gpu_info.cc 11.1 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
L
liaogang 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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
15
#include "paddle/fluid/platform/gpu_info.h"
16
#include <algorithm>
S
sneaxiy 已提交
17 18
#include <cstdlib>
#include <string>
L
liaogang 已提交
19

20
#include "gflags/gflags.h"
Y
Yi Wang 已提交
21
#include "paddle/fluid/platform/enforce.h"
22
#include "paddle/fluid/string/split.h"
L
liaogang 已提交
23

24
#ifndef _WIN32
P
peizhilin 已提交
25
constexpr static float fraction_of_gpu_memory_to_use = 0.92f;
26
#else
P
peizhilin 已提交
27 28 29
// fraction_of_gpu_memory_to_use cannot be too high on windows,
// since the win32 graphic sub-system can occupy some GPU memory
// which may lead to insufficient memory left for paddle
P
peizhilin 已提交
30
constexpr static float fraction_of_gpu_memory_to_use = 0.5f;
31 32
#endif

Z
zhhsplendid 已提交
33 34
constexpr static float fraction_reserve_gpu_memory = 0.05f;

35
DEFINE_double(fraction_of_gpu_memory_to_use, fraction_of_gpu_memory_to_use,
X
Xin Pan 已提交
36 37 38 39 40
              "Allocate a trunk of gpu memory that is this fraction of the "
              "total gpu memory size. Future memory usage will be allocated "
              "from the trunk. If the trunk doesn't have enough gpu memory, "
              "additional trunks of the same size will be requested from gpu "
              "until the gpu has no memory left for another trunk.");
L
liaogang 已提交
41

42 43 44 45
DEFINE_uint64(
    initial_gpu_memory_in_mb, 0ul,
    "Allocate a trunk of gpu memory whose byte size is specified by "
    "the flag. Future memory usage will be allocated from the "
46
    "trunk. If the trunk doesn't have enough gpu memory, additional "
47 48 49 50 51 52 53 54 55
    "trunks of the gpu memory will be requested from gpu with size "
    "specified by FLAGS_reallocate_gpu_memory_in_mb until the gpu has "
    "no memory left for the additional trunk. Note: if you set this "
    "flag, the memory size set by "
    "FLAGS_fraction_of_gpu_memory_to_use will be overrided by this "
    "flag. If you don't set this flag, PaddlePaddle will use "
    "FLAGS_fraction_of_gpu_memory_to_use to allocate gpu memory");

DEFINE_uint64(reallocate_gpu_memory_in_mb, 0ul,
Z
zhhsplendid 已提交
56 57 58 59
              "If this flag is set, Paddle will reallocate the gpu memory with "
              "size specified by this flag. Else Paddle will reallocate by "
              "FLAGS_fraction_of_gpu_memory_to_use");

60 61 62 63 64 65 66 67 68 69
DEFINE_bool(
    enable_cublas_tensor_op_math, false,
    "The enable_cublas_tensor_op_math indicate whether to use Tensor Core, "
    "but it may loss precision. Currently, There are two CUDA libraries that"
    " use Tensor Cores, cuBLAS and cuDNN. cuBLAS uses Tensor Cores to speed up"
    " GEMM computations(the matrices must be either half precision or single "
    "precision); cuDNN uses Tensor Cores to speed up both convolutions(the "
    "input and output must be half precision) and recurrent neural networks "
    "(RNNs).");

70 71 72 73 74 75 76 77 78
DEFINE_string(selected_gpus, "",
              "A list of device ids separated by comma, like: 0,1,2,3. "
              "This option is useful when doing multi process training and "
              "each process have only one device (GPU). If you want to use "
              "all visible devices, set this to empty string. NOTE: the "
              "reason of doing this is that we want to use P2P communication"
              "between GPU devices, use CUDA_VISIBLE_DEVICES can only use"
              "share-memory only.");

L
liaogang 已提交
79 80 81
namespace paddle {
namespace platform {

S
sneaxiy 已提交
82 83 84 85 86 87 88
static int GetCUDADeviceCountImpl() {
  const auto *cuda_visible_devices = std::getenv("CUDA_VISIBLE_DEVICES");
  if (cuda_visible_devices != nullptr) {
    std::string cuda_visible_devices_str(cuda_visible_devices);
    if (std::all_of(cuda_visible_devices_str.begin(),
                    cuda_visible_devices_str.end(),
                    [](char ch) { return ch == ' '; })) {
S
sneaxiy 已提交
89
      VLOG(2) << "CUDA_VISIBLE_DEVICES is set to be empty. No GPU detected.";
S
sneaxiy 已提交
90 91 92 93
      return 0;
    }
  }

L
liaogang 已提交
94
  int count;
L
liaogang 已提交
95
  PADDLE_ENFORCE(
L
liaogang 已提交
96
      cudaGetDeviceCount(&count),
97
      "cudaGetDeviceCount failed in paddle::platform::GetCUDADeviceCount");
L
liaogang 已提交
98 99 100
  return count;
}

S
sneaxiy 已提交
101 102 103 104 105
int GetCUDADeviceCount() {
  static auto dev_cnt = GetCUDADeviceCountImpl();
  return dev_cnt;
}

106 107 108 109 110 111 112 113 114
int GetCUDAComputeCapability(int id) {
  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
  cudaDeviceProp device_prop;
  PADDLE_ENFORCE(cudaGetDeviceProperties(&device_prop, id),
                 "cudaGetDeviceProperties failed in "
                 "paddle::platform::GetCUDAComputeCapability");
  return device_prop.major * 10 + device_prop.minor;
}

C
chengduo 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
int GetCUDARuntimeVersion(int id) {
  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
  int runtime_version = 0;
  PADDLE_ENFORCE(cudaRuntimeGetVersion(&runtime_version),
                 "cudaRuntimeGetVersion failed in "
                 "paddle::platform::cudaRuntimeGetVersion");
  return runtime_version;
}

int GetCUDADriverVersion(int id) {
  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
  int driver_version = 0;
  PADDLE_ENFORCE(cudaDriverGetVersion(&driver_version),
                 "cudaDriverGetVersion failed in "
                 "paddle::platform::GetCUDADriverVersion");
  return driver_version;
}

133 134 135 136 137 138 139 140 141 142
bool TensorCoreAvailable() {
#if CUDA_VERSION >= 9000
  int device = GetCurrentDeviceId();
  int driver_version = GetCUDAComputeCapability(device);
  return driver_version >= 70;
#else
  return false;
#endif
}

C
chengduoZH 已提交
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162
int GetCUDAMultiProcessors(int id) {
  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
  int count;
  PADDLE_ENFORCE(
      cudaDeviceGetAttribute(&count, cudaDevAttrMultiProcessorCount, id),
      "cudaDeviceGetAttribute failed in "
      "paddle::platform::GetCUDAMultiProcessors");
  return count;
}

int GetCUDAMaxThreadsPerMultiProcessor(int id) {
  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
  int count;
  PADDLE_ENFORCE(cudaDeviceGetAttribute(
                     &count, cudaDevAttrMaxThreadsPerMultiProcessor, id),
                 "cudaDeviceGetAttribute failed in "
                 "paddle::platform::GetCUDAMaxThreadsPerMultiProcessor");
  return count;
}

L
liaogang 已提交
163 164
int GetCurrentDeviceId() {
  int device_id;
L
liaogang 已提交
165
  PADDLE_ENFORCE(
L
liaogang 已提交
166 167 168 169 170
      cudaGetDevice(&device_id),
      "cudaGetDevice failed in paddle::platform::GetCurrentDeviceId");
  return device_id;
}

171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188
//! Get a list of device ids from environment variable or use all.
std::vector<int> GetSelectedDevices() {
  // use user specified GPUs in single-node multi-process mode.
  std::vector<int> devices;
  if (!FLAGS_selected_gpus.empty()) {
    auto devices_str = paddle::string::Split(FLAGS_selected_gpus, ',');
    for (auto id : devices_str) {
      devices.push_back(atoi(id.c_str()));
    }
  } else {
    int count = GetCUDADeviceCount();
    for (int i = 0; i < count; ++i) {
      devices.push_back(i);
    }
  }
  return devices;
}

L
liaogang 已提交
189
void SetDeviceId(int id) {
Q
qijun 已提交
190
  // TODO(qijun): find a better way to cache the cuda device count
Y
Yang Yang 已提交
191
  PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count");
L
liaogang 已提交
192
  PADDLE_ENFORCE(cudaSetDevice(id),
L
liaogang 已提交
193 194 195
                 "cudaSetDevice failed in paddle::platform::SetDeviceId");
}

196 197
void GpuMemoryUsage(size_t *available, size_t *total) {
  PADDLE_ENFORCE(cudaMemGetInfo(available, total),
L
liaogang 已提交
198 199 200 201
                 "cudaMemGetInfo failed in paddle::platform::GetMemoryUsage");
}

size_t GpuMaxAllocSize() {
Z
zhhsplendid 已提交
202 203 204 205
  return std::max(GpuInitAllocSize(), GpuReallocSize());
}

size_t GpuInitAllocSize() {
206 207 208
  if (FLAGS_initial_gpu_memory_in_mb > 0ul) {
    // Initial memory will be allocated by FLAGS_initial_gpu_memory_in_mb
    return static_cast<size_t>(FLAGS_initial_gpu_memory_in_mb << 20);
Z
zhhsplendid 已提交
209 210
  }

211
  // FLAGS_initial_gpu_memory_in_mb is 0, initial memory will be allocated by
Z
zhhsplendid 已提交
212
  // fraction
L
liaogang 已提交
213 214 215
  size_t total = 0;
  size_t available = 0;

216
  GpuMemoryUsage(&available, &total);
Z
zhhsplendid 已提交
217
  size_t reserving = static_cast<size_t>(fraction_reserve_gpu_memory * total);
L
liaogang 已提交
218

Z
zhhsplendid 已提交
219 220 221 222 223
  return static_cast<size_t>((total - reserving) *
                             FLAGS_fraction_of_gpu_memory_to_use);
}

size_t GpuReallocSize() {
224 225 226
  if (FLAGS_reallocate_gpu_memory_in_mb > 0ul) {
    // Additional memory will be allocated by FLAGS_reallocate_gpu_memory_in_mb
    return static_cast<size_t>(FLAGS_reallocate_gpu_memory_in_mb << 20);
Z
zhhsplendid 已提交
227 228
  }

229
  // FLAGS_reallocate_gpu_memory_in_mb is 0, additional memory will be allocated
Z
zhhsplendid 已提交
230 231 232 233 234 235 236 237 238
  // by fraction
  size_t total = 0;
  size_t available = 0;

  GpuMemoryUsage(&available, &total);
  size_t reserving = static_cast<size_t>(fraction_reserve_gpu_memory * total);

  return static_cast<size_t>((total - reserving) *
                             FLAGS_fraction_of_gpu_memory_to_use);
L
liaogang 已提交
239 240
}

L
liaogang 已提交
241 242 243 244 245 246 247
size_t GpuMinChunkSize() {
  // Allow to allocate the minimum chunk size is 256 bytes.
  return 1 << 8;
}

size_t GpuMaxChunkSize() {
  size_t total = 0;
C
chenweihang 已提交
248
  size_t available = 0;
L
liaogang 已提交
249

C
chenweihang 已提交
250
  GpuMemoryUsage(&available, &total);
M
minqiyang 已提交
251 252
  VLOG(10) << "GPU Usage " << available / 1024 / 1024 << "M/"
           << total / 1024 / 1024 << "M";
Z
zhhsplendid 已提交
253
  size_t reserving = static_cast<size_t>(fraction_reserve_gpu_memory * total);
L
liaogang 已提交
254
  // If available less than minimum chunk size, no usable memory exists.
C
chenweihang 已提交
255 256 257
  available =
      std::min(std::max(available, GpuMinChunkSize()) - GpuMinChunkSize(),
               total - reserving);
258

Z
zhhsplendid 已提交
259
  size_t allocating = GpuMaxAllocSize();
L
liaogang 已提交
260

C
chenweihang 已提交
261 262
  PADDLE_ENFORCE_LE(allocating, available,
                    "Insufficient GPU memory to allocation.");
263

C
chenweihang 已提交
264
  return allocating;
L
liaogang 已提交
265 266
}

L
liaogang 已提交
267 268
void GpuMemcpyAsync(void *dst, const void *src, size_t count,
                    enum cudaMemcpyKind kind, cudaStream_t stream) {
L
liaogang 已提交
269
  PADDLE_ENFORCE(cudaMemcpyAsync(dst, src, count, kind, stream),
270 271 272
                 "cudaMemcpyAsync failed in paddle::platform::GpuMemcpyAsync "
                 "(%p -> %p, length: %d)",
                 src, dst, static_cast<int>(count));
L
liaogang 已提交
273 274
}

275 276 277
void GpuMemcpySync(void *dst, const void *src, size_t count,
                   enum cudaMemcpyKind kind) {
  PADDLE_ENFORCE(cudaMemcpy(dst, src, count, kind),
278 279 280
                 "cudaMemcpy failed in paddle::platform::GpuMemcpySync (%p -> "
                 "%p, length: %d)",
                 src, dst, static_cast<int>(count));
281 282 283 284
}

void GpuMemcpyPeerAsync(void *dst, int dst_device, const void *src,
                        int src_device, size_t count, cudaStream_t stream) {
L
liaogang 已提交
285
  PADDLE_ENFORCE(
L
liaogang 已提交
286
      cudaMemcpyPeerAsync(dst, dst_device, src, src_device, count, stream),
287 288 289 290 291 292 293 294
      "cudaMemcpyPeerAsync failed in paddle::platform::GpuMemcpyPeerAsync");
}

void GpuMemcpyPeerSync(void *dst, int dst_device, const void *src,
                       int src_device, size_t count) {
  PADDLE_ENFORCE(
      cudaMemcpyPeer(dst, dst_device, src, src_device, count),
      "cudaMemcpyPeer failed in paddle::platform::GpuMemcpyPeerSync");
L
liaogang 已提交
295
}
D
dzhwinter 已提交
296 297 298 299 300

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
301 302
}  // namespace platform
}  // namespace paddle