Refactor device initialization

Some logic for device initialization was split across platform
init and device init. This duplicated some calls to get info
for max block dim in Y and Z dimension. This rectifies that
and makes sure the max block dims are only queried once, which
is called from the device constructor.

Author: hdelan

source/adapters/cuda/device.hpp

@@ -40,17 +40,21 @@ struct ur_device_handle_t_ {
       : CuDevice(cuDevice), CuContext(cuContext), EvBase(evBase), RefCount{1},
         Platform(platform) {
 
-    UR_CHECK_ERROR(cuDeviceGetAttribute(
-        &MaxBlockDimY, CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y, cuDevice));
-    UR_CHECK_ERROR(cuDeviceGetAttribute(
-        &MaxBlockDimZ, CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z, cuDevice));
     UR_CHECK_ERROR(cuDeviceGetAttribute(
         &MaxRegsPerBlock, CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK,
         cuDevice));
     UR_CHECK_ERROR(cuDeviceGetAttribute(
         &MaxCapacityLocalMem,
         CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN, cuDevice));
 
+    UR_CHECK_ERROR(urDeviceGetInfo(this, UR_DEVICE_INFO_MAX_WORK_ITEM_SIZES,
+                                   sizeof(MaxWorkItemSizes), MaxWorkItemSizes,
+                                   nullptr));
+
+    UR_CHECK_ERROR(urDeviceGetInfo(this, UR_DEVICE_INFO_MAX_WORK_GROUP_SIZE,
+                                   sizeof(MaxWorkGroupSize), &MaxWorkGroupSize,
+                                   nullptr));
+
     // Set local mem max size if env var is present
     static const char *LocalMemSizePtrUR =
         std::getenv("UR_CUDA_MAX_LOCAL_MEM_SIZE");
@@ -91,13 +95,6 @@ struct ur_device_handle_t_ {
 
   uint64_t getElapsedTime(CUevent) const;
 
-  void saveMaxWorkItemSizes(size_t Size,
-                            size_t *SaveMaxWorkItemSizes) noexcept {
-    memcpy(MaxWorkItemSizes, SaveMaxWorkItemSizes, Size);
-  };
-
-  void saveMaxWorkGroupSize(int Value) noexcept { MaxWorkGroupSize = Value; };
-
   void getMaxWorkItemSizes(size_t RetSize,
                            size_t *RetMaxWorkItemSizes) const noexcept {
     memcpy(RetMaxWorkItemSizes, MaxWorkItemSizes, RetSize);

source/adapters/cuda/platform.cpp

@@ -95,22 +95,6 @@ urPlatformGet(ur_adapter_handle_t *, uint32_t, uint32_t NumEntries,
 
               Platforms[i].Devices.emplace_back(new ur_device_handle_t_{
                   Device, Context, EvBase, &Platforms[i]});
-              {
-                const auto &Dev = Platforms[i].Devices.back().get();
-                size_t MaxWorkGroupSize = 0u;
-                size_t MaxThreadsPerBlock[3] = {};
-                UR_CHECK_ERROR(urDeviceGetInfo(
-                    Dev, UR_DEVICE_INFO_MAX_WORK_ITEM_SIZES,
-                    sizeof(MaxThreadsPerBlock), MaxThreadsPerBlock, nullptr));
-
-                UR_CHECK_ERROR(urDeviceGetInfo(
-                    Dev, UR_DEVICE_INFO_MAX_WORK_GROUP_SIZE,
-                    sizeof(MaxWorkGroupSize), &MaxWorkGroupSize, nullptr));
-
-                Dev->saveMaxWorkItemSizes(sizeof(MaxThreadsPerBlock),
-                                          MaxThreadsPerBlock);
-                Dev->saveMaxWorkGroupSize(MaxWorkGroupSize);
-              }
             }
           } catch (const std::bad_alloc &) {
             // Signal out-of-memory situation