sequence_padding.cu 6.9 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Y
Yiqun Liu 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

15
#include <algorithm>
Y
Yi Wang 已提交
16
#include "paddle/fluid/operators/math/sequence_padding.h"
Y
Yiqun Liu 已提交
17 18 19 20 21

namespace paddle {
namespace operators {
namespace math {

F
fengjiayi 已提交
22
template <typename T, CopyType Type>
23
__global__ void SequencePaddingKernel(
F
fengjiayi 已提交
24
    T* dst, const T* src, const T* pad_value, bool is_constant_pad,
F
bug fix  
fengjiayi 已提交
25 26
    const size_t* seq_offsets, const size_t seq_num, const size_t pad_seq_len,
    const size_t step_width, bool norm_by_len, const PadLayout layout) {
Y
yangyaming 已提交
27
  size_t seq_idx = blockIdx.y;
F
fengjiayi 已提交
28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
  size_t seq_len = seq_offsets[seq_idx + 1] - seq_offsets[seq_idx];

  size_t step_idx = blockIdx.x * blockDim.y + threadIdx.y;
  size_t seq_data_offset = (seq_offsets[seq_idx] + step_idx) * step_width;
  size_t pad_data_offset = layout == kBatchLengthWidth
                               ? (seq_idx * pad_seq_len + step_idx) * step_width
                               : (step_idx * seq_num + seq_idx) * step_width;

  T* dst_data = dst + (Type == kSeqToPad ? pad_data_offset : seq_data_offset);
  const T* src_data =
      src + (Type == kSeqToPad ? seq_data_offset : pad_data_offset);

  if (step_idx < seq_len) {
    float scale = norm_by_len ? (1.0f / static_cast<float>(seq_len)) : 1.0f;
    for (size_t i = threadIdx.x; i < step_width; i += blockDim.x) {
      dst_data[i] = scale * src_data[i];
Y
Yiqun Liu 已提交
44
    }
F
fengjiayi 已提交
45
  } else if (step_idx < pad_seq_len && Type == kSeqToPad) {
F
bug fix  
fengjiayi 已提交
46
    for (size_t i = threadIdx.x; i < step_width; i += blockDim.x) {
F
fengjiayi 已提交
47
      dst_data[i] = is_constant_pad ? pad_value[0] : pad_value[i];
Y
Yiqun Liu 已提交
48 49 50 51
    }
  }
}

Y
yangyaming 已提交
52 53
template <typename T>
class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
Y
Yiqun Liu 已提交
54 55
 public:
  void operator()(const platform::CUDADeviceContext& context,
56
                  const framework::LoDTensor& seq_tensor,
F
bug fix  
fengjiayi 已提交
57
                  framework::LoDTensor* pad_tensor,
F
fengjiayi 已提交
58 59 60 61 62 63 64
                  const framework::LoDTensor& pad_value, int pad_seq_len = -1,
                  int lod_level = 0, bool norm_by_times = false,
                  const PadLayout layout = kBatchLengthWidth) {
    auto seq_lod = seq_tensor.lod();
    const auto seq_offsets = framework::ToAbsOffset(seq_lod)[lod_level];
    const auto& seq_tensor_dims = seq_tensor.dims();
    const auto& pad_tensor_dims = pad_tensor->dims();
F
bug fix  
fengjiayi 已提交
65
    int max_seq_len = MaximumSequenceLength(seq_offsets);
F
fengjiayi 已提交
66
    if (pad_seq_len == -1) {
F
bug fix  
fengjiayi 已提交
67
      pad_seq_len = max_seq_len;
F
fengjiayi 已提交
68
    }
F
fengjiayi 已提交
69 70 71
    PADDLE_ENFORCE_GE(pad_seq_len, max_seq_len,
                      "The pad_seq_len must be equal to or greater than the "
                      "original max sequence length.");
F
fengjiayi 已提交
72
    int step_width = seq_tensor.numel() / seq_tensor_dims[0];
F
bug fix  
fengjiayi 已提交
73
    int seq_num = seq_offsets.size() - 1;
74

F
fengjiayi 已提交
75 76 77 78 79
    CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len,
              step_width, layout);
    PADDLE_ENFORCE(pad_value.numel() == 1 || pad_value.numel() == step_width,
                   "The numel of 'pad_value' can only be 1 or be equal to the "
                   "'step_width'.");
80

F
bug fix  
fengjiayi 已提交
81
    const int kBlockSize = 512;
Y
Yiqun Liu 已提交
82 83 84 85 86

    /* At least use 32 threads to copy sequence_width elements,
     * and at least 8 elements for each thread.
     */
    size_t block_dim_x =
F
fengjiayi 已提交
87
        std::min(((((step_width + 7) >> 3) + 31) >> 5) << 5, kBlockSize);
Y
Yiqun Liu 已提交
88 89 90
    size_t block_dim_y = kBlockSize / block_dim_x;
    dim3 threads(block_dim_x, block_dim_y);

F
fengjiayi 已提交
91
    size_t grid_dim_x = (pad_seq_len + block_dim_y - 1) / block_dim_y;
92
    size_t grid_dim_y = seq_num;
Y
Yiqun Liu 已提交
93 94
    dim3 grid(grid_dim_x, grid_dim_y);

95
    const T* seq_data = seq_tensor.data<T>();
Y
yangyaming 已提交
96
    T* pad_data = pad_tensor->data<T>();
F
fengjiayi 已提交
97
    const T* pad_value_data = pad_value.data<T>();
98

F
fengjiayi 已提交
99 100
    SequencePaddingKernel<T, kSeqToPad><<<grid, threads, 0, context.stream()>>>(
        pad_data, seq_data, pad_value_data, pad_value.numel() == 1,
F
bug fix  
fengjiayi 已提交
101
        seq_offsets.CUDAData(context.GetPlace()), seq_num, pad_seq_len,
F
fengjiayi 已提交
102
        step_width, norm_by_times, layout);
Y
Yiqun Liu 已提交
103 104 105
  }
};

Y
yangyaming 已提交
106 107
template <typename T>
class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
Y
Yiqun Liu 已提交
108 109
 public:
  void operator()(const platform::CUDADeviceContext& context,
F
fengjiayi 已提交
110 111 112 113 114 115 116
                  const framework::LoDTensor& pad_tensor,
                  framework::LoDTensor* seq_tensor, int pad_seq_len = -1,
                  int lod_level = 0, bool norm_by_times = false,
                  const PadLayout layout = kBatchLengthWidth) {
    auto seq_offsets = framework::ToAbsOffset(seq_tensor->lod())[lod_level];
    const auto& seq_tensor_dims = seq_tensor->dims();
    const auto& pad_tensor_dims = pad_tensor.dims();
F
bug fix  
fengjiayi 已提交
117
    int max_seq_len = MaximumSequenceLength(seq_offsets);
F
fengjiayi 已提交
118
    if (pad_seq_len == -1) {
F
bug fix  
fengjiayi 已提交
119
      pad_seq_len = max_seq_len;
F
fengjiayi 已提交
120 121
    }
    int step_width = seq_tensor->numel() / seq_tensor_dims[0];
F
bug fix  
fengjiayi 已提交
122
    int seq_num = seq_offsets.size() - 1;
Y
yangyaming 已提交
123

F
fengjiayi 已提交
124 125
    CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len,
              step_width, layout);
P
phlrain 已提交
126
    /*
F
bug fix  
fengjiayi 已提交
127
    if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) {
Y
yangyaming 已提交
128 129
      TensorCopy(pad_tensor, context.GetPlace(), context, seq_tensor);
      seq_tensor->Resize(seq_tensor_dims);
Y
Yiqun Liu 已提交
130 131
      return;
    }
P
phlrain 已提交
132
    */
Y
Yiqun Liu 已提交
133

F
bug fix  
fengjiayi 已提交
134
    const int kBlockSize = 512;
Y
Yiqun Liu 已提交
135 136 137 138 139

    /* At least use 32 threads to copy sequence_width elements,
     * and at least 8 elements for each thread.
     */
    size_t block_dim_x =
F
fengjiayi 已提交
140
        std::min(((((step_width + 7) >> 3) + 31) >> 5) << 5, kBlockSize);
Y
Yiqun Liu 已提交
141 142 143
    size_t block_dim_y = kBlockSize / block_dim_x;
    dim3 threads(block_dim_x, block_dim_y);

F
fengjiayi 已提交
144
    size_t grid_dim_x = (pad_seq_len + block_dim_y - 1) / block_dim_y;
145
    size_t grid_dim_y = seq_num;
Y
Yiqun Liu 已提交
146 147
    dim3 grid(grid_dim_x, grid_dim_y);

Y
yangyaming 已提交
148
    const T* pad_data = pad_tensor.data<T>();
149 150
    T* seq_data = seq_tensor->data<T>();

F
fengjiayi 已提交
151 152
    SequencePaddingKernel<T, kPadToSeq><<<grid, threads, 0, context.stream()>>>(
        seq_data, pad_data, nullptr, false,
F
bug fix  
fengjiayi 已提交
153
        seq_offsets.CUDAData(context.GetPlace()), seq_num, pad_seq_len,
F
fengjiayi 已提交
154
        step_width, norm_by_times, layout);
Y
Yiqun Liu 已提交
155 156 157
  }
};

Y
yangyaming 已提交
158 159 160 161 162 163 164 165 166
template class PaddingLoDTensorFunctor<platform::CUDADeviceContext, int>;
template class PaddingLoDTensorFunctor<platform::CUDADeviceContext, int64_t>;
template class PaddingLoDTensorFunctor<platform::CUDADeviceContext, float>;
template class PaddingLoDTensorFunctor<platform::CUDADeviceContext, double>;

template class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, int>;
template class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, int64_t>;
template class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, float>;
template class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, double>;
Y
Yiqun Liu 已提交
167 168 169 170

}  // namespace math
}  // namespace operators
}  // namespace paddle