[Cfp-interest] AI about reduction requirements

[Jim Thomas]

Action item:

>     Jim: Create a CR to change the reduction operations description from should->shall
> 

I believe we were looking at the TS 18661-4 specification for reduction functions in 7.12.13b, which says

[1] The functions in this subclause should be implemented so that intermediate computations do not overflow or underflow.

Note that this specification is in 7.12, hence for all implementations, not just IEC 60669. The IEC 60559 specification regarding overflow and underflow is  covered by:

After F.10.10a, insert:

F.10.10b Reduction functions

[1] The functions in this subclause return a NaN if any member of an array argument is a NaN, unless explicitly specified otherwise.

[2] The reduc_sum, reduc_sumabs, reduc_sumsq, and reduc_sumprod functions avoid overflow and underflow in intermediate computation. They raise the “overflow” or “underflow” floating-point exception if and only if the determination of the final result overflows or underflows.

[3] The scaled_prod, scaled_prodsum, and scaled_proddiff functions do not raise the “overflow” or “underflow” floating-point exceptions.

[2] is intended to be equivalent to IEC 60559’s “…, sums are computed in a manner that avoids overflow or underflow in the calculation and the final result is determined from that intermediate result. If that result overflows, signal overflow. If the result underflows, signal underflow.”

[3] is intended to be equivalent to IEC 60559’s “… the scaled result, pr, shall not be affected by overflow or underflow”. 

I don’t think we need to change anything.


-Jim Thomas
 




-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.oakapple.net/pipermail/cfp-interest/attachments/20180628/b26ddc1f/attachment-0001.html