Merge branch 'SYCLomatic' into fix_mma

oneapi-src · Jun 5, 2024 · b6329c8 · b6329c8
2 parents ece3a14 + 0be2d34
commit b6329c8
Show file tree

Hide file tree

Showing 6 changed files with 56 additions and 33 deletions.
diff --git a/clang/lib/DPCT/ASTTraversal.cpp b/clang/lib/DPCT/ASTTraversal.cpp
@@ -8575,14 +8575,11 @@ void KernelCallRule::runRule(
     }
     if (!LaunchKernelCall)
       return;
-
-    if (DpctGlobalInfo::getInstance().buildLaunchKernelInfo(LaunchKernelCall)) {
-      std::string Repl = "";
-      if (IsAssigned)
-        Repl = "0";
-      emplaceTransformation(
-          new ReplaceStmt(LaunchKernelCall, true, false, std::move(Repl)));
+    if (!IsAssigned)
       removeTrailingSemicolon(LaunchKernelCall, Result);
+    if (DpctGlobalInfo::getInstance().buildLaunchKernelInfo(LaunchKernelCall,
+                                                            IsAssigned)) {
+      emplaceTransformation(new ReplaceStmt(LaunchKernelCall, true, false, ""));
     }
   }
 }

diff --git a/clang/lib/DPCT/AnalysisInfo.cpp b/clang/lib/DPCT/AnalysisInfo.cpp
@@ -1971,14 +1971,17 @@ void DpctGlobalInfo::emplaceReplacements(ReplTy &ReplSetsCUDA /*out*/,
   }
 }
 std::shared_ptr<KernelCallExpr>
-DpctGlobalInfo::buildLaunchKernelInfo(const CallExpr *LaunchKernelCall) {
-  auto LocInfo = getLocInfo(LaunchKernelCall->getBeginLoc());
+DpctGlobalInfo::buildLaunchKernelInfo(const CallExpr *LaunchKernelCall,
+                                      bool IsAssigned) {
+  auto DefRange = getDefinitionRange(LaunchKernelCall->getBeginLoc(),
+                                     LaunchKernelCall->getEndLoc());
+  auto LocInfo = getLocInfo(DefRange.getBegin());
   auto FileInfo = insertFile(LocInfo.first);
   if (FileInfo->findNode<KernelCallExpr>(LocInfo.second))
     return std::shared_ptr<KernelCallExpr>();
 
-  auto KernelInfo =
-      KernelCallExpr::buildFromCudaLaunchKernel(LocInfo, LaunchKernelCall);
+  auto KernelInfo = KernelCallExpr::buildFromCudaLaunchKernel(
+      LocInfo, LaunchKernelCall, IsAssigned);
   if (KernelInfo) {
     FileInfo->insertNode(LocInfo.second, KernelInfo);
   } else {
@@ -5182,17 +5185,14 @@ const std::string &KernelCallExpr::ArgInfo::getTypeString() const {
 }
 void KernelCallExpr::print(KernelPrinter &Printer) {
   std::unique_ptr<KernelPrinter::Block> Block;
-  if (!OuterStmts.empty()) {
-    if (NeedBraces)
-      Block = std::move(Printer.block(true));
-    else
-      Block = std::move(Printer.block(false));
-    OuterStmts.print(Printer);
-  }
-  if (NeedLambda) {
+  if (NeedLambda)
     Block = std::move(Printer.block(true));
-  }
+  else if (!OuterStmts.empty())
+    Block = std::move(Printer.block(NeedBraces));
+  OuterStmts.print(Printer);
   printSubmit(Printer);
+  if (NeedDefaultRetValue)
+    Printer.line("return 0;");
   Block.reset();
   if (!getEvent().empty() && isSync())
     Printer.line(getEvent(), "->wait();");
@@ -5588,14 +5588,19 @@ std::string KernelCallExpr::getReplacement() {
 }
 std::shared_ptr<KernelCallExpr> KernelCallExpr::buildFromCudaLaunchKernel(
     const std::pair<clang::tooling::UnifiedPath, unsigned> &LocInfo,
-    const CallExpr *CE) {
+    const CallExpr *CE, bool IsAssigned) {
   auto LaunchFD = CE->getDirectCallee();
   if (!LaunchFD || (LaunchFD->getName() != "cudaLaunchKernel" &&
                     LaunchFD->getName() != "cudaLaunchCooperativeKernel")) {
     return std::shared_ptr<KernelCallExpr>();
   }
   auto Kernel = std::shared_ptr<KernelCallExpr>(
       new KernelCallExpr(LocInfo.second, LocInfo.first));
+  // Call the lambda function with default return value.
+  if (IsAssigned) {
+    Kernel->setNeedDefaultRet();
+    Kernel->setNeedAddLambda();
+  }
   Kernel->buildLocationInfo(CE);
   Kernel->buildExecutionConfig(
       ArrayRef<const Expr *>{CE->getArg(1), CE->getArg(2), CE->getArg(4),
@@ -5729,6 +5734,8 @@ void KernelCallExpr::setNeedAddLambda(const CUDAKernelCallExpr *KernelCall) {
     NeedLambda = true;
   }
 }
+void KernelCallExpr::setNeedAddLambda() { NeedLambda = true; }
+void KernelCallExpr::setNeedDefaultRet() { NeedDefaultRetValue = true; }
 void KernelCallExpr::buildNeedBracesInfo(const CallExpr *KernelCall) {
   NeedBraces = true;
   auto &Context = dpct::DpctGlobalInfo::getContext();

diff --git a/clang/lib/DPCT/AnalysisInfo.h b/clang/lib/DPCT/AnalysisInfo.h
@@ -1118,7 +1118,8 @@ class DpctGlobalInfo {
   // Emplace stored replacements into replacement set.
   void emplaceReplacements(ReplTy &ReplSetsCUDA /*out*/,
                            ReplTy &ReplSetsSYCL /*out*/);
-  std::shared_ptr<KernelCallExpr> buildLaunchKernelInfo(const CallExpr *);
+  std::shared_ptr<KernelCallExpr>
+  buildLaunchKernelInfo(const CallExpr *, bool IsAssigned = false);
   void insertCudaMalloc(const CallExpr *CE);
   void insertCublasAlloc(const CallExpr *CE);
   std::shared_ptr<CudaMallocInfo> findCudaMalloc(const Expr *CE);
@@ -2694,6 +2695,7 @@ class KernelCallExpr : public CallFunctionExpr {
 public:
   bool IsInMacroDefine = false;
   bool NeedLambda = false;
+  bool NeedDefaultRetValue = false;
 
 private:
   struct ArgInfo {
@@ -2760,7 +2762,7 @@ class KernelCallExpr : public CallFunctionExpr {
 
   static std::shared_ptr<KernelCallExpr> buildFromCudaLaunchKernel(
       const std::pair<clang::tooling::UnifiedPath, unsigned> &LocInfo,
-      const CallExpr *);
+      const CallExpr *, bool IsAssigned = false);
   static std::shared_ptr<KernelCallExpr>
   buildForWrapper(clang::tooling::UnifiedPath, const FunctionDecl *,
                   std::shared_ptr<DeviceFunctionInfo>);
@@ -2784,6 +2786,8 @@ class KernelCallExpr : public CallFunctionExpr {
   void buildKernelInfo(const CUDAKernelCallExpr *KernelCall);
   void setIsInMacroDefine(const CUDAKernelCallExpr *KernelCall);
   void setNeedAddLambda(const CUDAKernelCallExpr *KernelCall);
+  void setNeedAddLambda();
+  void setNeedDefaultRet();
   void buildNeedBracesInfo(const CallExpr *KernelCall);
   void buildLocationInfo(const CallExpr *KernelCall);
   template <class ArgsRange>

diff --git a/clang/runtime/dpct-rt/include/dpct/image.hpp b/clang/runtime/dpct-rt/include/dpct/image.hpp
@@ -797,7 +797,7 @@ template <class T, int dimensions, bool IsImageArray = false> class image_wrappe
                               IsImageArray>::accessor_t;
 
   image_wrapper() { set_channel(image_channel::create<T>()); }
-  ~image_wrapper() { detach(); }
+  ~image_wrapper() override { detach(); }
 
   /// Get image accessor.
   accessor_t get_access(sycl::handler &cgh, sycl::queue &q = get_default_queue()) {

diff --git a/clang/test/dpct/launch_kernel/k.inc b/clang/test/dpct/launch_kernel/k.inc
@@ -29,11 +29,14 @@ void foo38() {
   //CHECK-NEXT:/*
   //CHECK-NEXT:DPCT1123:{{[0-9]+}}: The kernel function pointer cannot be used in the device code. You need to call the kernel function with the correct argument(s) directly. According to the kernel function definition, adjusting the dimension of the sycl::nd_item may also be required.
   //CHECK-NEXT:*/
+  //CHECK-NEXT:CHECK_1([&](){
   //CHECK-NEXT:stream->parallel_for(
   //CHECK-NEXT:sycl::nd_range<3>(sycl::range<3>(z, y, x) * sycl::range<3>(1, 1, block), sycl::range<3>(1, 1, block)), 
   //CHECK-NEXT:[=](sycl::nd_item<3> item_ct1) {
-  //CHECK-NEXT:  ((void*)&kernel38<T>)();
-  //CHECK-NEXT:});CHECK_1(0);
+  //CHECK-NEXT: ((void*)&kernel38<T>)();
+  //CHECK-NEXT:});
+  //CHECK-NEXT:  return 0;
+  //CHECK-NEXT:}());
   cudaStream_t stream;
   CHECK_1(cudaLaunchKernel((void*)&kernel38<T>, dim3(x, y, z), block, args, shared, stream));
 }

diff --git a/clang/test/dpct/macro_test.cu b/clang/test/dpct/macro_test.cu
@@ -1328,14 +1328,26 @@ void foo38() {
   int z;
   int shared;
   cudaStream_t stream;
-  //     CHECK:stream->parallel_for(
-  //CHECK-NEXT:    sycl::nd_range<3>(sycl::range<3>(z, y, x) * sycl::range<3>(1, 1, block),
-  //CHECK-NEXT:                      sycl::range<3>(1, 1, block)),
-  //CHECK-NEXT:    [=](sycl::nd_item<3> item_ct1) {
-  //CHECK-NEXT:      ((void *)&kernel38<T>)();
-  //CHECK-NEXT:    });
-  //CHECK-NEXT:CHECK_1(0);
+  //CHECK:CHECK_1([&]() {
+  //CHECK-NEXT: stream->parallel_for(
+  //CHECK-NEXT: sycl::nd_range<3>(sycl::range<3>(z, y, x) * sycl::range<3>(1, 1, block),
+  //CHECK-NEXT:                   sycl::range<3>(1, 1, block)),
+  //CHECK-NEXT: [=](sycl::nd_item<3> item_ct1) {
+  //CHECK-NEXT:   ((void *)&kernel38<T>)();
+  //CHECK-NEXT: });
+  //CHECK-NEXT: return 0;
+  //CHECK-NEXT: }());
   CHECK_1(cudaLaunchKernel((void*)&kernel38<T>, dim3(x, y, z), block, args, shared, stream));
+  //CHECK:dpct::err0 status = [&]() {
+  //CHECK-NEXT:   stream->parallel_for(
+  //CHECK-NEXT:       sycl::nd_range<3>(sycl::range<3>(z, y, x) * sycl::range<3>(1, 1, block),
+  //CHECK-NEXT:                         sycl::range<3>(1, 1, block)),
+  //CHECK-NEXT:       [=](sycl::nd_item<3> item_ct1) {
+  //CHECK-NEXT:         ((void *)&kernel38<T>)();
+  //CHECK-NEXT:       });
+  //CHECK-NEXT:   return 0;
+  //CHECK-NEXT: }();
+  cudaError_t status = cudaLaunchKernel((void*)&kernel38<T>, dim3(x, y, z), block, args, shared, stream);
 }
 #undef CHECK_1
 #undef CHECK_2