forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Prelu.cpp
72 lines (57 loc) · 2.67 KB
/
Prelu.cpp
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
71
72
#include <ATen/ATen.h>
#include <ATen/NativeFunctions.h>
#include <ATen/Config.h>
#if !AT_MKLDNN_ENABLED()
namespace at { namespace native {
Tensor mkldnn_prelu(const Tensor& input, const Tensor& weight) {
TORCH_CHECK(false, "mkldnn_prelu: ATen not compiled with MKLDNN support");
}
std::tuple<Tensor, Tensor> mkldnn_prelu_backward(const Tensor& grad_output, const Tensor& input, const Tensor& weight) {
TORCH_CHECK(false, "mkldnn_prelu_backward: ATen not compiled with MKLDNN support");
}
}}
#else // AT_MKLDNN_ENABLED
#include <ATen/native/mkldnn/MKLDNNCommon.h>
#include <ATen/native/mkldnn/Utils.h>
namespace at { namespace native {
Tensor mkldnn_prelu(const Tensor& input, const Tensor& weight) {
if (input.scalar_type() == ScalarType::BFloat16) {
TORCH_CHECK(mkldnn_bf16_device_check(),
"mkldnn_relu: bf16 path needs the cpu support avx512bw, avx512vl and avx512dq");
}
const ideep::tensor& x = itensor_from_mkldnn(input);
const ideep::tensor& w = itensor_from_tensor(weight);
ideep::tensor y;
ideep::prelu_forward::compute(
x, w, y, ideep::prop_kind::forward_training);
return new_with_itensor_mkldnn(std::move(y), optTypeMetaToScalarType(input.options().dtype_opt()),
input.options().device_opt());
}
std::tuple<Tensor, Tensor> mkldnn_prelu_backward(const Tensor& grad_output, const Tensor& input, const Tensor& weight) {
const ideep::tensor& x = itensor_from_mkldnn(input);
const ideep::tensor& w = itensor_from_tensor(weight);
const ideep::tensor grady = itensor_from_mkldnn(grad_output);
ideep::tensor gradx;
ideep::tensor gradw;
ideep::prelu_backward::compute(
x, w, grady, gradx, gradw, ideep::prop_kind::backward);
if (weight.is_mkldnn()) {
return std::make_tuple(
new_with_itensor_mkldnn(std::move(gradx),
optTypeMetaToScalarType(grad_output.options().dtype_opt()),
grad_output.options().device_opt()),
new_with_itensor_mkldnn(std::move(gradw),
optTypeMetaToScalarType(weight.options().dtype_opt()),
weight.options().device_opt()));
} else {
return std::make_tuple(
new_with_itensor_mkldnn(std::move(gradx),
optTypeMetaToScalarType(grad_output.options().dtype_opt()),
grad_output.options().device_opt()),
mkldnn_to_dense(new_with_itensor_mkldnn(std::move(gradw),
optTypeMetaToScalarType(weight.options().dtype_opt()),
weight.options().device_opt())));
}
}
}}
#endif // AT_MKLDNN_ENABLED