Browse Source

x86/hevc_deblock: add add ff_hevc_[hv]_loop_filter_luma_{8, 10, 12}_avx

~5% faster than SSSE3

Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
tags/n2.4
James Almer Michael Niedermayer 11 years ago
parent
commit
73c4f63ba5
2 changed files with 39 additions and 1 deletions
  1. +9
    -1
      libavcodec/x86/hevc_deblock.asm
  2. +30
    -0
      libavcodec/x86/hevcdsp_init.c

+ 9
- 1
libavcodec/x86/hevc_deblock.asm View File

@@ -663,11 +663,11 @@ ALIGN 16
MASKED_COPY m4, m8
%endmacro

INIT_XMM sse2
;-----------------------------------------------------------------------------
; void ff_hevc_v_loop_filter_chroma(uint8_t *_pix, ptrdiff_t _stride, int *_tc,
; uint8_t *_no_p, uint8_t *_no_q);
;-----------------------------------------------------------------------------
%macro LOOP_FILTER_CHROMA 0
cglobal hevc_v_loop_filter_chroma_8, 3, 5, 7, pix, stride, tc, pix0, r3stride
sub pixq, 2
lea r3strideq, [3*strideq]
@@ -752,6 +752,12 @@ cglobal hevc_h_loop_filter_chroma_12, 3, 4, 7, pix, stride, tc, pix0
movu [pix0q+strideq], m1
movu [pixq], m2
RET
%endmacro

INIT_XMM sse2
LOOP_FILTER_CHROMA
INIT_XMM avx
LOOP_FILTER_CHROMA

%if ARCH_X86_64
%macro LOOP_FILTER_LUMA 0
@@ -903,4 +909,6 @@ INIT_XMM sse2
LOOP_FILTER_LUMA
INIT_XMM ssse3
LOOP_FILTER_LUMA
INIT_XMM avx
LOOP_FILTER_LUMA
%endif

+ 30
- 0
libavcodec/x86/hevcdsp_init.c View File

@@ -45,12 +45,18 @@ void ff_hevc_ ## DIR ## _loop_filter_luma_ ## DEPTH ## _ ## OPT(uint8_t *pix, pt
LFC_FUNCS(uint8_t, 8, sse2)
LFC_FUNCS(uint8_t, 10, sse2)
LFC_FUNCS(uint8_t, 12, sse2)
LFC_FUNCS(uint8_t, 8, avx)
LFC_FUNCS(uint8_t, 10, avx)
LFC_FUNCS(uint8_t, 12, avx)
LFL_FUNCS(uint8_t, 8, sse2)
LFL_FUNCS(uint8_t, 10, sse2)
LFL_FUNCS(uint8_t, 12, sse2)
LFL_FUNCS(uint8_t, 8, ssse3)
LFL_FUNCS(uint8_t, 10, ssse3)
LFL_FUNCS(uint8_t, 12, ssse3)
LFL_FUNCS(uint8_t, 8, avx)
LFL_FUNCS(uint8_t, 10, avx)
LFL_FUNCS(uint8_t, 12, avx)

#define IDCT_FUNCS(W, opt) \
void ff_hevc_idct##W##_dc_8_##opt(int16_t *coeffs); \
@@ -492,6 +498,14 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
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_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;
}
}
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;
@@ -528,6 +542,14 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
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(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;
}
}
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;
@@ -565,6 +587,14 @@ void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
QPEL_LINKS(c->put_hevc_qpel, 1, 0, qpel_v, 12, sse4);
QPEL_LINKS(c->put_hevc_qpel, 1, 1, qpel_hv, 12, sse4);
}
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;
}
}
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;


Loading…
Cancel
Save