Question

I have a cuda code which performs calculation on GPU.
I am using clock(); to find out timings

My code structure is

__global__ static void sum(){

// calculates sum 
}

extern "C"
int run_kernel(int array[],int nelements){
 clock_t start, end;
  start = clock();
  //perform operation on gpu - call sum
 end = clock();
 double elapsed_time = ((double) (end - start)) / CLOCKS_PER_SEC;
 printf("time required : %lf", elapsed_time);
}

But the time is always 0.0000
I checked printing start and end time. Start has some value but end time is always zero.

Any idea what might be the cause? Any alternatives to measure time.

Any help would be appreciated.

Thanks

Answer 1

There are two problems here:

1. The clock() function has too low resolution to measure the duration of the event you are trying to time
2. The CUDA kernel launch is an asynchronous operation, so it consumes almost no time (typically 10-20 microseconds on a sane platform). Unless you use a synchronous CUDA API call to force the host CPU to block until the kernel finishes running, you are not going to be measuring the execution time.

CUDA has its own high precision timing API, and it is the recommended way to time operations which run on the GPU. The code to use it would look something like this:

int run_kernel(int array[],int nelements){

    cudaEvent_t start,stop;
    cudaEventCreate(&start);
    cudaEventCreate(&stop);

    cudaEventRecord(start, 0);

    //
    //perform operation on gpu - call sum
    //

    cudaEventRecord(stop, 0); 
    cudaEventSynchronize(stop); 
    float elapsedTime; 
    cudaEventElapsedTime(&elapsedTime, start, stop); 
    printf("time required : %f", elapsed_time); 

    cudaEventDestroy(start);
    cudaEventDestroy(stop);
}