virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-08-04 22:03:09 +02:00
Code Issues Releases Activity

vf_bwdif: add AVX512 implementation

Browse Source 
I also tried replacing some of the instructions by more elaborate ones
using masks, but I found no performance gain significant enough to be worth
maintaining two code paths, so this implementation merely replaces the AVX2
implementation by drop-in AVX512 equivalents.

bwdif8_c:                                             6362.2 ( 1.00x)
bwdif8_sse2:                                          1004.9 ( 6.33x)
bwdif8_ssse3:                                          946.0 ( 6.73x)
bwdif8_avx2:                                           477.9 (13.31x)
bwdif8_avx512:                                         273.3 (23.28x)

bwdif10_c:                                            6341.5 ( 1.00x)
bwdif10_sse2:                                          872.4 ( 7.27x)
bwdif10_ssse3:                                         803.4 ( 7.89x)
bwdif10_avx2:                                          416.7 (15.22x)
bwdif10_avx512:                                        224.3 (28.27x)

Realtime test at 3840x2160 yuv420p:

avx2:   frame=20000 fps=3370 q=-0.0 Lsize=N/A time=00:06:40.00 bitrate=N/A speed=67.4x elapsed=0:00:05.93
avx512: frame=20000 fps=5077 q=-0.0 Lsize=N/A time=00:06:40.00 bitrate=N/A speed= 102x elapsed=0:00:03.93

The use of this function is gated behind avx512icl so that it doesn't
downclock on Skylake.
This commit is contained in:
Niklas Haas
2025-07-29 13:27:52 +02:00
committed by kierank
parent 586a1cd088
commit 7f00e24d70
2 changed files with 38 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
libavfilter/x86/vf_bwdif.asm
View File

@ -26,17 +26,17 @@
%include "libavutil/x86/x86util.asm"
SECTION_RODATA 32
SECTION_RODATA 64
pw_coefhf: times 8 dw 1016, 5570
pw_coefhf1: times 16 dw -3801
pw_coefsp: times 8 dw 5077, -981
pw_splfdif: times 8 dw -768, 768
pw_coefhf: times 16 dw 1016, 5570
pw_coefhf1: times 32 dw -3801
pw_coefsp: times 16 dw 5077, -981
pw_splfdif: times 16 dw -768, 768
SECTION .text
%macro LOAD8 2
%if mmsize == 32
%if mmsize >= 32
pmovzxbw %1, %2
%else
movh %1, %2
@ -49,7 +49,10 @@ SECTION .text
%endmacro
%macro DISP8 0
%if mmsize == 32
%if mmsize == 64
pmaxsw m2, m7
vpmovuswb [dstq], m2
%elif mmsize == 32
vextracti128 xm1, m2, 1
packuswb xm2, xm1
movu [dstq], xm2
@ -123,8 +126,13 @@ SECTION .text
mova m6, m7
psubw m6, m3
pmaxsw m6, m5
%if cpuflag(avx512)
pcmpgtw k1, m2, m7
vpmovm2w m3, k1
%else
mova m3, m2
pcmpgtw m3, m7
%endif
pand m6, m3
pmaxsw m2, m6
mova m11, m2
@ -169,7 +177,12 @@ SECTION .text
paddw m6, m5
psubw m1, m4
ABS1 m1, m4
%if cpuflag(avx512)
pcmpgtw k1, m1, m9
vpmovm2w m1, k1
%else
pcmpgtw m1, m9
%endif
mova m4, m1
punpcklwd m1, m4
punpckhwd m4, m4
@ -254,7 +267,7 @@ cglobal bwdif_filter_line_12bit, 4, 9, 13, 0, dst, prev, cur, next, w, \
prefs, mrefs, prefs2, mrefs2, \
prefs3, mrefs3, prefs4, \
mrefs4, parity, clip_max
%if mmsize == 32
%if mmsize >= 32
vpbroadcastw m12, WORD clip_maxm
%else
movd m12, DWORD clip_maxm
@ -283,3 +296,8 @@ BWDIF
INIT_YMM avx2
BWDIF
%endif
%if HAVE_AVX512_EXTERNAL && ARCH_X86_64
INIT_ZMM avx512icl
BWDIF
%endif

12
libavfilter/x86/vf_bwdif_init.c
View File

@ -36,6 +36,10 @@ void ff_bwdif_filter_line_avx2(void *dst, const void *prev, const void *cur, con
int w, int prefs, int mrefs, int prefs2,
int mrefs2, int prefs3, int mrefs3, int prefs4,
int mrefs4, int parity, int clip_max);
void ff_bwdif_filter_line_avx512icl(void *dst, const void *prev, const void *cur, const void *next,
int w, int prefs, int mrefs, int prefs2,
int mrefs2, int prefs3, int mrefs3, int prefs4,
int mrefs4, int parity, int clip_max);
void ff_bwdif_filter_line_12bit_sse2(void *dst, const void *prev, const void *cur, const void *next,
int w, int prefs, int mrefs, int prefs2,
@ -49,6 +53,10 @@ void ff_bwdif_filter_line_12bit_avx2(void *dst, const void *prev, const void *cu
int w, int prefs, int mrefs, int prefs2,
int mrefs2, int prefs3, int mrefs3, int prefs4,
int mrefs4, int parity, int clip_max);
void ff_bwdif_filter_line_12bit_avx512icl(void *dst, const void *prev, const void *cur, const void *next,
int w, int prefs, int mrefs, int prefs2,
int mrefs2, int prefs3, int mrefs3, int prefs4,
int mrefs4, int parity, int clip_max);
av_cold void ff_bwdif_init_x86(BWDIFDSPContext *bwdif, int bit_depth)
{
@ -61,6 +69,8 @@ av_cold void ff_bwdif_init_x86(BWDIFDSPContext *bwdif, int bit_depth)
bwdif->filter_line = ff_bwdif_filter_line_ssse3;
if (ARCH_X86_64 && EXTERNAL_AVX2_FAST(cpu_flags))
bwdif->filter_line = ff_bwdif_filter_line_avx2;
if (ARCH_X86_64 && EXTERNAL_AVX512ICL(cpu_flags))
bwdif->filter_line = ff_bwdif_filter_line_avx512icl;
} else if (bit_depth <= 12) {
if (EXTERNAL_SSE2(cpu_flags))
bwdif->filter_line = ff_bwdif_filter_line_12bit_sse2;
@ -68,5 +78,7 @@ av_cold void ff_bwdif_init_x86(BWDIFDSPContext *bwdif, int bit_depth)
bwdif->filter_line = ff_bwdif_filter_line_12bit_ssse3;
if (ARCH_X86_64 && EXTERNAL_AVX2_FAST(cpu_flags))
bwdif->filter_line = ff_bwdif_filter_line_12bit_avx2;
if (ARCH_X86_64 && EXTERNAL_AVX512ICL(cpu_flags))
bwdif->filter_line = ff_bwdif_filter_line_12bit_avx512icl;
}
}