multiplex_op.cu 3.5 KB
Newer Older
Y
Yibing Liu 已提交
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

L
Luo Tao 已提交
3 4 5
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
Y
Yibing Liu 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
Y
Yibing Liu 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
Y
Yibing Liu 已提交
14 15

#include "paddle/framework/op_registry.h"
16
#include "paddle/operators/multiplex_op.h"
Y
Yibing Liu 已提交
17

Y
Yibing Liu 已提交
18 19 20
namespace paddle {
namespace operators {

Y
Yibing Liu 已提交
21 22
using Tensor = framework::Tensor;

Y
Yibing Liu 已提交
23
template <typename Place, typename T>
Y
Yu Yang 已提交
24
class MultiplexGPUKernel : public framework::OpKernel<T> {
Y
Yibing Liu 已提交
25 26
 public:
  void Compute(const framework::ExecutionContext& ctx) const {
Y
Yibing Liu 已提交
27
    auto ins = ctx.MultiInput<Tensor>("X");
28
    auto* ids = ctx.Input<Tensor>("Ids");
Y
Yibing Liu 已提交
29
    auto* out = ctx.Output<Tensor>("Out");
Y
Yibing Liu 已提交
30 31
    out->mutable_data<T>(ctx.GetPlace());

32
    auto rows = ins[0]->dims()[0];
33
    auto cols = ins[0]->numel() / rows;
Y
Yibing Liu 已提交
34
    // copy index to cpu
Y
Yibing Liu 已提交
35
    Tensor index_t_cpu;
D
dzhwinter 已提交
36
    CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
37
    auto* index = index_t_cpu.data<int32_t>();
T
typhoonzero 已提交
38
    auto stream = ctx.cuda_device_context().stream();
D
dzhwinter 已提交
39
    platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
Y
Yibing Liu 已提交
40
    for (auto i = 0; i < rows; i++) {
41 42
      int32_t k = index[i];
      PADDLE_ENFORCE_GE(k, 0, "index must be nonnegative.");
43
      PADDLE_ENFORCE_LT((size_t)k, ins.size(),
Y
Yibing Liu 已提交
44 45 46 47 48 49 50 51
                        "index exceeds the number of candidate tensors.");
      memory::Copy(place, out->data<T>() + i * cols, place,
                   ins[k]->data<T>() + i * cols, cols * sizeof(T), stream);
    }
  }
};

template <typename Place, typename T>
Y
Yu Yang 已提交
52
class MultiplexGradGPUKernel : public framework::OpKernel<T> {
Y
Yibing Liu 已提交
53 54
 public:
  void Compute(const framework::ExecutionContext& ctx) const {
Y
Yibing Liu 已提交
55 56
    auto* d_out = ctx.Input<Tensor>(framework::GradVarName("Out"));
    auto ins = ctx.MultiInput<Tensor>("X");
57
    auto* ids = ctx.Input<Tensor>("Ids");
Y
Yibing Liu 已提交
58
    auto d_ins = ctx.MultiOutput<Tensor>(framework::GradVarName("X"));
59
    for (size_t i = 0; i < d_ins.size(); i++) {
Y
Yibing Liu 已提交
60 61 62
      if (d_ins[i]) {
        d_ins[i]->mutable_data<T>(ctx.GetPlace());
        auto t = framework::EigenVector<T>::Flatten(*d_ins[i]);
Q
QI JUN 已提交
63 64
        t.device(*ctx.template device_context<Place>().eigen_device()) =
            t.constant(static_cast<T>(0));
Y
Yibing Liu 已提交
65 66 67
      }
    }

68
    auto rows = ins[0]->dims()[0];
69
    auto cols = ins[0]->numel() / rows;
Y
Yibing Liu 已提交
70
    // copy index to cpu
Y
Yibing Liu 已提交
71
    Tensor index_t_cpu;
D
dzhwinter 已提交
72
    CopyFrom(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
73
    auto* index = index_t_cpu.data<int32_t>();
Y
Yibing Liu 已提交
74

T
typhoonzero 已提交
75
    auto stream = ctx.cuda_device_context().stream();
D
dzhwinter 已提交
76
    platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
Y
Yibing Liu 已提交
77
    for (auto i = 0; i < rows; i++) {
78
      size_t k = static_cast<size_t>(index[i]);
Y
Yibing Liu 已提交
79 80 81 82 83 84 85 86 87 88
      if (d_ins[k]) {
        memory::Copy(place, d_ins[k]->data<T>() + i * cols, place,
                     d_out->data<T>() + i * cols, cols * sizeof(T), stream);
      }
    }
  }
};
}  // namespace operators
}  // namespace paddle

Y
Yibing Liu 已提交
89 90
namespace ops = paddle::operators;

Q
QI JUN 已提交
91 92 93 94
REGISTER_OP_CUDA_KERNEL(
    multiplex,
    ops::MultiplexGPUKernel<paddle::platform::CUDADeviceContext, float>);
REGISTER_OP_CUDA_KERNEL(
95
    multiplex_grad,
Q
QI JUN 已提交
96
    ops::MultiplexGradGPUKernel<paddle::platform::CUDADeviceContext, float>);