[libomptarget] Fix unified shared memory testing with check-openmp.

This patch fixes the following issues:
- AOMP_GPU no longer needed to be specified
- correctly set architecture list which enabled unified memory support to be detected for AMD architectures
- add ability to select which architecture to use based on the ROCR_VISIBLE_DEVICES value for systems with multiple AMD GPU types
- remove additional entry for force-usm compilation
- fix the current force-usm compilation entry to pass the correct compiler flags to specify the architecture
- remove force_usm.h header file that's no longer needed
- remove inclusion of the force_usm.h file in the compile line

Change-Id: Id7ff15178bc2eab5aae583dda97c911cccd178d5

Author: doru1004

clang/lib/Driver/ToolChains/AMDGPUOpenMP.cpp

@@ -564,15 +564,6 @@ void AMDGPUOpenMPToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
   SmallString<128> P(HostTC.getDriver().ResourceDir);
   llvm::sys::path::append(P, "include/cuda_wrappers");
   CC1Args.push_back(DriverArgs.MakeArgString(P));
-
-  // Force APU mode will focefully include #pragma omp requires
-  // unified_shared_memory via the force_usm header
-  if (DriverArgs.hasArg(options::OPT_fopenmp_force_usm)) {
-    CC1Args.push_back("-include");
-    CC1Args.push_back(
-        DriverArgs.MakeArgString(HostTC.getDriver().ResourceDir +
-                                 "/include/openmp_wrappers/force_usm.h"));
-  }
 }
 
 /// Convert path list to Fortran frontend argument

clang/lib/Headers/CMakeLists.txt

@@ -324,7 +324,6 @@ set(openmp_wrapper_files
   openmp_wrappers/complex_cmath.h
   openmp_wrappers/new
   openmp_wrappers/hip/hip_runtime.h
-  openmp_wrappers/usm/force_usm.h
 )
 
 set(llvm_libc_wrapper_files

clang/lib/Headers/openmp_wrappers/usm/force_usm.h

@@ -101,12 +101,10 @@ if(LIBOMPTARGET_AMDGPU_ARCH)
   execute_process(COMMAND ${LIBOMPTARGET_AMDGPU_ARCH}
                   OUTPUT_VARIABLE LIBOMPTARGET_AMDGPU_ARCH_OUTPUT
                   OUTPUT_STRIP_TRAILING_WHITESPACE)
-  string(FIND "${LIBOMPTARGET_AMDGPU_ARCH_OUTPUT}" "\n" first_arch_string)
-  string(SUBSTRING "${LIBOMPTARGET_AMDGPU_ARCH_OUTPUT}" 0 ${first_arch_string}
-         arch_string)
-  if(arch_string)
+  string(REPLACE "\n" ";" amdgpu_arch_list "${LIBOMPTARGET_AMDGPU_ARCH_OUTPUT}")
+  if(amdgpu_arch_list)
     set(LIBOMPTARGET_FOUND_AMDGPU_GPU TRUE)
-    set(LIBOMPTARGET_DEP_AMDGPU_ARCH "${arch_string}")
+    set(LIBOMPTARGET_AMDGPU_DETECTED_ARCH_LIST "${amdgpu_arch_list}")
   endif()
 endif()
 

openmp/libomptarget/cmake/Modules/LibomptargetGetDependencies.cmake

@@ -189,11 +189,16 @@ def add_libraries(source):
 
 def get_arch_from_target(libomptarget_target):
     if libomptarget_target.startswith('amdgcn'):
-        aomp_gpu = os.environ.get("AOMP_GPU")
-        if not aomp_gpu:
-            print("WARNING: AOMP_GPU not set, defaulting to gfx90a")
-            return 'gfx90a'
-        return aomp_gpu
+        device = os.environ.get("ROCR_VISIBLE_DEVICES")
+        if not device:
+            device = 0
+        else:
+            device = int(device.split(",")[0])
+        amd_gpu_list = config.amdgpu_test_arch.split(";")
+        if len(amd_gpu_list) - 1 < device:
+            print("ERROR: Device selected via ROCR_VISIBLE_DEVICES exceeds number of available AMD devices")
+            os._exit(1)
+        return amd_gpu_list[device]
     if libomptarget_target.startswith('x86_64'):
         return 'x86-64'
     return 'native'
@@ -323,10 +328,9 @@ for libomptarget_target in config.libomptarget_all_targets:
                                      get_arch_from_target(libomptarget_target)))
         config.substitutions.append(("%clangxxx-force-usm-" + libomptarget_target, \
                                      "%clangxx %openmp_flags -fopenmp-force-usm  %cuda_flags %flags %flags_clang -fopenmp-targets=" +\
-                                     remove_suffix_if_present(libomptarget_target)))
-        config.substitutions.append(("%clangxxx-force-usm-" + libomptarget_target, \
-                                     "%clangxx %openmp_flags -fopenmp-force-usm  %cuda_flags %flags %flags_clang -fopenmp-targets=" +\
-                                     remove_suffix_if_present(libomptarget_target)))
+                                     remove_suffix_if_present(libomptarget_target) + " -Xopenmp-target=" +\
+                                     remove_suffix_if_present(libomptarget_target) + " -march=" +\
+                                     get_arch_from_target(libomptarget_target)))
         config.substitutions.append(("%clang-" + libomptarget_target, \
                                      "%clang %openmp_flags %cuda_flags %flags %flags_clang -fopenmp-targets=" +\
                                      remove_suffix_if_present(libomptarget_target) + " -Xopenmp-target=" +\