+1 Paul.
On Nov 23, 2013, at 8:06 AM, Joe Darcy <[email protected]> wrote: > Hello Mike and Paul, > > Thanks for reviewing the earlier versions of the change; I've posted an > update revision for your consideration at > > http://cr.openjdk.java.net/~darcy/8006572.4/ > > Two numerical aspects of the change are different in this iteration: > > * when two running compensated sums and combined, the compensation portion > of the second is now added in first. Since it should have smaller magnitude, > this is favorable in terms of error analysis > > * the final sum returned (or used to compute the average) is (sum + > sumCompensation). I checked one of my references, "Accuracy and Stability of > Numerical Algorithms" from Nicholas Higham, and that tweak gave a better > error bound > > The test code has been updated as you've suggested (other than porting to > TestNG ;-) and I also added a test case for a bug Paul spotted on an earlier > review. > > All the streams tests pass with the new code. > > Thanks, > > -Joe
