[UR] Make kernel launch global offset param optional. (#18573)

The offset is already deprecated so this will let sycl omit it entirely
in the future, for now it simplifies some adapter code by allowing the
check for whether we're working with an offset to happen once at the
sycl level with subsequent checks being simple nullptr comparisons.

Author: aarongreig

sycl/source/detail/scheduler/commands.cpp

@@ -2475,6 +2475,11 @@ static ur_result_t SetKernelParamsAndLaunch(
     if (EnforcedLocalSize)
       LocalSize = RequiredWGSize;
   }
+
+  const bool HasOffset = NDRDesc.GlobalOffset[0] != 0 ||
+                         NDRDesc.GlobalOffset[1] != 0 ||
+                         NDRDesc.GlobalOffset[2] != 0;
+
   std::vector<ur_exp_launch_property_t> property_list;
   if (KernelUsesClusterLaunch) {
     ur_exp_launch_property_value_t launch_property_value_cluster_range;
@@ -2505,8 +2510,9 @@ static ur_result_t SetKernelParamsAndLaunch(
     ur_result_t Error =
         Adapter->call_nocheck<UrApiKind::urEnqueueKernelLaunchCustomExp>(
             Queue->getHandleRef(), Kernel, NDRDesc.Dims,
-            &NDRDesc.GlobalOffset[0], &NDRDesc.GlobalSize[0], LocalSize,
-            property_list.size(), property_list.data(), RawEvents.size(),
+            HasOffset ? &NDRDesc.GlobalOffset[0] : nullptr,
+            &NDRDesc.GlobalSize[0], LocalSize, property_list.size(),
+            property_list.data(), RawEvents.size(),
             RawEvents.empty() ? nullptr : &RawEvents[0],
             OutEventImpl ? &UREvent : nullptr);
     if ((Error == UR_RESULT_SUCCESS) && OutEventImpl) {
@@ -2523,8 +2529,9 @@ static ur_result_t SetKernelParamsAndLaunch(
                   Args...);
         }
         return Adapter->call_nocheck<UrApiKind::urEnqueueKernelLaunch>(Args...);
-      }(Queue->getHandleRef(), Kernel, NDRDesc.Dims, &NDRDesc.GlobalOffset[0],
-        &NDRDesc.GlobalSize[0], LocalSize, RawEvents.size(),
+      }(Queue->getHandleRef(), Kernel, NDRDesc.Dims,
+        HasOffset ? &NDRDesc.GlobalOffset[0] : nullptr, &NDRDesc.GlobalSize[0],
+        LocalSize, RawEvents.size(),
         RawEvents.empty() ? nullptr : &RawEvents[0],
         OutEventImpl ? &UREvent : nullptr);
   if (Error == UR_RESULT_SUCCESS && OutEventImpl) {

unified-runtime/include/ur_api.h

@@ -33,7 +33,7 @@ params:
       desc: "[in] number of dimensions, from 1 to 3, to specify the global and work-group work-items"
     - type: "const size_t*"
       name: pGlobalWorkOffset
-      desc: "[in] pointer to an array of workDim unsigned values that specify the offset used to calculate the global ID of a work-item"
+      desc: "[in][optional] pointer to an array of workDim unsigned values that specify the offset used to calculate the global ID of a work-item"
     - type: "const size_t*"
       name: pGlobalWorkSize
       desc: "[in] pointer to an array of workDim unsigned values that specify the number of global work-items in workDim that will execute the kernel function"

unified-runtime/scripts/core/enqueue.yml

@@ -64,8 +64,8 @@ ur_result_t urEnqueueKernelLaunch(
     /// [in] number of dimensions, from 1 to 3, to specify the global and
     /// work-group work-items
     uint32_t WorkDim,
-    /// [in] pointer to an array of workDim unsigned values that specify the
-    /// offset used to calculate the global ID of a work-item
+    /// [in][optional] pointer to an array of workDim unsigned values that
+    /// specify the offset used to calculate the global ID of a work-item
     const size_t *GlobalWorkOffset,
     /// [in] pointer to an array of workDim unsigned values that specify the
     /// number of global work-items in workDim that will execute the kernel

unified-runtime/source/adapters/level_zero/kernel.cpp

@@ -220,16 +220,6 @@ ur_result_t ur_command_list_manager::appendKernelLaunch(
     waitListView.clear();
   };
 
-  // If the offset is {0, 0, 0}, pass NULL instead.
-  // This allows us to skip setting the offset.
-  bool hasOffset = false;
-  for (uint32_t i = 0; i < workDim; ++i) {
-    hasOffset |= pGlobalWorkOffset[i];
-  }
-  if (!hasOffset) {
-    pGlobalWorkOffset = NULL;
-  }
-
   UR_CALL(hKernel->prepareForSubmission(context, device, pGlobalWorkOffset,
                                         workDim, WG[0], WG[1], WG[2],
                                         memoryMigrate));

unified-runtime/source/adapters/level_zero/v2/command_list_manager.cpp

@@ -938,16 +938,6 @@ ur_result_t ur_queue_immediate_in_order_t::enqueueCooperativeKernelLaunchExp(
     waitListView.clear();
   };
 
-  // If the offset is {0, 0, 0}, pass NULL instead.
-  // This allows us to skip setting the offset.
-  bool hasOffset = false;
-  for (uint32_t i = 0; i < workDim; ++i) {
-    hasOffset |= pGlobalWorkOffset[i];
-  }
-  if (!hasOffset) {
-    pGlobalWorkOffset = NULL;
-  }
-
   UR_CALL(hKernel->prepareForSubmission(hContext, hDevice, pGlobalWorkOffset,
                                         workDim, WG[0], WG[1], WG[2],
                                         memoryMigrate));

unified-runtime/source/adapters/level_zero/v2/queue_immediate_in_order.cpp

@@ -30,7 +30,7 @@ struct NDRDescT {
            const size_t *GlobalWorkSize, const size_t *LocalWorkSize)
       : WorkDim(WorkDim) {
     for (uint32_t I = 0; I < WorkDim; I++) {
-      GlobalOffset[I] = GlobalWorkOffset[I];
+      GlobalOffset[I] = GlobalWorkOffset ? GlobalWorkOffset[I] : 0;
       GlobalSize[I] = GlobalWorkSize[I];
       LocalSize[I] = LocalWorkSize ? LocalWorkSize[I] : 1;
     }
@@ -72,7 +72,6 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueKernelLaunch(
   urEventWait(numEventsInWaitList, phEventWaitList);
   UR_ASSERT(hQueue, UR_RESULT_ERROR_INVALID_NULL_HANDLE);
   UR_ASSERT(hKernel, UR_RESULT_ERROR_INVALID_NULL_HANDLE);
-  UR_ASSERT(pGlobalWorkOffset, UR_RESULT_ERROR_INVALID_NULL_POINTER);
   UR_ASSERT(workDim > 0, UR_RESULT_ERROR_INVALID_WORK_DIMENSION);
   UR_ASSERT(workDim < 4, UR_RESULT_ERROR_INVALID_WORK_DIMENSION);