CUDA侵蚀算法

时间:2014-05-21 23:03:09

标签: c++ image-processing cuda

我是CUDA的新手,我正在尝试用结构元素3x3开发简单(天真)的侵蚀算法。至于现在,我已经开发了一个代码(它基于nVidia presentation):

#define bx (blockIdx.x)
#define by (blockIdx.y)
#define bdx (blockDim.x)
#define bdy (blockDim.y)
#define tx (threadIdx.x)
#define ty (threadIdx.y)
#define max( a, b ) ( ((a) > (b)) ? (a) : (b) )
#define min( a, b ) ( ((a) < (b)) ? (a) : (b) )

#define TILE_H 16
#define TILE_W 16
#define D 3    //structural element diameter
#define R 1    //structural element radius
#define BLOCK_W (TILE_W+D-1)
#define BLOCK_H (TILE_H+D-1)

__global__ void erosion(int *picture, unsigned int width, unsigned int height)
{    
    __shared__ int pixels[BLOCK_W*BLOCK_H];
    int x = bx*TILE_W + tx - R;
    int y = by*TILE_H + ty - R;

    x = max(0, x);
    x = min(x, (int)width-1);
    y = max(y,0);
    y = min(y, (int)height-1);

    unsigned int idx = y*width + x;
    unsigned int bidx = ty*bdy+tx;
    pixels[bidx] = picture[idx];
    __syncthreads();

    //compute pixels inside apron
    if (tx>=R && tx<BLOCK_W-R && ty>=R && ty < BLOCK_H-R)
    {
    //erode
    if (pixels[bidx] == 1)
        picture[idx] = pixels[ty*bdy+(tx+1)] & pixels[ty*bdy+(tx-1)] & pixels[(ty+1)*bdy+tx] & pixels[(ty-1)*bdy+tx];
    }
}

main()功能:

int main()
{
    //...    
    int *pixels;
    int img_width=M; int img_height=N;
    cudaMemcpy(dev_pixels, pixels, M*N*sizeof(int), cudaMemcpyHostToDevice);

    dim3 blocks(img_width/BLOCK_W, img_height/BLOCK_H);
    erosion<<<blocks, D*D>>>(dev_pixels, img_width, img_height);

    cudaMemcpy(output, dev_pixels, M*N*sizeof(int), cudaMemcpyDeviceToHost);
}

我的问题是:似乎erosion()永远不会到达if语句,我想在那里计算停机坪内的像素。你碰巧知道为什么会这样吗?我已经排除img_widht/BLOCK_W分区(它可以返回0值,但目前我已修复img_widht=54img_height=36)。

1 个答案:

答案 0 :(得分:2)

您正在启动一个内核,其网格由一个2D数组块组成,每个块都有一个 1D线程数组

dim3 blocks(img_width/BLOCK_W, img_height/BLOCK_H); // creates 2D blocks variable
erosion<<<blocks, D*D>>>(dev_pixels, img_width, img_height);
           ^       ^
           |       |
           |       1D array of threads
           2D array of blocks

由于您的threadblock是一维线程数组,threadIdx.y始终为零(对于每个块中的每个线程)。因此ty始终为零,此if-test始终失败:

if (tx>=R && tx<BLOCK_W-R && ty>=R && ty < BLOCK_H-R)

因为ty(==0)永远不会大于或等于R(==1)

您可以通过定义适当的dim3数量来启动每个块中的2D线程数组:

dim3 threads(D,D);

并在内核配置中传递:

erosion<<<blocks, threads>>>(dev_pixels, img_width, img_height);

我不能说这对你的其他代码是否合理,但是通过这种修改,我可以说你的if语句的内部(正文)将会被触及。