File: order_switch_ops.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 (70 lines) | stat: -rw-r--r-- 2,047 bytes parent folder | download
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
59
60
61
62
63
64
65
66
67
68
69
70
#include <caffe2/ideep/ideep_utils.h>

using namespace caffe2;

namespace {

class IDEEPNHWC2NCHWOp final : public IDEEPOperator {
 public:
  USE_IDEEP_DEF_ALIASES();
  USE_SIMPLE_IDEEP_CTOR_DTOR(IDEEPNHWC2NCHWOp);
  USE_IDEEP_OPERATOR_FUNCTIONS();

  bool RunOnDevice() override {
    const auto& X = Input(0);
    CAFFE_ENFORCE_EQ(X.ndims(), 4);
    CAFFE_ENFORCE(X.get_desc().is_nhwc());

    auto *Y = Output(OUTPUT);
    CAFFE_ENFORCE(Y != &X);

    // NOTE: NHWC changes the shape in framework, but not in MKL-DNN
    // Thus, for iDEEP tensor, the shapes of NCHW and NHWC are identical.
    Y->init({X.get_dims(), X.get_data_type(), iformat::nchw});
    Y->feed_from(X);
    // NOTE: This ops is only used to quantization path, setting scale
    // to distinguish with fp32 path activation(always return NCHW format
    // even ideep tensor has NHWC format) when convert to numpy memory.
    Y->set_scale({1.0});
    return true;
  }

 private:
  INPUT_TAGS(INPUT);
  OUTPUT_TAGS(OUTPUT);
};

class IDEEPNCHW2NHWCOp final : public IDEEPOperator {
 public:
  USE_IDEEP_DEF_ALIASES();
  USE_SIMPLE_IDEEP_CTOR_DTOR(IDEEPNCHW2NHWCOp);
  USE_IDEEP_OPERATOR_FUNCTIONS();

  bool RunOnDevice() override {
    const auto& X = Input(0);
    CAFFE_ENFORCE_EQ(X.ndims(), 4);
    CAFFE_ENFORCE(X.get_desc().is_nchw());

    auto *Y = Output(OUTPUT);
    CAFFE_ENFORCE(Y != &X);

    // NOTE: NHWC changes the shape in framework, but not in MKL-DNN
    // Thus, for iDEEP tensor, the shapes of NCHW and NHWC are identical.
    Y->init({X.get_dims(), X.get_data_type(), iformat::nhwc});
    Y->feed_from(X);
    // NOTE: This ops is only used to quantization path, setting scale
    // to distinguish with fp32 path activation(always return NCHW format
    // even ideep tensor has NHWC format) when convert to numpy memory.
    Y->set_scale({1.0});
    return true;
  }

 private:
  INPUT_TAGS(INPUT);
  OUTPUT_TAGS(OUTPUT);
};

REGISTER_IDEEP_OPERATOR(NHWC2NCHW, IDEEPNHWC2NCHWOp);
REGISTER_IDEEP_OPERATOR(NCHW2NHWC, IDEEPNCHW2NHWCOp);

} // namespace