Implement AVX512f floating point comparisons #869
Conversation
Co-authored-by: bjorn3 <[email protected]>
|
☔ The latest upstream changes (presumably b17efd8) made this pull request unmergeable. Please resolve the merge conflicts. |
Daniel-B-Smith
force-pushed
the
fp-cmp-512
branch
from
d110c09 to
f567b69
Compare
Daniel-B-Smith
changed the title
|
This should be ready for review. There's one remaining rough edge here. If you pass an invalid value to |
|
It seems that not all values are valid for the |
|
Done. Thanks! I tried to find where/how the sae parameter was validated, but I struggled to find it. |
crates/core_arch/src/x86/avx512f.rs
Outdated
| vcmpps(a.as_f32x16(), b.as_f32x16(), $imm5, neg_one, $imm4) | ||
| }; | ||
| } | ||
| let r = constify_imm5_sae!(op, _MM_FROUND_CUR_DIRECTION, call); |
There was a problem hiding this comment.
Why not just use constify_imm5 here and hard-code the rounding mode in the vcmpps call above? This code creates many unnecessary match branches.
There was a problem hiding this comment.
Just a bad refactoring. I had it like that originally. Should be fixed now.
|
CI is fixed now. |
|
Thanks! |
This adds the AVX 512f floating point comparison intrinsics. This should finish all of the AVX 512f comparison intrinsics.