Closed GoogleCodeExporter closed 9 years ago
There is a branch of CUDPP for this development. You can access it here:
http://code.google.com/p/cudpp/source/browse/#svn/branches/tridiagonal
Original comment by harr...@gmail.com
on 3 Nov 2010 at 11:50
Yes.We have CR and PCR but i'm looking for RD ( recursive doubling )
implementation. I don't think it is present in
http://code.google.com/p/cudpp/source/browse/#svn/branches/tridiagonal [ it
contains only CR,PCR and Hybrid CRPCR ]
Thanks in advance
Original comment by suyesh.t...@gmail.com
on 7 Jan 2011 at 4:59
I know this is old... but I'd like to warn anyone else looking to implement
recursive doubling:
Recursive doubling is inherently numerically unstable. The algorithm works with
arbitrary precision values. Examine the algorithm and you should notice that
elements within C_i get progressively large, ultimately resulting in numerical
instability. (In my implementation, numerical instability occurs when n > 128.)
It'd be nice to know if there is a solution to this problem. I couldn't find it
after spending quite some time implementing recursive doubling.
Original comment by djh.s...@gmail.com
on 11 Sep 2012 at 5:52
In our 2010 PPoPP paper ("Fast Tridiagonal Solvers on the GPU"), we noted the
same; both plain recursive doubling and RD+cyclic reduction overflowed in our
tests. I agree: this is not suitable for a GPU implementation.
Original comment by jow...@gmail.com
on 12 Sep 2012 at 4:08
Original issue reported on code.google.com by
suyesh.t...@gmail.com
on 15 Aug 2010 at 5:18