Update aom to slightly newer commit ID

This commit is contained in:
trav90 2018-10-17 05:59:08 -05:00 • committed by Roy Tam
commit 76f1e6edca
311 changed files with 55292 additions and 33790 deletions

View file

@ -2,8 +2,6 @@
%define ARCH_MIPS 0
%define ARCH_X86 0
%define ARCH_X86_64 1
%define HAVE_EDSP 0
%define HAVE_MEDIA 0
%define HAVE_NEON 0
%define HAVE_NEON_ASM 0
%define HAVE_MIPS32 0
@ -45,8 +43,6 @@
%define CONFIG_AV1_ENCODER 1
%define CONFIG_AV1_DECODER 1
%define CONFIG_AV1 1
%define CONFIG_ENCODERS 1
%define CONFIG_DECODERS 1
%define CONFIG_STATIC_MSVCRT 0
%define CONFIG_SPATIAL_RESAMPLING 1
%define CONFIG_REALTIME_ONLY 0
@ -66,36 +62,35 @@
%define CONFIG_ENCODE_PERF_TESTS 0
%define CONFIG_COEFFICIENT_RANGE_CHECKING 0
%define CONFIG_LOWBITDEPTH 1
%define CONFIG_HIGHBITDEPTH 0
%define CONFIG_HIGHBITDEPTH 1
%define CONFIG_EXPERIMENTAL 0
%define CONFIG_SIZE_LIMIT 1
%define CONFIG_FP_MB_STATS 0
%define CONFIG_CDEF 1
%define CONFIG_VAR_TX 0
%define CONFIG_VAR_TX 1
%define CONFIG_RECT_TX 1
%define CONFIG_REF_MV 1
%define CONFIG_RECT_TX_EXT 0
%define CONFIG_TPL_MV 0
%define CONFIG_DUAL_FILTER 1
%define CONFIG_CONVOLVE_ROUND 0
%define CONFIG_COMPOUND_ROUND 0
%define CONFIG_EXT_TX 0
%define CONFIG_EXT_TX 1
%define CONFIG_DPCM_INTRA 0
%define CONFIG_TX64X64 0
%define CONFIG_SUB8X8_MC 0
%define CONFIG_EXT_INTRA 1
%define CONFIG_INTRA_INTERP 0
%define CONFIG_FILTER_INTRA 0
%define CONFIG_INTRA_EDGE 0
%define CONFIG_INTRABC 0
%define CONFIG_EXT_INTER 0
%define CONFIG_INTERINTRA 0
%define CONFIG_WEDGE 0
%define CONFIG_COMPOUND_SEGMENT 0
%define CONFIG_EXT_INTER 1
%define CONFIG_INTERINTRA 1
%define CONFIG_WEDGE 1
%define CONFIG_COMPOUND_SEGMENT 1
%define CONFIG_EXT_REFS 1
%define CONFIG_GLOBAL_MOTION 1
%define CONFIG_NEW_QUANT 0
%define CONFIG_SUPERTX 0
%define CONFIG_ANS 0
%define CONFIG_EC_MULTISYMBOL 1
%define CONFIG_NEW_TOKENSET 1
%define CONFIG_LOOP_RESTORATION 0
%define CONFIG_EXT_PARTITION 0
%define CONFIG_EXT_PARTITION_TYPES 0
@ -105,20 +100,18 @@
%define CONFIG_NCOBMC 0
%define CONFIG_WARPED_MOTION 1
%define CONFIG_Q_ADAPT_PROBS 0
%define CONFIG_SUBFRAME_PROB_UPDATE 0
%define CONFIG_BITSTREAM_DEBUG 0
%define CONFIG_ALT_INTRA 1
%define CONFIG_PALETTE 1
%define CONFIG_PALETTE_DELTA_ENCODING 0
%define CONFIG_DAALA_EC 1
%define CONFIG_RAWBITS 0
%define CONFIG_EC_SMALLMUL 0
%define CONFIG_EC_SMALLMUL 1
%define CONFIG_PVQ 0
%define CONFIG_CFL 0
%define CONFIG_XIPHRC 0
%define CONFIG_CB4X4 1
%define CONFIG_CHROMA_2X2 0
%define CONFIG_CHROMA_SUB8X8 0
%define CONFIG_CHROMA_SUB8X8 1
%define CONFIG_FRAME_SIZE 0
%define CONFIG_DELTA_Q 1
%define CONFIG_EXT_DELTA_Q 1
@ -126,10 +119,10 @@
%define CONFIG_FILTER_7BIT 1
%define CONFIG_PARALLEL_DEBLOCKING 0
%define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
%define CONFIG_LOOPFILTERING_ACROSS_TILES 0
%define CONFIG_LOOPFILTERING_ACROSS_TILES 1
%define CONFIG_TILE_GROUPS 1
%define CONFIG_EC_ADAPT 1
%define CONFIG_TEMPMV_SIGNALING 0
%define CONFIG_TEMPMV_SIGNALING 1
%define CONFIG_RD_DEBUG 0
%define CONFIG_REFERENCE_BUFFER 1
%define CONFIG_COEF_INTERLEAVE 0
@ -147,5 +140,7 @@
%define CONFIG_NEW_MULTISYMBOL 0
%define CONFIG_COMPOUND_SINGLEREF 0
%define CONFIG_AOM_QM 0
%define CONFIG_LOWDELAY_COMPOUND 0
%define CONFIG_ONE_SIDED_COMPOUND 1
%define CONFIG_SMOOTH_HV 0
%define CONFIG_VAR_REFS 0
%define CONFIG_ANALYZER 0

View file

@ -15,8 +15,6 @@
#define ARCH_MIPS 0
#define ARCH_X86 0
#define ARCH_X86_64 1
#define HAVE_EDSP 0
#define HAVE_MEDIA 0
#define HAVE_NEON 0
#define HAVE_NEON_ASM 0
#define HAVE_MIPS32 0
@ -58,8 +56,6 @@
#define CONFIG_AV1_ENCODER 1
#define CONFIG_AV1_DECODER 1
#define CONFIG_AV1 1
#define CONFIG_ENCODERS 1
#define CONFIG_DECODERS 1
#define CONFIG_STATIC_MSVCRT 0
#define CONFIG_SPATIAL_RESAMPLING 1
#define CONFIG_REALTIME_ONLY 0
@ -79,36 +75,35 @@
#define CONFIG_ENCODE_PERF_TESTS 0
#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
#define CONFIG_LOWBITDEPTH 1
#define CONFIG_HIGHBITDEPTH 0
#define CONFIG_HIGHBITDEPTH 1
#define CONFIG_EXPERIMENTAL 0
#define CONFIG_SIZE_LIMIT 1
#define CONFIG_FP_MB_STATS 0
#define CONFIG_CDEF 1
#define CONFIG_VAR_TX 0
#define CONFIG_VAR_TX 1
#define CONFIG_RECT_TX 1
#define CONFIG_REF_MV 1
#define CONFIG_RECT_TX_EXT 0
#define CONFIG_TPL_MV 0
#define CONFIG_DUAL_FILTER 1
#define CONFIG_CONVOLVE_ROUND 0
#define CONFIG_COMPOUND_ROUND 0
#define CONFIG_EXT_TX 0
#define CONFIG_EXT_TX 1
#define CONFIG_DPCM_INTRA 0
#define CONFIG_TX64X64 0
#define CONFIG_SUB8X8_MC 0
#define CONFIG_EXT_INTRA 1
#define CONFIG_INTRA_INTERP 0
#define CONFIG_FILTER_INTRA 0
#define CONFIG_INTRA_EDGE 0
#define CONFIG_INTRABC 0
#define CONFIG_EXT_INTER 0
#define CONFIG_INTERINTRA 0
#define CONFIG_WEDGE 0
#define CONFIG_COMPOUND_SEGMENT 0
#define CONFIG_EXT_INTER 1
#define CONFIG_INTERINTRA 1
#define CONFIG_WEDGE 1
#define CONFIG_COMPOUND_SEGMENT 1
#define CONFIG_EXT_REFS 1
#define CONFIG_GLOBAL_MOTION 1
#define CONFIG_NEW_QUANT 0
#define CONFIG_SUPERTX 0
#define CONFIG_ANS 0
#define CONFIG_EC_MULTISYMBOL 1
#define CONFIG_NEW_TOKENSET 1
#define CONFIG_LOOP_RESTORATION 0
#define CONFIG_EXT_PARTITION 0
#define CONFIG_EXT_PARTITION_TYPES 0
@ -118,20 +113,18 @@
#define CONFIG_NCOBMC 0
#define CONFIG_WARPED_MOTION 1
#define CONFIG_Q_ADAPT_PROBS 0
#define CONFIG_SUBFRAME_PROB_UPDATE 0
#define CONFIG_BITSTREAM_DEBUG 0
#define CONFIG_ALT_INTRA 1
#define CONFIG_PALETTE 1
#define CONFIG_PALETTE_DELTA_ENCODING 0
#define CONFIG_DAALA_EC 1
#define CONFIG_RAWBITS 0
#define CONFIG_EC_SMALLMUL 0
#define CONFIG_EC_SMALLMUL 1
#define CONFIG_PVQ 0
#define CONFIG_CFL 0
#define CONFIG_XIPHRC 0
#define CONFIG_CB4X4 1
#define CONFIG_CHROMA_2X2 0
#define CONFIG_CHROMA_SUB8X8 0
#define CONFIG_CHROMA_SUB8X8 1
#define CONFIG_FRAME_SIZE 0
#define CONFIG_DELTA_Q 1
#define CONFIG_EXT_DELTA_Q 1
@ -139,10 +132,10 @@
#define CONFIG_FILTER_7BIT 1
#define CONFIG_PARALLEL_DEBLOCKING 0
#define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
#define CONFIG_LOOPFILTERING_ACROSS_TILES 0
#define CONFIG_LOOPFILTERING_ACROSS_TILES 1
#define CONFIG_TILE_GROUPS 1
#define CONFIG_EC_ADAPT 1
#define CONFIG_TEMPMV_SIGNALING 0
#define CONFIG_TEMPMV_SIGNALING 1
#define CONFIG_RD_DEBUG 0
#define CONFIG_REFERENCE_BUFFER 1
#define CONFIG_COEF_INTERLEAVE 0
@ -160,7 +153,9 @@
#define CONFIG_NEW_MULTISYMBOL 0
#define CONFIG_COMPOUND_SINGLEREF 0
#define CONFIG_AOM_QM 0
#define CONFIG_LOWDELAY_COMPOUND 0
#define CONFIG_ONE_SIDED_COMPOUND 1
#define CONFIG_SMOOTH_HV 0
#define CONFIG_VAR_REFS 0
#define CONFIG_ANALYZER 0
#define DECODE_WIDTH_LIMIT 8192
#define DECODE_HEIGHT_LIMIT 4608

File diff suppressed because it is too large Load diff

View file

@ -60,13 +60,7 @@ void aom_clpf_hblock_hbd_sse4_1(uint16_t *dst, const uint16_t *src, int dstride,
RTCD_EXTERN void (*aom_clpf_hblock_hbd)(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
int64_t av1_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
int64_t av1_block_error_fp_sse2(const int16_t *coeff, const int16_t *dqcoeff, int block_size);
#define av1_block_error_fp av1_block_error_fp_sse2
#define av1_block_error av1_block_error_c
void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
void av1_convolve_horiz_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
@ -80,9 +74,7 @@ int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_conf
#define av1_diamond_search_sad av1_diamond_search_sad_c
void av1_fdct8x8_quant_c(const int16_t *input, int stride, 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);
void av1_fdct8x8_quant_sse2(const int16_t *input, int stride, 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);
void av1_fdct8x8_quant_ssse3(const int16_t *input, int stride, 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);
RTCD_EXTERN void (*av1_fdct8x8_quant)(const int16_t *input, int stride, 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);
#define av1_fdct8x8_quant av1_fdct8x8_quant_c
void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
void av1_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, int tx_type);
@ -149,9 +141,174 @@ RTCD_EXTERN int (*av1_full_search_sad)(const struct macroblock *x, const struct
void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type);
#define av1_fwd_idtx av1_fwd_idtx_c
void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
#define av1_fwht4x4 av1_fwht4x4_c
int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
int64_t av1_highbd_block_error_sse2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
#define av1_highbd_block_error av1_highbd_block_error_sse2
void av1_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
void av1_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve8 av1_highbd_convolve8_sse2
void av1_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
void av1_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve8_avg av1_highbd_convolve8_avg_sse2
void av1_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
void av1_highbd_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve8_avg_horiz av1_highbd_convolve8_avg_horiz_sse2
void av1_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
void av1_highbd_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve8_avg_vert av1_highbd_convolve8_avg_vert_sse2
void av1_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
void av1_highbd_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve8_horiz av1_highbd_convolve8_horiz_sse2
void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2
void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
void av1_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
#define av1_highbd_convolve_copy av1_highbd_convolve_copy_c
void av1_highbd_convolve_horiz_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
void av1_highbd_convolve_horiz_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
RTCD_EXTERN void (*av1_highbd_convolve_horiz)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
void av1_highbd_convolve_init_c(void);
void av1_highbd_convolve_init_sse4_1(void);
RTCD_EXTERN void (*av1_highbd_convolve_init)(void);
void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
void av1_highbd_convolve_vert_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
RTCD_EXTERN void (*av1_highbd_convolve_vert)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
void av1_highbd_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht16x16 av1_highbd_fht16x16_c
void av1_highbd_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht16x32 av1_highbd_fht16x32_c
void av1_highbd_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht16x4 av1_highbd_fht16x4_c
void av1_highbd_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht16x8 av1_highbd_fht16x8_c
void av1_highbd_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht32x16 av1_highbd_fht32x16_c
void av1_highbd_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht32x32 av1_highbd_fht32x32_c
void av1_highbd_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht32x8 av1_highbd_fht32x8_c
void av1_highbd_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht4x16 av1_highbd_fht4x16_c
void av1_highbd_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
void av1_highbd_fht4x4_sse4_1(const int16_t *input, tran_low_t *output, int stride, int tx_type);
RTCD_EXTERN void (*av1_highbd_fht4x4)(const int16_t *input, tran_low_t *output, int stride, int tx_type);
void av1_highbd_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht4x8 av1_highbd_fht4x8_c
void av1_highbd_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht8x16 av1_highbd_fht8x16_c
void av1_highbd_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht8x32 av1_highbd_fht8x32_c
void av1_highbd_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht8x4 av1_highbd_fht8x4_c
void av1_highbd_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
#define av1_highbd_fht8x8 av1_highbd_fht8x8_c
void av1_highbd_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
#define av1_highbd_fwht4x4 av1_highbd_fwht4x4_c
void av1_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type, int bd);
#define av1_highbd_iht16x16_256_add av1_highbd_iht16x16_256_add_c
void av1_highbd_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht16x32_512_add av1_highbd_iht16x32_512_add_c
void av1_highbd_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht16x4_64_add av1_highbd_iht16x4_64_add_c
void av1_highbd_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht16x8_128_add av1_highbd_iht16x8_128_add_c
void av1_highbd_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht32x16_512_add av1_highbd_iht32x16_512_add_c
void av1_highbd_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht32x8_256_add av1_highbd_iht32x8_256_add_c
void av1_highbd_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht4x16_64_add av1_highbd_iht4x16_64_add_c
void av1_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht4x4_16_add av1_highbd_iht4x4_16_add_c
void av1_highbd_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht4x8_32_add av1_highbd_iht4x8_32_add_c
void av1_highbd_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht8x16_128_add av1_highbd_iht8x16_128_add_c
void av1_highbd_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht8x32_256_add av1_highbd_iht8x32_256_add_c
void av1_highbd_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht8x4_32_add av1_highbd_iht8x4_32_add_c
void av1_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
#define av1_highbd_iht8x8_64_add av1_highbd_iht8x8_64_add_c
void av1_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 log_scale);
#define av1_highbd_quantize_b av1_highbd_quantize_b_c
void av1_highbd_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 log_scale);
void av1_highbd_quantize_fp_sse4_1(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 log_scale);
RTCD_EXTERN void (*av1_highbd_quantize_fp)(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 log_scale);
void av1_highbd_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);
#define av1_highbd_temporal_filter_apply av1_highbd_temporal_filter_apply_c
void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
void av1_highbd_warp_affine_ssse3(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
void av1_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
void av1_iht16x16_256_add_avx2(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
@ -204,6 +361,25 @@ void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride
void av1_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
#define av1_iht8x8_64_add av1_iht8x8_64_add_sse2
void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_inv_txfm2d_add_16x16_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_inv_txfm2d_add_16x16)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_inv_txfm2d_add_32x32_avx2(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_inv_txfm2d_add_32x32)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
void av1_lowbd_convolve_init_c(void);
void av1_lowbd_convolve_init_ssse3(void);
RTCD_EXTERN void (*av1_lowbd_convolve_init)(void);
@ -213,20 +389,35 @@ void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_b
void av1_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);
void av1_quantize_fp_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);
void av1_quantize_fp_ssse3(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);
RTCD_EXTERN void (*av1_quantize_fp)(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);
#define av1_quantize_fp av1_quantize_fp_sse2
void av1_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);
void av1_quantize_fp_32x32_ssse3(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);
RTCD_EXTERN void (*av1_quantize_fp_32x32)(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);
#define av1_quantize_fp_32x32 av1_quantize_fp_32x32_c
void av1_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);
void av1_temporal_filter_apply_sse2(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);
#define av1_temporal_filter_apply av1_temporal_filter_apply_sse2
void av1_warp_affine_c(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
void av1_warp_affine_sse2(int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
#define av1_warp_affine av1_warp_affine_sse2
void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
void av1_warp_affine_sse2(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
void av1_warp_affine_ssse3(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a, const int16_t *b, int N);
#define av1_wedge_compute_delta_squares av1_wedge_compute_delta_squares_sse2
int av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
int av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
#define av1_wedge_sign_from_residuals av1_wedge_sign_from_residuals_sse2
uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
#define av1_wedge_sse_from_residuals av1_wedge_sse_from_residuals_sse2
double compute_cross_correlation_c(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
double compute_cross_correlation_sse4_1(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
RTCD_EXTERN double (*compute_cross_correlation)(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
void copy_4x4_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
void copy_4x4_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
@ -304,14 +495,10 @@ static void setup_rtcd_internal(void)
aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_sse2;
if (flags & HAS_SSSE3) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_ssse3;
if (flags & HAS_SSE4_1) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_sse4_1;
av1_block_error = av1_block_error_sse2;
if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
av1_convolve_horiz = av1_convolve_horiz_c;
if (flags & HAS_SSSE3) av1_convolve_horiz = av1_convolve_horiz_ssse3;
av1_convolve_vert = av1_convolve_vert_c;
if (flags & HAS_SSSE3) av1_convolve_vert = av1_convolve_vert_ssse3;
av1_fdct8x8_quant = av1_fdct8x8_quant_sse2;
if (flags & HAS_SSSE3) av1_fdct8x8_quant = av1_fdct8x8_quant_ssse3;
av1_fht16x16 = av1_fht16x16_sse2;
if (flags & HAS_AVX2) av1_fht16x16 = av1_fht16x16_avx2;
av1_fht32x32 = av1_fht32x32_sse2;
@ -319,14 +506,44 @@ static void setup_rtcd_internal(void)
av1_full_search_sad = av1_full_search_sad_c;
if (flags & HAS_SSE3) av1_full_search_sad = av1_full_search_sadx3;
if (flags & HAS_SSE4_1) av1_full_search_sad = av1_full_search_sadx8;
av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_c;
if (flags & HAS_SSE4_1) av1_fwd_txfm2d_16x16 = av1_fwd_txfm2d_16x16_sse4_1;
av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_c;
if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
av1_highbd_convolve_horiz = av1_highbd_convolve_horiz_c;
if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz = av1_highbd_convolve_horiz_sse4_1;
av1_highbd_convolve_init = av1_highbd_convolve_init_c;
if (flags & HAS_SSE4_1) av1_highbd_convolve_init = av1_highbd_convolve_init_sse4_1;
av1_highbd_convolve_vert = av1_highbd_convolve_vert_c;
if (flags & HAS_SSE4_1) av1_highbd_convolve_vert = av1_highbd_convolve_vert_sse4_1;
av1_highbd_fht4x4 = av1_highbd_fht4x4_c;
if (flags & HAS_SSE4_1) av1_highbd_fht4x4 = av1_highbd_fht4x4_sse4_1;
av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
if (flags & HAS_SSE4_1) av1_highbd_quantize_fp = av1_highbd_quantize_fp_sse4_1;
av1_highbd_warp_affine = av1_highbd_warp_affine_c;
if (flags & HAS_SSSE3) av1_highbd_warp_affine = av1_highbd_warp_affine_ssse3;
av1_iht16x16_256_add = av1_iht16x16_256_add_sse2;
if (flags & HAS_AVX2) av1_iht16x16_256_add = av1_iht16x16_256_add_avx2;
av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_c;
if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_sse4_1;
av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_c;
if (flags & HAS_AVX2) av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_avx2;
av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
av1_lowbd_convolve_init = av1_lowbd_convolve_init_c;
if (flags & HAS_SSSE3) av1_lowbd_convolve_init = av1_lowbd_convolve_init_ssse3;
av1_quantize_fp = av1_quantize_fp_sse2;
if (flags & HAS_SSSE3) av1_quantize_fp = av1_quantize_fp_ssse3;
av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
if (flags & HAS_SSSE3) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_ssse3;
av1_warp_affine = av1_warp_affine_sse2;
if (flags & HAS_SSSE3) av1_warp_affine = av1_warp_affine_ssse3;
compute_cross_correlation = compute_cross_correlation_c;
if (flags & HAS_SSE4_1) compute_cross_correlation = compute_cross_correlation_sse4_1;
copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse2;
if (flags & HAS_SSSE3) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_ssse3;
if (flags & HAS_SSE4_1) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse4_1;