For-Loop is not executed

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello!

I have a problem with a simple for-loop:

#include <stdio.h> 
#include <cuda.h>

__global__ void test(int *a) {
 for (int i = 0; i < 4; i++) {
   a[i] = i;
   printf("%d %d\n", threadIdx.x, i);
 }
}

int main(int argc, char **argv) {
  int *a, *b;
  cudaMalloc((void **) &a, 4 * sizeof(int));
  b = (int *) malloc(4 * sizeof(int));
  test <<< 1,1 >>>(a);
  cudaMemcpy(b, a, 4 * sizeof(int), cudaMemcpyDeviceToHost);
  for (int i = 0; i < 4; i++) printf("%d ", b[i]);
  printf("\n");
  cudaFree(a);
  free(b);
  return 0;
}

Basically, the kernel runs on one thread and executes a for loop that runs from 0 to 3 and sets a value in a global array. Also, some info is printed.

I compiled this code with toolkit 4.2 using
nvcc -gencode=arch=compute_20,code=compute_20 -gencode=arch=compute_20,code=sm_20 -lcudart

The output should be

0 0
0 1
0 2
0 3
0 1 2 3

but instead I get

0 0 
0 0 0 0

I tried the program on a GTX 580 and a GTX 550 Ti. The result is the same. Can anybody explain that to me? I thought about his for a long time, but I just don’t get it.

Volker

Edit: I changed the code to include the correct kernel call.

2

Have you called “test(a)” like a cuda kernel? I mean:

test <<< 1,1 >>>(a);

It works for me.

3

Yes, the kernel call is test <<< 1,1 >>>(a). That did not show up in the code. Sorry.

4

Sorry then. This code works for me in a GTX 580, Win7, 32/64 bits. Toolkit 4.2.

0 0
0 1
0 2
0 3
0 1 2 3
5

To get to the bottom of the failures, I would suggest adding error status checks after every CUDA API call and after every kernel launch.

6

Thank you for the replies. The problem was caused by the printf in the kernel loop. I ran this example on Linux with toolkit 4.2 and a GTX580. Removing the printf solved the issue.