【NPU】Support NPU kernel for reduce_sum op v2 (PaddlePaddle#31620)

* add reduce_sum

* fix broadcastd

* fix test

* fix

* add unsqueeze in reduce_sum

* add template

* add unittest for keep_dim

* test reduce_all

Co-authored-by: frankwhzhang <frankwhzhang@126.com>

paddle/fluid/operators/reduce_ops/reduce_sum_op_npu.cc

@@ -0,0 +1,109 @@
+/* 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 <memory>
+#include <string>
+
+#include "paddle/fluid/operators/npu_op_runner.h"
+#include "paddle/fluid/operators/reduce_ops/reduce_op.h"
+#include "paddle/fluid/operators/unsqueeze_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class ReduceSumNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::Tensor>("X");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    bool reduce_all = ctx.Attr<bool>("reduce_all");
+    bool keep_dims = ctx.Attr<bool>("keep_dim");
+    auto dims = ctx.Attr<std::vector<int>>("dim");
+
+    out->mutable_data<T>(ctx.GetPlace());
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    if (reduce_all) {
+      std::vector<int> dim_vec;
+      for (int i = 0; i < x->dims().size(); i++) {
+        dim_vec.push_back(i);
+      }
+      auto runner = NpuOpRunner("ReduceSumD", {*x}, {*out},
+                                {{"axes", dim_vec}, {"keep_dims", keep_dims}});
+      runner.Run(stream);
+
+    } else {
+      auto runner = NpuOpRunner("ReduceSumD", {*x}, {*out},
+                                {{"axes", dims}, {"keep_dims", keep_dims}});
+      runner.Run(stream);
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class ReduceSumGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::Tensor>("X");
+    auto* out_grad =
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    bool reduce_all = ctx.Attr<bool>("reduce_all");
+    bool keep_dims = ctx.Attr<bool>("keep_dim");
+    auto dims = ctx.Attr<std::vector<int>>("dim");
+
+    x_grad->mutable_data<T>(ctx.GetPlace());
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    if (keep_dims || reduce_all) {
+      auto runner = NpuOpRunner("BroadcastToD", {*out_grad}, {*x_grad},
+                                {{"shape", framework::vectorize(x->dims())}});
+      runner.Run(stream);
+    } else {
+      framework::DDim out_dims;
+      out_dims = UnsqueezeKernel<DeviceContext, T>::GetOutputShape(
+          dims, out_grad->dims());
+
+      Tensor out_grad_tmp(out_grad->type());
+      out_grad_tmp.Resize(out_dims);
+      out_grad_tmp.mutable_data<T>(ctx.GetPlace());
+
+      auto runner = NpuOpRunner("BroadcastToD", {out_grad_tmp}, {*x_grad},
+                                {{"shape", framework::vectorize(x->dims())}});
+      runner.Run(stream);
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_NPU_KERNEL(
+    reduce_sum,
+    ops::ReduceSumNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::ReduceSumNPUKernel<paddle::platform::NPUDeviceContext, int>,
+    ops::ReduceSumNPUKernel<paddle::platform::NPUDeviceContext,
+                            paddle::platform::float16>);
+REGISTER_OP_NPU_KERNEL(
+    reduce_sum_grad,
+    ops::ReduceSumGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::ReduceSumGradNPUKernel<paddle::platform::NPUDeviceContext, int>,
+    ops::ReduceSumGradNPUKernel<paddle::platform::NPUDeviceContext,
+                                paddle::platform::float16>);

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

@@ -0,0 +1,198 @@
+#  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
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestReduceSum(OpTest):
+    def setUp(self):
+        np.random.seed(SEED)
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.init_op_type()
+        self.initTestCase()
+
+        self.use_mkldnn = False
+        self.attrs = {
+            'dim': self.axis,
+            'keep_dim': self.keep_dim,
+            'reduce_all': self.reduce_all
+        }
+        self.inputs = {'X': np.random.random(self.shape).astype("float32")}
+        if self.attrs['reduce_all']:
+            self.outputs = {'Out': self.inputs['X'].sum()}
+        else:
+            self.outputs = {
+                'Out': self.inputs['X'].sum(axis=self.axis,
+                                            keepdims=self.attrs['keep_dim'])
+            }
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def init_op_type(self):
+        self.op_type = "reduce_sum"
+        self.use_mkldnn = False
+        self.keep_dim = False
+        self.reduce_all = False
+
+    def initTestCase(self):
+        self.shape = (5, 6)
+        self.axis = (0, )
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+    # TODO(ascendrc): Add grad test
+    # def test_check_grad(self):
+    #     if self.dtype == np.float16:
+    #         return
+    #     self.check_grad(['X'], 'Out')
+    #
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestReduceSumNet(unittest.TestCase):
+    def set_reduce_sum_function(self, x):
+        # keep_dim = False
+        return paddle.fluid.layers.reduce_sum(x, dim=-1)
+
+    def _test(self, run_npu=True):
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = SEED
+        startup_prog.random_seed = SEED
+        np.random.seed(SEED)
+
+        a_np = np.random.random(size=(2, 3, 4)).astype('float32')
+        b_np = np.random.random(size=(2, 3, 4)).astype('float32')
+        label_np = np.random.randint(2, size=(2, 1)).astype('int64')
+
+        with paddle.static.program_guard(main_prog, startup_prog):
+            a = paddle.static.data(name="a", shape=[2, 3, 4], dtype='float32')
+            b = paddle.static.data(name="b", shape=[2, 3, 4], dtype='float32')
+            label = paddle.static.data(
+                name="label", shape=[2, 1], dtype='int64')
+
+            z = paddle.add(a, b)
+            z_1 = self.set_reduce_sum_function(z)
+
+            prediction = fluid.layers.fc(input=z_1, size=2, act='softmax')
+
+            cost = fluid.layers.cross_entropy(input=prediction, label=label)
+            loss = fluid.layers.reduce_mean(cost)
+            sgd = fluid.optimizer.SGD(learning_rate=0.01)
+            sgd.minimize(loss)
+
+        if run_npu:
+            place = paddle.NPUPlace(0)
+        else:
+            place = paddle.CPUPlace()
+
+        exe = paddle.static.Executor(place)
+        exe.run(startup_prog)
+
+        print("Start run on {}".format(place))
+        for epoch in range(100):
+
+            pred_res, loss_res = exe.run(
+                main_prog,
+                feed={"a": a_np,
+                      "b": b_np,
+                      "label": label_np},
+                fetch_list=[prediction, loss])
+            if epoch % 10 == 0:
+                print("Epoch {} | Prediction[0]: {}, Loss: {}".format(
+                    epoch, pred_res[0], loss_res))
+
+        return pred_res, loss_res
+
+    def test_npu(self):
+        cpu_pred, cpu_loss = self._test(False)
+        npu_pred, npu_loss = self._test(True)
+
+        self.assertTrue(np.allclose(npu_pred, cpu_pred))
+        self.assertTrue(np.allclose(npu_loss, cpu_loss))
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestReduceSumNet2(TestReduceSumNet):
+    def set_reduce_sum_function(self, x):
+        # keep_dim = True
+        return paddle.fluid.layers.reduce_sum(x, dim=-1, keep_dim=True)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestReduceSumNet3(TestReduceSumNet):
+    def _test(self, run_npu=True):
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = SEED
+        startup_prog.random_seed = SEED
+        np.random.seed(SEED)
+
+        a_np = np.random.random(size=(2, 3, 4)).astype('float32')
+        b_np = np.random.random(size=(2, 3, 4)).astype('float32')
+
+        with paddle.static.program_guard(main_prog, startup_prog):
+            a = paddle.static.data(name="a", shape=[2, 3, 4], dtype='float32')
+            b = paddle.static.data(name="b", shape=[2, 3, 4], dtype='float32')
+
+            z = paddle.add(a, b)
+            loss = fluid.layers.reduce_sum(z)
+            sgd = fluid.optimizer.SGD(learning_rate=0.01)
+            sgd.minimize(loss)
+
+        if run_npu:
+            place = paddle.NPUPlace(0)
+        else:
+            place = paddle.CPUPlace()
+
+        exe = paddle.static.Executor(place)
+        exe.run(startup_prog)
+
+        print("Start run on {}".format(place))
+        for epoch in range(100):
+
+            loss_res = exe.run(main_prog,
+                               feed={"a": a_np,
+                                     "b": b_np},
+                               fetch_list=[loss])
+            if epoch % 10 == 0:
+                print("Epoch {} | Loss: {}".format(epoch, loss_res))
+
+        return loss_res, loss_res
+
+
+if __name__ == '__main__':
+    unittest.main()