merge.simd.hpp
6.21 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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
// 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"
namespace cv { namespace hal {
CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
void merge8u(const uchar** src, uchar* dst, int len, int cn);
void merge16u(const ushort** src, ushort* dst, int len, int cn);
void merge32s(const int** src, int* dst, int len, int cn);
void merge64s(const int64** src, int64* dst, int len, int cn);
#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
#if CV_SIMD
/*
The trick with STORE_UNALIGNED/STORE_ALIGNED_NOCACHE is the following:
on IA there are instructions movntps and such to which
v_store_interleave(...., STORE_ALIGNED_NOCACHE) is mapped.
Those instructions write directly into memory w/o touching cache
that results in dramatic speed improvements, especially on
large arrays (FullHD, 4K etc.).
Those intrinsics require the destination address to be aligned
by 16/32 bits (with SSE2 and AVX2, respectively).
So we potentially split the processing into 3 stages:
1) the optional prefix part [0:i0), where we use simple unaligned stores.
2) the optional main part [i0:len - VECSZ], where we use "nocache" mode.
But in some cases we have to use unaligned stores in this part.
3) the optional suffix part (the tail) (len - VECSZ:len) where we switch back to "unaligned" mode
to process the remaining len - VECSZ elements.
In principle there can be very poorly aligned data where there is no main part.
For that we set i0=0 and use unaligned stores for the whole array.
*/
template<typename T, typename VecT> static void
vecmerge_( const T** src, T* dst, int len, int cn )
{
const int VECSZ = VecT::nlanes;
int i, i0 = 0;
const T* src0 = src[0];
const T* src1 = src[1];
const int dstElemSize = cn * sizeof(T);
int r = (int)((size_t)(void*)dst % (VECSZ*sizeof(T)));
hal::StoreMode mode = hal::STORE_ALIGNED_NOCACHE;
if( r != 0 )
{
mode = hal::STORE_UNALIGNED;
if (r % dstElemSize == 0 && len > VECSZ*2)
i0 = VECSZ - (r / dstElemSize);
}
if( cn == 2 )
{
for( i = 0; i < len; i += VECSZ )
{
if( i > len - VECSZ )
{
i = len - VECSZ;
mode = hal::STORE_UNALIGNED;
}
VecT a = vx_load(src0 + i), b = vx_load(src1 + i);
v_store_interleave(dst + i*cn, a, b, mode);
if( i < i0 )
{
i = i0 - VECSZ;
mode = hal::STORE_ALIGNED_NOCACHE;
}
}
}
else if( cn == 3 )
{
const T* src2 = src[2];
for( i = 0; i < len; i += VECSZ )
{
if( i > len - VECSZ )
{
i = len - VECSZ;
mode = hal::STORE_UNALIGNED;
}
VecT a = vx_load(src0 + i), b = vx_load(src1 + i), c = vx_load(src2 + i);
v_store_interleave(dst + i*cn, a, b, c, mode);
if( i < i0 )
{
i = i0 - VECSZ;
mode = hal::STORE_ALIGNED_NOCACHE;
}
}
}
else
{
CV_Assert( cn == 4 );
const T* src2 = src[2];
const T* src3 = src[3];
for( i = 0; i < len; i += VECSZ )
{
if( i > len - VECSZ )
{
i = len - VECSZ;
mode = hal::STORE_UNALIGNED;
}
VecT a = vx_load(src0 + i), b = vx_load(src1 + i);
VecT c = vx_load(src2 + i), d = vx_load(src3 + i);
v_store_interleave(dst + i*cn, a, b, c, d, mode);
if( i < i0 )
{
i = i0 - VECSZ;
mode = hal::STORE_ALIGNED_NOCACHE;
}
}
}
vx_cleanup();
}
#endif
template<typename T> static void
merge_( const T** src, T* dst, int len, int cn )
{
int k = cn % 4 ? cn % 4 : 4;
int i, j;
if( k == 1 )
{
const T* src0 = src[0];
for( i = j = 0; i < len; i++, j += cn )
dst[j] = src0[i];
}
else if( k == 2 )
{
const T *src0 = src[0], *src1 = src[1];
i = j = 0;
for( ; i < len; i++, j += cn )
{
dst[j] = src0[i];
dst[j+1] = src1[i];
}
}
else if( k == 3 )
{
const T *src0 = src[0], *src1 = src[1], *src2 = src[2];
i = j = 0;
for( ; i < len; i++, j += cn )
{
dst[j] = src0[i];
dst[j+1] = src1[i];
dst[j+2] = src2[i];
}
}
else
{
const T *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
i = j = 0;
for( ; i < len; i++, j += cn )
{
dst[j] = src0[i]; dst[j+1] = src1[i];
dst[j+2] = src2[i]; dst[j+3] = src3[i];
}
}
for( ; k < cn; k += 4 )
{
const T *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
for( i = 0, j = k; i < len; i++, j += cn )
{
dst[j] = src0[i]; dst[j+1] = src1[i];
dst[j+2] = src2[i]; dst[j+3] = src3[i];
}
}
}
void merge8u(const uchar** src, uchar* dst, int len, int cn )
{
CV_INSTRUMENT_REGION();
#if CV_SIMD
if( len >= v_uint8::nlanes && 2 <= cn && cn <= 4 )
vecmerge_<uchar, v_uint8>(src, dst, len, cn);
else
#endif
merge_(src, dst, len, cn);
}
void merge16u(const ushort** src, ushort* dst, int len, int cn )
{
CV_INSTRUMENT_REGION();
#if CV_SIMD
if( len >= v_uint16::nlanes && 2 <= cn && cn <= 4 )
vecmerge_<ushort, v_uint16>(src, dst, len, cn);
else
#endif
merge_(src, dst, len, cn);
}
void merge32s(const int** src, int* dst, int len, int cn )
{
CV_INSTRUMENT_REGION();
#if CV_SIMD
if( len >= v_int32::nlanes && 2 <= cn && cn <= 4 )
vecmerge_<int, v_int32>(src, dst, len, cn);
else
#endif
merge_(src, dst, len, cn);
}
void merge64s(const int64** src, int64* dst, int len, int cn )
{
CV_INSTRUMENT_REGION();
#if CV_SIMD
if( len >= v_int64::nlanes && 2 <= cn && cn <= 4 )
vecmerge_<int64, v_int64>(src, dst, len, cn);
else
#endif
merge_(src, dst, len, cn);
}
#endif
CV_CPU_OPTIMIZATION_NAMESPACE_END
}} // namespace