void ff_hevcdsp_init_x86(HEVCDSPContext *c, const int bit_depth)
{
int mm_flags = av_get_cpu_flags();
if (bit_depth == 8) {
if (EXTERNAL_SSE2(mm_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
}
if (EXTERNAL_SSSE3(mm_flags) && ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
}
if (EXTERNAL_SSE4(mm_flags) && ARCH_X86_64) {
EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 8, sse4);
}
} else if (bit_depth == 10) {
if (EXTERNAL_SSE2(mm_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
}
if (EXTERNAL_SSSE3(mm_flags) && ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
}
if (EXTERNAL_SSE4(mm_flags) && ARCH_X86_64) {
EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 10, sse4);
}
}
}
void ff_hevcpred_init_x86(HEVCPredContext *c, const int bit_depth)
{
int mm_flags = av_get_cpu_flags();
if (bit_depth == 8) {
if (EXTERNAL_MMX(mm_flags)) {
if (EXTERNAL_MMXEXT(mm_flags)) {
if (EXTERNAL_SSE2(mm_flags)) {
}
if (EXTERNAL_SSSE3(mm_flags)) {
}
if (EXTERNAL_SSE4(mm_flags)) {
c->pred_planar[0]= pred_planar_0_8_sse;
c->pred_planar[1]= pred_planar_1_8_sse;
c->pred_planar[2]= pred_planar_2_8_sse;
c->pred_planar[3]= pred_planar_3_8_sse;
// c->pred_angular[0]= pred_angular_0_8_sse; //removed because too little data = bad performance
c->pred_angular[1]= pred_angular_1_8_sse;
c->pred_angular[2]= pred_angular_2_8_sse;
c->pred_angular[3]= pred_angular_3_8_sse;
}
if (EXTERNAL_AVX(mm_flags)) {
}
}
}
} else if (bit_depth == 10) {
if (EXTERNAL_MMX(mm_flags)) {
if (EXTERNAL_MMXEXT(mm_flags)) {
if (EXTERNAL_SSE2(mm_flags)) {
}
if (EXTERNAL_SSE4(mm_flags)) {
}
if (EXTERNAL_AVX(mm_flags)) {
}
}
}
}
}
void ff_llviddsp_init_x86(LLVidDSPContext *c, AVCodecContext *avctx)
{
int cpu_flags = av_get_cpu_flags();
const AVPixFmtDescriptor *pix_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
if (EXTERNAL_MMX(cpu_flags)) {
c->add_int16 = ff_add_int16_mmx;
c->diff_int16 = ff_diff_int16_mmx;
}
if (EXTERNAL_MMXEXT(cpu_flags) && pix_desc->comp[0].depth_minus1<15) {
c->add_hfyu_median_pred_int16 = ff_add_hfyu_median_pred_int16_mmxext;
c->sub_hfyu_median_pred_int16 = ff_sub_hfyu_median_pred_int16_mmxext;
}
if (EXTERNAL_SSE2(cpu_flags)) {
c->add_int16 = ff_add_int16_sse2;
c->diff_int16 = ff_diff_int16_sse2;
}
if (EXTERNAL_SSSE3(cpu_flags)) {
c->add_hfyu_left_pred_int16 = ff_add_hfyu_left_pred_int16_ssse3;
}
if (EXTERNAL_SSE4(cpu_flags)) {
c->add_hfyu_left_pred_int16 = ff_add_hfyu_left_pred_int16_sse4;
}
}
av_cold void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx,
unsigned high_bit_depth)
{
int cpu_flags = av_get_cpu_flags();
if (X86_MMX(cpu_flags))
dsputil_init_mmx(c, avctx, cpu_flags, high_bit_depth);
if (X86_MMXEXT(cpu_flags))
dsputil_init_mmxext(c, avctx, cpu_flags, high_bit_depth);
if (X86_SSE(cpu_flags))
dsputil_init_sse(c, avctx, cpu_flags, high_bit_depth);
if (X86_SSE2(cpu_flags))
dsputil_init_sse2(c, avctx, cpu_flags, high_bit_depth);
if (EXTERNAL_SSSE3(cpu_flags))
dsputil_init_ssse3(c, avctx, cpu_flags, high_bit_depth);
if (EXTERNAL_SSE4(cpu_flags))
dsputil_init_sse4(c, avctx, cpu_flags, high_bit_depth);
if (CONFIG_ENCODERS)
ff_dsputilenc_init_mmx(c, avctx, high_bit_depth);
}
av_cold void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx)
{
int cpu_flags = av_get_cpu_flags();
#if HAVE_7REGS && HAVE_INLINE_ASM
if (cpu_flags & AV_CPU_FLAG_CMOV)
c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_cmov;
#endif
if (X86_MMX(cpu_flags))
dsputil_init_mmx(c, avctx, cpu_flags);
if (X86_MMXEXT(cpu_flags))
dsputil_init_mmxext(c, avctx, cpu_flags);
if (X86_SSE(cpu_flags))
dsputil_init_sse(c, avctx, cpu_flags);
if (X86_SSE2(cpu_flags))
dsputil_init_sse2(c, avctx, cpu_flags);
if (EXTERNAL_SSSE3(cpu_flags))
dsputil_init_ssse3(c, avctx, cpu_flags);
if (EXTERNAL_SSE4(cpu_flags))
dsputil_init_sse4(c, avctx, cpu_flags);
if (CONFIG_ENCODERS)
ff_dsputilenc_init_mmx(c, avctx);
}
av_cold void ff_rv34dsp_init_x86(RV34DSPContext* c, DSPContext *dsp)
{
int mm_flags = av_get_cpu_flags();
if (EXTERNAL_MMX(mm_flags))
c->rv34_idct_dc_add = ff_rv34_idct_dc_add_mmx;
if (EXTERNAL_MMXEXT(mm_flags)) {
c->rv34_inv_transform_dc = ff_rv34_idct_dc_noround_mmx2;
c->rv34_idct_add = ff_rv34_idct_add_mmx2;
}
if (EXTERNAL_SSE4(mm_flags))
c->rv34_idct_dc_add = ff_rv34_idct_dc_add_sse4;
}
av_cold void ff_vc1dsp_init_x86(VC1DSPContext *dsp)
{
int cpu_flags = av_get_cpu_flags();
#if HAVE_6REGS && HAVE_INLINE_ASM
if (INLINE_MMX(cpu_flags))
ff_vc1dsp_init_mmx(dsp);
#endif
#if HAVE_6REGS && HAVE_MMX_EXTERNAL && HAVE_INLINE_ASM
if (INLINE_MMXEXT(cpu_flags))
ff_vc1dsp_init_mmxext(dsp);
#endif
#define ASSIGN_LF(EXT) \
dsp->vc1_v_loop_filter4 = ff_vc1_v_loop_filter4_ ## EXT; \
dsp->vc1_h_loop_filter4 = ff_vc1_h_loop_filter4_ ## EXT; \
dsp->vc1_v_loop_filter8 = ff_vc1_v_loop_filter8_ ## EXT; \
dsp->vc1_h_loop_filter8 = ff_vc1_h_loop_filter8_ ## EXT; \
dsp->vc1_v_loop_filter16 = vc1_v_loop_filter16_ ## EXT; \
dsp->vc1_h_loop_filter16 = vc1_h_loop_filter16_ ## EXT
#if HAVE_YASM
if (EXTERNAL_MMX(cpu_flags)) {
dsp->put_no_rnd_vc1_chroma_pixels_tab[0] = ff_put_vc1_chroma_mc8_nornd_mmx;
}
if (EXTERNAL_AMD3DNOW(cpu_flags)) {
dsp->avg_no_rnd_vc1_chroma_pixels_tab[0] = ff_avg_vc1_chroma_mc8_nornd_3dnow;
}
if (EXTERNAL_MMXEXT(cpu_flags)) {
ASSIGN_LF(mmxext);
dsp->avg_no_rnd_vc1_chroma_pixels_tab[0] = ff_avg_vc1_chroma_mc8_nornd_mmxext;
dsp->avg_vc1_mspel_pixels_tab[1][0] = avg_vc1_mspel_mc00_mmxext;
}
if (EXTERNAL_SSE2(cpu_flags)) {
dsp->vc1_v_loop_filter8 = ff_vc1_v_loop_filter8_sse2;
dsp->vc1_h_loop_filter8 = ff_vc1_h_loop_filter8_sse2;
dsp->vc1_v_loop_filter16 = vc1_v_loop_filter16_sse2;
dsp->vc1_h_loop_filter16 = vc1_h_loop_filter16_sse2;
dsp->avg_vc1_mspel_pixels_tab[0][0] = avg_vc1_mspel_mc00_16_sse2;
}
if (EXTERNAL_SSSE3(cpu_flags)) {
ASSIGN_LF(ssse3);
dsp->put_no_rnd_vc1_chroma_pixels_tab[0] = ff_put_vc1_chroma_mc8_nornd_ssse3;
dsp->avg_no_rnd_vc1_chroma_pixels_tab[0] = ff_avg_vc1_chroma_mc8_nornd_ssse3;
}
if (EXTERNAL_SSE4(cpu_flags)) {
dsp->vc1_h_loop_filter8 = ff_vc1_h_loop_filter8_sse4;
dsp->vc1_h_loop_filter16 = vc1_h_loop_filter16_sse4;
}
#endif /* HAVE_YASM */
}
av_cold void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx,
unsigned high_bit_depth)
{
int cpu_flags = av_get_cpu_flags();
#if HAVE_7REGS && HAVE_INLINE_ASM
if (HAVE_MMX && cpu_flags & AV_CPU_FLAG_CMOV)
c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_cmov;
#endif
if (X86_MMX(cpu_flags)) {
#if HAVE_INLINE_ASM
const int idct_algo = avctx->idct_algo;
if (avctx->lowres == 0 && !high_bit_depth) {
if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
c->idct_put = ff_simple_idct_put_mmx;
c->idct_add = ff_simple_idct_add_mmx;
c->idct = ff_simple_idct_mmx;
c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
} else if (idct_algo == FF_IDCT_XVIDMMX) {
c->idct_put = ff_idct_xvid_mmx_put;
c->idct_add = ff_idct_xvid_mmx_add;
c->idct = ff_idct_xvid_mmx;
}
}
#endif /* HAVE_INLINE_ASM */
dsputil_init_mmx(c, avctx, cpu_flags, high_bit_depth);
}
if (X86_MMXEXT(cpu_flags))
dsputil_init_mmxext(c, avctx, cpu_flags, high_bit_depth);
if (X86_SSE(cpu_flags))
dsputil_init_sse(c, avctx, cpu_flags, high_bit_depth);
if (X86_SSE2(cpu_flags))
dsputil_init_sse2(c, avctx, cpu_flags, high_bit_depth);
if (EXTERNAL_SSSE3(cpu_flags))
dsputil_init_ssse3(c, avctx, cpu_flags, high_bit_depth);
if (EXTERNAL_SSE4(cpu_flags))
dsputil_init_sse4(c, avctx, cpu_flags, high_bit_depth);
if (CONFIG_ENCODERS)
ff_dsputilenc_init_mmx(c, avctx, high_bit_depth);
}
void ff_framerate_init_x86(FrameRateContext *s)
{
int cpu_flags = av_get_cpu_flags();
if (s->bitdepth == 8) {
if (EXTERNAL_AVX2_FAST(cpu_flags))
s->blend = ff_blend_frames_avx2;
else if (EXTERNAL_SSSE3(cpu_flags))
s->blend = ff_blend_frames_ssse3;
} else {
if (EXTERNAL_AVX2_FAST(cpu_flags))
s->blend = ff_blend_frames16_avx2;
else if (EXTERNAL_SSE4(cpu_flags))
s->blend = ff_blend_frames16_sse4;
}
}
int ff_image_copy_plane_uc_from_x86(uint8_t *dst, ptrdiff_t dst_linesize,
const uint8_t *src, ptrdiff_t src_linesize,
ptrdiff_t bytewidth, int height)
{
int cpu_flags = av_get_cpu_flags();
ptrdiff_t bw_aligned = FFALIGN(bytewidth, 64);
if (EXTERNAL_SSE4(cpu_flags) &&
bw_aligned <= dst_linesize && bw_aligned <= src_linesize)
ff_image_copy_plane_uc_from_sse4(dst, dst_linesize, src, src_linesize,
bw_aligned, height);
else
return AVERROR(ENOSYS);
return 0;
}
void ff_diracdsp_init_x86(DiracDSPContext* c)
{
int mm_flags = av_get_cpu_flags();
if (EXTERNAL_MMX(mm_flags)) {
c->add_dirac_obmc[0] = ff_add_dirac_obmc8_mmx;
#if !ARCH_X86_64
c->add_dirac_obmc[1] = ff_add_dirac_obmc16_mmx;
c->add_dirac_obmc[2] = ff_add_dirac_obmc32_mmx;
c->dirac_hpel_filter = dirac_hpel_filter_mmx;
c->add_rect_clamped = ff_add_rect_clamped_mmx;
c->put_signed_rect_clamped[0] = (void *)ff_put_signed_rect_clamped_mmx;
#endif
PIXFUNC(put, 0, mmx);
PIXFUNC(avg, 0, mmx);
}
if (EXTERNAL_MMXEXT(mm_flags)) {
PIXFUNC(avg, 0, mmxext);
}
if (EXTERNAL_SSE2(mm_flags)) {
c->dirac_hpel_filter = dirac_hpel_filter_sse2;
c->add_rect_clamped = ff_add_rect_clamped_sse2;
c->put_signed_rect_clamped[0] = (void *)ff_put_signed_rect_clamped_sse2;
c->add_dirac_obmc[1] = ff_add_dirac_obmc16_sse2;
c->add_dirac_obmc[2] = ff_add_dirac_obmc32_sse2;
c->put_dirac_pixels_tab[1][0] = ff_put_dirac_pixels16_sse2;
c->avg_dirac_pixels_tab[1][0] = ff_avg_dirac_pixels16_sse2;
c->put_dirac_pixels_tab[2][0] = ff_put_dirac_pixels32_sse2;
c->avg_dirac_pixels_tab[2][0] = ff_avg_dirac_pixels32_sse2;
}
if (EXTERNAL_SSE4(mm_flags)) {
c->dequant_subband[1] = ff_dequant_subband_32_sse4;
#if ARCH_X86_64
c->put_signed_rect_clamped[1] = ff_put_signed_rect_clamped_10_sse4;
#endif
}
}
av_cold void ff_dcadsp_init_x86(DCADSPContext *s)
{
int cpu_flags = av_get_cpu_flags();
if (EXTERNAL_SSE(cpu_flags)) {
#if ARCH_X86_32
s->decode_hf = ff_decode_hf_sse;
#endif
s->lfe_fir[0] = ff_dca_lfe_fir0_sse;
s->lfe_fir[1] = ff_dca_lfe_fir1_sse;
}
if (EXTERNAL_SSE2(cpu_flags)) {
s->decode_hf = ff_decode_hf_sse2;
}
if (EXTERNAL_SSE4(cpu_flags)) {
s->decode_hf = ff_decode_hf_sse4;
}
}
av_cold void ff_psdsp_init_x86(PSDSPContext *s)
{
int cpu_flags = av_get_cpu_flags();
if (EXTERNAL_SSE(cpu_flags)) {
s->add_squares = ff_ps_add_squares_sse;
s->mul_pair_single = ff_ps_mul_pair_single_sse;
s->hybrid_analysis_ileave = ff_ps_hybrid_analysis_ileave_sse;
s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse;
s->hybrid_analysis = ff_ps_hybrid_analysis_sse;
}
if (EXTERNAL_SSE3(cpu_flags)) {
s->add_squares = ff_ps_add_squares_sse3;
s->stereo_interpolate[0] = ff_ps_stereo_interpolate_sse3;
s->stereo_interpolate[1] = ff_ps_stereo_interpolate_ipdopd_sse3;
s->hybrid_analysis = ff_ps_hybrid_analysis_sse3;
}
if (EXTERNAL_SSE4(cpu_flags)) {
s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse4;
}
}
av_cold void ff_proresdsp_init_x86(ProresDSPContext *dsp)
{
#if ARCH_X86_64
int cpu_flags = av_get_cpu_flags();
if (EXTERNAL_SSE2(cpu_flags)) {
dsp->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
dsp->idct_put = ff_prores_idct_put_10_sse2;
}
if (EXTERNAL_SSE4(cpu_flags)) {
dsp->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
dsp->idct_put = ff_prores_idct_put_10_sse4;
}
if (EXTERNAL_AVX(cpu_flags)) {
dsp->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
dsp->idct_put = ff_prores_idct_put_10_avx;
}
#endif /* ARCH_X86_64 */
}
av_cold void ff_yadif_init_x86(YADIFContext *yadif)
{
#if HAVE_YASM
int cpu_flags = av_get_cpu_flags();
int bit_depth = (!yadif->csp) ? 8
: yadif->csp->comp[0].depth_minus1 + 1;
if (bit_depth >= 15) {
#if ARCH_X86_32
if (EXTERNAL_MMXEXT(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_16bit_mmxext;
#endif /* ARCH_X86_32 */
if (EXTERNAL_SSE2(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_16bit_sse2;
if (EXTERNAL_SSSE3(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_16bit_ssse3;
if (EXTERNAL_SSE4(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_16bit_sse4;
} else if ( bit_depth >= 9 && bit_depth <= 14) {
#if ARCH_X86_32
if (EXTERNAL_MMXEXT(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_10bit_mmxext;
#endif /* ARCH_X86_32 */
if (EXTERNAL_SSE2(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_10bit_sse2;
if (EXTERNAL_SSSE3(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_10bit_ssse3;
} else {
#if ARCH_X86_32
if (EXTERNAL_MMXEXT(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_mmxext;
#endif /* ARCH_X86_32 */
if (EXTERNAL_SSE2(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_sse2;
if (EXTERNAL_SSSE3(cpu_flags))
yadif->filter_line = ff_yadif_filter_line_ssse3;
}
#endif /* HAVE_YASM */
}
void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
{
int cpu_flags = av_get_cpu_flags();
#define SET_LUMA_FUNCS(tabname, funcname, depth, cf) \
c->tabname[0] = funcname ## _4_ ## depth ## _ ## cf; \
c->tabname[1] = funcname ## _8_ ## depth ## _ ## cf; \
c->tabname[2] = funcname ## _12_ ## depth ## _ ## cf; \
c->tabname[3] = funcname ## _16_ ## depth ## _ ## cf; \
c->tabname[4] = funcname ## _24_ ## depth ## _ ## cf; \
c->tabname[5] = funcname ## _32_ ## depth ## _ ## cf; \
c->tabname[6] = funcname ## _48_ ## depth ## _ ## cf; \
c->tabname[7] = funcname ## _64_ ## depth ## _ ## cf;
#define SET_CHROMA_FUNCS(tabname, funcname, depth, cf) \
c->tabname[1] = funcname ## _4_ ## depth ## _ ## cf; \
c->tabname[3] = funcname ## _8_ ## depth ## _ ## cf; \
c->tabname[4] = funcname ## _12_ ## depth ## _ ## cf; \
c->tabname[5] = funcname ## _16_ ## depth ## _ ## cf; \
c->tabname[6] = funcname ## _24_ ## depth ## _ ## cf; \
c->tabname[7] = funcname ## _32_ ## depth ## _ ## cf;
#define SET_QPEL_FUNCS(v, h, depth, cf, name) SET_LUMA_FUNCS (put_hevc_qpel[v][h], name, depth, cf)
#define SET_EPEL_FUNCS(v, h, depth, cf, name) SET_CHROMA_FUNCS(put_hevc_epel[v][h], name, depth, cf)
if (bit_depth == 8) {
if (EXTERNAL_MMXEXT(cpu_flags)) {
c->idct_dc[0] = ff_hevc_idct_4x4_dc_8_mmxext;
c->idct_dc[1] = ff_hevc_idct_8x8_dc_8_mmxext;
}
if (EXTERNAL_SSE2(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
c->idct_dc[1] = ff_hevc_idct_8x8_dc_8_sse2;
c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_sse2;
c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_sse2;
c->idct[0] = ff_hevc_idct_4x4_8_sse2;
c->idct[1] = ff_hevc_idct_8x8_8_sse2;
SET_QPEL_FUNCS(0, 0, 8, sse2, ff_hevc_get_pixels);
SET_EPEL_FUNCS(0, 0, 8, sse2, ff_hevc_get_pixels);
SET_LUMA_FUNCS(put_unweighted_pred, ff_hevc_put_unweighted_pred, 8, sse2);
SET_LUMA_FUNCS(put_unweighted_pred_avg, ff_hevc_put_unweighted_pred_avg, 8, sse2);
SET_CHROMA_FUNCS(put_unweighted_pred_chroma, ff_hevc_put_unweighted_pred, 8, sse2);
SET_CHROMA_FUNCS(put_unweighted_pred_avg_chroma, ff_hevc_put_unweighted_pred_avg, 8, sse2);
}
if (EXTERNAL_SSSE3(cpu_flags)) {
SET_QPEL_FUNCS(0, 1, 8, ssse3, ff_hevc_qpel_h);
SET_QPEL_FUNCS(1, 0, 8, ssse3, ff_hevc_qpel_v);
SET_EPEL_FUNCS(0, 1, 8, ssse3, ff_hevc_epel_h);
SET_EPEL_FUNCS(1, 0, 8, ssse3, ff_hevc_epel_v);
}
if (EXTERNAL_AVX(cpu_flags)) {
c->idct[0] = ff_hevc_idct_4x4_8_avx;
c->idct[1] = ff_hevc_idct_8x8_8_avx;
}
} else if (bit_depth == 10) {
if (EXTERNAL_MMXEXT(cpu_flags)) {
c->idct_dc[0] = ff_hevc_idct_4x4_dc_10_mmxext;
c->idct_dc[1] = ff_hevc_idct_8x8_dc_10_mmxext;
}
if (EXTERNAL_SSE2(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
c->idct_dc[1] = ff_hevc_idct_8x8_dc_10_sse2;
c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_sse2;
c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_sse2;
c->idct[0] = ff_hevc_idct_4x4_10_sse2;
c->idct[1] = ff_hevc_idct_8x8_10_sse2;
SET_QPEL_FUNCS(0, 0, 10, sse2, ff_hevc_get_pixels);
SET_EPEL_FUNCS(0, 0, 10, sse2, ff_hevc_get_pixels);
SET_LUMA_FUNCS(put_unweighted_pred, ff_hevc_put_unweighted_pred, 10, sse2);
SET_LUMA_FUNCS(put_unweighted_pred_avg, ff_hevc_put_unweighted_pred_avg, 10, sse2);
SET_CHROMA_FUNCS(put_unweighted_pred_chroma, ff_hevc_put_unweighted_pred, 10, sse2);
SET_CHROMA_FUNCS(put_unweighted_pred_avg_chroma, ff_hevc_put_unweighted_pred_avg, 10, sse2);
}
if (EXTERNAL_AVX(cpu_flags)) {
c->idct[0] = ff_hevc_idct_4x4_10_avx;
c->idct[1] = ff_hevc_idct_8x8_10_avx;
}
}
#if ARCH_X86_64
if (bit_depth == 8) {
if (EXTERNAL_SSE2(cpu_flags)) {
c->idct[2] = ff_hevc_idct_16x16_8_sse2;
c->idct[3] = ff_hevc_idct_32x32_8_sse2;
}
if (EXTERNAL_SSSE3(cpu_flags)) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
}
if (EXTERNAL_SSE4(cpu_flags)) {
SET_LUMA_FUNCS(weighted_pred, ff_hevc_put_weighted_pred, 8, sse4);
SET_CHROMA_FUNCS(weighted_pred_chroma, ff_hevc_put_weighted_pred, 8, sse4);
SET_LUMA_FUNCS(weighted_pred_avg, ff_hevc_put_weighted_pred_avg, 8, sse4);
SET_CHROMA_FUNCS(weighted_pred_avg_chroma, ff_hevc_put_weighted_pred_avg, 8, sse4);
}
if (EXTERNAL_AVX(cpu_flags)) {
#if HAVE_AVX_EXTERNAL
SET_QPEL_FUNCS(1, 1, 8, avx, hevc_qpel_hv);
SET_EPEL_FUNCS(1, 1, 8, avx, hevc_epel_hv);
#endif /* HAVE_AVX_EXTERNAL */
c->idct[2] = ff_hevc_idct_16x16_8_avx;
c->idct[3] = ff_hevc_idct_32x32_8_avx;
}
if (EXTERNAL_AVX2(cpu_flags)) {
c->idct_dc[2] = ff_hevc_idct_16x16_dc_8_avx2;
c->idct_dc[3] = ff_hevc_idct_32x32_dc_8_avx2;
}
} else if (bit_depth == 10) {
if (EXTERNAL_SSE2(cpu_flags)) {
c->idct[2] = ff_hevc_idct_16x16_10_sse2;
c->idct[3] = ff_hevc_idct_32x32_10_sse2;
}
if (EXTERNAL_SSSE3(cpu_flags)) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
}
if (EXTERNAL_SSE4(cpu_flags)) {
SET_LUMA_FUNCS(weighted_pred, ff_hevc_put_weighted_pred, 10, sse4);
SET_CHROMA_FUNCS(weighted_pred_chroma, ff_hevc_put_weighted_pred, 10, sse4);
SET_LUMA_FUNCS(weighted_pred_avg, ff_hevc_put_weighted_pred_avg, 10, sse4);
SET_CHROMA_FUNCS(weighted_pred_avg_chroma, ff_hevc_put_weighted_pred_avg, 10, sse4);
}
if (EXTERNAL_AVX(cpu_flags)) {
#if HAVE_AVX_EXTERNAL
SET_QPEL_FUNCS(0, 1, 10, avx, ff_hevc_qpel_h);
SET_QPEL_FUNCS(1, 0, 10, avx, ff_hevc_qpel_v);
SET_QPEL_FUNCS(1, 1, 10, avx, hevc_qpel_hv);
SET_EPEL_FUNCS(0, 1, 10, avx, ff_hevc_epel_h);
SET_EPEL_FUNCS(1, 0, 10, avx, ff_hevc_epel_v);
SET_EPEL_FUNCS(1, 1, 10, avx, hevc_epel_hv);
#endif /* HAVE_AVX_EXTERNAL */
c->idct[2] = ff_hevc_idct_16x16_10_avx;
c->idct[3] = ff_hevc_idct_32x32_10_avx;
}
if (EXTERNAL_AVX2(cpu_flags)) {
c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_avx2;
c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_avx2;
}
}
#endif /* ARCH_X86_64 */
}
av_cold void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth,
const int chroma_format_idc)
{
#if HAVE_YASM
int mm_flags = av_get_cpu_flags();
if (chroma_format_idc == 1 && EXTERNAL_MMXEXT(mm_flags))
c->h264_loop_filter_strength = ff_h264_loop_filter_strength_mmxext;
if (bit_depth == 8) {
if (EXTERNAL_MMX(mm_flags)) {
c->h264_idct_dc_add =
c->h264_idct_add = ff_h264_idct_add_8_mmx;
c->h264_idct8_dc_add =
c->h264_idct8_add = ff_h264_idct8_add_8_mmx;
c->h264_idct_add16 = ff_h264_idct_add16_8_mmx;
c->h264_idct8_add4 = ff_h264_idct8_add4_8_mmx;
if (chroma_format_idc == 1)
c->h264_idct_add8 = ff_h264_idct_add8_8_mmx;
c->h264_idct_add16intra = ff_h264_idct_add16intra_8_mmx;
if (mm_flags & AV_CPU_FLAG_CMOV)
c->h264_luma_dc_dequant_idct = ff_h264_luma_dc_dequant_idct_mmx;
if (EXTERNAL_MMXEXT(mm_flags)) {
c->h264_idct_dc_add = ff_h264_idct_dc_add_8_mmxext;
c->h264_idct8_dc_add = ff_h264_idct8_dc_add_8_mmxext;
c->h264_idct_add16 = ff_h264_idct_add16_8_mmxext;
c->h264_idct8_add4 = ff_h264_idct8_add4_8_mmxext;
if (chroma_format_idc == 1)
c->h264_idct_add8 = ff_h264_idct_add8_8_mmxext;
c->h264_idct_add16intra = ff_h264_idct_add16intra_8_mmxext;
c->h264_v_loop_filter_chroma = ff_deblock_v_chroma_8_mmxext;
c->h264_v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_8_mmxext;
if (chroma_format_idc == 1) {
c->h264_h_loop_filter_chroma = ff_deblock_h_chroma_8_mmxext;
c->h264_h_loop_filter_chroma_intra = ff_deblock_h_chroma_intra_8_mmxext;
}
#if ARCH_X86_32 && HAVE_MMXEXT_EXTERNAL
c->h264_v_loop_filter_luma = ff_deblock_v_luma_8_mmxext;
c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_mmxext;
c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_mmxext;
c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_mmxext;
#endif /* ARCH_X86_32 && HAVE_MMXEXT_EXTERNAL */
c->weight_h264_pixels_tab[0] = ff_h264_weight_16_mmxext;
c->weight_h264_pixels_tab[1] = ff_h264_weight_8_mmxext;
c->weight_h264_pixels_tab[2] = ff_h264_weight_4_mmxext;
c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_mmxext;
c->biweight_h264_pixels_tab[1] = ff_h264_biweight_8_mmxext;
c->biweight_h264_pixels_tab[2] = ff_h264_biweight_4_mmxext;
if (EXTERNAL_SSE2(mm_flags)) {
c->h264_idct8_add = ff_h264_idct8_add_8_sse2;
c->h264_idct_add16 = ff_h264_idct_add16_8_sse2;
c->h264_idct8_add4 = ff_h264_idct8_add4_8_sse2;
if (chroma_format_idc == 1)
c->h264_idct_add8 = ff_h264_idct_add8_8_sse2;
c->h264_idct_add16intra = ff_h264_idct_add16intra_8_sse2;
c->h264_luma_dc_dequant_idct = ff_h264_luma_dc_dequant_idct_sse2;
c->weight_h264_pixels_tab[0] = ff_h264_weight_16_sse2;
c->weight_h264_pixels_tab[1] = ff_h264_weight_8_sse2;
c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_sse2;
c->biweight_h264_pixels_tab[1] = ff_h264_biweight_8_sse2;
c->h264_v_loop_filter_luma = ff_deblock_v_luma_8_sse2;
c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_sse2;
c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_sse2;
c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_sse2;
}
if (EXTERNAL_SSSE3(mm_flags)) {
c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_ssse3;
c->biweight_h264_pixels_tab[1] = ff_h264_biweight_8_ssse3;
}
if (EXTERNAL_AVX(mm_flags)) {
c->h264_v_loop_filter_luma = ff_deblock_v_luma_8_avx;
c->h264_h_loop_filter_luma = ff_deblock_h_luma_8_avx;
c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_avx;
c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_avx;
}
}
}
} else if (bit_depth == 10) {
if (EXTERNAL_MMX(mm_flags)) {
if (EXTERNAL_MMXEXT(mm_flags)) {
#if ARCH_X86_32
c->h264_v_loop_filter_chroma = ff_deblock_v_chroma_10_mmxext;
c->h264_v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_10_mmxext;
c->h264_v_loop_filter_luma = ff_deblock_v_luma_10_mmxext;
c->h264_h_loop_filter_luma = ff_deblock_h_luma_10_mmxext;
c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_mmxext;
c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_mmxext;
#endif /* ARCH_X86_32 */
c->h264_idct_dc_add = ff_h264_idct_dc_add_10_mmxext;
if (EXTERNAL_SSE2(mm_flags)) {
c->h264_idct_add = ff_h264_idct_add_10_sse2;
c->h264_idct8_dc_add = ff_h264_idct8_dc_add_10_sse2;
c->h264_idct_add16 = ff_h264_idct_add16_10_sse2;
if (chroma_format_idc == 1)
c->h264_idct_add8 = ff_h264_idct_add8_10_sse2;
c->h264_idct_add16intra = ff_h264_idct_add16intra_10_sse2;
#if HAVE_ALIGNED_STACK
c->h264_idct8_add = ff_h264_idct8_add_10_sse2;
c->h264_idct8_add4 = ff_h264_idct8_add4_10_sse2;
#endif /* HAVE_ALIGNED_STACK */
c->weight_h264_pixels_tab[0] = ff_h264_weight_16_10_sse2;
c->weight_h264_pixels_tab[1] = ff_h264_weight_8_10_sse2;
c->weight_h264_pixels_tab[2] = ff_h264_weight_4_10_sse2;
c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_10_sse2;
c->biweight_h264_pixels_tab[1] = ff_h264_biweight_8_10_sse2;
c->biweight_h264_pixels_tab[2] = ff_h264_biweight_4_10_sse2;
c->h264_v_loop_filter_chroma = ff_deblock_v_chroma_10_sse2;
c->h264_v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_10_sse2;
#if HAVE_ALIGNED_STACK
c->h264_v_loop_filter_luma = ff_deblock_v_luma_10_sse2;
c->h264_h_loop_filter_luma = ff_deblock_h_luma_10_sse2;
c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_sse2;
c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_sse2;
#endif /* HAVE_ALIGNED_STACK */
}
if (EXTERNAL_SSE4(mm_flags)) {
c->weight_h264_pixels_tab[0] = ff_h264_weight_16_10_sse4;
c->weight_h264_pixels_tab[1] = ff_h264_weight_8_10_sse4;
c->weight_h264_pixels_tab[2] = ff_h264_weight_4_10_sse4;
c->biweight_h264_pixels_tab[0] = ff_h264_biweight_16_10_sse4;
c->biweight_h264_pixels_tab[1] = ff_h264_biweight_8_10_sse4;
c->biweight_h264_pixels_tab[2] = ff_h264_biweight_4_10_sse4;
}
if (EXTERNAL_AVX(mm_flags)) {
c->h264_idct_dc_add =
c->h264_idct_add = ff_h264_idct_add_10_avx;
c->h264_idct8_dc_add = ff_h264_idct8_dc_add_10_avx;
c->h264_idct_add16 = ff_h264_idct_add16_10_avx;
if (chroma_format_idc == 1)
c->h264_idct_add8 = ff_h264_idct_add8_10_avx;
c->h264_idct_add16intra = ff_h264_idct_add16intra_10_avx;
#if HAVE_ALIGNED_STACK
c->h264_idct8_add = ff_h264_idct8_add_10_avx;
c->h264_idct8_add4 = ff_h264_idct8_add4_10_avx;
#endif /* HAVE_ALIGNED_STACK */
c->h264_v_loop_filter_chroma = ff_deblock_v_chroma_10_avx;
c->h264_v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_10_avx;
#if HAVE_ALIGNED_STACK
c->h264_v_loop_filter_luma = ff_deblock_v_luma_10_avx;
c->h264_h_loop_filter_luma = ff_deblock_h_luma_10_avx;
c->h264_v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_avx;
c->h264_h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_avx;
#endif /* HAVE_ALIGNED_STACK */
}
}
}
}
#endif
}
av_cold void ff_blend_init_x86(FilterParams *param, int is_16bit)
{
int cpu_flags = av_get_cpu_flags();
if (!is_16bit) {
if (EXTERNAL_SSE2(cpu_flags) && param->opacity == 1) {
switch (param->mode) {
case BLEND_ADDITION: param->blend = ff_blend_addition_sse2; break;
case BLEND_GRAINMERGE: param->blend = ff_blend_grainmerge_sse2; break;
case BLEND_AND: param->blend = ff_blend_and_sse2; break;
case BLEND_AVERAGE: param->blend = ff_blend_average_sse2; break;
case BLEND_DARKEN: param->blend = ff_blend_darken_sse2; break;
case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_sse2; break;
case BLEND_DIVIDE: param->blend = ff_blend_divide_sse2; break;
case BLEND_HARDMIX: param->blend = ff_blend_hardmix_sse2; break;
case BLEND_LIGHTEN: param->blend = ff_blend_lighten_sse2; break;
case BLEND_MULTIPLY: param->blend = ff_blend_multiply_sse2; break;
case BLEND_OR: param->blend = ff_blend_or_sse2; break;
case BLEND_PHOENIX: param->blend = ff_blend_phoenix_sse2; break;
case BLEND_SCREEN: param->blend = ff_blend_screen_sse2; break;
case BLEND_SUBTRACT: param->blend = ff_blend_subtract_sse2; break;
case BLEND_XOR: param->blend = ff_blend_xor_sse2; break;
case BLEND_DIFFERENCE: param->blend = ff_blend_difference_sse2; break;
case BLEND_EXTREMITY: param->blend = ff_blend_extremity_sse2; break;
case BLEND_NEGATION: param->blend = ff_blend_negation_sse2; break;
}
}
if (EXTERNAL_SSSE3(cpu_flags) && param->opacity == 1) {
switch (param->mode) {
case BLEND_DIFFERENCE: param->blend = ff_blend_difference_ssse3; break;
case BLEND_EXTREMITY: param->blend = ff_blend_extremity_ssse3; break;
case BLEND_NEGATION: param->blend = ff_blend_negation_ssse3; break;
}
}
if (EXTERNAL_AVX2_FAST(cpu_flags) && param->opacity == 1) {
switch (param->mode) {
case BLEND_ADDITION: param->blend = ff_blend_addition_avx2; break;
case BLEND_GRAINMERGE: param->blend = ff_blend_grainmerge_avx2; break;
case BLEND_AND: param->blend = ff_blend_and_avx2; break;
case BLEND_AVERAGE: param->blend = ff_blend_average_avx2; break;
case BLEND_DARKEN: param->blend = ff_blend_darken_avx2; break;
case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_avx2; break;
case BLEND_HARDMIX: param->blend = ff_blend_hardmix_avx2; break;
case BLEND_LIGHTEN: param->blend = ff_blend_lighten_avx2; break;
case BLEND_MULTIPLY: param->blend = ff_blend_multiply_avx2; break;
case BLEND_OR: param->blend = ff_blend_or_avx2; break;
case BLEND_PHOENIX: param->blend = ff_blend_phoenix_avx2; break;
case BLEND_SCREEN: param->blend = ff_blend_screen_avx2; break;
case BLEND_SUBTRACT: param->blend = ff_blend_subtract_avx2; break;
case BLEND_XOR: param->blend = ff_blend_xor_avx2; break;
case BLEND_DIFFERENCE: param->blend = ff_blend_difference_avx2; break;
case BLEND_EXTREMITY: param->blend = ff_blend_extremity_avx2; break;
case BLEND_NEGATION: param->blend = ff_blend_negation_avx2; break;
}
}
} else { /* is_16_bit */
#if ARCH_X86_64
if (EXTERNAL_SSE2(cpu_flags) && param->opacity == 1) {
switch (param->mode) {
case BLEND_ADDITION: param->blend = ff_blend_addition_16_sse2; break;
case BLEND_AND: param->blend = ff_blend_and_16_sse2; break;
case BLEND_AVERAGE: param->blend = ff_blend_average_16_sse2; break;
case BLEND_OR: param->blend = ff_blend_or_16_sse2; break;
case BLEND_SUBTRACT: param->blend = ff_blend_subtract_16_sse2; break;
case BLEND_XOR: param->blend = ff_blend_xor_16_sse2; break;
}
}
if (EXTERNAL_SSE4(cpu_flags) && param->opacity == 1) {
switch (param->mode) {
case BLEND_GRAINMERGE: param->blend = ff_blend_grainmerge_16_sse4; break;
case BLEND_DARKEN: param->blend = ff_blend_darken_16_sse4; break;
case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_16_sse4; break;
case BLEND_DIFFERENCE: param->blend = ff_blend_difference_16_sse4; break;
case BLEND_EXTREMITY: param->blend = ff_blend_extremity_16_sse4; break;
case BLEND_NEGATION: param->blend = ff_blend_negation_16_sse4; break;
case BLEND_LIGHTEN: param->blend = ff_blend_lighten_16_sse4; break;
case BLEND_PHOENIX: param->blend = ff_blend_phoenix_16_sse4; break;
}
}
if (EXTERNAL_AVX2_FAST(cpu_flags) && param->opacity == 1) {
switch (param->mode) {
case BLEND_ADDITION: param->blend = ff_blend_addition_16_avx2; break;
case BLEND_GRAINMERGE: param->blend = ff_blend_grainmerge_16_avx2; break;
case BLEND_AND: param->blend = ff_blend_and_16_avx2; break;
case BLEND_AVERAGE: param->blend = ff_blend_average_16_avx2; break;
case BLEND_DARKEN: param->blend = ff_blend_darken_16_avx2; break;
case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_16_avx2; break;
case BLEND_DIFFERENCE: param->blend = ff_blend_difference_16_avx2; break;
case BLEND_EXTREMITY: param->blend = ff_blend_extremity_16_avx2; break;
case BLEND_NEGATION: param->blend = ff_blend_negation_16_avx2; break;
case BLEND_LIGHTEN: param->blend = ff_blend_lighten_16_avx2; break;
case BLEND_OR: param->blend = ff_blend_or_16_avx2; break;
case BLEND_PHOENIX: param->blend = ff_blend_phoenix_16_avx2; break;
case BLEND_SUBTRACT: param->blend = ff_blend_subtract_16_avx2; break;
case BLEND_XOR: param->blend = ff_blend_xor_16_avx2; break;
}
}
#endif /* ARCH_X86_64 */
}
}
av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
{
int cpu_flags = av_get_cpu_flags();
if (EXTERNAL_MMX(cpu_flags)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S32,
0, 1, 8, "MMX", ff_conv_s32_to_s16_mmx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
6, 1, 4, "MMX", ff_conv_fltp_to_flt_6ch_mmx);
}
if (EXTERNAL_SSE(cpu_flags)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLTP,
6, 1, 2, "SSE", ff_conv_fltp_to_s16_6ch_sse);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
2, 16, 8, "SSE", ff_conv_fltp_to_flt_2ch_sse);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_FLT,
2, 16, 4, "SSE", ff_conv_flt_to_fltp_2ch_sse);
}
if (EXTERNAL_SSE2(cpu_flags)) {
if (!(cpu_flags & AV_CPU_FLAG_SSE2SLOW)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S32,
0, 16, 16, "SSE2", ff_conv_s32_to_s16_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
6, 16, 8, "SSE2", ff_conv_s16p_to_s16_6ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLTP,
6, 16, 4, "SSE2", ff_conv_fltp_to_s16_6ch_sse2);
} else {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
6, 1, 4, "SSE2SLOW", ff_conv_s16p_to_s16_6ch_sse2slow);
}
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_S16,
0, 16, 8, "SSE2", ff_conv_s16_to_s32_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16,
0, 16, 8, "SSE2", ff_conv_s16_to_flt_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S32,
0, 16, 8, "SSE2", ff_conv_s32_to_flt_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLT,
0, 16, 16, "SSE2", ff_conv_flt_to_s16_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_FLT,
0, 16, 16, "SSE2", ff_conv_flt_to_s32_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
2, 16, 16, "SSE2", ff_conv_s16p_to_s16_2ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16P,
2, 16, 8, "SSE2", ff_conv_s16p_to_flt_2ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16P,
6, 16, 4, "SSE2", ff_conv_s16p_to_flt_6ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLTP,
2, 16, 4, "SSE2", ff_conv_fltp_to_s16_2ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_S16,
2, 16, 8, "SSE2", ff_conv_s16_to_s16p_2ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_S16,
6, 16, 4, "SSE2", ff_conv_s16_to_s16p_6ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_S16,
2, 16, 8, "SSE2", ff_conv_s16_to_fltp_2ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_S16,
6, 16, 4, "SSE2", ff_conv_s16_to_fltp_6ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_FLT,
2, 16, 8, "SSE2", ff_conv_flt_to_s16p_2ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_FLT,
6, 16, 4, "SSE2", ff_conv_flt_to_s16p_6ch_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_FLT,
6, 16, 4, "SSE2", ff_conv_flt_to_fltp_6ch_sse2);
}
if (EXTERNAL_SSSE3(cpu_flags)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16P,
6, 16, 4, "SSSE3", ff_conv_s16p_to_flt_6ch_ssse3);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLTP,
2, 16, 4, "SSSE3", ff_conv_fltp_to_s16_2ch_ssse3);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_S16,
2, 16, 8, "SSSE3", ff_conv_s16_to_s16p_2ch_ssse3);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_S16,
6, 16, 4, "SSSE3", ff_conv_s16_to_s16p_6ch_ssse3);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_S16,
6, 16, 4, "SSSE3", ff_conv_s16_to_fltp_6ch_ssse3);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_FLT,
6, 16, 4, "SSSE3", ff_conv_flt_to_s16p_6ch_ssse3);
}
if (EXTERNAL_SSE4(cpu_flags)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16,
0, 16, 8, "SSE4", ff_conv_s16_to_flt_sse4);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
6, 16, 4, "SSE4", ff_conv_fltp_to_flt_6ch_sse4);
}
if (EXTERNAL_AVX_FAST(cpu_flags)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S32,
0, 32, 16, "AVX", ff_conv_s32_to_flt_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_FLT,
0, 32, 32, "AVX", ff_conv_flt_to_s32_avx);
}
if (EXTERNAL_AVX(cpu_flags)) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
2, 16, 16, "AVX", ff_conv_s16p_to_s16_2ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_S16P,
6, 16, 8, "AVX", ff_conv_s16p_to_s16_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16P,
2, 16, 8, "AVX", ff_conv_s16p_to_flt_2ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16P,
6, 16, 4, "AVX", ff_conv_s16p_to_flt_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLTP,
6, 16, 4, "AVX", ff_conv_fltp_to_s16_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
6, 16, 4, "AVX", ff_conv_fltp_to_flt_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_S16,
2, 16, 8, "AVX", ff_conv_s16_to_s16p_2ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_S16,
6, 16, 4, "AVX", ff_conv_s16_to_s16p_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_S16,
2, 16, 8, "AVX", ff_conv_s16_to_fltp_2ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_S16,
6, 16, 4, "AVX", ff_conv_s16_to_fltp_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_FLT,
2, 16, 8, "AVX", ff_conv_flt_to_s16p_2ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_FLT,
6, 16, 4, "AVX", ff_conv_flt_to_s16p_6ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_FLT,
2, 16, 4, "AVX", ff_conv_flt_to_fltp_2ch_avx);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_FLT,
6, 16, 4, "AVX", ff_conv_flt_to_fltp_6ch_avx);
}
}
void ff_hevcdsp_init_x86(HEVCDSPContext *c, const int bit_depth)
{
int mm_flags = av_get_cpu_flags();
if (bit_depth == 8) {
if (EXTERNAL_MMXEXT(mm_flags)) {
c->transform_dc_add[0] = ff_hevc_idct4_dc_add_8_mmxext;
c->transform_dc_add[1] = ff_hevc_idct8_dc_add_8_mmxext;
}
if (EXTERNAL_SSE2(mm_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
c->transform_dc_add[2] = ff_hevc_idct16_dc_add_8_sse2;
c->transform_dc_add[3] = ff_hevc_idct32_dc_add_8_sse2;
}
if (EXTERNAL_SSSE3(mm_flags) && ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
}
if (EXTERNAL_SSE4(mm_flags) && ARCH_X86_64) {
EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 8, sse4);
}
if (EXTERNAL_AVX2(mm_flags)) {
c->transform_dc_add[3] = ff_hevc_idct32_dc_add_8_avx2;
}
} else if (bit_depth == 10) {
if (EXTERNAL_MMXEXT(mm_flags)) {
c->transform_dc_add[0] = ff_hevc_idct4_dc_add_10_mmxext;
}
if (EXTERNAL_SSE2(mm_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
c->transform_dc_add[1] = ff_hevc_idct8_dc_add_10_sse2;
c->transform_dc_add[2] = ff_hevc_idct16_dc_add_10_sse2;
c->transform_dc_add[3] = ff_hevc_idct32_dc_add_10_sse2;
}
if (EXTERNAL_SSSE3(mm_flags) && ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
}
if (EXTERNAL_SSE4(mm_flags) && ARCH_X86_64) {
EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 10, sse4);
}
if (EXTERNAL_AVX(mm_flags)) {
c->transform_dc_add[1] = ff_hevc_idct8_dc_add_10_avx;
c->transform_dc_add[2] = ff_hevc_idct16_dc_add_10_avx;
c->transform_dc_add[3] = ff_hevc_idct32_dc_add_10_avx;
}
if (EXTERNAL_AVX2(mm_flags)) {
c->transform_dc_add[2] = ff_hevc_idct16_dc_add_10_avx2;
c->transform_dc_add[3] = ff_hevc_idct32_dc_add_10_avx2;
}
}
}
av_cold void ff_vp8dsp_init_x86(VP8DSPContext *c)
{
#if HAVE_YASM
int cpu_flags = av_get_cpu_flags();
if (EXTERNAL_MMX(cpu_flags)) {
c->vp8_idct_dc_add = ff_vp8_idct_dc_add_mmx;
c->vp8_idct_dc_add4uv = ff_vp8_idct_dc_add4uv_mmx;
#if ARCH_X86_32
c->vp8_idct_dc_add4y = ff_vp8_idct_dc_add4y_mmx;
c->vp8_idct_add = ff_vp8_idct_add_mmx;
c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_mmx;
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmx;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmx;
c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_mmx;
c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_mmx;
c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_mmx;
c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_mmx;
c->vp8_v_loop_filter16y = ff_vp8_v_loop_filter16y_mbedge_mmx;
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_mmx;
c->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_mbedge_mmx;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_mmx;
#endif
}
/* note that 4-tap width=16 functions are missing because w=16
* is only used for luma, and luma is always a copy or sixtap. */
if (EXTERNAL_MMXEXT(cpu_flags)) {
#if ARCH_X86_32
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmxext;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmxext;
c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_mmxext;
c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_mmxext;
c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_mmxext;
c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_mmxext;
c->vp8_v_loop_filter16y = ff_vp8_v_loop_filter16y_mbedge_mmxext;
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_mmxext;
c->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_mbedge_mmxext;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_mmxext;
#endif
}
if (EXTERNAL_SSE(cpu_flags)) {
c->vp8_idct_add = ff_vp8_idct_add_sse;
c->vp8_luma_dc_wht = ff_vp8_luma_dc_wht_sse;
}
if (HAVE_SSE2_EXTERNAL && cpu_flags & (AV_CPU_FLAG_SSE2 | AV_CPU_FLAG_SSE2SLOW)) {
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_sse2;
c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_sse2;
c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_sse2;
c->vp8_v_loop_filter16y = ff_vp8_v_loop_filter16y_mbedge_sse2;
c->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_mbedge_sse2;
}
if (EXTERNAL_SSE2(cpu_flags)) {
c->vp8_idct_dc_add4y = ff_vp8_idct_dc_add4y_sse2;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse2;
c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_sse2;
c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_sse2;
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_sse2;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_sse2;
}
if (EXTERNAL_SSSE3(cpu_flags)) {
c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_ssse3;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_ssse3;
c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_ssse3;
c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_ssse3;
c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_ssse3;
c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_ssse3;
c->vp8_v_loop_filter16y = ff_vp8_v_loop_filter16y_mbedge_ssse3;
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_ssse3;
c->vp8_v_loop_filter8uv = ff_vp8_v_loop_filter8uv_mbedge_ssse3;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_ssse3;
}
if (EXTERNAL_SSE4(cpu_flags)) {
c->vp8_idct_dc_add = ff_vp8_idct_dc_add_sse4;
c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse4;
c->vp8_h_loop_filter16y = ff_vp8_h_loop_filter16y_mbedge_sse4;
c->vp8_h_loop_filter8uv = ff_vp8_h_loop_filter8uv_mbedge_sse4;
}
#endif /* HAVE_YASM */
}
void ff_hevcdsp_init_x86(HEVCDSPContext *c, const int bit_depth)
{
int mm_flags = av_get_cpu_flags();
if (bit_depth == 8) {
if (EXTERNAL_MMX(mm_flags)) {
/*if (mm_flags & AV_CPU_FLAG_CMOV)
c->h264_luma_dc_dequant_idct = ff_h264_luma_dc_dequant_idct_mmx; */
if (EXTERNAL_MMXEXT(mm_flags)) {
#if ARCH_X86_32 && HAVE_MMXEXT_EXTERNAL
/* MMEXT optimizations */
#endif /* ARCH_X86_32 && HAVE_MMXEXT_EXTERNAL */
if (EXTERNAL_SSE2(mm_flags))
{
#if 0
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
#endif
}
if (EXTERNAL_SSSE3(mm_flags)) {
c->transform_4x4_luma_add = ff_hevc_transform_4x4_luma_add_8_sse4;
c->transform_add[0] = ff_hevc_transform_4x4_add_8_sse4;
c->transform_add[1] = ff_hevc_transform_8x8_add_8_sse4;
c->transform_add[2] = ff_hevc_transform_16x16_add_8_sse4;
c->transform_add[3] = ff_hevc_transform_32x32_add_8_sse4;
c->put_unweighted_pred = ff_hevc_put_unweighted_pred_8_sse;
c->put_hevc_qpel[0][0] = ff_hevc_put_hevc_qpel_pixels_8_sse;
c->put_hevc_qpel[0][1] = ff_hevc_put_hevc_qpel_h_1_8_sse;
c->put_hevc_qpel[0][2] = ff_hevc_put_hevc_qpel_h_2_8_sse;
c->put_hevc_qpel[0][3] = ff_hevc_put_hevc_qpel_h_3_8_sse;
c->put_hevc_qpel[1][0] = ff_hevc_put_hevc_qpel_v_1_8_sse;
c->put_hevc_qpel[2][0] = ff_hevc_put_hevc_qpel_v_2_8_sse;
c->put_hevc_qpel[3][0] = ff_hevc_put_hevc_qpel_v_3_8_sse;
//#if ARCH_X86_64
#if 0
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
#endif
}
if (EXTERNAL_SSE4(mm_flags)) {
#if GCC_VERSION > MIN_GCC_VERSION_MC || __APPLE__
c->put_weighted_pred_avg = ff_hevc_put_weighted_pred_avg_8_sse;
c->weighted_pred = ff_hevc_weighted_pred_8_sse;
c->weighted_pred_avg = ff_hevc_weighted_pred_avg_8_sse;
c->put_hevc_epel[0][0] = ff_hevc_put_hevc_epel_pixels_8_sse;
c->put_hevc_epel[0][1] = ff_hevc_put_hevc_epel_h_8_sse;
c->put_hevc_epel[1][0] = ff_hevc_put_hevc_epel_v_8_sse;
c->put_hevc_epel[1][1] = ff_hevc_put_hevc_epel_hv_8_sse;
#endif
c->sao_edge_filter[0] = ff_hevc_sao_edge_filter_0_8_sse;
c->sao_edge_filter[1] = ff_hevc_sao_edge_filter_1_8_sse;
c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_2_8_sse;
c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_3_8_sse;
c->sao_band_filter[0] = ff_hevc_sao_band_filter_0_8_sse;
c->sao_band_filter[1] = ff_hevc_sao_band_filter_1_8_sse;
c->sao_band_filter[2] = ff_hevc_sao_band_filter_2_8_sse;
c->sao_band_filter[3] = ff_hevc_sao_band_filter_3_8_sse;
c->put_hevc_qpel[1][1] = ff_hevc_put_hevc_qpel_h_1_v_1_sse;
c->put_hevc_qpel[1][2] = ff_hevc_put_hevc_qpel_h_2_v_1_sse;
c->put_hevc_qpel[1][3] = ff_hevc_put_hevc_qpel_h_3_v_1_sse;
c->put_hevc_qpel[2][1] = ff_hevc_put_hevc_qpel_h_1_v_2_sse;
c->put_hevc_qpel[2][2] = ff_hevc_put_hevc_qpel_h_2_v_2_sse;
c->put_hevc_qpel[2][3] = ff_hevc_put_hevc_qpel_h_3_v_2_sse;
c->put_hevc_qpel[3][1] = ff_hevc_put_hevc_qpel_h_1_v_3_sse;
c->put_hevc_qpel[3][2] = ff_hevc_put_hevc_qpel_h_2_v_3_sse;
c->put_hevc_qpel[3][3] = ff_hevc_put_hevc_qpel_h_3_v_3_sse;
}
if (EXTERNAL_AVX(mm_flags)) {
}
}
}
} else if (bit_depth == 10) {
if (EXTERNAL_MMX(mm_flags)) {
if (EXTERNAL_MMXEXT(mm_flags)) {
#if ARCH_X86_32
#endif /* ARCH_X86_32 */
if (EXTERNAL_SSE2(mm_flags)) {
#if HAVE_ALIGNED_STACK
/*stuff that requires aligned stack */
#endif /* HAVE_ALIGNED_STACK */
}
if (EXTERNAL_SSE4(mm_flags)) {
c->transform_4x4_luma_add = ff_hevc_transform_4x4_luma_add_10_sse4;
c->transform_add[0] = ff_hevc_transform_4x4_add_10_sse4;
c->transform_add[1] = ff_hevc_transform_8x8_add_10_sse4;
c->transform_add[2] = ff_hevc_transform_16x16_add_10_sse4;
c->transform_add[3] = ff_hevc_transform_32x32_add_10_sse4;
c->put_hevc_epel[0][0] = ff_hevc_put_hevc_epel_pixels_10_sse;
c->put_hevc_epel[0][1] = ff_hevc_put_hevc_epel_h_10_sse;
c->put_hevc_epel[1][0] = ff_hevc_put_hevc_epel_v_10_sse;
c->put_hevc_epel[1][1] = ff_hevc_put_hevc_epel_hv_10_sse;
c->put_hevc_qpel[0][0] = ff_hevc_put_hevc_qpel_pixels_10_sse;
c->put_hevc_qpel[0][1] = ff_hevc_put_hevc_qpel_h_1_10_sse;
c->put_hevc_qpel[1][0] = ff_hevc_put_hevc_qpel_v_1_10_sse;
c->put_hevc_qpel[2][0] = ff_hevc_put_hevc_qpel_v_2_10_sse;
c->put_hevc_qpel[3][0] = ff_hevc_put_hevc_qpel_v_3_10_sse;
}
if (EXTERNAL_AVX(mm_flags)) {
}
}
}
}
}
void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
{
int cpu_flags = av_get_cpu_flags();
if (bit_depth == 8) {
#if (HAVE_MMXEXT_EXTERNAL == 1)
if (EXTERNAL_MMXEXT(cpu_flags)) {
c->idct_dc[0] = ff_hevc_idct4x4_dc_8_mmxext;
c->idct_dc[1] = ff_hevc_idct8x8_dc_8_mmxext;
c->transform_add[0] = ff_hevc_transform_add4_8_mmxext;
}
#endif
#if (HAVE_SSE2_EXTERNAL == 1)
if (EXTERNAL_SSE2(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_sse2;
if (ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_sse2;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_sse2;
}
c->idct_dc[1] = ff_hevc_idct8x8_dc_8_sse2;
c->idct_dc[2] = ff_hevc_idct16x16_dc_8_sse2;
c->idct_dc[3] = ff_hevc_idct32x32_dc_8_sse2;
c->transform_add[1] = ff_hevc_transform_add8_8_sse2;
c->transform_add[2] = ff_hevc_transform_add16_8_sse2;
c->transform_add[3] = ff_hevc_transform_add32_8_sse2;
}
#endif
#if (HAVE_SSSE3_EXTERNAL == 1)
if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_ssse3;
}
#endif
#if (HAVE_SSE4_EXTERNAL == 1)
if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 8, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 8, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 8, sse4);
}
#endif
#if (HAVE_AVX_EXTERNAL == 1)
if (EXTERNAL_AVX(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_8_avx;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_8_avx;
if (ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_8_avx;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_8_avx;
}
c->transform_add[1] = ff_hevc_transform_add8_8_avx;
c->transform_add[2] = ff_hevc_transform_add16_8_avx;
c->transform_add[3] = ff_hevc_transform_add32_8_avx;
}
#endif
#if (HAVE_AVX2_EXTERNAL == 1)
if (EXTERNAL_AVX2(cpu_flags)) {
c->idct_dc[2] = ff_hevc_idct16x16_dc_8_avx2;
c->idct_dc[3] = ff_hevc_idct32x32_dc_8_avx2;
c->transform_add[3] = ff_hevc_transform_add32_8_avx2;
}
#endif
} else if (bit_depth == 10) {
#if (HAVE_MMXEXT_EXTERNAL == 1)
if (EXTERNAL_MMXEXT(cpu_flags)) {
c->transform_add[0] = ff_hevc_transform_add4_10_mmxext;
c->idct_dc[0] = ff_hevc_idct4x4_dc_10_mmxext;
c->idct_dc[1] = ff_hevc_idct8x8_dc_10_mmxext;
}
#endif
#if (HAVE_SSE2_EXTERNAL == 1)
if (EXTERNAL_SSE2(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_sse2;
if (ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_sse2;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_sse2;
}
c->idct_dc[1] = ff_hevc_idct8x8_dc_10_sse2;
c->idct_dc[2] = ff_hevc_idct16x16_dc_10_sse2;
c->idct_dc[3] = ff_hevc_idct32x32_dc_10_sse2;
c->transform_add[1] = ff_hevc_transform_add8_10_sse2;
c->transform_add[2] = ff_hevc_transform_add16_10_sse2;
c->transform_add[3] = ff_hevc_transform_add32_10_sse2;
}
#endif
#if (HAVE_SSSE3_EXTERNAL == 1)
if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_ssse3;
}
#endif
#if (HAVE_SSE4_EXTERNAL == 1)
if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 10, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 10, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 10, sse4);
}
#endif
#if (HAVE_AVX_EXTERNAL == 1)
if (EXTERNAL_AVX(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_10_avx;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_10_avx;
if (ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_10_avx;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_10_avx;
}
}
#endif
#if (HAVE_AVX2_EXTERNAL == 1)
if (EXTERNAL_AVX2(cpu_flags)) {
c->idct_dc[2] = ff_hevc_idct16x16_dc_10_avx2;
c->idct_dc[3] = ff_hevc_idct32x32_dc_10_avx2;
c->transform_add[2] = ff_hevc_transform_add16_10_avx2;
c->transform_add[3] = ff_hevc_transform_add32_10_avx2;
}
#endif
} else if (bit_depth == 12) {
#if (HAVE_MMXEXT_EXTERNAL == 1)
if (EXTERNAL_MMXEXT(cpu_flags)) {
c->idct_dc[0] = ff_hevc_idct4x4_dc_12_mmxext;
c->idct_dc[1] = ff_hevc_idct8x8_dc_12_mmxext;
}
#endif
#if (HAVE_SSE2_EXTERNAL == 1)
if (EXTERNAL_SSE2(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_sse2;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_sse2;
if (ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_sse2;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_sse2;
}
c->idct_dc[1] = ff_hevc_idct8x8_dc_12_sse2;
c->idct_dc[2] = ff_hevc_idct16x16_dc_12_sse2;
c->idct_dc[3] = ff_hevc_idct32x32_dc_12_sse2;
}
#endif
#if (HAVE_SSSE3_EXTERNAL == 1)
if (EXTERNAL_SSSE3(cpu_flags) && ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_ssse3;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_ssse3;
}
#endif
#if (HAVE_SSE4_EXTERNAL == 1)
if (EXTERNAL_SSE4(cpu_flags) && ARCH_X86_64) {
EPEL_LINKS(c->put_hevc_epel, 0, 0, pel_pixels, 12, sse4);
EPEL_LINKS(c->put_hevc_epel, 0, 1, epel_h, 12, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 0, epel_v, 12, sse4);
EPEL_LINKS(c->put_hevc_epel, 1, 1, epel_hv, 12, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 0, pel_pixels, 12, sse4);
QPEL_LINKS(c->put_hevc_qpel, 0, 1, qpel_h, 12, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 12, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 12, sse4);
}
#endif
#if (HAVE_AVX_EXTERNAL == 1)
if (EXTERNAL_AVX(cpu_flags)) {
c->hevc_v_loop_filter_chroma = ff_hevc_v_loop_filter_chroma_12_avx;
c->hevc_h_loop_filter_chroma = ff_hevc_h_loop_filter_chroma_12_avx;
if (ARCH_X86_64) {
c->hevc_v_loop_filter_luma = ff_hevc_v_loop_filter_luma_12_avx;
c->hevc_h_loop_filter_luma = ff_hevc_h_loop_filter_luma_12_avx;
}
}
#endif
#if (HAVE_AVX2_EXTERNAL == 1)
if (EXTERNAL_AVX2(cpu_flags)) {
c->idct_dc[2] = ff_hevc_idct16x16_dc_12_avx2;
c->idct_dc[3] = ff_hevc_idct32x32_dc_12_avx2;
}
#endif
}
}