Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2298
Star 20931
Fork 5422
P
Paddle
切换分支/标签
查找文件 普通视图历史永久链接Permalink
tensor_impl.h 9.1 KB
Newer Older
L
ENH: Refine Tensor and And CopyFrom  
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
fix some enforce (#3301)  
Yan Chunwei 已提交
17 
#include "paddle/platform/enforce.h"
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
18 19 20 21 

namespace paddle {
namespace framework {
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
22 23 24 
template <typename... T>
struct SizeOfTypeFunctor;
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
25 
template <typename T>
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 
struct SizeOfTypeFunctor<T> {
  size_t operator()(std::type_index type) const {
    if (typeid(T).hash_code() == type.hash_code()) {
      return sizeof(T);
    } else {
      return 0UL;
    }
  }
};

template <>
struct SizeOfTypeFunctor<> {
  size_t operator()(std::type_index type) const { return 0UL; }
};

template <typename HEAD, typename... TAIL>
struct SizeOfTypeFunctor<HEAD, TAIL...> {
  size_t operator()(std::type_index type) const {
    SizeOfTypeFunctor<HEAD> head;
    size_t head_size = head(type);
    if (head_size != 0) {
      return head_size;
    }
    SizeOfTypeFunctor<TAIL...> tail;
    return tail(type);
  }
};

static inline size_t SizeOfType(std::type_index type) {
  SizeOfTypeFunctor<int, float, double, int16_t, int64_t> functor;
  size_t size = functor(type);
  PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
  return size;
}
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
61 
inline void Tensor::check_memory_size() const {
Y
fix some enforce (#3301)  
Yan Chunwei 已提交
62 
  PADDLE_ENFORCE_NOT_NULL(
Z
tensor element size support  
Zhuoyuan 已提交
63 
      holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
S
add a error message to tensor  
Superjom 已提交
64 
  PADDLE_ENFORCE_GE(
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
65 
      holder_->size(), numel() * SizeOfType(type()) + offset_,
S
add a error message to tensor  
Superjom 已提交
66 67 68 
      "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
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
69 70 71 72 
}

template <typename T>
inline const T* Tensor::data() const {
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
73 74 75 76 77 78 
  check_memory_size();
  PADDLE_ENFORCE(std::is_same<T, void>::value ||
                     holder_->type().hash_code() == typeid(T).hash_code(),
                 "Tensor holds the wrong type, it holds %s",
                 this->holder_->type().name());
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
79 80 81 82 83 84 
  return reinterpret_cast<const T*>(
      reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
}

template <typename T>
inline T* Tensor::data() {
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
85 86 87 88 89 
  check_memory_size();
  PADDLE_ENFORCE(std::is_same<T, void>::value ||
                     holder_->type().hash_code() == typeid(T).hash_code(),
                 "Tensor holds the wrong type, it holds %s",
                 this->holder_->type().name());
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
90 91 92 93 94 95 
  return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                              offset_);
}

template <typename T>
inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
L
FIX: restricting c++ template usage to POD types  
liaogang 已提交
96 
  static_assert(std::is_pod<T>::value, "T must be POD");
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
97 98 99 100 101 102 
  Resize(dims);
  return mutable_data<T>(place);
}

template <typename T>
inline T* Tensor::mutable_data(platform::Place place) {
L
FIX: restricting c++ template usage to POD types  
liaogang 已提交
103 
  static_assert(std::is_pod<T>::value, "T must be POD");
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
104 105 106 107 108 109 110 
  return reinterpret_cast<T*>(mutable_data(place, typeid(T)));
}

inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
  if (holder_ != nullptr) {
    holder_->set_type(type);
  }
C
add gpu kernel by copying inputs/outputs between cpu and gpu.  
caoying03 已提交
111 112 113 114 
  PADDLE_ENFORCE_GT(
      numel(), 0,
      "When calling this method, the Tensor's numel must be larger than zero. "
      "Please check Tensor::Resize has been called first.");
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
115 
  int64_t size = numel() * SizeOfType(type);
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
116 117 118 119 
  /* some versions of boost::variant don't have operator!= */
  if (holder_ == nullptr || !(holder_->place() == place) ||
      holder_->size() < size + offset_) {
    if (platform::is_cpu_place(place)) {
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
120 121 
      holder_.reset(new PlaceholderImpl<platform::CPUPlace>(
          boost::get<platform::CPUPlace>(place), size, type));
Q
add gpu python op test  
qijun 已提交
122 
    } else if (platform::is_gpu_place(place)) {
Y
Use PADDLE_WITH_CUDA instead of PADDLE_WITH_GPU  
Yi Wang 已提交
123 
#ifndef PADDLE_WITH_CUDA
Q
add gpu python op test  
qijun 已提交
124 
      PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
125 
    }
Q
add gpu python op test  
qijun 已提交
126 
#else
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
127 128 
      holder_.reset(new PlaceholderImpl<platform::GPUPlace>(
          boost::get<platform::GPUPlace>(place), size, type));
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
129 130 131 132 
    }
#endif
    offset_ = 0;
  }
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
133 134 135 136 137 138 139 140 
  return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                                 offset_);
}

inline void* Tensor::mutable_data(platform::Place place) {
  PADDLE_ENFORCE(this->holder_ != nullptr,
                 "Cannot invoke mutable data if current hold nothing");
  return mutable_data(place, holder_->type());
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
141 142 
}
S
tensor slight improve  
Superjom 已提交
143 
inline Tensor& Tensor::ShareDataWith(const Tensor& src) {
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
144 
  src.check_memory_size();
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
145 
  *this = src;
S
tensor slight improve  
Superjom 已提交
146 
  return *this;
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
147 148 149 
}

inline void Tensor::CopyFrom(const Tensor& src,
Q
Unify CUDA stream in Tensor CopyFrom interface (#4692)  
QI JUN 已提交
150 151 
                             const platform::Place& dst_place,
                             const platform::DeviceContext& ctx) {
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
152 
  src.check_memory_size();
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
153 154 155 
  Resize(src.dims());

  auto src_place = src.holder_->place();
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
156 
  auto src_ptr = src.data<void>();
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
157
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
158 
  auto dst_ptr = mutable_data(dst_place, src.type());
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
159
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
160 
  auto size = src.numel() * SizeOfType(src.type());
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
161
Q
refine tensor copy from  
qijun 已提交
162 
  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
163 164 165 
    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                 boost::get<platform::CPUPlace>(src_place), src_ptr, size);
  }
Y
Use PADDLE_WITH_CUDA instead of PADDLE_WITH_GPU  
Yi Wang 已提交
166 
#ifdef PADDLE_WITH_CUDA
Q
refine tensor copy from  
qijun 已提交
167 168 
  else if (platform::is_gpu_place(src_place) &&
           platform::is_cpu_place(dst_place)) {
Q
Unify CUDA stream in Tensor CopyFrom interface (#4692)  
QI JUN 已提交
169 170 171 172 173 174 175 176 177 
    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
    auto dst_cpu_place = boost::get<platform::CPUPlace>(dst_place);
    auto ctx_place = ctx.GetPlace();
    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
    memory::Copy(
        dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size,
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
Q
refine tensor copy from  
qijun 已提交
178 179 
  } else if (platform::is_cpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
Q
Unify CUDA stream in Tensor CopyFrom interface (#4692)  
QI JUN 已提交
180 181 182 183 184 185 186 187 188 
    auto src_cpu_place = boost::get<platform::CPUPlace>(src_place);
    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
    auto ctx_place = ctx.GetPlace();
    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
    PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place);
    memory::Copy(
        dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size,
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
Q
refine tensor copy from  
qijun 已提交
189 190 
  } else if (platform::is_gpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
Q
Unify CUDA stream in Tensor CopyFrom interface (#4692)  
QI JUN 已提交
191 192 193 194 195 196 197 198 199 
    auto src_gpu_place = boost::get<platform::GPUPlace>(src_place);
    auto dst_gpu_place = boost::get<platform::GPUPlace>(dst_place);
    auto ctx_place = ctx.GetPlace();
    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
    auto ctx_gpu_place = boost::get<platform::GPUPlace>(ctx_place);
    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
    memory::Copy(
        dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
200 201 
  }
#endif
Q
refine tensor copy from  
qijun 已提交
202 
}
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
203
A
Adding implementation for copying a vector to a tensor (#4635)  
Abhinav Arora 已提交
204 205 
template <typename T>
inline void Tensor::CopyFromVector(const std::vector<T>& src,
Q
Unify CUDA stream in Tensor CopyFrom interface (#4692)  
QI JUN 已提交
206 207 
                                   const platform::DeviceContext& ctx) {
  auto dst_place = ctx.GetPlace();
A
Adding implementation for copying a vector to a tensor (#4635)  
Abhinav Arora 已提交
208 209 210 211 212 213 214 215 216 217 218 
  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)) {
Q
Unify CUDA stream in Tensor CopyFrom interface (#4692)  
QI JUN 已提交
219 220 221 222 
    memory::Copy(
        boost::get<platform::GPUPlace>(dst_place), dst_ptr, src_place, src_ptr,
        size,
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
A
Adding implementation for copying a vector to a tensor (#4635)  
Abhinav Arora 已提交
223 224 225 226 
  }
#endif
}
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
227 
inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
228 
  check_memory_size();
C
add forward computation of crf operator.  
caoying03 已提交
229 230 231 
  PADDLE_ENFORCE_GE(begin_idx, 0,
                    "The start row index must be greater than 0.");
  PADDLE_ENFORCE_LE(end_idx, dims_[0], "The end row index is out of bound.");
C
follow comments.  
caoying03 已提交
232 233 234 
  PADDLE_ENFORCE_LT(
      begin_idx, end_idx,
      "The start row index must be smaller than the end row index.");
H
Fix Tensor::Slice with dims[0] == 1.  
hedaoyuan 已提交
235 236 237 238 239 240 241 242 243 244 

  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);
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
245 
    dst.offset_ = offset_ + begin_idx * base * SizeOfType(type());
H
Fix Tensor::Slice with dims[0] == 1.  
hedaoyuan 已提交
246 247 
    return dst;
  }
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
248 249 
}
S
tensor slight improve  
Superjom 已提交
250 251 252 253 
inline Tensor& Tensor::Resize(const DDim& dims) {
  dims_ = dims;
  return *this;
}
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
254 255 256 

inline const DDim& Tensor::dims() const { return dims_; }
Y
Remove `numel` field in tensor  
Yu Yang 已提交
257 
inline int64_t Tensor::numel() const { return product(dims_); }
Q
Add function to get element count from tensor.  
qingqing01 已提交
258
F
Follow comments  
fengjiayi 已提交
259 
inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
F
WIP  
fengjiayi 已提交
260 
  Tensor res;
Y
Remove template parameter for Tensor methods (#4937)  
Yu Yang 已提交
261 
  res.ShareDataWith(src);
F
Follow comments  
fengjiayi 已提交
262 
  res.Resize(flatten_to_2d(src.dims(), num_col_dims));
F
WIP  
fengjiayi 已提交
263 264 265 
  return res;
}
L
ENH: Refine Tensor and And CopyFrom  
liaogang 已提交
266 267 
}  // namespace framework
}  // namespace paddle