opencv c++

opencv统一simd框架使用(一）

opencv universalsimd 简介

单指令多数据（SIMD，SingleInstruction Multiple Data）是非常强大的优化程序效率方法，顾名思义即可用一条指令运算多个数据，从而提升程序运行效率。举个例子， 假如有以下两个数组a、b,要求计算a、b数组一一对应相加结果存到c数组：

1
2

a = {1, 2, 3, 4}
b = {4, 3, 2, 1}

正常程序写法就是循环遍历a、b数组，一一对应相加存放到c

1
2
3
4

for(int i = 0; i < 4; i++)
{
    c[i] = a[i] + b[i];
}

但是simd可以这么做

1
2
3

//.. 将a、b一次性加载到va、vb
vc = va + vb;//一次运算4个数字相加
//..计算结果vc一次性存储至数组c

OpenCV 4.x(3.4.x版本以上也有)中提供了强大的统一向量指令（universal intrinsics），使用这些指令可以方便地为算法提速，计算密集型任务皆可使用这套指令加速，目前OpenCV源码算法都在用这套指令集框架进行优化。使用这套指令集感受最大的有点便是以下两点：

  • 跨平台：此跨平台指的是cpu指令集的跨平台，不同cpu提供不一样的加速指令集，如x86有sse、avx等，arm有neno，以及新出的Risc-V也有自己的指令集。有了OpenCV统一的Simd指令集库，就可以避免每个硬件平台单独优化算法的麻烦。
  • 自适应向量化内存位宽: 即便是相同架构的cpu也有不同的指令集标准，如早期x86的128bit位宽的sse，发展到后面有256bit位宽的avx,以及扩展到512bit的avx512。这意味着如果编写的代码有可能存在浪费cpu算力的情况（在运行机器可以支持更高的位宽，但是你只用了更低的），或者存在使用指令集标准过新导致运行设备不支持的情况。否则就需要编写面向多个标准的多套加速代码。

使用方法

  • 包含头文件
1

#include "opencv2/core/simd_intrinsics.hpp"
  • 编译选项勾选simd加速选项

Visual Studio 设置方法：属性->C/C++->代码生成->启用增强指令集

Linux Gcc编译选项设置：编译加上-mavx2、等选项

检查opencv支持simd信息

 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

//simd信息获取输出
void PrintOpenCVInfo()
{
    std::cout<<"--------------------------OpenCV informaintion--------------------------"<< std::endl;
    //OpenCV版本，Universal Simd在3.4.x版本就开始提供，不过建议使用4.x版本
    std::cout << "OpenCV version:" << cv::getVersionString() << std::endl;
    std::cout << "Simd info: "<< std::endl;
#ifdef CV_SIMD
    //是否支持simd优化
    std::cout << "CV_SIMD : " << CVAUX_STR(CV_SIMD)  << std::endl;
    //simd优化内存位宽
    std::cout << "CV_SIMD_WIDTH : " <<  CVAUX_STR(CV_SIMD_WIDTH) << std::endl;
    //128bit位宽优化是否支持，绝大部分x86架构cpu支持，如常用的sse指令集
    std::cout << "CV_SIMD128 : " <<CVAUX_STR(CV_SIMD128) << std::endl;
    //256bit位宽优化是否支持，大部分近几年的x86架构cpu支持， avxz指令集
    std::cout << "CV_SIMD256: " <<CVAUX_STR(CV_SIMD256) << std::endl;
    //512bit位宽是否支持，这个intel最先搞得，近几年的intel cpu与最新的AMD锐龙支持
    std::cout << "CV_SIMD512 : " CVAUX_STR(CV_SIMD512) << std::endl;
#else
    std::cout << "CV_SIMD is NOT defined." << std::endl;
#endif

#ifdef CV_SIMD
    std::cout << "sizeof(v_uint8) = " << sizeof(cv::v_uint8) << std::endl;
    std::cout << "sizeof(v_int32) = " << sizeof(cv::v_int32) << std::endl;
    std::cout << "sizeof(v_float32) = " << sizeof(cv::v_float32) << std::endl;
#endif
}

简单使用simd

下面代码介绍了

 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

//加载-计算-存储优化步骤
void SimdSample()
{
    std::cout << "----------------------OpenCV Simd simple sample--------------------------" << std::endl;
    
    // 构造一个常量
    cv::v_int32x4 v_c1(1, 1, 1, 1);             //方法一
    cv::v_int32x4 v_c2 = cv::v_setall(1);       //方法二
    cv::v_int32x4 v_c3 = cv::v_setzero_s32();   //方法三，但只能设置0

    // 构造一个simd长度的数组
    cv::v_int32x4 v_a(1, 2, 3, 4);
    cv::v_int32x4 v_b(1, 2, 3, 4);
    //v_a = [ a0, a1, a2, a3]
    //v_b = [ b0, b1, b2, b3]
    //v_c = [a0+b0, a1+b1, a2+b2, a3+b3]
    cv::v_int32x4 v_c = v_a + v_b;
    PrintVector(v_c, "v_c");
    
    //--------------数组循环优化示例--------------------
    const size_t  array_size = 400;
    int* int_array_a = new int[array_size];
    int* int_array_b = new int[array_size];
    int* int_array_c = new int[array_size];

    //原循环程序计算
    for (int i = 0; i < array_size; i++)
    {
        int_array_c[i] = int_array_a[i] * int_array_b[i];
    }

    //Simd优化:已知编译器向量化内存宽度（如常用的sse为128bit的向量化内存位宽）
    for (int i = 0; i < array_size; i += 4)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);//一次从数组中读取4个int变量存储到v_a
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);//一次从数组中读取4个int变量存储到v_b
        //计算
        cv::v_int32x4 v_c = v_a * v_b;//一次性计算4个变量乘法
        //v_a = [ a0, a1, a2, a3]
        //v_b = [ b0, b1, b2, b3]
        //v_c = [a0*b0, a1*b1, a2*b2, a3*b3]
        //储存
        cv::v_store(int_array_c + i, v_c);
    }

    //Simd自适位宽优化：未知编译器采用的向量化位宽
    for (size_t i = 0; i < array_size; i += cv::v_int32::nlanes)
    {
        //加载 n = 向量化内存位宽 / sizeof(int),如sse是 128 / 32 = 4, avx2 是256 / 32 = 8
        cv::v_int32 v_a = cv::vx_load(int_array_a + i);//一次从数组中读取n个int变量存储到v_a
        cv::v_int32 v_b = cv::vx_load(int_array_b + i);//一次从数组中读取n个int变量存储到v_b
        //计算
        cv::v_int32 v_c = v_a * v_b;//一次性计算n个变量乘法
        //v_a = [ a0, a1, a2, a3]
        //v_b = [ b0, b1, b2, b3]
        //v_c = [a0*b0, a1*b1, a2*b2, a3*b3]
        //储存
        cv::vx_store(int_array_c + i, v_c);
    }
}

常见类型数组for循环优化

 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

//uchar short int float类型数组优化
void BasicDataArrayOptimization()
{
    std::cout << "----------------------OpenCV Simd simple sample--------------------------" << std::endl;

    //uchar
    const size_t  array_size = 4000;
    uchar* uchar_array_a = new uchar[array_size];
    uchar* uchar_array_b = new uchar[array_size];
    uchar* uchar_array_c = new uchar[array_size];

    //一次性加载16个uchar数据进行计算，uchar占8bit, 8x16=128bit
    for (size_t i = 0; i < array_size; i += 16)
    {
        //加载
        cv::v_uint8x16 v_a = cv::v_load(uchar_array_a + i);
        cv::v_uint8x16 v_b = cv::v_load(uchar_array_b + i);
        //计算...
        cv::v_uint8x16 v_c = v_a + v_b;

        //存储
        cv::v_store(uchar_array_c + i, v_c);
    }

    //short
    short* short_arrary_a = new short[array_size];
    short* short_arrary_b = new short[array_size];
    short* short_arrary_c = new short[array_size];
    
    //一次性加载8个short数据进行计算，short占16bit, 16x8=128bit
    for (size_t i = 0; i < array_size; i += 8)
    {
        //加载
        cv::v_int16x8 v_a = cv::v_load(short_arrary_a + i);
        cv::v_int16x8 v_b = cv::v_load(short_arrary_b + i);
        //计算
        cv::v_int16x8 v_c = v_a + v_b;
        //存储
        cv::v_store(short_arrary_c + i, v_c);
    }


    //float
    float* float_array_a = new float[array_size];
    float* float_array_b = new float[array_size];
    float* float_array_c = new float[array_size];
    
    //一次性加载4个float数据进行计算，float占32bit, 32x4=128bit
    for (size_t i = 0; i < array_size; i += 4)
    {
        //加载
        cv::v_float32x4 v_a = cv::v_load(float_array_a + i);
        cv::v_float32x4 v_b = cv::v_load(float_array_a + i);
        //计算
        cv::v_float32x4 v_c = v_a + v_b;
        //存储
        cv::v_store(float_array_c + i, v_c);
    }
}

数组元素个数非位宽整数倍

 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

//顺组对齐示例，长度不是向量优化位宽时倍数
void LoopLenghtAligned()
{
    const size_t  array_size = 4007;
    uchar* uchar_array_a = new uchar[array_size];
    uchar* uchar_array_b = new uchar[array_size];
    uchar* uchar_array_c = new uchar[array_size];

    //一次性加载16个uchar数据进行计算，uchar占8bit, 8x16=128bit
    size_t i = 0;
    for (; i < array_size - 16; i += 16)
    {
        //加载
        cv::v_uint8x16 v_a = cv::v_load(uchar_array_a + i);
        cv::v_uint8x16 v_b = cv::v_load(uchar_array_b + i);
        //计算...
        cv::v_uint8x16 v_c = v_a + v_b;

        //存储
        cv::v_store(uchar_array_c + i, v_c);
    }
    for (; i < array_size; i++)
    {
        uchar_array_c[i] = uchar_array_a[i] + uchar_array_b[i];
    }
}

输出单个simd变量

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19

//输出向量
template<typename T>
void PrintVector(const T& simd_vector, const std::string& name)
{
    //获取向量数组长度
    int arr_size = int(T::nlanes);
    //申请对应类型c风格数组
    T::lane_type* arr = new T::lane_type[arr_size];
    //将向量装载到c风格数组
    cv::v_store(arr, simd_vector);

    //打印输出
    std::cout << name << " : ";
    for (size_t i = 0; i < arr_size; i++)
    {
        std::cout << arr[i] << " ";
    }   
    std::cout << std::endl;
}

数学运算方法

 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

//数学函数
void MathematicsFunctionSample()
{
    std::cout << "-----------------------Mathematics function sample------------------------------" << std::endl;
    cv::v_float32 v_float_arr(1.3, 1.4, 1.5, 1.7);

    //向下取整，处理结果为1 1 1 1
    cv::v_int32x4 v_int_arr = cv::v_floor(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_floor process: ");   


    //向上取整，处理结果为2 2 2 2
    v_int_arr = cv::v_ceil(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_ceil process: ");

    //四舍五入取整，处理结果为：1 1 2 2
    v_int_arr = cv::v_round(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_round process: ");

    //截断小数，目前只实现了保留0为小数，所以结果与v_floor一样
    v_int_arr = cv::v_trunc(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_trunc process: ");


    //最大最小值
    cv::v_int32x4 v1(1, 5, 2, 6);
    cv::v_int32x4 v2(5, 1, 6, 2);
    cv::v_int32x4 v3 = cv::v_min(v1, v2);//1 1 2 2
    cv::v_int32x4 v4 = cv::v_max(v1, v2);//5 5 6 6 
    int min = cv::v_reduce_min(v1);//1
    int max = cv::v_reduce_max(v2);//6
    PrintVector<cv::v_int32x4>(v1, "v1");
    PrintVector<cv::v_int32x4>(v2, "v2");
    PrintVector<cv::v_int32x4>(v3, "cv::v_min(v1, v2)");
    PrintVector<cv::v_int32x4>(v4, "cv::v_max(v1, v2)");
    std::cout << "cv::v_reduce_min(v1) : " << min << std::endl;
    std::cout << "cv::v_reduce_max(v2) : " << max << std::endl;

    //绝对值|a|
    cv::v_int32x4 v5(-1, 1, -2, 0);
    cv::v_uint32x4 v6 = cv::v_abs(v5);// 1 1 2 0
    PrintVector<cv::v_int32x4>(v5, "v5");
    PrintVector<cv::v_uint32x4>(v6, "cv::v_abs(v5)");
    
    //求取差值的绝对值|a - b|
    cv::v_uint32x4 v7 = cv::v_absdiff(v1, v2);
    PrintVector<cv::v_uint32x4>(v7, "cv::v_absdiff(v1, v2)");

    //求平方根
    cv::v_float32x4 v8(1.0f, 4.0f, 9.0f, 16.0f);
    cv::v_float32x4 v9 = cv::v_sqrt(v8);//数据类型只能为float
    PrintVector(v8, "v8");
    PrintVector(v9, "cv::v_sqrt(v8)");

    //求平方和（幅度值）
    cv::v_float32x4 v10(1.0f, 1.0f, 1.0f, 1.0f);
    cv::v_float32x4 v11(1.0f, 1.0f, 1.0f, 1.0f);
    cv::v_float32x4 v12 = cv::v_magnitude(v10, v11);
    cv::v_float32x4 v13 = cv::v_sqr_magnitude(v10, v11);
    PrintVector(v10, "v10");
    PrintVector(v11, "v11");
    PrintVector(v12, "cv::v_magnitude(v10, v11)");
    PrintVector(v13, "cv::v_sqr_magnitude(v10, v11)");

    //求和
    cv::v_int32x4 v14(1, 2, 3, 4);
    int sum = cv::v_reduce_sum(v14);
    PrintVector(v14, "v14");
    std::cout << "cv::v_reduce_sum(v14) : "<< sum << std::endl;
}

兼容OpenMP并行

 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

//结合openmp并行使用
void OpitmizeWithOpenMP()
{
    std::cout << "----------------------OpenCV Simd with OpenMP sample--------------------------" << std::endl;

    //初始化数组
    const size_t  array_size = 40000000;
    int* int_array_a = new int[array_size];
    int* int_array_b = new int[array_size];
    int* int_array_c = new int[array_size];


    //原循环程序计算
    for (size_t i = 0; i < array_size; i++)
    {
        int_array_c[i] = int_array_a[i] * int_array_b[i];
    }

    //Simd优化
    for (size_t i = 0; i < array_size; i += cv::v_int32x4::nlanes)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);
        //计算
        cv::v_int32x4 v_c = v_a * v_b;
        //储存
        cv::v_store(int_array_c + i, v_c);
    }

    //simd + OpenMP优化
    int step = cv::v_int32x4::nlanes;
    #pragma omp parallel for
    for (int i = 0; i < array_size; i += step)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);
        //计算
        cv::v_int32x4 v_c = v_a * v_b;
        //储存
        cv::v_store(int_array_c + i, v_c);
    }
}

条件分支

 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

void ConditionSample()
{
    std::cout << "------------------------------Condition sample-----------------------------------" << std::endl;
    const size_t array_size = 400;
    int* int_array_a = new int[array_size];
    int* int_array_b = new int[array_size];
    
    //原程序
    for (size_t i = 0; i < array_size; i++)
    {
        if (int_array_a[i] > 0)
        {
            int_array_b[i] = int_array_a[i];
        }
    }

    //simd优化版本
    for (size_t i = 0; i < array_size; i += 4)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);

        //计算:构造条件
        auto condiction_mask = v_a > cv::v_setzero_s32();
        // b = mask ? a : b
        v_b = cv::v_select(condiction_mask, v_a, v_b);

        //存储
        cv::v_store(int_array_b + i, v_b);
    }
    
    //simd优化自适应加速内存位宽版本
    for (size_t i = 0; i < array_size; i += cv::v_int32::nlanes)
    {
        //加载
        cv::v_int32 v_a = cv::vx_load(int_array_a + i);
        cv::v_int32 v_b = cv::vx_load(int_array_b + i);
        //计算:构造条件
        auto condiction_mask = v_a > cv::v_setzero_s32();
        // b = mask ? a : b
        v_b = cv::v_select(condiction_mask, v_a, v_b);
        //存储
        cv::vx_store(int_array_b + i, v_b);
    }
    
    //simd + openmp优化版本
#pragma omp parallel for
    for (int i = 0; i < array_size; i += int(cv::v_int32::nlanes))
    {
        //加载
        cv::v_int32 v_a = cv::vx_load(int_array_a + i);
        cv::v_int32 v_b = cv::vx_load(int_array_b + i);
        //计算:构造条件
        auto condiction_mask = v_a > cv::v_setzero_s32();
        // b = mask ? a : b
        v_b = cv::v_select(condiction_mask, v_a, v_b);
        //存储
        cv::vx_store(int_array_b + i, v_b);
    }
}

完整代码

  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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446

/**
 * @file main.cpp
 * @author mango ([email protected])
 * @brief OpenCV夸平台Simd学习
 * @version 0.1
 * @date 2021-10-15
 * 
 * @copyright Copyright (c) 2021
 * 
 */

#include <iostream>
#include <chrono>
#include <random>

#include "opencv2/opencv.hpp"
#include "opencv2/core/simd_intrinsics.hpp"

#include "omp.h"


//测试系统平台信息输出
void PrintTestPlatform()
{
    std::cout << "------------------------test platform informaintion----------------------------"<<std::endl;

    std::string system_name = 
#if defined(__Apple__)
    "Mac Os"
#elif defined(__linux__)
    "Linux"
#else 
    "Windows";
#endif 
    std::cout<< "System : "<< system_name << std::endl;
    std::string compiler =
#if defined(__clang__)
    "clang " + ver_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
#elif defined(__GNUC__)
    "gcc " + ver_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
#else
    "msvc " + std::to_string(_MSC_VER);
#endif

    std::cout<< "Compiler : " << compiler << std::endl;
}


//simd信息获取输出
void PrintOpenCVInfo()
{
    std::cout<<"--------------------------OpenCV informaintion--------------------------"<< std::endl;
    //OpenCV版本，Universal Simd在3.4.x版本就开始提供，不过建议使用4.x版本
    std::cout << "OpenCV version:" << cv::getVersionString() << std::endl;
    std::cout << "Simd info: "<< std::endl;
#ifdef CV_SIMD
    //是否支持simd优化
    std::cout << "CV_SIMD : " << CVAUX_STR(CV_SIMD)  << std::endl;
    //simd优化内存位宽
    std::cout << "CV_SIMD_WIDTH : " <<  CVAUX_STR(CV_SIMD_WIDTH) << std::endl;
    //128bit位宽优化是否支持，绝大部分x86架构cpu支持，如常用的sse指令集
    std::cout << "CV_SIMD128 : " <<CVAUX_STR(CV_SIMD128) << std::endl;
    //256bit位宽优化是否支持，大部分近几年的x86架构cpu支持， avxz指令集
    std::cout << "CV_SIMD256: " <<CVAUX_STR(CV_SIMD256) << std::endl;
    //512bit位宽是否支持，这个intel最先搞得，近几年的intel cpu与最新的AMD锐龙支持
    std::cout << "CV_SIMD512 : " CVAUX_STR(CV_SIMD512) << std::endl;
#else
    std::cout << "CV_SIMD is NOT defined." << std::endl;
#endif

#ifdef CV_SIMD
    std::cout << "sizeof(v_uint8) = " << sizeof(cv::v_uint8) << std::endl;
    std::cout << "sizeof(v_int32) = " << sizeof(cv::v_int32) << std::endl;
    std::cout << "sizeof(v_float32) = " << sizeof(cv::v_float32) << std::endl;
#endif
}

//输出向量
template<typename T>
void PrintVector(const T& simd_vector, const std::string& name)
{
    //获取向量数组长度
    int arr_size = int(T::nlanes);
    //申请对应类型c风格数组
    T::lane_type* arr = new T::lane_type[arr_size];
    //将向量装载到c风格数组
    cv::v_store(arr, simd_vector);

    //打印输出
    std::cout << name << " : ";
    for (size_t i = 0; i < arr_size; i++)
    {
        std::cout << arr[i] << " ";
    }   
    std::cout << std::endl;
}

//加载-计算-存储优化步骤
void SimdSample()
{
    std::cout << "----------------------OpenCV Simd simple sample--------------------------" << std::endl;
    
    // 构造一个常量
    cv::v_int32x4 v_c1(1, 1, 1, 1);             //方法一
    cv::v_int32x4 v_c2 = cv::v_setall(1);       //方法二
    cv::v_int32x4 v_c3 = cv::v_setzero_s32();   //方法三，但只能设置0

    // 构造一个simd长度的数组
    cv::v_int32x4 v_a(1, 2, 3, 4);
    cv::v_int32x4 v_b(1, 2, 3, 4);
    //v_a = [ a0, a1, a2, a3]
    //v_b = [ b0, b1, b2, b3]
    //v_c = [a0+b0, a1+b1, a2+b2, a3+b3]
    cv::v_int32x4 v_c = v_a + v_b;
    PrintVector(v_c, "v_c");
    
    //--------------数组循环优化示例--------------------
    const size_t  array_size = 400;
    int* int_array_a = new int[array_size];
    int* int_array_b = new int[array_size];
    int* int_array_c = new int[array_size];

    //原循环程序计算
    for (int i = 0; i < array_size; i++)
    {
        int_array_c[i] = int_array_a[i] * int_array_b[i];
    }

    //Simd优化:已知编译器向量化内存宽度（如常用的sse为128bit的向量化内存位宽）
    for (int i = 0; i < array_size; i += 4)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);//一次从数组中读取4个int变量存储到v_a
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);//一次从数组中读取4个int变量存储到v_b
        //计算
        cv::v_int32x4 v_c = v_a * v_b;//一次性计算4个变量乘法
        //v_a = [ a0, a1, a2, a3]
        //v_b = [ b0, b1, b2, b3]
        //v_c = [a0*b0, a1*b1, a2*b2, a3*b3]
        //储存
        cv::v_store(int_array_c + i, v_c);
    }

    //Simd自适位宽优化：未知编译器采用的向量化位宽
    for (size_t i = 0; i < array_size; i += cv::v_int32::nlanes)
    {
        //加载 n = 向量化内存位宽 / sizeof(int),如sse是 128 / 32 = 4, avx2 是256 / 32 = 8
        cv::v_int32 v_a = cv::vx_load(int_array_a + i);//一次从数组中读取n个int变量存储到v_a
        cv::v_int32 v_b = cv::vx_load(int_array_b + i);//一次从数组中读取n个int变量存储到v_b
        //计算
        cv::v_int32 v_c = v_a * v_b;//一次性计算n个变量乘法
        //v_a = [ a0, a1, a2, a3]
        //v_b = [ b0, b1, b2, b3]
        //v_c = [a0*b0, a1*b1, a2*b2, a3*b3]
        //储存
        cv::vx_store(int_array_c + i, v_c);
    }
}

//顺组对齐示例，长度不是向量优化位宽时倍数
void LoopLenghtAligned()
{
    const size_t  array_size = 4007;
    uchar* uchar_array_a = new uchar[array_size];
    uchar* uchar_array_b = new uchar[array_size];
    uchar* uchar_array_c = new uchar[array_size];

    //一次性加载16个uchar数据进行计算，uchar占8bit, 8x16=128bit
    size_t i = 0;
    for (; i < array_size - 1; i += 16)
    {
        //加载
        cv::v_uint8x16 v_a = cv::v_load(uchar_array_a + i);
        cv::v_uint8x16 v_b = cv::v_load(uchar_array_b + i);
        //计算...
        cv::v_uint8x16 v_c = v_a + v_b;

        //存储
        cv::v_store(uchar_array_c + i, v_c);
    }
    for (; i < array_size; i++)
    {
        uchar_array_c[i] = uchar_array_a[i] + uchar_array_b[i];
    }
}

//uchar short int float类型数组优化
void BasicDataArrayOptimization()
{
    std::cout << "----------------------OpenCV Simd simple sample--------------------------" << std::endl;

    //uchar
    const size_t  array_size = 4000;
    uchar* uchar_array_a = new uchar[array_size];
    uchar* uchar_array_b = new uchar[array_size];
    uchar* uchar_array_c = new uchar[array_size];

    //一次性加载16个uchar数据进行计算，uchar占8bit, 8x16=128bit
    for (size_t i = 0; i < array_size; i += 16)
    {
        //加载
        cv::v_uint8x16 v_a = cv::v_load(uchar_array_a + i);
        cv::v_uint8x16 v_b = cv::v_load(uchar_array_b + i);
        //计算...
        cv::v_uint8x16 v_c = v_a + v_b;

        //存储
        cv::v_store(uchar_array_c + i, v_c);
    }

    //short
    short* short_arrary_a = new short[array_size];
    short* short_arrary_b = new short[array_size];
    short* short_arrary_c = new short[array_size];
    
    //一次性加载8个short数据进行计算，short占16bit, 16x8=128bit
    for (size_t i = 0; i < array_size; i += 8)
    {
        //加载
        cv::v_int16x8 v_a = cv::v_load(short_arrary_a + i);
        cv::v_int16x8 v_b = cv::v_load(short_arrary_b + i);
        //计算
        cv::v_int16x8 v_c = v_a + v_b;
        //存储
        cv::v_store(short_arrary_c + i, v_c);
    }


    //float
    float* float_array_a = new float[array_size];
    float* float_array_b = new float[array_size];
    float* float_array_c = new float[array_size];
    
    //一次性加载4个float数据进行计算，float占32bit, 32x4=128bit
    for (size_t i = 0; i < array_size; i += 4)
    {
        //加载
        cv::v_float32x4 v_a = cv::v_load(float_array_a + i);
        cv::v_float32x4 v_b = cv::v_load(float_array_a + i);
        //计算
        cv::v_float32x4 v_c = v_a + v_b;
        //存储
        cv::v_store(float_array_c + i, v_c);
    }
}


//if条件分支
void ConditionSample()
{
    std::cout << "------------------------------Condition sample-----------------------------------" << std::endl;
    const size_t array_size = 400;
    int* int_array_a = new int[array_size];
    int* int_array_b = new int[array_size];
    
    //原程序
    for (size_t i = 0; i < array_size; i++)
    {
        if (int_array_a[i] > 0)
        {
            int_array_b[i] = int_array_a[i];
        }
    }

    //simd优化版本
    for (size_t i = 0; i < array_size; i += 4)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);

        //计算:构造条件
        auto condiction_mask = v_a > cv::v_setzero_s32();
        // b = mask ? a : b
        v_b = cv::v_select(condiction_mask, v_a, v_b);

        //存储
        cv::v_store(int_array_b + i, v_b);
    }
    
    //simd优化自适应加速内存位宽版本
    for (size_t i = 0; i < array_size; i += cv::v_int32::nlanes)
    {
        //加载
        cv::v_int32 v_a = cv::vx_load(int_array_a + i);
        cv::v_int32 v_b = cv::vx_load(int_array_b + i);
        //计算:构造条件
        auto condiction_mask = v_a > cv::v_setzero_s32();
        // b = mask ? a : b
        v_b = cv::v_select(condiction_mask, v_a, v_b);
        //存储
        cv::vx_store(int_array_b + i, v_b);
    }
    
    //simd + openmp优化版本
#pragma omp parallel for
    for (int i = 0; i < array_size; i += int(cv::v_int32::nlanes))
    {
        //加载
        cv::v_int32 v_a = cv::vx_load(int_array_a + i);
        cv::v_int32 v_b = cv::vx_load(int_array_b + i);
        //计算:构造条件
        auto condiction_mask = v_a > cv::v_setzero_s32();
        // b = mask ? a : b
        v_b = cv::v_select(condiction_mask, v_a, v_b);
        //存储
        cv::vx_store(int_array_b + i, v_b);
    }
}

//数学函数
void MathematicsFunctionSample()
{
    std::cout << "-----------------------Mathematics function sample------------------------------" << std::endl;
    cv::v_float32 v_float_arr(1.3, 1.4, 1.5, 1.7);

    //向下取整，处理结果为1 1 1 1
    cv::v_int32x4 v_int_arr = cv::v_floor(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_floor process: ");   


    //向上取整，处理结果为2 2 2 2
    v_int_arr = cv::v_ceil(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_ceil process: ");

    //四舍五入取整，处理结果为：1 1 2 2
    v_int_arr = cv::v_round(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_round process: ");

    //截断小数，目前只实现了保留0为小数，所以结果与v_floor一样
    v_int_arr = cv::v_trunc(v_float_arr);
    PrintVector<cv::v_int32x4>(v_int_arr, "(1.3, 1.4, 1.5, 1.7) after v_trunc process: ");


    //最大最小值
    cv::v_int32x4 v1(1, 5, 2, 6);
    cv::v_int32x4 v2(5, 1, 6, 2);
    cv::v_int32x4 v3 = cv::v_min(v1, v2);//1 1 2 2
    cv::v_int32x4 v4 = cv::v_max(v1, v2);//5 5 6 6 
    int min = cv::v_reduce_min(v1);//1
    int max = cv::v_reduce_max(v2);//6
    PrintVector<cv::v_int32x4>(v1, "v1");
    PrintVector<cv::v_int32x4>(v2, "v2");
    PrintVector<cv::v_int32x4>(v3, "cv::v_min(v1, v2)");
    PrintVector<cv::v_int32x4>(v4, "cv::v_max(v1, v2)");
    std::cout << "cv::v_reduce_min(v1) : " << min << std::endl;
    std::cout << "cv::v_reduce_max(v2) : " << max << std::endl;

    //绝对值|a|
    cv::v_int32x4 v5(-1, 1, -2, 0);
    cv::v_uint32x4 v6 = cv::v_abs(v5);// 1 1 2 0
    PrintVector<cv::v_int32x4>(v5, "v5");
    PrintVector<cv::v_uint32x4>(v6, "cv::v_abs(v5)");
    
    //求取差值的绝对值|a - b|
    cv::v_uint32x4 v7 = cv::v_absdiff(v1, v2);
    PrintVector<cv::v_uint32x4>(v7, "cv::v_absdiff(v1, v2)");

    //求平方根
    cv::v_float32x4 v8(1.0f, 4.0f, 9.0f, 16.0f);
    cv::v_float32x4 v9 = cv::v_sqrt(v8);//数据类型只能为float
    PrintVector(v8, "v8");
    PrintVector(v9, "cv::v_sqrt(v8)");

    //求平方和（幅度值）
    cv::v_float32x4 v10(1.0f, 1.0f, 1.0f, 1.0f);
    cv::v_float32x4 v11(1.0f, 1.0f, 1.0f, 1.0f);
    cv::v_float32x4 v12 = cv::v_magnitude(v10, v11);
    cv::v_float32x4 v13 = cv::v_sqr_magnitude(v10, v11);
    PrintVector(v10, "v10");
    PrintVector(v11, "v11");
    PrintVector(v12, "cv::v_magnitude(v10, v11)");
    PrintVector(v13, "cv::v_sqr_magnitude(v10, v11)");

    //求和
    cv::v_int32x4 v14(1, 2, 3, 4);
    int sum = cv::v_reduce_sum(v14);
    PrintVector(v14, "v14");
    std::cout << "cv::v_reduce_sum(v14) : "<< sum << std::endl;
}

//结合openmp并行使用
void OpitmizeWithOpenMP()
{
    std::cout << "----------------------OpenCV Simd with OpenMP sample--------------------------" << std::endl;

    //初始化数组
    const size_t  array_size = 40000000;
    int* int_array_a = new int[array_size];
    int* int_array_b = new int[array_size];
    int* int_array_c = new int[array_size];


    //原循环程序计算
    for (size_t i = 0; i < array_size; i++)
    {
        int_array_c[i] = int_array_a[i] * int_array_b[i];
    }

    //Simd优化
    for (size_t i = 0; i < array_size; i += cv::v_int32x4::nlanes)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);
        //计算
        cv::v_int32x4 v_c = v_a * v_b;
        //储存
        cv::v_store(int_array_c + i, v_c);
    }

    //simd + OpenMP优化
    int step = cv::v_int32x4::nlanes;
    #pragma omp parallel for
    for (int i = 0; i < array_size; i += step)
    {
        //加载
        cv::v_int32x4 v_a = cv::v_load(int_array_a + i);
        cv::v_int32x4 v_b = cv::v_load(int_array_b + i);
        //计算
        cv::v_int32x4 v_c = v_a * v_b;
        //储存
        cv::v_store(int_array_c + i, v_c);
    }
}

int main(int argc, char** argv)
{
    //测试平台信息
    PrintTestPlatform();
    //OpenCV信息
    PrintOpenCVInfo();
    //简单使用OpenCV统一simd框架示例
    SimdSample();
    //集成OpenMP并行优化示例
    OpitmizeWithOpenMP();
    //顺组对齐示例，长度不是向量优化位宽时倍数
    LoopLenghtAligned();
    //基础数据类型数组循环优化示例
    BasicDataArrayOptimization();
    //条件分支示例
    ConditionSample();
    //数学函数操作相关示例
    MathematicsFunctionSample();
    return 0;
}

本文由芒果浩明发布，转载请注明出处。 本文链接：https://blog.mangoeffect.net/opencv/opencv-universal-intrinsics-simd-1.html

微信公众号