Ensure that the various Min and Max APIs are accelerated where possible #87641
Conversation
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsThis mostly does AVX-512 based acceleration but it also ensures that these are supported on Arm64 where that's trivial to do.
|
|
|
||
| #if defined(FEATURE_HW_INTRINSICS) && defined(TARGET_XARCH) | ||
| case NI_System_Math_Max: | ||
| case NI_System_Math_Min: |
There was a problem hiding this comment.
This logic was all extracted to a new helper function impMinMaxIntrinsic, to avoid code duplication.
There was a problem hiding this comment.
Doing so made the diff unhappy, and so it's showing a bunch of code as "new" where it actually isn't
4cce32d to
0e6353c
Compare
|
CC. @dotnet/jit-contrib, @dotnet/avx512-contrib As per the top post this ensures various Min/Max APIs are accelerated using available instructions on both Arm64 and x86/x64. For the latter it utilizes AVX-512 where possible. Diffs look good with us saving 16k bytes of codegen on Linux x64 and 29k bytes of codegen on Windows x64 (significantly simpler code that is now branch-free). Cleaning up the intrinsic recognition path also brings a |
This mostly does AVX-512 based acceleration but it also ensures that these are supported on Arm64 where that's trivial to do.