Timing CUDA operations

benchmarking, c, cuda

Solution

You could do something along the lines of :

#include <sys/time.h>

struct timeval t1, t2;

gettimeofday(&t1, 0);

kernel_call<<<dimGrid, dimBlock, 0>>>();

HANDLE_ERROR(cudaThreadSynchronize();)

gettimeofday(&t2, 0);

double time = (1000000.0*(t2.tv_sec-t1.tv_sec) + t2.tv_usec-t1.tv_usec)/1000.0;

printf("Time to generate:  %3.1f ms \n", time);

or:

float time;
cudaEvent_t start, stop;

HANDLE_ERROR( cudaEventCreate(&start) );
HANDLE_ERROR( cudaEventCreate(&stop) );
HANDLE_ERROR( cudaEventRecord(start, 0) );

kernel_call<<<dimGrid, dimBlock, 0>>>();

HANDLE_ERROR( cudaEventRecord(stop, 0) );
HANDLE_ERROR( cudaEventSynchronize(stop) );
HANDLE_ERROR( cudaEventElapsedTime(&time, start, stop) );

printf("Time to generate:  %3.1f ms \n", time);

Problem

I need to time a CUDA kernel execution. The Best Practices Guide says that we can use either events or standard timing functions like `clock()` in Windows. My problem is that using these two functions gives me a totally different result. In fact, the result given by events seems to be huge compared to the actual speed in practice. What I actually need all this for is to be able to predict the running time of a computation by first running a reduced version of it on a smaller data set. Unfortunately, the results of this benchmark are totally unrealistic, being either too optimistic (`clock()`) or waaaay too pessimistic (events).

Original source

Related problems