Reduction question

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I was reading through the reduction topics in this forum and got a question.

I want to “sort” or “map” a huge amount of data (like a million float2 values) into a huge array of floats (like 8080, or 200200 values). I can’t just “+=” add the values on the float array, because of the parallel memory acces problems.
But if I use a reduction, i think i need one of those float-arrays where i sort the float2s in per thread and after that i can reduce them, that would mean an enormous amount of data…
And anyways i think reduction only works within the threads of a block and i would need way more than just one threadblock.

So the question is, how can I solve this problem?
Thx for any help.

2

btw…

I could start a thread for every point on the (200*200) float grid and check for every float2 value, whether the particle has an influence on it or not, but that can’t be the best solution, right?

3

It’s really unclear what you are trying to do. Does each output depend on every input? What would a sequential algorithm look like?

4

What are you trying to write? Some sort of particle dynamics code or n-body simulation?