32(a, b) \
+  _mm_set_epi32((int)(b), (int)(a), (int)(b), (int)(a))
+
+void vp9_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
+void vp9_highbd_fdct4x4_sse2(const int16_t *input, tran_low_t *output,
+                            int stride);
+void vp9_highbd_fdct8x8_sse2(const int16_t *input, tran_low_t *output,
+                            int stride);
+void vp9_highbd_fdct16x16_sse2(const int16_t *input, tran_low_t *output,
+                               int stride);
+
+static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
+  __m128i buf0, buf1;
+  buf0 = _mm_mul_epu32(a, b);
+  a = _mm_srli_epi64(a, 32);
+  b = _mm_srli_epi64(b, 32);
+  buf1 = _mm_mul_epu32(a, b);
+  return _mm_add_epi64(buf0, buf1);
+}
+
+static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) {
+  __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm_unpacklo_epi64(buf0, buf1);
+}
+
+static INLINE int check_epi16_overflow_x2(const __m128i *preg0,
+                                          const __m128i *preg1) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  cmp0 = _mm_or_si128(cmp0, cmp1);
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x4(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow),
+                              _mm_cmpeq_epi16(*preg2, min_overflow));
+  __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow),
+                              _mm_cmpeq_epi16(*preg3, min_overflow));
+  cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3));
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x8(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3,
+                                          const __m128i *preg4,
+                                          const __m128i *preg5,
+                                          const __m128i *preg6,
+                                          const __m128i *preg7) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x12(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0)
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x16(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11,
+                                           const __m128i *preg12,
+                                           const __m128i *preg13,
+                                           const __m128i *preg14,
+                                           const __m128i *preg15) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1)
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+  }
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x32(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11,
+                                           const __m128i *preg12,
+                                           const __m128i *preg13,
+                                           const __m128i *preg14,
+                                           const __m128i *preg15,
+                                           const __m128i *preg16,
+                                           const __m128i *preg17,
+                                           const __m128i *preg18,
+                                           const __m128i *preg19,
+                                           const __m128i *preg20,
+                                           const __m128i *preg21,
+                                           const __m128i *preg22,
+                                           const __m128i *preg23,
+                                           const __m128i *preg24,
+                                           const __m128i *preg25,
+                                           const __m128i *preg26,
+                                           const __m128i *preg27,
+                                           const __m128i *preg28,
+                                           const __m128i *preg29,
+                                           const __m128i *preg30,
+                                           const __m128i *preg31) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1) {
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+      if (!res0) {
+        res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19);
+        if (!res1) {
+          res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23);
+          if (!res0) {
+            res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27);
+            if (!res1)
+              res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31);
+          }
+        }
+      }
+    }
+  }
+  return res0 + res1;
+}
+
+static INLINE int k_check_epi32_overflow_4(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *zero) {
+  __m128i minus_one = _mm_set1_epi32(-1);
+  // Check for overflows
+  __m128i reg0_shifted = _mm_slli_epi64(*preg0, 1);
+  __m128i reg1_shifted = _mm_slli_epi64(*preg1, 1);
+  __m128i reg2_shifted = _mm_slli_epi64(*preg2, 1);
+  __m128i reg3_shifted = _mm_slli_epi64(*preg3, 1);
+  __m128i reg0_top_dwords = _mm_shuffle_epi32(
+      reg0_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg1_top_dwords = _mm_shuffle_epi32(
+      reg1_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg2_top_dwords = _mm_shuffle_epi32(
+      reg2_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg3_top_dwords = _mm_shuffle_epi32(
+      reg3_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i top_dwords_01 = _mm_unpacklo_epi64(reg0_top_dwords, reg1_top_dwords);
+  __m128i top_dwords_23 = _mm_unpacklo_epi64(reg2_top_dwords, reg3_top_dwords);
+  __m128i valid_positve_01 = _mm_cmpeq_epi32(top_dwords_01, *zero);
+  __m128i valid_positve_23 = _mm_cmpeq_epi32(top_dwords_23, *zero);
+  __m128i valid_negative_01 = _mm_cmpeq_epi32(top_dwords_01, minus_one);
+  __m128i valid_negative_23 = _mm_cmpeq_epi32(top_dwords_23, minus_one);
+  int overflow_01 = _mm_movemask_epi8(
+      _mm_cmpeq_epi32(valid_positve_01, valid_negative_01));
+  int overflow_23 = _mm_movemask_epi8(
+      _mm_cmpeq_epi32(valid_positve_23, valid_negative_23));
+  return (overflow_01 + overflow_23);
+}
+
+static INLINE int k_check_epi32_overflow_8(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_16(const __m128i *preg0,
+                                            const __m128i *preg1,
+                                            const __m128i *preg2,
+                                            const __m128i *preg3,
+                                            const __m128i *preg4,
+                                            const __m128i *preg5,
+                                            const __m128i *preg6,
+                                            const __m128i *preg7,
+                                            const __m128i *preg8,
+                                            const __m128i *preg9,
+                                            const __m128i *preg10,
+                                            const __m128i *preg11,
+                                            const __m128i *preg12,
+                                            const __m128i *preg13,
+                                            const __m128i *preg14,
+                                            const __m128i *preg15,
+                                            const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11,
+                                          zero);
+      if (!overflow) {
+        overflow = k_check_epi32_overflow_4(preg12, preg13, preg14, preg15,
+                                            zero);
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_32(const __m128i *preg0,
+                                            const __m128i *preg1,
+                                            const __m128i *preg2,
+                                            const __m128i *preg3,
+                                            const __m128i *preg4,
+                                            const __m128i *preg5,
+                                            const __m128i *preg6,
+                                            const __m128i *preg7,
+                                            const __m128i *preg8,
+                                            const __m128i *preg9,
+                                            const __m128i *preg10,
+                                            const __m128i *preg11,
+                                            const __m128i *preg12,
+                                            const __m128i *preg13,
+                                            const __m128i *preg14,
+                                            const __m128i *preg15,
+                                            const __m128i *preg16,
+                                            const __m128i *preg17,
+                                            const __m128i *preg18,
+                                            const __m128i *preg19,
+                                            const __m128i *preg20,
+                                            const __m128i *preg21,
+                                            const __m128i *preg22,
+                                            const __m128i *preg23,
+                                            const __m128i *preg24,
+                                            const __m128i *preg25,
+                                            const __m128i *preg26,
+                                            const __m128i *preg27,
+                                            const __m128i *preg28,
+                                            const __m128i *preg29,
+                                            const __m128i *preg30,
+                                            const __m128i *preg31,
+                                            const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11, zero);
+      if (!overflow) {
+        overflow = k_check_epi32_overflow_4(preg12, preg13, preg14, preg15,
+                                            zero);
+        if (!overflow) {
+          overflow = k_check_epi32_overflow_4(preg16, preg17, preg18, preg19,
+                                              zero);
+          if (!overflow) {
+            overflow = k_check_epi32_overflow_4(preg20, preg21,
+                                                preg22, preg23, zero);
+            if (!overflow) {
+              overflow = k_check_epi32_overflow_4(preg24, preg25,
+                                                  preg26, preg27, zero);
+              if (!overflow) {
+                overflow = k_check_epi32_overflow_4(preg28, preg29,
+                                                    preg30, preg31, zero);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE void store_output(const __m128i *poutput, tran_low_t* dst_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_store_si128((__m128i *)(dst_ptr), out0);
+  _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_store_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+static INLINE void storeu_output(const __m128i *poutput, tran_low_t* dst_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_storeu_si128((__m128i *)(dst_ptr), out0);
+  _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_storeu_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+
+static INLINE __m128i mult_round_shift(const __m128i *pin0,
+                                       const __m128i *pin1,
+                                       const __m128i *pmultiplier,
+                                       const __m128i *prounding,
+                                       const int shift) {
+  const __m128i u0 = _mm_madd_epi16(*pin0, *pmultiplier);
+  const __m128i u1 = _mm_madd_epi16(*pin1, *pmultiplier);
+  const __m128i v0 = _mm_add_epi32(u0, *prounding);
+  const __m128i v1 = _mm_add_epi32(u1, *prounding);
+  const __m128i w0 = _mm_srai_epi32(v0, shift);
+  const __m128i w1 = _mm_srai_epi32(v1, shift);
+  return _mm_packs_epi32(w0, w1);
+}
+
+static INLINE void transpose_and_output8x8(
+    const __m128i *pin00, const __m128i *pin01,
+    const __m128i *pin02, const __m128i *pin03,
+    const __m128i *pin04, const __m128i *pin05,
+    const __m128i *pin06, const __m128i *pin07,
+    const int pass, int16_t* out0_ptr,
+    tran_low_t* out1_ptr) {
+  // 00 01 02 03 04 05 06 07
+  // 10 11 12 13 14 15 16 17
+  // 20 21 22 23 24 25 26 27
+  // 30 31 32 33 34 35 36 37
+  // 40 41 42 43 44 45 46 47
+  // 50 51 52 53 54 55 56 57
+  // 60 61 62 63 64 65 66 67
+  // 70 71 72 73 74 75 76 77
+  const __m128i tr0_0 = _mm_unpacklo_epi16(*pin00, *pin01);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(*pin02, *pin03);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(*pin00, *pin01);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(*pin02, *pin03);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(*pin04, *pin05);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(*pin06, *pin07);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(*pin04, *pin05);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(*pin06, *pin07);
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  // 04 14 05 15 06 16 07 17
+  // 24 34 25 35 26 36 27 37
+  // 40 50 41 51 42 52 43 53
+  // 60 70 61 71 62 72 63 73
+  // 54 54 55 55 56 56 57 57
+  // 64 74 65 75 66 76 67 77
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+  // 00 10 20 30 01 11 21 31
+  // 40 50 60 70 41 51 61 71
+  // 02 12 22 32 03 13 23 33
+  // 42 52 62 72 43 53 63 73
+  // 04 14 24 34 05 15 21 36
+  // 44 54 64 74 45 55 61 76
+  // 06 16 26 36 07 17 27 37
+  // 46 56 66 76 47 57 67 77
+  const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+  const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+  const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+  const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+  const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+  // 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
+  if (pass == 0) {
+    _mm_storeu_si128((__m128i*)(out0_ptr + 0 * 16), tr2_0);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 1 * 16), tr2_1);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 2 * 16), tr2_2);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 3 * 16), tr2_3);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 4 * 16), tr2_4);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 5 * 16), tr2_5);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 6 * 16), tr2_6);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 7 * 16), tr2_7);
+  } else {
+    storeu_output(&tr2_0, (out1_ptr + 0 * 16));
+    storeu_output(&tr2_1, (out1_ptr + 1 * 16));
+    storeu_output(&tr2_2, (out1_ptr + 2 * 16));
+    storeu_output(&tr2_3, (out1_ptr + 3 * 16));
+    storeu_output(&tr2_4, (out1_ptr + 4 * 16));
+    storeu_output(&tr2_5, (out1_ptr + 5 * 16));
+    storeu_output(&tr2_6, (out1_ptr + 6 * 16));
+    storeu_output(&tr2_7, (out1_ptr + 7 * 16));
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_X86_VP9_DCT_SSE2_H_

media/libvpx/vp9/encoder/x86/vp9_dct_ssse3.c

@@ -0,0 +1,471 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#if defined(_MSC_VER) && _MSC_VER <= 1500
+// Need to include math.h before calling tmmintrin.h/intrin.h
+// in certain versions of MSVS.
+#include <math.h>
+#endif
+#include <tmmintrin.h>  // SSSE3
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/x86/vp9_idct_intrin_sse2.h"
+
+void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
+                             int16_t* coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t* zbin_ptr,
+                             const int16_t* round_ptr, const int16_t* quant_ptr,
+                             const int16_t* quant_shift_ptr,
+                             int16_t* qcoeff_ptr,
+                             int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                             uint16_t* eob_ptr,
+                             const int16_t* scan_ptr,
+                             const int16_t* iscan_ptr) {
+  __m128i zero;
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__dual_p16_p16 = dual_set_epi16(23170, 23170);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // Load input
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  __m128i *in[8];
+  int index = 0;
+
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)coeff_ptr;
+
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  in[0] = &in0;
+  in[1] = &in1;
+  in[2] = &in2;
+  in[3] = &in3;
+  in[4] = &in4;
+  in[5] = &in5;
+  in[6] = &in6;
+  in[7] = &in7;
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = _mm_add_epi16(in0, in7);
+    const __m128i q1 = _mm_add_epi16(in1, in6);
+    const __m128i q2 = _mm_add_epi16(in2, in5);
+    const __m128i q3 = _mm_add_epi16(in3, in4);
+    const __m128i q4 = _mm_sub_epi16(in3, in4);
+    const __m128i q5 = _mm_sub_epi16(in2, in5);
+    const __m128i q6 = _mm_sub_epi16(in1, in6);
+    const __m128i q7 = _mm_sub_epi16(in0, in7);
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = _mm_add_epi16(q0, q3);
+      const __m128i r1 = _mm_add_epi16(q1, q2);
+      const __m128i r2 = _mm_sub_epi16(q1, q2);
+      const __m128i r3 = _mm_sub_epi16(q0, q3);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+      // dct_const_round_shift
+
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+
+      res0 = _mm_packs_epi32(w0, w1);
+      res4 = _mm_packs_epi32(w2, w3);
+      res2 = _mm_packs_epi32(w4, w5);
+      res6 = _mm_packs_epi32(w6, w7);
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_sub_epi16(q6, q5);
+      const __m128i d1 = _mm_add_epi16(q6, q5);
+      const __m128i r0 = _mm_mulhrs_epi16(d0, k__dual_p16_p16);
+      const __m128i r1 = _mm_mulhrs_epi16(d1, k__dual_p16_p16);
+
+      // Add/subtract
+      const __m128i x0 = _mm_add_epi16(q4, r0);
+      const __m128i x1 = _mm_sub_epi16(q4, r0);
+      const __m128i x2 = _mm_sub_epi16(q7, r1);
+      const __m128i x3 = _mm_add_epi16(q7, r1);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res1 = _mm_packs_epi32(w0, w1);
+      res7 = _mm_packs_epi32(w2, w3);
+      res5 = _mm_packs_epi32(w4, w5);
+      res3 = _mm_packs_epi32(w6, w7);
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 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
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+  }
+
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant, thr;
+    int16_t nzflag;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = *in[0];
+        coeff1 = *in[1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    index = 2;
+    thr = _mm_srai_epi16(dequant, 1);
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
+        coeff0 = *in[index];
+        coeff1 = *in[index + 1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+            _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+      index += 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}

media/libvpx/vp9/encoder/x86/vp9_dct_ssse3_x86_64.asm

@@ -0,0 +1,255 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the forward transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_%2_m%1:  dw  %2, -%1,  %2, -%1,  %2, -%1,  %2, -%1
+%endmacro
+
+TRANSFORM_COEFFS 11585,  11585
+TRANSFORM_COEFFS 15137,   6270
+TRANSFORM_COEFFS 16069,   3196
+TRANSFORM_COEFFS  9102,  13623
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_%4_%3], [pw_%3_m%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_%4_%3], [pw_%3_m%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+; 1D forward 8x8 DCT transform
+%macro FDCT8_1D 1
+  SUM_SUB            0,  7,  9
+  SUM_SUB            1,  6,  9
+  SUM_SUB            2,  5,  9
+  SUM_SUB            3,  4,  9
+
+  SUM_SUB            0,  3,  9
+  SUM_SUB            1,  2,  9
+  SUM_SUB            6,  5,  9
+%if %1 == 0
+  SUM_SUB            0,  1,  9
+%endif
+
+  BUTTERFLY_4X       2,  3,  6270,  15137,  m8,  9,  10
+
+  pmulhrsw           m6, m12
+  pmulhrsw           m5, m12
+%if %1 == 0
+  pmulhrsw           m0, m12
+  pmulhrsw           m1, m12
+%else
+  BUTTERFLY_4X       1,  0,  11585, 11585,  m8,  9,  10
+  SWAP               0,  1
+%endif
+
+  SUM_SUB            4,  5,  9
+  SUM_SUB            7,  6,  9
+  BUTTERFLY_4X       4,  7,  3196,  16069,  m8,  9,  10
+  BUTTERFLY_4X       5,  6,  13623,  9102,  m8,  9,  10
+  SWAP               1,  4
+  SWAP               3,  6
+%endmacro
+
+%macro DIVIDE_ROUND_2X 4 ; dst1, dst2, tmp1, tmp2
+  psraw              m%3, m%1, 15
+  psraw              m%4, m%2, 15
+  psubw              m%1, m%3
+  psubw              m%2, m%4
+  psraw              m%1, 1
+  psraw              m%2, 1
+%endmacro
+
+INIT_XMM ssse3
+cglobal fdct8x8, 3, 5, 13, input, output, stride
+
+  mova               m8, [pd_8192]
+  mova              m12, [pw_11585x2]
+  pxor              m11, m11
+
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  ; left shift by 2 to increase forward transformation precision
+  psllw              m0, 2
+  psllw              m1, 2
+  psllw              m2, 2
+  psllw              m3, 2
+  psllw              m4, 2
+  psllw              m5, 2
+  psllw              m6, 2
+  psllw              m7, 2
+
+  ; column transform
+  FDCT8_1D  0
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  FDCT8_1D  1
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  DIVIDE_ROUND_2X   0, 1, 9, 10
+  DIVIDE_ROUND_2X   2, 3, 9, 10
+  DIVIDE_ROUND_2X   4, 5, 9, 10
+  DIVIDE_ROUND_2X   6, 7, 9, 10
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+
+%macro HMD8_1D 0
+  psubw              m8, m0, m1
+  psubw              m9, m2, m3
+  paddw              m0, m1
+  paddw              m2, m3
+  SWAP               1, 8
+  SWAP               3, 9
+  psubw              m8, m4, m5
+  psubw              m9, m6, m7
+  paddw              m4, m5
+  paddw              m6, m7
+  SWAP               5, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m2
+  psubw              m9, m1, m3
+  paddw              m0, m2
+  paddw              m1, m3
+  SWAP               2, 8
+  SWAP               3, 9
+  psubw              m8, m4, m6
+  psubw              m9, m5, m7
+  paddw              m4, m6
+  paddw              m5, m7
+  SWAP               6, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m4
+  psubw              m9, m1, m5
+  paddw              m0, m4
+  paddw              m1, m5
+  SWAP               4, 8
+  SWAP               5, 9
+  psubw              m8, m2, m6
+  psubw              m9, m3, m7
+  paddw              m2, m6
+  paddw              m3, m7
+  SWAP               6, 8
+  SWAP               7, 9
+%endmacro
+
+INIT_XMM ssse3
+cglobal hadamard_8x8, 3, 5, 10, input, stride, output
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  HMD8_1D
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+  HMD8_1D
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+%endif

media/libvpx/vp9/encoder/x86/vp9_denoiser_sse2.c

@@ -0,0 +1,375 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx_ports/emmintrin_compat.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+#include "vpx_mem/vpx_mem.h"
+
+// Compute the sum of all pixel differences of this MB.
+static INLINE int sum_diff_16x1(__m128i acc_diff) {
+  const __m128i k_1 = _mm_set1_epi16(1);
+  const __m128i acc_diff_lo =
+      _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_hi =
+      _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
+  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
+  const __m128i hgfe_dcba =
+      _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
+  const __m128i hgfedcba =
+      _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
+  return _mm_cvtsi128_si32(hgfedcba);
+}
+
+// Denoise a 16x1 vector.
+static INLINE __m128i vp9_denoiser_16x1_sse2(const uint8_t *sig,
+                                             const uint8_t *mc_running_avg_y,
+                                             uint8_t *running_avg_y,
+                                             const __m128i *k_0,
+                                             const __m128i *k_4,
+                                             const __m128i *k_8,
+                                             const __m128i *k_16,
+                                             const __m128i *l3,
+                                             const __m128i *l32,
+                                             const __m128i *l21,
+                                             __m128i acc_diff) {
+  // Calculate differences
+  const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y =
+      _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
+  __m128i v_running_avg_y;
+  const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+  const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+  // Obtain the sign. FF if diff is negative.
+  const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, *k_0);
+  // Clamp absolute difference to 16 to be used to get mask. Doing this
+  // allows us to use _mm_cmpgt_epi8, which operates on signed byte.
+  const __m128i clamped_absdiff =
+      _mm_min_epu8(_mm_or_si128(pdiff, ndiff), *k_16);
+  // Get masks for l2 l1 and l0 adjustments.
+  const __m128i mask2 = _mm_cmpgt_epi8(*k_16, clamped_absdiff);
+  const __m128i mask1 = _mm_cmpgt_epi8(*k_8, clamped_absdiff);
+  const __m128i mask0 = _mm_cmpgt_epi8(*k_4, clamped_absdiff);
+  // Get adjustments for l2, l1, and l0.
+  __m128i adj2 = _mm_and_si128(mask2, *l32);
+  const __m128i adj1 = _mm_and_si128(mask1, *l21);
+  const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+  __m128i adj,  padj, nadj;
+
+  // Combine the adjustments and get absolute adjustments.
+  adj2 = _mm_add_epi8(adj2, adj1);
+  adj = _mm_sub_epi8(*l3, adj2);
+  adj = _mm_andnot_si128(mask0, adj);
+  adj = _mm_or_si128(adj, adj0);
+
+  // Restore the sign and get positive and negative adjustments.
+  padj = _mm_andnot_si128(diff_sign, adj);
+  nadj = _mm_and_si128(diff_sign, adj);
+
+  // Calculate filtered value.
+  v_running_avg_y = _mm_adds_epu8(v_sig, padj);
+  v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
+  _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+  // Adjustments <=7, and each element in acc_diff can fit in signed
+  // char.
+  acc_diff = _mm_adds_epi8(acc_diff, padj);
+  acc_diff = _mm_subs_epi8(acc_diff, nadj);
+  return acc_diff;
+}
+
+// Denoise a 16x1 vector with a weaker filter.
+static INLINE __m128i vp9_denoiser_adj_16x1_sse2(
+    const uint8_t *sig, const uint8_t *mc_running_avg_y,
+    uint8_t *running_avg_y, const __m128i k_0,
+    const __m128i k_delta, __m128i acc_diff) {
+  __m128i v_running_avg_y = _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
+  // Calculate differences.
+  const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y =
+      _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
+  const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+  const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+  // Obtain the sign. FF if diff is negative.
+  const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+  // Clamp absolute difference to delta to get the adjustment.
+  const __m128i adj =
+      _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+  // Restore the sign and get positive and negative adjustments.
+  __m128i padj, nadj;
+  padj = _mm_andnot_si128(diff_sign, adj);
+  nadj = _mm_and_si128(diff_sign, adj);
+  // Calculate filtered value.
+  v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
+  v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
+  _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+  // Accumulate the adjustments.
+  acc_diff = _mm_subs_epi8(acc_diff, padj);
+  acc_diff = _mm_adds_epi8(acc_diff, nadj);
+  return acc_diff;
+}
+
+// Denoiser for 4xM and 8xM blocks.
+static int vp9_denoiser_NxM_sse2_small(
+    const uint8_t *sig, int sig_stride, const uint8_t *mc_running_avg_y,
+    int mc_avg_y_stride, uint8_t *running_avg_y, int avg_y_stride,
+    int increase_denoising, BLOCK_SIZE bs, int motion_magnitude, int width) {
+  int sum_diff_thresh, r, sum_diff = 0;
+  const int shift_inc  = (increase_denoising &&
+                          motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                         1 : 0;
+  uint8_t sig_buffer[8][16], mc_running_buffer[8][16], running_buffer[8][16];
+  __m128i acc_diff = _mm_setzero_si128();
+  const __m128i k_0 = _mm_setzero_si128();
+  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+  const __m128i k_8 = _mm_set1_epi8(8);
+  const __m128i k_16 = _mm_set1_epi8(16);
+  // Modify each level's adjustment according to motion_magnitude.
+  const __m128i l3 = _mm_set1_epi8(
+      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
+  // Difference between level 3 and level 2 is 2.
+  const __m128i l32 = _mm_set1_epi8(2);
+  // Difference between level 2 and level 1 is 1.
+  const __m128i l21 = _mm_set1_epi8(1);
+  const uint8_t shift = (width == 4) ? 2 : 1;
+
+  for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> shift); ++r) {
+    memcpy(sig_buffer[r], sig, width);
+    memcpy(sig_buffer[r] + width, sig + sig_stride, width);
+    memcpy(mc_running_buffer[r], mc_running_avg_y, width);
+    memcpy(mc_running_buffer[r] + width,
+           mc_running_avg_y + mc_avg_y_stride, width);
+    memcpy(running_buffer[r], running_avg_y, width);
+    memcpy(running_buffer[r] + width, running_avg_y + avg_y_stride, width);
+    if (width == 4) {
+      memcpy(sig_buffer[r] + width * 2, sig + sig_stride * 2, width);
+      memcpy(sig_buffer[r] + width * 3, sig + sig_stride * 3, width);
+      memcpy(mc_running_buffer[r] + width * 2,
+             mc_running_avg_y + mc_avg_y_stride * 2, width);
+      memcpy(mc_running_buffer[r] + width * 3,
+             mc_running_avg_y + mc_avg_y_stride * 3, width);
+      memcpy(running_buffer[r] + width * 2,
+             running_avg_y + avg_y_stride * 2, width);
+      memcpy(running_buffer[r] + width * 3,
+             running_avg_y + avg_y_stride * 3, width);
+    }
+    acc_diff = vp9_denoiser_16x1_sse2(sig_buffer[r],
+                                      mc_running_buffer[r],
+                                      running_buffer[r],
+                                      &k_0, &k_4, &k_8, &k_16,
+                                      &l3, &l32, &l21, acc_diff);
+    memcpy(running_avg_y, running_buffer[r], width);
+    memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, width);
+    if (width == 4) {
+      memcpy(running_avg_y + avg_y_stride * 2,
+             running_buffer[r] + width * 2, width);
+      memcpy(running_avg_y + avg_y_stride * 3,
+             running_buffer[r] + width * 3, width);
+    }
+    // Update pointers for next iteration.
+    sig += (sig_stride << shift);
+    mc_running_avg_y += (mc_avg_y_stride << shift);
+    running_avg_y += (avg_y_stride << shift);
+  }
+
+  {
+    sum_diff = sum_diff_16x1(acc_diff);
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising),
+      // check if we can still apply some (weaker) temporal filtering to
+      // this block, that would otherwise not be denoised at all. Simplest
+      // is to apply an additional adjustment to running_avg_y to bring it
+      // closer to sig. The adjustment is capped by a maximum delta, and
+      // chosen such that in most cases the resulting sum_diff will be
+      // within the acceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over the
+      // threshold.
+      const int delta = ((abs(sum_diff) - sum_diff_thresh) >>
+                         num_pels_log2_lookup[bs]) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        running_avg_y -= avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> shift); ++r) {
+          acc_diff = vp9_denoiser_adj_16x1_sse2(
+              sig_buffer[r], mc_running_buffer[r], running_buffer[r],
+              k_0, k_delta, acc_diff);
+          memcpy(running_avg_y, running_buffer[r], width);
+          memcpy(running_avg_y + avg_y_stride,
+                 running_buffer[r] + width, width);
+          if (width == 4) {
+            memcpy(running_avg_y + avg_y_stride * 2,
+                   running_buffer[r] + width * 2, width);
+            memcpy(running_avg_y + avg_y_stride * 3,
+                   running_buffer[r] + width * 3, width);
+          }
+          // Update pointers for next iteration.
+          running_avg_y += (avg_y_stride << shift);
+        }
+        sum_diff = sum_diff_16x1(acc_diff);
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+// Denoiser for 16xM, 32xM and 64xM blocks
+static int vp9_denoiser_NxM_sse2_big(const uint8_t *sig, int sig_stride,
+                                     const uint8_t *mc_running_avg_y,
+                                     int mc_avg_y_stride,
+                                     uint8_t *running_avg_y,
+                                     int avg_y_stride,
+                                     int increase_denoising, BLOCK_SIZE bs,
+                                     int motion_magnitude) {
+  int sum_diff_thresh, r, c, sum_diff = 0;
+  const int shift_inc  = (increase_denoising &&
+                          motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                         1 : 0;
+  __m128i acc_diff[4][4];
+  const __m128i k_0 = _mm_setzero_si128();
+  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+  const __m128i k_8 = _mm_set1_epi8(8);
+  const __m128i k_16 = _mm_set1_epi8(16);
+  // Modify each level's adjustment according to motion_magnitude.
+  const __m128i l3 = _mm_set1_epi8(
+      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
+  // Difference between level 3 and level 2 is 2.
+  const __m128i l32 = _mm_set1_epi8(2);
+  // Difference between level 2 and level 1 is 1.
+  const __m128i l21 = _mm_set1_epi8(1);
+
+  for (c = 0; c < 4; ++c) {
+    for (r = 0; r < 4; ++r) {
+      acc_diff[c][r] = _mm_setzero_si128();
+    }
+  }
+
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+      acc_diff[c>>4][r>>4] = vp9_denoiser_16x1_sse2(
+          sig, mc_running_avg_y, running_avg_y, &k_0, &k_4,
+          &k_8, &k_16, &l3, &l32, &l21, acc_diff[c>>4][r>>4]);
+      // Update pointers for next iteration.
+      sig += 16;
+      mc_running_avg_y += 16;
+      running_avg_y += 16;
+    }
+
+    if ((r + 1) % 16 == 0 || (bs == BLOCK_16X8 && r == 7)) {
+      for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+        sum_diff += sum_diff_16x1(acc_diff[c>>4][r>>4]);
+      }
+    }
+
+    // Update pointers for next iteration.
+    sig = sig - 16 * ((4 << b_width_log2_lookup[bs]) >> 4) + sig_stride;
+    mc_running_avg_y = mc_running_avg_y -
+                       16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                       mc_avg_y_stride;
+    running_avg_y = running_avg_y -
+                    16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                    avg_y_stride;
+  }
+
+  {
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      const int delta = ((abs(sum_diff) - sum_diff_thresh) >>
+                         num_pels_log2_lookup[bs]) + 1;
+
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        sig -= sig_stride * (4 << b_height_log2_lookup[bs]);
+        mc_running_avg_y -= mc_avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        running_avg_y -= avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        sum_diff = 0;
+        for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+          for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+            acc_diff[c>>4][r>>4] = vp9_denoiser_adj_16x1_sse2(
+                sig, mc_running_avg_y, running_avg_y, k_0,
+                k_delta, acc_diff[c>>4][r>>4]);
+            // Update pointers for next iteration.
+            sig += 16;
+            mc_running_avg_y += 16;
+            running_avg_y += 16;
+          }
+
+          if ((r + 1) % 16 == 0 || (bs == BLOCK_16X8 && r == 7)) {
+            for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+              sum_diff += sum_diff_16x1(acc_diff[c>>4][r>>4]);
+            }
+          }
+          sig = sig - 16 * ((4 << b_width_log2_lookup[bs]) >> 4) + sig_stride;
+          mc_running_avg_y = mc_running_avg_y -
+                             16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                             mc_avg_y_stride;
+          running_avg_y = running_avg_y -
+                          16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                          avg_y_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+int vp9_denoiser_filter_sse2(const uint8_t *sig, int sig_stride,
+                             const uint8_t *mc_avg,
+                             int mc_avg_stride,
+                             uint8_t *avg, int avg_stride,
+                             int increase_denoising,
+                             BLOCK_SIZE bs,
+                             int motion_magnitude) {
+  if (bs == BLOCK_4X4 || bs == BLOCK_4X8) {
+    return vp9_denoiser_NxM_sse2_small(sig, sig_stride,
+                                       mc_avg, mc_avg_stride,
+                                       avg, avg_stride,
+                                       increase_denoising,
+                                       bs, motion_magnitude, 4);
+  } else if (bs == BLOCK_8X4 || bs == BLOCK_8X8 || bs == BLOCK_8X16) {
+    return vp9_denoiser_NxM_sse2_small(sig, sig_stride,
+                                       mc_avg, mc_avg_stride,
+                                       avg, avg_stride,
+                                       increase_denoising,
+                                       bs, motion_magnitude, 8);
+  } else if (bs == BLOCK_16X8 || bs == BLOCK_16X16 || bs == BLOCK_16X32 ||
+             bs == BLOCK_32X16|| bs == BLOCK_32X32 || bs == BLOCK_32X64 ||
+             bs == BLOCK_64X32 || bs == BLOCK_64X64) {
+    return vp9_denoiser_NxM_sse2_big(sig, sig_stride,
+                                     mc_avg, mc_avg_stride,
+                                     avg, avg_stride,
+                                     increase_denoising,
+                                     bs, motion_magnitude);
+  } else {
+    return COPY_BLOCK;
+  }
+}

media/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c

@@ -0,0 +1,73 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Usee of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./vp9_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+int64_t vp9_block_error_avx2(const int16_t *coeff,
+                             const int16_t *dqcoeff,
+                             intptr_t block_size,
+                             int64_t *ssz) {
+  __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
+  __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
+  __m256i sse_reg_64hi, ssz_reg_64hi;
+  __m128i sse_reg128, ssz_reg128;
+  int64_t sse;
+  int i;
+  const __m256i zero_reg = _mm256_set1_epi16(0);
+
+  // init sse and ssz registerd to zero
+  sse_reg = _mm256_set1_epi16(0);
+  ssz_reg = _mm256_set1_epi16(0);
+
+  for (i = 0 ; i < block_size ; i+= 16) {
+    // load 32 bytes from coeff and dqcoeff
+    coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
+    dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
+    // dqcoeff - coeff
+    dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
+    // madd (dqcoeff - coeff)
+    dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
+    // madd coeff
+    coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
+    // expand each double word of madd (dqcoeff - coeff) to quad word
+    exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
+    exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
+    // expand each double word of madd (coeff) to quad word
+    exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
+    exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
+    // add each quad word of madd (dqcoeff - coeff) and madd (coeff)
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
+  }
+  // save the higher 64 bit of each 128 bit lane
+  sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
+  ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
+  // add the higher 64 bit to the low 64 bit
+  sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
+  ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
+
+  // add each 64 bit from each of the 128 bit lane of the 256 bit
+  sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
+                             _mm256_extractf128_si256(sse_reg, 1));
+
+  ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
+                             _mm256_extractf128_si256(ssz_reg, 1));
+
+  // store the results
+  _mm_storel_epi64((__m128i*)(&sse), sse_reg128);
+
+  _mm_storel_epi64((__m128i*)(ssz), ssz_reg128);
+  return sse;
+}

media/libvpx/vp9/encoder/x86/vp9_error_sse2.asm

@@ -0,0 +1,120 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; int64_t vp9_block_error(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size,
+;                         int64_t *ssz)
+
+INIT_XMM sse2
+cglobal block_error, 3, 3, 8, uqc, dqc, size, ssz
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m6, m6                 ; ssz accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  pmaddwd   m2, m2
+  pmaddwd   m3, m3
+  ; accumulate in 64bit
+  punpckldq m7, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m7
+  punpckldq m7, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m7
+  punpckldq m7, m2, m5
+  paddq     m4, m1
+  punpckhdq m2, m5
+  paddq     m6, m7
+  punpckldq m7, m3, m5
+  paddq     m6, m2
+  punpckhdq m3, m5
+  paddq     m6, m7
+  paddq     m6, m3
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  movhlps   m7, m6
+  paddq     m4, m5
+  paddq     m6, m7
+%if ARCH_X86_64
+  movq    rax, m4
+  movq [sszq], m6
+%else
+  mov     eax, sszm
+  pshufd   m5, m4, 0x1
+  movq  [eax], m6
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
+
+; Compute the sum of squared difference between two int16_t vectors.
+; int64_t vp9_block_error_fp(int16_t *coeff, int16_t *dqcoeff,
+;                            intptr_t block_size)
+
+INIT_XMM sse2
+cglobal block_error_fp, 3, 3, 6, uqc, dqc, size
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  ; accumulate in 64bit
+  punpckldq m3, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m3
+  punpckldq m3, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m3
+  paddq     m4, m1
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  paddq     m4, m5
+%if ARCH_X86_64
+  movq    rax, m4
+%else
+  pshufd   m5, m4, 0x1
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET

media/libvpx/vp9/encoder/x86/vp9_highbd_block_error_intrin_sse2.c

@@ -0,0 +1,71 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+#include <stdio.h>
+
+#include "vp9/common/vp9_common.h"
+
+int64_t vp9_highbd_block_error_sse2(tran_low_t *coeff, tran_low_t *dqcoeff,
+                                    intptr_t block_size, int64_t *ssz,
+                                    int bps) {
+  int i, j, test;
+  uint32_t temp[4];
+  __m128i max, min, cmp0, cmp1, cmp2, cmp3;
+  int64_t error = 0, sqcoeff = 0;
+  const int shift = 2 * (bps - 8);
+  const int rounding = shift > 0 ? 1 << (shift - 1) : 0;
+
+  for (i = 0; i < block_size; i+=8) {
+    // Load the data into xmm registers
+    __m128i mm_coeff = _mm_load_si128((__m128i*) (coeff + i));
+    __m128i mm_coeff2 = _mm_load_si128((__m128i*) (coeff + i + 4));
+    __m128i mm_dqcoeff = _mm_load_si128((__m128i*) (dqcoeff + i));
+    __m128i mm_dqcoeff2 = _mm_load_si128((__m128i*) (dqcoeff + i + 4));
+    // Check if any values require more than 15 bit
+    max = _mm_set1_epi32(0x3fff);
+    min = _mm_set1_epi32(0xffffc000);
+    cmp0 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff, max),
+            _mm_cmplt_epi32(mm_coeff, min));
+    cmp1 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff2, max),
+            _mm_cmplt_epi32(mm_coeff2, min));
+    cmp2 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff, max),
+            _mm_cmplt_epi32(mm_dqcoeff, min));
+    cmp3 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff2, max),
+            _mm_cmplt_epi32(mm_dqcoeff2, min));
+    test = _mm_movemask_epi8(_mm_or_si128(_mm_or_si128(cmp0, cmp1),
+            _mm_or_si128(cmp2, cmp3)));
+
+    if (!test) {
+      __m128i mm_diff, error_sse2, sqcoeff_sse2;;
+      mm_coeff = _mm_packs_epi32(mm_coeff, mm_coeff2);
+      mm_dqcoeff = _mm_packs_epi32(mm_dqcoeff, mm_dqcoeff2);
+      mm_diff = _mm_sub_epi16(mm_coeff, mm_dqcoeff);
+      error_sse2 = _mm_madd_epi16(mm_diff, mm_diff);
+      sqcoeff_sse2 = _mm_madd_epi16(mm_coeff, mm_coeff);
+      _mm_storeu_si128((__m128i*)temp, error_sse2);
+      error = error + temp[0] + temp[1] + temp[2] + temp[3];
+      _mm_storeu_si128((__m128i*)temp, sqcoeff_sse2);
+      sqcoeff += temp[0] + temp[1] + temp[2] + temp[3];
+    } else {
+      for (j = 0; j < 8; j++) {
+        const int64_t diff = coeff[i + j] - dqcoeff[i + j];
+        error +=  diff * diff;
+        sqcoeff += (int64_t)coeff[i + j] * (int64_t)coeff[i + j];
+      }
+    }
+  }
+  assert(error >= 0 && sqcoeff >= 0);
+  error = (error + rounding) >> shift;
+  sqcoeff = (sqcoeff + rounding) >> shift;
+
+  *ssz = sqcoeff;
+  return error;
+}

media/libvpx/vp9/encoder/x86/vp9_highbd_quantize_intrin_sse2.c

@@ -0,0 +1,181 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_common.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// from vp9_idct.h: typedef int32_t tran_low_t;
+void vp9_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr,
+                                intptr_t count,
+                                int skip_block,
+                                const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr,
+                                tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr,
+                                uint16_t *eob_ptr,
+                                const int16_t *scan,
+                                const int16_t *iscan) {
+  int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+  __m128i zbins[2];
+  __m128i nzbins[2];
+
+  zbins[0] = _mm_set_epi32((int)zbin_ptr[1],
+                           (int)zbin_ptr[1],
+                           (int)zbin_ptr[1],
+                           (int)zbin_ptr[0]);
+  zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  (void)scan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = ((int)count / 4) - 1; i >= 0; i--) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (test == 0xffff)
+        non_zero_regs--;
+      else
+        break;
+    }
+
+    // Quantization pass:
+    for (i = 0; i < non_zero_regs; i++) {
+      __m128i coeffs, coeffs_sign, tmp1, tmp2;
+      int test;
+      int abs_coeff[4];
+      int coeff_sign[4];
+
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      coeffs_sign = _mm_srai_epi32(coeffs, 31);
+      coeffs = _mm_sub_epi32(
+            _mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+      tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+      tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+      tmp1 = _mm_or_si128(tmp1, tmp2);
+      test = _mm_movemask_epi8(tmp1);
+      _mm_storeu_si128((__m128i*)abs_coeff, coeffs);
+      _mm_storeu_si128((__m128i*)coeff_sign, coeffs_sign);
+
+      for (j = 0; j < 4; j++) {
+        if (test & (1 << (4 * j))) {
+          int k = 4 * i + j;
+          int64_t tmp = clamp(abs_coeff[j] + round_ptr[k != 0],
+                              INT32_MIN, INT32_MAX);
+          tmp = ((((tmp * quant_ptr[k != 0]) >> 16) + tmp) *
+                    quant_shift_ptr[k != 0]) >> 16;  // quantization
+          qcoeff_ptr[k] = (tmp ^ coeff_sign[j]) - coeff_sign[j];
+          dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+          if (tmp)
+            eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
+        }
+      }
+    }
+  }
+  *eob_ptr = eob_i + 1;
+}
+
+
+void vp9_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr,
+                                      intptr_t n_coeffs,
+                                      int skip_block,
+                                      const int16_t *zbin_ptr,
+                                      const int16_t *round_ptr,
+                                      const int16_t *quant_ptr,
+                                      const int16_t *quant_shift_ptr,
+                                      tran_low_t *qcoeff_ptr,
+                                      tran_low_t *dqcoeff_ptr,
+                                      const int16_t *dequant_ptr,
+                                      uint16_t *eob_ptr,
+                                      const int16_t *scan,
+                                      const int16_t *iscan) {
+  __m128i zbins[2];
+  __m128i nzbins[2];
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
+  const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
+  (void)scan;
+  zbins[0] = _mm_set_epi32(zbin1_tmp,
+                           zbin1_tmp,
+                           zbin1_tmp,
+                           zbin0_tmp);
+  zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs / 4; i++) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (!(test & 0xf))
+        idx_arr[idx++] = i * 4;
+      if (!(test & 0xf0))
+        idx_arr[idx++] = i * 4 + 1;
+      if (!(test & 0xf00))
+        idx_arr[idx++] = i * 4 + 2;
+      if (!(test & 0xf000))
+        idx_arr[idx++] = i * 4 + 3;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = idx_arr[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      int64_t tmp = clamp(abs_coeff +
+                          ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                          INT32_MIN, INT32_MAX);
+      tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+               quant_shift_ptr[rc != 0]) >> 15;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+
+      if (tmp)
+        eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif

media/libvpx/vp9/encoder/x86/vp9_highbd_variance_sse2.c

@@ -0,0 +1,349 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+
+#include "vp9/encoder/vp9_variance.h"
+#include "vpx_ports/mem.h"
+
+#define DECL(w, opt) \
+int vp9_highbd_sub_pixel_variance##w##xh_##opt(const uint16_t *src, \
+                                               ptrdiff_t src_stride, \
+                                               int x_offset, int y_offset, \
+                                               const uint16_t *dst, \
+                                               ptrdiff_t dst_stride, \
+                                               int height, unsigned int *sse);
+#define DECLS(opt1, opt2) \
+DECL(8, opt1); \
+DECL(16, opt1)
+
+DECLS(sse2, sse);
+// DECLS(ssse3, ssse3);
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+uint32_t vp9_highbd_sub_pixel_variance##w##x##h##_##opt(const uint8_t *src8, \
+                                                        int src_stride, \
+                                                        int x_offset, \
+                                                        int y_offset, \
+                                                        const uint8_t *dst8, \
+                                                        int dst_stride, \
+                                                        uint32_t *sse_ptr) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  int se = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst, dst_stride, h, \
+                                                       &sse); \
+  if (w > wf) { \
+    unsigned int sse2; \
+    int se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 16, \
+                                                          src_stride, \
+                                                          x_offset, y_offset, \
+                                                          dst + 16, \
+                                                          dst_stride, \
+                                                          h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 32, dst_stride, \
+                                                        h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+          src + 48, src_stride, x_offset, y_offset, \
+          dst + 48, dst_stride, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_10_sub_pixel_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  int se = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst, dst_stride, \
+                                                       h, &sse); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 16, \
+                                                          src_stride, \
+                                                          x_offset, y_offset, \
+                                                          dst + 16, \
+                                                          dst_stride, \
+                                                          h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 32, dst_stride, \
+                                                        h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 48, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 48, dst_stride, \
+                                                        h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 2); \
+  sse = ROUND_POWER_OF_TWO(sse, 4); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_12_sub_pixel_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \
+  int start_row; \
+  uint32_t sse; \
+  int se = 0; \
+  uint64_t long_sse = 0; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  for (start_row = 0; start_row < h; start_row +=16) { \
+    uint32_t sse2; \
+    int height = h - start_row < 16 ? h - start_row : 16; \
+    int se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+        src + (start_row * src_stride), src_stride, \
+        x_offset, y_offset, dst + (start_row * dst_stride), \
+        dst_stride, height, &sse2); \
+    se += se2; \
+    long_sse += sse2; \
+    if (w > wf) { \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+          src + 16 + (start_row * src_stride), src_stride, \
+          x_offset, y_offset, dst + 16 + (start_row * dst_stride), \
+          dst_stride, height, &sse2); \
+      se += se2; \
+      long_sse += sse2; \
+      if (w > wf * 2) { \
+        se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+            src + 32 + (start_row * src_stride), src_stride, \
+            x_offset, y_offset, dst + 32 + (start_row * dst_stride), \
+            dst_stride, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+        se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+            src + 48 + (start_row * src_stride), src_stride, \
+            x_offset, y_offset, dst + 48 + (start_row * dst_stride), \
+            dst_stride, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+      }\
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 4); \
+  sse = ROUND_POWER_OF_TWO(long_sse, 8); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+#define FNS(opt1, opt2) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
+FN(16, 8, 16, 4, 3, opt1, (int64_t)); \
+FN(8, 16, 8, 3, 4, opt1, (int64_t)); \
+FN(8, 8, 8, 3, 3, opt1, (int64_t)); \
+FN(8, 4, 8, 3, 2, opt1, (int64_t));
+
+
+FNS(sse2, sse);
+
+#undef FNS
+#undef FN
+
+#define DECL(w, opt) \
+int vp9_highbd_sub_pixel_avg_variance##w##xh_##opt(const uint16_t *src, \
+                                                   ptrdiff_t src_stride, \
+                                                   int x_offset, int y_offset, \
+                                                   const uint16_t *dst, \
+                                                   ptrdiff_t dst_stride, \
+                                                   const uint16_t *sec, \
+                                                   ptrdiff_t sec_stride, \
+                                                   int height, \
+                                                   unsigned int *sse);
+#define DECLS(opt1) \
+DECL(16, opt1) \
+DECL(8, opt1)
+
+DECLS(sse2);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+uint32_t vp9_highbd_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  int se = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+               src, src_stride, x_offset, \
+               y_offset, dst, dst_stride, sec, w, h, &sse); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                  src + 16, src_stride, x_offset, y_offset, \
+                  dst + 16, dst_stride, sec + 16, w, h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 32, src_stride, x_offset, y_offset, \
+                dst + 32, dst_stride, sec + 32, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 48, src_stride, x_offset, y_offset, \
+                dst + 48, dst_stride, sec + 48, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  int se = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src, src_stride, x_offset, \
+                                            y_offset, dst, dst_stride, \
+                                            sec, w, h, &sse); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 16, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 16, dst_stride, \
+                                            sec + 16, w, h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 32, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 32, dst_stride, \
+                                            sec + 32, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 48, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 48, dst_stride, \
+                                            sec + 48, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 2); \
+  sse = ROUND_POWER_OF_TWO(sse, 4); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  int start_row; \
+  uint32_t sse; \
+  int se = 0; \
+  uint64_t long_sse = 0; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  for (start_row = 0; start_row < h; start_row +=16) { \
+    uint32_t sse2; \
+    int height = h - start_row < 16 ? h - start_row : 16; \
+    int se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + (start_row * src_stride), src_stride, x_offset, \
+                y_offset, dst + (start_row * dst_stride), dst_stride, \
+                sec + (start_row * w), w, height, &sse2); \
+    se += se2; \
+    long_sse += sse2; \
+    if (w > wf) { \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 16 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 16 + (start_row * dst_stride), dst_stride, \
+                sec + 16 + (start_row * w), w, height, &sse2); \
+      se += se2; \
+      long_sse += sse2; \
+      if (w > wf * 2) { \
+        se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 32 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 32 + (start_row * dst_stride), dst_stride, \
+                sec + 32 + (start_row * w), w, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+        se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 48 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 48 + (start_row * dst_stride), dst_stride, \
+                sec + 48 + (start_row * w), w, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+      } \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 4); \
+  sse = ROUND_POWER_OF_TWO(long_sse, 8); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+
+#define FNS(opt1) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
+FN(16, 8, 16, 4, 3, opt1, (int64_t)); \
+FN(8, 16, 8, 4, 3, opt1, (int64_t)); \
+FN(8, 8, 8, 3, 3, opt1, (int64_t)); \
+FN(8, 4, 8, 3, 2, opt1, (int64_t));
+
+FNS(sse2);
+
+#undef FNS
+#undef FN