Is it possible?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, everyone.

I would like to know is it possible bellow explained algorithm in CUDA:

Given N points, I give each the cuda threads one point and have programming to find the local maxima about something object function in each thread.

Namely,

mycuda_kernelfunction(...) {

while(1) {

 Â  Â  1. given one point in this thread, find the next point to minimize the object function.

 Â  Â  2....N-1. something jobs...

 Â  Â  N. stop the while loop if some condition is true, otherwise continue.

}

Is it possible above the codes…

In this above codes, each threads will run respective difference(not equal) step. For example, some thread is processing at 2 step, another one is processing at 14 step, some of the others is processing at Nth step.

As I understood, SIMD is to be one single instructment multiple data, so the group threads in SMs(or warp) must be processing at the same steps. I am confusing these words and the concept of thread programming model in CUDA.

I think the above code is impossible under threads of the CUDA.

Is it correct? Help me, please!!

Thank you for reading my question.

2

I think it is possible for different thread having different steps