diff options
Diffstat (limited to 'llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h')
| -rw-r--r-- | llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h | 20 |
1 files changed, 20 insertions, 0 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h index 6878744496cf..57b757c990e1 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h +++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h @@ -100,6 +100,26 @@ public: /// be converted to integer, or violate subtarget's specifications. std::pair<unsigned, unsigned> getFlatWorkGroupSizes(const Function &F) const; + /// \returns The required size of workgroups that will be used to execute \p F + /// in the \p Dim dimension, if it is known (from `!reqd_work_group_size` + /// metadata. Otherwise, returns std::nullopt. + std::optional<unsigned> getReqdWorkGroupSize(const Function &F, + unsigned Dim) const; + + /// \returns true if \p F will execute in a manner that leaves the X + /// dimensions of the workitem ID evenly tiling wavefronts - that is, if X / + /// wavefrontsize is uniform. This is true if either the Y and Z block + /// dimensions are known to always be 1 or if the X dimension will always be a + /// power of 2. If \p RequireUniformYZ is true, it also ensures that the Y and + /// Z workitem IDs will be uniform (so, while a (32, 2, 1) launch with + /// wavesize64 would ordinarily pass this test, it won't with + /// \pRequiresUniformYZ). + /// + /// This information is currently only gathered from the !reqd_work_group_size + /// metadata on \p F, but this may be improved in the future. + bool hasWavefrontsEvenlySplittingXDim(const Function &F, + bool REquiresUniformYZ = false) const; + /// \returns Subtarget's default pair of minimum/maximum number of waves per /// execution unit for function \p F, or minimum/maximum number of waves per /// execution unit explicitly requested using "amdgpu-waves-per-eu" attribute |