Variables persisting after program executes

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, I’m running the code below on a tesla node in our cluster, and I’m finding that every time I run it, the variables C and D persist after the run. running it multiple times causes the variables to just keep incrementing. This happens even if I compile the code into two different binaries, say ‘test1’ and ‘test2’. Running test2 will still increment the same variables that test1 used. This seems very odd to me. Anyone have a good guess at what’s going on here?

Thanks,

Paul

[codebox]/*

  • functor.cu

  • functor

*/

#include <stdio.h>

device struct myFex {

__device__ void operator()(int *bob, float *joe, int art, float big) {

*bob += art;

*joe += big;

}

};

global void Bill(int* A, float* B, int C, float D){

myFex bogoFex;

bogoFex(A,B,C,D);

}

int main (int argc, char * const argv) {

// insert code here...

printf("Hello World\n");

int *A, C;

float *B, D;

cudaMalloc((void**)&A,sizeof(int));

cudaMalloc((void**)&B,sizeof(float));

int* Ah;

float* Bh;

cudaMallocHost((void**)&Ah,sizeof(int));

cudaMallocHost((void**)&Bh,sizeof(float));



C = 5;

D= 7.2;

for(int itr = 0; itr <20; itr++){

Bill<<<1,1>>>(A,B,C,D);

cudaMemcpy(Ah,A,sizeof(int),cudaMemcpyDeviceToHost);

cudaMemcpy(Bh,B,sizeof(float),cudaMemcpyDeviceToHost);

printf("%i\t%f\n",*Ah,*Bh);

}

cudaFree(A);

cudaFree(B);

cudaFreeHost(Ah);

cudaFreeHost(Bh);

return 0;

}

[/codebox]

2

It sounds like you mean A and B are persisting, since C and D are just local variables in your host code.

Since you are not initializing the memory you cudaMalloc() on the device, you are getting whatever data is left there in GPU memory from the previous run. As it happens, the allocation is predictable enough, you get the same memory location in every run.

3

I did some more fiddling with the code, and it looks like you’re exactly right. Thanks for the quick reply. :)

Paul