[SYCL] Optimize handler::StoreLambda implementation (#17669)

Current implementation creates descriptions of all kernel params, then
process them one by one. It's possible to process each param right away.

Author: Alexandr-Konovalov

sycl/include/sycl/handler.hpp

@@ -384,16 +384,6 @@ template <int Dims> bool range_size_fits_in_size_t(const range<Dims> &r) {
   return true;
 }
 
-template <typename KernelNameType>
-std::vector<kernel_param_desc_t> getKernelParamDescs() {
-  std::vector<kernel_param_desc_t> Result;
-  int NumParams = getKernelNumParams<KernelNameType>();
-  Result.reserve(NumParams);
-  for (int I = 0; I < NumParams; ++I) {
-    Result.push_back(getKernelParamDesc<KernelNameType>(I));
-  }
-  return Result;
-}
 } // namespace detail
 
 /// Command group handler class.
@@ -485,20 +475,30 @@ class __SYCL_EXPORT handler {
                             "a single kernel or explicit memory operation.");
   }
 
-  /// Extracts and prepares kernel arguments from the lambda using information
-  /// from the built-ins or integration header.
+#ifndef __INTEL_PREVIEW_BREAKING_CHANGES
+  // TODO: Those functions are not used anymore, remove it in the next
+  // ABI-breaking window.
   void extractArgsAndReqsFromLambda(
       char *LambdaPtr,
       const std::vector<detail::kernel_param_desc_t> &ParamDescs, bool IsESIMD);
-  // TODO Unused, remove during ABI breaking window
   void
   extractArgsAndReqsFromLambda(char *LambdaPtr, size_t KernelArgsNum,
                                const detail::kernel_param_desc_t *KernelArgs,
                                bool IsESIMD);
+#endif
+  /// Extracts and prepares kernel arguments from the lambda using information
+  /// from the built-ins or integration header.
+  void extractArgsAndReqsFromLambda(
+      char *LambdaPtr, detail::kernel_param_desc_t (*ParamDescGetter)(int),
+      size_t NumKernelParams, bool IsESIMD);
 
   /// Extracts and prepares kernel arguments set via set_arg(s).
   void extractArgsAndReqs();
 
+#if defined(__INTEL_PREVIEW_BREAKING_CHANGES)
+  // TODO: processArg need not to be public
+  __SYCL_DLL_LOCAL
+#endif
   void processArg(void *Ptr, const detail::kernel_param_kind_t &Kind,
                   const int Size, const size_t Index, size_t &IndexShift,
                   bool IsKernelCreatedFromSource, bool IsESIMD);
@@ -770,9 +770,11 @@ class __SYCL_EXPORT handler {
     // header, so don't perform things that require it.
     if constexpr (KernelHasName) {
       // TODO support ESIMD in no-integration-header case too.
+
       clearArgs();
       extractArgsAndReqsFromLambda(MHostKernel->getPtr(),
-                                   detail::getKernelParamDescs<KernelName>(),
+                                   &(detail::getKernelParamDesc<KernelName>),
+                                   detail::getKernelNumParams<KernelName>(),
                                    detail::isKernelESIMD<KernelName>());
       MKernelName = detail::getKernelName<KernelName>();
     } else {

sycl/source/handler.cpp

@@ -1115,6 +1115,43 @@ void handler::extractArgsAndReqs() {
   }
 }
 
+void handler::extractArgsAndReqsFromLambda(
+    char *LambdaPtr, detail::kernel_param_desc_t (*ParamDescGetter)(int),
+    size_t NumKernelParams, bool IsESIMD) {
+  size_t IndexShift = 0;
+  impl->MArgs.reserve(MaxNumAdditionalArgs * NumKernelParams);
+
+  for (size_t I = 0; I < NumKernelParams; ++I) {
+    detail::kernel_param_desc_t ParamDesc = ParamDescGetter(I);
+    void *Ptr = LambdaPtr + ParamDesc.offset;
+    const detail::kernel_param_kind_t &Kind = ParamDesc.kind;
+    const int &Size = ParamDesc.info;
+    if (Kind == detail::kernel_param_kind_t::kind_accessor) {
+      // For args kind of accessor Size is information about accessor.
+      // The first 11 bits of Size encodes the accessor target.
+      const access::target AccTarget =
+          static_cast<access::target>(Size & AccessTargetMask);
+      if ((AccTarget == access::target::device ||
+           AccTarget == access::target::constant_buffer) ||
+          (AccTarget == access::target::image ||
+           AccTarget == access::target::image_array)) {
+        detail::AccessorBaseHost *AccBase =
+            static_cast<detail::AccessorBaseHost *>(Ptr);
+        Ptr = detail::getSyclObjImpl(*AccBase).get();
+      } else if (AccTarget == access::target::local) {
+        detail::LocalAccessorBaseHost *LocalAccBase =
+            static_cast<detail::LocalAccessorBaseHost *>(Ptr);
+        Ptr = detail::getSyclObjImpl(*LocalAccBase).get();
+      }
+    }
+    processArg(Ptr, Kind, Size, I, IndexShift,
+               /*IsKernelCreatedFromSource=*/false, IsESIMD);
+  }
+}
+
+#ifndef __INTEL_PREVIEW_BREAKING_CHANGES
+// TODO: Those functions are not used anymore, remove it in the next
+// ABI-breaking window.
 void handler::extractArgsAndReqsFromLambda(
     char *LambdaPtr, const std::vector<detail::kernel_param_desc_t> &ParamDescs,
     bool IsESIMD) {
@@ -1149,14 +1186,14 @@ void handler::extractArgsAndReqsFromLambda(
   }
 }
 
-// TODO Unused, remove during ABI breaking window
 void handler::extractArgsAndReqsFromLambda(
     char *LambdaPtr, size_t KernelArgsNum,
     const detail::kernel_param_desc_t *KernelArgs, bool IsESIMD) {
   std::vector<detail::kernel_param_desc_t> ParamDescs(
       KernelArgs, KernelArgs + KernelArgsNum);
   extractArgsAndReqsFromLambda(LambdaPtr, ParamDescs, IsESIMD);
 }
+#endif // __INTEL_PREVIEW_BREAKING_CHANGES
 
 // Calling methods of kernel_impl requires knowledge of class layout.
 // As this is impossible in header, there's a function that calls necessary

sycl/test/abi/sycl_symbols_linux.dump

@@ -3567,6 +3567,7 @@ _ZN4sycl3_V17handler27addLifetimeSharedPtrStorageESt10shared_ptrIKvE
 _ZN4sycl3_V17handler27computeFallbackKernelBoundsEmm
 _ZN4sycl3_V17handler28extractArgsAndReqsFromLambdaEPcRKSt6vectorINS0_6detail19kernel_param_desc_tESaIS5_EEb
 _ZN4sycl3_V17handler28extractArgsAndReqsFromLambdaEPcmPKNS0_6detail19kernel_param_desc_tEb
+_ZN4sycl3_V17handler28extractArgsAndReqsFromLambdaEPcPFNS0_6detail19kernel_param_desc_tEiEmb
 _ZN4sycl3_V17handler28memcpyToHostOnlyDeviceGlobalEPKvS3_mbmm
 _ZN4sycl3_V17handler28setArgsToAssociatedAccessorsEv
 _ZN4sycl3_V17handler28setStateExplicitKernelBundleEv

sycl/test/abi/sycl_symbols_windows.dump

@@ -3987,6 +3987,7 @@
 ?ext_oneapi_wait_external_semaphore@queue@_V1@sycl@@QEAA?AVevent@23@Uexternal_semaphore@experimental@oneapi@ext@23@_KV423@AEBUcode_location@detail@23@@Z
 ?extractArgsAndReqs@handler@_V1@sycl@@AEAAXXZ
 ?extractArgsAndReqsFromLambda@handler@_V1@sycl@@AEAAXPEADAEBV?$vector@Ukernel_param_desc_t@detail@_V1@sycl@@V?$allocator@Ukernel_param_desc_t@detail@_V1@sycl@@@std@@@std@@_N@Z
+?extractArgsAndReqsFromLambda@handler@_V1@sycl@@AEAAXPEADP6A?AUkernel_param_desc_t@detail@23@H@Z_K_N@Z
 ?extractArgsAndReqsFromLambda@handler@_V1@sycl@@AEAAXPEAD_KPEBUkernel_param_desc_t@detail@23@_N@Z
 ?fill_impl@handler@_V1@sycl@@AEAAXPEAXPEBX_K2@Z
 ?finalize@handler@_V1@sycl@@AEAA?AVevent@23@XZ