Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Motion 24.03: NO



I vote NO for Motion 24.03

I agree with Vincent Lefevre and Nathalie Revol.

Additionally I think the requirement of this operations with directed rounding and maximal accuracy will block (efficient) implementations on some systems. For example on GPU's it is necessary for the maximal performance that all threads perform the same operations simultaneously. If we use algorithms (on GPU's without directed rounding) to simulate the directed rounding with one ulp accuracy and as long as this algorithms use branches than the computation is performed sequentially. Hence, we lose the power of this highly parallel system. On the other hand if we use round to nearest for the operation and then multiply by (1+\epsilon) we lose the maximal accuracy but we can perform this operations in parallel.

And as mentioned in my mail from July 20 I am in doubt that this motion fits into the specifications of other languages/systems like Java (or all the other Java VM languages).

Therefore I think this motion work against early and widespread acceptance of the standard.

Best regards

Marco

--
     o           Marco Nehmeier, Lehrstuhl fuer Informatik II
    / \          Universitaet Wuerzburg, Am Hubland, D-97074 Wuerzburg
InfoII o         Tel.: +49 931 / 31 88684
  / \  Uni       E-Mail: nehmeier@xxxxxxxxxxxxxxxxxxxxxxxxxxx
 o   o Wuerzburg