| Message ID | 20210930015612.29608-3-jianhua.wu@intel.com |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel,v2,1/4] libswscale/x86/rgb2rgb: add shuffle_bytes avx2 | expand |
| Context | Check | Description |
|---|---|---|
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
| andriy/make_ppc | success | Make finished |
| andriy/make_fate_ppc | success | Make fate finished |
Hello, >+pb_shuffle_low: times 4 db 1, 3, 5, 7, 9, 11, 13, 15, -1, -1, -1, -1, -1, -1, -1, -1 Why we times 4? AVX2 provided instruction VPBROADCASTQ to load these constant into SIMD register. Moreover, the plane U/V also apply same algorithm to get improve. Regards, Min Chen At 2021-09-30 09:56:11, "Wu Jianhua" <jianhua.wu@intel.com> wrote: >With the accelerating by means of AVX2, the uyvytoyuv422 can be faster > >Performance data(Less is better): > uyvytoyuv422_sse2 0.50388 > uyvytoyuv422_avx 0.46132 > uyvytoyuv422_avx2 0.27309 > >Signed-off-by: Wu Jianhua <jianhua.wu@intel.com> >--- > libswscale/x86/rgb2rgb.c | 6 ++++ > libswscale/x86/rgb_2_rgb.asm | 60 ++++++++++++++++++++++++++++-------- > 2 files changed, 53 insertions(+), 13 deletions(-) > >diff --git a/libswscale/x86/rgb2rgb.c b/libswscale/x86/rgb2rgb.c >index c9ff33ab77..a965a1755c 100644 >--- a/libswscale/x86/rgb2rgb.c >+++ b/libswscale/x86/rgb2rgb.c >@@ -164,6 +164,9 @@ void ff_uyvytoyuv422_sse2(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, > void ff_uyvytoyuv422_avx(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, > const uint8_t *src, int width, int height, > int lumStride, int chromStride, int srcStride); >+void ff_uyvytoyuv422_avx2(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, >+ const uint8_t *src, int width, int height, >+ int lumStride, int chromStride, int srcStride); > #endif > > >_______________________________________________ >ffmpeg-devel mailing list >ffmpeg-devel@ffmpeg.org >https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > >To unsubscribe, visit link above, or email >ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
Min Chen wrote: > Sent: Thursday, September 30, 2021 10:29 AM > To: FFmpeg development discussions and patches <ffmpeg- > devel@ffmpeg.org> > Subject: Re: [FFmpeg-devel] [PATCH v2 3/4] libswscale/x86/rgb2rgb: add > uyvytoyuv422 avx2 > > Hello, > > >+pb_shuffle_low: times 4 db 1, 3, 5, 7, 9, 11, 13, 15, -1, -1, -1, -1, > >+-1, -1, -1, -1 > Why we times 4? > AVX2 provided instruction VPBROADCASTQ to load these constant into SIMD > register. > > Moreover, the plane U/V also apply same algorithm to get improve. > > Regards, > Min Chen > Hi Min Chen, Much appreciated your helpful suggestions. Correct! It's not necessary to use time 4 here. It's funny that I did try to avoid using it here when writing the codes and get no way because I ignored the VBROADCASTI128 instruction. About the UV extracting, I have estimated the new method before making a decision to keep using the masterpiece of the previous author. The former is better, and pand instruction has a better reciprocal throughput, or issue latency. Best regards, Jianhua
At 2021-09-30 15:23:08, "Wu, Jianhua" <jianhua.wu@intel.com> wrote: >Min Chen wrote: >> Sent: Thursday, September 30, 2021 10:29 AM >> To: FFmpeg development discussions and patches <ffmpeg- >> devel@ffmpeg.org> >> Subject: Re: [FFmpeg-devel] [PATCH v2 3/4] libswscale/x86/rgb2rgb: add >> uyvytoyuv422 avx2 >> >> Hello, >> >> >+pb_shuffle_low: times 4 db 1, 3, 5, 7, 9, 11, 13, 15, -1, -1, -1, -1, >> >+-1, -1, -1, -1 >> Why we times 4? >> AVX2 provided instruction VPBROADCASTQ to load these constant into SIMD >> register. >> >> Moreover, the plane U/V also apply same algorithm to get improve. >> >> Regards, >> Min Chen >> >Hi Min Chen, > >Much appreciated your helpful suggestions. > >Correct! It's not necessary to use time 4 here. It's funny that I did try to avoid using it here >when writing the codes and get no way because I ignored the VBROADCASTI128 instruction. > >About the UV extracting, I have estimated the new method before making a decision to keep >using the masterpiece of the previous author. The former is better, and pand instruction has a better >reciprocal throughput, or issue latency. > >Best regards, >Jianhua For VBROADCASTI128, we don't care high part of result, so we just need lowest 64-bits constant table. VPBROADCASTQ enough. Regards, Min Chen
Min Chen wrote: > At 2021-09-30 15:23:08, "Wu, Jianhua" <jianhua.wu@intel.com> wrote: > >Min Chen wrote: > >> Sent: Thursday, September 30, 2021 10:29 AM > >> To: FFmpeg development discussions and patches <ffmpeg- > >> devel@ffmpeg.org> > >> Subject: Re: [FFmpeg-devel] [PATCH v2 3/4] libswscale/x86/rgb2rgb: > >> add > >> uyvytoyuv422 avx2 > >> > >> Hello, > >> > >> >+pb_shuffle_low: times 4 db 1, 3, 5, 7, 9, 11, 13, 15, -1, -1, -1, > >> >+-1, -1, -1, -1, -1 > >> Why we times 4? > >> AVX2 provided instruction VPBROADCASTQ to load these constant into > >> SIMD register. > >> > >> Moreover, the plane U/V also apply same algorithm to get improve. > >> > >> Regards, > >> Min Chen > >> > >Hi Min Chen, > > > >Much appreciated your helpful suggestions. > > > >Correct! It's not necessary to use time 4 here. It's funny that I did > >try to avoid using it here when writing the codes and get no way because I > ignored the VBROADCASTI128 instruction. > > > >About the UV extracting, I have estimated the new method before making > >a decision to keep using the masterpiece of the previous author. The > >former is better, and pand instruction has a better reciprocal throughput, or > issue latency. > > > >Best regards, > >Jianhua > > > > For VBROADCASTI128, we don't care high part of result, so we just need > lowest 64-bits constant table. VPBROADCASTQ enough. > > Definitely make sense. Thanks.
diff --git a/libswscale/x86/rgb2rgb.c b/libswscale/x86/rgb2rgb.c index c9ff33ab77..a965a1755c 100644 --- a/libswscale/x86/rgb2rgb.c +++ b/libswscale/x86/rgb2rgb.c @@ -164,6 +164,9 @@ void ff_uyvytoyuv422_sse2(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, void ff_uyvytoyuv422_avx(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src, int width, int height, int lumStride, int chromStride, int srcStride); +void ff_uyvytoyuv422_avx2(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, + const uint8_t *src, int width, int height, + int lumStride, int chromStride, int srcStride); #endif av_cold void rgb2rgb_init_x86(void) @@ -216,5 +219,8 @@ av_cold void rgb2rgb_init_x86(void) if (EXTERNAL_AVX(cpu_flags)) { uyvytoyuv422 = ff_uyvytoyuv422_avx; } + if (EXTERNAL_AVX2_FAST(cpu_flags)) { + uyvytoyuv422 = ff_uyvytoyuv422_avx2; + } #endif } diff --git a/libswscale/x86/rgb_2_rgb.asm b/libswscale/x86/rgb_2_rgb.asm index 3380a1272c..41ac91a747 100644 --- a/libswscale/x86/rgb_2_rgb.asm +++ b/libswscale/x86/rgb_2_rgb.asm @@ -26,14 +26,22 @@ SECTION_RODATA pb_mask_shuffle2103_mmx times 8 dw 255 +pb_shuffle_low: times 4 db 1, 3, 5, 7, 9, 11, 13, 15, -1, -1, -1, -1, -1, -1, -1, -1 pb_shuffle2103: db 2, 1, 0, 3, 6, 5, 4, 7, 10, 9, 8, 11, 14, 13, 12, 15 pb_shuffle0321: db 0, 3, 2, 1, 4, 7, 6, 5, 8, 11, 10, 9, 12, 15, 14, 13 pb_shuffle1230: db 1, 2, 3, 0, 5, 6, 7, 4, 9, 10, 11, 8, 13, 14, 15, 12 pb_shuffle3012: db 3, 0, 1, 2, 7, 4, 5, 6, 11, 8, 9, 10, 15, 12, 13, 14 pb_shuffle3210: db 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12 +pd_permd256_uv: dd 0, 4, 1, 5, 2, 6, 3, 7 SECTION .text +%macro VPERM 5 +%if mmsize == %2 + vperm%1 %3, %4, %5 +%endif +%endmacro + %macro RSHIFT_COPY 3 ; %1 dst ; %2 src ; %3 shift %if cpuflag(avx) @@ -198,11 +206,16 @@ cglobal uyvytoyuv422, 9, 14, 8, ydst, udst, vdst, src, w, h, lum_stride, chrom_s mov whalfq, wq shr whalfq, 1 ; whalf = width / 2 - lea srcq, [srcq + wq * 2] + lea srcq, [srcq + wq * 2] add ydstq, wq add udstq, whalfq add vdstq, whalfq +%if mmsize > 16 + movu m13, [pb_shuffle_low] + movu m15, [pd_permd256_uv] +%endif + .loop_line: mov xq, wq mov wtwoq, wq @@ -246,23 +259,35 @@ cglobal uyvytoyuv422, 9, 14, 8, ydst, udst, vdst, src, w, h, lum_stride, chrom_s movu m5, [srcq + wtwoq + mmsize * 3] ; extract y part 1 +%if mmsize < 32 RSHIFT_COPY m6, m2, 1 ; UYVY UYVY -> YVYU YVY... pand m6, m1; YxYx YxYx... RSHIFT_COPY m7, m3, 1 ; UYVY UYVY -> YVYU YVY... pand m7, m1 ; YxYx YxYx... - packuswb m6, m7 ; YYYY YYYY... +%else + pshufb m6, m2, m13 + pshufb m7, m3, m13 + punpcklqdq m6, m6, m7 + VPERM q, 32, m6, m6, 0xd8 +%endif movu [ydstq + wq], m6 ; extract y part 2 +%if mmsize < 32 RSHIFT_COPY m6, m4, 1 ; UYVY UYVY -> YVYU YVY... pand m6, m1; YxYx YxYx... RSHIFT_COPY m7, m5, 1 ; UYVY UYVY -> YVYU YVY... pand m7, m1 ; YxYx YxYx... - - packuswb m6, m7 ; YYYY YYYY... + packuswb m6, m7 ; YYYY YYYY... +%else + pshufb m6, m4, m13 + pshufb m7, m5, m13 + punpcklqdq m6, m6, m7 + VPERM q, 32, m6, m6, 0xd8 +%endif movu [ydstq + wq + mmsize], m6 ; extract uv @@ -275,17 +300,21 @@ cglobal uyvytoyuv422, 9, 14, 8, ydst, udst, vdst, src, w, h, lum_stride, chrom_s packuswb m4, m5 ; UVUV... ; U - pand m6, m2, m1 ; UxUx... - pand m7, m4, m1 ; UxUx... + pand m6, m2, m1 ; UxUx... + pand m7, m4, m1 ; UxUx... + packuswb m6, m7 ; UUUU - packuswb m6, m7 ; UUUU - movu [udstq + whalfq], m6 + VPERM d, 32, m6, m15, m6 + movu [udstq + whalfq], m6 ; V - psrlw m2, 8 ; VxVx... - psrlw m4, 8 ; VxVx... - packuswb m2, m4 ; VVVV + psrlw m2, 8 ; VxVx... + psrlw m4, 8 ; VxVx... + packuswb m2, m4 ; VVVV + + VPERM d, 32, m2, m15, m2 + movu [vdstq + whalfq], m2 add whalfq, mmsize @@ -294,13 +323,13 @@ cglobal uyvytoyuv422, 9, 14, 8, ydst, udst, vdst, src, w, h, lum_stride, chrom_s jl .loop_simd .end_line: - add srcq, src_strideq + add srcq, src_strideq add ydstq, lum_strideq add udstq, chrom_strideq add vdstq, chrom_strideq ;restore initial state of line variable - mov wq, back_wq + mov wq, back_wq mov xq, wq mov whalfq, wq shr whalfq, 1 ; whalf = width / 2 @@ -316,4 +345,9 @@ UYVY_TO_YUV422 INIT_XMM avx UYVY_TO_YUV422 + +%if HAVE_AVX2_EXTERNAL +INIT_YMM avx2 +UYVY_TO_YUV422 +%endif %endif
With the accelerating by means of AVX2, the uyvytoyuv422 can be faster Performance data(Less is better): uyvytoyuv422_sse2 0.50388 uyvytoyuv422_avx 0.46132 uyvytoyuv422_avx2 0.27309 Signed-off-by: Wu Jianhua <jianhua.wu@intel.com> --- libswscale/x86/rgb2rgb.c | 6 ++++ libswscale/x86/rgb_2_rgb.asm | 60 ++++++++++++++++++++++++++++-------- 2 files changed, 53 insertions(+), 13 deletions(-)