[flang-commits] [flang] a9c73f6 - [flang][cuda] Add fir.cuda_alloc/fir.cuda_free operations (#90525)
via flang-commits
flang-commits at lists.llvm.org
Tue Apr 30 08:29:55 PDT 2024
Author: Valentin Clement (バレンタイン クレメン)
Date: 2024-04-30T08:29:51-07:00
New Revision: a9c73f66ce96421ba1e8952950d6df74e9043589
URL: https://github.com/llvm/llvm-project/commit/a9c73f66ce96421ba1e8952950d6df74e9043589
DIFF: https://github.com/llvm/llvm-project/commit/a9c73f66ce96421ba1e8952950d6df74e9043589.diff
LOG: [flang][cuda] Add fir.cuda_alloc/fir.cuda_free operations (#90525)
This patch introduces fir.cuda_alloc/fir.cuda_free. These operations
will be used instead of fir.alloca for local CUDA device, managed and
unified variables.
Added:
Modified:
flang/include/flang/Optimizer/Dialect/FIROps.td
flang/test/Fir/cuf.mlir
Removed:
################################################################################
diff --git a/flang/include/flang/Optimizer/Dialect/FIROps.td b/flang/include/flang/Optimizer/Dialect/FIROps.td
index 496193e25cab61..24950322c3ca9e 100644
--- a/flang/include/flang/Optimizer/Dialect/FIROps.td
+++ b/flang/include/flang/Optimizer/Dialect/FIROps.td
@@ -3247,4 +3247,58 @@ def fir_CUDADeallocateOp : fir_Op<"cuda_deallocate",
let hasVerifier = 1;
}
+def fir_CUDAAllocOp : fir_Op<"cuda_alloc", [AttrSizedOperandSegments,
+ MemoryEffects<[MemAlloc]>]> {
+ let summary = "Allocate an object on device";
+
+ let description = [{
+ This is a drop in replacement for fir.alloca and fir.allocmem for device
+ object. Any device, managed or unified object declared in an host
+ subprogram needs to be allocated in the device memory through runtime calls.
+ The fir.cuda_alloc is an abstraction to the runtime calls and works together
+ with fir.cuda_free.
+ }];
+
+ let arguments = (ins
+ TypeAttr:$in_type,
+ OptionalAttr<StrAttr>:$uniq_name,
+ OptionalAttr<StrAttr>:$bindc_name,
+ Variadic<AnyIntegerType>:$typeparams,
+ Variadic<AnyIntegerType>:$shape,
+ fir_CUDADataAttributeAttr:$cuda_attr
+ );
+
+ let results = (outs fir_ReferenceType:$ptr);
+
+ let assemblyFormat = [{
+ $in_type (`(` $typeparams^ `:` type($typeparams) `)`)?
+ (`,` $shape^ `:` type($shape) )? attr-dict `->` qualified(type($ptr))
+ }];
+
+ let builders = [
+ OpBuilder<(ins "mlir::Type":$inType, "llvm::StringRef":$uniqName,
+ "llvm::StringRef":$bindcName,
+ "fir::CUDADataAttributeAttr":$cudaAttr,
+ CArg<"mlir::ValueRange", "{}">:$typeparams,
+ CArg<"mlir::ValueRange", "{}">:$shape,
+ CArg<"llvm::ArrayRef<mlir::NamedAttribute>", "{}">:$attributes)>];
+}
+
+def fir_CUDAFreeOp : fir_Op<"cuda_free", [MemoryEffects<[MemFree]>]> {
+ let summary = "Free a device allocated object";
+
+ let description = [{
+ The fir.cuda_free operation frees the memory allocated by fir.cuda_alloc.
+ This is used for non-allocatable device, managed and unified device
+ variables declare in host subprogram.
+ }];
+
+ let arguments = (ins
+ Arg<AnyReferenceLike, "", [MemFree]>:$devptr,
+ fir_CUDADataAttributeAttr:$cuda_attr
+ );
+
+ let assemblyFormat = "$devptr `:` qualified(type($devptr)) attr-dict";
+}
+
#endif
diff --git a/flang/test/Fir/cuf.mlir b/flang/test/Fir/cuf.mlir
index 71f0652067facf..8e2346def43ea8 100644
--- a/flang/test/Fir/cuf.mlir
+++ b/flang/test/Fir/cuf.mlir
@@ -74,3 +74,15 @@ func.func @_QPsub1() {
// CHECK: fir.cuda_allocate %{{.*}} : !fir.ref<!fir.box<none>> errmsg(%{{.*}} : !fir.box<none>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
// CHECK: fir.cuda_deallocate %{{.*}} : !fir.ref<!fir.box<none>> errmsg(%{{.*}} : !fir.box<none>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
+
+// -----
+
+func.func @_QPsub1() {
+ %0 = fir.cuda_alloc f32 {bindc_name = "r", cuda_attr = #fir.cuda<device>, uniq_name = "_QFsub1Er"} -> !fir.ref<f32>
+ fir.cuda_free %0 : !fir.ref<f32> {cuda_attr = #fir.cuda<device>}
+ return
+}
+
+// CHECK: fir.cuda_alloc
+// CHECK: fir.cuda_free
+
More information about the flang-commits
mailing list