Re: [Gzz] Benchmarks

[Tuomas Lukka]

On Mon, Mar 17, 2003 at 08:20:37PM +0200, Matti Katila wrote:
> On Mon, 17 Mar 2003, Tuomas Lukka wrote:
> >> I added some benchmarks for pp[press Ctrl+b] and the results are 
> >> interesting, though:
> >> 
> >> 1 : Benchmark - leave out paper :      140  frames.
> >> 2 : Benchmark - leave out shortcuts :  175  frames.
> >> 3 : Benchmark - leave out text :       252  frames.
> >> 4 : Benchmark - leave out background : 269  frames.
> >> 5 : Benchmark - leave out fillet :     122  frames.
> >> 6 : Benchmark - leave out nothing :    122  frames.
> > 
> > What is the "frame" unit? What kind of a period do you use for measurement?
> 
> How many frames are shown on screen from time and place a to time and 
> place b. Where from a to b is deterministic and 'normal' use case of pp.

You mean interpolation? How much Java code is counted in?

This is not good - you should use timeRender() with a specific vobscene.

> Since OpenGL tries to show as many frames as possible between a and b the 
> result should be some way ok?

No. It's not, since you don't know what comes from Java putting
the vobs and what comes from rendering the vobs. This makes the data useless.

There are way too few frames if you're counting them; the difference between
122 and 140 is not statistically really significant, for example.

> > If it's "fps" i.e. frames per second, then you'd actually be better 
> > off calculating the number of milliseconds per frame, as that would
> > be more linear and easily understandable.
> 
> No, it isn't "fps". I let UpdateManager to render as many frames as 
> possible between two scenes.

No, please don't use UpdateManager when doing benchmarking. Too much
uncertainty due to threads, wakeups etc.

> >> It seems that background, which is the calendar thing, eats a lot of 
> >> power. 
> > The calendar has a lot of text, right? What if you disable text in calendar?
> 
> Yes, we are already a lot ahead from the starting point where
> we did know nothing about what's eating all the cpu powers.

No, we don't really know much. You're still shooting blindly, not really
*measuring* anything concrete.

        Tuomas