/* 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/gather_op.h" #include #include #include #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/operators/kron_op.h" #include "paddle/fluid/operators/npu_op_runner.h" namespace paddle { namespace operators { template class GatherOpNPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { auto *x = ctx.Input("X"); auto *index = ctx.Input("Index"); auto *out = ctx.Output("Out"); out->mutable_data(ctx.GetPlace()); auto runner = NpuOpRunner("Gather", {*x, *index}, {*out}, {{"validate_indices", true}}); auto stream = ctx.template device_context() .stream(); runner.Run(stream); } }; template class GatherGradOpNPUKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &ctx) const override { auto *index = ctx.Input("Index"); auto *x = ctx.Input("X"); auto *dout = ctx.Input(framework::GradVarName("Out")); auto *dx = ctx.Output(framework::GradVarName("X")); // step1: Unsqueeze index framework::Tensor tmp_tensor(index->type()); const auto index_dims = index->dims(); if (index_dims.size() == 1) { tmp_tensor.ShareDataWith(*index); std::vector new_dim = {index_dims[0], 1}; tmp_tensor.Resize(framework::make_ddim(new_dim)); index = &tmp_tensor; } auto stream = ctx.template device_context() .stream(); // step2: ZerosLike x in device Tensor *tmp_zerox = const_cast(x); Tensor zeroslike_xout(x->type()); zeroslike_xout.Resize(x->dims()); zeroslike_xout.mutable_data(ctx.GetPlace()); auto runner_zeroslike = NpuOpRunner("ZerosLike", {*x}, {zeroslike_xout}, {}); runner_zeroslike.Run(stream); tmp_zerox = &zeroslike_xout; // step3: scatter(x_grad) dx->mutable_data(ctx.GetPlace()); auto runner_scatter = NpuOpRunner("TensorScatterUpdate", {*tmp_zerox, *index, *dout}, {*dx}, {}); runner_scatter.Run(stream); } }; } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP_NPU_KERNEL( gather, ops::GatherOpNPUKernel, ops::GatherOpNPUKernel); REGISTER_OP_NPU_KERNEL( gather_grad, ops::GatherGradOpNPUKernel, ops::GatherGradOpNPUKernel);