Emulated double precision Double single routine header

Yes it can. The original double-single routines come from a Fortran library here (DSFUN90):

http://crd.lbl.gov/~dhbailey/mpdist/

They also have a quad-double library and a double-double library. Unfortunately, they don’t have a quad-single library, but it should be possible if not a bit tedious to put together your own.

In my experiments with the double-singles, the standard arithmetic were accurate to almost the full 46-bits of precision, but algebraic or transcendental operations like sqrt and exp, while still better than 23-bits of precision, were a lot less accurate.

Hi,

I haven’t done any work on emulated double precision so I would be very curious to hear some performance figures for emulated DP for the DP disabled commercial GF100 cards.

Seems like StickGuy has done some very interesting work!

regards,
Jim

Hi,

I haven’t done any work on emulated double precision so I would be very curious to hear some performance figures for emulated DP for the DP disabled commercial GF100 cards.

Seems like StickGuy has done some very interesting work!

regards,
Jim

In my experience emulated double precision has about one eighth of the performance of single precision in normal applications. Therefore you get just about what you get with the castrated hardware.

In my experience emulated double precision has about one eighth of the performance of single precision in normal applications. Therefore you get just about what you get with the castrated hardware.