Make UrArgs thread_local

Author: ldorau

sycl/source/detail/scheduler/commands.cpp

@@ -2395,6 +2395,8 @@ static void GetUrArgsBasedOnType(
   }
 }
 
+thread_local std::vector<ur_exp_kernel_arg_properties_t> UrArgs;
+
 static ur_result_t SetKernelParamsAndLaunch(
     queue_impl &Queue, std::vector<ArgDesc> &Args,
     device_image_impl *DeviceImageImpl, ur_kernel_handle_t Kernel,
@@ -2416,12 +2418,11 @@ static ur_result_t SetKernelParamsAndLaunch(
         DeviceImageImpl ? DeviceImageImpl->get_spec_const_blob_ref() : Empty);
   }
 
-  std::vector<ur_exp_kernel_arg_properties_t> UrArgs;
   UrArgs.reserve(Args.size());
+  UrArgs.clear();
 
   if (KernelFuncPtr && !KernelHasSpecialCaptures) {
-    auto setFunc = [&UrArgs,
-                    KernelFuncPtr](const detail::kernel_param_desc_t &ParamDesc,
+    auto setFunc = [KernelFuncPtr](const detail::kernel_param_desc_t &ParamDesc,
                                    size_t NextTrueIndex) {
       const void *ArgPtr = (const char *)KernelFuncPtr + ParamDesc.offset;
       switch (ParamDesc.kind) {
@@ -2451,8 +2452,8 @@ static ur_result_t SetKernelParamsAndLaunch(
     applyFuncOnFilteredArgs(EliminatedArgMask, KernelNumArgs,
                             KernelParamDescGetter, setFunc);
   } else {
-    auto setFunc = [&DeviceImageImpl, &getMemAllocationFunc, &Queue,
-                    &UrArgs](detail::ArgDesc &Arg, size_t NextTrueIndex) {
+    auto setFunc = [&DeviceImageImpl, &getMemAllocationFunc,
+                    &Queue](detail::ArgDesc &Arg, size_t NextTrueIndex) {
       GetUrArgsBasedOnType(DeviceImageImpl, getMemAllocationFunc,
                            Queue.getContextImpl(), Arg, NextTrueIndex, UrArgs);
     };