Red/21.x

clang/include/clang/Basic/Attr.td

@@ -5229,6 +5229,14 @@ def NonString : InheritableAttr {
   let Documentation = [NonStringDocs];
 }
 
+// +===== kitsune-/tapir-centric attributes
+
+def KitsuneReduction : InheritableAttr {
+  let Spellings = [Clang<"kitsune_reduction">];
+  let Subjects = SubjectList<[FunctionLike]>;
+  let Documentation = [TapirRTDocs];
+}
+
 def TapirTarget : StmtAttr {
   let Spellings = [CXX11<"tapir","target">];
 
@@ -5247,9 +5255,9 @@ def TapirTarget : StmtAttr {
                              ErrorDiag, "'parallel' statements">;
   let Args = [
     EnumArgument<"TapirTargetAttrType", "TapirTargetAttrTy", /*is_string=*/ true,
-                 ["nolo", "serial", "cuda", "hip", "opencilk", "openmp",
+                 ["nolo", "serial", "cuda", "hip", "gpu",  "opencilk", "openmp",
                   "qthreads", "realm"],
-                 ["Nolo", "Serial", "Cuda", "Hip", "OpenCilk", "OpenMP",
+                 ["Nolo", "Serial", "Cuda", "Hip", "GPU", "OpenCilk", "OpenMP",
                   "Qthreads", "Realm"],
                  0>
   ];

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -901,6 +901,9 @@ def err_drv_unsupported_option_argument_for_frontend : Error<
 def err_drv_kitsune_missing_required : Error<
   "missing required option '--%0'">;
 
+def err_drv_kitsune_tapir_required : Error<
+  "--tapir is required with '%0'">;
+
 def err_drv_kitsune_optzns_required : Error<
   "--tapir requires optimization level O1 or higher">;
 def err_drv_kitsune_lto_o2_required : Error<
@@ -944,8 +947,6 @@ def err_drv_opencilk_missing_abi_bitcode: Error<
 
 def err_drv_kitsune_kokkos_disabled : Error<
   "kokkos support was not enabled when kitsune was built">;
-def err_drv_kitsune_kokkos_no_tapir : Error<
-  "--tapir is required with '%0'">;
 
 def err_drv_kitsune_target_not_enabled: Error<
   "tapir target '%0' was not enabled when kitsune was built">;

clang/include/clang/Driver/Options.td

@@ -4283,7 +4283,7 @@ def fno_trigraphs : Flag<["-"], "fno-trigraphs">, Group<f_Group>,
   HelpText<"Do not process trigraph sequences">,
   Visibility<[ClangOption, CC1Option]>;
 def funique_source_file_names: Flag<["-"], "funique-source-file-names">, Group<f_Group>,
-  HelpText<"Allow the compiler to assume that each translation unit has a unique "                       
+  HelpText<"Allow the compiler to assume that each translation unit has a unique "
            "source file identifier (see -funique-source-file-identifier) at link time">;
 def fno_unique_source_file_names: Flag<["-"], "fno-unique-source-file-names">;
 def unique_source_file_identifier_EQ: Joined<["-"], "funique-source-file-identifier=">, Group<f_Group>,
@@ -7084,7 +7084,7 @@ defm android_pad_segment : BooleanFFlag<"android-pad-segment">, Group<f_Group>;
 def shared_libflangrt : Flag<["-"], "shared-libflangrt">,
   HelpText<"Link the flang-rt shared library">, Group<Link_Group>,
   Visibility<[FlangOption]>, Flags<[NoArgumentUnused]>;
-def static_libflangrt : Flag<["-"], "static-libflangrt">, 
+def static_libflangrt : Flag<["-"], "static-libflangrt">,
   HelpText<"Link the flang-rt static library">, Group<Link_Group>,
   Visibility<[FlangOption]>, Flags<[NoArgumentUnused]>;
 
@@ -9508,14 +9508,23 @@ def fkokkos_no_init : Flag<["-"], "fkokkos-no-init">, Alias<kokkos_no_init>,
   Visibility<[ClangOption, CC1Option]>,
   HelpText<"DEPRECATED: Use --kokkos-no-init">;
 
+// Tapir currently requires at least -O1. This makes it difficult to write tests
+// that check that the frontend generates the expected LLVM IR. If this
+// restriction is ever removed, this option will no longer be necessary since
+// we could just pass -O0 to examine the IR generated by the frontend.
+def print_before_first: Joined<["--"], "print-before-first">,
+  Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>,
+  HelpText<"Print the LLVM Module to stderr before running the optimization "
+           "pipeline">;
+
 def kitrt_verbose: Joined<["--"], "kitrt-verbose">,
   Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>,
   HelpText<"Enable verbose mode in kitsune's runtime">;
 
 def tapir_EQ : Joined<["--"], "tapir=">,
   Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>,
   MetaVarName<"<target>">,
-  Values<"none,serial,cuda,hip,opencilk">,
+  Values<"nolo,serial,cuda,hip,opencilk">,
   HelpText<"The primary tapir target">;
 
 def ftapir_EQ : Joined<["-"], "ftapir=">, Alias<tapir_EQ>,

clang/lib/CodeGen/BackendUtil.cpp

@@ -1216,6 +1216,10 @@ void EmitAssemblyHelper::RunOptimizationPipeline(
     return;
   }
 
+  const KitsuneOptions& kitOpts = CI.getKitsuneOpts();
+  if (kitOpts.hasTTID() && kitOpts.getPrintBeforeFirst())
+    llvm::errs() << *TheModule << "\n";
+
   // Now that we have all of the passes ready, run them.
   {
     PrettyStackTraceString CrashInfo("Optimizer");

clang/lib/CodeGen/CGCall.cpp

@@ -2512,6 +2512,9 @@ void CodeGenModule::ConstructAttributeList(StringRef Name,
       FuncAttrs.addMemoryAttr(llvm::MemoryEffects::inaccessibleOrArgMemOnly());
       FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
     }
+    if (TargetDecl->hasAttr<KitsuneReductionAttr>()) {
+      FuncAttrs.addAttribute(llvm::Attribute::KitsuneReduction);
+    }
     if (const auto *RA = TargetDecl->getAttr<RestrictAttr>();
         RA && RA->getDeallocator() == nullptr)
       RetAttrs.addAttribute(llvm::Attribute::NoAlias);

clang/lib/CodeGen/CGKitsune.cpp

@@ -59,6 +59,7 @@
 #include "clang/CodeGen/CGFunctionInfo.h"
 #include "clang/Frontend/FrontendDiagnostic.h"
 #include "llvm/IR/ValueMap.h"
+#include "llvm/IR/FixedPointBuilder.h"
 
 using namespace clang;
 using namespace CodeGen;
@@ -106,6 +107,8 @@ CodeGenFunction::GetTapirTargetAttr(ArrayRef<const Attr *> Attrs) {
         return llvm::TTID::Cuda;
       case TapirTargetAttr::Hip:
         return llvm::TTID::Hip;
+      case TapirTargetAttr::GPU:
+        return llvm::TTID::GPU; 
       case TapirTargetAttr::OpenCilk:
         return llvm::TTID::OpenCilk;
       case TapirTargetAttr::OpenMP:
@@ -326,6 +329,12 @@ void CodeGenFunction::EmitForallStmt(const ForallStmt &S,
   // Evaluate the initialization before the loop.
   EmitStmt(S.getInit());
 
+  // TODO: explain more
+  // We assume that the boolean is a binary operator and pre-compute RHS
+  auto* BO = dyn_cast<BinaryOperator>(S.getCond()); 
+  llvm::Value *RHS = EmitScalarExpr(BO->getRHS()); 
+
+
   // In a parallel loop there will always be a condition block
   // so there is no need to test
   JumpDest Condition = getJumpDestInCurrentScope("forall.cond");
@@ -364,7 +373,59 @@ void CodeGenFunction::EmitForallStmt(const ForallStmt &S,
 
   // C99 6.8.5p2/p4: The first substatement is executed if the expression
   // compares unequal to 0.  The condition must be a scalar type.
-  llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
+
+  //llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
+  llvm::Value *LHS = EmitScalarExpr(BO->getLHS()); 
+  QualType LHSTy = BO->getLHS()->getType();
+  llvm::Value *BoolCondVal;
+  if (LHSTy->hasSignedIntegerRepresentation()) {
+    switch(BO->getOpcode()) {
+      case BO_GT:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_SGT, LHS, RHS, "cmp");
+        break;
+      case BO_GE:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_SGE, LHS, RHS, "cmp");
+        break;
+      case BO_LT:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_SLT, LHS, RHS, "cmp");
+        break;
+      case BO_LE:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_SLE, LHS, RHS, "cmp");
+        break;
+      case BO_NE:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_NE, LHS, RHS, "cmp");
+        break;
+      case BO_EQ:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, LHS, RHS, "cmp");
+        break;
+      default:
+        llvm_unreachable("Invalid comparison in forall");
+    }
+  } else {
+    switch(BO->getOpcode()){
+      case BO_GT:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_UGT, LHS, RHS, "cmp");
+        break;
+      case BO_GE:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_UGE, LHS, RHS, "cmp");
+        break;
+      case BO_LT:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_ULT, LHS, RHS, "cmp");
+        break;
+      case BO_LE:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_ULE, LHS, RHS, "cmp");
+        break;
+      case BO_NE:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_NE, LHS, RHS, "cmp");
+        break;
+      case BO_EQ:
+        BoolCondVal = Builder.CreateICmp(llvm::ICmpInst::ICMP_EQ, LHS, RHS, "cmp");
+        break;
+      default:
+        llvm_unreachable("Invalid comparison in forall");
+    }
+  }
+
   Builder.CreateCondBr(
       BoolCondVal, Detach, Sync.getBlock(),
       createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));

clang/lib/Driver/Driver.cpp

@@ -161,6 +161,9 @@ static void CheckTTEnabled(const Driver &D, llvm::TTID TT) {
     if (!KITSUNE_REALM_ENABLED)
       D.Diag(diag::err_drv_kitsune_target_not_enabled) << llvm::toString(TT);
     return;
+  case llvm::TTID::GPU:
+    // TODO: Check this
+    return;
   case llvm::TTID::Serial:
     // The serial tapir target is always enabled
     return;
@@ -229,13 +232,17 @@ static void CheckKitsuneOptions(const Driver &D, const ArgList &Args,
 
     // If --kokkos is provided, then a tapir target must also be provided.
     if (!Args.hasArg(options::OPT_tapir_EQ)) {
-      D.Diag(diag::err_drv_kitsune_kokkos_no_tapir)
+      D.Diag(diag::err_drv_kitsune_tapir_required)
           << Args.getLastArg(options::OPT_kokkos, options::OPT_kokkos_no_init)
                  ->getSpelling();
       return;
     }
   }
 
+  if (const Arg *A = Args.getLastArg(options::OPT_print_before_first))
+    if (!Args.hasArg(options::OPT_tapir_EQ))
+      D.Diag(diag::err_drv_kitsune_tapir_required) << A->getSpelling();
+
   // Check that the -ftapir flag has a valid value. This stops us from
   // reporting multiple errors because the flag is examined in several places.
   if (const Arg *A = Args.getLastArg(options::OPT_tapir_EQ)) {

clang/lib/Driver/KitsuneOptionUtils.cpp

@@ -112,6 +112,8 @@ clang::getTapirTargetConfigFileName(const opt::ArgList &args) {
     return "qthreads.cfg";
   case TTID::Realm:
     return "realm.cfg";
+  case TTID::GPU:
+    return "gpu.cfg";
   }
   llvm_unreachable("getTapirTargetConfigFile: TTID not handled");
 }
@@ -323,6 +325,7 @@ static bool parseKitsuneTTArgs(KitsuneOptions &kitOpts, TTID tt,
   case llvm::TTID::Realm:
     return parseKitsuneRealmArgs(kitOpts, args, optTable, diags);
   case llvm::TTID::Serial:
+  case llvm::TTID::GPU: // TODO:?
     return true;
   }
   llvm_unreachable("ParseKitsuneTTArgs: TTID not handled");
@@ -335,6 +338,7 @@ bool clang::parseKitsuneArgs(KitsuneOptions &kitOpts, const char *argv0,
 
   kitOpts.setKitsuneFrontend(IsKitsuneFrontend(argv0));
   kitOpts.setStripmineLoops(args.hasArg(OPT_fstripmine));
+  kitOpts.setPrintBeforeFirst(args.hasArg(OPT_print_before_first));
 
   kitOpts.setTapirVerbose(args.hasArg(OPT_tapir_verbose));
   kitOpts.setKitrtVerbose(args.hasArg(OPT_kitrt_verbose));

clang/lib/Driver/ToolChain.cpp

@@ -2277,6 +2277,8 @@ void ToolChain::AddKitsunePreprocessorArgs(const ArgList &Args,
     switch (TT) {
     case TTID::Nolo:
       return;
+    case TTID::GPU:
+      return;
     case TTID::Cuda:
       return ExtractArgsFromString(KITSUNE_CUDA_EXTRA_PREPROCESSOR_FLAGS,
                                    CmdArgs, Args);
@@ -2356,6 +2358,8 @@ void ToolChain::AddKitsuneCompilerArgs(const ArgList &Args,
       ExtractArgsFromString(KITSUNE_OPENCILK_EXTRA_COMPILER_FLAGS, CmdArgs,
                             Args);
       return;
+    case llvm::TTID::GPU:
+      return;
     case TTID::OpenMP:
       AddKitsuneOpenMPCommonArgs(Args, CmdArgs);
       ExtractArgsFromString(KITSUNE_OPENMP_EXTRA_COMPILER_FLAGS, CmdArgs, Args);
@@ -2385,6 +2389,7 @@ void ToolChain::AddKitsuneCompilerArgs(const ArgList &Args,
 
   if (std::optional<TTID> TT = parseTapirTargetIfValid(Args)) {
     Args.AddLastArg(CmdArgs, options::OPT_ffp_contract);
+    Args.AddLastArg(CmdArgs, options::OPT_print_before_first);
     Args.AddLastArg(CmdArgs, options::OPT_kitrt_verbose);
     Args.AddLastArg(CmdArgs, options::OPT_tapir_verbose);
     Args.AddLastArg(CmdArgs, options::OPT_tapir_EQ);
@@ -2563,6 +2568,8 @@ void ToolChain::AddKitsuneLinkerArgs(const ArgList &Args,
       AddKitsuneOpenCilkLinkerArgs(Args, CmdArgs);
       ExtractArgsFromString(KITSUNE_OPENCILK_EXTRA_LINKER_FLAGS, CmdArgs, Args);
       return;
+    case TTID::GPU:
+      return;
     case TTID::OpenMP:
       AddKitsuneOpenMPLinkerArgs(Args, CmdArgs);
       ExtractArgsFromString(KITSUNE_OPENMP_EXTRA_LINKER_FLAGS, CmdArgs, Args);

clang/lib/Frontend/CompilerInvocation.cpp

@@ -4813,6 +4813,7 @@ void CompilerInvocationBase::GenerateKitsuneArgs(const KitsuneOptions &Opts,
     case llvm::TTID::Qthreads:
     case llvm::TTID::Realm:
     case llvm::TTID::Serial:
+    case llvm::TTID::GPU:
       return;
     }
     llvm_unreachable("GenerateKitsuneArgs: TTID not handled");
@@ -4849,6 +4850,9 @@ void CompilerInvocationBase::GenerateKitsuneArgs(const KitsuneOptions &Opts,
 
   if (Opts.getStripmineLoops())
     GenerateArg(Consumer, OPT_fstripmine);
+
+  if (Opts.getPrintBeforeFirst())
+    GenerateArg(Consumer, OPT_print_before_first);
 }
 
 bool CompilerInvocation::CheckKitsuneArgs(const ArgList &Args,

clang/lib/Sema/SemaDeclAttr.cpp

@@ -7902,10 +7902,12 @@ ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D, const ParsedAttr &AL,
   case ParsedAttr::AT_TypeNullable:
     handleNullableTypeAttr(S, D, AL);
     break;
-
   case ParsedAttr::AT_VTablePointerAuthentication:
     handleVTablePointerAuthentication(S, D, AL);
     break;
+  case ParsedAttr::AT_KitsuneReduction:
+    handleSimpleAttribute<KitsuneReductionAttr>(S, D, AL);
+    break;
   }
 }
 

flang/lib/Frontend/FrontendActions.cpp

@@ -1053,6 +1053,9 @@ void CodeGenAction::runOptimizationPipeline(llvm::raw_pwrite_stream &os) {
     return;
   }
 
+  if (kitsuneOpts.hasTTID() && kitsuneOpts.getPrintBeforeFirst())
+    llvm::errs() << *llvmModule << "\n";
+
   // Run the passes.
   mpm.run(*llvmModule, mam);
 

kitsune-tests/reductions/l2.c

@@ -0,0 +1,49 @@
+#include<time.h>
+#include<math.h>
+#include<stdio.h>
+#include<stdlib.h>
+#include<kitsune.h>
+#include<gpu.h>
+#include<stdint.h>
+#include<assert.h>
+#include<omp.h>
+
+reduction
+void sum(double *a, double b, double unit){
+  *a += b;
+}
+
+double l2(uint64_t n, double* a){
+  double red = 0; 
+  forall(uint64_t i=0; i<n; i++){
+    sum(&red, a[i] * a[i], 0.0); 
+  }
+
+  return sqrt(red);
+}
+
+int main(int argc, char** argv){
+  int e = argc > 1 ? atoi(argv[1]) : 28; 
+  int niter = argc > 2 ? atoi(argv[2]) : 100; 
+  uint64_t n = 1ULL<<e; 
+  double* arr = (double*)gpuManagedMalloc(sizeof(double) * n); 
+
+  forall(uint64_t i=0; i<n; i++){
+    arr[i] = i; 
+  }
+
+  l2(n, arr);
+
+  double par = 0; 
+  double before = omp_get_wtime();
+  for(int i=0; i<niter; i++){
+    par = l2(n, arr);
+  }
+  double after = omp_get_wtime(); 
+  double partime = after - before; 
+
+  printf("%f in %f s\n" , par, partime);
+  double bw = (double)((1ULL<<e) * niter * sizeof(double)) / (1000000000.0 * partime);  
+  printf("bandwidth: %f GB/s \n" , bw);
+}
+