gpu_info.cc 4.1 KB
Newer Older
L
liaogang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/platform/gpu_info.h"
L
liaogang 已提交
16

L
liaogang 已提交
17
#include "gflags/gflags.h"
L
liaogang 已提交
18

L
liaogang 已提交
19
#include "paddle/platform/enforce.h"
L
liaogang 已提交
20

21 22
DEFINE_double(fraction_of_gpu_memory_to_use, 0.92,
              "Default use 92% of GPU memory for PaddlePaddle,"
L
liaogang 已提交
23 24 25 26 27
              "reserve the rest for page tables, etc");

namespace paddle {
namespace platform {

28
int GetCUDADeviceCount() {
L
liaogang 已提交
29
  int count;
L
liaogang 已提交
30
  PADDLE_ENFORCE(
L
liaogang 已提交
31
      cudaGetDeviceCount(&count),
32
      "cudaGetDeviceCount failed in paddle::platform::GetCUDADeviceCount");
L
liaogang 已提交
33 34 35
  return count;
}

L
liaogang 已提交
36 37
int GetCurrentDeviceId() {
  int device_id;
L
liaogang 已提交
38
  PADDLE_ENFORCE(
L
liaogang 已提交
39 40 41 42 43 44
      cudaGetDevice(&device_id),
      "cudaGetDevice failed in paddle::platform::GetCurrentDeviceId");
  return device_id;
}

void SetDeviceId(int id) {
Q
qijun 已提交
45
  // TODO(qijun): find a better way to cache the cuda device count
Y
Yang Yang 已提交
46
  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
L
liaogang 已提交
47
  PADDLE_ENFORCE(cudaSetDevice(id),
L
liaogang 已提交
48 49 50
                 "cudaSetDevice failed in paddle::platform::SetDeviceId");
}

L
liaogang 已提交
51
void GpuMemoryUsage(size_t &available, size_t &total) {
L
liaogang 已提交
52
  PADDLE_ENFORCE(cudaMemGetInfo(&available, &total),
L
liaogang 已提交
53 54 55 56 57 58 59 60 61
                 "cudaMemGetInfo failed in paddle::platform::GetMemoryUsage");
}

size_t GpuMaxAllocSize() {
  size_t total = 0;
  size_t available = 0;

  GpuMemoryUsage(available, total);

L
liaogang 已提交
62
  // Reserve the rest for page tables, etc.
L
liaogang 已提交
63
  return static_cast<size_t>(total * FLAGS_fraction_of_gpu_memory_to_use);
L
liaogang 已提交
64 65
}

L
liaogang 已提交
66 67 68 69 70 71 72 73 74 75
size_t GpuMinChunkSize() {
  // Allow to allocate the minimum chunk size is 256 bytes.
  return 1 << 8;
}

size_t GpuMaxChunkSize() {
  size_t total = 0;
  size_t available = 0;

  GpuMemoryUsage(available, total);
76 77 78
  VLOG(10) << "GPU Usage " << available / 1024 / 1024 << "M/"
           << total / 1024 / 1024 << "M";
  size_t reserving = static_cast<size_t>(0.05 * total);
L
liaogang 已提交
79
  // If available less than minimum chunk size, no usable memory exists.
80
  available =
81 82 83 84
      std::min(std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(),
               total - reserving);

  // Reserving the rest memory for page tables, etc.
L
liaogang 已提交
85

86 87
  size_t allocating = static_cast<size_t>(FLAGS_fraction_of_gpu_memory_to_use *
                                          (total - reserving));
L
liaogang 已提交
88

89
  PADDLE_ENFORCE_LE(allocating, available);
90 91

  return allocating;
L
liaogang 已提交
92 93
}

L
liaogang 已提交
94 95
void GpuMemcpyAsync(void *dst, const void *src, size_t count,
                    enum cudaMemcpyKind kind, cudaStream_t stream) {
L
liaogang 已提交
96 97
  PADDLE_ENFORCE(cudaMemcpyAsync(dst, src, count, kind, stream),
                 "cudaMemcpyAsync failed in paddle::platform::GpuMemcpyAsync");
L
liaogang 已提交
98 99 100 101
}

void GpuMemcpySync(void *dst, const void *src, size_t count,
                   enum cudaMemcpyKind kind) {
L
liaogang 已提交
102 103
  PADDLE_ENFORCE(cudaMemcpy(dst, src, count, kind),
                 "cudaMemcpy failed in paddle::platform::GpuMemcpySync");
L
liaogang 已提交
104 105
  // note: cudaMemcpy may actually be asynchronous with respect to the caller,
  //       block on stream 0 to make sure the copy has completed
L
liaogang 已提交
106 107 108
  PADDLE_ENFORCE(
      cudaStreamSynchronize(0),
      "cudaStreamSynchronize failed in paddle::platform::GpuMemcpySync");
L
liaogang 已提交
109 110 111 112 113
}

void GpuMemcpyPeer(void *dst, int dst_device, const void *src, int src_device,
                   size_t count, cudaStream_t stream) {
  PADDLE_ENFORCE(
L
liaogang 已提交
114 115
      cudaMemcpyPeerAsync(dst, dst_device, src, src_device, count, stream),
      "cudaMemcpyPeerAsync failed in paddle::platform::GpuMemcpyPeer");
L
liaogang 已提交
116
}
D
dzhwinter 已提交
117 118 119 120 121

void GpuMemsetAsync(void *dst, int value, size_t count, cudaStream_t stream) {
  PADDLE_ENFORCE(cudaMemsetAsync(dst, value, count, stream),
                 "cudaMemsetAsync failed in paddle::platform::GpuMemsetAsync");
}
L
liaogang 已提交
122 123
}  // namespace platform
}  // namespace paddle