[flang-commits] [flang] [flang][cuda] Add fir.deallocate operation (PR #88839)

Valentin Clement バレンタイン クレメン via flang-commits flang-commits at lists.llvm.org
Mon Apr 15 20:58:57 PDT 2024


https://github.com/clementval created https://github.com/llvm/llvm-project/pull/88839

Add the fir.cuda_deallocate operation that perform device deallocation of data hold by a descriptor. This will replace the call to AllocatableDeallocate from the runtime. 

This is a companion operation to the one added in #88586 

>From b29a6f7759cc986f7c460d7d2d7612f6c65adc19 Mon Sep 17 00:00:00 2001
From: Valentin Clement <clementval at gmail.com>
Date: Mon, 15 Apr 2024 14:31:00 -0700
Subject: [PATCH] [flang][cuda] Add fir.deallocate operation

---
 .../include/flang/Optimizer/Dialect/FIROps.td | 25 +++++++++++++
 flang/lib/Optimizer/Dialect/FIROps.cpp        | 13 +++++++
 flang/test/Fir/cuf-invalid.fir                | 37 +++++++++++++++++++
 flang/test/Fir/cuf.mlir                       |  6 +++
 4 files changed, 81 insertions(+)

diff --git a/flang/include/flang/Optimizer/Dialect/FIROps.td b/flang/include/flang/Optimizer/Dialect/FIROps.td
index c181c7ed62dff3..0d7f45eb6255e9 100644
--- a/flang/include/flang/Optimizer/Dialect/FIROps.td
+++ b/flang/include/flang/Optimizer/Dialect/FIROps.td
@@ -3222,4 +3222,29 @@ def fir_CUDAAllocateOp : fir_Op<"cuda_allocate", [AttrSizedOperandSegments,
   let hasVerifier = 1;
 }
 
+def fir_CUDADeallocateOp : fir_Op<"cuda_deallocate",
+    [MemoryEffects<[MemFree<DefaultResource>]>]> {
+  let summary = "Perform the device deallocation of data of an allocatable";
+
+  let description = [{
+    The fir.cuda_deallocate operation performs the deallocation on the device
+    of the data of an allocatable.
+  }];
+
+  let arguments = (ins Arg<AnyRefOrBoxType, "", [MemWrite]>:$box,
+                       Arg<Optional<AnyRefOrBoxType>, "", [MemWrite]>:$errmsg,
+                       fir_CUDADataAttributeAttr:$cuda_attr,
+                       UnitAttr:$hasStat);
+
+  let results = (outs AnyIntegerType:$stat);
+
+  let assemblyFormat = [{
+    $box `:` qualified(type($box))
+    ( `errmsg` `(` $errmsg^ `:` type($errmsg) `)` )?
+    attr-dict `->` type($stat)
+  }];
+
+  let hasVerifier = 1;
+}
+
 #endif
diff --git a/flang/lib/Optimizer/Dialect/FIROps.cpp b/flang/lib/Optimizer/Dialect/FIROps.cpp
index 88710880174d21..927e167a9a8b66 100644
--- a/flang/lib/Optimizer/Dialect/FIROps.cpp
+++ b/flang/lib/Optimizer/Dialect/FIROps.cpp
@@ -4012,6 +4012,19 @@ mlir::LogicalResult fir::CUDAAllocateOp::verify() {
   return mlir::success();
 }
 
+mlir::LogicalResult fir::CUDADeallocateOp::verify() {
+  if (!fir::unwrapRefType(getBox().getType()).isa<fir::BaseBoxType>())
+    return emitOpError(
+        "expect box to be a reference to/or a class or box type value");
+  if (getErrmsg() &&
+      !fir::unwrapRefType(getErrmsg().getType()).isa<fir::BoxType>())
+    return emitOpError(
+        "expect errmsg to be a reference to/or a box type value");
+  if (getErrmsg() && !getHasStat())
+    return emitOpError("expect stat attribute when errmsg is provided");
+  return mlir::success();
+}
+
 //===----------------------------------------------------------------------===//
 // FIROpsDialect
 //===----------------------------------------------------------------------===//
diff --git a/flang/test/Fir/cuf-invalid.fir b/flang/test/Fir/cuf-invalid.fir
index 9c5ffe7176a3bd..ee31515134c469 100644
--- a/flang/test/Fir/cuf-invalid.fir
+++ b/flang/test/Fir/cuf-invalid.fir
@@ -48,3 +48,40 @@ func.func @_QPsub1() {
   %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> errmsg(%1 : !fir.ref<i32>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
   return
 }
+
+// -----
+
+func.func @_QPsub1() {
+  %1 = fir.alloca i32
+  // expected-error at +1{{'fir.cuda_deallocate' op expect box to be a reference to/or a class or box type value}}
+  %2 = fir.cuda_deallocate %1 : !fir.ref<i32> {cuda_attr = #fir.cuda<device>} -> i32
+  return
+}
+
+// -----
+
+func.func @_QPsub1() {
+  %0 = fir.alloca !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "a", uniq_name = "_QFsub1Ea"}
+  %4:2 = hlfir.declare %0 {cuda_attr = #fir.cuda<device>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>)
+  %1 = fir.alloca i32
+  %11 = fir.convert %4#1 : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> !fir.ref<!fir.box<none>>
+  // expected-error at +1{{'fir.cuda_deallocate' op expect errmsg to be a reference to/or a box type value}}
+  %13 = fir.cuda_deallocate %11 : !fir.ref<!fir.box<none>> errmsg(%1 : !fir.ref<i32>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
+  return
+}
+
+// -----
+
+func.func @_QPsub1() {
+  %0 = fir.alloca !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "a", uniq_name = "_QFsub1Ea"}
+  %4:2 = hlfir.declare %0 {cuda_attr = #fir.cuda<device>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>)
+  %c100 = arith.constant 100 : index
+  %7 = fir.alloca !fir.char<1,100> {bindc_name = "msg", uniq_name = "_QFsub1Emsg"}
+  %8:2 = hlfir.declare %7 typeparams %c100 {uniq_name = "_QFsub1Emsg"} : (!fir.ref<!fir.char<1,100>>, index) -> (!fir.ref<!fir.char<1,100>>, !fir.ref<!fir.char<1,100>>)
+  %9 = fir.embox %8#1 : (!fir.ref<!fir.char<1,100>>) -> !fir.box<!fir.char<1,100>>
+  %11 = fir.convert %4#1 : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> !fir.ref<!fir.box<none>>
+  %16 = fir.convert %9 : (!fir.box<!fir.char<1,100>>) -> !fir.box<none>
+  // expected-error at +1{{'fir.cuda_deallocate' op expect stat attribute when errmsg is provided}}
+  %13 = fir.cuda_deallocate %11 : !fir.ref<!fir.box<none>> errmsg(%16 : !fir.box<none>) {cuda_attr = #fir.cuda<device>} -> i32
+  return
+}
diff --git a/flang/test/Fir/cuf.mlir b/flang/test/Fir/cuf.mlir
index 67eff31b35b2b8..71f0652067facf 100644
--- a/flang/test/Fir/cuf.mlir
+++ b/flang/test/Fir/cuf.mlir
@@ -7,10 +7,12 @@ func.func @_QPsub1() {
   %4:2 = hlfir.declare %0 {cuda_attr = #fir.cuda<device>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>)
   %11 = fir.convert %4#1 : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> !fir.ref<!fir.box<none>>
   %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> {cuda_attr = #fir.cuda<device>} -> i32
+  %14 = fir.cuda_deallocate %11 : !fir.ref<!fir.box<none>> {cuda_attr = #fir.cuda<device>} -> i32
   return
 }
 
 // CHECK: fir.cuda_allocate %{{.*}} : !fir.ref<!fir.box<none>> {cuda_attr = #fir.cuda<device>} -> i32
+// CHECK: fir.cuda_deallocate %{{.*}} : !fir.ref<!fir.box<none>> {cuda_attr = #fir.cuda<device>} -> i32
 
 // -----
 
@@ -66,5 +68,9 @@ func.func @_QPsub1() {
   %11 = fir.convert %4#1 : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> !fir.ref<!fir.box<none>>
   %16 = fir.convert %9 : (!fir.box<!fir.char<1,100>>) -> !fir.box<none>
   %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> errmsg(%16 : !fir.box<none>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
+  %14 = fir.cuda_deallocate %11 : !fir.ref<!fir.box<none>> errmsg(%16 : !fir.box<none>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
   return
 }
+
+// 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



More information about the flang-commits mailing list