OpenCL 图像卷积 3 使用 CPU

▶ CPU 图像卷积，共四种方法。分别为基本串行，使用模板，使用局部内存，使用AVX指令优化

● 全部的代码，仅在主函数中选择调用的函数名即可。

 #include <stdio.h>
 #include <stdlib.h>
 #include <time.h>
 #include <opencv2/opencv.hpp>
 
 const char *inputFile = "R:/1.png";
 const char *outputFile = "R:/2.png";
 
 bool floatEq(const float a, const float b)// 相等返回 1
 {
     if (b == )
         return fabs(a) < 0.001;
     return fabs(a / b - ) < 0.001;
 }
 
 void convolution01(const float *input, float *output, const int inputRow, const int inputCol,
     const float *filter, const int filterWidth)
 {
     const int halfFilterWidth = filterWidth / ;
     int row, col, rr, cc;
     float sum;
 
     //memset(output, 0, sizeof(float) * inputRow * inputCol);             // 使用 memset 将 output 全部凃成 0
 
 #pragma omp parallel for num_threads(8) default(none) shared(halfFilterWidth, output, inputRow, inputCol) private(row, col)
     for (row = ; row < halfFilterWidth; row++)                         // 人工将边角涂成 0
     {
         for (col = ; col < inputCol; col++)
             output[row*inputCol + col] = output[(inputRow -  - row)*inputCol + col] = ;
     }
 #pragma omp parallel for num_threads(8) default(none) shared(halfFilterWidth, output, inputRow, inputCol) private(row, col)
     for (row = halfFilterWidth; row < inputRow - halfFilterWidth; row++)
     {
         for (col = ; col < halfFilterWidth; col++)
             output[row*inputCol + col] = output[row*inputCol + inputCol -  - col] = ;
     }
 
 #pragma omp parallel for num_threads(8) default(none) shared(halfFilterWidth, input, output, inputRow, inputCol, filter) private(row, col, rr, cc, sum)
     for (row = halfFilterWidth; row < inputRow - halfFilterWidth; row++)// 内部计算部分
     {
         for (col = halfFilterWidth; col < inputCol - halfFilterWidth; col++)
         {
             for (sum = 0.0f, rr = -halfFilterWidth; rr <= halfFilterWidth; rr++)
             {
                 for (cc = -halfFilterWidth; cc <= halfFilterWidth; cc++)
                     sum += filter[(rr + halfFilterWidth) * filterWidth + cc + halfFilterWidth] * input[(row + rr)*inputCol + col + cc];
             }
             output[row * inputCol + col] = sum;
         }
     }
     /*
     for (row = 0; row < inputRow; row++)                                // 全范围循环，在最里层判断
     {
         for (col = 0; col < inputCol; col++)
         {
             if (row < halfFilterWidth || row >= inputRow - halfFilterWidth || col < halfFilterWidth || col >= inputCol - halfFilterWidth)
             {
                output[row * inputCol + col] = 0;
                 continue;
             }
             for (sum = 0.0f, rr = -halfFilterWidth; rr <= halfFilterWidth; rr++)
             {
                 for (cc = -halfFilterWidth; cc <= halfFilterWidth; cc++)
                     sum += filter[(rr + halfFilterWidth) * filterWidth + cc + halfFilterWidth] * input[(row + rr)*inputCol + col + cc];
             }
             output[row * inputCol + col] = sum;
 
         }
     }
     */
 }
 
 template<int filterWidth>
 void convolution02(const float *input, float *output, const int inputRow, const int inputCol, const float *filter)// 卷积宽度作为模板
 {
     const int halfFilterWidth = filterWidth / ;
     int row, col, rr, cc;
     float sum;
 
     memset(output, , sizeof(float) * inputRow * inputCol);
 #pragma omp parallel for num_threads(8) default(none) shared(halfFilterWidth, input, output, inputRow, inputCol, filter) private(row, col, rr, cc, sum)
     for (row = halfFilterWidth; row < inputRow - halfFilterWidth; row++)
     {
         for (col = halfFilterWidth; col < inputCol - halfFilterWidth; col++)
         {
             for (sum = 0.0f, rr = -halfFilterWidth; rr <= halfFilterWidth; rr++)
             {
                 for (cc = -halfFilterWidth; cc <= halfFilterWidth; cc++)
                     sum += filter[(rr + halfFilterWidth) * filterWidth + cc + halfFilterWidth] * input[(row + rr)*inputCol + col + cc];
             }
             output[row * inputCol + col] = sum;
         }
     }
 }
 
 template<int filterWidth, int blockRow, int blockCol>
 void convolution03(const float *input, float *output, const int inputRow, const int inputCol, const float *filter)// 使用局部内存块
 {
     const int halfFilterWidth = filterWidth / ;
     int row, col, rr, cc, i, j;
     float filterElement;
 
     if (inputRow % blockRow || inputCol % blockCol) // 要求图片长宽为局部内存块的整数倍
     {
         printf("Failed, outputRow %% blockRow || outputCol %% blockCol\n");
         return;
     }
 
     memset(output, , sizeof(float) * inputRow * inputCol);
 #pragma omp parallel for num_threads(8)
     for (row = halfFilterWidth; row < inputRow - halfFilterWidth; row += blockRow)
     {
         for (col = halfFilterWidth; col < inputCol - halfFilterWidth; col += blockCol)
         {
             float sum[blockRow * blockCol] = { 0.0f };
             for (rr = -halfFilterWidth; rr <= halfFilterWidth; rr++)
             {
                 for (cc = -halfFilterWidth; cc <= halfFilterWidth; cc++)
                 {
                     filterElement = filter[(rr + halfFilterWidth) * filterWidth + cc + halfFilterWidth];
                     for (i = ; i < blockRow; i++)
                     {
                         for (j = ; j < blockCol; j++)
                         {
                             if (row + rr + i >= inputRow || col + cc + j >= inputCol)
                                 break;
                             sum[i * blockCol + j] += filterElement * input[(row + rr + i) * inputCol + col + cc + j];
                         }
                     }
                 }
             }
             for (i = ; i < blockRow; i++)
             {
                 for (j = ; j < blockCol; j++)
                 {
                     if (row + i >= inputRow || col + j >= inputCol)
                         continue;
                     output[(row + i) * inputCol + col + j] = sum[i * blockCol + j];
                 }
             }
         }
     }
 }
 
 template<int filterWidth, int blockRow, int blockCol>
 void convolution04(const float *input, float *output, const int inputRow, const int inputCol, const float *filter)// 使用 AVX 指令扩展
 {
     const int halfFilterWidth = filterWidth / ;
     int row, col, rr, cc, i, j;
 
     if (inputRow % blockRow || inputCol % (blockCol * ))
     {
         printf("Failed, inputRow %% blockRow || inputCol %% blockCol\n");
         return;
     }
 
     memset(output, , sizeof(float) * inputRow * inputCol);
 #pragma omp parallel for num_threads(8)
     for (row = halfFilterWidth; row < inputRow - halfFilterWidth; row += blockRow)
     {
         for (col = halfFilterWidth; col < inputCol - halfFilterWidth; col += blockCol * )
         {
             __m256 sum[blockRow * blockCol] = {_mm256_setzero_ps()};
             for (rr = -halfFilterWidth; rr <= halfFilterWidth; rr++)
             {
                 for (cc = -halfFilterWidth; cc <= halfFilterWidth; cc++)
                 {
                     __m256 filterElement = _mm256_broadcast_ss(filter + (rr + halfFilterWidth) * filterWidth + cc + halfFilterWidth);
                     for (i = ; i < blockRow; i++)
                     {
                         for (j = ; j < blockCol; j++)
                         {
                             //if (row + rr + i >= inputRow || col + cc + j * 8 >= inputCol)// 在局部内存块较大时需要越界检查
                             //    continue;
                             __m256 imageElement = _mm256_loadu_ps(input + (row + rr + i)*inputCol + col + cc + j * );
                             sum[i * blockCol + j] = _mm256_fmadd_ps(filterElement, imageElement, sum[i * blockCol + j]);
                         }
                     }
                 }
             }
             for (i = ; i < blockRow; i++)
             {
                 for (j = ; j < blockCol; j++)
                 {
                     //if (row + i >= inputRow || col + j * 8 >= inputCol)
                     //    continue;
                     _mm256_storeu_ps(output + (row + i)*inputCol + col + j * , sum[i * blockCol + j]);
                 }
             }
         }
     }
 }
 
 int main()
 {
     int i, k;
     clock_t time;
     float filterSum;
 
     // 卷积窗口相关
     const int filterWidth = , filterSize = filterWidth * filterWidth, halfFilterWidth = filterWidth / ;
     float filter[filterSize] =
     {// 模糊窗口
         ,       ,       ,       , ,
         , .f / , .f / , .f / , ,
         , .f / , .f / , .f / , ,
         , .f / , .f / , .f / , ,
         ,       ,       ,       ,
     };
     for (filterSum = 0.0f, i = ; i < filterSize; filterSum += filter[i++]);
     if (!floatEq(filterSum, ))// 非归零的卷积窗口（如模糊）需要归一化
         for (i = ; i < filterSize; filter[i] /= filterSum, i++);
 
     // 图片相关
     cv::Mat input = cv::imread(inputFile), output = input, channel[];
     cv::split(input, channel);
     const int inputRow = input.rows, inputCol = input.cols, inputDataSize = inputRow * inputCol;
     float *inputData = (float*)malloc(sizeof(float) * inputDataSize);
     float *outputData = (float*)malloc(sizeof(float) * inputDataSize);
 
     for (k = ; k < ; k++)// 三个通道，分别为蓝、绿、红
     {
         for (i = ; i < inputRow * inputCol; inputData[i] = (float)channel[k].data[i], i++);
         time = clock();
         convolution01(inputData, outputData, inputRow, inputCol, (const float *)filter, filterWidth);
         //convolution02<filterWidth>(inputData, outputData, inputRow, inputCol, filter);
         //convolution03<filterWidth, 4, 4>(inputData, outputData, inputRow, inputCol, filter);
         //convolution04<filterWidth, 4, 4>(inputData, outputData, inputRow, inputCol, filter);
         time = clock() - time;
         printf("Time for channel[%d]:%d ms\n", k, time);
         for (i = ; i < inputRow * inputCol; channel[k].data[i] = (unsigned char)outputData[i], i++);
     }
 
     cv::merge(channel, , output);
     cv::imwrite(outputFile, output);
     //imshow("merge", output);
     //cv::waitKey(0);
     free(inputData);
     free(outputData);
     printf("\nFinished.\n");
     getchar();
     return ;
 }

● 输出结果，使用一张 4608 × 6656 的图片（bmp87.7MB）进行测试，使用主函数中那个边长为5、实际窗口长度为 3 的均值窗口。图片太大喘不上来，偷梁换柱成小图看效果

　　　　　

■ 计时结果

// convolution01，memset + 内部计算，无 OpenMP
Time for channel[]: ms
Time for channel[]: ms
Time for channel[]: ms
 
Finished.
 
// convolution01，手动除边 + 内部计算，无 OpenMP
Time for channel[]: ms
Time for channel[]: ms
Time for channel[]: ms
 
Finished.
 
// convolution01，循环内判断，无 OpenMP
Time for channel[]: ms
Time for channel[]: ms
Time for channel[]: ms
 
Finished.
 
// convolution01，手动除边 + 内部计算，有 OpenMP
Time for channel[]: ms
Time for channel[]: ms
Time for channel[]: ms
 
Finished.
 
// convolution02，有 OpenMP
Time for channel[]: ms
Time for channel[]: ms
Time for channel[]: ms
 
Finished.
 
// convolution03<filterWidth, 4, 4>，无 OpenMP
Time for channel[]: ms
Time for channel[]: ms
Time for channel[]: ms
 
Finished.
 
// convolution04<filterWidth, 4, 4>，无 OpenMP
Time for channel[]: ms
Time for channel[]: ms
Time for channel[]: ms
 
Finished.

■ 没法给 convolution03 和 convolution04 加 OpenMP，一加就各种内存冲突，便捷判断都挡不住。

■