elementwise_layers.cpp
6.54 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
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
// 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 "../precomp.hpp"
#include "layers_common.hpp"
#include <opencv2/dnn/shape_utils.hpp>
#include <iostream>
namespace cv
{
namespace dnn
{
class ActivationLayerInt8Impl CV_FINAL : public ActivationLayerInt8
{
public:
ActivationLayerInt8Impl(const LayerParams ¶ms)
{
setParamsFrom(params);
activationLUT = !blobs.empty() ? blobs[0] : Mat();
}
virtual bool supportBackend(int backendId) CV_OVERRIDE
{
return backendId == DNN_BACKEND_OPENCV;
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
const int requiredOutputs,
std::vector<MatShape> &outputs,
std::vector<MatShape> &internals) const CV_OVERRIDE
{
Layer::getMemoryShapes(inputs, requiredOutputs, outputs, internals);
return true;
}
class Activation : public cv::ParallelLoopBody
{
public:
const Mat* src;
const Mat* lut;
Mat* dst;
int nstripes;
Activation() : src(0), lut(0), dst(0), nstripes(0){}
static void run(const Mat& src, const Mat& lut, Mat& dst, int nstripes)
{
Activation p;
p.src = &src;
p.lut = &lut;
p.dst = &dst;
p.nstripes = nstripes;
parallel_for_(Range(0, nstripes), p, nstripes);
}
void operator()(const Range &r) const CV_OVERRIDE
{
const int8_t* table = lut->ptr<int8_t>();
int nsamples = 1, outCn = 1;
size_t planeSize = 1;
if (src->dims > 1)
{
nsamples = src->size[0];
outCn = src->size[1];
}
else
outCn = src->size[0];
for (int i = 2; i < src->dims; ++i)
planeSize *= src->size[i];
size_t stripeSize = (planeSize + nstripes - 1)/nstripes;
size_t stripeStart = r.start*stripeSize;
size_t stripeEnd = std::min(r.end*stripeSize, planeSize);
int len = (int)(stripeEnd - stripeStart);
for( int i = 0; i < nsamples; i++ )
{
const int8_t* srcptr = src->ptr<int8_t>(i) + stripeStart;
int8_t* dstptr = dst->ptr<int8_t>(i) + stripeStart;
for( int cn = 0; cn < outCn; cn++, srcptr += planeSize, dstptr += planeSize )
{
int i = 0;
#if CV_SIMD128
for( ; i <= len - 16; i += 16 )
{
v_int8x16 out(table[srcptr[i] + 128], table[srcptr[i+1] + 128], table[srcptr[i+2] + 128], table[srcptr[i+3] + 128],
table[srcptr[i+4] + 128], table[srcptr[i+5] + 128], table[srcptr[i+6] + 128], table[srcptr[i+7] + 128],
table[srcptr[i+8] + 128], table[srcptr[i+9] + 128], table[srcptr[i+10] + 128], table[srcptr[i+11] + 128],
table[srcptr[i+12] + 128], table[srcptr[i+13] + 128], table[srcptr[i+14] + 128], table[srcptr[i+15] + 128]);
v_store(dstptr + i, out);
}
#endif
for( ; i < len; i++ )
{
dstptr[i] = table[srcptr[i] + 128];
}
}
}
}
};
void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
{
CV_TRACE_FUNCTION();
std::vector<Mat> inputs, outputs;
inputs_arr.getMatVector(inputs);
outputs_arr.getMatVector(outputs);
for (size_t i = 0; i < inputs.size(); i++)
{
const Mat &src = inputs[i];
if (!activationLUT.empty())
{
const int nstripes = getNumThreads();
Mat &dst = outputs[i];
CV_Assert(src.size == dst.size && src.type() == dst.type() &&
src.isContinuous() && dst.isContinuous() && src.type() == CV_8S);
Activation::run(src, activationLUT, dst, nstripes);
}
else
{
src.copyTo(outputs[i]);
}
}
}
void forwardSlice(const int8_t* src, const int8_t* lut, int8_t* dst, int len, size_t planeSize, int cn0, int cn1) const CV_OVERRIDE
{
for( int cn = cn0; cn < cn1; cn++, src += planeSize, dst += planeSize )
{
int i = 0;
#if CV_SIMD128
for( ; i <= len - 16; i += 16 )
{
v_int8x16 out(lut[src[i] + 128], lut[src[i+1] + 128], lut[src[i+2] + 128], lut[src[i+3] + 128],
lut[src[i+4] + 128], lut[src[i+5] + 128], lut[src[i+6] + 128], lut[src[i+7] + 128],
lut[src[i+8] + 128], lut[src[i+9] + 128], lut[src[i+10] + 128], lut[src[i+11] + 128],
lut[src[i+12] + 128], lut[src[i+13] + 128], lut[src[i+14] + 128], lut[src[i+15] + 128]);
v_store(dst + i, out);
}
#endif
for( ; i < len; i++ )
dst[i] = lut[src[i] + 128];
}
}
void forwardSlice(const int* src, const int* lut, int* dst, int len, size_t planeSize, int cn0, int cn1) const CV_OVERRIDE
{
for( int cn = cn0; cn < cn1; cn++, src += planeSize, dst += planeSize )
{
int i = 0;
#if CV_SIMD128
for( ; i <= len - 16; i += 16 )
{
v_int32x4 out0(lut[src[i] + 128], lut[src[i+1] + 128], lut[src[i+2] + 128], lut[src[i+3] + 128]);
v_int32x4 out1(lut[src[i+4] + 128], lut[src[i+5] + 128], lut[src[i+6] + 128], lut[src[i+7] + 128]);
v_int32x4 out2(lut[src[i+8] + 128], lut[src[i+9] + 128], lut[src[i+10] + 128], lut[src[i+11] + 128]);
v_int32x4 out3(lut[src[i+12] + 128], lut[src[i+13] + 128], lut[src[i+14] + 128], lut[src[i+15] + 128]);
v_store(dst + i, out0);
v_store(dst + i + 4, out1);
v_store(dst + i + 8, out2);
v_store(dst + i + 12, out3);
}
#endif
for( ; i < len; i++ )
dst[i] = lut[src[i] + 128];
}
}
Mat activationLUT;
};
Ptr<ActivationLayerInt8> ActivationLayerInt8::create(const LayerParams& params)
{
return Ptr<ActivationLayerInt8>(new ActivationLayerInt8Impl(params));
}
}
}