[NPU] add npu kernel for sgd (#31639)

paddle/fluid/operators/optimizers/sgd_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 <memory>
+#include <string>
+
+#include "paddle/fluid/operators/npu_op_runner.h"
+#include "paddle/fluid/operators/optimizers/sgd_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class SGDNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* learning_rate = ctx.Input<framework::LoDTensor>("LearningRate");
+    auto* param_var = ctx.Input<framework::LoDTensor>("Param");
+    auto* grad_var = ctx.Input<framework::LoDTensor>("Grad");
+    auto* param_out = ctx.Output<framework::LoDTensor>("ParamOut");
+
+    param_out->mutable_data<T>(ctx.GetPlace());
+
+    auto runner =
+        NpuOpRunner("ApplyGradientDescent",
+                    {*param_var, *learning_rate, *grad_var}, {*param_out}, {});
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+
+    // NOTE(zhiqiu): ApplyGradientDescent updates params inplace, so
+    // if param and param_out is not same, we need to do copy.
+    if (param_out->data<T>() != param_var->data<T>()) {
+      ctx.template device_context<paddle::platform::NPUDeviceContext>().Wait();
+      framework::TensorCopySync(*param_var, ctx.GetPlace(), param_out);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_NPU_KERNEL(
+    sgd, ops::SGDNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::SGDNPUKernel<paddle::platform::NPUDeviceContext, double>,
+    ops::SGDNPUKernel<paddle::platform::NPUDeviceContext,
+                      paddle::platform::float16>);

python/paddle/fluid/tests/unittests/npu/test_sgd_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.
+
+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 TestSGD(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "sgd"
+        self.conf()
+        w = np.random.random((self.h, self.w)).astype("float32")
+        g = np.random.random((self.h, self.w)).astype("float32")
+        lr = np.array([0.1]).astype("float32")
+
+        self.inputs = {'Param': w, 'Grad': g, 'LearningRate': lr}
+        self.outputs = {'ParamOut': w - lr * g}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def conf(self):
+        self.h = 12
+        self.w = 15
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestNet(unittest.TestCase):
+    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=(32, 32)).astype('float32')
+        b_np = np.random.random(size=(32, 32)).astype('float32')
+        label_np = np.random.randint(2, size=(32, 1)).astype('int64')
+
+        with paddle.static.program_guard(main_prog, startup_prog):
+            a = paddle.static.data(name="a", shape=[32, 32], dtype='float32')
+            b = paddle.static.data(name="b", shape=[32, 32], dtype='float32')
+            label = paddle.static.data(
+                name="label", shape=[32, 1], dtype='int64')
+
+            sum = paddle.add(a, b)
+            z = paddle.pow(sum, 2.0)
+
+            fc_1 = fluid.layers.fc(input=z, size=128)
+            prediction = fluid.layers.fc(input=fc_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))
+
+
+if __name__ == '__main__':
+    unittest.main()