On Sun, 14 Nov 2010 16:03:57 +0100 Arnold Krille <[email protected]> wrote:
> On Sunday 14 November 2010 06:14:48 Ralf Mardorf wrote: > > On Sat, 2010-11-13 at 19:30 +0100, Arnold Krille wrote: > > > On Saturday 13 November 2010 18:51:17 Ralf Mardorf wrote: > > > > On Sat, 2010-11-13 at 16:26 +0100, [email protected] wrote: > > > > > On Sat, Nov 13, 2010 at 02:57:44PM +0000, Folderol wrote: > > > > > > Paul Davis <[email protected]> wrote: > > > > > > > there's an awful lot of math for which a modern processor can > > > > > > > compute the answer faster than it can look it up. this is one > > > > > > > such example (as fons noted), but there are many others. this > > > > > > > started changing about 8 years ago, and its only gotten more > > > > > > > true since then. > > > > > > > > > > > > Oh dear. I didn't realise I was so far out of date :( > > > > > > > > > > It's absolutely true. > > > > > > > > > > Note that the calculation I posted earlier runs entirely on > > > > > the FP processor, and two variables, q and d, will probably > > > > > only exist there and never be written to memory. If it takes > > > > > any time at all, the CPU can meanwhile do something else, > > > > > such as getting the pointers to the audio data it will need > > > > > later. Using a LUT will mean the CPU has to do all the work, > > > > > which includes getting the base pointer and calculating the > > > > > required offset(s). > > > > > > > > Just to ensure you aren't talking about different claims. > > > > > > > > Is math even faster, as e.g. 4 states for left, 1 for centre and 4 > > > > states for right, provided e.g. by an array? Or are you thinking about > > > > nearly stepless panning? > > > > > > Unless your memory runs (and is connected) at the same speed of the > > > processor (which afaik no platform has, even the first and second cache > > > run at lower speeds), simple math of only a limited number of operations > > > is always faster then getting stuff from memory. Todays processors > > > contain even more optimizations for fast and parallelized math. > > > The trick Fons showed is about doing the pan calculations in simple math > > > instead of using trigonometry which would require complicated functions > > > like sin/cos to be called or tables from memory to be checked. > > > > > > Have fun, > > > > > > Arnold > > > > Using sin/cos still is very time-consuming? > > > > I do understand The Wiki on German better, but the Wiki on English, > > pardon: > > > > "Die Tabellierung aller Werte ist angezeigt bei > > geschwindigkeitskritischen Echtzeit-Anwendungen, wenn diese nur eine > > recht kleine Winkelauflösung > > benötigen." (http://de.wikipedia.org/wiki/Sinus_und_Kosinus#Berechnung) > > > > Loosely translated: Tabulation for rt (when using sin/cos) for small > > angles (what ever a small angle might be) is still faster. > > "Winkelauflösung" != "Winkel", that is the angle resolution is not the angle! > So if you want results with bad resolution for your panning, using tabulated > values will work. But that will result in "jumps" in the panning. Better let > the cpu calculate the sin/cos directly. And then realize that approximating > the sinus in that range by another, simpler function speeds up the process to > the point where its faster then looking up a value in a table... > If it was just about small angles less then say 3 degree, the best > approximation of sin(x) is x itself :-) > > Have fun, > > Arnold I don't know if this is at all relevant (prolly not!) but I dimly remember, from my BBC B days, there was a way of drawing circles using Pythagoras. This was dramatically faster than using sin/cos. -- Will J Godfrey http://www.musically.me.uk Say you have a poem and I have a tune. Exchange them and we can both have a poem, a tune, and a song. _______________________________________________ Linux-audio-dev mailing list [email protected] http://lists.linuxaudio.org/listinfo/linux-audio-dev
