shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity

[OpenMP] Remove 'minncta' attributes from NVPTX kernels (#88398)

Browse Source 
Summary:
Currently we treat this attribute as a minimum number for the amount of
blocks scheduled on the kernel. However, the doucmentation states that
this applies to CTA's mapped onto a *single* SM. Currently we just set
it to the total number of blocks, which will almost always result in a
warning that the value is out of range and will be ignored. We don't
have a good way to automatically know how many CTAs can be put on a
single SM nor if we should do this, so we should probably leave this up
to users manually adding it.


https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#performance-tuning-directives-minnctapersm
This commit is contained in:
Joseph Huber 2024-04-15 15:28:05 -05:00 committed by GitHub
parent 071ac0ae30
commit 0287a5cc4e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 2 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
clang/test/OpenMP/ompx_attributes_codegen.cpp
View File

@ -36,6 +36,5 @@ void func() {
// NVIDIA: "omp_target_thread_limit"="45"
// NVIDIA: "omp_target_thread_limit"="17"
// NVIDIA: !{ptr @__omp_offloading[[HASH1:.*]]_l16, !"maxntidx", i32 20}
// NVIDIA: !{ptr @__omp_offloading[[HASH2:.*]]_l18, !"minctasm", i32 90}
// NVIDIA: !{ptr @__omp_offloading[[HASH2]]_l18, !"maxntidx", i32 45}
// NVIDIA: !{ptr @__omp_offloading[[HASH2:.*]]_l18, !"maxntidx", i32 45}
// NVIDIA: !{ptr @__omp_offloading[[HASH3:.*]]_l20, !"maxntidx", i32 17}

4
llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
View File

@ -4786,11 +4786,9 @@ OpenMPIRBuilder::readTeamBoundsForKernel(const Triple &, Function &Kernel) {
void OpenMPIRBuilder::writeTeamsForKernel(const Triple &T, Function &Kernel,
int32_t LB, int32_t UB) {
if (T.isNVPTX()) {
if (T.isNVPTX())
if (UB > 0)
updateNVPTXMetadata(Kernel, "maxclusterrank", UB, true);
updateNVPTXMetadata(Kernel, "minctasm", LB, false);
}
if (T.isAMDGPU())
Kernel.addFnAttr("amdgpu-max-num-workgroups", llvm::utostr(LB) + ",1,1");