On 08.12.2021 22:13, Ichthyostega wrote: >>> Am 05.12.21 um 21:44 schrieb Will Godfrey: >>>> I'm still seeing a very poor performance with BasicSUB (-48.5722dB). > >> On 06.12.2021 23:21, Ichthyostega wrote: >>> ... These (obviously rather limited) tests indicate that possibly "build >>> for this CPU" might give us the best match of results on different >>> compilers. All with -O3 of course. > > Am 07.12.21 um 22:43 schrieb Kristian Amlie: >> So do I understand it correctly: The best thing is for each to use >> -march and >> set it to the CPU type of their PC? > > > yes: -march=native -mtune=native -O3 > > This allows the compiler to autodetect the best matching sub architecture > for your CPU. Incidentally, the -tune setting is redundant, because > -march=xxx always implies -mtune=xxx > > > The default if you do not specify -march at all is to use the > compiled-in default architecture of your compiler executable. > Typically this is a very broad, generic setting. And thus the > latter is what pretty much all distros do, fall back on the > conservative compiler default. > > > You can try it out: > > gcc -Q --help=target > > > then, to the contrary, try > > gcc -Q -march=native --help=target | grep march > > > > OK, for sake of completeness: we seemingly have here one of these > extremely rare cases, where these settings /do/ have /any/ tangible > effect. Other than creating an executable that possibly crashes > on any other CPU. And, as expected, my run time measurements > indicate that the most generic case (not setting -march) vs > -march=native gives *not the slightest measurable* runtime benefit. > Just it seems to handle rounding on the last place of floating > point values better. And our beast, the AnalogFilter amplifies > this minute effect to produce the differences we see. > > With the basic generic setting (O3 but no -march), I see differences > ~ -70dB .. -110dB on most tests over my three machines. With > -march=native and O3 only two tests give any difference at all here, > all way below -120dB
Indeed, in general I'm getting better results, but there are some outliers. I've posted a new thread, since I think it's unrelated to the float precision; the difference is too big; so it must be something in my setup. -- Kristian _______________________________________________ Yoshimi-devel mailing list Yoshimi-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/yoshimi-devel
