gpu_info.cc 3.8 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
#include "paddle/platform/environment.h"
L
liaogang 已提交
21 22 23 24 25 26 27 28

DEFINE_double(fraction_of_gpu_memory_to_use, 0.95,
              "Default use 95% of GPU memory for PaddlePaddle,"
              "reserve the rest for page tables, etc");

namespace paddle {
namespace platform {

L
liaogang 已提交
29
int GetDeviceCount() {
L
liaogang 已提交
30
  int count;
L
liaogang 已提交
31
  PADDLE_ENFORCE(
L
liaogang 已提交
32
      cudaGetDeviceCount(&count),
L
liaogang 已提交
33
      "cudaGetDeviceCount failed in paddle::platform::GetDeviceCount");
L
liaogang 已提交
34 35 36
  return count;
}

L
liaogang 已提交
37 38
int GetCurrentDeviceId() {
  int device_id;
L
liaogang 已提交
39
  PADDLE_ENFORCE(
L
liaogang 已提交
40 41 42 43 44 45
      cudaGetDevice(&device_id),
      "cudaGetDevice failed in paddle::platform::GetCurrentDeviceId");
  return device_id;
}

void SetDeviceId(int id) {
L
liaogang 已提交
46
  PADDLE_ENFORCE(cudaSetDevice(id),
L
liaogang 已提交
47 48 49
                 "cudaSetDevice failed in paddle::platform::SetDeviceId");
}

L
liaogang 已提交
50
void GpuMemoryUsage(size_t &available, size_t &total) {
L
liaogang 已提交
51
  PADDLE_ENFORCE(cudaMemGetInfo(&available, &total),
L
liaogang 已提交
52 53 54 55 56 57 58 59 60
                 "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
61
  // Reserve the rest for page tables, etc.
L
liaogang 已提交
62
  return static_cast<size_t>(total * FLAGS_fraction_of_gpu_memory_to_use);
L
liaogang 已提交
63 64
}

L
liaogang 已提交
65 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);

L
liaogang 已提交
76 77 78 79 80 81 82
  if (IsEnvVarDefined(kEnvFractionGpuMemoryToUse)) {
    auto val = std::stod(GetEnvValue(kEnvFractionGpuMemoryToUse));
    PADDLE_ENFORCE_GT(val, 0.0);
    PADDLE_ENFORCE_LE(val, 1.0);
    FLAGS_fraction_of_gpu_memory_to_use = val;
  }

L
liaogang 已提交
83 84 85 86 87 88 89 90 91 92 93 94
  // Reserving the rest memory for page tables, etc.
  size_t reserving = (1 - FLAGS_fraction_of_gpu_memory_to_use) * total;

  // If available less than minimum chunk size, no usable memory exists.
  available = std::max(available, GpuMinChunkSize()) - GpuMinChunkSize();

  // If available less than reserving, no usable memory exists.
  size_t usable = std::max(available, reserving) - reserving;

  return usable;
}

L
liaogang 已提交
95 96
void GpuMemcpyAsync(void *dst, const void *src, size_t count,
                    enum cudaMemcpyKind kind, cudaStream_t stream) {
L
liaogang 已提交
97 98
  PADDLE_ENFORCE(cudaMemcpyAsync(dst, src, count, kind, stream),
                 "cudaMemcpyAsync failed in paddle::platform::GpuMemcpyAsync");
L
liaogang 已提交
99 100 101 102
}

void GpuMemcpySync(void *dst, const void *src, size_t count,
                   enum cudaMemcpyKind kind) {
L
liaogang 已提交
103 104
  PADDLE_ENFORCE(cudaMemcpy(dst, src, count, kind),
                 "cudaMemcpy failed in paddle::platform::GpuMemcpySync");
L
liaogang 已提交
105 106
  // 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
107 108 109
  PADDLE_ENFORCE(
      cudaStreamSynchronize(0),
      "cudaStreamSynchronize failed in paddle::platform::GpuMemcpySync");
L
liaogang 已提交
110 111 112 113 114
}

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
115 116
      cudaMemcpyPeerAsync(dst, dst_device, src, src_device, count, stream),
      "cudaMemcpyPeerAsync failed in paddle::platform::GpuMemcpyPeer");
L
liaogang 已提交
117
}
L
liaogang 已提交
118 119
}  // namespace platform
}  // namespace paddle