[SYCL] Optimize handler::setNDRangeDescriptor (intel#17969)

Author: slawekptak

sycl/include/sycl/detail/array.hpp

@@ -58,6 +58,10 @@ template <int dimensions = 1> class array {
     return result;
   }
 
+  void reset(size_t value) {
+    std::fill(common_array, common_array + dimensions, value);
+  }
+
   size_t get(int dimension) const {
     check_dimension(dimension);
     return common_array[dimension];

sycl/include/sycl/handler.hpp

@@ -3718,7 +3718,8 @@ class __SYCL_EXPORT handler {
   bool HasAssociatedAccessor(detail::AccessorImplHost *Req,
                              access::target AccessTarget) const;
 
-  template <int Dims> static sycl::range<3> padRange(sycl::range<Dims> Range) {
+  template <int Dims>
+  static sycl::range<3> padRange(const sycl::range<Dims> &Range) {
     if constexpr (Dims == 3) {
       return Range;
     } else {
@@ -3729,7 +3730,7 @@ class __SYCL_EXPORT handler {
     }
   }
 
-  template <int Dims> static sycl::id<3> padId(sycl::id<Dims> Id) {
+  template <int Dims> static sycl::id<3> padId(const sycl::id<Dims> &Id) {
     if constexpr (Dims == 3) {
       return Id;
     } else {

sycl/source/detail/cg.hpp

@@ -108,9 +108,6 @@ class NDRDescT {
     setNDRangeLeftover();
   }
 
-  NDRDescT(sycl::range<3> NumWorkItems, sycl::id<3> Offset, int DimsArg)
-      : GlobalSize{NumWorkItems}, GlobalOffset{Offset}, Dims{size_t(DimsArg)} {}
-
   NDRDescT(sycl::range<3> NumWorkItems, sycl::range<3> LocalSize,
            sycl::id<3> Offset, int DimsArg)
       : GlobalSize{NumWorkItems}, LocalSize{LocalSize}, GlobalOffset{Offset},
@@ -144,6 +141,43 @@ class NDRDescT {
       ClusterDimensions[I] = (I < Dims) ? N[I] : 1;
   }
 
+  void reset(const sycl::range<3> &N, bool SetNumWorkGroups, int DimsArg) {
+    if (SetNumWorkGroups) {
+      GlobalSize.reset(0);
+      NumWorkGroups = N;
+    } else {
+      GlobalSize = N;
+      NumWorkGroups.reset(0);
+    }
+    LocalSize.reset(0);
+    GlobalOffset.reset(0);
+    ClusterDimensions.reset(1);
+    Dims = size_t(DimsArg);
+    setNDRangeLeftover();
+  }
+
+  void reset(const sycl::range<3> &NumWorkItems, const sycl::id<3> &Offset,
+             int DimsArg) {
+    GlobalSize = NumWorkItems;
+    GlobalOffset = Offset;
+    NumWorkGroups.reset(0);
+    LocalSize.reset(0);
+    ClusterDimensions.reset(1);
+    Dims = size_t(DimsArg);
+  }
+
+  void reset(const sycl::range<3> &NumWorkItems,
+             const sycl::range<3> &LocalSizeArg, const sycl::id<3> &Offset,
+             int DimsArg) {
+    GlobalSize = NumWorkItems;
+    LocalSize = LocalSizeArg;
+    GlobalOffset = Offset;
+    NumWorkGroups.reset(0);
+    ClusterDimensions.reset(1);
+    Dims = size_t(DimsArg);
+    setNDRangeLeftover();
+  }
+
   NDRDescT &operator=(const NDRDescT &Desc) = default;
   NDRDescT &operator=(NDRDescT &&Desc) = default;
 

sycl/source/handler.cpp

@@ -2218,16 +2218,16 @@ sycl::detail::CGType handler::getType() const { return impl->MCGType; }
 
 void handler::setNDRangeDescriptorPadded(sycl::range<3> N,
                                          bool SetNumWorkGroups, int Dims) {
-  impl->MNDRDesc = NDRDescT{N, SetNumWorkGroups, Dims};
+  impl->MNDRDesc.reset(N, SetNumWorkGroups, Dims);
 }
 void handler::setNDRangeDescriptorPadded(sycl::range<3> NumWorkItems,
                                          sycl::id<3> Offset, int Dims) {
-  impl->MNDRDesc = NDRDescT{NumWorkItems, Offset, Dims};
+  impl->MNDRDesc.reset(NumWorkItems, Offset, Dims);
 }
 void handler::setNDRangeDescriptorPadded(sycl::range<3> NumWorkItems,
                                          sycl::range<3> LocalSize,
                                          sycl::id<3> Offset, int Dims) {
-  impl->MNDRDesc = NDRDescT{NumWorkItems, LocalSize, Offset, Dims};
+  impl->MNDRDesc.reset(NumWorkItems, LocalSize, Offset, Dims);
 }
 
 void handler::saveCodeLoc(detail::code_location CodeLoc, bool IsTopCodeLoc) {