Borrowed registers NVCC using Shared Memory for Registers

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

Is it possible to ask NVCC to utilise unused shared memory as registers?

Thanks,
Paul

2

No. You have to force the variable to shared mem in your code.

nvcc handles register spills always by putting the variable in local memory.

Peter

3

Thanks Peter.

Is it the same when too much shared memory is requested? If I crank up some shared variable sizes, a kernel still executes. But if I make Ns, the third kernel execution configuration parameter too large, it doesn’t; cudaGetLastError returns cudaErrorLaunchOutOfResources.

Paul

4

Shared memory is max 16k per block. That goes for static as dynamic shared mem. So static+dynamic < 16k. If you augment the shared mem requirements of your code from say a few bytes to several kbytes, the runtime will first schedule fewer and fewer blocks on each multiprocessor until it runs out of memory at 16k. Then you get a launch failure. Check the occupancy calculator how the shared mem size influences the multiprocessor block scheduling.

Peter