OpenCL 第7课:旋转变换(1)

旋转是一个常用的处理功能。图片中所有的点以某一个点为轴，顺时或逆时方向旋转N个角度。我们利用OpenCL就可以对图片中所有的点进行并行转换，大大提高效率。

上两节中，我们编写了CL文件来传递数组的地址，这一节中我们会多加入几个参数传递。

首先我们先来看下图片旋转的原理。这里我们假设图片的旋转是以图片的中心点为轴。也就是（width/2,height/2）这个点。旋转的角度是任意值。图片旋转会出现这两种情况。

TM截图20130327230153

第一幅是原图，第二幅是旋转了30度，大家可以看到旋转后图片的一部份数据已经超出了原来图片的大小范围。怎么处理超出部份的数据呢。因为我们的图片数据是用数组来存储的，一般我们有两个方法，1、存储图片的数组做大些。2、超出图片部份的数据不显示。这里我们选择后者，方便。

下面我们来看下rotate.cl程序和主程序。

rotate.cl程序

?
__kernel void rotation(__global int* A,
                    __global int* B,
                    int width,
                    int height,
                    float sinangle,
                    float cosangle)
{
    //获取索引号，这里是二维的，所以可以取两个
    //否则另一个永远是0
    int col = get_global_id(0);
    int row = get_global_id(1);
    //计算图形中心点
    float cx = ((float)width)/2;
    float cy = ((float)height)/2;
    int nx = (int)(cx + cosangle * ((float)col-cx) + sinangle * ((float)row-cy));
    int ny = (int)(cy + (-1*sinangle) * ((float)col-cx) + cosangle * ((float)row-cy));
    //边界检测
    if(nx>=0 && nx<width && ny>=0 && ny<height)
    {
        B[nx + ny*width] = A[col + row*width];
    }
}

主程序

?
#include <iostream>
#include <stdio.h>
#include <string.h>
#include <string>
#include <conio.h>
#include <math.h>//数学库
#include <CL/cl.h>//包含CL的头文件
using namespace std;
//8x8数组
const int dim_x = 8;
const int dim_y = 8;
//45度的弧度
const float angle = 3.1415926f/4.0f;
//从外部文件获取cl内核代码
bool GetFileData(const char* fname,string& str)
{
    FILE* fp = fopen(fname,"r");
    if(fp==NULL)
    {
        printf("no found file\n");
        return false;
    }
    while(feof(fp)==0)
    {
        str += fgetc(fp);
    }
    return true;
}
int  main()
{
    //先读外部CL核心代码，如果失败则退出。
    //代码存buf_code里面
    string code_file;
    if(false == GetFileData("rotate.cl",code_file))
    {
        printf("Open rotate.cl error\n");
        return 0;
    }
    char* buf_code = new char[code_file.size()];
    strcpy(buf_code,code_file.c_str());
    buf_code[code_file.size()-1] = NULL;
    //声明CL所需变量。
    cl_device_id device;
    cl_platform_id platform_id = NULL;
    cl_context context;
    cl_command_queue cmdQueue;
    cl_mem bufferA,bufferB;
    cl_program program;
    cl_kernel kernel = NULL;
    //我们使用的是二维向量
    //设定向量大小（维数）
    size_t globalWorkSize[2];
    globalWorkSize[0] = dim_x ;
    globalWorkSize[1] = dim_y;
    cl_int err;
    /*
        定义输入变量和输出变量，并设定初值
    */
    size_t datasize = sizeof(int) * dim_x * dim_y;
    int m,n;
    int buf_A[] = {0,0,0,1,1,0,0,0,
                   0,0,1,0,0,1,0,0,
                   0,1,0,0,0,0,1,0,
                   1,1,1,0,0,1,1,1,
                   0,0,1,0,0,1,0,0,
                   0,0,1,0,0,1,0,0,
                   0,0,1,0,0,1,0,0,
                   0,0,1,1,1,1,0,0,};
//输出数组初始为-1,即该位置不显示数字
    int buf_B[] = {-1,-1,-1,-1,-1,-1,-1,-1,
                   -1,-1,-1,-1,-1,-1,-1,-1,
                   -1,-1,-1,-1,-1,-1,-1,-1,
                   -1,-1,-1,-1,-1,-1,-1,-1,
                   -1,-1,-1,-1,-1,-1,-1,-1,
                   -1,-1,-1,-1,-1,-1,-1,-1,
                   -1,-1,-1,-1,-1,-1,-1,-1,
                   -1,-1,-1,-1,-1,-1,-1,-1};
    //step 1:初始化OpenCL
    err = clGetPlatformIDs(1,&platform_id,NULL);
    if(err!=CL_SUCCESS)
    {
        cout<<"clGetPlatformIDs error:"<<err<<endl;
        return 0;
    }
    //这次我们只用CPU来进行并行运算，当然你也可以该成GPU
    clGetDeviceIDs(platform_id,CL_DEVICE_TYPE_GPU,1,&device,NULL);
    //step 2:创建上下文
    context = clCreateContext(NULL,1,&device,NULL,NULL,NULL);
    //step 3:创建命令队列
    cmdQueue = clCreateCommandQueue(context,device,0,NULL);
    //step 4:创建数据缓冲区
    bufferA = clCreateBuffer(context,
                             CL_MEM_READ_ONLY,
                             datasize,NULL,NULL);
    bufferB = clCreateBuffer(context,
                             CL_MEM_WRITE_ONLY,
                             datasize,NULL,NULL);
    //step 5:将数据上传到缓冲区
    clEnqueueWriteBuffer(cmdQueue,
                         bufferA,CL_FALSE,
                         0,datasize,
                         buf_A,0,
                         NULL,NULL);
    //step 6:加载编译代码,创建内核调用函数
    program = clCreateProgramWithSource(context,1,
                                        (const char**)&buf_code,
                                        NULL,NULL);
    clBuildProgram(program,1,&device,NULL,NULL,NULL);
    kernel = clCreateKernel(program,"rotation",NULL);
    //step 7:设置参数，执行内核
    float sinangle = sinf(angle);
    float cosangle = cosf(angle);
    clSetKernelArg(kernel,0,sizeof(cl_mem),&bufferA);
    clSetKernelArg(kernel,1,sizeof(cl_mem),&bufferB);
    clSetKernelArg(kernel,2,sizeof(cl_int),&dim_x);
    clSetKernelArg(kernel,3,sizeof(cl_int),&dim_y);
    clSetKernelArg(kernel,4,sizeof(cl_float),&sinangle);
    clSetKernelArg(kernel,5,sizeof(cl_float),&cosangle);
    //注意这里第三个参数已经改成2，表示二维数据。
    clEnqueueNDRangeKernel(cmdQueue,kernel,
                           2,NULL,
                           globalWorkSize,
                           NULL,0,NULL,NULL);
    //step 8:取回计算结果
    clEnqueueReadBuffer(cmdQueue,bufferB,CL_TRUE,0,
                        datasize,buf_B,0,NULL,NULL);
    //输出计算结果
    for(n=0;n<dim_x;n++)
    {
        for(m=0;m<dim_y;m++)
        {
            if(buf_A[m+dim_x*n]==0)
                cout<<"  ";
            else
                cout<< buf_A[m+dim_x*n] <<" ";
        }
        cout<<endl;
    }
    cout<<endl<<"====Rotate 45===="<<endl<<endl;
    for(n=0;n<dim_x;n++)
    {
        for(m=0;m<dim_y;m++)
        {
            if(buf_B[m+dim_x*n]<=0)
                cout<<"  ";
            else
                cout<< buf_B[m+dim_x*n] <<" ";
        }
        cout<<endl;
    }
    //释放所有调用和内存
    clReleaseKernel(kernel);
    clReleaseProgram(program);
    clReleaseCommandQueue(cmdQueue);
    clReleaseMemObject(bufferA);
    clReleaseMemObject(bufferB);
    clReleaseContext(context);
    delete buf_code;
    return 0;
}