Thanks, interesting numbers. - Alon
On Tue, Dec 24, 2013 at 2:02 AM, <[email protected]> wrote: > The percent number is calculated from the time it takes to generate 1 > second of sound data, so it's a pretty accurate measure of CPU time for the > emulation. > > > On Tuesday, December 24, 2013 5:59:32 AM UTC+1, azakai wrote: > >> I was hoping that getting the compiler to emit different stuff would help >> the JITs. But it sounds like in this case hand-optimizing as you are doing >> makes the most sense, if it is one big and important switch at the core. >> >> What does the 40% on Android mean? What is compared to what? >> >> - Alon >> >> >> >> On Sun, Dec 22, 2013 at 12:37 PM, <[email protected]> wrote: >> >>> I got it a lot faster by breaking up the switch statement manually >>> (Around 4x - 5x). Interestingly, even though this is quite CPU-heavy code >>> (emulation) the difference between Chrome and Firefox is marginal (Around >>> 10%) >>> >>> Outlining does not seem to help, neither with compilation speed nor run >>> speed. >>> >>> I compile with normal optimizations (-O2) and no extra flags that may >>> slow things down as far as I know. >>> >>> Have not checked the useIfs-thing, becuase this was a JIT problem and >>> that seems to be about the compiler. Also, my switch statement has close to >>> 256 statements so I don't think it applies anyway. >>> >>> -- Sasq >>> >>> On Sunday, December 22, 2013 12:11:39 AM UTC+1, azakai wrote: >>> >>>> Yes, outliner is one option here. If that doesn't help, you can >>>> manually change the heuristics for switches vs ifs in the compiler, look >>>> for `useIfs` in src/jsifier.js. >>>> >>>> Was this with optimizations on, or not? >>>> >>>> - Alon >>>> >>>> >>>> >>>> On Sat, Dec 21, 2013 at 2:35 PM, Sören Balko <[email protected]>wrote: >>>> >>>>> Why don't you change the source to split up that single large switch >>>>> statement across multiple functions? The Chrome profiler should then tell >>>>> you if that was effective, no? You could also try Alon's outlining feature >>>>> and see if that does the trick without touching the source. I think this >>>>> was introduced to split large functions into smaller ones in order to make >>>>> it easier (and faster) for JITs to optimize the code. >>>>> >>>>> >>>>> Am Samstag, 21. Dezember 2013 20:11:10 UTC+10 schrieb [email protected] >>>>> : >>>>> >>>>>> So the SID-chip/C64 emulation part of my music player runs very slow >>>>>> - about 10x native. Checking in the Chrome profiler I see that the main >>>>>> function has an exclamation mark with: >>>>>> Not Optimized: SwitchStatement: Too many clauses >>>>>> >>>>>> The 6510 emulation core is one big switch statement, so this is true. >>>>>> I realize this is a JIT problem but I wonder if you can get the >>>>>> emscripten >>>>>> compiler to help in this situation? >>>>>> >>>>>> Otherwise maybe I can try breaking it up manually, but would that >>>>>> help? >>>>>> >>>>>> -- Sasq >>>>>> >>>>> -- >>>>> You received this message because you are subscribed to the Google >>>>> Groups "emscripten-discuss" group. >>>>> To unsubscribe from this group and stop receiving emails from it, send >>>>> an email to [email protected]. >>>>> For more options, visit https://groups.google.com/groups/opt_out. >>>>> >>>> >>>> -- >>> You received this message because you are subscribed to the Google >>> Groups "emscripten-discuss" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to [email protected]. >>> For more options, visit https://groups.google.com/groups/opt_out. >>> >> >> -- > You received this message because you are subscribed to the Google Groups > "emscripten-discuss" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/groups/opt_out. > -- You received this message because you are subscribed to the Google Groups "emscripten-discuss" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/groups/opt_out.
