File: common_world_ops_gpu.cc

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (35 lines) | stat: -rw-r--r-- 789 bytes parent folder | download | duplicates (2)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
#include "caffe2/contrib/gloo/common_world_ops.h"

#include "caffe2/core/context_gpu.h"

#include <gloo/cuda.h>
#include <gloo/transport/tcp/device.h>

namespace caffe2 {
namespace gloo {

template <>
void CreateCommonWorld<CUDAContext>::initializeForContext() {
  static std::once_flag once;
  std::call_once(once, [&]() {
      // This is the first time we call Gloo code for a CUDAContext.
      // Share Caffe2 CUDA mutex with Gloo.
      ::gloo::CudaShared::setMutex(&CUDAContext::mutex());
    });
}

namespace {

REGISTER_CUDA_OPERATOR_WITH_ENGINE(
    CreateCommonWorld,
    GLOO,
    CreateCommonWorld<CUDAContext>);

REGISTER_CUDA_OPERATOR_WITH_ENGINE(
    CloneCommonWorld,
    GLOO,
    CloneCommonWorld<CUDAContext>);

} // namespace
} // namespace gloo
} // namespace caffe2