add swish_op for npu (#36579)

paddle/fluid/operators/activation_op_npu.cc

@@ -459,6 +459,78 @@ class SigmoidGradNPUKernel : public framework::OpKernel<T> {
   }
 };
 
+// Swish = x * sigmoid(beta * x)
+template <typename T>
+class SwishNPUKernel : 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");
+    float beta = ctx.Attr<float>("beta");
+
+    out->mutable_data<T>(ctx.GetPlace());
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    const auto& muls_runner =
+        NpuOpRunner("Muls", {*x}, {*out}, {{"value", beta}});
+    muls_runner.Run(stream);
+
+    const auto& sigmoid_runner = NpuOpRunner("Sigmoid", {*out}, {*out}, {});
+    sigmoid_runner.Run(stream);
+
+    const auto& mul_runner = NpuOpRunner("Mul", {*x, *out}, {*out});
+    mul_runner.Run(stream);
+  }
+};
+
+template <typename T>
+class SwishGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    float beta = ctx.Attr<float>("beta");
+
+    dx->mutable_data<T>(ctx.GetPlace());
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    Tensor beta_x, sigmoid_out, swish_out;
+    beta_x.mutable_data<T>(x->dims(), ctx.GetPlace());
+    sigmoid_out.mutable_data<T>(x->dims(), ctx.GetPlace());
+    swish_out.mutable_data<T>(x->dims(), ctx.GetPlace());
+
+    const auto& muls_runner =
+        NpuOpRunner("Muls", {*x}, {beta_x}, {{"value", beta}});
+    muls_runner.Run(stream);
+
+    const auto& sigmoid_runner =
+        NpuOpRunner("Sigmoid", {beta_x}, {sigmoid_out}, {});
+    sigmoid_runner.Run(stream);
+
+    const auto& mul_runner =
+        NpuOpRunner("Mul", {sigmoid_out, *x}, {swish_out}, {});
+    mul_runner.Run(stream);
+
+    const auto& mul_runner1 =
+        NpuOpRunner("Mul", {sigmoid_out, swish_out}, {*dx}, {});
+    mul_runner1.Run(stream);
+
+    const auto& sub_runner = NpuOpRunner("Sub", {swish_out, *dx}, {*dx}, {});
+    sub_runner.Run(stream);
+
+    const auto& add_runner = NpuOpRunner("Add", {sigmoid_out, *dx}, {*dx}, {});
+    add_runner.Run(stream);
+
+    const auto& mul_runner2 = NpuOpRunner("Mul", {*dout, *dx}, {*dx}, {});
+    mul_runner2.Run(stream);
+  }
+};
+
 // HardSwish = min(max(0, x+offset), threshold) * x / scale
 template <typename T>
 class HardSwishNPUKernel : public framework::OpKernel<T> {
@@ -936,6 +1008,12 @@ REGISTER_OP_NPU_KERNEL(
     ops::SigmoidGradNPUKernel<paddle::platform::NPUDeviceContext,
                               paddle::platform::float16>);
 
+REGISTER_OP_NPU_KERNEL(swish, ops::SwishNPUKernel<float>,
+                       ops::SwishNPUKernel<paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(swish_grad, ops::SwishGradNPUKernel<float>,
+                       ops::SwishGradNPUKernel<paddle::platform::float16>);
+
 REGISTER_OP_NPU_KERNEL(hard_swish, ops::HardSwishNPUKernel<float>,
                        ops::HardSwishNPUKernel<paddle::platform::float16>);
 

python/paddle/fluid/tests/unittests/npu/test_swish_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 paddle.fluid.tests.unittests.op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+from test_activation_op import ref_swish, expit
+
+paddle.enable_static()
+SEED = 1024
+
+
+class TestSwishOp(OpTest):
+    def setUp(self):
+        self.op_type = "swish"
+        self.set_npu()
+        self.init_dtype()
+        np.random.seed(2048)
+        x = np.random.uniform(-1, 1, [10, 12]).astype(self.dtype)
+        out = ref_swish(x)
+        self.inputs = {'X': x}
+        self.attrs = {'beta': 1.0}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad(self):
+        beta = self.attrs['beta']
+        out = self.outputs['Out']
+        x = self.inputs['X']
+        dx = beta * out + expit(x) * (1 - beta * out)
+        dx = dx / x.size
+
+        self.check_grad_with_place(
+            self.place, ['X'],
+            'Out',
+            max_relative_error=0.01,
+            user_defined_grads=[dx])
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.place = paddle.NPUPlace(0)
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+
+class TestSwishOpFp16(TestSwishOp):
+    def test_check_output(self):
+        self.check_output_with_place(self.place, atol=1e-3)
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+
+if __name__ == '__main__':
+    unittest.main()