tensor_impl.h 6.3 KB
Newer Older
L
liaogang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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. */

#pragma once
#include "paddle/memory/memcpy.h"
Y
Yan Chunwei 已提交
17
#include "paddle/platform/enforce.h"
L
liaogang 已提交
18 19 20 21 22 23

namespace paddle {
namespace framework {

template <typename T>
inline void Tensor::check_memory_size() const {
Y
Yan Chunwei 已提交
24
  PADDLE_ENFORCE_NOT_NULL(
Z
Zhuoyuan 已提交
25
      holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
S
Superjom 已提交
26
  PADDLE_ENFORCE_GE(
Q
qingqing01 已提交
27
      holder_->size(), numel() * sizeof(T) + offset_,
S
Superjom 已提交
28 29 30
      "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
      "first to re-allocate memory.\n"
      "or maybe the required data-type mismatches the data already stored.");
L
liaogang 已提交
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48
}

template <typename T>
inline const T* Tensor::data() const {
  check_memory_size<T>();
  return reinterpret_cast<const T*>(
      reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
}

template <typename T>
inline T* Tensor::data() {
  check_memory_size<T>();
  return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                              offset_);
}

template <typename T>
inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
49
  static_assert(std::is_pod<T>::value, "T must be POD");
L
liaogang 已提交
50 51 52 53 54 55
  Resize(dims);
  return mutable_data<T>(place);
}

template <typename T>
inline T* Tensor::mutable_data(platform::Place place) {
56
  static_assert(std::is_pod<T>::value, "T must be POD");
Q
qingqing01 已提交
57
  PADDLE_ENFORCE_GT(numel(), 0,
Y
Yan Chunwei 已提交
58 59
                    "Tensor's numel must be larger than zero to call "
                    "Tensor::mutable_data. Call Tensor::set_dim first.");
L
liaogang 已提交
60
  /* some versions of boost::variant don't have operator!= */
Q
qingqing01 已提交
61
  int64_t size = numel() * sizeof(T);
L
liaogang 已提交
62 63 64 65 66
  if (holder_ == nullptr || !(holder_->place() == place) ||
      holder_->size() < size + offset_) {
    if (platform::is_cpu_place(place)) {
      holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
          boost::get<platform::CPUPlace>(place), size));
Q
qijun 已提交
67
    } else if (platform::is_gpu_place(place)) {
68
#ifndef PADDLE_WITH_CUDA
Q
qijun 已提交
69
      PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
L
liaogang 已提交
70
    }
Q
qijun 已提交
71
#else
L
liaogang 已提交
72 73 74 75 76 77 78 79 80 81 82
      holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
          boost::get<platform::GPUPlace>(place), size));
    }
#endif
    offset_ = 0;
  }
  return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                              offset_);
}

template <typename T>
S
Superjom 已提交
83
inline Tensor& Tensor::ShareDataWith(const Tensor& src) {
L
liaogang 已提交
84 85
  src.check_memory_size<T>();
  *this = src;
S
Superjom 已提交
86
  return *this;
L
liaogang 已提交
87 88 89 90
}

template <typename T>
inline void Tensor::CopyFrom(const Tensor& src,
Q
qijun 已提交
91
                             const platform::Place& dst_place) {
L
liaogang 已提交
92 93 94 95 96 97 98 99
  src.check_memory_size<T>();
  Resize(src.dims());

  auto src_place = src.holder_->place();
  auto src_ptr = static_cast<const void*>(src.data<T>());

  auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));

100
  auto size = src.numel() * sizeof(T);
L
liaogang 已提交
101

Q
qijun 已提交
102
  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
L
liaogang 已提交
103 104 105
    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                 boost::get<platform::CPUPlace>(src_place), src_ptr, size);
  }
106
#ifdef PADDLE_WITH_CUDA
Q
qijun 已提交
107 108
  else if (platform::is_gpu_place(src_place) &&
           platform::is_cpu_place(dst_place)) {
L
liaogang 已提交
109 110
    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                 boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
Q
qijun 已提交
111 112
  } else if (platform::is_cpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
L
liaogang 已提交
113
    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
Q
qijun 已提交
114
                 boost::get<platform::CPUPlace>(src_place), src_ptr, size, 0);
Q
qijun 已提交
115 116
  } else if (platform::is_gpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
L
liaogang 已提交
117
    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
Q
qijun 已提交
118
                 boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
L
liaogang 已提交
119
  }
Q
qijun 已提交
120 121
  PADDLE_ENFORCE(cudaStreamSynchronize(0),
                 "cudaStreamSynchronize failed in Tensor CopyFrom");
Q
qijun 已提交
122

L
liaogang 已提交
123
#endif
Q
qijun 已提交
124
}
L
liaogang 已提交
125

126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148
template <typename T>
inline void Tensor::CopyFromVector(const std::vector<T>& src,
                                   const platform::Place& dst_place) {
  auto src_ptr = static_cast<const void*>(src.data());
  platform::CPUPlace src_place;
  auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
  auto size = src.size() * sizeof(T);

  if (platform::is_cpu_place(dst_place)) {
    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr, src_place,
                 src_ptr, size);
  }
#ifdef PADDLE_WITH_CUDA
  else if (platform::is_gpu_place(dst_place)) {
    memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place,
                 src_ptr, size, 0);
  }
  PADDLE_ENFORCE(cudaStreamSynchronize(0),
                 "cudaStreamSynchronize failed in Tensor CopyFromVector");

#endif
}

L
liaogang 已提交
149 150 151
template <typename T>
inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
  check_memory_size<T>();
Y
Yan Chunwei 已提交
152 153 154 155
  PADDLE_ENFORCE_GE(begin_idx, 0, "Slice begin index is less than zero.");
  PADDLE_ENFORCE_LE(end_idx, dims_[0], "Slice end index is out of bound.");
  PADDLE_ENFORCE_LT(begin_idx, end_idx,
                    "Begin index must be less than end index.");
H
hedaoyuan 已提交
156 157 158 159 160 161 162 163 164 165 166 167 168

  if (dims_[0] == 1) {
    return *this;
  } else {
    size_t base = numel() / dims_[0];
    Tensor dst;
    dst.holder_ = holder_;
    DDim dst_dims = dims_;
    dst_dims[0] = end_idx - begin_idx;
    dst.Resize(dst_dims);
    dst.offset_ = offset_ + begin_idx * base * sizeof(T);
    return dst;
  }
L
liaogang 已提交
169 170
}

S
Superjom 已提交
171 172 173 174
inline Tensor& Tensor::Resize(const DDim& dims) {
  dims_ = dims;
  return *this;
}
L
liaogang 已提交
175 176 177

inline const DDim& Tensor::dims() const { return dims_; }

Y
Yu Yang 已提交
178
inline int64_t Tensor::numel() const { return product(dims_); }
179

F
WIP  
fengjiayi 已提交
180
template <typename T>
F
fengjiayi 已提交
181
inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
F
WIP  
fengjiayi 已提交
182 183
  Tensor res;
  res.ShareDataWith<T>(src);
F
fengjiayi 已提交
184
  res.Resize(flatten_to_2d(src.dims(), num_col_dims));
F
WIP  
fengjiayi 已提交
185 186 187
  return res;
}

L
liaogang 已提交
188 189
}  // namespace framework
}  // namespace paddle