diff --git a/flang/include/flang/Optimizer/Dialect/FIROps.td b/flang/include/flang/Optimizer/Dialect/FIROps.td
index dff1cdb20cbfe..c181c7ed62dff 100644
--- a/flang/include/flang/Optimizer/Dialect/FIROps.td
+++ b/flang/include/flang/Optimizer/Dialect/FIROps.td
@@ -3190,4 +3190,36 @@ def fir_CUDADataTransferOp : fir_Op<"cuda_data_transfer", []> {
   }];
 }
 
+def fir_CUDAAllocateOp : fir_Op<"cuda_allocate", [AttrSizedOperandSegments,
+    MemoryEffects<[MemAlloc<DefaultResource>]>]> {
+  let summary = "Perform the device allocation of data of an allocatable";
+
+  let description = [{
+    The fir.cuda_allocate operation performs the allocation on the device
+    of the data of an allocatable. The descriptor passed to the operation
+    is initialized before with the standard flang runtime calls.
+  }];
+
+  let arguments = (ins Arg<AnyRefOrBoxType, "", [MemWrite]>:$box,
+                       Arg<Optional<AnyRefOrBoxType>, "", [MemWrite]>:$errmsg,
+                       Optional<AnyIntegerType>:$stream,
+                       Arg<Optional<AnyRefOrBoxType>, "", [MemWrite]>:$pinned,
+                       Arg<Optional<AnyRefOrBoxType>, "", [MemRead]>:$source,
+                       fir_CUDADataAttributeAttr:$cuda_attr,
+                       UnitAttr:$hasStat);
+
+  let results = (outs AnyIntegerType:$stat);
+
+  let assemblyFormat = [{
+    $box `:` qualified(type($box))
+    ( `source` `(` $source^ `:` qualified(type($source) )`)` )?
+    ( `errmsg` `(` $errmsg^ `:` type($errmsg) `)` )?
+    ( `stream` `(` $stream^ `:` type($stream) `)` )?
+    ( `pinned` `(` $pinned^ `:` type($pinned) `)` )?
+    attr-dict `->` type($stat)
+  }];
+
+  let hasVerifier = 1;
+}
+
 #endif
diff --git a/flang/include/flang/Optimizer/Dialect/FIRTypes.td b/flang/include/flang/Optimizer/Dialect/FIRTypes.td
index 4c6a8064991ab..3b876e4642da9 100644
--- a/flang/include/flang/Optimizer/Dialect/FIRTypes.td
+++ b/flang/include/flang/Optimizer/Dialect/FIRTypes.td
@@ -625,6 +625,7 @@ def AnyRefOrBoxLike : TypeConstraint<Or<[AnyReferenceLike.predicate,
 def AnyRefOrBox : TypeConstraint<Or<[fir_ReferenceType.predicate,
     fir_HeapType.predicate, fir_PointerType.predicate,
     IsBaseBoxTypePred]>, "any reference or box">;
+def AnyRefOrBoxType : Type<AnyRefOrBox.predicate, "any legal ref or box type">;
 
 def AnyShapeLike : TypeConstraint<Or<[fir_ShapeType.predicate,
     fir_ShapeShiftType.predicate]>, "any legal shape type">;
diff --git a/flang/lib/Optimizer/Dialect/FIROps.cpp b/flang/lib/Optimizer/Dialect/FIROps.cpp
index 8ab74103cb6a8..88710880174d2 100644
--- a/flang/lib/Optimizer/Dialect/FIROps.cpp
+++ b/flang/lib/Optimizer/Dialect/FIROps.cpp
@@ -3993,6 +3993,25 @@ mlir::LogicalResult fir::CUDAKernelOp::verify() {
   return mlir::success();
 }
 
+mlir::LogicalResult fir::CUDAAllocateOp::verify() {
+  if (getPinned() && getStream())
+    return emitOpError("pinned and stream cannot appears at the same time");
+  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 (getSource() &&
+      !fir::unwrapRefType(getSource().getType()).isa<fir::BaseBoxType>())
+    return emitOpError(
+        "expect source 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
new file mode 100644
index 0000000000000..9c5ffe7176a3b
--- /dev/null
+++ b/flang/test/Fir/cuf-invalid.fir
@@ -0,0 +1,50 @@
+// RUN: fir-opt -split-input-file -verify-diagnostics %s
+
+func.func @_QPsub1() {
+  %0 = fir.alloca !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "a", uniq_name = "_QFsub1Ea"}
+  %1 = fir.alloca i32
+  %pinned = fir.alloca i1
+  %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>>
+  %s = fir.load %1 : !fir.ref<i32>
+  // expected-error@+1{{'fir.cuda_allocate' op pinned and stream cannot appears at the same time}}
+  %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> stream(%s : i32) pinned(%pinned : !fir.ref<i1>) {cuda_attr = #fir.cuda<device>} -> i32
+  return
+}
+
+// -----
+
+func.func @_QPsub1() {
+  %1 = fir.alloca i32
+  // expected-error@+1{{'fir.cuda_allocate' op expect box to be a reference to/or a class or box type value}}
+  %2 = fir.cuda_allocate %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>>>>)
+  %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@+1{{'fir.cuda_allocate' op expect stat attribute when errmsg is provided}}
+  %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> errmsg(%16 : !fir.box<none>) {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@+1{{'fir.cuda_allocate' op expect errmsg to be a reference to/or a box type value}}
+  %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> errmsg(%1 : !fir.ref<i32>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
+  return
+}
diff --git a/flang/test/Fir/cuf.mlir b/flang/test/Fir/cuf.mlir
new file mode 100644
index 0000000000000..67eff31b35b2b
--- /dev/null
+++ b/flang/test/Fir/cuf.mlir
@@ -0,0 +1,70 @@
+// RUN: fir-opt --split-input-file %s | fir-opt --split-input-file | FileCheck %s
+
+// Simple round trip test of operations.
+
+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>>>>)
+  %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
+  return
+}
+
+// CHECK: fir.cuda_allocate %{{.*}} : !fir.ref<!fir.box<none>> {cuda_attr = #fir.cuda<device>} -> i32
+
+// -----
+
+func.func @_QPsub1() {
+  %0 = fir.alloca !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "a", uniq_name = "_QFsub1Ea"}
+  %1 = fir.alloca i32
+  %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>>
+  %s = fir.load %1 : !fir.ref<i32>
+  %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> stream(%s : i32) {cuda_attr = #fir.cuda<device>} -> i32
+  return
+}
+
+// CHECK: fir.cuda_allocate %{{.*}} : !fir.ref<!fir.box<none>> stream(%{{.*}} : i32) {cuda_attr = #fir.cuda<device>} -> i32
+
+// -----
+
+func.func @_QPsub1() {
+  %0 = fir.alloca !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "a", uniq_name = "_QFsub1Ea"}
+  %1 = fir.alloca !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "b", uniq_name = "_QFsub1Eb"}
+  %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>>>>)
+  %5:2 = hlfir.declare %1 {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>>
+  %12 = fir.convert %5#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>> source(%12 : !fir.ref<!fir.box<none>>) {cuda_attr = #fir.cuda<device>} -> i32
+  return
+}
+
+// CHECK: fir.cuda_allocate %{{.*}} : !fir.ref<!fir.box<none>> source(%{{.*}} : !fir.ref<!fir.box<none>>) {cuda_attr = #fir.cuda<device>} -> i32
+
+// -----
+
+func.func @_QPsub1() {
+  %0 = fir.alloca !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "a", uniq_name = "_QFsub1Ea"}
+  %pinned = fir.alloca i1
+  %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>> pinned(%pinned : !fir.ref<i1>) {cuda_attr = #fir.cuda<device>} -> i32
+  return
+}
+
+// CHECK: fir.cuda_allocate %{{.*}} : !fir.ref<!fir.box<none>> pinned(%{{.*}} : !fir.ref<i1>) {cuda_attr = #fir.cuda<device>} -> i32
+
+// -----
+
+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>
+  %13 = fir.cuda_allocate %11 : !fir.ref<!fir.box<none>> errmsg(%16 : !fir.box<none>) {cuda_attr = #fir.cuda<device>, hasStat} -> i32
+  return
+}