[flang][cuda] Support any_sync and ballot_sync

[clementval]

No description provided.

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

Author: Valentin Clement (バレンタイン クレメン) (clementval)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/134135.diff

4 Files Affected:

• (modified) flang/include/flang/Optimizer/Builder/IntrinsicCall.h (+2)
• (modified) flang/lib/Optimizer/Builder/IntrinsicCall.cpp (+33-6)
• (modified) flang/module/cudadevice.f90 (+14)
• (modified) flang/test/Lower/CUDA/cuda-device-proc.cuf (+4-2)

diff --git a/flang/include/flang/Optimizer/Builder/IntrinsicCall.h b/flang/include/flang/Optimizer/Builder/IntrinsicCall.h
index a31bbd0a1bd88..4cbef141ced94 100644
--- a/flang/include/flang/Optimizer/Builder/IntrinsicCall.h
+++ b/flang/include/flang/Optimizer/Builder/IntrinsicCall.h
@@ -442,6 +442,8 @@ struct IntrinsicLibrary {
   fir::ExtendedValue genUnpack(mlir::Type, llvm::ArrayRef<fir::ExtendedValue>);
   fir::ExtendedValue genVerify(mlir::Type, llvm::ArrayRef<fir::ExtendedValue>);
   mlir::Value genVoteAllSync(mlir::Type, llvm::ArrayRef<mlir::Value>);
+  mlir::Value genVoteAnySync(mlir::Type, llvm::ArrayRef<mlir::Value>);
+  mlir::Value genVoteBallotSync(mlir::Type, llvm::ArrayRef<mlir::Value>);
 
   /// Implement all conversion functions like DBLE, the first argument is
   /// the value to convert. There may be an additional KIND arguments that
diff --git a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
index 9029ea69dd5c4..8aed288d128b6 100644
--- a/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
+++ b/flang/lib/Optimizer/Builder/IntrinsicCall.cpp
@@ -273,6 +273,10 @@ static constexpr IntrinsicHandler handlers[]{
      &I::genAny,
      {{{"mask", asAddr}, {"dim", asValue}}},
      /*isElemental=*/false},
+    {"any_sync",
+     &I::genVoteAnySync,
+     {{{"mask", asValue}, {"pred", asValue}}},
+     /*isElemental=*/false},
     {"asind", &I::genAsind},
     {"associated",
      &I::genAssociated,
@@ -335,6 +339,10 @@ static constexpr IntrinsicHandler handlers[]{
     {"atomicsubi", &I::genAtomicSub, {{{"a", asAddr}, {"v", asValue}}}, false},
     {"atomicsubl", &I::genAtomicSub, {{{"a", asAddr}, {"v", asValue}}}, false},
     {"atomicxori", &I::genAtomicXor, {{{"a", asAddr}, {"v", asValue}}}, false},
+    {"ballot_sync",
+     &I::genVoteBallotSync,
+     {{{"mask", asValue}, {"pred", asValue}}},
+     /*isElemental=*/false},
     {"bessel_jn",
      &I::genBesselJn,
      {{{"n1", asValue}, {"n2", asValue}, {"x", asValue}}},
@@ -6499,12 +6507,9 @@ IntrinsicLibrary::genMatchAllSync(mlir::Type resultType,
   return value;
 }
 
-// ALL_SYNC
-mlir::Value IntrinsicLibrary::genVoteAllSync(mlir::Type resultType,
-                                             llvm::ArrayRef<mlir::Value> args) {
-  assert(args.size() == 2);
-
-  llvm::StringRef funcName = "llvm.nvvm.vote.all.sync";
+static mlir::Value genVoteSync(fir::FirOpBuilder &builder, mlir::Location loc,
+                               llvm::StringRef funcName,
+                               llvm::ArrayRef<mlir::Value> args) {
   mlir::MLIRContext *context = builder.getContext();
   mlir::Type i32Ty = builder.getI32Type();
   mlir::FunctionType ftype =
@@ -6514,6 +6519,28 @@ mlir::Value IntrinsicLibrary::genVoteAllSync(mlir::Type resultType,
   return builder.create<fir::CallOp>(loc, funcOp, args).getResult(0);
 }
 
+// ALL_SYNC
+mlir::Value IntrinsicLibrary::genVoteAllSync(mlir::Type resultType,
+                                             llvm::ArrayRef<mlir::Value> args) {
+  assert(args.size() == 2);
+  return genVoteSync(builder, loc, "llvm.nvvm.vote.all.sync", args);
+}
+
+// ANY_SYNC
+mlir::Value IntrinsicLibrary::genVoteAnySync(mlir::Type resultType,
+                                             llvm::ArrayRef<mlir::Value> args) {
+  assert(args.size() == 2);
+  return genVoteSync(builder, loc, "llvm.nvvm.vote.any.sync", args);
+}
+
+// BALLOT_SYNC
+mlir::Value
+IntrinsicLibrary::genVoteBallotSync(mlir::Type resultType,
+                                    llvm::ArrayRef<mlir::Value> args) {
+  assert(args.size() == 2);
+  return genVoteSync(builder, loc, "llvm.nvvm.vote.ballot.sync", args);
+}
+
 // MATCH_ANY_SYNC
 mlir::Value
 IntrinsicLibrary::genMatchAnySync(mlir::Type resultType,
diff --git a/flang/module/cudadevice.f90 b/flang/module/cudadevice.f90
index 6b8aa4de74240..591e25e4108b2 100644
--- a/flang/module/cudadevice.f90
+++ b/flang/module/cudadevice.f90
@@ -1022,6 +1022,20 @@ attributes(device) integer function all_sync(mask, pred)
     end function
   end interface
 
+  interface any_sync
+    attributes(device) integer function any_sync(mask, pred)
+      !dir$ ignore_tkr(d) mask, (td) pred
+      integer, value :: mask, pred
+    end function
+  end interface
+
+  interface ballot_sync
+    attributes(device) integer function ballot_sync(mask, pred)
+      !dir$ ignore_tkr(d) mask, (td) pred
+      integer, value :: mask, pred
+    end function
+  end interface
+
   ! LDCG
   interface __ldcg
     attributes(device) pure integer(4) function __ldcg_i4(x) bind(c)
diff --git a/flang/test/Lower/CUDA/cuda-device-proc.cuf b/flang/test/Lower/CUDA/cuda-device-proc.cuf
index 9758107c84031..6a7fee73f338a 100644
--- a/flang/test/Lower/CUDA/cuda-device-proc.cuf
+++ b/flang/test/Lower/CUDA/cuda-device-proc.cuf
@@ -299,12 +299,14 @@ end
 attributes(device) subroutine testVote()
   integer :: a, ipred, mask, v32
   a = all_sync(mask, v32)
-  
+  a = any_sync(mask, v32)
+  a = ballot_sync(mask, v32)
 end subroutine
 
 ! CHECK-LABEL: func.func @_QPtestvote()
 ! CHECK: fir.call @llvm.nvvm.vote.all.sync
-
+! CHECK: fir.call @llvm.nvvm.vote.any.sync
+! CHECK: fir.call @llvm.nvvm.vote.ballot.sync
 
 ! CHECK-DAG: func.func private @__ldca_i4x4_(!fir.ref<!fir.array<4xi32>>, !fir.ref<!fir.array<4xi32>>)
 ! CHECK-DAG: func.func private @__ldcg_i4x4_(!fir.ref<!fir.array<4xi32>>, !fir.ref<!fir.array<4xi32>>)