AVX512_changes

[deeptiag1]

Signed-off-by: deeptiag1 [email protected]

[1480c1]

@deeptiag1, you may want to check your git config, it seems you might have misspelled your email, unless your email domain is intel.comm

[tianjunwork]

Hi @deeptiag1 , since AVX512 code doesn't execute when running encoder, our CI can't validate your code. Could you let us know what tests you did and the result? Thank you very much for your contribution. Could you also give a better title to this PR which be more specific. Thank you.

[deeptiag1]

Hi @tianjunwork, Enable the AVX512 flag in Ebdefinitions and then run the same test as one would run for any build. Let me know if all test pass.

[intelmark]

From @hassount 3/17/2020:

From the results below, it seems that we need to disable the optimization for the LumaInterpolationFilterOneDOutRawHorizontal_AVX512 kernel and make sure that the documentation contains a clear reference to the gcc version requirement when vnni is on, then I don’t see any blockers to this PR moving forward assuming functional testing is passing._

The current PR shows:

#ifdef VNNI_SUPPORT #ifndef NON_AVX512_SUPPORT #define LumaInterpolationFilterOneDOutRawHorizontalLumaInterpolationFilterOneDOutRawHorizontal_AVX512 #else #define LumaInterpolationFilterOneDOutRawHorizontalLumaInterpolationFilterOneDOutRawHorizontal_SSSE3 #endif

But isn't VNNI a subset of AVX512 - Are all combinations of the VNNI_SUPPORT and AVX512_SUPPORT defines possible?

Also there doesn't seem to be a mention of the gcc requirements for compiling the VNNI specific code anywhere in the PR

[deeptiag1]

#define vnni in EbDefinitions enable vnni code committed in this PR.

Perf Data; Encoder mode :0 reference_avx512 : branch vnni_changes_1 @ d92a7c4 vnni_avx512: reference_avx512 + VNNI_SUPPORT + LumaInterpolationFilterOneDOutRawHorizontal_SSSE3 speed reference_avx512 | 0.95fps vnni_avx512 | 0.96fps Speed gain: 1%

Encoder mode :0 reference_avx2 : branch vnni_changes_1 @ d92a7c4 + NON_AVX512_SUPPORT vnni_avx2: reference_avx2 + VNNI_SUPPORT + LumaInterpolationFilterOneDOutRawHorizontal_SSSE3 speed reference_avx2 | 0.96fps vnni_avx2 | 0.97fps Speed gain: 1%

[intelmark]

@deeptiag1 Regarding the performance data - great to see the speed up, but what size video and is there a command line that was used?