lookup_table_op.cu 6.3 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
L
Luo Tao 已提交
2 3 4 5 6 7 8 9 10 11 12 13

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. */
14

Y
Yi Wang 已提交
15 16 17 18 19
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/lookup_table_op.h"
#include "paddle/fluid/platform/assert.h"
#include "paddle/fluid/platform/cuda_helper.h"
20 21 22 23

namespace paddle {
namespace operators {

24 25
template <typename T, int BlockDimX, int BlockDimY, int GridDimX,
          bool PaddingFlag>
26
__global__ void LookupTable(T* output, const T* table, const int64_t* ids,
27 28
                            const int64_t N, const int64_t K, const int64_t D,
                            const int64_t padding_idx) {
29
  int idx = threadIdx.x;
30
  int idy = blockIdx.x + threadIdx.y * GridDimX;
31 32

  while (idy < K) {
33
    int64_t id = ids[idy];
34 35
    PADDLE_ASSERT(id >= 0);
    PADDLE_ASSERT(id < N);
D
dangqingqing 已提交
36 37
    T* out = output + idy * D;
    const T* tab = table + id * D;
38
    for (int i = idx; i < D; i += BlockDimX) {
39
      if (PaddingFlag) {
40
        if (id == padding_idx)
41 42 43 44 45 46
          out[i] = static_cast<T>(0);
        else
          out[i] = tab[i];
      } else {
        out[i] = tab[i];
      }
47
    }
48
    idy += BlockDimY * GridDimX;
49 50 51
  }
}

52
template <typename T, int BlockDimX, int BlockDimY, int GridDimX>
53 54 55
__global__ void LookupTableGrad(T* table, const T* output, const int64_t* ids,
                                const int64_t N, const int64_t K,
                                const int64_t D) {
56
  int idx = threadIdx.x;
57
  int idy = blockIdx.x + threadIdx.y * GridDimX;
58 59 60 61 62

  while (idy < K) {
    int id = ids[idy];
    PADDLE_ASSERT(id >= 0);
    PADDLE_ASSERT(id < N);
D
dangqingqing 已提交
63 64
    const T* out = output + idy * D;
    T* tab = table + id * D;
65
    for (int i = idx; i < D; i += BlockDimX) {
D
dangqingqing 已提交
66
      paddle::platform::CudaAtomicAdd(&tab[i], out[i]);
67
    }
68
    idy += BlockDimY * GridDimX;
69 70 71 72
  }
}

template <typename T>
Y
Yu Yang 已提交
73
class LookupTableCUDAKernel : public framework::OpKernel<T> {
74 75
 public:
  void Compute(const framework::ExecutionContext& context) const override {
F
fengjiayi 已提交
76 77 78
    auto* table_t = context.Input<LoDTensor>("W");
    auto* ids_t = context.Input<LoDTensor>("Ids");
    auto* output_t = context.Output<LoDTensor>("Out");
79
    int64_t padding_idx = context.Attr<int64_t>("padding_idx");
80 81 82

    size_t N = table_t->dims()[0];
    size_t D = table_t->dims()[1];
83
    size_t K = ids_t->numel();
F
fengjiayi 已提交
84 85 86
    auto* ids = ids_t->data<int64_t>();
    auto* table = table_t->data<T>();
    auto* output = output_t->mutable_data<T>(context.GetPlace());
87 88 89

    dim3 threads(128, 8);
    dim3 grids(8, 1);
90 91 92 93 94 95 96 97 98 99 100

    if (padding_idx == -1)
      LookupTable<
          T, 128, 8, 8,
          false><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
          output, table, ids, N, K, D, padding_idx);
    else
      LookupTable<
          T, 128, 8, 8,
          true><<<grids, threads, 0, context.cuda_device_context().stream()>>>(
          output, table, ids, N, K, D, padding_idx);
101 102 103 104
  }
};

template <typename T>
Y
Yu Yang 已提交
105
class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
106 107
 public:
  void Compute(const framework::ExecutionContext& context) const override {
Q
QI JUN 已提交
108 109
    auto& dev_ctx =
        context.template device_context<platform::CUDADeviceContext>();
110
    bool is_sparse = context.Attr<bool>("is_sparse");
111 112
    // Since paddings are not trainable and fixed in forward, the gradient of
    // paddings makes no sense and we don't deal with it in backward.
113
    if (is_sparse) {
F
fengjiayi 已提交
114 115 116
      auto* ids = context.Input<LoDTensor>("Ids");
      auto* table = context.Input<LoDTensor>("W");
      auto* d_output = context.Input<LoDTensor>(framework::GradVarName("Out"));
117 118 119 120 121
      auto* d_table = context.Output<SelectedRows>(framework::GradVarName("W"));

      auto* ids_data = ids->data<int64_t>();
      auto ids_dim = ids->dims();

Q
QI JUN 已提交
122
      auto stream = dev_ctx.stream();
123 124 125
      // copy GPU memory to CPU pinned memory
      framework::Vector<int64_t> new_rows;
      new_rows.resize(ids_dim[0]);
D
dzhwinter 已提交
126
      auto gpu_place = boost::get<platform::CUDAPlace>(context.GetPlace());
127

Y
Yu Yang 已提交
128 129 130
      // TODO(yuyang18): Strange code here.
      memory::Copy(platform::CPUPlace(),
                   new_rows.CUDAMutableData(context.GetPlace()), gpu_place,
D
dzhwinter 已提交
131
                   ids_data, ids_dim[0] * sizeof(int64_t), stream);
132 133 134 135 136 137 138 139 140 141 142

      d_table->set_rows(new_rows);

      auto* d_table_value = d_table->mutable_value();
      d_table_value->Resize({ids_dim[0], table->dims()[1]});
      d_table_value->mutable_data<T>(context.GetPlace());

      auto* d_table_data = d_table_value->data<T>();
      auto* d_output_data = d_output->data<T>();
      PADDLE_ENFORCE_EQ(d_table_value->dims(), d_output->dims());
      memory::Copy(gpu_place, d_table_data, gpu_place, d_output_data,
143
                   d_output->numel() * sizeof(T), stream);
144 145

    } else {
F
fengjiayi 已提交
146 147 148
      auto ids_t = context.Input<LoDTensor>("Ids");
      auto d_output_t = context.Input<LoDTensor>(framework::GradVarName("Out"));
      auto d_table_t = context.Output<LoDTensor>(framework::GradVarName("W"));
149 150 151 152 153 154 155 156 157

      int N = d_table_t->dims()[0];
      int D = d_table_t->dims()[1];
      int K = ids_t->numel();
      const int64_t* ids = ids_t->data<int64_t>();
      const T* d_output = d_output_t->data<T>();
      T* d_table = d_table_t->mutable_data<T>(context.GetPlace());

      auto t = framework::EigenVector<T>::Flatten(*d_table_t);
Q
QI JUN 已提交
158
      t.device(*dev_ctx.eigen_device()) = t.constant(static_cast<T>(0));
159 160 161

      dim3 threads(128, 8);
      dim3 grids(8, 1);
Q
QI JUN 已提交
162
      LookupTableGrad<T, 128, 8, 8><<<grids, threads, 0, dev_ctx.stream()>>>(
T
typhoonzero 已提交
163
          d_table, d_output, ids, N, K, D);
164
    }
165 166 167 168 169 170 171
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
Q
QI JUN 已提交
172 173 174 175 176
REGISTER_OP_CUDA_KERNEL(lookup_table, ops::LookupTableCUDAKernel<float>,
                        ops::LookupTableCUDAKernel<double>);
REGISTER_OP_CUDA_KERNEL(lookup_table_grad,
                        ops::LookupTableGradCUDAKernel<float>,
                        ops::LookupTableGradCUDAKernel<double>);