Execution configuration question

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey guys I had a question about the dimension and size of the grid ie Db in <<<Dg, Db, Ns>>>

for my program when I allocate more threads in the x and y direction I get better performance compared to when I give more to in z.
For example (16,16,2) is much faster compared to (2,2,16)

Could somebody explain the reasoning behind this.

Thanks in advance

2

How are you using the thread indices in your kernel? One configuration might be looping over memory in a different order than the other.

Also, your two example cases have different numbers of threads. Do you mean (16,2,2) compared to (2,2,16)?

3

Yeah sorry that was a typo, it is (16,16,2) & (2,16,16) or a better example could be (64,2,2) is much faster than (2,2,64). What exactly do you mean by using thread indices in the kernel? I am dealing with 3D input streams and I get their indices from the resp directions and put them in a 1D array and then execute.

Thanks

4

How do you combine the the x,y,z thread indices into the 1D array index? I’m wondering if there is something about the indexing which is breaking coalescing.