[PD-dev] Was: Re: algo for sequencing merges

[IOhannes m zmölnig]

Am 22. Dezember 2019 01:10:47 MEZ schrieb Claude Heiland-Allen <claude at mathr.co.uk>:
>Hi,
>
>On 19/12/2019 20:01, IOhannes m zmölnig wrote:
>> the high-performance phase-wrapping algorithm
>For reference: 
>https://github.com/pure-data/pure-data/blob/e3a78da5b7d825bf8b9f5d313e4ea55e5e32e8f1/src/d_osc.c#L91-L97
>
>(current master at time of writing)
>
>Using a [phasor~] with x -= floor(x); for wrapping is about twice as 
>slow as the built-in tabfudge version, provided I compile with 
>-march=native on my AMD Ryzen 7 2700X Eight-Core Processor.  When 
>compiled with pd-lib-builder's default -march=core2, the floor version 
>is over five times slower.  I can run tests on other hardware in early 
>January if more data points are desired.
>
>Benchmark methodology: time how long running 1000 copies of [phasor~ 
>440] 

thanks for the benchmarks.

as katja pointed out, due to branch-prediction the actual frequency of `[phasor~]` might have an impact as well.


>I suppose it's not so hard to load a custom phasor~ library (source 
>attached), if you need the extra accuracy of the floor version, and can
>
>live with the slowdown. 

i think this is a good (intermediate?) solution.
as long as the core framework works with double-precision and no internal object produces actual garbage when compiled with double-precision (as is the case right now), Pd-dbl might be ready to fly...



mfg.hft.fsl
IOhannes