f4334277
 Hu Chunming
提交3rdparty
|
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
|
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.
#ifndef OPENCV_DNN_CUDA4DNN_CSL_CUDNN_SOFTMAX_HPP
#define OPENCV_DNN_CUDA4DNN_CSL_CUDNN_SOFTMAX_HPP
#include "cudnn.hpp"
#include "../pointer.hpp"
#include <cudnn.h>
namespace cv { namespace dnn { namespace cuda4dnn { namespace csl { namespace cudnn {
/** @brief computes softmax (or log softmax)
*
* @tparam T element type (must be `half` or `float`)
*
* @param handle valid cuDNN handle
* @param outputDesc tensor descriptor for A
* @param[out] output pointer to tensor in device memory
* @param inputDesc tensor descriptor for C
* @param[in] input pointer to tensor in device memory
* @param log apply log on probabilities
*
* Exception Guarantee: Basic
*/
template <class T>
void softmax(const cudnn::Handle& handle,
const TensorDescriptor<T>& outputDesc, DevicePtr<T> output,
const TensorDescriptor<T>& inputDesc, DevicePtr<const T> input,
bool log)
{
T alpha = 1.0, beta = 0.0;
cudnnSoftmaxAlgorithm_t algo = log ? CUDNN_SOFTMAX_LOG : CUDNN_SOFTMAX_ACCURATE;
CUDA4DNN_CHECK_CUDNN(
cudnnSoftmaxForward(
handle.get(),
algo, CUDNN_SOFTMAX_MODE_CHANNEL,
&alpha, inputDesc.get(), input.get(),
&beta, outputDesc.get(), output.get()
)
);
}
template <> inline
void softmax(const cudnn::Handle& handle,
const TensorDescriptor<half>& outputDesc, DevicePtr<half> output,
const TensorDescriptor<half>& inputDesc, DevicePtr<const half> input,
bool log)
{
/* we specalize for fp16 as the scaling factors must be provided as `float` */
float alpha = 1.0, beta = 0.0;
cudnnSoftmaxAlgorithm_t algo = log ? CUDNN_SOFTMAX_LOG : CUDNN_SOFTMAX_ACCURATE;
CUDA4DNN_CHECK_CUDNN(
cudnnSoftmaxForward(
handle.get(),
algo, CUDNN_SOFTMAX_MODE_CHANNEL,
&alpha, inputDesc.get(), input.get(),
&beta, outputDesc.get(), output.get()
)
);
}
}}}}} /* namespace cv::dnn::cuda4dnn::csl::cudnn */
#endif /* OPENCV_DNN_CUDA4DNN_CSL_CUDNN_SOFTMAX_HPP */
|
