egg changed the topic of #kspacademia to: https://git.io/JqLs2 | Dogs are cats. Spiders are cat interferometers. | Document well, for tomorrow you may get mauled by a ネコバス. | <UmbralRaptor> egg|nomz|egg: generally if your eyes are dewing over, that's not the weather. | <ferram4> I shall beat my problems to death with an engineer. | We can haz pdf | Logs: https://esper.irclog.whitequark.org/kspacademia
egg|anbo|egg has quit [Remote host closed the connection]
egg|anbo|egg has joined #kspacademia
<egg|anbo|egg__>
Iskierka: nobody has a hardware *cube* root
<egg|anbo|egg__>
but if you are talking about the rsqrt, see Steve's point that absent the fast inverse cube root in hardware, a precise square root and division are generally faster
<egg|anbo|egg__>
the tradeoff of the fast inverse square root thing relies on the other FP operations being much faster than the division and square root, which is not as much the case as it used to be
<egg|anbo|egg__>
Iskierka: the broader point is that these tradeoffs change all the time, so must be re-evaluated if they are used after a long time
<galois>
title: Principia/cbrt.pdf at cbrt-documentation · eggrobin/Principia · GitHub
<egg|anbo|egg__>
31 references
<egg|anbo|egg__>
two of which are twitter threads,
<raptop>
Remove Before Flight
<raptop>
hah
e_14159_ has joined #kspacademia
e_14159 has quit [Ping timeout: 189 seconds]
<Iskierka>
part of the point that there are faster computations now is suggesting that pipelining is much of the advantage. I'd debate whether a processor that lacks a sqrt (while having float support at all at all) would be modern enough to pipeline given it's at least 35 years old
<Iskierka>
though fair that combined ops like FMA do reduce the advantage anyway
egg|anbo|egg_ has joined #kspacademia
egg|anbo|egg___ has joined #kspacademia
egg|anbo|egg__ has quit [Ping timeout: 189 seconds]
egg|anbo|egg has quit [Ping timeout: 189 seconds]
egg|anbo|egg___ has quit [Remote host closed the connection]
<egg|cell|egg>
Iskierka: to the contrary, pipelining plays against a sqrt followed by div, and in favour of polynomials
<egg|cell|egg>
But div and sqrt are not as slow as they used to be in comparison to mul and add
<egg|cell|egg>
1/sqrt cannot benefit from pipelining, it is a serial dependency
<egg|cell|egg>
But with Estrin evaluation polynomials can
<egg|cell|egg>
(avoid Horner like the plague obviously)
<egg|cell|egg>
The question is not sqrt support mind you; it is rsqrt
<egg|cell|egg>
(yes you will have that if you have SSE, but even if you don't use it you still shouldn't do that ancient trick anymore)
<egg|cell|egg>
(rsqrtss)
<egg|cell|egg>
Whitequark: how are the cats
egg|anbo|egg has joined #kspacademia
egg|anbo|egg has quit [Remote host closed the connection]