Monitor performance

[Ashurek85]

I'm working on a project with ngdevkit and would like some feedback on how to monitor performance

I currently have an FPS meter that stays stable in the gngeo emulator. I don't have a flashcard or the actual hardware to test it on a NeoGeo. Based on your experience, if it works well in Gngeo, does it also work well on actual hardware?

I want to shed some light on a few points and hear everyone’s personal experiences; this might help us

I’ve read that:

• Don’t use C++; it’s better to stick with C
• No dynamic memory (malloc, calloc, etc.)

What assumptions do you use?

Thanks

[dciabrin]

My 2c below

I currently have an FPS meter that stays stable in the gngeo emulator. I don't have a flashcard or the actual hardware to test it on a NeoGeo. Based on your experience, if it works well in Gngeo, does it also work well on actual hardware?

From experience, you want to use MAME for precise measurement. GnGeo's strength is that it's simple to compile and understand, but from an emulation's standpoint, MAME is more accurate, both in term of features and timings. I now exclusively use MAME.

There are at least three things I would envision for profiling:

• the crude approach we used do with @yo6snap back in the demoscene days: change the background color when the VBlank starts, and revert it to back when you finished your computation. This works great on MAME and gives you a simple measurement of available CPU resource every VBlank.

• the customizable apprach, again with MAME: I haven't used it yet, but @city41 has a project at https://github.com/city41/mameNeoGeoDevPlugin, a series of lua extensions for MAME that can totally be the basis to a nice resource profiler.

• alpine9000 https://github.com/alpine9000/engine9000-public is probably something to support in the devkit eventually, as it seems to be a useful way of profiling 68k code. Though I have never tested it personnaly, and I'm still in the process of tweaking the debugging symbols that are generated by ngdevkit at linktime. So YMMV.

I want to shed some light on a few points and hear everyone’s personal experiences; this might help us

I’ve read that:

* Don’t use C++; it’s better to stick with C

I think that one stems from the fact that depending on how you use C++ it may have hidden runtime costs (e.g. class initialization, virtual function calls, non trivial allocators...). That being said, ngdevkit should support it.

* No dynamic memory (malloc, calloc, etc.)

The default newlib allocator is not meant to be fast, just to support the C runtime. So if you plan to do any dynamic allocation, you should definitely rely on a custom allocator.

[sasja-san]

Here's a short demo of the e9k-debugger, mentioned above. Skip to about two minutes in to see some of the performance measuring it provides:
https://www.youtube.com/watch?v=Q24F6S8J57U

I asked the creator and it should be quite easy to get started with it. At least he managed to do it. I have yet to sit down and test this out.

As for C vs C++, I don't think there will be too much difference between the two. Are there some special C++ features you are looking to use? Namespaces are pretty neat to create cleaner enums. And maybe classes might be useful in some instances. I'm sticking with C and might inline ASM if I run into problem which I can't solve otherwise.