[SYCL] Enable SYCL integration header support with kernel arg optimization table

Author: srividya-sundaram

clang/include/clang/Sema/Sema.h

@@ -336,7 +336,8 @@ class SYCLIntegrationHeader {
 
   /// Adds a kernel parameter descriptor to current kernel invocation
   /// descriptor.
-  void addParamDesc(kernel_param_kind_t Kind, int Info, unsigned Offset);
+  void addParamDesc(kernel_param_kind_t Kind, int Info, unsigned Offset,
+                    unsigned NumOpenCLParams);
 
   /// Signals that addition of parameter descriptors to current kernel
   /// invocation descriptor has finished.
@@ -358,9 +359,9 @@ class SYCLIntegrationHeader {
     int Info = 0;
     // Offset of the captured parameter value in the lambda or function object.
     unsigned Offset = 0;
-    // bool to mark kernel parameters that are used
-    bool KernelParamUsed;    
-    KernelParamDesc() = default;
+    // NumOpenCLParams to capture number of parameters in Accessor/Sampler init
+    // method.
+    unsigned NumOpenCLParams = 1;
   };
 
   // Kernel invocation descriptor

clang/lib/Sema/SemaSYCL.cpp

@@ -1765,14 +1765,16 @@ class SyclKernelIntHeaderCreator : public SyclKernelFieldHandler {
 
   void addParam(const FieldDecl *FD, QualType ArgTy,
                 SYCLIntegrationHeader::kernel_param_kind_t Kind) {
+    assert(Kind != SYCLIntegrationHeader::kind_accessor &&
+           "Kernel parameter should not be an Accessor");
     uint64_t Size;
     const ConstantArrayType *CAT =
         SemaRef.getASTContext().getAsConstantArrayType(ArgTy);
     if (CAT)
       ArgTy = CAT->getElementType();
     Size = SemaRef.getASTContext().getTypeSizeInChars(ArgTy).getQuantity();
     Header.addParamDesc(Kind, static_cast<unsigned>(Size),
-                        static_cast<unsigned>(CurOffset));
+                        static_cast<unsigned>(CurOffset), 1);
   }
 
 public:
@@ -1783,6 +1785,16 @@ class SyclKernelIntHeaderCreator : public SyclKernelFieldHandler {
     Header.startKernel(Name, NameType, StableName, KernelObj->getLocation());
   }
 
+  unsigned getNumOpenCLParams(const CXXRecordDecl *SamplerOrAccessorTy) {
+    assert(SamplerOrAccessorTy &&
+           "Sampler/Accessor type must be a C++ record type");
+    CXXMethodDecl *InitMethod =
+        getMethodByName(SamplerOrAccessorTy, InitMethodName);
+    assert(InitMethod && "sampler/accessor must have __init method");
+    unsigned NumOpenCLParams;
+    return NumOpenCLParams = InitMethod->param_size();
+  }
+
   bool handleSyclAccessorType(const CXXBaseSpecifier &BC,
                               QualType FieldTy) final {
     const auto *AccTy =
@@ -1792,7 +1804,8 @@ class SyclKernelIntHeaderCreator : public SyclKernelFieldHandler {
     int Dims = static_cast<int>(
         AccTy->getTemplateArgs()[1].getAsIntegral().getExtValue());
     int Info = getAccessTarget(AccTy) | (Dims << 11);
-    Header.addParamDesc(SYCLIntegrationHeader::kind_accessor, Info, CurOffset);
+    Header.addParamDesc(SYCLIntegrationHeader::kind_accessor, Info, CurOffset,
+                        getNumOpenCLParams(AccTy));
     return true;
   }
 
@@ -1804,20 +1817,21 @@ class SyclKernelIntHeaderCreator : public SyclKernelFieldHandler {
     int Dims = static_cast<int>(
         AccTy->getTemplateArgs()[1].getAsIntegral().getExtValue());
     int Info = getAccessTarget(AccTy) | (Dims << 11);
-    Header.addParamDesc(SYCLIntegrationHeader::kind_accessor, Info, CurOffset);
+    Header.addParamDesc(SYCLIntegrationHeader::kind_accessor, Info, CurOffset,
+                        getNumOpenCLParams(AccTy));
     return true;
   }
-
+ 
   bool handleSyclSamplerType(FieldDecl *FD, QualType FieldTy) final {
     const auto *SamplerTy = FieldTy->getAsCXXRecordDecl();
     assert(SamplerTy && "Sampler type must be a C++ record type");
     CXXMethodDecl *InitMethod = getMethodByName(SamplerTy, InitMethodName);
     assert(InitMethod && "sampler must have __init method");
-
     // sampler __init method has only one argument
+    assert((InitMethod->param_size() == 1) &&
+           "sampler __init method should have only one argument");
     const ParmVarDecl *SamplerArg = InitMethod->getParamDecl(0);
     assert(SamplerArg && "sampler __init method must have sampler parameter");
-
     addParam(FD, SamplerArg->getType(), SYCLIntegrationHeader::kind_sampler);
     return true;
   }
@@ -2686,34 +2700,42 @@ void SYCLIntegrationHeader::emit(raw_ostream &O) {
        "the\n";
   O << "// corresponding source\n";
   O << "static constexpr\n";
-  O << "const bool kernel_param_used[] = {\n";
-  O << "  //PARAM_USED_TABLE_BEGIN "
-    << "\n";
-  for (unsigned I = 0; I < KernelDescs.size(); I++) {
-    auto &K = KernelDescs[I];
+  O << "const bool param_omit_table[] = {\n";
+  O << "  // OMIT_TABLE_BEGIN\n";
+  for (const KernelDesc &K : KernelDescs) {
     O << "  //--- " << K.Name << "\n";
-    // bool value for each param
+    O << "  ";
+    for (const KernelParamDesc &P : K.Params) 
+      for (unsigned J = 0; J < P.NumOpenCLParams; J++)
+        O << "false, ";
+    O << "\n";
   }
-  O << "  //PARAM_USED_TABLE_END "
-    << "\n";
-  O << "  } \n\n";
+  O << "  // OMIT_TABLE_END\n";
+  O << "  }; \n\n";
 
   O << "// array representing signatures of all kernels defined in the\n";
   O << "// corresponding source\n";
   O << "static constexpr\n";
   O << "const kernel_param_desc_t kernel_signatures[] = {\n";
 
-  for (unsigned I = 0; I < KernelDescs.size(); I++) {
-    auto &K = KernelDescs[I];
+  unsigned CurIndex = 0;
+  for (const KernelDesc &K : KernelDescs) {
     O << "  //--- " << K.Name << "\n";
-
-    for (unsigned I = 0; I < K.Params.size(); I++) {
-      auto &P = K.Params[I];
+    for (const KernelParamDesc &P : K.Params) {
       std::string TyStr = paramKind2Str(P.Kind);
       O << "  { kernel_param_kind_t::" << TyStr << ", ";
-      O << P.Info << ", " << P.Offset << ", "
-        << "kernel_param_used[" << I << "]"
-        << " },\n";
+      O << P.Info << ", " << P.Offset << ", ";
+      O << "param_omit_table[" << CurIndex << "]";
+      for (unsigned X = 1; X < P.NumOpenCLParams; X++) {
+        ++CurIndex;
+        if (X < P.NumOpenCLParams)
+          O << " | ";
+        O << "(param_omit_table[" << CurIndex << "]"
+          << " << " << X << ")";
+      }
+      ++CurIndex;
+      O << "}";
+      O << ",\n";
     }
     O << "\n";
   }
@@ -2808,14 +2830,11 @@ void SYCLIntegrationHeader::startKernel(StringRef KernelName,
 }
 
 void SYCLIntegrationHeader::addParamDesc(kernel_param_kind_t Kind, int Info,
-                                         unsigned Offset) {
+                                         unsigned Offset,
+                                         unsigned NumOpenCLParams) {
   auto *K = getCurKernelDesc();
   assert(K && "no kernels");
-  K->Params.push_back(KernelParamDesc());
-  KernelParamDesc &PD = K->Params.back();
-  PD.Kind = Kind;
-  PD.Info = Info;
-  PD.Offset = Offset;
+  K->Params.push_back({Kind, Info, Offset, NumOpenCLParams});
 }
 
 void SYCLIntegrationHeader::endKernel() {

clang/test/CodeGenSYCL/integration_header.cpp

@@ -24,52 +24,57 @@
 // CHECK-NEXT:   "_ZTSZ4mainE16accessor_in_base"
 // CHECK-NEXT: };
 //
-// CHECK:  static constexpr
-// CHECK-NEXT:  const bool kernel_param_used[] = {
-// CHECK-NEXT:  //PARAM_USED_TABLE_BEGIN 
-// CHECK-NEXT:  //kernel_name_1 
-// CHECK-NEXT:  true, true, true 
-// CHECK-NEXT:  //kernel_name_2 
-// CHECK-NEXT:  true, true, true 
-// CHECK-NEXT:  //PARAM_USED_TABLE_END 
-// CHECK-NEXT:  } 
-
+// CHECK: static constexpr
+// CHECK-NEXT: const bool param_omit_table[] = {
+// CHECK-NEXT:   // OMIT_TABLE_BEGIN
+// CHECK-NEXT:   //--- _ZTSZ4mainE12first_kernel
+// CHECK-NEXT:   false, false, false, false, false, false, false, false, false, false, false, false,
+// CHECK-NEXT:   //--- _ZTSN16second_namespace13second_kernelIcEE
+// CHECK-NEXT:   false, false, false, false, false, false,
+// CHECK-NEXT:   //--- _ZTS12third_kernelILi1Ei5pointIZ4mainE1XEE
+// CHECK-NEXT:   false, false, false, false, false, false,
+// CHECK-NEXT:   //--- _ZTS13fourth_kernelIJN15template_arg_ns14namespaced_argILi1EEEEE
+// CHECK-NEXT:   false, false, false, false, false,
+// CHECK-NEXT:   //--- _ZTSZ4mainE16accessor_in_base
+// CHECK-NEXT:   false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false,
+// CHECK-NEXT:   // OMIT_TABLE_END
+// CHECK-NEXT:   };
 // CHECK: static constexpr
 // CHECK-NEXT: const kernel_param_desc_t kernel_signatures[] = {
 // CHECK-NEXT:   //--- _ZTSZ4mainE12first_kernel
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, kernel_param_used[0] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 1, 4, kernel_param_used[1] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 8, kernel_param_used[2] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 12, kernel_param_used[3] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 24, kernel_param_used[4] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_sampler, 8, 40, kernel_param_used[5] },
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, param_omit_table[0]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 1, 4, param_omit_table[1]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 8, param_omit_table[2]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 12, param_omit_table[3] | (param_omit_table[4] << 1) | (param_omit_table[5] << 2) | (param_omit_table[6] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 24, param_omit_table[7] | (param_omit_table[8] << 1) | (param_omit_table[9] << 2) | (param_omit_table[10] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_sampler, 8, 40, param_omit_table[11]},
 // CHECK-EMPTY:
 // CHECK-NEXT:   //--- _ZTSN16second_namespace13second_kernelIcEE
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, kernel_param_used[0] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 4, kernel_param_used[1] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_sampler, 8, 16, kernel_param_used[2] },
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, param_omit_table[12]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 4, param_omit_table[13] | (param_omit_table[14] << 1) | (param_omit_table[15] << 2) | (param_omit_table[16] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_sampler, 8, 16, param_omit_table[17]},
 // CHECK-EMPTY:
 // CHECK-NEXT:   //--- _ZTS12third_kernelILi1Ei5pointIZ4mainE1XEE
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, kernel_param_used[0] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 4, kernel_param_used[1] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_sampler, 8, 16, kernel_param_used[2] },
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, param_omit_table[18]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 4, param_omit_table[19] | (param_omit_table[20] << 1) | (param_omit_table[21] << 2) | (param_omit_table[22] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_sampler, 8, 16, param_omit_table[23]},
 // CHECK-EMPTY:
 // CHECK-NEXT:   //--- _ZTS13fourth_kernelIJN15template_arg_ns14namespaced_argILi1EEEEE
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, kernel_param_used[0] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 4, kernel_param_used[1] },
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, param_omit_table[24]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 6112, 4, param_omit_table[25] | (param_omit_table[26] << 1) | (param_omit_table[27] << 2) | (param_omit_table[28] << 3)},
 // CHECK-EMPTY:
-// CHECK-NEXT:  //--- _ZTSZ4mainE16accessor_in_base
-// CHECK-NEXT:  { kernel_param_kind_t::kind_std_layout, 4, 0, kernel_param_used[0] },
-// CHECK-NEXT:  { kernel_param_kind_t::kind_std_layout, 4, 4, kernel_param_used[1] },
-// CHECK-NEXT:  { kernel_param_kind_t::kind_accessor, 4062, 8, kernel_param_used[2] },
-// CHECK-NEXT:  { kernel_param_kind_t::kind_std_layout, 4, 20, kernel_param_used[3] },
-// CHECK-NEXT:  { kernel_param_kind_t::kind_accessor, 4062, 24, kernel_param_used[4] },
-// CHECK-NEXT:  { kernel_param_kind_t::kind_std_layout, 4, 36, kernel_param_used[5] },
-// CHECK-NEXT:  { kernel_param_kind_t::kind_accessor, 4062, 40, kernel_param_used[6] },
-// CHECK-NEXT:  { kernel_param_kind_t::kind_accessor, 4062, 52, kernel_param_used[7] },
+// CHECK-NEXT:   //--- _ZTSZ4mainE16accessor_in_base
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, param_omit_table[29]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 4, param_omit_table[30]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 8, param_omit_table[31] | (param_omit_table[32] << 1) | (param_omit_table[33] << 2) | (param_omit_table[34] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 20, param_omit_table[35]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 24, param_omit_table[36] | (param_omit_table[37] << 1) | (param_omit_table[38] << 2) | (param_omit_table[39] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 36, param_omit_table[40]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 40, param_omit_table[41] | (param_omit_table[42] << 1) | (param_omit_table[43] << 2) | (param_omit_table[44] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 52, param_omit_table[45] | (param_omit_table[46] << 1) | (param_omit_table[47] << 2) | (param_omit_table[48] << 3)},
 // CHECK-EMPTY:
 // CHECK-NEXT: };
-//
+
 // CHECK: template <> struct KernelInfo<class first_kernel> {
 // CHECK: template <> struct KernelInfo<::second_namespace::second_kernel<char>> {
 // CHECK: template <> struct KernelInfo<::third_kernel<1, int, ::point<X>>> {

clang/test/CodeGenSYCL/kernel-param-acc-array-ih.cpp

@@ -16,30 +16,26 @@
 // CHECK-NEXT: const char* const kernel_names[] = {
 // CHECK-NEXT:   "_ZTSZ4mainE8kernel_A"
 // CHECK-NEXT: };
-
+//
 // CHECK: static constexpr
-// CHECK-NEXT: const bool kernel_param_used[] = {
-// CHECK-NEXT:  //PARAM_USED_TABLE_BEGIN
-// CHECK-NEXT:  //kernel_name_1
-// CHECK-NEXT:  true, true, true
-// CHECK-NEXT:  //kernel_name_2
-// CHECK-NEXT:  true, true, true
-// CHECK-NEXT:  //PARAM_USED_TABLE_END
-// CHECK-NEXT:  }
-
+// CHECK-NEXT: const bool param_omit_table[] = {
+// CHECK-NEXT:   // OMIT_TABLE_BEGIN
+// CHECK-NEXT:   //--- _ZTSZ4mainE8kernel_A
+// CHECK-NEXT:   false, false, false, false, false, false, false, false,
+// CHECK-NEXT:   // OMIT_TABLE_END
+// CHECK-NEXT:   };
+//
 // CHECK: static constexpr
 // CHECK-NEXT: const kernel_param_desc_t kernel_signatures[] = {
 // CHECK-NEXT:   //--- _ZTSZ4mainE8kernel_A
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 0, kernel_param_used[0] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 12, kernel_param_used[1] },
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 0, param_omit_table[0] | (param_omit_table[1] << 1) | (param_omit_table[2] << 2) | (param_omit_table[3] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 12, param_omit_table[4] | (param_omit_table[5] << 1) | (param_omit_table[6] << 2) | (param_omit_table[7] << 3)},
 // CHECK-EMPTY:
 // CHECK-NEXT: };
-
 // CHECK: static constexpr
 // CHECK-NEXT: const unsigned kernel_signature_start[] = {
 // CHECK-NEXT:  0 // _ZTSZ4mainE8kernel_A
 // CHECK-NEXT: };
-
 // CHECK: template <> struct KernelInfo<class kernel_A> {
 
 #include <sycl.hpp>

clang/test/CodeGenSYCL/kernel-param-member-acc-array-ih.cpp

@@ -18,20 +18,18 @@
 // CHECK-NEXT: };
 
 // CHECK: static constexpr
-// CHECK-NEXT: const bool kernel_param_used[] = {
-// CHECK-NEXT:  //PARAM_USED_TABLE_BEGIN
-// CHECK-NEXT:  //kernel_name_1
-// CHECK-NEXT:  true, true, true
-// CHECK-NEXT:  //kernel_name_2
-// CHECK-NEXT:  true, true, true
-// CHECK-NEXT:  //PARAM_USED_TABLE_END
-// CHECK-NEXT:  }
+// CHECK-NEXT: const bool param_omit_table[] = {
+// CHECK-NEXT:   // OMIT_TABLE_BEGIN
+// CHECK-NEXT:   //--- _ZTSZ4mainE8kernel_C
+// CHECK-NEXT:   false, false, false, false, false, false, false, false,
+// CHECK-NEXT:   // OMIT_TABLE_END
+// CHECK-NEXT:   };
 
 // CHECK: static constexpr
 // CHECK-NEXT: const kernel_param_desc_t kernel_signatures[] = {
-// CHECK-NEXT: //--- _ZTSZ4mainE8kernel_C
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 0, kernel_param_used[0] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 12, kernel_param_used[1] },
+// CHECK-NEXT:   //--- _ZTSZ4mainE8kernel_C
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 0, param_omit_table[0] | (param_omit_table[1] << 1) | (param_omit_table[2] << 2) | (param_omit_table[3] << 3)},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_accessor, 4062, 12, param_omit_table[4] | (param_omit_table[5] << 1) | (param_omit_table[6] << 2) | (param_omit_table[7] << 3)},
 // CHECK-EMPTY:
 // CHECK-NEXT: };
 

clang/test/CodeGenSYCL/kernel-param-pod-array-ih.cpp

@@ -17,23 +17,21 @@
 // CHECK-NEXT: };
 
 // CHECK: static constexpr
-// CHECK-NEXT: const bool kernel_param_used[] = {
-// CHECK-NEXT:  //PARAM_USED_TABLE_BEGIN
-// CHECK-NEXT:  //kernel_name_1
-// CHECK-NEXT:  true, true, true
-// CHECK-NEXT:  //kernel_name_2
-// CHECK-NEXT:  true, true, true
-// CHECK-NEXT:  //PARAM_USED_TABLE_END
-// CHECK-NEXT:  }
+// CHECK-NEXT: const bool param_omit_table[] = {
+// CHECK-NEXT:   // OMIT_TABLE_BEGIN
+// CHECK-NEXT:   //--- _ZTSZ4mainE8kernel_B
+// CHECK-NEXT:   false, false, false, false, false,
+// CHECK-NEXT:   // OMIT_TABLE_END
+// CHECK-NEXT:   };
 
 // CHECK: static constexpr
 // CHECK-NEXT: const kernel_param_desc_t kernel_signatures[] = {
-// CHECK-NEXT: //--- _ZTSZ4mainE8kernel_B
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, kernel_param_used[0] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 4, kernel_param_used[1] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 8, kernel_param_used[2] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 12, kernel_param_used[3] },
-// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 16, kernel_param_used[4] },
+// CHECK-NEXT:   //--- _ZTSZ4mainE8kernel_B
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 0, param_omit_table[0]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 4, param_omit_table[1]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 8, param_omit_table[2]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 12, param_omit_table[3]},
+// CHECK-NEXT:   { kernel_param_kind_t::kind_std_layout, 4, 16, param_omit_table[4]},
 // CHECK-EMPTY:
 // CHECK-NEXT: };