scatter_op.cu 2.3 KB
Newer Older
Z
zchen0211 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

   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 "gather.cu.h"
#include "paddle/operators/gather_op.h"
#include "scatter.cu.h"

namespace paddle {
namespace operators {

template <typename T>
Z
zchen0211 已提交
23
class ScatterOpCUDAKernel : public framework::OpKernel<T> {
Z
zchen0211 已提交
24 25 26 27 28 29 30 31 32 33 34
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                   "This kernel only runs on GPU device.");
    auto *Ref = ctx.Input<Tensor>("Ref");
    auto *Index = ctx.Input<Tensor>("Index");
    auto *Updates = ctx.Input<Tensor>("Updates");
    auto *Out = ctx.Output<Tensor>("Out");

    Out->ShareDataWith<T>(*Ref);

Z
1 api  
zchen0211 已提交
35
    GPUScatterAssign<T>(ctx.GetPlace(), Updates, Index, Out);
Z
zchen0211 已提交
36 37 38 39
  }
};

template <typename T>
Z
zchen0211 已提交
40
class ScatterGradOpCUDAKernel : public framework::OpKernel<T> {
Z
zchen0211 已提交
41 42 43 44 45 46 47 48 49 50 51 52 53
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                   "This kernel only runs on GPU device.");
    auto *dRef = ctx.Output<Tensor>(framework::GradVarName("Ref"));
    auto *dUpdates = ctx.Output<Tensor>(framework::GradVarName("Updates"));
    auto *Index = ctx.Input<Tensor>("Index");
    auto *dOut = ctx.Input<Tensor>(framework::GradVarName("Out"));

    // In place gradient: dRef = dO
    dRef->ShareDataWith<T>(*dOut);
    dUpdates->mutable_data<T>(ctx.GetPlace());
    // Gradient by Gather: dUpdates = dO[Index]
Z
1 api  
zchen0211 已提交
54
    GPUGather<T>(ctx.GetPlace(), dOut, Index, dUpdates);
Z
zchen0211 已提交
55 56 57 58 59 60 61 62 63
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(scatter, ops::ScatterOpCUDAKernel<float>);
REGISTER_OP_GPU_KERNEL(scatter_grad, ops::ScatterGradOpCUDAKernel<float>);