From 17862b725f36eb88acd370f3a9e086ddc9fed863 Mon Sep 17 00:00:00 2001 From: OleNet Date: Fri, 19 Mar 2021 14:28:53 +0800 Subject: [PATCH] [NPU] Support mean npu kernel (#31729) --- paddle/fluid/operators/CMakeLists.txt | 2 + paddle/fluid/operators/mean_op_npu.cc | 117 ++++++++++++++++++ .../tests/unittests/npu/test_mean_op_npu.py | 89 +++++++++++++ 3 files changed, 208 insertions(+) create mode 100644 paddle/fluid/operators/mean_op_npu.cc create mode 100644 python/paddle/fluid/tests/unittests/npu/test_mean_op_npu.py diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index 6fe18f24794..618d3109856 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -187,4 +187,6 @@ endif() if(WITH_ASCEND_CL) cc_test(gelu_op_npu_test SRCS gelu_op_npu_test.cc DEPS op_registry gelu_op scope device_context enforce executor) +cc_test(mean_op_npu_test SRCS mean_op_npu_test.cc DEPS op_registry mean_op scope device_context enforce executor) endif() + diff --git a/paddle/fluid/operators/mean_op_npu.cc b/paddle/fluid/operators/mean_op_npu.cc new file mode 100644 index 00000000000..a577da80de4 --- /dev/null +++ b/paddle/fluid/operators/mean_op_npu.cc @@ -0,0 +1,117 @@ +/* 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 "paddle/fluid/operators/mean_op.h" +#include "paddle/fluid/platform/float16.h" +#include "paddle/fluid/operators/npu_op_runner.h" + + +namespace paddle { +namespace operators { + +template +class MeanNPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* out = ctx.Output("Out"); + + std::vector axes; + + framework::NPUAttributeMap attr_input = { + {"keep_dims", false}, + {"axes", axes}}; + + out->mutable_data(ctx.GetPlace()); + + auto runner = NpuOpRunner("ReduceMeanD", + {*x}, + {*out}, + attr_input); + + auto stream = + ctx.template device_context< + paddle::platform::NPUDeviceContext>() + .stream(); + runner.Run(stream); + } +}; + + +template +class MeanGradNPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto stream = + context.template device_context< + paddle::platform::NPUDeviceContext>() + .stream(); + + auto grad = context.Input(framework::GradVarName("Out")); + + PADDLE_ENFORCE_EQ(grad->numel(), 1, + platform::errors::InvalidArgument( + "Mean Gradient Input Tensor len should be 1. But " + "received Out@Grad's elements num is %d.", + grad->numel())); + + auto IG = context.Output(framework::GradVarName("X")); + IG->mutable_data(context.GetPlace()); + + // ones + Tensor ones(grad->type()); + ones.mutable_data(IG->dims(), context.GetPlace()); + auto runner_ones = NpuOpRunner("OnesLike", {*IG}, {ones}, {}); + runner_ones.Run(stream); + + // means + Tensor mean_tensor(grad->type()); + mean_tensor.Resize({1}); + mean_tensor.mutable_data(context.GetPlace()); + std::vector mean_vec; + mean_vec.push_back(1.0/static_cast(IG->numel())); + framework::TensorFromVector(mean_vec, + context.device_context(), + &mean_tensor); + + // means mul ones + Tensor mean_ma(grad->type()); + mean_ma.Resize(IG->dims()); + mean_ma.mutable_data(context.GetPlace()); + auto runner_mul_1 = NpuOpRunner("Mul", {mean_tensor, ones}, {mean_ma}, {}); + runner_mul_1.Run(stream); + + // and mul grad + auto runner_mul_2 = NpuOpRunner("Mul", {mean_ma, *grad}, {*IG}, {}); + runner_mul_2.Run(stream); + } +}; + + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +namespace plat = paddle::platform; +REGISTER_OP_NPU_KERNEL( + mean, + ops::MeanNPUKernel, + ops::MeanNPUKernel, + ops::MeanNPUKernel, + ops::MeanNPUKernel) + + +REGISTER_OP_NPU_KERNEL( + mean_grad, + ops::MeanGradNPUKernel, + ops::MeanGradNPUKernel, + ops::MeanGradNPUKernel, + ops::MeanGradNPUKernel) diff --git a/python/paddle/fluid/tests/unittests/npu/test_mean_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_mean_op_npu.py new file mode 100644 index 00000000000..60357d2c937 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/npu/test_mean_op_npu.py @@ -0,0 +1,89 @@ +# 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 +from paddle.fluid import core + +paddle.enable_static() +SEED = 2021 + + +@unittest.skipIf(not paddle.is_compiled_with_npu(), + "core is not compiled with NPU") +class TestMean(OpTest): + def setUp(self): + self.set_npu() + self.place = paddle.NPUPlace(0) + self.op_type = "mean" + self.init_dtype() + + x = np.random.random([1, 100]).astype(self.dtype) + self.inputs = {'X': x} + + self.attrs = {} + np_out = np.mean(x) + self.outputs = {'Out': np_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) + + def test_check_grad(self): + self.check_grad_with_place(self.place, ['X'], 'Out', check_dygraph=False) + + +@unittest.skipIf(not paddle.is_compiled_with_npu(), + "core is not compiled with NPU") +class TestMeanFP16(OpTest): + def setUp(self): + self.set_npu() + self.place = paddle.NPUPlace(0) + self.op_type = "mean" + self.init_dtype() + + x = np.random.random([3, 200]).astype(self.dtype) + self.inputs = {'X': x} + + self.attrs = {} + np_out = np.mean(x) + self.outputs = {'Out': np_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) + + + +if __name__ == '__main__': + unittest.main() + -- GitLab