[NPU] Support npu op log, log_grad, sqrt, sqrt_grad, square, tanh and tanh_grad (#31600)

paddle/fluid/operators/activation_op_npu.cc

@@ -143,23 +143,186 @@ class ReluGradNPUKernel : public framework::OpKernel<T> {
     runner.Run(stream);
   }
 };
+
+template <typename DeviceContext, typename T>
+class SqrtNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto place = ctx.GetPlace();
+
+    out->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    auto runner = NpuOpRunner("Sqrt", {*x}, {*out}, {});
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class SqrtGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* out = ctx.Input<Tensor>("Out");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    auto place = ctx.GetPlace();
+
+    dx->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    auto dx_runner = NpuOpRunner("SqrtGrad", {*out, *dout}, {*dx}, {});
+    dx_runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class LogNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto place = ctx.GetPlace();
+
+    out->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    Tensor one(x->type());
+    one.mutable_data<T>(x->dims(), place);
+    auto one_runner = NpuOpRunner("OnesLike", {*x}, {one}, {});
+    one_runner.Run(stream);
+
+    Tensor sub(x->type());
+    sub.mutable_data<T>(x->dims(), place);
+    auto sub_runner = NpuOpRunner("Sub", {*x, one}, {sub}, {});
+    sub_runner.Run(stream);
+
+    auto out_runner = NpuOpRunner("Log1p", {sub}, {*out}, {});
+    out_runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class LogGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* x = ctx.Input<Tensor>("X");
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    auto place = ctx.GetPlace();
+
+    dx->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    auto runner = NpuOpRunner("DivNoNan", {*dout, *x}, {*dx}, {});
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class TanhNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto place = ctx.GetPlace();
+
+    out->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    auto runner = NpuOpRunner("Tanh", {*x}, {*out}, {});
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class TanhGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* out = ctx.Input<Tensor>("Out");
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    auto place = ctx.GetPlace();
+
+    dx->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    auto dx_runner = NpuOpRunner("TanhGrad", {*out, *dout}, {*dx}, {});
+    dx_runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class SquareNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto place = ctx.GetPlace();
+
+    out->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    auto runner = NpuOpRunner("Square", {*x}, {*out}, {});
+    runner.Run(stream);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 
 REGISTER_OP_NPU_KERNEL(
-    pow, ops::PowNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    pow,
+    ops::PowNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::PowNPUKernel<paddle::platform::NPUDeviceContext,
                       paddle::platform::float16>);
 
 REGISTER_OP_NPU_KERNEL(
-    pow_grad, ops::PowGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    pow_grad,
+    ops::PowGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::PowGradNPUKernel<paddle::platform::NPUDeviceContext,
                           paddle::platform::float16>);
 
 REGISTER_OP_NPU_KERNEL(
-    relu, ops::ReluNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    relu,
+    ops::ReluNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::ReluNPUKernel<paddle::platform::NPUDeviceContext,
                        paddle::platform::float16>);
 
@@ -168,3 +331,46 @@ REGISTER_OP_NPU_KERNEL(
     ops::ReluGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::ReluGradNPUKernel<paddle::platform::NPUDeviceContext,
                            paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    sqrt,
+    ops::SqrtNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::SqrtNPUKernel<paddle::platform::NPUDeviceContext,
+                      paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    sqrt_grad,
+    ops::SqrtGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::SqrtGradNPUKernel<paddle::platform::NPUDeviceContext,
+                          paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    log,
+    ops::LogNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::LogNPUKernel<paddle::platform::NPUDeviceContext,
+                      paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    log_grad,
+    ops::LogGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::LogGradNPUKernel<paddle::platform::NPUDeviceContext,
+                          paddle::platform::float16>);
+
+
+REGISTER_OP_NPU_KERNEL(
+    tanh,
+    ops::TanhNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::TanhNPUKernel<paddle::platform::NPUDeviceContext,
+                       paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    tanh_grad,
+    ops::TanhGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::TanhGradNPUKernel<paddle::platform::NPUDeviceContext,
+                           paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    square,
+    ops::SquareNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::SquareNPUKernel<paddle::platform::NPUDeviceContext,
+                         paddle::platform::float16>);

python/paddle/fluid/tests/unittests/npu/test_log_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
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestLog(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "log"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
+        out = np.log(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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)
+
+    # 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 TestLogFp16(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "log"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
+        out = np.log(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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, atol=1e-5)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestLogNet(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')
+
+            c = paddle.multiply(a, b)
+            d = paddle.log(c)
+
+            fc_1 = fluid.layers.fc(input=d, 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, atol=1e-4))
+        self.assertTrue(np.allclose(npu_loss, cpu_loss, atol=1e-4))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/npu/test_sqrt_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
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSqrt(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "sqrt"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
+        out = np.sqrt(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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)
+
+    # 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 TestSqrtFp16(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "sqrt"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
+        out = np.sqrt(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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, atol=1e-5)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSqrtNet(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')
+
+            c = paddle.multiply(a, b)
+            d = paddle.sqrt(c)
+
+            fc_1 = fluid.layers.fc(input=d, 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()

python/paddle/fluid/tests/unittests/npu/test_square_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
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSquare(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "square"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
+        out = np.square(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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)
+
+    # 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 TestSquareFp16(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "square"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
+        out = np.square(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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, atol=1e-5)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSquareNet(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')
+
+            c = paddle.multiply(a, b)
+            d = paddle.square(c)
+
+            fc_1 = fluid.layers.fc(input=d, 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()

python/paddle/fluid/tests/unittests/npu/test_tanh_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
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestTanh(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "tanh"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype)
+        out = np.tanh(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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)
+
+    # 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 TestTanhFp16(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "tanh"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.uniform(1, 2, [3, 4]).astype(self.dtype)
+        out = np.tanh(x)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {}
+        self.outputs = {'Out': 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, atol=1e-3)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestTanhNet(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')
+
+            c = paddle.multiply(a, b)
+            d = paddle.tanh(c)
+
+            fc_1 = fluid.layers.fc(input=d, 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()