The future is 24bit

[preda]

Originally Posted by Mark Rose

Niall Emmart's dissertation A Study of High Performance Multiple Precision Arithmetic on
Graphics Processing Units may be useful.

Thank you. While the thesys contains a wide overview, it's does not go too deep into multi-precision FP (at least based on my cursory inspection).

[Mark Rose]

Originally Posted by preda

Thank you. While the thesys contains a wide overview, it's does not go too deep into multi-precision FP (at least based on my cursory inspection).

Agreed, but I was thinking section 5.1's techniques starting on page 104 for large integer addition and subtraction may also be useful for floating point? He found performance benefits using a second kernel for carry propogation, and I wonder if there's an opportunity to do something similar here. Though you may well already be doing something similar.

[preda]

Originally Posted by Mark Rose

Agreed, but I was thinking section 5.1's techniques starting on page 104 for large integer addition and subtraction may also be useful for floating point? He found performance benefits using a second kernel for carry propogation, and I wonder if there's an opportunity to do something similar here. Though you may well already be doing something similar.

Yes I think we're the state of the art in what concerns carry propagation :)