forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
AbsCriterion.cu
62 lines (54 loc) · 1.4 KB
/
AbsCriterion.cu
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
#include "THCUNN.h"
#include "common.h"
#include "TH/THHalf.h"
#include "THCHalfAutoNumerics.cuh"
#include "THCApply.cuh"
#include <thrust/fill.h>
#include <thrust/functional.h>
#include <thrust/device_ptr.h>
#include <thrust/reduce.h>
#include <thrust/inner_product.h>
template <typename Dtype, typename Acctype>
struct abs_functor
{
__host__ __device__ Acctype operator()(const Dtype& x, const Dtype& y) const
{
Dtype z = x-y;
return ScalarConvert<Dtype, Acctype>::to(z >= 0 ? z : -z);
}
};
template <typename Dtype>
struct abs_updateOutput_no_reduce_functor
{
__host__ __device__ void operator()(const Dtype* x, const Dtype* y, Dtype *out)
{
Dtype z = *x - *y;
*out = z >= 0 ? z : -z;
}
};
template <typename Dtype>
struct abs_updateGradInput_no_reduce_functor
{
__forceinline__ __host__ __device__ void operator()(
const Dtype *x,
const Dtype *y,
Dtype *gradInput)
{
*gradInput = ScalarConvert<int, Dtype>::to(*x >= *y ? 1 : -1);
}
};
template <typename Dtype>
struct abs_updateGradInput_functor
{
const Dtype norm;
const Dtype gradOutput;
abs_updateGradInput_functor(Dtype norm_, Dtype gradOutput_)
: norm(norm_), gradOutput(gradOutput_)
{}
__host__ __device__ Dtype operator()(const Dtype& x, const Dtype& y) const
{
return ((x - y) >= 0 ? norm : -norm) * gradOutput;
}
};
#include "generic/AbsCriterion.cu"
#include "THCGenerateFloatTypes.h"