experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Remove "+code-object-v3" from the target feature_string used during XLA codegen.

Browse Source 
Starting with ROCm 4.0, "+code-object-v3" is no longer recognized as a valid feature-string from ADGPU targets, and result in the following warning for each HSACO codegen

```
'+code-object-v3' is not a recognized feature for this target (ignoring feature)
```

Since the hipclang compiler generates code-object-v3 objects by default now, we no longer need to the explicitly specify "+code-object-v3" as a feature-string. Removing taht feature string also gets rid of the warning.
This commit is contained in:
Deven Desai 2020-11-30 18:36:40 +00:00
parent 358cbc9e53
commit 80548e833f
1 changed files with 9 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc
View File

@ -800,8 +800,16 @@ Status AMDGPUTargetModuleLinker(llvm::Module* module, GpuVersion gpu_version,
std::unique_ptr<llvm::TargetMachine> AMDGPUGetTargetMachine(
llvm::Triple target_triple, int amdgpu_version,
const HloModuleConfig& hlo_module_config) {
string feature_str = "+code-object-v3";
#if TF_ROCM_VERSION >= 30900
// code-object-v3 is default, so no need to expliticitly specify it
// in the feature string. Also, starting with ROCm 4.0, this feature string
// is deprecated, and we get a warning to that effect. So removing that
// feature string
feature_str = "";
#endif
return GetTargetMachine(target_triple, absl::StrCat("gfx", amdgpu_version),
hlo_module_config, "+code-object-v3");
hlo_module_config, feature_str);
}
void AMDGPUBackendInit(const HloModuleConfig& hlo_module_config) {