Shared and Global memory

[Old posts from the commercial version of ArrayFire] Discussion of ArrayFire using CUDA or OpenCL.

Moderator: pavanky

Shared and Global memory

Postby Great » Thu Aug 19, 2010 7:09 am


I read that in the CUDA Architecture (Programming model) can be used different memorys; Global-, Shared-memory, and Registers.

In a few words: All the threads have their own Register, which only it can use. A block of threads can use a shared memory, but a thread in a other block can't see this. And all threads can use the Global memory. The host can read & write global memory but not shared, nor register memory. The shared memory as fast as the registers, but the global memory has higher latency.

So, is it possible to afford each of the variables that are stored in the Shared or the Global memory in Jacket? And if not, which one to use by default?
Posts: 1
Joined: Thu Aug 19, 2010 4:39 am

Return to [archive-commercial] Programming & Development with ArrayFire