File: conv_op_shared_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 (27 lines) | stat: -rw-r--r-- 910 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
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/conv_op_shared.h"

namespace caffe2 {

template <>
void createSharedBuffer<CUDAContext>(Workspace* ws) {
  auto* mutexPtr = ws->CreateBlob("__CAFFE2_SHARED_CONV_BUFFER_CUDA_MUTEX__")
                       ->GetMutable<std::unique_ptr<std::mutex>>();
  mutexPtr->reset(new std::mutex());
  ws->CreateBlob("__CAFFE2_SHARED_CONV_BUFFER_CUDA__");
}

template <>
void runWithSharedBuffer<CUDAContext>(
    Workspace* ws,
    std::function<void(Tensor* buffer)> f) {
  auto* mutexBlob = ws->GetBlob("__CAFFE2_SHARED_CONV_BUFFER_CUDA_MUTEX__");
  CAFFE_ENFORCE(mutexBlob, "Must call createSharedBuffer() first");

  auto* mutexPtr = mutexBlob->GetMutable<std::unique_ptr<std::mutex>>();
  std::lock_guard<std::mutex> g(**mutexPtr);
  auto* buffer = BlobGetMutableTensor(
      ws->GetBlob("__CAFFE2_SHARED_CONV_BUFFER_CUDA__"), CUDA);
  f(buffer);
}
}