CUDA: How do I use L2 cache in Fermi?

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

I need to be able to transfer data between different blocks but without using global memory because it’s way too slow. In the NVIDIA Fermi card, there’s a new L2 cache that goes across the whole GPU. Is it possible to use this cache, and if it is, how?
Thanks

2

Hi Tom,

The L2 cache is hardware managed so you don’t have direct control on how the data is stored. I’ve been poking around the web to see if I can find details on how the L2 caching policy works, but haven’t found anything concrete. As best as I can tell, data in the L2 cache is visible to all executing blocks and I assume if the hardware will cache frequently used variables especially if used across blocks.

  • Mat
3

I posted a similar question in the NVIDIA forums: The Official NVIDIA Forums | NVIDIA. Apparently the only control you have to the L2 cache is through modifiers in PTX instructions. This can be done on the CUDA C compiler using inline PTX.

What I’m hoping is that by not going over the 768KB limit, every read/write I make to global memory will be done through the L2 cache. I’ll be frequently transferring data between blocks by writing to a global memory and then reading from that array with the other blocks. Is is wrong to assume that the L2 cache will always be used if I don’t go over 768KB?

4

Is is wrong to assume that the L2 cache will always be used if I don’t go over 768KB?

Unfortunately, I don’t know. Hopefully the NVIDIA forums can be more helpful with this one.

  • Mat