Sequential call of kernels

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

Another simple question:

In sequential call of different kernels (Grid dim and block dim are different), I observed there is serious slow-down.

For example,

Kernel 1 only => 0.001 sec
Kernel 2 only => 0.130 sec

But kernel1 + kernel2 would take 0.450 sec.

Can I have some advice on this?

Thanks.

2

You probably make a mistake when timing your kernels. Kernel calls are asynchronous so they return immediately after you called them. Use cudaThreadSynchronize() before starting and before stopping the timer to get accurate results.

3

You should do it in following way:

cudaThreadSyncronize();

t1 = clock();

kernel1<<>>();

cudaThreadSYncronize();

t2=clock();

kernel2<<>>();

cudaThreadSyncronize();

t3=clock();

Replace clock() with appropriate timing function (such as QueryPerformanceCounter() on WIndows). kernel1 execution time will be (t2-t1) and kernel2 execution time will be (t3-t2).