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
|
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/leaky_relu_op.h"
#include "caffe2/utils/math.h"
namespace caffe2 {
namespace {
template <typename T>
__global__ void LeakyReluKernel(const int N, const T alpha, const T* X, T* Y) {
CUDA_1D_KERNEL_LOOP(i, N) {
Y[i] = X[i] >= 0 ? X[i] : X[i] * alpha;
}
}
template <typename T>
__global__ void LeakyReluGradientKernel(
const int N,
const T alpha,
const T* Y,
const T* dY,
T* dX) {
CUDA_1D_KERNEL_LOOP(i, N) {
dX[i] = Y[i] >= 0 ? dY[i] : dY[i] * alpha;
}
}
} // namespace
template <>
bool LeakyReluOp<float, CUDAContext>::RunOnDevice() {
const auto& X = Input(0);
CAFFE_ENFORCE_GT(X.numel(), 0);
auto* Y = Output(0, X.sizes(), at::dtype<float>());
LeakyReluKernel<<<
CAFFE_GET_BLOCKS(X.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
X.numel(), alpha_, X.data<float>(), Y->template mutable_data<float>());
return true;
}
template <>
bool LeakyReluGradientOp<float, CUDAContext>::RunOnDevice() {
const auto& Y = Input(0);
const auto& dY = Input(1);
auto* dX = Output(0, Y.sizes(), at::dtype<float>());
CAFFE_ENFORCE_EQ(Y.numel(), dY.numel());
LeakyReluGradientKernel<<<
CAFFE_GET_BLOCKS(Y.numel()),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
Y.numel(),
alpha_,
Y.data<float>(),
dY.data<float>(),
dX->template mutable_data<float>());
return true;
}
REGISTER_CUDA_OPERATOR(LeakyRelu, LeakyReluOp<float, CUDAContext>);
REGISTER_CUDA_OPERATOR(
LeakyReluGradient,
LeakyReluGradientOp<float, CUDAContext>);
} // namespace caffe2
|
