On Feb 18, 2007, at 20:48 , Herbert König wrote:

> Hello Bert,
>
> thank you, your tips are welcome as always.
>
> maybe its clearer if I state what I want to achieve in the beginning.
>
> I have a neural network (self organizing feature map) and I want to
> watch it learn.
>
> Right now it has 100 nodes and 156 inputs, that's determining the
> size. I start out with a neighbourhood of 32 which means that 65 of
> the nodes will get trained for each of my 800+ samples. That was my
> Float array question last time.
>
> If everything goes well, after every training of any node I want to
> redraw the line of 156 coefficients which are between 0 and 1.
>
> I planned to have a white Form and a blue Form, and use
> copyBitsTranslucent: coefficientMappedTo255..
>
> Or I might use 255 Forms of the right colours and blit them to the
> appropriate coordinates.
>>>         rectangleCoords do: [:koords| bb destX: koords x; destY:
>>> koords y; copyBits.]
>
> BF> Wouldn't "Form over" be sufficient?
>
> How to understand this? Speed is about the same. Is there some place
> to read up on what the different modes actually do?

The Blue Book for example. I usually look at the BitBltPlugin  
sources ...

> BF> Try not drawing to Display but to a temporary form, and then  
> blt this
> BF> one.
>
> Thanks I'll try. Do you mean that one big blt is faster than many
> small blt*s

Yes. The way to optimize drawing performance is to let bitblt do most  
of the work in as few calls as possible.

> and that a blt to a physical display is slower than to a
> Form?

Yes, because it is also copied to the OS window.

> I didn't yet because it involved an extra step.
>
> BF> BitBlt has nothing to do with OpenGL or D3D.
>
> Just didn't know what ate the cycles so I tried everything that could
> be tried easily before asking here.

Well, the best way is to measure.

> BF> What are you actually trying to achieve? It's probably much more
> BF> efficient to update a single form and then enlarge it by 4 using
> BF> warpblt.
>
> See above and I'll try warpblt too.
>
> BF> Also, the Kedama Plugin has primitives for drawing large  
> numbers of
> BF> particles, if that is what you're after.
>
> I'll look again but I think that's different from what I want. Squeak
> is big, so give me a decade or two :-))
>
> BTW, others here seem to be into neural nets too, I don't mind to
> share what I have and what I have learned.

Profile your performance. And, for a real speedup, try to rework the  
algorithm to not work on individual elements, or at least keep the  
per-element cost to a minimum.

- Bert -