[UR] Support 2D and 3D enqueue kernel launches in offload

A (rather trivial) test has also been added since Offload currently has
issues with arguments.

Author: RossBrunton

unified-runtime/source/adapters/offload/enqueue.cpp

@@ -28,22 +28,27 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   //
 
   (void)pGlobalWorkOffset;
-  (void)pLocalWorkSize;
 
-  if (workDim == 1) {
-    std::cerr
-        << "UR Offload adapter only supports 1d kernel launches at the moment";
-    return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  if (!pLocalWorkSize) {
+    // TODO: This is not optimal, but it is legal
+    static size_t DefaultWorkSize[3] = {1, 1, 1};
+    pLocalWorkSize = DefaultWorkSize;
+  }
+
+  if (pLocalWorkSize[0] > pGlobalWorkSize[0] ||
+      pLocalWorkSize[1] > pGlobalWorkSize[1] ||
+      pLocalWorkSize[2] > pGlobalWorkSize[2]) {
+    return UR_RESULT_ERROR_INVALID_WORK_GROUP_SIZE;
   }
 
   ol_kernel_launch_size_args_t LaunchArgs;
   LaunchArgs.Dimensions = workDim;
-  LaunchArgs.NumGroupsX = pGlobalWorkSize[0];
-  LaunchArgs.NumGroupsY = 1;
-  LaunchArgs.NumGroupsZ = 1;
-  LaunchArgs.GroupSizeX = 1;
-  LaunchArgs.GroupSizeY = 1;
-  LaunchArgs.GroupSizeZ = 1;
+  LaunchArgs.NumGroupsX = pGlobalWorkSize[0] / pLocalWorkSize[0];
+  LaunchArgs.NumGroupsY = pGlobalWorkSize[1] / pLocalWorkSize[1];
+  LaunchArgs.NumGroupsZ = pGlobalWorkSize[2] / pLocalWorkSize[2];
+  LaunchArgs.GroupSizeX = pLocalWorkSize[0];
+  LaunchArgs.GroupSizeY = pLocalWorkSize[1];
+  LaunchArgs.GroupSizeZ = pLocalWorkSize[2];
   LaunchArgs.DynSharedMemory = 0;
 
   ol_event_handle_t EventOut;

unified-runtime/test/conformance/device_code/CMakeLists.txt

@@ -163,6 +163,7 @@ add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/standard_types.cpp)
 add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/subgroup.cpp)
 add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/linker_error.cpp)
 add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/saxpy_usm_local_mem.cpp)
+add_device_binary(${CMAKE_CURRENT_SOURCE_DIR}/no_args.cpp)
 
 set(KERNEL_HEADER ${UR_CONFORMANCE_DEVICE_BINARIES_DIR}/kernel_entry_points.h)
 add_custom_command(OUTPUT ${KERNEL_HEADER}

unified-runtime/test/conformance/device_code/no_args.cpp

@@ -0,0 +1,17 @@
+// Copyright (C) 2025 Intel Corporation
+// Part of the Unified-Runtime Project, under the Apache License v2.0 with LLVM
+// Exceptions. See LICENSE.TXT
+//
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include <sycl/sycl.hpp>
+
+int main() {
+  sycl::queue sycl_queue;
+  sycl_queue.submit([&](sycl::handler &cgh) {
+    cgh.parallel_for<class no_args>(
+        sycl::range<3>{128, 128, 128},
+        [](sycl::item<3> itemId) { itemId.get_id(0); });
+  });
+  return 0;
+}

unified-runtime/test/conformance/enqueue/urEnqueueKernelLaunch.cpp

@@ -9,6 +9,18 @@
 #include <uur/fixtures.h>
 #include <uur/known_failure.h>
 
+struct urEnqueueKernelLaunchNoArgs3DTest : uur::urKernelExecutionTest {
+  void SetUp() override {
+    program_name = "no_args";
+    UUR_RETURN_ON_FATAL_FAILURE(urKernelExecutionTest::SetUp());
+  }
+
+  size_t global_size[3] = {32, 16, 8};
+  size_t global_offset[3] = {0, 0, 0};
+  size_t n_dimensions = 3;
+};
+UUR_INSTANTIATE_DEVICE_TEST_SUITE(urEnqueueKernelLaunchNoArgs3DTest);
+
 struct urEnqueueKernelLaunchTest : uur::urKernelExecutionTest {
   void SetUp() override {
     program_name = "fill";
@@ -67,6 +79,13 @@ struct urEnqueueKernelLaunchKernelStandardTest : uur::urKernelExecutionTest {
 };
 UUR_INSTANTIATE_DEVICE_TEST_SUITE(urEnqueueKernelLaunchKernelStandardTest);
 
+TEST_P(urEnqueueKernelLaunchNoArgs3DTest, Success) {
+  ASSERT_SUCCESS(urEnqueueKernelLaunch(queue, kernel, n_dimensions,
+                                       global_offset, global_size, nullptr, 0,
+                                       nullptr, nullptr));
+  ASSERT_SUCCESS(urQueueFinish(queue));
+}
+
 TEST_P(urEnqueueKernelLaunchTest, Success) {
   ur_mem_handle_t buffer = nullptr;
   AddBuffer1DArg(sizeof(val) * global_size, &buffer);