double and integer 64 bits

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone,

I have an algorithm that is using 64bits integer and double (it is a benchmark). I would like to know if there is a way to implement these types of data. I know that we only can use single precision on CUDA. Could I use two floats to computing a double? Do you have an idea about how I can address that without loss accuracy in my results?

Thanks a lot,

Diego

2

One person presented a poster on this at SIGGRAPH 2006:

http://cs.allegheny.edu/~thall/

If you zoom in and read the PDF, it gives some hints for how this was done, but I haven’t found any code yet.

3

Actually, (if you don’t mind reading FORTRAN), the dsfun90 package on this page looks promising:

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

It at least shows how to implement double precision routines in terms of single precision variables (REAL*4 is a 32-bit float in FORTRAN).

4

Thanks a lot seibert, I will send an email to the guy from the conference.

Diego

5

If he gets back to you, let us know what his status is on releasing the code (if he is planning to do that). I realized today that I might need to work in double precision for some calculations, and other people probably have similar questions.

6

Hi seibert I have an answer from him, but it is not hope. If you write me an email I could replay his answer. my email is : drivera@ece.neu.edu.

Thanks

Diego