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 "caffe2/operators/rsqrt_op.h"
#include <algorithm>
#include <functional>
#include "caffe2/core/context_gpu.h"
#include "caffe2/utils/math.h"
namespace caffe2 {
namespace {
template <typename T>
__global__ void
RsqrtGradientCUDAKernel(const int size, const T* dY, const T* Y, T* dX) {
CUDA_1D_KERNEL_LOOP(i, size) {
#if __CUDA_ARCH__ >= 350
dX[i] = __ldg(dY + i) * math::utils::Cube<T>(__ldg(Y + i)) *
static_cast<T>(-0.5);
#else
dX[i] = dY[i] * math::utils::Cube<T>(Y[i]) * static_cast<T>(-0.5);
#endif
}
}
} // namespace
template <>
template <typename T>
bool RsqrtGradientFunctor<CUDAContext>::Forward(
const std::vector<int>& dY_dims,
const std::vector<int>& /* Y_dims */,
const T* dY,
const T* Y,
T* dX,
CUDAContext* context) const {
const int size = std::accumulate(
dY_dims.cbegin(), dY_dims.cend(), 1, std::multiplies<int>());
RsqrtGradientCUDAKernel<T>
<<<CAFFE_GET_BLOCKS(size),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(size, dY, Y, dX);
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
REGISTER_CUDA_OPERATOR(
Rsqrt,
UnaryElementwiseOp<
TensorTypes<float>,
CUDAContext,
RsqrtFunctor<CUDAContext>>);
REGISTER_CUDA_OPERATOR(
RsqrtGradient,
BinaryElementwiseOp<
TensorTypes<float>,
CUDAContext,
RsqrtGradientFunctor<CUDAContext>>);
} // namespace caffe2
|
