[flang][cuda] Introduce cuf.set_allocator_idx operation (#148717)

Author: clementval

flang-rt/lib/cuda/descriptor.cpp

@@ -62,6 +62,15 @@ void RTDEF(CUFDescriptorCheckSection)(
   }
 }
 
+void RTDEF(CUFSetAllocatorIndex)(
+    Descriptor *, int index, const char *sourceFile, int sourceLine) {
+  if (!desc) {
+    Terminator terminator{sourceFile, sourceLine};
+    terminator.Crash("descriptor is null");
+  }
+  desc->SetAllocIdx(index);
+}
+
 RT_EXT_API_GROUP_END
 }
 } // namespace Fortran::runtime::cuda

flang-rt/unittests/Runtime/CUDA/AllocatorCUF.cpp

@@ -72,3 +72,13 @@ TEST(AllocatableCUFTest, DescriptorAllocationTest) {
   EXPECT_TRUE(desc != nullptr);
   RTNAME(CUFFreeDescriptor)(desc);
 }
+
+TEST(AllocatableCUFTest, CUFSetAllocatorIndex) {
+  using Fortran::common::TypeCategory;
+  RTNAME(CUFRegisterAllocator)();
+  // REAL(4), DEVICE, ALLOCATABLE :: a(:)
+  auto a{createAllocatable(TypeCategory::Real, 4)};
+  EXPECT_EQ((int)kDefaultAllocator, a->GetAllocIdx());
+  RTNAME(CUFSetAllocatorIndex)(*a, kDeviceAllocatorPos, __FILE__, __LINE__);
+  EXPECT_EQ((int)kDeviceAllocatorPos, a->GetAllocIdx());
+}

flang/include/flang/Optimizer/Builder/Runtime/CUDA/Descriptor.h

@@ -31,6 +31,10 @@ void genSyncGlobalDescriptor(fir::FirOpBuilder &builder, mlir::Location loc,
 void genDescriptorCheckSection(fir::FirOpBuilder &builder, mlir::Location loc,
                                mlir::Value desc);
 
+/// Generate runtime call to set the allocator index in the descriptor.
+void genSetAllocatorIndex(fir::FirOpBuilder &builder, mlir::Location loc,
+                          mlir::Value desc, mlir::Value index);
+
 } // namespace fir::runtime::cuda
 
 #endif // FORTRAN_OPTIMIZER_BUILDER_RUNTIME_CUDA_DESCRIPTOR_H_

flang/include/flang/Optimizer/Dialect/CUF/CUFOps.td

@@ -388,4 +388,25 @@ def cuf_StreamCastOp : cuf_Op<"stream_cast", [NoMemoryEffect]> {
   let hasVerifier = 1;
 }
 
+def cuf_SetAllocatorIndexOp : cuf_Op<"set_allocator_idx", []> {
+  let summary = "Set the allocator index in a descriptor";
+
+  let description = [{
+    Allocator index in the Fortran descriptor is used to retrived the correct
+    CUDA allocator to allocate the memory on the device.
+    In many cases the allocator index is set when the descriptor is created. For
+    device components, the descriptor is part of the derived-type itself and
+    needs to be set after the derived-type is allocated in managed memory.
+  }];
+
+  let arguments = (ins Arg<fir_ReferenceType, "", [MemRead, MemWrite]>:$box,
+      cuf_DataAttributeAttr:$data_attr);
+
+  let assemblyFormat = [{
+    $box `:` qualified(type($box)) attr-dict
+  }];
+
+  let hasVerifier = 1;
+}
+
 #endif // FORTRAN_DIALECT_CUF_CUF_OPS

flang/include/flang/Runtime/CUDA/descriptor.h

@@ -41,6 +41,10 @@ void RTDECL(CUFSyncGlobalDescriptor)(
 void RTDECL(CUFDescriptorCheckSection)(
     const Descriptor *, const char *sourceFile = nullptr, int sourceLine = 0);
 
+/// Set the allocator index with the provided value.
+void RTDECL(CUFSetAllocatorIndex)(Descriptor *, int index,
+    const char *sourceFile = nullptr, int sourceLine = 0);
+
 } // extern "C"
 
 } // namespace Fortran::runtime::cuda

flang/lib/Optimizer/Builder/Runtime/CUDA/Descriptor.cpp

@@ -47,3 +47,18 @@ void fir::runtime::cuda::genDescriptorCheckSection(fir::FirOpBuilder &builder,
       builder, loc, fTy, desc, sourceFile, sourceLine)};
   builder.create<fir::CallOp>(loc, func, args);
 }
+
+void fir::runtime::cuda::genSetAllocatorIndex(fir::FirOpBuilder &builder,
+                                              mlir::Location loc,
+                                              mlir::Value desc,
+                                              mlir::Value index) {
+  mlir::func::FuncOp func =
+      fir::runtime::getRuntimeFunc<mkRTKey(CUFSetAllocatorIndex)>(loc, builder);
+  auto fTy = func.getFunctionType();
+  mlir::Value sourceFile = fir::factory::locationToFilename(builder, loc);
+  mlir::Value sourceLine =
+      fir::factory::locationToLineNo(builder, loc, fTy.getInput(3));
+  llvm::SmallVector<mlir::Value> args{fir::runtime::createArguments(
+      builder, loc, fTy, desc, index, sourceFile, sourceLine)};
+  builder.create<fir::CallOp>(loc, func, args);
+}

flang/lib/Optimizer/Dialect/CUF/CUFOps.cpp

@@ -345,6 +345,17 @@ llvm::LogicalResult cuf::StreamCastOp::verify() {
   return checkStreamType(*this);
 }
 
+//===----------------------------------------------------------------------===//
+// SetAllocatorOp
+//===----------------------------------------------------------------------===//
+
+llvm::LogicalResult cuf::SetAllocatorIndexOp::verify() {
+  if (!mlir::isa<fir::BaseBoxType>(fir::unwrapRefType(getBox().getType())))
+    return emitOpError(
+        "expect box to be a reference to class or box type value");
+  return mlir::success();
+}
+
 // Tablegen operators
 
 #define GET_OP_CLASSES

flang/lib/Optimizer/Transforms/CUFOpConversion.cpp

@@ -22,6 +22,7 @@
 #include "flang/Runtime/CUDA/memory.h"
 #include "flang/Runtime/CUDA/pointer.h"
 #include "flang/Runtime/allocatable.h"
+#include "flang/Runtime/allocator-registry-consts.h"
 #include "flang/Support/Fortran.h"
 #include "mlir/Conversion/LLVMCommon/Pattern.h"
 #include "mlir/Dialect/DLTI/DLTI.h"
@@ -923,6 +924,34 @@ struct CUFSyncDescriptorOpConversion
   }
 };
 
+struct CUFSetAllocatorIndexOpConversion
+    : public mlir::OpRewritePattern<cuf::SetAllocatorIndexOp> {
+  using OpRewritePattern::OpRewritePattern;
+
+  mlir::LogicalResult
+  matchAndRewrite(cuf::SetAllocatorIndexOp op,
+                  mlir::PatternRewriter &rewriter) const override {
+    auto mod = op->getParentOfType<mlir::ModuleOp>();
+    fir::FirOpBuilder builder(rewriter, mod);
+    mlir::Location loc = op.getLoc();
+    int idx = kDefaultAllocator;
+    if (op.getDataAttr() == cuf::DataAttribute::Device) {
+      idx = kDeviceAllocatorPos;
+    } else if (op.getDataAttr() == cuf::DataAttribute::Managed) {
+      idx = kManagedAllocatorPos;
+    } else if (op.getDataAttr() == cuf::DataAttribute::Unified) {
+      idx = kUnifiedAllocatorPos;
+    } else if (op.getDataAttr() == cuf::DataAttribute::Pinned) {
+      idx = kPinnedAllocatorPos;
+    }
+    mlir::Value index =
+        builder.createIntegerConstant(loc, builder.getI32Type(), idx);
+    fir::runtime::cuda::genSetAllocatorIndex(builder, loc, op.getBox(), index);
+    op.erase();
+    return mlir::success();
+  }
+};
+
 class CUFOpConversion : public fir::impl::CUFOpConversionBase<CUFOpConversion> {
 public:
   void runOnOperation() override {
@@ -984,8 +1013,8 @@ void cuf::populateCUFToFIRConversionPatterns(
     const mlir::SymbolTable &symtab, mlir::RewritePatternSet &patterns) {
   patterns.insert<CUFAllocOpConversion>(patterns.getContext(), &dl, &converter);
   patterns.insert<CUFAllocateOpConversion, CUFDeallocateOpConversion,
-                  CUFFreeOpConversion, CUFSyncDescriptorOpConversion>(
-      patterns.getContext());
+                  CUFFreeOpConversion, CUFSyncDescriptorOpConversion,
+                  CUFSetAllocatorIndexOpConversion>(patterns.getContext());
   patterns.insert<CUFDataTransferOpConversion>(patterns.getContext(), symtab,
                                                &dl, &converter);
   patterns.insert<CUFLaunchOpConversion, CUFDeviceAddressOpConversion>(

flang/test/Fir/CUDA/cuda-alloc-free.fir

@@ -94,4 +94,19 @@ func.func @_QQalloc_char() attributes {fir.bindc_name = "alloc_char"} {
 // CHECK: %[[BYTES_CONV:.*]] = fir.convert %[[BYTES]] : (index) -> i64
 // CHECK: fir.call @_FortranACUFMemAlloc(%[[BYTES_CONV]], %c0{{.*}}, %{{.*}}, %{{.*}}) {cuf.data_attr = #cuf.cuda<device>} : (i64, i32, !fir.ref<i8>, i32) -> !fir.llvm_ptr<i8>
 
+
+func.func @_QQsetalloc() {
+  %0 = cuf.alloc !fir.type<_QMm1Tdt1{a2:!fir.box<!fir.heap<!fir.array<?xf32>>>}> {bindc_name = "d1", data_attr = #cuf.cuda<managed>, uniq_name = "_QFEd1"} -> !fir.ref<!fir.type<_QMm1Tdt1{a2:!fir.box<!fir.heap<!fir.array<?xf32>>>}>>
+  %1 = fir.coordinate_of %0, a2 : (!fir.ref<!fir.type<_QMm1Tdt1{a2:!fir.box<!fir.heap<!fir.array<?xf32>>>}>>) -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>
+  cuf.set_allocator_idx %1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {data_attr = #cuf.cuda<device>}
+  return
+}
+
+// CHECK-LABEL:   func.func @_QQsetalloc() {
+// CHECK: %[[DT:.*]] = fir.call @_FortranACUFMemAlloc
+// CHECK: %[[CONV:.*]] = fir.convert %[[DT]] : (!fir.llvm_ptr<i8>) -> !fir.ref<!fir.type<_QMm1Tdt1{a2:!fir.box<!fir.heap<!fir.array<?xf32>>>}>>
+// CHECK: %[[COMP:.*]] = fir.coordinate_of %[[CONV]], a2 : (!fir.ref<!fir.type<_QMm1Tdt1{a2:!fir.box<!fir.heap<!fir.array<?xf32>>>}>>) -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>
+// CHECK: %[[DESC:.*]] = fir.convert %[[COMP]] : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> !fir.ref<!fir.box<none>>
+// CHECK: fir.call @_FortranACUFSetAllocatorIndex(%[[DESC]], %c2{{.*}}, %{{.*}}, %{{.*}}) : (!fir.ref<!fir.box<none>>, i32, !fir.ref<i8>, i32) -> ()
+
 } // end module