cherry picked from commit 02ce785

CMakeLists.txt

@@ -6,7 +6,53 @@ else()
 endif()
 cmake_minimum_required(VERSION 3.5)
 
-project(TransferBench VERSION 1.51.0 LANGUAGES CXX)
+project(TransferBench VERSION 1.52.0 LANGUAGES CXX)
+
+# Default GPU architectures to build
+#==================================================================================================
+set(DEFAULT_GPUS
+      gfx906
+      gfx908
+      gfx90a
+      gfx942
+      gfx1030
+      gfx1100
+      gfx1101
+      gfx1102
+      gfx1200
+      gfx1201)
+
+# Build only for local GPU architecture
+if (BUILD_LOCAL_GPU_TARGET_ONLY)
+  message(STATUS "Building only for local GPU target")
+  if (COMMAND rocm_local_targets)
+    rocm_local_targets(DEFAULT_GPUS)
+  else()
+    message(WARNING "Unable to determine local GPU targets. Falling back to default GPUs.")
+  endif()
+endif()
+
+# Determine which GPU architectures to build for
+set(AMDGPU_TARGETS "${DEFAULT_GPUS}" CACHE STRING "Target default GPUs if AMDGPU_TARGETS is not defined.")
+
+# Check if clang compiler can offload to AMDGPU_TARGETS
+if (COMMAND rocm_check_target_ids)
+  message(STATUS "Checking for ROCm support for GPU targets: " "${AMDGPU_TARGETS}")
+  rocm_check_target_ids(SUPPORTED_GPUS TARGETS ${AMDGPU_TARGETS})
+else()
+  message(WARNING "Unable to check for supported GPU targets. Falling back to default GPUs.")
+  set(SUPPORTED_GPUS ${DEFAULT_GPUS})
+endif()
+
+set(GPU_TARGETS "${SUPPORTED_GPUS}" CACHE STRING "GPU targets to compile for.")
+message(STATUS "Compiling for ${GPU_TARGETS}")
+
+foreach(target ${GPU_TARGETS})
+ list(APPEND static_link_flags --offload-arch=${target})
+endforeach()
+list(JOIN static_link_flags " " flags_str)
+set( CMAKE_CXX_FLAGS "${flags_str} ${CMAKE_CXX_FLAGS}")
+
 set( CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -L${ROCM_PATH}/lib")
 include_directories(${ROCM_PATH}/include)
 link_libraries(numa hsa-runtime64 pthread)

src/include/Kernels.hpp

@@ -62,7 +62,7 @@ struct SubExecParam
 };
 
 // Macro for collecting HW_REG_HW_ID
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1200__) || defined(__gfx1201__)
 #define GetHwId(hwId) \
   hwId = 0
 #elif defined(__NVCC__)