GPU results related to printf() GPU bug?

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi!

I run into a strange case where I use GeForce 8800 GT to perform inner product of two vectors. If I add a printf() right before I use the results, it works fine. But if I don’t add printf(), the result is wrong.

float xTy (float *x, float *y, int n)
{
float gpu = GPU_inner (float *x, float *y, int n);

    printf ("gpu = %g\n, gpu); // remove this line will make
                                           // value of next f(gpu) wrong

    f (gpu) ...             // gpu value is correct only if above printf is present

}

First I suspect I didn’t synchronize the return data from GPU, then I realize the gpu value has been saved, before the printf statment. I also tried to substitute the printf statement with some busy loop, but no matter how much delay I add in between, it won’t make any difference. The only difference is a printf statement, which I suspect has to do with the graphics card.

Thanks,
Peter

#2

Try running in debug mode and see if you get any asserts or other errors. If you are running in linux, try valgrind.

I’ve seen this kind of error occur when writing past the end of an array in a seemingly unrelated part of the application.