[Offload] Add liboffload unit tests for shared/local memory

offload/unittests/OffloadAPI/device_code/CMakeLists.txt

@@ -2,6 +2,8 @@ add_offload_test_device_code(foo.c foo)
 add_offload_test_device_code(bar.c bar)
 # Compile with optimizations to eliminate AMDGPU implicit arguments.
 add_offload_test_device_code(noargs.c noargs -O3)
+add_offload_test_device_code(localmem.c localmem)
+add_offload_test_device_code(localmem_reduction.c localmem_reduction)
 
-add_custom_target(offload_device_binaries DEPENDS foo.bin bar.bin noargs.bin)
+add_custom_target(offload_device_binaries DEPENDS foo.bin bar.bin noargs.bin localmem.bin localmem_reduction.bin)
 set(OFFLOAD_TEST_DEVICE_CODE_PATH ${CMAKE_CURRENT_BINARY_DIR} PARENT_SCOPE)

offload/unittests/OffloadAPI/device_code/localmem.c

@@ -0,0 +1,11 @@
+#include <gpuintrin.h>
+#include <stdint.h>
+
+extern __gpu_local uint32_t shared_mem[];
+
+__gpu_kernel void localmem(uint32_t *out) {
+  shared_mem[__gpu_thread_id(0)] = __gpu_thread_id(0);
+  shared_mem[__gpu_thread_id(0)] *= 2;
+  out[__gpu_thread_id(0) + (__gpu_num_threads(0) * __gpu_block_id(0))] =
+      shared_mem[__gpu_thread_id(0)];
+}

offload/unittests/OffloadAPI/device_code/localmem_reduction.c

@@ -0,0 +1,16 @@
+#include <gpuintrin.h>
+#include <stdint.h>
+
+extern __gpu_local uint32_t shared_mem[];
+
+__gpu_kernel void localmem_reduction(uint32_t *out) {
+  shared_mem[__gpu_thread_id(0)] = 2;
+
+  __gpu_sync_threads();
+
+  if (__gpu_thread_id(0) == 0) {
+    out[__gpu_block_id(0)] = 0;
+    for (uint32_t i = 0; i < __gpu_num_threads(0); i++)
+      out[__gpu_block_id(0)] += shared_mem[i];
+  }
+}

offload/unittests/OffloadAPI/kernel/olLaunchKernel.cpp

@@ -52,6 +52,21 @@ struct olLaunchKernelNoArgsTest : LaunchKernelTestBase {
 };
 OFFLOAD_TESTS_INSTANTIATE_DEVICE_FIXTURE(olLaunchKernelNoArgsTest);
 
+struct olLaunchKernelLocalMemTest : LaunchKernelTestBase {
+  void SetUp() override {
+    RETURN_ON_FATAL_FAILURE(LaunchKernelTestBase::SetUpKernel("localmem"));
+  }
+};
+OFFLOAD_TESTS_INSTANTIATE_DEVICE_FIXTURE(olLaunchKernelLocalMemTest);
+
+struct olLaunchKernelLocalMemReductionTest : LaunchKernelTestBase {
+  void SetUp() override {
+    RETURN_ON_FATAL_FAILURE(
+        LaunchKernelTestBase::SetUpKernel("localmem_reduction"));
+  }
+};
+OFFLOAD_TESTS_INSTANTIATE_DEVICE_FIXTURE(olLaunchKernelLocalMemReductionTest);
+
 TEST_P(olLaunchKernelTest, Success) {
   void *Mem;
   ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
@@ -99,3 +114,51 @@ TEST_P(olLaunchKernelTest, SuccessSynchronous) {
 
   ASSERT_SUCCESS(olMemFree(Mem));
 }
+
+TEST_P(olLaunchKernelLocalMemTest, Success) {
+  LaunchArgs.NumGroups.x = 4;
+  LaunchArgs.DynSharedMemory = 64 * sizeof(uint32_t);
+
+  void *Mem;
+  ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
+                            LaunchArgs.GroupSize.x * LaunchArgs.NumGroups.x *
+                                sizeof(uint32_t),
+                            &Mem));
+  struct {
+    void *Mem;
+  } Args{Mem};
+
+  ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
+                                &LaunchArgs, nullptr));
+
+  ASSERT_SUCCESS(olWaitQueue(Queue));
+
+  uint32_t *Data = (uint32_t *)Mem;
+  for (uint32_t i = 0; i < LaunchArgs.GroupSize.x * LaunchArgs.NumGroups.x; i++)
+    ASSERT_EQ(Data[i], (i % 64) * 2);
+
+  ASSERT_SUCCESS(olMemFree(Mem));
+}
+
+TEST_P(olLaunchKernelLocalMemReductionTest, Success) {
+  LaunchArgs.NumGroups.x = 4;
+  LaunchArgs.DynSharedMemory = 64 * sizeof(uint32_t);
+
+  void *Mem;
+  ASSERT_SUCCESS(olMemAlloc(Device, OL_ALLOC_TYPE_MANAGED,
+                            LaunchArgs.NumGroups.x * sizeof(uint32_t), &Mem));
+  struct {
+    void *Mem;
+  } Args{Mem};
+
+  ASSERT_SUCCESS(olLaunchKernel(Queue, Device, Kernel, &Args, sizeof(Args),
+                                &LaunchArgs, nullptr));
+
+  ASSERT_SUCCESS(olWaitQueue(Queue));
+
+  uint32_t *Data = (uint32_t *)Mem;
+  for (uint32_t i = 0; i < LaunchArgs.NumGroups.x; i++)
+    ASSERT_EQ(Data[i], 2 * LaunchArgs.GroupSize.x);
+
+  ASSERT_SUCCESS(olMemFree(Mem));
+}