File: copy_op.cu

package info (click to toggle)
pytorch 1.7.1-7
  • links: PTS, VCS
  • area: main
  • in suites: bullseye
  • size: 80,340 kB
  • sloc: cpp: 670,830; python: 343,991; ansic: 67,845; asm: 5,503; sh: 2,924; java: 2,888; xml: 266; makefile: 244; ruby: 148; yacc: 144; objc: 51; lex: 44
file content (58 lines) | stat: -rw-r--r-- 1,943 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
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/copy_op.h"

namespace caffe2 {

template <>
class CopyOnDeviceLikeOp<CUDAContext, CUDAContext, CUDAContext>
    : public Operator<CUDAContext> {
 public:
  template <class... Args>
  explicit CopyOnDeviceLikeOp(Args&&... args)
      : Operator<CUDAContext>(std::forward<Args>(args)...) {}
  USE_OPERATOR_FUNCTIONS(CUDAContext);

  bool RunOnDevice() override {
    auto& input = Input(0);
    auto* output = OperatorBase::Output<Tensor>(0, CUDA);
    CUDAContext context(GetGPUIDForPointer(Input(1).raw_data()));
    output->ResizeLike(input);
    context.template CopyItems<CUDAContext, CUDAContext>(
        input.meta(),
        input.numel(),
        input.raw_data(),
        output->raw_mutable_data(input.meta()));
    return true;
  }
};

// From CPU, copy it to whatever the current context
REGISTER_CUDA_OPERATOR(
    CopyFromCPUInput,
    CopyOp<CUDAContext, CUDAContext, CPUContext>);

// CopyGPUToCPU and CopyCPUToGPU should both be carried out in a cuda context,
// since gpu code will be involved.
REGISTER_CUDA_OPERATOR(
    CopyGPUToCPU,
    CopyOp<CUDAContext, CPUContext, CUDAContext>);
REGISTER_CUDA_OPERATOR(
    CopyCPUToGPU,
    CopyOp<CUDAContext, CUDAContext, CPUContext>);
// If we only specify Copy, we assume that it is a gpu to gpu copy - maybe
// involving different GPUs.
REGISTER_CUDA_OPERATOR(Copy, CopyOp<CUDAContext, CUDAContext, CUDAContext>);

REGISTER_CUDA_OPERATOR(
    CopyOnDeviceLike,
    CopyOnDeviceLikeOp<CUDAContext, CUDAContext, CUDAContext>);
} // namespace caffe2

using CopyGPUToCPU_CUDA = caffe2::
    CopyOp<caffe2::CUDAContext, caffe2::CPUContext, caffe2::CUDAContext>;
using CopyCPUToGPU_CUDA = caffe2::
    CopyOp<caffe2::CUDAContext, caffe2::CUDAContext, caffe2::CPUContext>;

C10_EXPORT_CAFFE2_OP_TO_C10_CUDA(CopyGPUToCPU, CopyGPUToCPU_CUDA);

C10_EXPORT_CAFFE2_OP_TO_C10_CPU_KERNEL_ONLY(CopyCPUToGPU, CopyCPUToGPU_CUDA);