Re: Decimal Review Result and Summary.

From: Christopher Kormanyos (e_float_at_[hidden])
Date: 2025-02-01 20:41:53

 >> NO: We will NOT be able to* No,>> we will not be able to meet the>> factor of 10 speed exhibited by>> one particular architecture-specific>> library.
> I must be missing something here,> because when I look at
No Peter, you are not missinganything. It is I who lackedclarity-of response.
I should be more clear and thisis a personal response. Thishas not been filtered through Matt.
I, known as Chris, do not identifymaximum performance as a top-levelrequirement for an IEEE-754 conformingdecimal. I am fascinated by thepotential in language andherenceand potential specification.
I will not try to reproduce thatlibrary's magnificent performance,as this holds no technical interestfor me whatsoever.
This might be a stark answerto your query, but I'd betterbe clear about that up front.I won't spend much time on that.
Maybe a wrapper of sime othertype like we do in Multiprecisioncould give wou the
"speed that you need"
- Chris

    On Saturday, February 1, 2025 at 09:22:54 PM GMT+1, Peter Dimov via Boost <boost_at_[hidden]> wrote:  
 
 Christopher Kormanyos wrote:
> NO: We will NOT be able to* No, we will not be able to meet the factor of 10
> speed exhibited by one particular architecture-specific library.

I must be missing something here, because when I look at

https://www.netlib.org/misc/intel/

which is linked from

https://www.intel.com/content/www/us/en/developer/articles/tool/intel-decimal-floating-point-math-library.html

I don't see anything architecture-specific in the library, everything looks like portable C.

I'm also not sure why this library would be so much faster, because the number format
seems to be standard IEEE BID

https://github.com/featuremine/IntelRDFPMathLib/blob/edda5420528167b14b775cc871e673d561a03da9/LIBRARY/src/bid_internal.h#L2263-L2297
https://github.com/featuremine/IntelRDFPMathLib/blob/edda5420528167b14b775cc871e673d561a03da9/LIBRARY/src/bid_internal.h#L772-L783

and when I look at the comparison

https://github.com/featuremine/IntelRDFPMathLib/blob/edda5420528167b14b775cc871e673d561a03da9/LIBRARY/src/bid32_compare.c#L37

I see nothing that could possibly explain an order of magnitude difference in performance.

Am I looking at the wrong thing?

_______________________________________________
Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost