how to compute time on device

Accelerated Computing CUDA CUDA Programming and Performance
1

hi,
if I compute time on device ,can I do this as following? Firstly I get the GPU clock with clock() function,and then I divide the value that the clock() return by GPU runtime frequency.is the method right?and how can I get the GPU runtime frequency,is it the GPU speed clock of the SDK ?

2

If you want to find the time to complete some kernel use this lines:

float gputime;

    cudaEvent_t start,stop;

    cudaEventCreate(&start);

    cudaEventCreate(&stop);

cudaEventRecord(start,0);

// kernel calls 

cudaEventRecord(stop,0);

    cudaEventSynchronize(stop);

    cudaEventElapsedTime(&gputime,start,stop);

cudaEventDestroy(start);

    cudaEventDestroy(stop) ;   

    printf(" \n");

printf("Time = %g \n",  gputime/1000.0f);  

printf(" \n");
3

But I think you can use cudaEvent_t only in the host code. Am I right?
If you want to compute different execution time inside the kernel, you may use clock(). Does clock compute CPU or GPU time?

4

You can get the clock frequency in kilohertz from the [font=“Courier New”]cudaDeviceProp.clockRate[/font] as returned by [font=“Courier New”]cudaGetDeviceProperties()[/font].

5

Yes. The code above it was only for calculating how much time does it take to compute the whole kernel and it was in host code.