我打算使用CUDA构建图像处理。为了表示图像,我使用矩阵(随机生成值)。我想对这个矩阵应用平均滤波器。我使用的过滤器大小是3.这是我写的代码。当数字(N = 10)小于块尺寸大小(BLOCK_DIM = 32)时,这可以正常工作。我试过N = 5和BLOCK_DIM = 3.它工作正常。
为什么当BLOCK_DIM增加时,此代码会导致意外结果(0而不是平均值),我该如何解决?
#include <stdio.h>
#include <stdlib.h>
#define N 10
#define BLOCK_DIM 32
__global__ void averageKernel (int *a, int *c) {
int col = blockIdx.x * blockDim.x + threadIdx.x;
int row = blockIdx.y * blockDim.y + threadIdx.y;
int index = col + row * N;
c[index] = 1;
int sum = 0;
int avg = 0;
if (row > 0 && col > 0 && col < N-1 && row < N-1 ) {
sum = sum + a[index - 1];
sum = sum + a[index + 1];
sum = sum + a[index - N-1];
sum = sum + a[index - N];
sum = sum + a[index - N+1];
sum = sum + a[index + N-1];
sum = sum + a[index + N];
sum = sum + a[index + N+1];
sum = sum + a[index];
avg = sum/9;
}
c[index] = avg;
}
void printMatrix(int a[N][N] )
{
for(int i=0; i<N; i++){
for (int j=0; j<N; j++){
printf("%d\t", a[i][j] );
}
printf("\n");
}
}
int main() {
int a[N][N], c[N][N];
int *dev_a, *dev_c;
int size = N * N * sizeof(int);
for(int i=0; i<N; i++)
for (int j=0; j<N; j++){
a[i][j] = rand() % 256;
}
printf("Matrix A\n");
printMatrix(a);
cudaMalloc((void**)&dev_a, size);
cudaMalloc((void**)&dev_c, size);
cudaMemcpy(dev_a, a, size, cudaMemcpyHostToDevice);
dim3 dimBlock(BLOCK_DIM, BLOCK_DIM);
dim3 dimGrid((N+dimBlock.x-1)/dimBlock.x, (N+dimBlock.y-1)/dimBlock.y);
printf("dimGrid.x = %d, dimGrid.y = %d\n", dimGrid.x, dimGrid.y);
averageKernel<<<dimGrid,dimBlock>>>(dev_a,dev_c);
cudaDeviceSynchronize();
cudaMemcpy(c, dev_c, size, cudaMemcpyDeviceToHost);
printf("Matrix c\n");
printMatrix(c);
cudaFree(dev_a);
cudaFree(dev_c);
}
答案 0 :(得分:2)
您收到“意外结果”,因为您的内核失败并且超出内存访问权限。如果您在代码中添加了error checking和/或使用了cuda-memcheck
,那么您就已经知道了。
问题的根源是这两行:
c[index] = 1;
....
c[index] = avg;
是无条件执行的,当你运行的线程数超过输出矩阵的大小时,它将产生超出内存访问。如果修改内核以便只对输出矩阵范围内的线程执行,那么问题就会消失。