tensor_util.cc 24.2 KB
Newer Older
Y
Yang Yu 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

   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. */
Y
Yi Wang 已提交
14
#include "paddle/fluid/framework/tensor_util.h"
C
chengduoZH 已提交
15 16
#include <algorithm>
#include <limits>
C
chengduo 已提交
17 18
#include <memory>
#include <utility>
C
chengduoZH 已提交
19
#include <vector>
Y
yuyang18 已提交
20
#include "paddle/fluid/framework/data_type.h"
21
#include "paddle/fluid/platform/profiler.h"
Y
Yang Yu 已提交
22 23 24

namespace paddle {
namespace framework {
Y
Yi Wang 已提交
25 26

void TensorCopy(const Tensor& src, const platform::Place& dst_place,
F
fengjiayi 已提交
27
                const platform::DeviceContext& ctx, Tensor* dst) {
28 29 30 31 32 33
  if (&src == dst) {
    auto src_copy = src;
    TensorCopy(src_copy, dst_place, ctx, dst);
    return;
  }

M
minqiyang 已提交
34 35
  VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
          << dst_place;
Y
Yi Wang 已提交
36 37 38 39 40 41 42 43
  src.check_memory_size();

  dst->Resize(src.dims());
  dst->set_layout(src.layout());
  auto src_place = src.place();
  auto src_ptr = src.data<void>();
  auto dst_ptr = dst->mutable_data(dst_place, src.type());

44 45 46 47 48 49
  if (src_ptr == dst_ptr && src_place == dst_place) {
    VLOG(3) << "Skip copy the same data async from " << src_place << " to "
            << dst_place;
    return;
  }

Y
Yi Wang 已提交
50 51 52 53 54 55 56 57 58 59 60 61
  auto size = src.numel() * SizeOfType(src.type());

  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                 boost::get<platform::CPUPlace>(src_place), src_ptr, size);
  }
#ifdef PADDLE_WITH_CUDA
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
    auto dst_cpu_place = boost::get<platform::CPUPlace>(dst_place);
    auto ctx_place = ctx.GetPlace();
62
    PADDLE_ENFORCE_EQ(platform::is_gpu_place(ctx_place), true);
Y
Yi Wang 已提交
63 64
    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
65
    auto stream =
F
fengjiayi 已提交
66
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
67
    memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream);
Y
Yi Wang 已提交
68 69 70 71 72
  } else if (platform::is_cpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
    auto src_cpu_place = boost::get<platform::CPUPlace>(src_place);
    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
    auto ctx_place = ctx.GetPlace();
73
    PADDLE_ENFORCE_EQ(platform::is_gpu_place(ctx_place), true);
Y
Yi Wang 已提交
74 75
    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
    PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place);
76
    auto stream =
F
fengjiayi 已提交
77
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
78
    memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, stream);
Y
Yi Wang 已提交
79 80 81 82 83
  } else if (platform::is_gpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
    auto ctx_place = ctx.GetPlace();
84
    PADDLE_ENFORCE_EQ(platform::is_gpu_place(ctx_place), true);
85
    auto stream =
F
fengjiayi 已提交
86
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
C
chengduo 已提交
87 88 89 90 91 92 93
    if (platform::is_same_place(src_place, dst_place)) {
      memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                   stream);
    } else {
      if (platform::is_same_place(ctx_place, src_place)) {
        memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                     stream);
C
chengduo 已提交
94
        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
C
chengduo 已提交
95
      } else if (platform::is_same_place(ctx_place, dst_place)) {
C
chengduo 已提交
96
        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
C
chengduo 已提交
97 98 99 100 101 102
        memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                     stream);
      } else {
        PADDLE_THROW("ctx is not belong to dst_gpu_place or src_gpu_place.");
      }
    }
103 104
  } else {
    PADDLE_THROW("Copy from %s to %s is not supported.", src_place, dst_place);
Y
Yi Wang 已提交
105 106 107 108 109 110 111 112
  }
#endif
}

void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                Tensor* dst) {
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
  const platform::DeviceContext* dev_ctx;
C
chengduo 已提交
113
  if (platform::is_gpu_place(dst_place)) {
Y
Yi Wang 已提交
114
    dev_ctx = pool.Get(dst_place);
C
chengduo 已提交
115 116
  } else {
    dev_ctx = pool.Get(src.place());
Y
Yi Wang 已提交
117 118 119 120
  }
  TensorCopy(src, dst_place, *dev_ctx, dst);
}

F
fengjiayi 已提交
121 122
void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
                    Tensor* dst) {
123 124 125 126 127 128
  if (&src == dst) {
    auto src_copy = src;
    TensorCopySync(src_copy, dst_place, dst);
    return;
  }

M
minqiyang 已提交
129 130
  VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place()
          << " to " << dst_place;
F
fengjiayi 已提交
131 132 133 134 135 136
  src.check_memory_size();
  dst->Resize(src.dims());
  dst->set_layout(src.layout());
  auto src_place = src.place();
  auto src_ptr = src.data<void>();
  auto dst_ptr = dst->mutable_data(dst_place, src.type());
137 138 139 140 141 142 143

  if (src_ptr == dst_ptr && src_place == dst_place) {
    VLOG(3) << "Skip copy the same data from " << src_place << " to "
            << dst_place;
    return;
  }

F
fengjiayi 已提交
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
  auto size = src.numel() * SizeOfType(src.type());
  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
    memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
                 boost::get<platform::CPUPlace>(src_place), src_ptr, size);
  }
#ifdef PADDLE_WITH_CUDA
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
    auto dst_cpu_place = boost::get<platform::CPUPlace>(dst_place);
    memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
  } else if (platform::is_cpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
    auto src_cpu_place = boost::get<platform::CPUPlace>(src_place);
    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
    memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr);
  } else if (platform::is_gpu_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
M
minqiyang 已提交
162 163
    auto src_gpu_place = boost::get<platform::CUDAPlace>(src_place);
    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
F
fengjiayi 已提交
164
    memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
W
Wu Yi 已提交
165 166 167 168 169 170
  } else if (platform::is_cuda_pinned_place(src_place) &&
             platform::is_gpu_place(dst_place)) {
    auto src_pinned_place = boost::get<platform::CUDAPinnedPlace>(src_place);
    auto dst_gpu_place = boost::get<platform::CUDAPlace>(dst_place);
    memory::Copy(dst_gpu_place, dst_ptr, src_pinned_place, src_ptr, size,
                 nullptr);
171 172
  } else {
    PADDLE_THROW("Copy from %s to %s is not supported.", src_place, dst_place);
F
fengjiayi 已提交
173 174 175 176
  }
#endif
}

Y
Yang Yu 已提交
177 178 179 180 181 182 183 184 185 186 187 188
template <typename Predicate, typename DevCtx>
struct AnyDTypeVisitor {
  Predicate predicate_;
  const Tensor& tensor_;
  const DevCtx& ctx_;
  Tensor* out_;

  AnyDTypeVisitor(Predicate predicate, const Tensor& tensor, const DevCtx& ctx,
                  Tensor* out)
      : predicate_(predicate), tensor_(tensor), ctx_(ctx), out_(out) {}

  template <typename T>
D
dzhwinter 已提交
189
  void apply() const {
Y
Yang Yu 已提交
190 191
    auto t = EigenVector<T>::Flatten(tensor_);
    auto o = EigenScalar<bool>::From(*out_);
Y
Yang Yu 已提交
192
    // return any of predicate_(t) is true.
Y
Yang Yu 已提交
193 194 195 196 197 198 199
    o.device(*ctx_.eigen_device()) = predicate_(t).any();
  }
};

template <typename Predicate, typename DevCtx>
inline void AnyImpl(Predicate predicate, const framework::Tensor& tensor,
                    const DevCtx& ctx, framework::Tensor* out) {
Y
Yu Yang 已提交
200 201
  VisitDataType(tensor.type(), AnyDTypeVisitor<Predicate, DevCtx>(
                                   predicate, tensor, ctx, out));
Y
Yang Yu 已提交
202 203 204
}

template <typename Predicate>
205 206
class AnyVisitor : public boost::static_visitor<bool> {
 private:
Y
Yang Yu 已提交
207 208 209
  const framework::Tensor& tensor_;
  Predicate predicate_;

210
 public:
Y
Yang Yu 已提交
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229
  AnyVisitor(const framework::Tensor& tensor, Predicate predicate)
      : tensor_(tensor), predicate_(std::move(predicate)) {}

  template <typename Place>
  bool operator()(const Place& place) const {
    framework::Tensor out;
    out.Resize({1});
    out.mutable_data<bool>(place);
    auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(place);
    AnyImpl(predicate_, tensor_, *ctx, &out);
    return this->GetResult(out, place);
  }

  bool GetResult(const framework::Tensor& out,
                 const platform::CUDAPlace& gpu) const {
    platform::CPUPlace cpu;
    framework::Tensor tmp;
    tmp.Resize({1});
    tmp.mutable_data<bool>(cpu);
Y
Yang Yu 已提交
230 231
    auto gpuctx = platform::DeviceContextPool::Instance().Get(gpu);
    gpuctx->Wait();
Y
Yi Wang 已提交
232
    TensorCopy(out, cpu, *gpuctx, &tmp);
Y
Yang Yu 已提交
233
    gpuctx->Wait();
Y
Yang Yu 已提交
234 235 236 237 238 239 240
    return GetResult(tmp, cpu);
  }

  bool GetResult(const framework::Tensor& out,
                 const platform::CPUPlace& cpu) const {
    return *out.data<bool>();
  }
C
chengduoZH 已提交
241 242 243 244 245

  bool GetResult(const framework::Tensor& out,
                 const platform::CUDAPinnedPlace& cpu) const {
    return *out.data<bool>();
  }
Y
Yang Yu 已提交
246 247
};

248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
template <typename Predicate>
class AnyOutVisitor : public boost::static_visitor<> {
 private:
  const framework::Tensor& tensor_;
  mutable framework::Tensor* out_;
  Predicate predicate_;

 public:
  AnyOutVisitor(const framework::Tensor& tensor, Predicate predicate,
                framework::Tensor* out)
      : tensor_(tensor), out_(out), predicate_(std::move(predicate)) {}

  template <typename Place>
  void operator()(const Place& place) const {
    auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(place);
    out_->Resize({1});
    out_->mutable_data<bool>(place);
    AnyImpl(predicate_, tensor_, *ctx, out_);
  }
};

Y
Yang Yu 已提交
269 270 271 272 273 274 275
template <typename Predicate>
inline bool Any(const framework::Tensor& tensor, Predicate predicate) {
  AnyVisitor<Predicate> visitor(tensor, predicate);
  auto place = tensor.place();
  return platform::VisitPlace(place, visitor);
}

276 277 278 279 280 281 282 283
template <typename Predicate>
inline void Any(const framework::Tensor& tensor, Predicate predicate,
                framework::Tensor* out) {
  AnyOutVisitor<Predicate> visitor(tensor, predicate, out);
  auto place = tensor.place();
  platform::VisitPlace(place, visitor);
}

Y
Yi Wang 已提交
284
struct ContainsNANPredicate {
Y
Yang Yu 已提交
285 286 287
  template <typename T>
  auto operator()(const T& eigen_vec) const
      -> decltype(std::declval<T>().isnan()) {
Y
Yang Yu 已提交
288
    // Cast eigen_vector to vector of bool. true if is inf.
Y
Yang Yu 已提交
289 290 291 292
    return eigen_vec.isnan();
  }
};

Y
Yi Wang 已提交
293 294
bool TensorContainsNAN(const framework::Tensor& tensor) {
  ContainsNANPredicate predicate;
Y
Yang Yu 已提交
295 296 297
  return Any(tensor, predicate);
}

298 299 300 301 302 303
void TensorContainsNAN(const framework::Tensor& tensor,
                       framework::Tensor* out) {
  ContainsNANPredicate predicate;
  Any(tensor, predicate, out);
}

Y
Yi Wang 已提交
304
struct ContainsInfPredicate {
Y
Yang Yu 已提交
305 306 307
  template <typename T>
  auto operator()(const T& eigen_vec) const
      -> decltype(std::declval<T>().isinf()) {
Y
Yang Yu 已提交
308
    // Cast eigen_vector to vector of bool. true if is inf.
Y
Yang Yu 已提交
309 310 311 312
    return eigen_vec.isinf();
  }
};

Y
Yi Wang 已提交
313 314
bool TensorContainsInf(const framework::Tensor& tensor) {
  ContainsInfPredicate predicate;
Y
Yang Yu 已提交
315 316 317
  return Any(tensor, predicate);
}

318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382
void TensorContainsInf(const framework::Tensor& tensor,
                       framework::Tensor* out) {
  ContainsInfPredicate predicate;
  Any(tensor, predicate, out);
}

// NOTE(dzhwinter):
// Isfinite need a AllVisitor to loop through all the elements.
// We choose two cuda call instead of one allvisitor. The AllVisitor
// should be implemented if the performance hurts.
bool TensorIsfinite(const framework::Tensor& tensor) {
  ContainsInfPredicate pred_inf;
  ContainsNANPredicate pred_nan;
  return !Any(tensor, pred_inf) && !Any(tensor, pred_nan);
}

#ifdef PADDLE_WITH_CUDA
template <typename T>
static inline void __global__ BothFalse(const T* cmp, T* out) {
  out[0] = (!cmp[0]) && (!out[0]);
}
#endif

struct BothFalseVisitor : public boost::static_visitor<> {
  const framework::Tensor& in_;
  mutable framework::Tensor* out_;
  BothFalseVisitor(const framework::Tensor& in, framework::Tensor* out)
      : in_(in), out_(out) {}

  template <typename Place>
  void operator()(const Place& place) const {
    VisitorImpl(place);
  }

  void VisitorImpl(const platform::CUDAPlace& gpu) const {
#ifdef PADDLE_WITH_CUDA
    auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(gpu);
    BothFalse<bool><<<1, 1, 0, ctx->stream()>>>(in_.data<bool>(),
                                                out_->mutable_data<bool>(gpu));
#endif
  }

  void VisitorImpl(const platform::CPUPlace& cpu) const {
    bool lhs = !in_.data<bool>()[0];
    bool rhs = !out_->mutable_data<bool>(cpu)[0];
    out_->mutable_data<bool>(cpu)[0] = lhs && rhs;
  }

  void VisitorImpl(
      const platform::CUDAPinnedPlace& cpu /* equals to cpu*/) const {
    bool lhs = !in_.data<bool>()[0];
    bool rhs = !out_->mutable_data<bool>(cpu)[0];
    out_->mutable_data<bool>(cpu)[0] = lhs && rhs;
  }
};

void TensorIsfinite(const framework::Tensor& tensor, framework::Tensor* out) {
  framework::Tensor tmp;
  TensorContainsInf(tensor, &tmp);
  TensorContainsNAN(tensor, out);
  BothFalseVisitor visitor(tmp, out);
  auto place = tensor.place();
  platform::VisitPlace(place, visitor);
}

Y
Yi Wang 已提交
383 384 385 386 387 388 389 390 391 392
void TensorToStream(std::ostream& os, const Tensor& tensor,
                    const platform::DeviceContext& dev_ctx) {
  {  // the 1st field, uint32_t version
    constexpr uint32_t version = 0;
    os.write(reinterpret_cast<const char*>(&version), sizeof(version));
  }
  {  // the 2nd field, tensor description
     // int32_t  size
     // void*    protobuf message
    proto::VarType::TensorDesc desc;
Y
Yu Yang 已提交
393
    desc.set_data_type(tensor.type());
Y
Yi Wang 已提交
394 395 396 397 398 399 400 401 402 403
    auto dims = framework::vectorize(tensor.dims());
    auto* pb_dims = desc.mutable_dims();
    pb_dims->Resize(static_cast<int>(dims.size()), 0);
    std::copy(dims.begin(), dims.end(), pb_dims->begin());
    int32_t size = desc.ByteSize();
    os.write(reinterpret_cast<const char*>(&size), sizeof(size));
    auto out = desc.SerializeAsString();
    os.write(out.data(), size);
  }
  {  // the 3rd field, tensor data
Y
yuyang18 已提交
404 405
    uint64_t size = tensor.numel() * framework::SizeOfType(tensor.type());

Y
Yi Wang 已提交
406
    auto* data_ptr = tensor.data<void>();
T
tangwei12 已提交
407 408 409
    PADDLE_ENFORCE_LT(size, std::numeric_limits<std::streamsize>::max(),
                      platform::errors::ResourceExhausted(
                          "tensor size %d overflow when writing tensor", size));
Y
Yi Wang 已提交
410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429
    if (platform::is_gpu_place(tensor.place())) {
#ifdef PADDLE_WITH_CUDA
      constexpr size_t kBufSize = 1024 * 1024 * 64;  // 64MB
      std::unique_ptr<char[]> buf(new char[kBufSize]);
      auto& gpu_dev_ctx =
          static_cast<const platform::CUDADeviceContext&>(dev_ctx);
      platform::CPUPlace cpu;
      uintptr_t data = reinterpret_cast<uintptr_t>(data_ptr);
      while (size != 0) {
        size_t size_to_write = std::min(kBufSize, static_cast<size_t>(size));
        memory::Copy(cpu, buf.get(),
                     boost::get<platform::CUDAPlace>(tensor.place()),
                     reinterpret_cast<const void*>(data), size_to_write,
                     gpu_dev_ctx.stream());
        gpu_dev_ctx.Wait();
        os.write(buf.get(), size_to_write);
        data += size_to_write;
        size -= size_to_write;
      }
#else
T
tangwei12 已提交
430 431
      PADDLE_THROW(platform::errors::Unimplemented(
          "CUDAPlace is not supported when not compiled with CUDA"));
Y
Yi Wang 已提交
432 433 434 435 436 437 438 439 440 441 442 443 444 445
#endif
    } else {
      os.write(static_cast<const char*>(data_ptr),
               static_cast<std::streamsize>(size));
    }
  }
}

struct DeserializedDataFunctor {
  DeserializedDataFunctor(void** buf, Tensor* tensor,
                          const platform::Place& place)
      : buf_(buf), tensor_(tensor), place_(place) {}

  template <typename T>
D
dzhwinter 已提交
446
  void apply() {
Y
Yi Wang 已提交
447 448 449 450 451 452 453 454
    *buf_ = tensor_->mutable_data<T>(place_);
  }

  void** buf_;
  Tensor* tensor_;
  platform::Place place_;
};

T
tangwei12 已提交
455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508
void TensorFromStream(std::istream& is, Tensor* tensor,
                      const platform::DeviceContext& dev_ctx,
                      const size_t& seek, const std::vector<int64_t>& shape) {
  uint32_t version;
  is.read(reinterpret_cast<char*>(&version), sizeof(version));

  PADDLE_ENFORCE_EQ(
      version, 0U,
      platform::errors::InvalidArgument(
          "tensor version %u is not supported, Only version 0 is supported",
          version));

  proto::VarType::TensorDesc desc;
  {  // int32_t size
    // proto buffer
    int32_t size;
    is.read(reinterpret_cast<char*>(&size), sizeof(size));
    std::unique_ptr<char[]> buf(new char[size]);
    is.read(reinterpret_cast<char*>(buf.get()), size);
    PADDLE_ENFORCE_EQ(
        desc.ParseFromArray(buf.get(), size), true,
        platform::errors::InvalidArgument("Cannot parse tensor desc"));
  }
  {  // read tensor
    tensor->Resize(framework::make_ddim(shape));
    size_t seekg = seek * framework::SizeOfType(desc.data_type());
    is.seekg(seekg, is.cur);

    void* buf;
    auto ctx = platform::CPUDeviceContext();
    size_t size = tensor->numel() * framework::SizeOfType(desc.data_type());
    if (platform::is_gpu_place(dev_ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
      Tensor cpu_tensor;
      cpu_tensor.Resize(framework::make_ddim(shape));
      framework::VisitDataType(
          desc.data_type(),
          DeserializedDataFunctor(&buf, &cpu_tensor, ctx.GetPlace()));
      is.read(static_cast<char*>(buf), size);
      auto dst_place = dev_ctx.GetPlace();
      framework::TensorCopy(cpu_tensor, dst_place, dev_ctx, tensor);
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "CUDAPlace is not supported when not compiled with CUDA"));
#endif
    } else {
      framework::VisitDataType(
          desc.data_type(),
          DeserializedDataFunctor(&buf, tensor, ctx.GetPlace()));
      is.read(static_cast<char*>(buf), size);
    }
  }
}

Y
Yi Wang 已提交
509 510 511 512
void TensorFromStream(std::istream& is, Tensor* tensor,
                      const platform::DeviceContext& dev_ctx) {
  uint32_t version;
  is.read(reinterpret_cast<char*>(&version), sizeof(version));
T
tangwei12 已提交
513 514 515 516 517
  PADDLE_ENFORCE_EQ(
      version, 0U,
      platform::errors::InvalidArgument(
          "tensor version %u is not supported, Only version 0 is supported",
          version));
Y
Yi Wang 已提交
518 519 520 521 522 523 524
  proto::VarType::TensorDesc desc;
  {  // int32_t size
     // proto buffer
    int32_t size;
    is.read(reinterpret_cast<char*>(&size), sizeof(size));
    std::unique_ptr<char[]> buf(new char[size]);
    is.read(reinterpret_cast<char*>(buf.get()), size);
T
tangwei12 已提交
525 526 527
    PADDLE_ENFORCE_EQ(
        desc.ParseFromArray(buf.get(), size), true,
        platform::errors::InvalidArgument("Cannot parse tensor desc"));
Y
Yi Wang 已提交
528 529 530 531 532 533 534 535
  }
  {  // read tensor
    std::vector<int64_t> dims;
    dims.reserve(static_cast<size_t>(desc.dims().size()));
    std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
    tensor->Resize(framework::make_ddim(dims));
    void* buf;
    auto ctx = platform::CPUDeviceContext();
Y
Yu Yang 已提交
536
    size_t size = tensor->numel() * framework::SizeOfType(desc.data_type());
Y
Yi Wang 已提交
537 538 539 540 541 542 543
    if (platform::is_gpu_place(dev_ctx.GetPlace())) {
#ifdef PADDLE_WITH_CUDA
      Tensor cpu_tensor;
      cpu_tensor.Resize(framework::make_ddim(dims));
      framework::VisitDataType(
          desc.data_type(),
          DeserializedDataFunctor(&buf, &cpu_tensor, ctx.GetPlace()));
Y
yuyang18 已提交
544
      is.read(static_cast<char*>(buf), size);
Y
Yi Wang 已提交
545 546 547
      auto dst_place = dev_ctx.GetPlace();
      framework::TensorCopy(cpu_tensor, dst_place, dev_ctx, tensor);
#else
T
tangwei12 已提交
548 549
      PADDLE_THROW(platform::errors::Unimplemented(
          "CUDAPlace is not supported when not compiled with CUDA"));
Y
Yi Wang 已提交
550 551 552 553 554
#endif
    } else {
      framework::VisitDataType(
          desc.data_type(),
          DeserializedDataFunctor(&buf, tensor, ctx.GetPlace()));
Y
yuyang18 已提交
555
      is.read(static_cast<char*>(buf), size);
Y
Yi Wang 已提交
556 557 558 559
    }
  }
}

6
633WHU 已提交
560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635
// get tensor data point by DLDataType
void* GetDstPtrByDLDataType(DLDataType type, framework::Tensor* dst,
                            const platform::Place& dst_place) {
  // vector types not currently supported
  PADDLE_ENFORCE_LE(type.lanes, 1, "vector types not currently supported");

  switch (type.bits) {
    case 8:
      if (type.code == kDLInt)
        return static_cast<void*>(dst->mutable_data<int8_t>(dst_place));
      if (type.code == kDLUInt)
        return static_cast<void*>(dst->mutable_data<uint8_t>(dst_place));
      PADDLE_THROW("There is no this type.code <%d> when type.bits is <%d>.",
                   type.code, type.bits);
    case 16:
      if (type.code == kDLInt)
        return static_cast<void*>(dst->mutable_data<int16_t>(dst_place));
      if (type.code == kDLFloat)
        return static_cast<void*>(
            dst->mutable_data<paddle::platform::float16>(dst_place));
      PADDLE_THROW("There is no this type.code <%d> when type.bits is <%d>.",
                   type.code, type.bits);
    case 32:
      if (type.code == kDLInt)
        return static_cast<void*>(dst->mutable_data<int32_t>(dst_place));
      if (type.code == kDLFloat)
        return static_cast<void*>(dst->mutable_data<float>(dst_place));
      PADDLE_THROW("There is no this type.code <%d> when type.bits is <%d>.",
                   type.code, type.bits);
    case 64:
      if (type.code == kDLInt)
        return static_cast<void*>(dst->mutable_data<int64_t>(dst_place));
      if (type.code == kDLFloat)
        return static_cast<void*>(dst->mutable_data<double>(dst_place));
      PADDLE_THROW("There is no this type.code <%d> when type.bits is <%d>.",
                   type.code, type.bits);
    default:
      PADDLE_THROW("Unsupport type.bits %d", type.bits);
  }
}

void TensorFromDLPack(const ::DLTensor& dl_tensor, framework::Tensor* dst) {
  platform::CPUPlace dst_place = platform::CPUPlace();
  platform::CPUPlace src_place = platform::CPUPlace();

  std::vector<int64_t> vec;
  std::copy(dl_tensor.shape, dl_tensor.shape + dl_tensor.ndim,
            std::back_inserter(vec));

  framework::DDim vddim = framework::make_ddim(vec);

  dst->Resize(vddim);
  ::DLDataType type = dl_tensor.dtype;
  void* dst_ptr = GetDstPtrByDLDataType(type, dst, dst_place);

  auto src_ptr = static_cast<const void*>(dl_tensor.data);
  auto size = paddle::framework::product(vddim) * type.bits / 8;

  if (dl_tensor.ctx.device_type == kDLCPU) {
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
  }
#ifdef PADDLE_WITH_CUDA
  if (dl_tensor.ctx.device_type == kDLGPU) {
    platform::CUDAPlace dst_place =
        platform::CUDAPlace(dl_tensor.ctx.device_id);
    platform::CUDAPlace src_place =
        platform::CUDAPlace(dl_tensor.ctx.device_id);
    dst_ptr = GetDstPtrByDLDataType(type, dst, dst_place);
    auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(dst_place);
    memory::Copy(
        dst_place, dst_ptr, src_place, src_ptr, size,
        reinterpret_cast<const platform::CUDADeviceContext&>(*ctx).stream());
  }
#endif
}

636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681
template <typename T>
std::ostream& print_tensor(std::ostream& os, const framework::Tensor& tensor) {
  auto inspect = tensor.data<T>();
  auto element_num = tensor.numel();

  os << "\tdata: [";
  if (element_num > 0) {
    os << inspect[0];
    for (int j = 1; j < element_num; ++j) {
      os << " " << inspect[j];
    }
  }
  os << "]";
  return os;
}

std::ostream& operator<<(std::ostream& os, const Tensor& t) {
  os << "\tdim: " << t.dims() << "\n";
  os << "\tlayout: " << DataLayoutToString(t.layout()) << "\n";

  Tensor tensor;
  tensor.Resize(t.dims());
  if (platform::is_cpu_place(t.place())) {
    tensor.ShareDataWith(t);
  } else {
    platform::CPUPlace place;
    framework::TensorCopy(t, place, &tensor);
    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
    auto& dev_ctx = *pool.Get(t.place());
    dev_ctx.Wait();
  }

#define PrintTensorCallback(cpp_type, proto_type) \
  do {                                            \
    if (tensor.type() == proto_type) {            \
      os << "\tdtype: " << proto_type << "\n";    \
      print_tensor<cpp_type>(os, tensor);         \
      return os;                                  \
    }                                             \
  } while (0)

  _ForEachDataType_(PrintTensorCallback);
  VLOG(1) << "PrintVar: unrecognized data type:" << t.type();
  return os;
}

Y
Yang Yu 已提交
682 683
}  // namespace framework
}  // namespace paddle