Avx512: Abs, ceil, floor, min, max #84937
Conversation
ghost
added
the
area-CodeGen-coreclr
# Conflicts: # src/coreclr/jit/hwintrinsiclistxarch.h # src/coreclr/jit/instrsxarch.h
Removing unnecessary HW_Flag_NoEvexSemantics flag Co-authored-by: Tanner Gooding <tagoo@outlook.com>
Co-authored-by: Tanner Gooding <tagoo@outlook.com>
Co-authored-by: Tanner Gooding <tagoo@outlook.com>
Co-authored-by: Tanner Gooding <tagoo@outlook.com>
Co-authored-by: Tanner Gooding <tagoo@outlook.com>
# Conflicts: # src/coreclr/jit/hwintrinsiclistxarch.h
|
Resolved the merge conflict caused by other merged PRs |
There was a problem hiding this comment.
Actually, on second thought, remembered that there is still a need to fixup the roundps and roundpd instructions in instrsxarch.h as they are still marked as EVEX incompatible.
They do actually, in that they become the vrndscaleps/vrndscalepd instructions and the upper 4-bits of the immediate can now be non-zero.
That also means we probably need to update instr.cpp to ensure that roundps/roundpd are emitted as vrndscale in the disassembly
Without this fix, Ceiling and Floor are going to fail in their tests
-- The scalar versions, roundss and roundsd probably need the same treatment.
ghost
added
the
needs-author-action
Co-authored-by: Tanner Gooding <tagoo@outlook.com>
ghost
removed
the
needs-author-action
We have tried to add vrndscaleps/vrndscalepd instructions. hope this will find the round instruction |
… the code gen part with EVEX encoding.
Co-authored-by: Tanner Gooding <tagoo@outlook.com>
|
for some weird reason, git + was added to the code change. making an additional push to remove that. |
|
CC. @dotnet/jit-contrib, @dotnet/avx512-contrib this needs a secondary sign-off from the JIT side |
3 participants
Adding support for AVX 512:
Abs, ceil, floor, min, max, sqrt, Negate, and Unary Addition.
Covered cases:
Abs:
Abs(Vector512), Abs(Vector512), Abs(Vector512), Abs(Vector512), Abs(Vector512)
, Abs(Vector512)
ceil:
ceil(Vector512), ceil(Vector512)
Floor:
Floor(Vector512), Floor(Vector512)
Min
Min(Vector512, Vector512), Min(Vector512, Vector512),
Min(Vector512, Vector512), Min(Vector512, Vector512),
Min(Vector512, Vector512), Min(Vector512, Vector512)
Max
Max(Vector512, Vector512), Max(Vector512, Vector512),
Max(Vector512, Vector512), Max(Vector512, Vector512),
Max(Vector512, Vector512), Max(Vector512, Vector512)
Negate & Unary Addition:
Negate(Vector512), Negate(Vector512), Negate(Vector512), Negate(Vector512), Negate(Vector512), Negate(Vector512)
Operator for Negate and Unary Addition.
Introduced a few new instructions that were not supported by lower vector intrinsics: