[UR] Add default implementation for cooperative kernel functions

Cooperative kernels can synchronize using device-scope barriers. These kernels
use `urKernelSuggestMaxCooperativeGroupCountExp` to ensure that all work groups
can run concurrently. When the maximum number of work groups is 1, these kernels
behave the same as regular kernels.

This PR adds a default implementation for
`urKernelSuggestMaxCooperativeGroupCountExp` that returns 1. Also, it adds a
default implementation for `urEnqueueCooperativeKernelLaunchExp` that calls
`urEnqueueKernelLaunch`.

Signed-off-by: Michael Aziz <[email protected]>

Author: 0x12CC

source/adapters/cuda/enqueue.cpp

@@ -490,6 +490,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   return Result;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
+    ur_queue_handle_t hQueue, ur_kernel_handle_t hKernel, uint32_t workDim,
+    const size_t *pGlobalWorkOffset, const size_t *pGlobalWorkSize,
+    const size_t *pLocalWorkSize, uint32_t numEventsInWaitList,
+    const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent) {
+  return urEnqueueKernelLaunch(hQueue, hKernel, workDim, pGlobalWorkOffset,
+                               pGlobalWorkSize, pLocalWorkSize,
+                               numEventsInWaitList, phEventWaitList, phEvent);
+}
+
 /// Set parameters for general 3D memory copy.
 /// If the source and/or destination is on the device, SrcPtr and/or DstPtr
 /// must be a pointer to a CUdeviceptr

source/adapters/cuda/kernel.cpp

@@ -169,6 +169,13 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelGetNativeHandle(
   return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
+    ur_kernel_handle_t hKernel, uint32_t *pGroupCountRet) {
+  (void)hKernel;
+  *pGroupCountRet = 1;
+  return UR_RESULT_SUCCESS;
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urKernelSetArgValue(
     ur_kernel_handle_t hKernel, uint32_t argIndex, size_t argSize,
     const ur_kernel_arg_value_properties_t *pProperties,

source/adapters/hip/enqueue.cpp

@@ -475,6 +475,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   return Result;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
+    ur_queue_handle_t hQueue, ur_kernel_handle_t hKernel, uint32_t workDim,
+    const size_t *pGlobalWorkOffset, const size_t *pGlobalWorkSize,
+    const size_t *pLocalWorkSize, uint32_t numEventsInWaitList,
+    const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent) {
+  return urEnqueueKernelLaunch(hQueue, hKernel, workDim, pGlobalWorkOffset,
+                               pGlobalWorkSize, pLocalWorkSize,
+                               numEventsInWaitList, phEventWaitList, phEvent);
+}
+
 /// Enqueues a wait on the given queue for all events.
 /// See \ref enqueueEventWait
 ///

source/adapters/hip/kernel.cpp

@@ -158,6 +158,13 @@ urKernelGetNativeHandle(ur_kernel_handle_t, ur_native_handle_t *) {
   return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
+    ur_kernel_handle_t hKernel, uint32_t *pGroupCountRet) {
+  (void)hKernel;
+  *pGroupCountRet = 1;
+  return UR_RESULT_SUCCESS;
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urKernelSetArgValue(
     ur_kernel_handle_t hKernel, uint32_t argIndex, size_t argSize,
     const ur_kernel_arg_value_properties_t *, const void *pArgValue) {

source/adapters/level_zero/kernel.cpp

@@ -253,6 +253,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   return UR_RESULT_SUCCESS;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
+    ur_queue_handle_t hQueue, ur_kernel_handle_t hKernel, uint32_t workDim,
+    const size_t *pGlobalWorkOffset, const size_t *pGlobalWorkSize,
+    const size_t *pLocalWorkSize, uint32_t numEventsInWaitList,
+    const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent) {
+  return urEnqueueKernelLaunch(hQueue, hKernel, workDim, pGlobalWorkOffset,
+                               pGlobalWorkSize, pLocalWorkSize,
+                               numEventsInWaitList, phEventWaitList, phEvent);
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urEnqueueDeviceGlobalVariableWrite(
     ur_queue_handle_t Queue,     ///< [in] handle of the queue to submit to.
     ur_program_handle_t Program, ///< [in] handle of the program containing the
@@ -736,6 +746,13 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelGetNativeHandle(
   return UR_RESULT_SUCCESS;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
+    ur_kernel_handle_t hKernel, uint32_t *pGroupCountRet) {
+  (void)hKernel;
+  *pGroupCountRet = 1;
+  return UR_RESULT_SUCCESS;
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urKernelCreateWithNativeHandle(
     ur_native_handle_t NativeKernel, ///< [in] the native handle of the kernel.
     ur_context_handle_t Context,     ///< [in] handle of the context object

source/adapters/opencl/enqueue.cpp

@@ -41,6 +41,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   return UR_RESULT_SUCCESS;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urEnqueueCooperativeKernelLaunchExp(
+    ur_queue_handle_t hQueue, ur_kernel_handle_t hKernel, uint32_t workDim,
+    const size_t *pGlobalWorkOffset, const size_t *pGlobalWorkSize,
+    const size_t *pLocalWorkSize, uint32_t numEventsInWaitList,
+    const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent) {
+  return urEnqueueKernelLaunch(hQueue, hKernel, workDim, pGlobalWorkOffset,
+                               pGlobalWorkSize, pLocalWorkSize,
+                               numEventsInWaitList, phEventWaitList, phEvent);
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urEnqueueEventsWait(
     ur_queue_handle_t hQueue, uint32_t numEventsInWaitList,
     const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent) {

source/adapters/opencl/kernel.cpp

@@ -376,6 +376,13 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelGetNativeHandle(
   return UR_RESULT_SUCCESS;
 }
 
+UR_APIEXPORT ur_result_t UR_APICALL urKernelSuggestMaxCooperativeGroupCountExp(
+    ur_kernel_handle_t hKernel, uint32_t *pGroupCountRet) {
+  (void)hKernel;
+  *pGroupCountRet = 1;
+  return UR_RESULT_SUCCESS;
+}
+
 UR_APIEXPORT ur_result_t UR_APICALL urKernelCreateWithNativeHandle(
     ur_native_handle_t hNativeKernel, ur_context_handle_t, ur_program_handle_t,
     const ur_kernel_native_properties_t *pProperties,