For Turing, if shared memory is larger than 32KB but smaller than 64KB, can L1 cache use the remaining part?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi! I am using Turing for shared memory allocation. I notice two allocation choices: shared memory 32KB L1 cache 64KB, or shared memory 64KB L1 cache 32KB. So if I allocate 49KB, 64-49, still remain 15KB! Will this part be used by L1 cache? Or still stubbonly stay there for possible future shared memory allocation?

Thank you!!!

2

I don’t think this is documented. I interpret the word “dedicated” to mean that if you choose 64KB for shared, it is “dedicated” meaning it is not available for L1. Perhaps someone has written a microbenchmark.

1 Like