AnsweredAssumed Answered

double-double for opencl

Question asked by moozoo on Apr 19, 2015
Latest reply on Apr 21, 2015 by jtrudeau

I'm after more precision than double.

 

There is a couple of double-double libraries for CUDA

One from Nvidia https://devtalk.nvidia.com/default/topic/509693/double-double-precision-arithmetic-library-now-available/

Another is open source https://code.google.com/p/gpuprec/

 

Are there any for opencl?

I only need basic arithmetic and inverse square root.

 

What I really needs is the vector form (i.e.doubledouble4).

 

I believe the older amd graphics cards used 4 float ALU's to do double precision.

Is this exposed in some low level manner that would lend it to higher levels of precision.

i.e. quad from 4 double precision ALU's

Outcomes