how to avoid race condition without using the Atomic function

Please understand my low command of english.(I’m not native)

Hello I’m studying physical simulation and I have trouble in sorting.

When I change my source code to CUDA version, there was race condition problem.

I had been finding the sort algorithm with cuda, and I decide. I have to develop the parallel algorithm to understand my code.

So my question is that, how I can avoid race condition without using the atomic function.

Or how I can implement the read - after - write function with simple code, or device function.

ArrayFire will be helpful to you. Use this:

Thank you for your reply.

and Is there any hardware restriction? My gpu board is Tesla C870,

It can’t use some additional function of CUDA latest version.

You can also do reductions. Google a little:

CUDA provides several functions which are useful. YOu need to take a look at the Programming guide and see a list of availabele functions.

There is no restriction. ArrayFire will run on the Tesla C870.

You can use the shared memory or just use the thrust library.

Thank you for your replies.

I’m still studying. It will be very helpful.

When I solve this problem I’ll inform my solution. :)