diff --git a/lib/Transforms/InsertGPUAllocs.cpp b/lib/Transforms/InsertGPUAllocs.cpp
index 893be344f..bab5ad173 100644
--- a/lib/Transforms/InsertGPUAllocs.cpp
+++ b/lib/Transforms/InsertGPUAllocs.cpp
@@ -360,8 +360,10 @@ class InsertGPUAllocsPass final
             auto newAlloc = builder.create<mlir::memref::AllocOp>(
                 loc, alloc.getType(), alloc.getDynamicSizes(),
                 alloc.getSymbolOperands());
-            builder.create<mlir::memref::CopyOp>(loc, allocResult,
-                                                 newAlloc.getResult());
+            builder.create<mlir::gpu::MemcpyOp>(
+                loc, /*asyncToken*/ static_cast<mlir::Type>(nullptr),
+                /*asyncDependencies*/ std::nullopt, newAlloc.getResult(),
+                allocResult);
             use.set(newAlloc.getResult());
           }
         }
@@ -401,8 +403,9 @@ class InsertGPUAllocsPass final
             /*symbolOperands*/ std::nullopt, hostShared);
         auto allocResult = gpuAlloc.getResult(0);
         if (access.hostWrite && access.deviceRead) {
-          auto copy =
-              builder.create<mlir::memref::CopyOp>(loc, op, allocResult);
+          auto copy = builder.create<mlir::gpu::MemcpyOp>(
+              loc, /*asyncToken*/ static_cast<mlir::Type>(nullptr),
+              /*asyncDependencies*/ std::nullopt, allocResult, op);
           filter.insert(copy);
         }
 
@@ -421,7 +424,9 @@ class InsertGPUAllocsPass final
           op.replaceAllUsesExcept(allocResult, filter);
           builder.setInsertionPoint(term);
           if (access.hostRead && access.deviceWrite) {
-            builder.create<mlir::memref::CopyOp>(loc, allocResult, op);
+            builder.create<mlir::gpu::MemcpyOp>(
+                loc, /*asyncToken*/ static_cast<mlir::Type>(nullptr),
+                /*asyncDependencies*/ std::nullopt, op, allocResult);
           }
           builder.create<mlir::gpu::DeallocOp>(loc, std::nullopt, allocResult);
         }
diff --git a/test/Transforms/InsertGpuAllocs/add-gpu-alloc.mlir b/test/Transforms/InsertGpuAllocs/add-gpu-alloc.mlir
index f7beea259..bc9b661bd 100644
--- a/test/Transforms/InsertGpuAllocs/add-gpu-alloc.mlir
+++ b/test/Transforms/InsertGpuAllocs/add-gpu-alloc.mlir
@@ -7,9 +7,9 @@ func.func @addt(%arg0: memref<2x5xf32>, %arg1: memref<2x5xf32>) -> memref<2x5xf3
   %c1 = arith.constant 1 : index
   %c5 = arith.constant 5 : index
   // OPENCL: %[[MEMREF0:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
-  // OPENCL: memref.copy %arg1, %[[MEMREF0]] : memref<2x5xf32> to memref<2x5xf32>
+  // OPENCL: gpu.memcpy %[[MEMREF0]], %arg1 : memref<2x5xf32>, memref<2x5xf32>
   // OPENCL: %[[MEMREF1:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
-  // OPENCL: memref.copy %arg0, %[[MEMREF1]] : memref<2x5xf32> to memref<2x5xf32>
+  // OPENCL: gpu.memcpy %[[MEMREF1]], %arg0 : memref<2x5xf32>, memref<2x5xf32>
   // VULKAN: %[[MEMREF0:.*]] = memref.alloc() : memref<2x5xf32>
   // VULKAN: memref.copy %arg1, %[[MEMREF0]] : memref<2x5xf32> to memref<2x5xf32>
   // VULKAN: %[[MEMREF1:.*]] = memref.alloc() : memref<2x5xf32>
diff --git a/test/Transforms/InsertGpuAllocs/memref-get-global.mlir b/test/Transforms/InsertGpuAllocs/memref-get-global.mlir
index 2d94ebfee..5e32a61a9 100644
--- a/test/Transforms/InsertGpuAllocs/memref-get-global.mlir
+++ b/test/Transforms/InsertGpuAllocs/memref-get-global.mlir
@@ -17,10 +17,10 @@ func.func @addt(%arg0: memref<2x5xf32>, %arg1: memref<2x5xf32>) -> memref<2x5xf3
 
   // OPENCL: [[VAR0:%.*]] = memref.get_global @__constant_2x5xf32 : memref<2x5xf32>
   // OPENCL: %[[MEMREF0:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
-  // OPENCL: memref.copy [[VAR0]], %[[MEMREF0]] : memref<2x5xf32> to memref<2x5xf32>
+  // OPENCL: gpu.memcpy %[[MEMREF0]], [[VAR0]] : memref<2x5xf32>, memref<2x5xf32>
   // OPENCL: [[VAR1:%.*]] = memref.get_global @__constant_2x5xf32_0 : memref<2x5xf32>
   // OPENCL: %[[MEMREF1:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
-  // OPENCL: memref.copy [[VAR1]], %[[MEMREF1]] : memref<2x5xf32> to memref<2x5xf32>
+  // OPENCL: gpu.memcpy %[[MEMREF1]], [[VAR1]] : memref<2x5xf32>, memref<2x5xf32>
   // OPENCL: %[[MEMREF2:.*]] = gpu.alloc host_shared () : memref<2x5xf32>
   // VULKAN: [[VAR0:%.*]] = memref.get_global @__constant_2x5xf32 : memref<2x5xf32>
   // VULKAN: %[[MEMREF0:.*]] = memref.alloc() : memref<2x5xf32>
diff --git a/test/Transforms/InsertGpuAllocs/memref-returned-from-call.mlir b/test/Transforms/InsertGpuAllocs/memref-returned-from-call.mlir
index 62af82f4f..3d0e3d50e 100644
--- a/test/Transforms/InsertGpuAllocs/memref-returned-from-call.mlir
+++ b/test/Transforms/InsertGpuAllocs/memref-returned-from-call.mlir
@@ -12,7 +12,7 @@ func.func @main() {
   // OPENCL: func.func @main()
   %0 = func.call @alloc_buffer() :  () -> memref<8xf32>
   // OPENCL: %[[MEMREF:.*]] = gpu.alloc  host_shared () : memref<8xf32>
-  // OPENCL: memref.copy %0, %[[MEMREF]] : memref<8xf32> to memref<8xf32>
+  // OPENCL: gpu.memcpy %[[MEMREF]], %0 : memref<8xf32>, memref<8xf32>
   %1 = memref.alloc() : memref<8xf32>
   %2 = memref.alloc() : memref<8xf32>
   gpu.launch blocks(%arg0, %arg1, %arg2) in (%arg6 = %c8, %arg7 = %c1, %arg8 = %c1) threads(%arg3, %arg4, %arg5) in (%arg9 = %c1, %arg10 = %c1, %arg11 = %c1) {