[FFmpeg-devel] [PATCH] avcodec/x86/vp9: Add AVX-512ICL for 16x16 and 32x32 8bpc inverse transforms

[FFmpeg-devel] [PATCH] avcodec/x86/vp9: Add AVX-512ICL for 16x16 and 32x32 8bpc inverse transforms

[Henrik Gramner via ffmpeg-devel]

[-- Attachment #1: Type: text/plain, Size: 911 bytes --]

Placed in a new separate file as the existing combined MMX/SSE/AVX
file is humongous and takes forever to assemble as is.

This adds ~16 KiB of .text. The existing 8bpc asm is ~240 KiB of which
the corresponding AVX2 functions makes up ~42 KiB.

Tested to pass FATE on Linux and Windows.

Checkasm numbers vs AVX2 on Zen 5 (Strix Halo):
  vp9_inv_adst_adst_16x16_sub16_add_8_avx2:        209.3
  vp9_inv_adst_adst_16x16_sub16_add_8_avx512icl:    99.5

  vp9_inv_adst_dct_16x16_sub16_add_8_avx2:         165.2
  vp9_inv_adst_dct_16x16_sub16_add_8_avx512icl:     89.7

  vp9_inv_dct_adst_16x16_sub16_add_8_avx2:         165.9
  vp9_inv_dct_adst_16x16_sub16_add_8_avx512icl:     87.7

  vp9_inv_dct_dct_16x16_sub16_add_8_avx2:          121.3
  vp9_inv_dct_dct_16x16_sub16_add_8_avx512icl:      79.2

  vp9_inv_dct_dct_32x32_sub32_add_8_avx2:          745.5
  vp9_inv_dct_dct_32x32_sub32_add_8_avx512icl:     285.5

[-- Attachment #2: vp9_itx_avx512.patch --]
[-- Type: application/octet-stream, Size: 73066 bytes --]

[-- Attachment #3: Type: text/plain, Size: 251 bytes --]

_______________________________________________
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".

Re: [FFmpeg-devel] [PATCH] avcodec/x86/vp9: Add AVX-512ICL for 16x16 and 32x32 8bpc inverse transforms

[Henrik Gramner via ffmpeg-devel]

On Sat, May 17, 2025 at 12:59 AM Henrik Gramner <henrik@gramner.com> wrote:
>
> Placed in a new separate file as the existing combined MMX/SSE/AVX
> file is humongous and takes forever to assemble as is.
>
> This adds ~16 KiB of .text. The existing 8bpc asm is ~240 KiB of which
> the corresponding AVX2 functions makes up ~42 KiB.
>
> Tested to pass FATE on Linux and Windows.

Pushed.
_______________________________________________
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".