From e15ccafb8410bb2329c32490e308269fb2957bba Mon Sep 17 00:00:00 2001 From: oyjxer <1728722986@qq.com> Date: Fri, 12 Mar 2021 12:00:00 +0800 Subject: [PATCH] [NPU] Support npu op elementwise_mul and elementwise_mul_grad (#31571) --- .../elementwise/elementwise_mul_op_npu.cc | 90 ++++++++++ .../npu/test_elementwise_mul_op_npu.py | 170 ++++++++++++++++++ 2 files changed, 260 insertions(+) create mode 100644 paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc create mode 100644 python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py diff --git a/paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc b/paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc new file mode 100644 index 00000000000..cecb62d3580 --- /dev/null +++ b/paddle/fluid/operators/elementwise/elementwise_mul_op_npu.cc @@ -0,0 +1,90 @@ +/* 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 +#include + +#include "paddle/fluid/operators/elementwise/elementwise_mul_op.h" +#include "paddle/fluid/operators/npu_op_runner.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class ElementwiseMulNPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + + auto* out = ctx.Output("Out"); + + auto place = ctx.GetPlace(); + + out->mutable_data(place); + + auto stream = + ctx.template device_context() + .stream(); + + auto runner = NpuOpRunner("Mul", {*x, *y}, {*out}, {}); + runner.Run(stream); + } +}; + +template +class ElementwiseMulGradNPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* dout = ctx.Input(framework::GradVarName("Out")); + + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dy = ctx.Output(framework::GradVarName("Y")); + + auto place = ctx.GetPlace(); + + dx->mutable_data(place); + dy->mutable_data(place); + + auto stream = + ctx.template device_context() + .stream(); + + auto dx_runner = NpuOpRunner("Mul", {*dout, *y}, {*dx}, {}); + dx_runner.Run(stream); + auto dy_runner = NpuOpRunner("Mul", {*x, *dout}, {*dy}, {}); + dy_runner.Run(stream); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_NPU_KERNEL( + elementwise_mul, + ops::ElementwiseMulNPUKernel, + ops::ElementwiseMulNPUKernel); + +REGISTER_OP_NPU_KERNEL( + elementwise_mul_grad, + ops::ElementwiseMulGradNPUKernel, + ops::ElementwiseMulGradNPUKernel); diff --git a/python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py new file mode 100644 index 00000000000..d0baa5877f2 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/npu/test_elementwise_mul_op_npu.py @@ -0,0 +1,170 @@ +# 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 TestElementwiseMul(OpTest): + def setUp(self): + self.set_npu() + self.op_type = "elementwise_mul" + self.place = paddle.NPUPlace(0) + + self.init_dtype() + np.random.seed(SEED) + x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype) + y = np.random.uniform(1, 2, [11, 17]).astype(self.dtype) + out = np.multiply(x, y) + + self.inputs = { + 'X': OpTest.np_dtype_to_fluid_dtype(x), + 'Y': OpTest.np_dtype_to_fluid_dtype(y) + } + 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): Mul 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 TestElementwiseMulFp16(OpTest): + def setUp(self): + self.set_npu() + self.op_type = "elementwise_mul" + self.place = paddle.NPUPlace(0) + + self.init_dtype() + np.random.seed(SEED) + x = np.random.uniform(1, 2, [3, 4]).astype(self.dtype) + y = np.random.uniform(1, 2, [3, 4]).astype(self.dtype) + out = np.multiply(x, y) + + self.inputs = { + 'X': OpTest.np_dtype_to_fluid_dtype(x), + 'Y': OpTest.np_dtype_to_fluid_dtype(y) + } + 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 TestElementwiseMulNet(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') + c_np = np.random.random(size=(32, 32)).astype('float32') + d_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') + c = paddle.static.data(name="c", shape=[32, 32], dtype='float32') + d = paddle.static.data(name="d", shape=[32, 32], dtype='float32') + label = paddle.static.data( + name="label", shape=[32, 1], dtype='int64') + + e = paddle.multiply(a, b) + f = paddle.multiply(c, d) + g = paddle.multiply(e, f) + + fc_1 = fluid.layers.fc(input=g, 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, + "c": c_np, + "d": d_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() -- GitLab