================
@@ -310,24 +312,31 @@ int main(int argc, const char **argv) {
IEB->SlabAllocateSize = *SizeOrErr;
IEB->UseSharedMemory = UseSharedMemory;
- std::unique_ptr<clang::CompilerInstance> DeviceCI;
- if (CudaEnabled) {
- if (!CudaPath.empty())
- CB.SetCudaSDK(CudaPath);
+ if (HipEnabled && CudaEnabled) {
+ if (HipEnabled.getPosition() > CudaEnabled.getPosition())
+ CudaEnabled = false;
+ else
+ HipEnabled = false;
+ }
- if (OffloadArch.empty()) {
- OffloadArch = "sm_35";
- }
- CB.SetOffloadArch(OffloadArch);
+ bool DeviceEnabled = HipEnabled || CudaEnabled;
+ llvm::StringRef DevicePath = HipEnabled ? RocmPath : CudaPath;
+ llvm::StringRef DeviceOffloadArch = !OffloadArch.empty()
+ ? llvm::StringRef(OffloadArch)
+ : (HipEnabled ? "gfx90a" : "sm_35");
----------------
yxsamliu wrote:
Hard-coding `gfx90a` will fail on other AMD GPUs. Could HIP use
`--offload-arch=native` by default? The driver already detects the architecture
with `amdgpu-arch`. If it finds no GPU or multiple architectures, Clang-Repl
can report an error since it currently supports one device architecture.
https://github.com/llvm/llvm-project/pull/217582
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
