Separating elements in CUDA

Question

I have a vector of structure A in the device memory of GPU.

struct A{
    int a;
    int type;
}

I am trying to split each type and store it in an array.

For example if the types are 0,1 and 2,then split all 0's,1's and 2's 
in a separate array.

The approach I am thinking is to implement it using switch case in parallel but it might lead to a lot of divergence and thus may not be an efficient.

Is there any other approach that would help ?

Answer 1

I would start by checking out thrust:

http://docs.nvidia.com/cuda/thrust/

how fast is thrust::sort and what is the fastest radix sort implementation

There is a way to sort with key/value pairs in case you need something else to be sorted along with it:

Thrust Sort by key on the fly or different approach?

I have implemented my own cuda radix sort. I needed to keep track of an index number while I sorted on something specific. If these options don't work for you I can try to find some time to document it.