Add both absolute and relative rocm paths to include clang headers

[hsharsha]

Motivation

On MI355 with new 7.1.1 and 7.2 docker images TF 2.18 fails to build with clang include path not found. Adding rocm relative path /opt/rocm is not sufficient as it looks for absolute path as well.

Technical Details

Add both relative and absolute paths for clang headers to be included

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[i-chaochen]

Seems CI still has these build errors?

[2026-03-13T01:22:02.471Z] //tensorflow/compiler/mlir/tools/kernel_gen/tests:buffer_deallocation.mlir.test FAILED TO BUILD
[2026-03-13T01:22:02.471Z] //tensorflow/compiler/mlir/tools/kernel_gen/tests:buffer_reuse.mlir.test FAILED TO BUILD
[2026-03-13T01:22:02.471Z] //tensorflow/compiler/mlir/tools/kernel_gen/tests:bufferize.mlir.test FAILED TO BUILD
[2026-03-13T01:22:02.471Z] //tensorflow/compiler/mlir/tools/kernel_gen/tests:copy_cleanup.mlir.test FAILED TO BUILD
[2026-03-13T01:22:02.471Z] //tensorflow/compiler/mlir/tools/kernel_gen/tests:embed_tf_framework.mlir.test FAILED TO BUILD
[2026-03-13T01:22:02.471Z] (Skipping other failed to build tests)

[hsharsha]

Failures are due to

# Configuration: 220b3b198b74c1f779fd5797602095dbe99c565ac10f4f60431d7f246d347cb6
# Execution platform: @local_execution_config_platform//:platform
external/local_xla/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc:963:10: fatal error: 'amdgpu_device_lib_data.inc' file not found
  963 | #include "amdgpu_device_lib_data.inc"
      |   

related to embedded device libs. Looks like data is not generated.

[hsharsha]

@alekstheod Looks like r2.18 still has a mix of gcc and clang support. So -no-canonical-prefixes doesn't seems to work.

[hsharsha]

This happens because config=rocm is not passed to cpu tests

[i-chaochen]

# Configuration: 220b3b198b74c1f779fd5797602095dbe99c565ac10f4f60431d7f246d347cb6
# Execution platform: @local_execution_config_platform//:platform
external/local_xla/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc:963:10: fatal error: 'amdgpu_device_lib_data.inc' file not found
  963 | #include "amdgpu_device_lib_data.inc"
      |   

we shouldn't not run these gpu related tests in our CPU pipeline, isn't? I thought we already fixed this issue before?

@mmakevic-amd does it ring any bell?

[hsharsha]

# Configuration: 220b3b198b74c1f779fd5797602095dbe99c565ac10f4f60431d7f246d347cb6
# Execution platform: @local_execution_config_platform//:platform
external/local_xla/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc:963:10: fatal error: 'amdgpu_device_lib_data.inc' file not found
  963 | #include "amdgpu_device_lib_data.inc"
      |   

we shouldn't not run these gpu related tests in our CPU pipeline, isn't? I thought we already fixed this issue before?

@mmakevic-amd does it ring any bell?

These were marked manual and the code is refactored in r2.19 and above into amdgpu_backend.cc. For r2.18 it is still part of llvm_backend_lib.cc. This can be fixed for r2.18 passing config=rocm so that if_rocm_configured path in bazel is taken.