Double in C1060 - shared variable splitting

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi ,

am using a double shared variable in y kernel … and according to the programming guide on page 61. I would have bank conflicts when I access data from it.

he solution they state is to split the variable into hi and lo variables. I was not able to fully understand how it would work ?

Also I wondering if there is a better way to handle that problem? as my current sp code is little messy when it comes to shared memory access, I dont wana make it more messy with this hi and lo thing.

thanks all…

2

Then dont make it more messier… Shared memory conflicts are the last thing to look at for speedup.

3

Thanks Sarnath…

Why do you say they shared memory conflicts are the last thing to consider ? I am getting a performance of 45gflops … its a test Nbody algorithm, which I have written to learn cuda with double precision. The card is Tesla c1060 (max dp gflops of ~ 80 )

Pardon my lack of know how … but am not from computer science background.

Thanks,