|
From: Nicholas N. <nj...@cs...> - 2005-12-05 16:51:53
|
On Mon, 5 Dec 2005, Zhongxing Xu wrote: > I am doing a tool for valgrind named: floatcheck. > > I am a graduate student. I concern about floating-point roundoff error > for a long time, and wanted to doing something for it. I just found > valgrind a week ago. I think it may be suitable for the job. > > I don't have a very clear idea yet. The basic idea is to replicate the > storage of the original program and do a higher precision version with > that. Then we can compare the final results or give out warnings at > some critical point(e.g. disagreed branch point). > > I began reading documents and a little code of valgrind. I think the > floating-point part really needs to be strengthened. > > To my surprise, I find a similar idea on he project list today. > http://www.valgrind.org/help/projects.html > > Is there anybody doing related things to this? One or two other people have thought about it, but I don't think anybody made any real progress. One issue you may hit is that although x86/AMD64 have 80-bit double-precision floats, Valgrind only does them to 64-bit accuracy. Good luck with this. It would be an excellent tool if you could get it working well. I hope you have a good understanding of floating point arithmetic :) Nick |