WANTED: cuPrintf that uses local memory instead of registers is that possible?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

My kernels are often under register pressure and adding a cuPrintf in there crosses the threshold so the kernels won’t launch anymore.

Could nVidia provide a cuPrintf that uses local memory instead of registers and thus does not add extra registers to a kernel? Would
this even be remotely possible, given the current state of the compiler?

Also what’s holding you back from making cuPrintf entirely public and accessible for everyone?

Christian

2

Can’t you just set a -maxregcount parameter? Every excessive registers will be spilled to local memory automatically. Won’t be fast but will work…

3

Isn’t everything in a register before it gets written to global/local or host memory? As far as I know, only shared memory can be directly used.