activation_eltwise.cu
6.4 KB
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
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
// 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.
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include "functors.hpp"
#include "vector_traits.hpp"
#include "grid_stride_range.hpp"
#include "execution.hpp"
#include "../cuda4dnn/csl/stream.hpp"
#include "../cuda4dnn/csl/span.hpp"
using namespace cv::dnn::cuda4dnn::csl;
using namespace cv::dnn::cuda4dnn::csl::device;
namespace cv { namespace dnn { namespace cuda4dnn { namespace kernels {
namespace raw {
template <class T, class ActivationOp, class EltwiseOp, std::size_t N>
__global__ void generic_op_eltwise_op_inplace_vec(Span<T> inplace_output, View<T> eltwise, const typename ActivationOp::Params act_params, const typename EltwiseOp::Params eltwise_params) {
using vector_type = get_vector_type_t<T, N>;
auto inplace_output_vPtr = vector_type::get_pointer(inplace_output.data());
auto eltwise_vPtr = vector_type::get_pointer(eltwise.data());
ActivationOp activation_op(act_params);
EltwiseOp eltwise_op(eltwise_params);
for (auto i : grid_stride_range(inplace_output.size() / vector_type::size())) {
vector_type output_vec, eltwise_vec;
v_load(output_vec, inplace_output_vPtr[i]);
v_load(eltwise_vec, eltwise_vPtr[i]);
for(int j = 0; j < output_vec.size(); j++)
output_vec.data[j] = eltwise_op(activation_op(output_vec.data[j]), eltwise_vec.data[j]);
v_store(inplace_output_vPtr[i], output_vec);
}
}
}
template <class T, class ActivationOp, class EltwiseOp, std::size_t N> static
void launch_vectorized_generic_op_eltwise_op_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise, const typename ActivationOp::Params& act_params, const typename EltwiseOp::Params& eltwise_params) {
CV_Assert(is_fully_aligned<T>(inplace_output, N));
CV_Assert(is_fully_aligned<T>(eltwise, N));
auto kernel = raw::generic_op_eltwise_op_inplace_vec<T, ActivationOp, EltwiseOp, N>;
auto policy = make_policy(kernel, inplace_output.size() / N, 0, stream);
launch_kernel(kernel, policy, inplace_output, eltwise, act_params, eltwise_params);
}
template <class T, class ActivationOp, class EltwiseOp> static
void generic_op_eltwise_op_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise, const typename ActivationOp::Params& act_params = {}, const typename EltwiseOp::Params& eltwise_params = {}) {
CV_Assert(inplace_output.size() == eltwise.size());
if (is_fully_aligned<T>(inplace_output, 4) && is_fully_aligned<T>(eltwise, 4)) {
launch_vectorized_generic_op_eltwise_op_inplace<T, ActivationOp, EltwiseOp, 4>(stream, inplace_output, eltwise, act_params, eltwise_params);
} else if (is_fully_aligned<T>(inplace_output, 2) && is_fully_aligned<T>(eltwise, 2)) {
launch_vectorized_generic_op_eltwise_op_inplace<T, ActivationOp, EltwiseOp, 2>(stream, inplace_output, eltwise, act_params, eltwise_params);
} else {
launch_vectorized_generic_op_eltwise_op_inplace<T, ActivationOp, EltwiseOp, 1>(stream, inplace_output, eltwise, act_params, eltwise_params);
}
}
template <class T>
void relu_eltwise_sum_2_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise, T slope) {
generic_op_eltwise_op_inplace<T, ReLUFunctor<T>, SumFunctor<T>>(stream, inplace_output, eltwise, {slope});
}
template <class T>
void clipped_relu_eltwise_sum_2_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise, T floor, T ceiling) {
CV_Assert(static_cast<double>(floor) <= static_cast<double>(ceiling));
generic_op_eltwise_op_inplace<T, ClippedReLUFunctor<T>, SumFunctor<T>>(stream, inplace_output, eltwise, {floor, ceiling});
}
template <class T>
void tanh_eltwise_sum_2_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise) {
generic_op_eltwise_op_inplace<T, TanHFunctor<T>, SumFunctor<T>>(stream, inplace_output, eltwise);
}
template <class T>
void swish_eltwise_sum_2_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise) {
generic_op_eltwise_op_inplace<T, SwishFunctor<T>, SumFunctor<T>>(stream, inplace_output, eltwise);
}
template <class T>
void mish_eltwise_sum_2_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise) {
generic_op_eltwise_op_inplace<T, MishFunctor<T>, SumFunctor<T>>(stream, inplace_output, eltwise);
}
template <class T>
void sigmoid_eltwise_sum_2_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise) {
generic_op_eltwise_op_inplace<T, SigmoidFunctor<T>, SumFunctor<T>>(stream, inplace_output, eltwise);
}
template <class T>
void power_eltwise_sum_2_inplace(const Stream& stream, Span<T> inplace_output, View<T> eltwise, T exp, T scale, T shift) {
generic_op_eltwise_op_inplace<T, PowerFunctor<T>, SumFunctor<T>>(stream, inplace_output, eltwise, {exp, scale, shift});
}
#if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 530)
template void relu_eltwise_sum_2_inplace<__half>(const Stream&, Span<__half>, View<__half>, __half);
template void clipped_relu_eltwise_sum_2_inplace<__half>(const Stream&, Span<__half>, View<__half>, __half, __half);
template void tanh_eltwise_sum_2_inplace<__half>(const Stream&, Span<__half>, View<__half>);
template void swish_eltwise_sum_2_inplace<__half>(const Stream&, Span<__half>, View<__half>);
template void mish_eltwise_sum_2_inplace<__half>(const Stream&, Span<__half>, View<__half>);
template void sigmoid_eltwise_sum_2_inplace<__half>(const Stream&, Span<__half>, View<__half>);
template void power_eltwise_sum_2_inplace<__half>(const Stream&, Span<__half>, View<__half>, __half, __half, __half);
#endif
template void relu_eltwise_sum_2_inplace<float>(const Stream&, Span<float>, View<float>, float);
template void clipped_relu_eltwise_sum_2_inplace<float>(const Stream&, Span<float>, View<float>, float, float);
template void tanh_eltwise_sum_2_inplace<float>(const Stream&, Span<float>, View<float>);
template void swish_eltwise_sum_2_inplace<float>(const Stream&, Span<float>, View<float>);
template void mish_eltwise_sum_2_inplace<float>(const Stream&, Span<float>, View<float>);
template void sigmoid_eltwise_sum_2_inplace<float>(const Stream&, Span<float>, View<float>);
template void power_eltwise_sum_2_inplace<float>(const Stream&, Span<float>, View<float>, float, float, float);
}}}} /* namespace cv::dnn::cuda4dnn::kernels */