[FFmpeg-devel] libavcodec/huffyuvdsp(enc) : add avx2 version for add(diff)_int16

[Martin Vignali]

2017-10-22 0:26 GMT+02:00 Martin Vignali <martin.vignali at gmail.com>:

> Hello,
>
> In attach patch to add avx2 version for huffyuv dsp and huffyuvdsp enc
> for add_int16 and diff_int16 func
>
> Check asm result for add_int16 (Kaby Lake, os 10.12)
> add_int16_128_c: 1607.9
> add_int16_128_sse2: 442.7
> add_int16_128_avx2: 218.9
>
> Pass fate test for me
>
>
> 0001-checkasm-add-test-for-huffyuvdsp-add_int16 :
> add a checkasm test for add_int16
> base on lossless_videodsp checkasm test
>
> i add a test with a fix size, to make speed test more easy to compare
>
> 0002-libavcodec-huffyuvdsp-enc-move-duplicate-macro-to-a-
> huffyuvdsp.asm and huffyuvdspenc.asm use the same INT16_LOOP macro
> with arg add for dec and sub for encoder
>
> this patch move this macro in an asm file in order to be share by both dsp
> asm
>
> 0003-libavcodec-huffyuvdsp-reorganize-add_int16-asm
> 0005-libavcodec-huffyuvdspenc-reorganize-diff_int16
> Code reorganization
>
>
> 0004-libavcodec-huffyuvdsp-add-add_int16-AVX2-func
> 0006-libavcodec-huffyuvdspenc-add-diff_int16-AVX2-func
> AVX2 version for each func
>
>
>
> ping