Upgrading Vector256/512 Shuffle() with VBMI support #87083
Conversation
dotnet-issue-labeler bot added the area-CodeGen-coreclr 
ghost added the community-contribution 
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsUpgrading Vector256/512 Shuffle() with VBMI support
|
DeepakRajendrakumaran changed the title
DeepakRajendrakumaran force-pushed the MRT143 branch 2 times, most recently from bd850a2 to 5885bac Compare
DeepakRajendrakumaran force-pushed the MRT143 branch 2 times, most recently from 4bdbd2a to 8eecf8f Compare
I can't figure out if the CI fails are known issues or build specific Looks like infrastructure from this |
There's currently an infrastructure issue causing a large number of CI machine timeouts. It's likely not related to your PR. |
Please let me know you there is anything else to be done on this to be merged. |
| @@ -23482,6 +23509,15 @@ GenTree* Compiler::gtNewSimdShuffleNode( | |||
| // swap the operands to match the encoding requirements | |||
| retNode = gtNewSimdHWIntrinsicNode(type, op2, op1, NI_AVX512BW_PermuteVar32x16, simdBaseJitType, simdSize); | |||
| } | |||
| else if (elementSize == 1) | |||
There was a problem hiding this comment.
There's potentially a missing "todo" here to handle this using pshufb if we don't cross lanes and VBMI is not available.
I don't think it needs to be handled as part of this PR (or even for .NET 8 necessarily) since that scenario should only be limited to pre-cannon lake (and therefore first generation AVX-512 hardware). But, since it should just be refactoring the logic under the if (simdSize == 32) path to be shared; its probably worth having still.
CC. @dotnet/jit-contrib, @dotnet/avx512-contrib for secondary sign-off. |
DeepakRajendrakumaran force-pushed the MRT143 branch from 3a2e180 to 17fb7dc Compare
This change does the following
1: Accelerates Vector512.Shuffle() using AVX512VBMI(permb)
2. Accelerates Vector256.Shuffle() using AVX512VBMI_VL
3. Accelerates Vector256.Shuffle() using AVX512BW_VL
@dotnet/avx512-contrib