tensor_util.cc 53.3 KB
Newer Older
Y
Yang Yu 已提交
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

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
Y
Yang Yu 已提交
6

7 8 9 10 11 12 13
    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
Yang Yu 已提交
14

C
chengduoZH 已提交
15 16
#include <algorithm>
#include <limits>
C
chengduo 已提交
17
#include <memory>
18
#include <string>
C
chengduo 已提交
19
#include <utility>
C
chengduoZH 已提交
20
#include <vector>
21

22
#include "paddle/fluid/framework/convert_utils.h"
Y
yuyang18 已提交
23
#include "paddle/fluid/framework/data_type.h"
S
Steffy-zxf 已提交
24
#include "paddle/fluid/framework/tensor_util.h"
25
#include "paddle/fluid/platform/complex.h"
26
#include "paddle/fluid/platform/profiler.h"
27 28 29

#include "paddle/pten/core/dense_tensor.h"

30
#ifdef PADDLE_WITH_MKLDNN
31
#include "dnnl_debug.h"  // NOLINT
32
#endif
Y
Yang Yu 已提交
33 34 35

namespace paddle {
namespace framework {
Y
Yi Wang 已提交
36

37 38 39
template <typename TENSOR>
void TensorCopyImpl(const TENSOR& src, const platform::Place& dst_place,
                    const platform::DeviceContext& ctx, TENSOR* dst) {
40 41
  if (&src == dst) {
    auto src_copy = src;
42
    TensorCopyImpl(src_copy, dst_place, ctx, dst);
43 44 45
    return;
  }

M
minqiyang 已提交
46 47
  VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to "
          << dst_place;
Y
Yi Wang 已提交
48 49 50 51
  src.check_memory_size();
  dst->Resize(src.dims());
  dst->set_layout(src.layout());
  auto src_place = src.place();
52
  auto src_ptr = src.data();
53 54 55 56 57 58
#ifdef PADDLE_WITH_MKLDNN
  dst->set_format(src.format());
  // oneDNN tensors due to padding may be of bigger size
  // than numel()*size(type())
  auto dst_ptr =
      src.layout() == DataLayout::kMKLDNN
59 60
          ? dst->mutable_data(dst_place, src.dtype(), src.memory_size())
          : dst->mutable_data(dst_place, src.dtype());
61
#else
62
  auto dst_ptr = dst->mutable_data(dst_place, src.dtype());
63
#endif
64 65 66 67 68
  if (src_ptr == dst_ptr && src_place == dst_place) {
    VLOG(3) << "Skip copy the same data async from " << src_place << " to "
            << dst_place;
    return;
  }
69
  VLOG(4) << "src:" << src_ptr << ", dst:" << dst_ptr;
70

71 72 73
#ifdef PADDLE_WITH_MKLDNN
  auto size = src.layout() == DataLayout::kMKLDNN
                  ? src.memory_size()
74
                  : src.numel() * framework::DataTypeSize(src.dtype());
75
#else
76
  auto size = src.numel() * framework::DataTypeSize(src.dtype());
77
#endif
Y
Yi Wang 已提交
78 79

  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
80
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
Y
Yi Wang 已提交
81
  }
J
jianghaicheng 已提交
82 83 84
#ifdef PADDLE_WITH_IPU
  else if (platform::is_ipu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
85
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
J
jianghaicheng 已提交
86 87
  } else if (platform::is_cpu_place(src_place) &&
             platform::is_ipu_place(dst_place)) {
88
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
J
jianghaicheng 已提交
89 90
  } else if (platform::is_ipu_place(src_place) &&
             platform::is_ipu_place(dst_place)) {
91
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
J
jianghaicheng 已提交
92 93 94
  }
#endif

95 96 97
#ifdef PADDLE_WITH_XPU
  else if (platform::is_xpu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
98
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
99 100
  } else if (platform::is_cpu_place(src_place) &&
             platform::is_xpu_place(dst_place)) {
101
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
102 103 104 105 106 107 108
  } else if (platform::is_xpu_place(src_place) &&
             platform::is_xpu_place(dst_place)) {
    if (src_ptr == dst_ptr) {
      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
              << dst_place;
      return;
    }
109
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
110 111 112 113 114
  } else {
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copy from %s to %s is not supported.", src_place, dst_place));
  }
#endif
115 116 117 118 119 120
#ifdef PADDLE_WITH_ASCEND_CL
  // TODO(zhiqiu): handle different condition like CUDA code below
  else if (platform::is_npu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
    auto stream =
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream();
121
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, stream);
122 123 124
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_npu_place(dst_place)) {
125 126 127 128 129
    //  1. cpu tensor -> npu pinned tensor
    platform::NPUPinnedPlace npu_pinned_place;
    Tensor npu_pinned_tensor;
    npu_pinned_tensor.Resize(src.dims());
    auto npu_pinned_ptr =
130
        npu_pinned_tensor.mutable_data(npu_pinned_place, src.dtype());
131
    memory::Copy(npu_pinned_place, npu_pinned_ptr, src_place, src_ptr, size);
132 133 134

    //  2. async copy npu pinned tensor -> npu tensor
    memory::Copy(
135
        dst_place, dst_ptr, npu_pinned_place, npu_pinned_ptr, size,
136 137 138 139 140 141 142 143
        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());
144
    pten::Allocation* allocation = npu_pinned_tensor.Holder().get();
145 146 147
    npu_pinned_allocator->RecordEvent(
        allocation,
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream());
148 149 150 151 152 153 154 155 156 157
  }
  else if (platform::is_npu_place(src_place) &&  // NOLINT
           platform::is_npu_place(dst_place)) {
    if (src_ptr == dst_ptr) {
      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
              << dst_place;
      return;
    }
    auto stream =
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream();
158
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, stream);
159
  }
W
WangXi 已提交
160 161
  else if (platform::is_npu_pinned_place(src_place) &&  // NOLINT
           platform::is_npu_place(dst_place)) {         /* npu_pinned->npu */
162 163
    auto src_npu_pinned_place = src_place;
    auto dst_npu_place = dst_place;
W
WangXi 已提交
164 165 166 167 168 169
    auto ctx_place = ctx.GetPlace();
    PADDLE_ENFORCE_EQ(platform::is_npu_place(ctx_place), true,
                      platform::errors::PreconditionNotMet(
                          "Device context place mismatch. When copying Tensor "
                          "data from NPU Pinned memory to NPU memory, current "
                          "device context place should be NPU."));
170
    auto ctx_npu_place = ctx_place;
W
WangXi 已提交
171 172 173 174 175 176 177 178 179 180 181 182 183
    PADDLE_ENFORCE_EQ(dst_npu_place, ctx_npu_place,
                      platform::errors::PreconditionNotMet(
                          "The target NPU device and current device context do "
                          "not match. The target NPU device number is %d, but "
                          "device context NPU number is %d.",
                          dst_npu_place.device, ctx_npu_place.device));
    auto stream =
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream();
    memory::Copy(dst_npu_place, dst_ptr, src_npu_pinned_place, src_ptr, size,
                 stream);
  }
  else if (platform::is_npu_place(src_place) &&        // NOLINT
           platform::is_npu_pinned_place(dst_place)) { /* npu->npu_pinned */
184 185
    auto src_npu_place = src_place;
    auto dst_npu_pinned_place = dst_place;
W
WangXi 已提交
186 187 188 189 190 191
    auto ctx_place = ctx.GetPlace();
    PADDLE_ENFORCE_EQ(platform::is_npu_place(ctx_place), true,
                      platform::errors::PreconditionNotMet(
                          "Device context place mismatch. When copying Tensor "
                          "data from NPU memory to NPU Pinned memory, current "
                          "device context place should be NPU."));
192
    auto ctx_npu_place = ctx_place;
W
WangXi 已提交
193 194 195 196 197 198 199 200 201 202 203
    PADDLE_ENFORCE_EQ(src_place, ctx_npu_place,
                      platform::errors::PreconditionNotMet(
                          "The source NPU device and current device context do "
                          "not match. The source NPU device number is %d, but "
                          "device context NPU number is %d.",
                          src_npu_place.device, ctx_npu_place.device));
    auto stream =
        reinterpret_cast<const platform::NPUDeviceContext&>(ctx).stream();
    memory::Copy(dst_npu_pinned_place, dst_ptr, src_npu_place, src_ptr, size,
                 stream);
  }
204 205 206 207 208
  else {  // NOLINT
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copy from %s to %s is not supported.", src_place, dst_place));
  }
#endif
209
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
210 211
  else if (platform::is_cuda_pinned_place(src_place) &&  // NOLINT
           platform::is_cuda_pinned_place(dst_place)) {
212
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
213
  }
214
  else if (platform::is_cuda_pinned_place(src_place) &&  // NOLINT
Y
Yi Wang 已提交
215
           platform::is_cpu_place(dst_place)) {
216
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
217 218 219
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_cuda_pinned_place(dst_place)) {
220
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
221 222 223
  }
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
224 225
    auto src_gpu_place = src_place;
    auto dst_cpu_place = dst_place;
Y
Yi Wang 已提交
226
    auto ctx_place = ctx.GetPlace();
227 228 229 230 231
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx_place), true,
        platform::errors::PreconditionNotMet(
            "Context place error, excepted GPUPlace, but actually %s.",
            ctx_place));
232
    auto ctx_gpu_place = ctx_place;
233 234 235 236 237
    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place,
                      platform::errors::Unavailable(
                          "Source place and context place do not match, source "
                          "place is %s, context place is %s.",
                          src_gpu_place, ctx_gpu_place));
238
    auto stream =
F
fengjiayi 已提交
239
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
240
    memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream);
241 242 243
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_gpu_place(dst_place)) {
244 245
    auto src_cpu_place = src_place;
    auto dst_gpu_place = dst_place;
Y
Yi Wang 已提交
246
    auto ctx_place = ctx.GetPlace();
247 248 249 250 251
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx_place), true,
        platform::errors::PreconditionNotMet(
            "Context place error, excepted GPUPlace, but actually %s.",
            ctx_place));
252
    auto ctx_gpu_place = ctx_place;
253 254 255 256 257
    PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place,
                      platform::errors::Unavailable(
                          "Destination place and context place do not match, "
                          "destination place is %s, context place is %s.",
                          dst_gpu_place, ctx_gpu_place));
258
    auto stream =
F
fengjiayi 已提交
259
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
260
    memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, stream);
261 262 263
  }
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_cuda_pinned_place(dst_place)) {
264 265
    auto src_gpu_place = src_place;
    auto dst_cuda_pinned_place = dst_place;
266 267 268 269 270 271
    auto ctx_place = ctx.GetPlace();
    PADDLE_ENFORCE_EQ(platform::is_gpu_place(ctx_place), true,
                      platform::errors::PreconditionNotMet(
                          "Device context place mismatch. When copying Tensor "
                          "data from GPU memory to CUDA Pinned memory, current "
                          "device context place should be GPU."));
272
    auto ctx_gpu_place = ctx_place;
273 274 275 276 277 278 279 280 281 282
    PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place,
                      platform::errors::PreconditionNotMet(
                          "The source GPU device and current device context do "
                          "not match. The source GPU device number is %d, but "
                          "device context GPU number is %d.",
                          src_gpu_place.device, ctx_gpu_place.device));
    auto stream =
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
    memory::Copy(dst_cuda_pinned_place, dst_ptr, src_gpu_place, src_ptr, size,
                 stream);
283 284 285
  }
  else if (platform::is_cuda_pinned_place(src_place) &&  // NOLINT
           platform::is_gpu_place(dst_place)) {
286 287
    auto src_cuda_pinned_place = src_place;
    auto dst_gpu_place = dst_place;
288 289 290 291 292 293
    auto ctx_place = ctx.GetPlace();
    PADDLE_ENFORCE_EQ(platform::is_gpu_place(ctx_place), true,
                      platform::errors::PreconditionNotMet(
                          "Device context place mismatch. When copying Tensor "
                          "data from CUDA Pinned memory to GPU memory, current "
                          "device context place should be GPU."));
294
    auto ctx_gpu_place = ctx_place;
295 296 297 298 299 300 301 302 303 304
    PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place,
                      platform::errors::PreconditionNotMet(
                          "The target GPU device and current device context do "
                          "not match. The target GPU device number is %d, but "
                          "device context GPU number is %d.",
                          dst_gpu_place.device, ctx_gpu_place.device));
    auto stream =
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
    memory::Copy(dst_gpu_place, dst_ptr, src_cuda_pinned_place, src_ptr, size,
                 stream);
305 306 307
  }
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_gpu_place(dst_place)) {
308 309
    auto src_gpu_place = src_place;
    auto dst_gpu_place = dst_place;
Y
Yi Wang 已提交
310
    auto ctx_place = ctx.GetPlace();
311 312 313 314 315
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx_place), true,
        platform::errors::PreconditionNotMet(
            "Context place error, excepted GPUPlace, but actually %s.",
            ctx_place));
316
    auto stream =
F
fengjiayi 已提交
317
        reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
C
chengduo 已提交
318 319 320 321 322 323 324
    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
325
        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
C
chengduo 已提交
326
      } else if (platform::is_same_place(ctx_place, dst_place)) {
C
chengduo 已提交
327
        platform::DeviceContextPool::Instance().Get(src.place())->Wait();
C
chengduo 已提交
328 329 330
        memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
                     stream);
      } else {
331 332
        PADDLE_THROW(platform::errors::Unavailable(
            "Context place dose not match the source and destination place."));
C
chengduo 已提交
333 334
      }
    }
335 336
  }
  else {  // NOLINT
337 338
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copying from %s to %s is not supported.", src_place, dst_place));
Y
Yi Wang 已提交
339 340
  }
#endif
341 342 343
#ifdef PADDLE_WITH_MLU
  else if (platform::is_mlu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
344 345
    auto src_mlu_place = src_place;
    auto dst_cpu_place = dst_place;
346 347 348 349 350 351
    auto stream =
        reinterpret_cast<const platform::MLUDeviceContext&>(ctx).stream();
    memory::Copy(dst_cpu_place, dst_ptr, src_mlu_place, src_ptr, size, stream);
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_mlu_place(dst_place)) {
352 353
    auto src_cpu_place = src_place;
    auto dst_mlu_place = dst_place;
354 355 356 357 358 359
    auto stream =
        reinterpret_cast<const platform::MLUDeviceContext&>(ctx).stream();
    memory::Copy(dst_mlu_place, dst_ptr, src_cpu_place, src_ptr, size, stream);
  }
  else if (platform::is_mlu_place(src_place) &&  // NOLINT
           platform::is_mlu_place(dst_place)) {
360 361
    auto src_mlu_place = src_place;
    auto dst_mlu_place = dst_place;
362 363 364 365 366 367 368 369 370
    auto stream =
        reinterpret_cast<const platform::MLUDeviceContext&>(ctx).stream();
    memory::Copy(dst_mlu_place, dst_ptr, src_mlu_place, src_ptr, size, stream);
  }
  else {  // NOLINT
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copying from %s to %s is not supported.", src_place, dst_place));
  }
#endif
Y
Yi Wang 已提交
371 372
}

373 374 375
template <typename TENSOR>
void TensorCopyImpl(const TENSOR& src, const platform::Place& dst_place,
                    TENSOR* dst) {
Y
Yi Wang 已提交
376 377
  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
  const platform::DeviceContext* dev_ctx;
378 379
  if (platform::is_gpu_place(dst_place) || platform::is_npu_place(dst_place) ||
      platform::is_mlu_place(dst_place)) {
Y
Yi Wang 已提交
380
    dev_ctx = pool.Get(dst_place);
C
chengduo 已提交
381 382
  } else {
    dev_ctx = pool.Get(src.place());
Y
Yi Wang 已提交
383
  }
384 385 386 387 388 389 390 391 392 393 394
  TensorCopyImpl(src, dst_place, *dev_ctx, dst);
}

void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                Tensor* dst) {
  TensorCopyImpl<Tensor>(src, dst_place, dst);
}
void TensorCopy(const Tensor& src, const platform::Place& dst_place,
                const platform::DeviceContext& ctx, Tensor* dst) {
  TensorCopyImpl<Tensor>(src, dst_place, ctx, dst);
}
Y
Yi Wang 已提交
395

F
fengjiayi 已提交
396 397
void TensorCopySync(const Tensor& src, const platform::Place& dst_place,
                    Tensor* dst) {
398 399 400 401 402 403
  if (&src == dst) {
    auto src_copy = src;
    TensorCopySync(src_copy, dst_place, dst);
    return;
  }

M
minqiyang 已提交
404 405
  VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place()
          << " to " << dst_place;
F
fengjiayi 已提交
406 407 408
  src.check_memory_size();
  dst->Resize(src.dims());
  dst->set_layout(src.layout());
J
Jacek Czaja 已提交
409 410 411
#ifdef PADDLE_WITH_MKLDNN
  dst->set_format(src.format());
#endif
F
fengjiayi 已提交
412
  auto src_place = src.place();
413
  auto src_ptr = src.data();
414
  auto dst_ptr = dst->mutable_data(dst_place, src.dtype());
415
  VLOG(4) << "src:" << src_ptr << ", dst:" << dst_ptr;
416 417 418 419 420 421 422

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

423
  auto size = src.numel() * framework::DataTypeSize(src.dtype());
F
fengjiayi 已提交
424
  if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
425
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
F
fengjiayi 已提交
426
  }
J
jianghaicheng 已提交
427 428 429
#ifdef PADDLE_WITH_IPU
  else if (platform::is_ipu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
430
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
J
jianghaicheng 已提交
431 432
  } else if (platform::is_cpu_place(src_place) &&  // NOLINT
             platform::is_ipu_place(dst_place)) {
433
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
J
jianghaicheng 已提交
434 435 436 437 438
  } else {  // NOLINT
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copy from %s to %s is not supported.", src_place, dst_place));
  }
#endif
439 440 441
#ifdef PADDLE_WITH_XPU
  else if (platform::is_xpu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
442
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
J
jianghaicheng 已提交
443 444 445
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_xpu_place(dst_place)) {
446
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
J
jianghaicheng 已提交
447 448 449
  }
  else if (platform::is_xpu_place(src_place) &&  // NOLINT
           platform::is_xpu_place(dst_place)) {
450 451 452 453 454
    if (src_ptr == dst_ptr) {
      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
              << dst_place;
      return;
    }
455 456 457
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
    platform::XPUPlace xpu_dst_place = dst_place;
    platform::XPUPlace xpu_src_place = src_place;
458 459 460 461
    if (xpu_dst_place.device == xpu_src_place.device) {
      auto xpu_ctx = platform::DeviceContextPool::Instance().Get(xpu_dst_place);
      xpu_ctx->Wait();
    }
J
jianghaicheng 已提交
462 463
  }
  else {  // NOLINT
464 465 466 467
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copy from %s to %s is not supported.", src_place, dst_place));
  }
#endif
468 469 470
#ifdef PADDLE_WITH_ASCEND_CL
  else if (platform::is_npu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {  /* npu -> cpu*/
471
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, nullptr);
472 473 474
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_npu_place(dst_place)) {  /* cpu -> npu*/
475
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, nullptr);
476 477 478 479 480 481 482 483
  }
  else if (platform::is_npu_place(src_place) &&  // NOLINT
           platform::is_npu_place(dst_place)) {  /* npu -> npu*/
    if (src_ptr == dst_ptr) {
      VLOG(3) << "Skip copy the same data sync from " << src_place << " to "
              << dst_place;
      return;
    }
484
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, nullptr);
485 486 487 488 489 490
  }
  else {  // NOLINT
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copy from %s to %s is not supported.", src_place, dst_place));
  }
#endif
491
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
492 493
  else if (platform::is_cuda_pinned_place(src_place) &&  // NOLINT
           platform::is_cuda_pinned_place(dst_place)) {
494
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
495
  }
496
  else if (platform::is_cuda_pinned_place(src_place) &&  // NOLINT
F
fengjiayi 已提交
497
           platform::is_cpu_place(dst_place)) {
498
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
499 500 501
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_cuda_pinned_place(dst_place)) {
502
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
503 504 505
  }
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_cuda_pinned_place(dst_place)) {
506
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, nullptr);
507 508 509
  }
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
510 511
    auto src_gpu_place = src_place;
    auto dst_cpu_place = dst_place;
F
fengjiayi 已提交
512
    memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
513 514 515
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_gpu_place(dst_place)) {
516 517
    auto src_cpu_place = src_place;
    auto dst_gpu_place = dst_place;
F
fengjiayi 已提交
518
    memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr);
519 520 521
  }
  else if (platform::is_gpu_place(src_place) &&  // NOLINT
           platform::is_gpu_place(dst_place)) {
522 523
    auto src_gpu_place = src_place;
    auto dst_gpu_place = dst_place;
F
fengjiayi 已提交
524
    memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr);
525 526 527
  }
  else if (platform::is_cuda_pinned_place(src_place) &&  // NOLINT
           platform::is_gpu_place(dst_place)) {
528 529
    auto src_pinned_place = src_place;
    auto dst_gpu_place = dst_place;
W
Wu Yi 已提交
530 531
    memory::Copy(dst_gpu_place, dst_ptr, src_pinned_place, src_ptr, size,
                 nullptr);
532 533
  }
  else {  // NOLINT
534 535
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copy from %s to %s is not supported.", src_place, dst_place));
F
fengjiayi 已提交
536 537
  }
#endif
538 539 540
#ifdef PADDLE_WITH_MLU
  else if (platform::is_mlu_place(src_place) &&  // NOLINT
           platform::is_cpu_place(dst_place)) {
541
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, nullptr);
542 543 544
  }
  else if (platform::is_cpu_place(src_place) &&  // NOLINT
           platform::is_mlu_place(dst_place)) {
545
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, nullptr);
546 547 548 549 550 551 552 553
  }
  else if (platform::is_mlu_place(src_place) &&  // NOLINT
           platform::is_mlu_place(dst_place)) {
    if (src_ptr == dst_ptr) {
      VLOG(3) << "Skip copy the same data async from " << src_place << " to "
              << dst_place;
      return;
    }
554
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size, nullptr);
555 556 557 558 559 560
  }
  else {  // NOLINT
    PADDLE_THROW(platform::errors::Unimplemented(
        "Copy from %s to %s is not supported.", src_place, dst_place));
  }
#endif
F
fengjiayi 已提交
561 562
}

Y
Yang Yu 已提交
563 564 565 566 567 568 569 570 571 572 573 574
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
575
  void apply() const {
Y
Yang Yu 已提交
576 577
    auto t = EigenVector<T>::Flatten(tensor_);
    auto o = EigenScalar<bool>::From(*out_);
Y
Yang Yu 已提交
578
    // return any of predicate_(t) is true.
Y
Yang Yu 已提交
579 580 581 582 583 584 585
    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) {
586 587 588
  VisitDataType(
      framework::TransToProtoVarType(tensor.dtype()),
      AnyDTypeVisitor<Predicate, DevCtx>(predicate, tensor, ctx, out));
Y
Yang Yu 已提交
589 590 591
}

template <typename Predicate>
592 593
class AnyVisitor : public boost::static_visitor<bool> {
 private:
Y
Yang Yu 已提交
594 595 596
  const framework::Tensor& tensor_;
  Predicate predicate_;

597 598 599 600 601 602 603 604 605 606 607 608 609
  bool GetResultHelper(const framework::Tensor& out,
                       const platform::Place& place) const {
    platform::CPUPlace cpu;
    framework::Tensor tmp;
    tmp.Resize({1});
    tmp.mutable_data<bool>(cpu);
    auto ctx = platform::DeviceContextPool::Instance().Get(place);
    ctx->Wait();
    TensorCopy(out, cpu, *ctx, &tmp);
    ctx->Wait();
    return GetResult(tmp, cpu);
  }

610
 public:
Y
Yang Yu 已提交
611 612 613 614 615 616 617 618 619 620 621 622 623
  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);
  }

624 625 626 627 628
  bool GetResult(const framework::Tensor& out,
                 const platform::XPUPlace& xpu) const {
    return GetResultHelper(out, xpu);
  }

F
fwenguang 已提交
629 630 631 632 633 634 635
  bool GetResult(const framework::Tensor& out,
                 const platform::MLUPlace& mlu) const {
    PADDLE_THROW(
        platform::errors::Unimplemented("Not supported on place (%s) ", mlu));
    return true;
  }

Y
Yang Yu 已提交
636 637
  bool GetResult(const framework::Tensor& out,
                 const platform::CUDAPlace& gpu) const {
638
    return GetResultHelper(out, gpu);
Y
Yang Yu 已提交
639 640
  }

641 642 643 644 645 646
  bool GetResult(const framework::Tensor& out,
                 const platform::NPUPlace& npu) const {
    PADDLE_THROW(
        platform::errors::Unimplemented("Not supported on place (%s) ", npu));
    // return GetResultHelper(out, npu);
  }
J
jianghaicheng 已提交
647 648 649 650 651
  bool GetResult(const framework::Tensor& out,
                 const platform::IPUPlace& ipu) const {
    PADDLE_THROW(
        platform::errors::Unimplemented("Not supported on place (%s) ", ipu));
  }
652

653 654 655 656 657
  bool GetResult(const framework::Tensor& out,
                 const platform::NPUPinnedPlace& cpu) const {
    return *out.data<bool>();
  }

Y
Yang Yu 已提交
658 659 660 661
  bool GetResult(const framework::Tensor& out,
                 const platform::CPUPlace& cpu) const {
    return *out.data<bool>();
  }
C
chengduoZH 已提交
662 663 664 665 666

  bool GetResult(const framework::Tensor& out,
                 const platform::CUDAPinnedPlace& cpu) const {
    return *out.data<bool>();
  }
Y
Yang Yu 已提交
667 668
};

669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689
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
690 691 692 693 694 695 696
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);
}

697 698 699 700 701 702 703 704
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);
}

J
Jack Zhou 已提交
705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726
template <typename Predicate, typename DevCtx>
struct AllDTypeVisitor {
  Predicate predicate_;
  const Tensor& tensor_;
  const DevCtx& ctx_;
  Tensor* out_;

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

  template <typename T>
  void apply() const {
    auto t = EigenVector<T>::Flatten(tensor_);
    auto o = EigenVector<bool>::Flatten(*out_);
    o.device(*ctx_.eigen_device()) = predicate_(t);
  }
};

template <typename Predicate, typename DevCtx>
inline void AllImpl(Predicate predicate, const framework::Tensor& tensor,
                    const DevCtx& ctx, framework::Tensor* out) {
727 728 729
  VisitDataType(
      framework::TransToProtoVarType(tensor.dtype()),
      AllDTypeVisitor<Predicate, DevCtx>(predicate, tensor, ctx, out));
J
Jack Zhou 已提交
730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760
}

template <typename Predicate>
class AllOutVisitor : public boost::static_visitor<> {
 private:
  const framework::Tensor& tensor_;
  mutable framework::Tensor* out_;
  Predicate predicate_;

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

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

template <typename Predicate>
inline void All(const framework::Tensor& tensor, Predicate predicate,
                framework::Tensor* out) {
  AllOutVisitor<Predicate> visitor(tensor, predicate, out);
  auto place = tensor.place();
  platform::VisitPlace(place, visitor);
}

Y
Yi Wang 已提交
761
struct ContainsNANPredicate {
Y
Yang Yu 已提交
762 763 764
  template <typename T>
  auto operator()(const T& eigen_vec) const
      -> decltype(std::declval<T>().isnan()) {
Y
Yang Yu 已提交
765
    // Cast eigen_vector to vector of bool. true if is inf.
Y
Yang Yu 已提交
766 767 768 769
    return eigen_vec.isnan();
  }
};

Y
Yi Wang 已提交
770 771
bool TensorContainsNAN(const framework::Tensor& tensor) {
  ContainsNANPredicate predicate;
Y
Yang Yu 已提交
772 773 774
  return Any(tensor, predicate);
}

775 776 777 778 779 780
void TensorContainsNAN(const framework::Tensor& tensor,
                       framework::Tensor* out) {
  ContainsNANPredicate predicate;
  Any(tensor, predicate, out);
}

J
Jack Zhou 已提交
781 782 783 784 785 786
void TensorContainsNANV2(const framework::Tensor& tensor,
                         framework::Tensor* out) {
  ContainsNANPredicate predicate;
  All(tensor, predicate, out);
}

Y
Yi Wang 已提交
787
struct ContainsInfPredicate {
Y
Yang Yu 已提交
788 789 790
  template <typename T>
  auto operator()(const T& eigen_vec) const
      -> decltype(std::declval<T>().isinf()) {
Y
Yang Yu 已提交
791
    // Cast eigen_vector to vector of bool. true if is inf.
Y
Yang Yu 已提交
792 793 794 795
    return eigen_vec.isinf();
  }
};

Y
Yi Wang 已提交
796 797
bool TensorContainsInf(const framework::Tensor& tensor) {
  ContainsInfPredicate predicate;
Y
Yang Yu 已提交
798 799 800
  return Any(tensor, predicate);
}

801 802 803 804 805 806
void TensorContainsInf(const framework::Tensor& tensor,
                       framework::Tensor* out) {
  ContainsInfPredicate predicate;
  Any(tensor, predicate, out);
}

J
Jack Zhou 已提交
807 808 809 810 811 812
void TensorContainsInfV2(const framework::Tensor& tensor,
                         framework::Tensor* out) {
  ContainsInfPredicate predicate;
  All(tensor, predicate, out);
}

813 814 815 816 817 818 819 820 821 822
// 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);
}

823
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
824
template <typename T>
J
Jack Zhou 已提交
825 826
static inline void __global__ BothFalse(const T* cmp, T* out, int element_num) {
  CUDA_KERNEL_LOOP(i, element_num) { out[i] = (!cmp[i]) && (!out[i]); }
827 828 829 830 831 832 833 834 835 836 837 838 839 840
}
#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);
  }

841 842 843
  void VisitorImpl(const platform::XPUPlace& xpu) const {
    PADDLE_THROW(platform::errors::Unimplemented("XPUPlace is not supported"));
  }
J
jianghaicheng 已提交
844 845 846
  void VisitorImpl(const platform::IPUPlace& ipu) const {
    PADDLE_THROW(platform::errors::Unimplemented("IPUPlace is not supported"));
  }
847

848
  void VisitorImpl(const platform::CUDAPlace& gpu) const {
849
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
850
    auto* ctx = platform::DeviceContextPool::Instance().GetByPlace(gpu);
J
Jack Zhou 已提交
851 852 853 854 855 856 857 858 859 860
    constexpr int MAX_BLOCK_DIM = 512;
    const int MAX_GRID_DIM = ctx->GetMaxPhysicalThreadCount() / MAX_BLOCK_DIM;
    int element_num = in_.numel();
    int block_size = (element_num >= MAX_BLOCK_DIM)
                         ? MAX_BLOCK_DIM
                         : (1 << static_cast<int>(std::log2(element_num)));
    int grid_size = element_num / block_size;
    grid_size = (grid_size >= MAX_GRID_DIM) ? MAX_GRID_DIM : grid_size;
    BothFalse<bool><<<grid_size, block_size, 0, ctx->stream()>>>(
        in_.data<bool>(), out_->mutable_data<bool>(gpu), element_num);
861 862 863
#endif
  }

864 865 866 867
  void VisitorImpl(const platform::NPUPlace& npu) const {
    // TODO(zhiqiu)
  }

F
fwenguang 已提交
868 869 870 871
  void VisitorImpl(const platform::MLUPlace& mlu) const {
    PADDLE_THROW(platform::errors::Unimplemented("MLUPlace is not supported"));
  }

872
  void VisitorImpl(const platform::CPUPlace& cpu) const {
J
Jack Zhou 已提交
873 874 875 876 877 878 879 880
    int num = in_.numel();
    const bool* in_ptr = in_.data<bool>();
    bool* out_ptr = out_->data<bool>();
    for (int i = 0; i < num; ++i) {
      bool lhs = !in_ptr[i];
      bool rhs = !out_ptr[i];
      out_ptr[i] = lhs && rhs;
    }
881 882 883 884
  }

  void VisitorImpl(
      const platform::CUDAPinnedPlace& cpu /* equals to cpu*/) const {
J
Jack Zhou 已提交
885 886 887 888 889 890 891 892
    int num = in_.numel();
    const bool* in_ptr = in_.data<bool>();
    bool* out_ptr = out_->data<bool>();
    for (int i = 0; i < num; ++i) {
      bool lhs = !in_ptr[i];
      bool rhs = !out_ptr[i];
      out_ptr[i] = lhs && rhs;
    }
893
  }
894 895 896 897 898 899 900 901 902 903 904 905

  void VisitorImpl(
      const platform::NPUPinnedPlace& cpu /* equals to cpu*/) const {
    int num = in_.numel();
    const bool* in_ptr = in_.data<bool>();
    bool* out_ptr = out_->data<bool>();
    for (int i = 0; i < num; ++i) {
      bool lhs = !in_ptr[i];
      bool rhs = !out_ptr[i];
      out_ptr[i] = lhs && rhs;
    }
  }
906 907 908 909 910 911 912 913 914 915 916
};

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);
}

J
Jack Zhou 已提交
917 918 919 920 921 922 923 924 925
void TensorIsfiniteV2(const framework::Tensor& tensor, framework::Tensor* out) {
  framework::Tensor tmp;
  TensorContainsInfV2(tensor, &tmp);
  TensorContainsNANV2(tensor, out);
  BothFalseVisitor visitor(tmp, out);
  auto place = tensor.place();
  platform::VisitPlace(place, visitor);
}

Y
Yi Wang 已提交
926 927 928 929 930 931 932 933 934 935
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;
936
    desc.set_data_type(framework::TransToProtoVarType(tensor.dtype()));
Y
Yi Wang 已提交
937 938 939 940 941 942 943 944 945 946
    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
947
    uint64_t size = tensor.numel() * framework::DataTypeSize(tensor.dtype());
Y
yuyang18 已提交
948

949
    auto* data_ptr = tensor.data();
W
wanghuancoder 已提交
950
    PADDLE_ENFORCE_LT(size, (std::numeric_limits<std::streamsize>::max)(),
T
tangwei12 已提交
951 952
                      platform::errors::ResourceExhausted(
                          "tensor size %d overflow when writing tensor", size));
Y
Yi Wang 已提交
953
    if (platform::is_gpu_place(tensor.place())) {
954
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
Y
Yi Wang 已提交
955 956 957 958 959 960 961 962
      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));
963
        memory::Copy(cpu, buf.get(), tensor.place(),
Y
Yi Wang 已提交
964 965 966 967 968 969 970 971
                     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
972 973
      PADDLE_THROW(platform::errors::Unimplemented(
          "CUDAPlace is not supported when not compiled with CUDA"));
974 975 976 977 978 979 980 981 982 983 984
#endif
    } else if (platform::is_xpu_place(tensor.place())) {
#ifdef PADDLE_WITH_XPU
      constexpr size_t kBufSize = 1024 * 1024 * 64;  // 64MB
      std::unique_ptr<char[]> buf(new char[kBufSize]);
      auto& xpu_dev_ctx =
          static_cast<const platform::XPUDeviceContext&>(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));
985
        memory::Copy(cpu, buf.get(), tensor.place(),
986 987 988 989 990 991 992 993 994
                     reinterpret_cast<const void*>(data), size_to_write);
        xpu_dev_ctx.Wait();
        os.write(buf.get(), size_to_write);
        data += size_to_write;
        size -= size_to_write;
      }
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "XPUPlace is not supported when not compiled with XPU"));
995 996 997 998 999 1000 1001 1002 1003 1004 1005
#endif
    } else if (platform::is_mlu_place(tensor.place())) {
#ifdef PADDLE_WITH_MLU
      constexpr size_t kBufSize = 1024 * 1024 * 64;  // 64MB
      std::unique_ptr<char[]> buf(new char[kBufSize]);
      auto& mlu_dev_ctx =
          static_cast<const platform::MLUDeviceContext&>(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));
1006
        memory::Copy(cpu, buf.get(), tensor.place(),
1007 1008 1009 1010 1011 1012 1013 1014 1015 1016
                     reinterpret_cast<const void*>(data), size_to_write,
                     mlu_dev_ctx.stream());
        mlu_dev_ctx.Wait();
        os.write(buf.get(), size_to_write);
        data += size_to_write;
        size -= size_to_write;
      }
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "MLUPlace is not supported when not compiled with MLU"));
1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027
#endif
    } else if (platform::is_npu_place(tensor.place())) {
#ifdef PADDLE_WITH_ASCEND_CL
      constexpr size_t kBufSize = 1024 * 1024 * 64;  // 64MB
      std::unique_ptr<char[]> buf(new char[kBufSize]);
      auto& npu_dev_ctx =
          static_cast<const platform::NPUDeviceContext&>(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));
1028
        memory::Copy(cpu, buf.get(), tensor.place(),
1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
                     reinterpret_cast<const void*>(data), size_to_write,
                     npu_dev_ctx.stream());
        npu_dev_ctx.Wait();
        os.write(buf.get(), size_to_write);
        data += size_to_write;
        size -= size_to_write;
      }
#else
      PADDLE_THROW(platform::errors::Unimplemented(
          "NPUPlace is not supported when not compiled with NPU"));
Y
Yi Wang 已提交
1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052
#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
1053
  void apply() {
Y
Yi Wang 已提交
1054 1055 1056 1057 1058 1059 1060 1061
    *buf_ = tensor_->mutable_data<T>(place_);
  }

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

T
tangwei12 已提交
1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
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;
W
Wilber 已提交
1091
    platform::CPUDeviceContext ctx;
T
tangwei12 已提交
1092
    size_t size = tensor->numel() * framework::SizeOfType(desc.data_type());
1093
    if (platform::is_gpu_place(dev_ctx.GetPlace()) ||
1094
        platform::is_xpu_place(dev_ctx.GetPlace()) ||
1095
        platform::is_mlu_place(dev_ctx.GetPlace()) ||
1096
        platform::is_npu_place(dev_ctx.GetPlace())) {
1097
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
1098 1099
    defined(PADDLE_WITH_XPU) || defined(PADDLE_WITH_MLU) ||  \
    defined(PADDLE_WITH_ASCEND_CL)
T
tangwei12 已提交
1100 1101 1102 1103 1104 1105 1106 1107
      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);
1108 1109 1110
      if (platform::is_npu_place(dev_ctx.GetPlace())) {
        dev_ctx.Wait();
      }
T
tangwei12 已提交
1111
#else
1112 1113 1114
      if (platform::is_gpu_place(dev_ctx.GetPlace())) {
        PADDLE_THROW(platform::errors::Unimplemented(
            "CUDAPlace is not supported when not compiled with CUDA"));
1115
      } else if (platform::is_xpu_place(dev_ctx.GetPlace())) {
1116 1117
        PADDLE_THROW(platform::errors::Unimplemented(
            "XPUPlace is not supported when not compiled with XPU"));
1118 1119 1120
      } else if (platform::is_mlu_place(dev_ctx.GetPlace())) {
        PADDLE_THROW(platform::errors::Unimplemented(
            "MLUPlace is not supported when not compiled with MLU"));
1121 1122 1123
      } else {
        PADDLE_THROW(platform::errors::Unimplemented(
            "NPUPlace is not supported when not compiled with NPU"));
1124
      }
T
tangwei12 已提交
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134
#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
1135 1136 1137 1138
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
1139 1140 1141 1142 1143
  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
1144 1145 1146 1147 1148 1149 1150
  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
1151 1152 1153
    PADDLE_ENFORCE_EQ(
        desc.ParseFromArray(buf.get(), size), true,
        platform::errors::InvalidArgument("Cannot parse tensor desc"));
Y
Yi Wang 已提交
1154 1155 1156 1157 1158 1159 1160
  }
  {  // 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;
W
Wilber 已提交
1161
    platform::CPUDeviceContext ctx;
Y
Yu Yang 已提交
1162
    size_t size = tensor->numel() * framework::SizeOfType(desc.data_type());
1163
    if (platform::is_gpu_place(dev_ctx.GetPlace()) ||
1164
        platform::is_xpu_place(dev_ctx.GetPlace()) ||
1165
        platform::is_mlu_place(dev_ctx.GetPlace()) ||
1166
        platform::is_npu_place(dev_ctx.GetPlace())) {
1167
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP) || \
1168 1169
    defined(PADDLE_WITH_XPU) || defined(PADDLE_WITH_MLU) ||  \
    defined(PADDLE_WITH_ASCEND_CL)
Y
Yi Wang 已提交
1170 1171 1172 1173 1174
      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
1175
      is.read(static_cast<char*>(buf), size);
Y
Yi Wang 已提交
1176 1177
      auto dst_place = dev_ctx.GetPlace();
      framework::TensorCopy(cpu_tensor, dst_place, dev_ctx, tensor);
1178 1179 1180
      if (platform::is_npu_place(dev_ctx.GetPlace())) {
        dev_ctx.Wait();
      }
Y
Yi Wang 已提交
1181
#else
1182 1183 1184
      if (platform::is_gpu_place(dev_ctx.GetPlace())) {
        PADDLE_THROW(platform::errors::Unimplemented(
            "CUDAPlace is not supported when not compiled with CUDA"));
1185
      } else if (platform::is_xpu_place(dev_ctx.GetPlace())) {
1186 1187
        PADDLE_THROW(platform::errors::Unimplemented(
            "XPUPlace is not supported when not compiled with XPU"));
1188 1189 1190
      } else if (platform::is_mlu_place(dev_ctx.GetPlace())) {
        PADDLE_THROW(platform::errors::Unimplemented(
            "MLUPlace is not supported when not compiled with MLU"));
1191 1192 1193
      } else {
        PADDLE_THROW(platform::errors::Unimplemented(
            "NPUPlace is not supported when not compiled with NPU"));
1194
      }
Y
Yi Wang 已提交
1195 1196 1197 1198 1199
#endif
    } else {
      framework::VisitDataType(
          desc.data_type(),
          DeserializedDataFunctor(&buf, tensor, ctx.GetPlace()));
Y
yuyang18 已提交
1200
      is.read(static_cast<char*>(buf), size);
Y
Yi Wang 已提交
1201 1202 1203 1204
    }
  }
}

6
633WHU 已提交
1205 1206 1207 1208
// get tensor data point by DLDataType
void* GetDstPtrByDLDataType(DLDataType type, framework::Tensor* dst,
                            const platform::Place& dst_place) {
  // vector types not currently supported
1209 1210 1211
  PADDLE_ENFORCE_LE(type.lanes, 1,
                    platform::errors::Unimplemented(
                        "Vector type is not supported currently."));
6
633WHU 已提交
1212 1213 1214 1215 1216 1217 1218

  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));
1219 1220 1221
      PADDLE_THROW(platform::errors::Unimplemented(
          "DLDataType code <%d> is illegal when DLDataType.bits is <%d>.",
          type.code, type.bits));
6
633WHU 已提交
1222 1223 1224 1225 1226 1227
    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));
S
Siming Dai 已提交
1228 1229 1230
      if (type.code == kDLBfloat)
        return static_cast<void*>(
            dst->mutable_data<paddle::platform::bfloat16>(dst_place));
1231 1232 1233
      PADDLE_THROW(platform::errors::Unimplemented(
          "DLDataType code <%d> is illegal when DLDataType.bits is <%d>.",
          type.code, type.bits));
6
633WHU 已提交
1234 1235 1236 1237 1238
    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));
1239 1240 1241
      PADDLE_THROW(platform::errors::Unimplemented(
          "DLDataType code <%d> is illegal when DLDataType.bits is <%d>.",
          type.code, type.bits));
6
633WHU 已提交
1242 1243 1244 1245 1246
    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));
S
Siming Dai 已提交
1247 1248 1249 1250 1251 1252 1253 1254 1255 1256
      if (type.code == kDLComplex)
        return static_cast<void*>(
            dst->mutable_data<paddle::platform::complex<float>>(dst_place));
      PADDLE_THROW(platform::errors::Unimplemented(
          "DLDataType code <%d> is illegal when DLDataType.bits is <%d>.",
          type.code, type.bits));
    case 128:
      if (type.code == kDLComplex)
        return static_cast<void*>(
            dst->mutable_data<paddle::platform::complex<double>>(dst_place));
1257 1258 1259
      PADDLE_THROW(platform::errors::Unimplemented(
          "DLDataType code <%d> is illegal when DLDataType.bits is <%d>.",
          type.code, type.bits));
6
633WHU 已提交
1260
    default:
1261 1262
      PADDLE_THROW(platform::errors::Unimplemented(
          "Unsupported DLDataType.bits %d.", type.bits));
6
633WHU 已提交
1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282
  }
}

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;

S
Siming Dai 已提交
1283
  if (dl_tensor.device.device_type == kDLCPU) {
6
633WHU 已提交
1284 1285
    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
  }
1286
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
S
Siming Dai 已提交
1287
  if (dl_tensor.device.device_type == kDLGPU) {
6
633WHU 已提交
1288
    platform::CUDAPlace dst_place =
S
Siming Dai 已提交
1289
        platform::CUDAPlace(dl_tensor.device.device_id);
6
633WHU 已提交
1290
    platform::CUDAPlace src_place =
S
Siming Dai 已提交
1291
        platform::CUDAPlace(dl_tensor.device.device_id);
6
633WHU 已提交
1292 1293 1294 1295 1296 1297 1298
    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
1299 1300 1301
#ifdef PADDLE_WITH_XPU
  PADDLE_THROW(platform::errors::Unimplemented("XPUPlace is not supported"));
#endif
6
633WHU 已提交
1302 1303
}

1304 1305 1306 1307 1308 1309
template <typename T>
std::string format_tensor(const framework::Tensor& tensor) {
  // TODO(zhiqiu): use the print option to format tensor.
  return "NOT IMPLEMENTED";
}

1310 1311 1312 1313 1314
template <typename T>
std::ostream& print_tensor(std::ostream& os, const framework::Tensor& tensor) {
  auto inspect = tensor.data<T>();
  auto element_num = tensor.numel();

1315
  os << "  - data: [";
1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329
  // Note: int8_t && uint8_t is typedf of char, ostream unable to print properly
  if (typeid(int8_t) == typeid(T) || typeid(uint8_t) == typeid(T)) {
    if (element_num > 0) {
      os << signed(inspect[0]);
      for (int j = 1; j < element_num; ++j) {
        os << " " << signed(inspect[j]);
      }
    }
  } else {
    if (element_num > 0) {
      os << inspect[0];
      for (int j = 1; j < element_num; ++j) {
        os << " " << inspect[j];
      }
1330 1331 1332 1333 1334 1335
    }
  }
  os << "]";
  return os;
}

1336
template <>
1337
std::ostream& print_tensor<paddle::platform::complex<float>>(
1338
    std::ostream& os, const framework::Tensor& tensor) {
1339
  auto inspect = tensor.data<paddle::platform::complex<float>>();
1340 1341 1342 1343
  auto element_num = tensor.numel();

  os << "  - data: [";
  if (element_num > 0) {
1344
    os << signed(inspect[0].real) << "+" << signed(inspect[0].imag) << "j";
1345
    for (int j = 1; j < element_num; ++j) {
1346 1347
      os << " " << signed(inspect[j].real) << "+" << signed(inspect[j].imag)
         << "j";
1348 1349 1350 1351 1352 1353 1354
    }
  }
  os << "]";
  return os;
}

template <>
1355
std::ostream& print_tensor<paddle::platform::complex<double>>(
1356
    std::ostream& os, const framework::Tensor& tensor) {
1357
  auto inspect = tensor.data<paddle::platform::complex<double>>();
1358 1359 1360 1361
  auto element_num = tensor.numel();

  os << "  - data: [";
  if (element_num > 0) {
1362
    os << signed(inspect[0].real) << "+" << signed(inspect[0].imag) << "j";
1363
    for (int j = 1; j < element_num; ++j) {
1364 1365
      os << " " << signed(inspect[j].real) << "+" << signed(inspect[j].imag)
         << "j";
1366 1367 1368 1369 1370 1371
    }
  }
  os << "]";
  return os;
}

1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391
std::ostream& operator<<(std::ostream& os, const LoD& lod) {
  os << "{";
  for (auto& v : lod) {
    os << "{";
    bool is_first = true;
    for (auto& i : v) {
      if (is_first) {
        os << i;
        is_first = false;
      } else {
        os << ", " << i;
      }
    }
    os << "}";
  }
  os << "}";

  return os;
}

1392 1393 1394 1395 1396 1397
}  // namespace framework
}  // namespace paddle

namespace pten {

std::ostream& operator<<(std::ostream& os, const pten::DenseTensor& t) {
1398 1399 1400 1401
  if (t.lod().size() > 0) {
    os << "  - lod: " << t.lod() << "\n";
  }

1402 1403
  os << "  - place: " << t.place() << "\n";
  os << "  - shape: [" << t.dims() << "]\n";
1404 1405
  os << "  - layout: " << paddle::framework::DataLayoutToString(t.layout())
     << "\n";
1406

1407 1408 1409 1410 1411
#ifdef PADDLE_WITH_MKLDNN
  os << "  - format: "
     << dnnl_fmt_tag2str(static_cast<dnnl_format_tag_t>(t.format())) << "\n";
#endif

1412
  DenseTensor tensor;
1413
  tensor.Resize(t.dims());
1414
  if (paddle::platform::is_cpu_place(t.place())) {
1415 1416
    tensor.ShareDataWith(t);
  } else {
1417 1418 1419 1420
    paddle::platform::CPUPlace place;
    paddle::framework::TensorCopy(t, place, &tensor);
    paddle::platform::DeviceContextPool& pool =
        paddle::platform::DeviceContextPool::Instance();
1421 1422 1423 1424
    auto& dev_ctx = *pool.Get(t.place());
    dev_ctx.Wait();
  }

1425 1426 1427 1428 1429 1430 1431 1432
#define PrintTensorCallback(cpp_type, proto_type)                 \
  do {                                                            \
    if (paddle::framework::TransToProtoVarType(tensor.dtype()) == \
        proto_type) {                                             \
      os << "  - dtype: " << proto_type << "\n";                  \
      paddle::framework::print_tensor<cpp_type>(os, tensor);      \
      return os;                                                  \
    }                                                             \
1433 1434 1435 1436 1437 1438
  } while (0)

  _ForEachDataType_(PrintTensorCallback);
  VLOG(1) << "PrintVar: unrecognized data type:" << t.type();
  return os;
}
W
Wilber 已提交
1439
}  // namespace pten