On Sun, 24 Oct 2021 02:43:43 +0200 Ichthyostega <p...@ichthyostega.de> wrote:
>Hi Will, >Hi Kristian, > >yesterday I missed the forest for the trees... >What is actually going on here is that recent Compilers/Optimisers >are clever enough to perform that transformation: replace a powf() >with constant base by a expf() with precomputed logarithm. > >Just depending on the compiler and seemingly the way how SSE is used, >this leads to various degrees of slightly different numerical results. >The generated assembly really looks quite different on GCC-7 vs. GCC-8. >And the O0-case still uses the powf() variant and thus gets significant >different results. > >Today, for this experiment, I have now replaced that code explicitly >to always use the expf() with a precomputed logarithm. For this test case, >I also needed to do the same on the global volume (and I see various further >instances of the same pattern, e.g. in LFO or with punch, which we just do not >touch with that simple sine test). > >After those adjustments I now get precisely the same numbers with the >Release builds both on my PC and my laptop. With that first simple >test with just a single sine note of course (I expect more such hickups >down the road) > > >However, I still get different results with Debug builds vs Release >builds. And this is due to the AnalogFilter. Since we're using >--fast-math, the optimiser is allowed to reorder statements, which >seemingly leads to some build-up of differences in the feedback >loop of that filter. But, as said, the results on the Release >builds as such do match now on my Desktop and Laptop. > >-- Hermann That's a good bit of sleuthing. Well done! I've found in the past that AnalogFilter was 'interesting' and it gets used in a *lot* of places. Is this ready to go now? -- Will J Godfrey https://willgodfrey.bandcamp.com/ http://yoshimi.github.io Say you have a poem and I have a tune. Exchange them and we can both have a poem, a tune, and a song. _______________________________________________ Yoshimi-devel mailing list Yoshimi-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/yoshimi-devel
