commit e03c5517bbd3409c079c78c193f94ac608d75713
author Deven Desai <deven.desai.amd@gmail.com> Thu Mar 11 13:46:26 2021 +0000
committer Frederic Bastien <fbastien@nvidia.com> Mon Mar 15 10:57:18 2021 -0700
tree 79f8b3c42e94ded8200659be6aa781bde368c79b
parent ddbc68f5598d7400cef45ad4b4f8d52ddd51de87

[ROCm] Fix for ROCm CSB breakage - 210311

The following commit breaks the ROCm CSB (errors while running GPU unit tests)

https://github.com/tensorflow/tensorflow/pull/47398

```
ERROR: /root/tensorflow/tensorflow/compiler/xla/service/gpu/tests/BUILD:573:13: undeclared inclusion(s) in rule '//tensorflow/compiler/xla/service/gpu/tests:hlo_to_llvm_ir':
this rule is missing dependency declarations for the following files included by 'tensorflow/compiler/xla/service/gpu/tests/hlo_to_llvm_ir.cc':
  'tensorflow/compiler/xla/service/gpu/nvptx_compiler.h'
Target //tensorflow/compiler/xla/service/gpu/tests:constant.hlo.test failed to build
```

This PR/commit fixes the ROCm build, by making the newly added functionality specific to the TF CUDA build.

Long term, we need to add similar capability to the ROCM side, but for now this workaround is needed to get the ROCm CSB working again.

tensorflow/compiler/xla/service/gpu/tests/hlo_to_llvm_ir.cc

diff --git a/tensorflow/compiler/xla/service/gpu/tests/hlo_to_llvm_ir.cc b/tensorflow/compiler/xla/service/gpu/tests/hlo_to_llvm_ir.cc
index 47f18c8..7a36011 100644
--- a/tensorflow/compiler/xla/service/gpu/tests/hlo_to_llvm_ir.cc
+++ b/tensorflow/compiler/xla/service/gpu/tests/hlo_to_llvm_ir.cc
@@ -17,7 +17,9 @@
 #include "tensorflow/compiler/xla/service/gpu/gpu_compiler.h"
 #include "tensorflow/compiler/xla/service/gpu/gpu_device_info.h"
 #include "tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.h"
+#if GOOGLE_CUDA
 #include "tensorflow/compiler/xla/service/gpu/nvptx_compiler.h"
+#endif
 #include "tensorflow/compiler/xla/service/gpu/target_constants.h"
 #include "tensorflow/compiler/xla/service/hlo_module.h"
 #include "tensorflow/compiler/xla/status.h"
@@ -77,6 +79,7 @@
   if (!generate_ptx) {
     llvm_module->print(llvm::outs(), nullptr);
   } else {
+#if GOOGLE_CUDA
     std::pair<int, int> gpu_version = std::make_pair(
         cuda_compute_capability.cc_major, cuda_compute_capability.cc_minor);
     std::string libdevice_dir = xla::gpu::GetLibdeviceDir(hlo_module->config());
@@ -85,6 +88,10 @@
         xla::gpu::nvptx::CompileToPtx(llvm_module.get(), gpu_version,
                                       hlo_module->config(), libdevice_dir));
     std::cout << ptx << std::endl;
+#else
+    return xla::Status(tensorflow::error::UNIMPLEMENTED,
+                       "Feature not yet implemented in ROCm");
+#endif
   }
   return xla::Status::OK();
 }