[FFmpeg-devel] swscale/rgb2rgb : add X86_64 SIMD (SSSE3 and AVX2) for shuffly_bytes func

[Martin Vignali]

2018-03-18 16:49 GMT+01:00 James Almer <jamrial at gmail.com>:

> On 3/18/2018 12:08 PM, Martin Vignali wrote:
> > 2018-03-03 18:20 GMT+01:00 Martin Vignali <martin.vignali at gmail.com>:
> >
> >> Hello,
> >>
> >> Patch in attach add SIMD for the 5 shuffle_bytes func for rgb2rgb
> >> The new SIMD are write using external ASM.
> >>
> >> Also add checkasm test for theses func
> >> Restricted to x86_64, because the scalar part doesn't compile on x86_32
> >>
> >> I consider for the scalar part that the src_size value is a multiple of
> 4
> >> (because the shuffle is for 4 bytes)
> >>
> >> Pass fate test on X86_64 and X86_32 (os 10.12)
> >>
> >>
> >>
> >>
> >> New patchs in attach :
> > - Now compile on x86_32 and x86_64
> > - Add cosmetic patch to put all shuffle_bytes declaration in the same
> place
> >
> > Tested on X86_64 and X86_32 (os 10.12)
> >
> > Checkasm result :  ./tests/checkasm/checkasm --test=sw_rgb --bench
> >
> > checkasm: using random seed 292997963
> > MMX:
> >  - sw_rgb.shuffle_bytes_2103 [OK]
> > MMXEXT:
> >  - sw_rgb.shuffle_bytes_2103 [OK]
> > SSSE3:
> >  - sw_rgb.shuffle_bytes_2103 [OK]
> >  - sw_rgb.shuffle_bytes_0321 [OK]
> >  - sw_rgb.shuffle_bytes_1230 [OK]
> >  - sw_rgb.shuffle_bytes_3012 [OK]
> >  - sw_rgb.shuffle_bytes_3210 [OK]
> > AVX2:
> >  - sw_rgb.shuffle_bytes_2103 [OK]
> >  - sw_rgb.shuffle_bytes_0321 [OK]
> >  - sw_rgb.shuffle_bytes_1230 [OK]
> >  - sw_rgb.shuffle_bytes_3012 [OK]
> >  - sw_rgb.shuffle_bytes_3210 [OK]
> > checkasm: all 12 tests passed
> > shuffle_bytes_0321_c: 51.4
> > shuffle_bytes_0321_ssse3: 18.7
> > shuffle_bytes_0321_avx2: 12.7
> > shuffle_bytes_1230_c: 126.9
> > shuffle_bytes_1230_ssse3: 16.7
> > shuffle_bytes_1230_avx2: 12.9
> > shuffle_bytes_2103_c: 52.4
> > shuffle_bytes_2103_mmx: 76.7
> > shuffle_bytes_2103_mmxext: 197.2
> > shuffle_bytes_2103_ssse3: 17.4
> > shuffle_bytes_2103_avx2: 12.4
> > shuffle_bytes_3012_c: 127.4
> > shuffle_bytes_3012_ssse3: 14.7
> > shuffle_bytes_3012_avx2: 12.4
> > shuffle_bytes_3210_c: 127.4
> > shuffle_bytes_3210_ssse3: 18.2
> > shuffle_bytes_3210_avx2: 12.9
>
> These AVX2 numbers are not worth it. Some CPU archs throttle down the
> frequency when using ymm instructions, so unless the function is
> considerably faster than the SSE* version then it's usually not worth
> adding.
>
>
I run the test again with a bigger width (512 instead of 128)
This is my result :
shuffle_bytes_0321_c: 128.6
shuffle_bytes_0321_ssse3: 41.6
shuffle_bytes_0321_avx2: 23.4
shuffle_bytes_1230_c: 626.4
shuffle_bytes_1230_ssse3: 41.6
shuffle_bytes_1230_avx2: 23.9
shuffle_bytes_2103_c: 128.4
shuffle_bytes_2103_mmx: 307.1
shuffle_bytes_2103_mmxext: 224.6
shuffle_bytes_2103_ssse3: 72.9
shuffle_bytes_2103_avx2: 32.9
shuffle_bytes_3012_c: 620.9
shuffle_bytes_3012_ssse3: 40.6
shuffle_bytes_3012_avx2: 36.1
shuffle_bytes_3210_c: 602.6
shuffle_bytes_3210_ssse3: 75.4
shuffle_bytes_3210_avx2: 33.6


So except for the 3012 version (don't know why), we are around x2 in AVX2.
Do you still think, it's need to remove AVX2 version ?


Martin