[L0][UR] Do not set global offset unless required

sycl/source/detail/cg.hpp

@@ -134,6 +134,13 @@ class NDRDescT {
       ClusterDimensions[I] = (I < Dims) ? N[I] : 1;
   }
 
+  bool hasOffset() const {
+    bool HasOffset = false;
+    for (int I = 0; I < 3; ++I)
+      HasOffset |= GlobalOffset[I];
+    return HasOffset;
+  }
+
   NDRDescT &operator=(const NDRDescT &Desc) = default;
   NDRDescT &operator=(NDRDescT &&Desc) = default;
 

sycl/source/detail/scheduler/commands.cpp

@@ -2431,6 +2431,10 @@ static ur_result_t SetKernelParamsAndLaunch(
 
   size_t RequiredWGSize[3] = {0, 0, 0};
   size_t *LocalSize = nullptr;
+  size_t *GlobalOffset = nullptr;
+
+  if (NDRDesc.hasOffset())
+    GlobalOffset = &NDRDesc.GlobalOffset[0];
 
   if (HasLocalSize)
     LocalSize = &NDRDesc.LocalSize[0];
@@ -2478,9 +2482,9 @@ static ur_result_t SetKernelParamsAndLaunch(
     ur_event_handle_t UREvent = nullptr;
     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(),
+            Queue->getHandleRef(), Kernel, NDRDesc.Dims, GlobalOffset,
+            &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) {
@@ -2497,7 +2501,7 @@ static ur_result_t SetKernelParamsAndLaunch(
                   Args...);
         }
         return Adapter->call_nocheck<UrApiKind::urEnqueueKernelLaunch>(Args...);
-      }(Queue->getHandleRef(), Kernel, NDRDesc.Dims, &NDRDesc.GlobalOffset[0],
+      }(Queue->getHandleRef(), Kernel, NDRDesc.Dims, GlobalOffset,
         &NDRDesc.GlobalSize[0], LocalSize, RawEvents.size(),
         RawEvents.empty() ? nullptr : &RawEvents[0],
         OutEventImpl ? &UREvent : nullptr);
@@ -2605,6 +2609,10 @@ ur_result_t enqueueImpCommandBufferKernel(
 
   size_t RequiredWGSize[3] = {0, 0, 0};
   size_t *LocalSize = nullptr;
+  size_t *GlobalOffset = nullptr;
+
+  if (NDRDesc.hasOffset())
+    GlobalOffset = &NDRDesc.GlobalOffset[0];
 
   if (HasLocalSize)
     LocalSize = &NDRDesc.LocalSize[0];
@@ -2633,7 +2641,7 @@ ur_result_t enqueueImpCommandBufferKernel(
 
   ur_result_t Res =
       Adapter->call_nocheck<UrApiKind::urCommandBufferAppendKernelLaunchExp>(
-          CommandBuffer, UrKernel, NDRDesc.Dims, &NDRDesc.GlobalOffset[0],
+          CommandBuffer, UrKernel, NDRDesc.Dims, GlobalOffset,
           &NDRDesc.GlobalSize[0], LocalSize, AltUrKernels.size(),
           AltUrKernels.size() ? AltUrKernels.data() : nullptr,
           SyncPoints.size(), SyncPoints.size() ? SyncPoints.data() : nullptr, 0,

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

@@ -280,9 +280,19 @@ ur_result_t ur_kernel_handle_t_::prepareForSubmission(
     std::function<void(void *, void *, size_t)> migrate) {
   auto hZeKernel = getZeHandle(hDevice);
 
+  // If we have a new offset, update the zeKernel object
   if (pGlobalWorkOffset != NULL) {
     UR_CALL(
         setKernelGlobalOffset(hContext, hZeKernel, workDim, pGlobalWorkOffset));
+    hadOffset = true;
+  }
+
+  // If we have no offset now but the last submission had one, clear it
+  else if (hadOffset) {
+    size_t zeroOffset[3] = {0, 0, 0};
+    UR_CALL(
+        setKernelGlobalOffset(hContext, hZeKernel, workDim, zeroOffset));
+    hadOffset = false;
   }
 
   ZE2UR_CALL(zeKernelSetGroupSize,

unified-runtime/source/adapters/level_zero/v2/kernel.hpp

@@ -115,4 +115,7 @@ struct ur_kernel_handle_t_ : ur_object {
 
   // pointer to any non-null kernel in deviceKernels
   ur_single_device_kernel_t *nonEmptyKernel;
+
+  // keep track of whether the offset is non-zero
+  bool hadOffset = false;
 };