Avx512 avx512vl #999
Avx512 avx512vl #999
Conversation
|
r? @Amanieu (rust-highfive has picked a reviewer for you, use r? to override) |
crates/core_arch/src/x86/avx512f.rs
Outdated
| }; | ||
| } | ||
| let r = constify_imm8_gather!(scale, call); | ||
| transmute(simd_select_bitmask(mask, r.as_f32x8(), src.as_f32x8())) |
There was a problem hiding this comment.
simd_select_bitmask is incorrect in this case. We need to use the special LLVM intrinsic llvm.x86.avx512.mask.gather3siv4.sf.
The difference here is that the LLVM intrinsic will only access memory addresses which are not masked out. However your version will access memory addresses that are masked out, which could cause a crash if a masked value is an invalid offset.
There was a problem hiding this comment.
simd_select_bitmaskis incorrect in this case. We need to use the special LLVM intrinsicllvm.x86.avx512.mask.gather3siv4.sf.The difference here is that the LLVM intrinsic will only access memory addresses which are not masked out. However your version will access memory addresses that are masked out, which could cause a crash if a masked value is an invalid offset.
True, Thanks for your remind.
llvm.x86.avx512.mask.gather3siv4.sf will requires i1, so I need to remove those mask gather functions
permute_ps,pd: mm256,mm; permutevar_ps,pd: mm256,mm
permutex_epi64,pd: mm256
permutexvar_epi32,epi64,ps,pd: mm256
permutex2var_epi32,epi64,ps,pd: mm256,mm
i32gather_epi32,epi64,ps,pd: mm256,mm
i64gather_epi32,epi64,ps,pd: mm256,mm