Cuda Fix for command-buffer local argument upate

After setting kernel arguments during update, we
need to reset the amount of local memory used.

Author: Ewan Crawford

source/adapters/cuda/command_buffer.cpp

@@ -1396,14 +1396,22 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferUpdateKernelLaunchExp(
 
   CUDA_KERNEL_NODE_PARAMS &Params = KernelCommandHandle->Params;
 
+  const auto LocalSize = KernelCommandHandle->Kernel->getLocalSize();
+  if (LocalSize != 0) {
+    // Clean the local size, otherwise calling updateKernelArguments() in
+    // future updates with local arguments will incorrectly increase the
+    // size further.
+    KernelCommandHandle->Kernel->clearLocalSize();
+  }
+
   Params.func = CuFunc;
   Params.gridDimX = BlocksPerGrid[0];
   Params.gridDimY = BlocksPerGrid[1];
   Params.gridDimZ = BlocksPerGrid[2];
   Params.blockDimX = ThreadsPerBlock[0];
   Params.blockDimY = ThreadsPerBlock[1];
   Params.blockDimZ = ThreadsPerBlock[2];
-  Params.sharedMemBytes = KernelCommandHandle->Kernel->getLocalSize();
+  Params.sharedMemBytes = LocalSize;
   Params.kernelParams =
       const_cast<void **>(KernelCommandHandle->Kernel->getArgIndices().data());