[SYCL] do not set zeroed GlobalOffset in kernel parameters (#20082)

Do not set a zeroed GlobalOffset in kernel parameters - the Unified
Runtime layer will handle this correctly.
Note that there is already a similar optimization in
https://github.com/intel/llvm/blob/sycl/sycl/source/detail/scheduler/commands.cpp#L2477

Author: bratpiorka

sycl/source/detail/scheduler/commands.cpp

@@ -2468,14 +2468,16 @@ static ur_result_t SetKernelParamsAndLaunch(
         /* pPropSizeRet = */ nullptr);
 
     const bool EnforcedLocalSize =
-        (RequiredWGSize[0] != 0 || RequiredWGSize[1] != 0 ||
-         RequiredWGSize[2] != 0);
+        (RequiredWGSize[0] != 0 &&
+         (NDRDesc.Dims < 2 || RequiredWGSize[1] != 0) &&
+         (NDRDesc.Dims < 3 || RequiredWGSize[2] != 0));
     if (EnforcedLocalSize)
       LocalSize = RequiredWGSize;
   }
-  const bool HasOffset = NDRDesc.GlobalOffset[0] != 0 ||
-                         NDRDesc.GlobalOffset[1] != 0 ||
-                         NDRDesc.GlobalOffset[2] != 0;
+
+  const bool HasOffset = NDRDesc.GlobalOffset[0] != 0 &&
+                         (NDRDesc.Dims < 2 || NDRDesc.GlobalOffset[1] != 0) &&
+                         (NDRDesc.Dims < 3 || NDRDesc.GlobalOffset[2] != 0);
 
   std::vector<ur_kernel_launch_property_t> property_list;
 
@@ -2610,6 +2612,10 @@ ur_result_t enqueueImpCommandBufferKernel(
   size_t RequiredWGSize[3] = {0, 0, 0};
   size_t *LocalSize = nullptr;
 
+  const bool HasOffset = NDRDesc.GlobalOffset[0] != 0 &&
+                         (NDRDesc.Dims < 2 || NDRDesc.GlobalOffset[1] != 0) &&
+                         (NDRDesc.Dims < 3 || NDRDesc.GlobalOffset[2] != 0);
+
   if (HasLocalSize)
     LocalSize = &NDRDesc.LocalSize[0];
   else {
@@ -2620,8 +2626,9 @@ ur_result_t enqueueImpCommandBufferKernel(
         /* pPropSizeRet = */ nullptr);
 
     const bool EnforcedLocalSize =
-        (RequiredWGSize[0] != 0 || RequiredWGSize[1] != 0 ||
-         RequiredWGSize[2] != 0);
+        (RequiredWGSize[0] != 0 &&
+         (NDRDesc.Dims < 2 || RequiredWGSize[1] != 0) &&
+         (NDRDesc.Dims < 3 || RequiredWGSize[2] != 0));
     if (EnforcedLocalSize)
       LocalSize = RequiredWGSize;
   }
@@ -2637,7 +2644,8 @@ ur_result_t enqueueImpCommandBufferKernel(
 
   ur_result_t Res =
       Adapter.call_nocheck<UrApiKind::urCommandBufferAppendKernelLaunchExp>(
-          CommandBuffer, UrKernel, NDRDesc.Dims, &NDRDesc.GlobalOffset[0],
+          CommandBuffer, UrKernel, NDRDesc.Dims,
+          HasOffset ? &NDRDesc.GlobalOffset[0] : nullptr,
           &NDRDesc.GlobalSize[0], LocalSize, AltUrKernels.size(),
           AltUrKernels.size() ? AltUrKernels.data() : nullptr,
           SyncPoints.size(), SyncPoints.size() ? SyncPoints.data() : nullptr, 0,

unified-runtime/include/ur_api.h

@@ -353,7 +353,7 @@ params:
       desc: "[in] Dimension of the kernel execution."
     - type: "const size_t*"
       name: pGlobalWorkOffset
-      desc: "[in] Offset to use when executing kernel."
+      desc: "[in][optional] Offset to use when executing kernel."
     - type: "const size_t*"
       name: pGlobalWorkSize
       desc: "[in] Global work size to use when executing kernel."

unified-runtime/scripts/core/exp-command-buffer.yml

@@ -109,8 +109,15 @@ kernel_command_data::kernel_command_data(
     ur_kernel_handle_t *KernelAlternatives)
     : Kernel(Kernel), Params(Params), WorkDim(WorkDim) {
   const size_t CopySize = sizeof(size_t) * WorkDim;
-  std::memcpy(GlobalWorkOffset, GlobalWorkOffsetPtr, CopySize);
   std::memcpy(GlobalWorkSize, GlobalWorkSizePtr, CopySize);
+
+  // GlobalWorkOffsetPtr may be nullptr
+  if (GlobalWorkOffsetPtr) {
+    std::memcpy(GlobalWorkOffset, GlobalWorkOffsetPtr, CopySize);
+  } else {
+    std::memset(GlobalWorkOffset, 0, sizeof(size_t) * 3);
+  }
+
   // Local work size may be nullptr
   if (LocalWorkSizePtr) {
     std::memcpy(LocalWorkSize, LocalWorkSizePtr, CopySize);

unified-runtime/source/adapters/cuda/command_buffer.cpp

@@ -55,8 +55,15 @@ ur_exp_command_buffer_command_handle_t_::
     : handle_base(), CommandBuffer(CommandBuffer), Kernel(Kernel), Node(Node),
       Params(Params), WorkDim(WorkDim) {
   const size_t CopySize = sizeof(size_t) * WorkDim;
-  std::memcpy(GlobalWorkOffset, GlobalWorkOffsetPtr, CopySize);
   std::memcpy(GlobalWorkSize, GlobalWorkSizePtr, CopySize);
+
+  // GlobalWorkOffsetPtr may be nullptr
+  if (GlobalWorkOffsetPtr) {
+    std::memcpy(GlobalWorkOffset, GlobalWorkOffsetPtr, CopySize);
+  } else {
+    std::memset(GlobalWorkOffset, 0, sizeof(size_t) * 3);
+  }
+
   // Local work size may be nullptr
   if (LocalWorkSizePtr) {
     std::memcpy(LocalWorkSize, LocalWorkSizePtr, CopySize);