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
73
74
75
76
77
78
79
80
81
82
83
84
85
|
#include "caffe2/operators/sigmoid_op.h"
#include <algorithm>
#include <functional>
#include "caffe2/core/context_gpu.h"
namespace caffe2 {
namespace {
template <typename T>
__global__ void SigmoidCUDAKernel(const int N, const T* X, T* Y);
template <>
__global__ void
SigmoidCUDAKernel<float>(const int N, const float* X, float* Y) {
CUDA_1D_KERNEL_LOOP(i, N) {
#if __CUDA_ARCH__ >= 350
Y[i] = 1.0f / (1.0f + expf(-__ldg(X + i)));
#else
Y[i] = 1.0f / (1.0f + expf(-X[i]));
#endif
}
}
template <typename T>
__global__ void
SigmoidGradientCUDAKernel(const int N, const T* dY, const T* Y, T* dX) {
CUDA_1D_KERNEL_LOOP(i, N) {
#if __CUDA_ARCH__ >= 350
dX[i] = __ldg(dY + i) * __ldg(Y + i) * (T(1) - __ldg(Y + i));
#else
dX[i] = dY[i] * Y[i] * (T(1) - Y[i]);
#endif
}
}
} // namespace
template <>
template <typename T>
bool SigmoidFunctor<CUDAContext>::
operator()(const int N, const T* X, T* Y, CUDAContext* context) const {
SigmoidCUDAKernel<T>
<<<CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(N, X, Y);
return true;
}
template <>
template <typename T>
bool SigmoidGradientFunctor<CUDAContext>::Forward(
const std::vector<int>& Y_dims,
const std::vector<int>& /* dY_dims */,
const T* Y,
const T* dY,
T* dX,
CUDAContext* context) const {
const int size = std::accumulate(
Y_dims.cbegin(), Y_dims.cend(), 1, std::multiplies<int>());
SigmoidGradientCUDAKernel<T>
<<<CAFFE_GET_BLOCKS(size),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(size, dY, Y, dX);
return true;
}
REGISTER_CUDA_OPERATOR(
Sigmoid,
UnaryElementwiseOp<
TensorTypes<float>,
CUDAContext,
SigmoidFunctor<CUDAContext>>);
REGISTER_CUDA_OPERATOR(
SigmoidGradient,
BinaryElementwiseOp<
TensorTypes<float>,
CUDAContext,
SigmoidGradientFunctor<CUDAContext>>);
} // namespace caffe2
|
