[ROCm] Remove usage of deprecated ROCm component header includes (#97620)

- clang parameter 'amdgpu-target' changed to 'offload-arch'
- HIP and MIOpen includes path updated for extensions

Pull Request resolved: https://github.com/pytorch/pytorch/pull/97620
Approved by: https://github.com/ezyang, https://github.com/jithunnair-amd

cmake/Dependencies.cmake

@@ -1288,7 +1288,7 @@ if(USE_ROCM)
     # host linker to link.
     list(APPEND HIP_CLANG_FLAGS -fno-gpu-rdc)
     foreach(pytorch_rocm_arch ${PYTORCH_ROCM_ARCH})
-      list(APPEND HIP_CLANG_FLAGS --amdgpu-target=${pytorch_rocm_arch})
+      list(APPEND HIP_CLANG_FLAGS --offload-arch=${pytorch_rocm_arch})
     endforeach()
 
     set(Caffe2_HIP_INCLUDE

cmake/public/LoadHIP.cmake

@@ -313,5 +313,4 @@ if(HIP_FOUND)
   find_library(ROCM_HIPRTC_LIB ${hip_library_name} HINTS ${HIP_PATH}/lib)
   # roctx is part of roctracer
   find_library(ROCM_ROCTX_LIB roctx64 HINTS ${ROCTRACER_PATH}/lib)
-  set(roctracer_INCLUDE_DIRS ${ROCTRACER_PATH}/include)
 endif()

torch/CMakeLists.txt

@@ -139,7 +139,6 @@ if(USE_ROCM)
       __HIP_PLATFORM_HCC__
       )
     list(APPEND TORCH_PYTHON_LINK_LIBRARIES ${ROCM_ROCTX_LIB})
-    list(APPEND TORCH_PYTHON_INCLUDE_DIRECTORIES ${roctracer_INCLUDE_DIRS})
 endif()
 
 if(USE_EXPERIMENTAL_CUDNN_V8_API)

torch/utils/cpp_extension.py

@@ -200,7 +200,6 @@ CUDA was not found on the system, please set the CUDA_HOME or the CUDA_PATH
 environment variable or add NVCC to your system PATH. The extension compilation will fail.
 '''
 ROCM_HOME = _find_rocm_home()
-MIOPEN_HOME = _join_rocm_home('miopen') if ROCM_HOME else None
 HIP_HOME = _join_rocm_home('hip') if ROCM_HOME else None
 IS_HIP_EXTENSION = True if ((ROCM_HOME is not None) and (torch.version.hip is not None)) else False
 ROCM_VERSION = None
@@ -1142,10 +1141,6 @@ def include_paths(cuda: bool = False) -> List[str]:
     if cuda and IS_HIP_EXTENSION:
         paths.append(os.path.join(lib_include, 'THH'))
         paths.append(_join_rocm_home('include'))
-        if MIOPEN_HOME is not None:
-            paths.append(os.path.join(MIOPEN_HOME, 'include'))
-        if HIP_HOME is not None:
-            paths.append(os.path.join(HIP_HOME, 'include'))
     elif cuda:
         cuda_home_include = _join_cuda_home('include')
         # if we have the Debian/Ubuntu packages for cuda, we get /usr as cuda home.
@@ -1804,7 +1799,7 @@ def _get_rocm_arch_flags(cflags: Optional[List[str]] = None) -> List[str]:
     # (from `extra_compile_args`)
     if cflags is not None:
         for flag in cflags:
-            if 'amdgpu-target' in flag:
+            if 'amdgpu-target' in flag or 'offload-arch' in flag:
                 return ['-fno-gpu-rdc']
     # Use same defaults as used for building PyTorch
     # Allow env var to override, just like during initial cmake build.
@@ -1817,7 +1812,7 @@ def _get_rocm_arch_flags(cflags: Optional[List[str]] = None) -> List[str]:
             archs = []
     else:
         archs = _archs.replace(' ', ';').split(';')
-    flags = ['--amdgpu-target=%s' % arch for arch in archs]
+    flags = ['--offload-arch=%s' % arch for arch in archs]
     flags += ['-fno-gpu-rdc']
     return flags