[NPU] fix cast op (#32121)

* fix npu kernel of cast op to handle casting to same dtype

* add comments

paddle/fluid/operators/cast_op_npu.cc

@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#ifdef PADDLE_WITH_ASCEND_CL
 #include <memory>
 #include <string>
 
@@ -41,12 +40,22 @@ class CastNPUKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* x = ctx.Input<Tensor>("X");
     int dtype = ctx.Attr<int>("out_dtype");
-
     auto* out = ctx.Output<Tensor>("Out");
-
     auto place = ctx.GetPlace();
 
-    auto iter = DTYPE_2_ACL_DTYPE.find(static_cast<framework::proto::VarType::Type>(dtype));
+    if (x->type() == dtype) {
+      // NOTE(zhiqiu): NPU cast op may result in wrong value, so
+      // add special case here.
+      VLOG(4) << "cast to same dtype:" << dtype;
+      out->mutable_data(place, x->type());
+      framework::TensorCopy(
+          *x, ctx.GetPlace(),
+          ctx.template device_context<platform::DeviceContext>(), out);
+      return;
+    }
+
+    auto iter = DTYPE_2_ACL_DTYPE.find(
+        static_cast<framework::proto::VarType::Type>(dtype));
     int aclDtype = iter->second;
 
     if (dtype == framework::proto::VarType::FP32) {
@@ -69,18 +78,18 @@ class CastNPUKernel : public framework::OpKernel<T> {
         ctx.template device_context<paddle::platform::NPUDeviceContext>()
             .stream();
 
-    auto runner = NpuOpRunner("Cast", {*x}, {*out}, {{"dst_type", static_cast<int32_t>(aclDtype)}});
+    auto runner = NpuOpRunner("Cast", {*x}, {*out},
+                              {{"dst_type", static_cast<int32_t>(aclDtype)}});
     runner.Run(stream);
   }
 };
 }  // namespace operators
-}  // namespace paddleaclDtype
+}  // namespace paddle
 
 namespace ops = paddle::operators;
 
 REGISTER_OP_NPU_KERNEL(
-    cast,
-    ops::CastNPUKernel<paddle::platform::NPUDeviceContext, int16_t>,
+    cast, ops::CastNPUKernel<paddle::platform::NPUDeviceContext, int16_t>,
     ops::CastNPUKernel<paddle::platform::NPUDeviceContext, int32_t>,
     ops::CastNPUKernel<paddle::platform::NPUDeviceContext, int64_t>,
     ops::CastNPUKernel<paddle::platform::NPUDeviceContext, int>,
@@ -89,4 +98,3 @@ REGISTER_OP_NPU_KERNEL(
     ops::CastNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::CastNPUKernel<paddle::platform::NPUDeviceContext,
                        paddle::platform::float16>);
-#endif

python/paddle/fluid/tests/unittests/npu/test_cast_op_npu.py

@@ -50,6 +50,7 @@ class TestCast1(OpTest):
     def test_check_output(self):
         self.check_output_with_place(self.place, check_dygraph=False)
 
+
 class TestCast2(OpTest):
     def setUp(self):
         self.set_npu()
@@ -71,5 +72,28 @@ class TestCast2(OpTest):
     def test_check_output(self):
         self.check_output_with_place(self.place, check_dygraph=False, atol=1e-3)
 
+
+class TestCast3(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "cast"
+        self.place = paddle.NPUPlace(0)
+
+        ipt = np.random.random(size=[10, 10]) + 1
+        self.inputs = {'X': ipt.astype('int32')}
+        self.outputs = {'Out': ipt.astype('int32')}
+
+        self.attrs = {
+            'in_dtype': int(core.VarDesc.VarType.INT32),
+            'out_dtype': int(core.VarDesc.VarType.INT32)
+        }
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False, atol=1e-3)
+
+
 if __name__ == '__main__':
     unittest.main()