[NPU] Support mean npu kernel (#31729)

paddle/fluid/operators/CMakeLists.txt

@@ -187,4 +187,6 @@ endif()
 
 if(WITH_ASCEND_CL)
 cc_test(gelu_op_npu_test SRCS gelu_op_npu_test.cc DEPS op_registry gelu_op scope device_context enforce executor)
+cc_test(mean_op_npu_test SRCS mean_op_npu_test.cc DEPS op_registry mean_op scope device_context enforce executor)
 endif()
+

paddle/fluid/operators/mean_op_npu.cc

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+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. */
+
+#include "paddle/fluid/operators/mean_op.h"
+#include "paddle/fluid/platform/float16.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class MeanNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto* out = ctx.Output<framework::LoDTensor>("Out");
+
+    std::vector<int> axes;
+
+    framework::NPUAttributeMap attr_input = {
+                  {"keep_dims", false},
+                  {"axes", axes}};
+
+    out->mutable_data<T>(ctx.GetPlace());
+
+    auto runner = NpuOpRunner("ReduceMeanD",
+                              {*x},
+                              {*out},
+                              attr_input);
+
+    auto stream =
+      ctx.template device_context<
+                     paddle::platform::NPUDeviceContext>()
+                .stream();
+    runner.Run(stream);
+  }
+};
+
+
+template <typename DeviceContext, typename T>
+class MeanGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto stream =
+      context.template device_context<
+                          paddle::platform::NPUDeviceContext>()
+                          .stream();
+
+    auto grad = context.Input<Tensor>(framework::GradVarName("Out"));
+
+    PADDLE_ENFORCE_EQ(grad->numel(), 1,
+                      platform::errors::InvalidArgument(
+                          "Mean Gradient Input Tensor len should be 1. But "
+                          "received Out@Grad's elements num is %d.",
+                          grad->numel()));
+
+    auto IG = context.Output<Tensor>(framework::GradVarName("X"));
+    IG->mutable_data<T>(context.GetPlace());
+
+    // ones
+    Tensor ones(grad->type());
+    ones.mutable_data<T>(IG->dims(), context.GetPlace());
+    auto runner_ones = NpuOpRunner("OnesLike", {*IG}, {ones}, {});
+    runner_ones.Run(stream);
+
+    // means
+    Tensor mean_tensor(grad->type());
+    mean_tensor.Resize({1});
+    mean_tensor.mutable_data<T>(context.GetPlace());
+    std::vector<float> mean_vec;
+    mean_vec.push_back(1.0/static_cast<float>(IG->numel()));
+    framework::TensorFromVector(mean_vec,
+                                context.device_context(),
+                                &mean_tensor);
+
+    // means mul ones
+    Tensor mean_ma(grad->type());
+    mean_ma.Resize(IG->dims());
+    mean_ma.mutable_data<T>(context.GetPlace());
+    auto runner_mul_1 = NpuOpRunner("Mul", {mean_tensor, ones}, {mean_ma}, {});
+    runner_mul_1.Run(stream);
+
+    // and mul grad
+    auto runner_mul_2 = NpuOpRunner("Mul", {mean_ma, *grad}, {*IG}, {});
+    runner_mul_2.Run(stream);
+  }
+};
+
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+REGISTER_OP_NPU_KERNEL(
+    mean,
+    ops::MeanNPUKernel<paddle::platform::NPUDeviceContext, int>,
+    ops::MeanNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::MeanNPUKernel<paddle::platform::NPUDeviceContext, double>,
+    ops::MeanNPUKernel<paddle::platform::NPUDeviceContext, plat::float16>)
+
+
+REGISTER_OP_NPU_KERNEL(
+    mean_grad,
+    ops::MeanGradNPUKernel<paddle::platform::NPUDeviceContext, int>,
+    ops::MeanGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::MeanGradNPUKernel<paddle::platform::NPUDeviceContext, double>,
+    ops::MeanGradNPUKernel<paddle::platform::NPUDeviceContext, plat::float16>)

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

+#  Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+from __future__ import print_function
+
+import numpy as np
+import unittest
+import sys
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid import core
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestMean(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "mean"
+        self.init_dtype()
+
+        x = np.random.random([1, 100]).astype(self.dtype)
+        self.inputs = {'X': x}
+
+        self.attrs = {}
+        np_out = np.mean(x)
+        self.outputs = {'Out': np_out}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+    def test_check_grad(self):
+        self.check_grad_with_place(self.place, ['X'], 'Out', check_dygraph=False)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestMeanFP16(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "mean"
+        self.init_dtype()
+
+        x = np.random.random([3, 200]).astype(self.dtype)
+        self.inputs = {'X': x}
+
+        self.attrs = {}
+        np_out = np.mean(x)
+        self.outputs = {'Out': np_out}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.__class__.no_need_check_grad = True
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+
+
+if __name__ == '__main__':
+    unittest.main()
+