Details
Details
- Reviewers
b-sumner arsenm - Commits
- rGe668b1cd1e78: [AMDGPU][GFX9] Set +fp32-denormals for >=gfx900 unless -cl-denorms-are-zero is…
rC300306: [AMDGPU][GFX9] Set +fp32-denormals for >=gfx900 unless -cl-denorms-are-zero is…
rL300306: [AMDGPU][GFX9] Set +fp32-denormals for >=gfx900 unless -cl-denorms-are-zero is…
Diff Detail
Diff Detail
- Repository
- rL LLVM
Event Timeline
llvm/tools/clang/lib/Basic/Targets.cpp | ||
---|---|---|
2114–2116 ↗ | (On Diff #93420) | This is misleading since it was true on VI as well. I think just FMA rate changed |
llvm/tools/clang/lib/Basic/Targets.cpp | ||
---|---|---|
2114–2116 ↗ | (On Diff #93420) | Yes, GFX8 supports f32 denorms at full speed too. However, it doesn't have a full speed fma, so we didh't enable it then since it caused too many mad-heavy apps to slow down. |
llvm/tools/clang/lib/Basic/Targets.cpp | ||
---|---|---|
2114–2116 ↗ | (On Diff #93420) | Yes, so the name should refer to FMA rather than just fp32 denorms |
lib/Basic/Targets.cpp | ||
---|---|---|
2114 ↗ | (On Diff #93422) | Actually this also needs to specify full speed FMA with denorms. Full rate FMA is already a subtarget feature when denorms are disabled |
lib/Basic/Targets.cpp | ||
---|---|---|
2114 ↗ | (On Diff #93422) | I will do it in a separate patch. Thanks. |