Reducing register usage

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I have a large kernel that uses too many registers (46) that kill occupancy. I’ve been hand-optimizing for days but don’t seem to be able to beat the compiler :-) It is possible to split the kernel into two disjunct ones with 24 and 47 registers, but that introduces global memory reads and writes and the overhead of calling two kernels that i don’t really like either.

I was wondering if any clever general principles for reducing register usage exist. Or that it always depends on a particular kernel. The programming and best practices guide don’t say too much on the subject.

Kind Regards,
Daniel Dekkers

2

To limit you register usage, declare volatile variables when you use them more than one time. Avoid to use double numbers.

For example, instead of:

if ( i < 0 )

for ( i = 0 ... )

use this code:

volatile int zero = 0.0f;

if ( i < zero )

for ( i = zero ... )

Cheers,

luca