forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
erf_op.cu
60 lines (51 loc) · 1.29 KB
/
erf_op.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
#include "caffe2/operators/erf_op.h"
#include <algorithm>
#include <functional>
#include "caffe2/core/context_gpu.h"
namespace caffe2 {
namespace {
__global__ void ErfGradientCUDAKernel(
const int N,
const float* dY,
const float* X,
float* dX) {
CUDA_1D_KERNEL_LOOP(i, N) {
#if __CUDA_ARCH__ >= 350
dX[i] = 2.0f / sqrtf(PI) * expf(-powf(__ldg(X+i), 2.0f)) * __ldg(dY + i);
#else
dX[i] = 2.0f / sqrtf(PI) * expf(-powf(X[i], 2.0f)) * dY[i];
#endif
}
}
} // namespace
template <>
template <typename T>
bool ErfGradientFunctor<CUDAContext>::Forward(
const std::vector<int>& X_dims,
const std::vector<int>& /* dY_dims */,
const T* X,
const T* dY,
T* dX,
CUDAContext* context) const {
const int size = std::accumulate(
X_dims.cbegin(), X_dims.cend(), 1, std::multiplies<int>());
ErfGradientCUDAKernel<<<
CAFFE_GET_BLOCKS(size),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(size, dY, X, dX);
return true;
}
REGISTER_CUDA_OPERATOR(
Erf,
UnaryElementwiseOp<
TensorTypes<float>,
CUDAContext,
ErfFunctor<CUDAContext>>);
REGISTER_CUDA_OPERATOR(
ErfGradient,
BinaryElementwiseOp<
TensorTypes<float>,
CUDAContext,
ErfGradientFunctor<CUDAContext>>);
} // namespace caffe2