Question

我有以下内核

   __global__ void filter(unsigned char *image, unsigned char *out, int n, int m)
    {
        int x = blockIdx.x * blockDim.x + threadIdx.x;
        int y = blockIdx.y * blockDim.y + threadIdx.y;
        int offset = x + y * blockDim.x * gridDim.x;
        int sumx, sumy, sumz, k, l;

        __shared__ float shared[16][16];

        shared[threadIdx.x][threadIdx.y] = image[offset];
            out[offset] = shared[threadIdx.x][threadIdx.y]; 

    }

我称之为filter<<<dimGrid, dimBlock>>>(dev_image, dev_out, n, m);。

奇怪的是，即使我对内核的调用进行注释并进行编译，图像也保持不变。知道为什么会这样吗？是不是释放了gpu上的内存？

void Draw()
{
    unsigned char *image, *out;
    int n, m;
    unsigned char *dev_image, *dev_out;
    image = readppm("maskros512.ppm", &n, &m);
    out = (unsigned char*) malloc(n*m*3);
    printf("%d %d\n",n,m );
    cudaMalloc( (void**)&dev_image, n*m*3);
    cudaMalloc( (void**)&dev_out, n*m*3);
    cudaMemcpy( dev_image, image, n*m*3, cudaMemcpyHostToDevice);
    dim3 threads( 1, 256 );
    dim3 blocks( 32, 32 );
    filter<<<blocks, threads>>>(dev_image, dev_out, n, m);
    cudaMemcpy( out, dev_out, n*m*3, cudaMemcpyDeviceToHost );
    cudaFree(dev_image);
    cudaFree(dev_out);
    glClearColor( 0.0, 0.0, 0.0, 1.0 );
    glClear( GL_COLOR_BUFFER_BIT );
    glRasterPos2f(-1, -1);
    glDrawPixels( n, m, GL_RGB, GL_UNSIGNED_BYTE, image );
    glRasterPos2i(0, -1);
    glDrawPixels( n, m, GL_RGB, GL_UNSIGNED_BYTE, out );
    glFlush();
}

Answer 1

如果您只是注释掉filter行，则无法填充dev_out。因此，如果您将dev_out复制到out，那么您将获得垃圾，这可能是最后一次dev_out中的任何内容。

这些行不对：

dim3 threads( 1, 256 );
dim3 blocks( 32, 32 );

您正在启动线程块，该线程块是x中的1个线程，y中包含256个线程。这对你的内核没有意义。你的内核需要每个像素启动一个线程，并且它期望x和y中的足够的线程数组以像素为单位覆盖图像空间。此外，您的共享内存分配期望一个16x16的线程块。试试这个：

dim3 threads(16,16);
dim3 blocks((n+threads.x-1)/threads.x, (m+threads.y-1)/threads.y);

此外，您的图像似乎由3字节像素组成。但是你每个像素只启动一个线程。因此，您需要复制每个像素3个字节，而不是一个。像这样：

#define RED 0
#define GRN 1
#define BLU 2

__global__ void filter(unsigned char *image, unsigned char *out, int n, int m)
{
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    int y = blockIdx.y * blockDim.y + threadIdx.y;
    int offset = x + y * blockDim.x * gridDim.x;
    // the above numbers are all pixel dimensions.  To convert to byte dimensions, 
    // we must multiply by 3
    int sumx, sumy, sumz, k, l;

    __shared__ unsigned char shared[16][16*3];

    shared[threadIdx.x][(threadIdx.y*3)+RED] = image[(offset*3)+RED]; // pick up red
    shared[threadIdx.x][(threadIdx.y*3)+GRN] = image[(offset*3)+GRN]; // pick up green
    shared[threadIdx.x][(threadIdx.y*3)+BLU] = image[(offset*3)+BLU]; // pick up blue
    out[(offset*3)+RED] = shared[threadIdx.x][(threadIdx.y*3)+RED]; 
    out[(offset*3)+GRN] = shared[threadIdx.x][(threadIdx.y*3)+GRN]; 
    out[(offset*3)+BLU] = shared[threadIdx.x][(threadIdx.y*3)+BLU]; 
}

最后你应该做正确的cuda error checking

用cuda进行图像处理，图像不会改变

1 个答案: