OpenCL: Blur image with Gaussian filter, 高斯模糊

Question

我计算了 n = 5 的高斯滤波器掩码 nxn。看起来像这样：

int gauss5[] = { 
    1, 4, 7, 4, 1,  
    4, 20, 33, 20, 4,
    7, 33, 55, 33, 7,
    4, 20, 33, 20, 4,
    1, 4, 7, 4, 1
};

int gauss5_summ = 331;

在opencl中计算后图像变暗。是什么原因？

size_t global_wblur[2];
size_t local_wblur[2];

global_wblur[0] = h;
global_wblur[1] = w;

local_wblur[0] = local_wblur[1] = 32;

err = clEnqueueNDRangeKernel(queue, cl_img_gaussian_blur, 2, NULL, global_wblur, local_wblur, 0, NULL, NULL);

if (err != CL_SUCCESS) {
   fprintf(stderr, "error: clEnqueueNDRangeKernel() blur %d %s\n", err, cl_strerror(err));
   exit(EXIT_FAILURE);
}

内核源代码：

 __kernel void cl_img_gaussian_blur(__global const uchar *gray, __global uchar *out, __global const uchar *gbox, uint n, uint sum, uint w, uint h)
{
    int i, j, offset;
    uint x, y, summ;

    y = get_global_id(0);
    x = get_global_id(1);

    offset = n/2;

    /* ignore border pixels 
     */
    if (y - offset < 0 || y + offset > h || x - offset < 0 || x + offset > w) {
        out[y*w + x] = gray[y*w + x];
        return;
    }

    summ = 0;

    for (j = -offset; j <= offset; j++) {
        for (i = -offset; i <= offset; i++) {
            summ += gray[(y + j)*w + x + i]*gbox[(j + offset)*n + i + offset];
        }
    }

    out[y*w + x] = summ/sum;
}

模糊图像：

深色(opencl) -- https://github.com/apetrunev/imgalg-opencl/blob/master/dark.png

正常 -- https://github.com/apetrunev/imgalg-opencl/blob/master/test-blur/out.png

Answer 1

你说 gbox 数据初始化如下：

clEnqueueWriteBuffer(queue, gauss_buf, CL_FALSE, 0, 5*5, gauss5, 0, NULL, NULL);

这是错误的，因为您正在复制实际内存量的 1/4。正确的方法是：

clEnqueueWriteBuffer(queue, gauss_buf, CL_FALSE, 0, 5*5*sizeof(cl_int), gauss5, 0, NULL, NULL);

否则，其余为0，导致输出值偏低。

OpenCL: Blur image with Gaussian filter, 高斯模糊

OpenCL: Blur image with Gaussian filter, Gaussian blur

gaussian

opencl