I need something between scan and largescan

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

I need to perform a scan on an array of size 2048. As far as I know the scan example goes till 1024 (with 512 threads) and the scanlargearray example is a little overkill for this (and slow, because data is written to global memory, while 2048 floats fit in shared memory)

Anyone ever tried to adjust the scan example, or maybe someone can already tell me that it is impossible? I don’t have the Gems3 book with me unfortunately, because I need the book every time to understand the scan example… External Image

2

Just extend the small scan (single threadblock) to handle 2048 instead of 1024 elements. It’s a straightforward extension - you’re only doubling the intput size.

Paulius

3

Thanks, that will be the path that I take then.