tensor_util.cc 17.5 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 17
#include <algorithm>
#include <limits>
#include <vector>
Y
yuyang18 已提交
18
#include "paddle/fluid/framework/data_type.h"
Y
Yang Yu 已提交
19 20 21

namespace paddle {
namespace framework {
Y
Yi Wang 已提交
22 23

void TensorCopy(const Tensor& src, const platform::Place& dst_place,
F
fengjiayi 已提交
24
                const platform::DeviceContext& ctx, Tensor* dst) {
Y
Yi Wang 已提交
25 26 27 28 29 30 31 32 33 34 35 36 37 38
  VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
          << dst_place;
  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());

  auto size = src.numel() * SizeOfType(src.type());

  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
M
minqiyang 已提交
39 40 41 42 43
    if (src_ptr == dst_ptr) {
      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
              << dst_place;
      return;
    }
Y
Yi Wang 已提交
44 45 46 47 48 49 50 51 52 53 54 55
    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();
    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place);
56
    auto stream =
F
fengjiayi 已提交
57
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
58
    memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream);
Y
Yi Wang 已提交
59 60 61 62 63 64 65 66
  } 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();
    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
    auto ctx_gpu_place = boost::get<platform::CUDAPlace>(ctx_place);
    PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place);
67
    auto stream =
F
fengjiayi 已提交
68
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
69
    memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, stream);
Y
Yi Wang 已提交
70 71 72 73 74 75
  } 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();
    PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
76
    auto stream =
F
fengjiayi 已提交
77
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
C
chengduo 已提交
78
    if (platform::is_same_place(src_place, dst_place)) {
M
minqiyang 已提交
79 80 81 82 83
      if (src_ptr == dst_ptr) {
        VLOG(3) << "Skip copy the same data async from " << src_place << " to "
                << dst_place;
        return;
      }
C
chengduo 已提交
84 85 86 87 88 89
      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
90
        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
C
chengduo 已提交
91
      } else if (platform::is_same_place(ctx_place, dst_place)) {
C
chengduo 已提交
92
        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
C
chengduo 已提交
93 94 95 96 97 98
        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.");
      }
    }
Y
Yi Wang 已提交
99 100 101 102 103 104 105 106
  }
#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
107
  if (platform::is_gpu_place(dst_place)) {
Y
Yi Wang 已提交
108
    dev_ctx = pool.Get(dst_place);
C
chengduo 已提交
109 110
  } else {
    dev_ctx = pool.Get(src.place());
Y
Yi Wang 已提交
111 112 113 114
  }
  TensorCopy(src, dst_place, *dev_ctx, dst);
}

F
fengjiayi 已提交
115 116 117 118 119 120 121 122 123 124 125 126
void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
                    Tensor* dst) {
  VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place()
          << " to " << dst_place;
  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());
  auto size = src.numel() * SizeOfType(src.type());
  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
127
    if (src_ptr == dst_ptr) {
M
minqiyang 已提交
128
      VLOG(3) << "Skip copy the same data from " << src_place << " to "
129 130 131
              << dst_place;
      return;
    }
F
fengjiayi 已提交
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147
    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
148 149
    if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) {
      VLOG(3) << "Skip copy the same data from " << src_place << " to "
150 151 152
              << dst_place;
      return;
    }
M
minqiyang 已提交
153 154
    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
155
    memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
W
Wu Yi 已提交
156 157 158 159 160 161
  } 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);
F
fengjiayi 已提交
162 163 164 165
  }
#endif
}

Y
Yang Yu 已提交
166 167 168 169 170 171 172 173 174 175 176 177
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
178
  void apply() const {
Y
Yang Yu 已提交
179 180
    auto t = EigenVector<T>::Flatten(tensor_);
    auto o = EigenScalar<bool>::From(*out_);
Y
Yang Yu 已提交
181
    // return any of predicate_(t) is true.
Y
Yang Yu 已提交
182 183 184 185 186 187 188 189 190 191 192 193
    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) {
  VisitDataType(ToDataType(tensor.type()), AnyDTypeVisitor<Predicate, DevCtx>(
                                               predicate, tensor, ctx, out));
}

template <typename Predicate>
194 195
class AnyVisitor : public boost::static_visitor<bool> {
 private:
Y
Yang Yu 已提交
196 197 198
  const framework::Tensor& tensor_;
  Predicate predicate_;

199
 public:
Y
Yang Yu 已提交
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
  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
219 220
    auto gpuctx = platform::DeviceContextPool::Instance().Get(gpu);
    gpuctx->Wait();
Y
Yi Wang 已提交
221
    TensorCopy(out, cpu, *gpuctx, &tmp);
Y
Yang Yu 已提交
222
    gpuctx->Wait();
Y
Yang Yu 已提交
223 224 225 226 227 228 229
    return GetResult(tmp, cpu);
  }

  bool GetResult(const framework::Tensor& out,
                 const platform::CPUPlace& cpu) const {
    return *out.data<bool>();
  }
C
chengduoZH 已提交
230 231 232 233 234

  bool GetResult(const framework::Tensor& out,
                 const platform::CUDAPinnedPlace& cpu) const {
    return *out.data<bool>();
  }
Y
Yang Yu 已提交
235 236
};

237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257
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
258 259 260 261 262 263 264
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);
}

265 266 267 268 269 270 271 272
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
273
struct ContainsNANPredicate {
Y
Yang Yu 已提交
274 275 276
  template <typename T>
  auto operator()(const T& eigen_vec) const
      -> decltype(std::declval<T>().isnan()) {
Y
Yang Yu 已提交
277
    // Cast eigen_vector to vector of bool. true if is inf.
Y
Yang Yu 已提交
278 279 280 281
    return eigen_vec.isnan();
  }
};

Y
Yi Wang 已提交
282 283
bool TensorContainsNAN(const framework::Tensor& tensor) {
  ContainsNANPredicate predicate;
Y
Yang Yu 已提交
284 285 286
  return Any(tensor, predicate);
}

287 288 289 290 291 292
void TensorContainsNAN(const framework::Tensor& tensor,
                       framework::Tensor* out) {
  ContainsNANPredicate predicate;
  Any(tensor, predicate, out);
}

Y
Yi Wang 已提交
293
struct ContainsInfPredicate {
Y
Yang Yu 已提交
294 295 296
  template <typename T>
  auto operator()(const T& eigen_vec) const
      -> decltype(std::declval<T>().isinf()) {
Y
Yang Yu 已提交
297
    // Cast eigen_vector to vector of bool. true if is inf.
Y
Yang Yu 已提交
298 299 300 301
    return eigen_vec.isinf();
  }
};

Y
Yi Wang 已提交
302 303
bool TensorContainsInf(const framework::Tensor& tensor) {
  ContainsInfPredicate predicate;
Y
Yang Yu 已提交
304 305 306
  return Any(tensor, predicate);
}

307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371
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
372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392
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;
    desc.set_data_type(framework::ToDataType(tensor.type()));
    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
393 394
    uint64_t size = tensor.numel() * framework::SizeOfType(tensor.type());

Y
Yi Wang 已提交
395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
    auto* data_ptr = tensor.data<void>();
    PADDLE_ENFORCE(size < std::numeric_limits<std::streamsize>::max(),
                   "Index overflow when writing tensor");
    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
      PADDLE_THROW("Unexpected branch");
#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
433
  void apply() {
Y
Yi Wang 已提交
434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
    *buf_ = tensor_->mutable_data<T>(place_);
  }

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

void TensorFromStream(std::istream& is, Tensor* tensor,
                      const platform::DeviceContext& dev_ctx) {
  uint32_t version;
  is.read(reinterpret_cast<char*>(&version), sizeof(version));
  PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
  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(desc.ParseFromArray(buf.get(), size),
                   "Cannot parse tensor desc");
  }
  {  // 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
yuyang18 已提交
464 465 466
    size_t size =
        tensor->numel() *
        framework::SizeOfType(framework::ToTypeIndex(desc.data_type()));
Y
Yi Wang 已提交
467 468 469 470 471 472 473
    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
474
      is.read(static_cast<char*>(buf), size);
Y
Yi Wang 已提交
475 476 477 478 479 480 481 482 483
      auto dst_place = dev_ctx.GetPlace();
      framework::TensorCopy(cpu_tensor, dst_place, dev_ctx, tensor);
#else
      PADDLE_THROW("Unexpected branch");
#endif
    } else {
      framework::VisitDataType(
          desc.data_type(),
          DeserializedDataFunctor(&buf, tensor, ctx.GetPlace()));
Y
yuyang18 已提交
484
      is.read(static_cast<char*>(buf), size);
Y
Yi Wang 已提交
485 486 487 488
    }
  }
}

Y
Yang Yu 已提交
489 490
}  // namespace framework
}  // namespace paddle