tensor_util.h 15.6 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2

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
D
dzhwinter 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 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. */
D
dzhwinter 已提交
14 15

#pragma once
16
#include <vector>
W
wanghuancoder 已提交
17

Y
Yi Wang 已提交
18
#include "paddle/fluid/framework/data_type.h"
6
633WHU 已提交
19
#include "paddle/fluid/framework/dlpack_tensor.h"
Y
Yi Wang 已提交
20 21
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/tensor.h"
22 23 24 25
#include "paddle/fluid/memory/allocation/allocator_facade.h"
#ifdef PADDLE_WITH_ASCEND_CL
#include "paddle/fluid/memory/allocation/npu_pinned_allocator.h"
#endif
Y
Yi Wang 已提交
26
#include "paddle/fluid/platform/device_context.h"
D
dzhwinter 已提交
27 28 29 30

namespace paddle {
namespace framework {

31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
class PrintOptions {
 public:
  static PrintOptions& Instance() {
    static PrintOptions instance;
    return instance;
  }
  ~PrintOptions() {}
  PrintOptions(const PrintOptions& o) = delete;
  const PrintOptions& operator=(const PrintOptions& o) = delete;

  int precision = 8;
  int threshold = 1000;
  int edgeitems = 3;
  int linewidth = 75;
  bool sci_mode = false;

 private:
  PrintOptions() {}
};

C
chengduo 已提交
51 52 53 54 55 56
// NOTE(zcd): Because TensorCopy is an async operation, when the src_place
// and dst_place are two different GPU, to ensure that the operation can
// be carried out correctly, there is a src_ctx wait operation in TensorCopy.
// If ctx_place and src_place are the same, src_ctx.Wait() is added
// after memory::Copy; if ctx_place and dst_place are the same,
// src_ctx.Wait() is added before memory::Copy.
W
wanghuancoder 已提交
57 58
class Tensor;

Y
Yi Wang 已提交
59
void TensorCopy(const Tensor& src, const platform::Place& dst_place,
F
fengjiayi 已提交
60
                const platform::DeviceContext& ctx, Tensor* dst);
C
chengduo 已提交
61 62 63 64 65 66 67 68

// NOTE(zcd): If the src.place() and dst_place are two different GPU,
// the copy operation is carried out on the dst_place's stream. This is
// very important, because TensorCopy is an async operator, and in most
// case, once this copy operator returns, dst is to be used in dst_place's
// stream, if this copy operation is carried out on the src_place's stream,
// when dst is used in dst_place's stream the copy operation may be
// not completed.
Y
Yi Wang 已提交
69 70
void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                Tensor* dst);
C
chengduo 已提交
71

F
fengjiayi 已提交
72 73
void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
                    Tensor* dst);
D
dzhwinter 已提交
74

Y
Yi Wang 已提交
75 76 77 78 79
template <typename T>
void TensorFromVector(const std::vector<T>& src,
                      const platform::DeviceContext& ctx, Tensor* dst);
template <typename T>
void TensorFromVector(const std::vector<T>& src, Tensor* dst);
D
dzhwinter 已提交
80

Y
Yi Wang 已提交
81 82 83 84 85
template <typename T>
void TensorToVector(const Tensor& src, const platform::DeviceContext& ctx,
                    std::vector<T>* dst);
template <typename T>
void TesnorToVector(const Tensor& src, std::vector<T>* dst);
D
dzhwinter 已提交
86

87
// copy the result bool to cpu
Y
Yi Wang 已提交
88 89
bool TensorContainsNAN(const framework::Tensor& tensor);
bool TensorContainsInf(const framework::Tensor& tensor);
90 91 92 93 94 95
bool TensorIsfinite(const framework::Tensor& tensor);

// store the result bool in gpu tensor, async operation. Faster than above ones.
void TensorContainsNAN(const framework::Tensor& tensor, framework::Tensor* out);
void TensorContainsInf(const framework::Tensor& tensor, framework::Tensor* out);
void TensorIsfinite(const framework::Tensor& tensor, framework::Tensor* out);
D
dzhwinter 已提交
96

Y
Yi Wang 已提交
97 98 99 100
void TensorToStream(std::ostream& os, const Tensor& tensor,
                    const platform::DeviceContext& dev_ctx);
void TensorFromStream(std::istream& is, Tensor* tensor,
                      const platform::DeviceContext& dev_ctx);
T
tangwei12 已提交
101 102 103
void TensorFromStream(std::istream& is, Tensor* tensor,
                      const platform::DeviceContext& dev_ctx,
                      const size_t& seek, const std::vector<int64_t>& shape);
D
dzhwinter 已提交
104

J
Jack Zhou 已提交
105 106 107 108 109 110 111
// store the bool result tensor in out tensor
void TensorContainsNANV2(const framework::Tensor& tensor,
                         framework::Tensor* out);
void TensorContainsInfV2(const framework::Tensor& tensor,
                         framework::Tensor* out);
void TensorIsfiniteV2(const framework::Tensor& tensor, framework::Tensor* out);

6
633WHU 已提交
112 113 114
// convert dlpack's DLTensor to tensor
void TensorFromDLPack(const ::DLTensor& dl_tensor, framework::Tensor* dst);

Y
Yi Wang 已提交
115 116 117
//
// The implementation of template functions.
//
D
dzhwinter 已提交
118

119 120 121 122 123 124 125 126 127 128 129
template <typename T>
void TensorFromArray(const T* src, const size_t& array_size,
                     const platform::DeviceContext& ctx, Tensor* dst) {
  auto dst_place = ctx.GetPlace();
  auto src_ptr = static_cast<const void*>(src);
  platform::CPUPlace src_place;
  dst->Resize({static_cast<int64_t>(array_size)});
  auto dst_ptr = static_cast<void*>(dst->mutable_data<T>(dst_place));
  auto size = array_size * sizeof(T);

  if (platform::is_cpu_place(dst_place)) {
130 131
    memory::Copy(BOOST_GET_CONST(platform::CPUPlace, dst_place), dst_ptr,
                 src_place, src_ptr, size);
132
  }
133
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
134 135
  else if (platform::is_gpu_place(dst_place)) {  // NOLINT
    memory::Copy(
136 137
        BOOST_GET_CONST(platform::CUDAPlace, dst_place), dst_ptr, src_place,
        src_ptr, size,
138 139 140 141
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
  }
#endif
}
142

D
dzhwinter 已提交
143
template <typename T>
Y
Yi Wang 已提交
144 145
void TensorFromVector(const std::vector<T>& src,
                      const platform::DeviceContext& ctx, Tensor* dst) {
D
dzhwinter 已提交
146 147 148 149 150 151 152 153
  auto dst_place = ctx.GetPlace();
  auto src_ptr = static_cast<const void*>(src.data());
  platform::CPUPlace src_place;
  dst->Resize({static_cast<int64_t>(src.size())});
  auto dst_ptr = static_cast<void*>(dst->mutable_data<T>(dst_place));
  auto size = src.size() * sizeof(T);

  if (platform::is_cpu_place(dst_place)) {
154 155
    memory::Copy(BOOST_GET_CONST(platform::CPUPlace, dst_place), dst_ptr,
                 src_place, src_ptr, size);
D
dzhwinter 已提交
156
  }
157
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
D
dzhwinter 已提交
158 159
  else if (platform::is_gpu_place(dst_place)) {  // NOLINT
    memory::Copy(
160 161
        BOOST_GET_CONST(platform::CUDAPlace, dst_place), dst_ptr, src_place,
        src_ptr, size,
D
dzhwinter 已提交
162 163 164
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
  }
#endif
165
#ifdef PADDLE_WITH_ASCEND_CL
166 167 168 169 170 171
  // NOTE(zhiqiu): Becareful that aclrtMemcpyAsync is different from
  // cudaMemcpyAsync.
  // cudaMemcpyAsync is actually "sync" between cpu <-> gpu.
  // aclrtMemcpyAsync is really "async" between cpu <-> npu.
  // Since vector is on cpu, I think this function should be a "sync" operation,
  // so pass nullptr as stream to  memory::Copy().
172
  else if (platform::is_npu_place(dst_place)) {  // NOLINT
173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
    //  1. vector -> npu pinned tensor
    Tensor npu_pinned_tensor(dst->type());
    platform::NPUPinnedPlace npu_pinned_place;
    auto npu_pinned_ptr =
        npu_pinned_tensor.mutable_data<T>(dst->dims(), npu_pinned_place);
    memory::Copy(npu_pinned_place, npu_pinned_ptr, src_place, src_ptr, size);

    //  2. async copy npu pinned tensor -> npu tensor
    memory::Copy(
        BOOST_GET_CONST(platform::NPUPlace, dst_place), dst_ptr,
        npu_pinned_place, npu_pinned_ptr, size,
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream());

    //  3. record event
    auto npu_pinned_allocator =
        static_cast<paddle::memory::allocation::NPUPinnedAllocator*>(
            paddle::memory::allocation::AllocatorFacade::Instance()
                .GetAllocator(npu_pinned_place)
                .get());
    paddle::memory::allocation::Allocation* allocation =
        npu_pinned_tensor.Holder().get();
    npu_pinned_allocator->RecordEvent(
        allocation,
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream());
197 198
  }
#endif
D
dzhwinter 已提交
199 200
}

201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234
// The fully specialized function should be inline to avoid
// multi-definition.
template <>
inline void TensorFromVector(const std::vector<bool>& src,
                             const platform::DeviceContext& ctx, Tensor* dst) {
  // vector<bool> has no data() member, use array instead.
  // See details:
  // https://stackoverflow.com/questions/46115669/why-does-stdvectorbool-have-no-data/46115714
  bool* array = new bool[src.size()];
  for (unsigned int i = 0; i < src.size(); i++) {
    array[i] = static_cast<bool>(src[i]);
  }

  auto dst_place = ctx.GetPlace();
  auto src_ptr = static_cast<const void*>(array);
  platform::CPUPlace src_place;
  dst->Resize({static_cast<int64_t>(src.size())});
  auto dst_ptr = static_cast<void*>(dst->mutable_data<bool>(dst_place));
  auto size = src.size() * sizeof(bool);

  if (platform::is_cpu_place(dst_place)) {
    memory::Copy(BOOST_GET_CONST(platform::CPUPlace, dst_place), dst_ptr,
                 src_place, src_ptr, size);
  }
#ifdef PADDLE_WITH_CUDA
  else if (platform::is_gpu_place(dst_place)) {  // NOLINT
    memory::Copy(
        BOOST_GET_CONST(platform::CUDAPlace, dst_place), dst_ptr, src_place,
        src_ptr, size,
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
  }
#endif
#ifdef PADDLE_WITH_ASCEND_CL
  else if (platform::is_npu_place(dst_place)) {  // NOLINT
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
    //  1. vector -> npu pinned tensor
    platform::NPUPinnedPlace npu_pinned_place;
    Tensor npu_pinned_tensor;
    npu_pinned_tensor.Resize(dst->dims());
    auto npu_pinned_ptr =
        npu_pinned_tensor.mutable_data(npu_pinned_place, dst->type());
    memory::Copy(npu_pinned_place, npu_pinned_ptr, src_place, src_ptr, size);

    //  2. async copy npu pinned tensor -> npu tensor
    memory::Copy(
        BOOST_GET_CONST(platform::NPUPlace, dst_place), dst_ptr,
        npu_pinned_place, npu_pinned_ptr, size,
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream());

    //  3. record event
    auto npu_pinned_allocator =
        static_cast<paddle::memory::allocation::NPUPinnedAllocator*>(
            paddle::memory::allocation::AllocatorFacade::Instance()
                .GetAllocator(npu_pinned_place)
                .get());
    paddle::memory::allocation::Allocation* allocation =
        npu_pinned_tensor.Holder().get();
    npu_pinned_allocator->RecordEvent(
        allocation,
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream());
260 261 262 263 264
  }
#endif
  delete[] array;
}

D
dzhwinter 已提交
265
template <typename T>
Y
Yi Wang 已提交
266
void TensorFromVector(const std::vector<T>& src, Tensor* dst) {
D
dzhwinter 已提交
267 268 269 270 271 272 273 274 275 276
  platform::CPUPlace dst_place = platform::CPUPlace();
  auto src_ptr = static_cast<const void*>(src.data());
  platform::CPUPlace src_place;
  dst->Resize({static_cast<int64_t>(src.size())});
  auto dst_ptr = static_cast<void*>(dst->mutable_data<T>(dst_place));
  auto size = src.size() * sizeof(T);

  memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
}

277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
template <>
inline void TensorFromVector(const std::vector<bool>& src, Tensor* dst) {
  bool* array = new bool[src.size()];
  for (unsigned int i = 0; i < src.size(); i++) {
    array[i] = static_cast<bool>(src[i]);
  }
  platform::CPUPlace dst_place = platform::CPUPlace();
  auto src_ptr = static_cast<const void*>(array);
  platform::CPUPlace src_place;
  dst->Resize({static_cast<int64_t>(src.size())});
  auto dst_ptr = static_cast<void*>(dst->mutable_data<bool>(dst_place));
  auto size = src.size() * sizeof(bool);

  memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
  delete[] array;
}

D
dzhwinter 已提交
294
template <typename T>
Y
Yi Wang 已提交
295 296
void TensorToVector(const Tensor& src, const platform::DeviceContext& ctx,
                    std::vector<T>* dst) {
D
dzhwinter 已提交
297 298 299 300 301 302 303 304
  auto src_ptr = static_cast<const void*>(src.data<T>());
  auto size = src.numel() * sizeof(T);

  platform::CPUPlace dst_place;
  dst->resize(src.numel());
  auto dst_ptr = static_cast<void*>(dst->data());

  if (platform::is_cpu_place(src.place())) {
305
    memory::Copy(dst_place, dst_ptr,
306 307
                 BOOST_GET_CONST(platform::CPUPlace, src.place()), src_ptr,
                 size);
D
dzhwinter 已提交
308
  }
309
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
D
dzhwinter 已提交
310 311
  else if (platform::is_gpu_place(src.place())) {  // NOLINT
    memory::Copy(
312
        dst_place, dst_ptr, BOOST_GET_CONST(platform::CUDAPlace, src.place()),
313
        src_ptr, size,
D
dzhwinter 已提交
314 315 316
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
  }
#endif
317 318 319 320 321 322 323
#if defined(PADDLE_WITH_XPU)
  else if (platform::is_xpu_place(src.place())) {  // NOLINT
    memory::Copy(dst_place, dst_ptr,
                 BOOST_GET_CONST(platform::XPUPlace, src.place()), src_ptr,
                 size);
  }
#endif
324 325
#ifdef PADDLE_WITH_ASCEND_CL
  else if (platform::is_npu_place(src.place())) {  // NOLINT
326 327 328
    memory::Copy(dst_place, dst_ptr,
                 BOOST_GET_CONST(platform::NPUPlace, src.place()), src_ptr,
                 size, nullptr);
329 330
  }
#endif
D
dzhwinter 已提交
331 332
}

333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350
template <>
inline void TensorToVector(const Tensor& src,
                           const platform::DeviceContext& ctx,
                           std::vector<bool>* dst) {
  auto src_ptr = static_cast<const void*>(src.data<bool>());
  auto size = src.numel() * sizeof(bool);

  bool* array = new bool[src.numel()];

  platform::CPUPlace dst_place;
  dst->resize(src.numel());
  auto dst_ptr = static_cast<void*>(array);

  if (platform::is_cpu_place(src.place())) {
    memory::Copy(dst_place, dst_ptr,
                 BOOST_GET_CONST(platform::CPUPlace, src.place()), src_ptr,
                 size);
  }
351
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
352 353 354 355 356 357 358
  else if (platform::is_gpu_place(src.place())) {  // NOLINT
    memory::Copy(
        dst_place, dst_ptr, BOOST_GET_CONST(platform::CUDAPlace, src.place()),
        src_ptr, size,
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream());
  }
#endif
359 360 361 362 363 364 365
#if defined(PADDLE_WITH_XPU)
  else if (platform::is_xpu_place(src.place())) {  // NOLINT
    memory::Copy(dst_place, dst_ptr,
                 BOOST_GET_CONST(platform::XPUPlace, src.place()), src_ptr,
                 size);
  }
#endif
366 367
#ifdef PADDLE_WITH_ASCEND_CL
  else if (platform::is_npu_place(src.place())) {  // NOLINT
368 369 370
    memory::Copy(dst_place, dst_ptr,
                 BOOST_GET_CONST(platform::NPUPlace, src.place()), src_ptr,
                 size, nullptr);
371 372 373 374 375 376 377 378
  }
#endif
  for (unsigned int i = 0; i < src.numel(); i++) {
    (*dst)[i] = static_cast<bool>(array[i]);
  }
  delete[] array;
}

D
dzhwinter 已提交
379
template <typename T>
Y
Yi Wang 已提交
380
void TensorToVector(const Tensor& src, std::vector<T>* dst) {
D
dzhwinter 已提交
381 382 383 384 385 386 387
  auto src_ptr = static_cast<const void*>(src.data<T>());
  auto size = src.numel() * sizeof(T);

  platform::CPUPlace dst_place;
  dst->resize(src.numel());
  auto dst_ptr = static_cast<void*>(dst->data());

388 389 390 391 392
  PADDLE_ENFORCE_EQ(
      platform::is_cpu_place(src.place()), true,
      platform::errors::InvalidArgument(
          "The input tensor should be CPU device, but actually it is in %s.",
          src.place()));
D
dzhwinter 已提交
393

394 395
  memory::Copy(dst_place, dst_ptr,
               BOOST_GET_CONST(platform::CPUPlace, src.place()), src_ptr, size);
D
dzhwinter 已提交
396
}
397

398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423
template <>
inline void TensorToVector(const Tensor& src, std::vector<bool>* dst) {
  auto src_ptr = static_cast<const void*>(src.data<bool>());
  auto size = src.numel() * sizeof(bool);

  bool* array = new bool[src.numel()];

  platform::CPUPlace dst_place;
  dst->resize(src.numel());
  auto dst_ptr = static_cast<void*>(array);

  PADDLE_ENFORCE_EQ(
      platform::is_cpu_place(src.place()), true,
      platform::errors::InvalidArgument(
          "The input tensor should be CPU device, but actually it is in %s.",
          src.place()));

  memory::Copy(dst_place, dst_ptr,
               BOOST_GET_CONST(platform::CPUPlace, src.place()), src_ptr, size);

  for (unsigned int i = 0; i < src.numel(); i++) {
    (*dst)[i] = static_cast<bool>(array[i]);
  }
  delete[] array;
}

424
std::ostream& operator<<(std::ostream& os, const Tensor& t);
D
dzhwinter 已提交
425 426
}  // namespace framework
}  // namespace paddle