Fix for AMD

Author: RossBrunton

offload/plugins-nextgen/amdgpu/src/rtl.cpp

@@ -1261,8 +1261,9 @@ struct AMDGPUStreamTy {
     auto [Curr, InputSignal] = consume(OutputSignal);
 
     // Setup the post action to release the kernel args buffer.
-    if (auto Err = Slots[Curr].schedReleaseBuffer(KernelArgs, MemoryManager))
-      return Err;
+    if (KernelArgs)
+      if (auto Err = Slots[Curr].schedReleaseBuffer(KernelArgs, MemoryManager))
+        return Err;
 
     // If we are running an RPC server we want to wake up the server thread
     // whenever there is a kernel running and let it sleep otherwise.
@@ -3375,8 +3376,9 @@ Error AMDGPUKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
   AMDGPUMemoryManagerTy &ArgsMemoryManager = HostDevice.getArgsMemoryManager();
 
   void *AllArgs = nullptr;
-  if (auto Err = ArgsMemoryManager.allocate(ArgsSize, &AllArgs))
-    return Err;
+  if (ArgsSize)
+    if (auto Err = ArgsMemoryManager.allocate(ArgsSize, &AllArgs))
+      return Err;
 
   // Account for user requested dynamic shared memory.
   uint32_t GroupSize = getGroupSize();

offload/unittests/OffloadAPI/device_code/CMakeLists.txt

@@ -7,6 +7,7 @@ macro(add_offload_test_device_code test_filename test_name)
         add_custom_command(OUTPUT ${BIN_PATH}
             COMMAND
             ${CMAKE_C_COMPILER} --target=nvptx64-nvidia-cuda
+            ${ARGN}
             -march=${LIBOMPTARGET_DEP_CUDA_ARCH}
             --cuda-path=${CUDA_ROOT}
             ${SRC_PATH} -o ${BIN_PATH}
@@ -21,6 +22,7 @@ macro(add_offload_test_device_code test_filename test_name)
         add_custom_command(OUTPUT ${BIN_PATH}
             COMMAND
             ${CMAKE_C_COMPILER} --target=amdgcn-amd-amdhsa -nogpulib
+            ${ARGN}
             -mcpu=${LIBOMPTARGET_DEP_AMDGPU_ARCH}
             ${SRC_PATH} -o ${BIN_PATH}
             DEPENDS ${SRC_PATH}
@@ -61,7 +63,9 @@ endif()
 
 add_offload_test_device_code(foo.c foo)
 add_offload_test_device_code(bar.c bar)
-add_offload_test_device_code(noargs.c noargs)
+# By default, amdhsa will add a number of "hidden" arguments to the kernel defintion
+# O3 disables this, and results in a kernel function with actually no arguments as seen by liboffload
+add_offload_test_device_code(noargs.c noargs -O3)
 
 add_custom_target(OffloadUnitTestsDeviceBins DEPENDS ${BIN_PATHS})