Re: Image transforms as a benchmark?

[Alan Third <alan@idiocy.org>]

On Sun, Sep 12, 2021 at 03:28:09PM +0200, Arthur Miller wrote:
> Alan Third <alan@idiocy.org> writes:
> 
> > On Sun, Sep 12, 2021 at 01:45:39PM +0200, Arthur Miller wrote:
> >> 
> >> I tried to make another little benchmark, I saw with optimization flags, that
> >> quite some loops have got unrolled and vectorized in image.c, so I wanted to see
> >> if it matters when doing some transforms on images. I tested so far just with
> >> svg.
> >> 
> >> I wonder if image-rotate is handled completely by external libraries? I see
> >> no effect on performance, regardless of how many time I rotate some image. Is it
> >> same situation for scaling down? I see big difference when scaling up images so
> >> I guess that is handled by Emacs own code?
> >
> > SVG is probably not a great example for testing image transforms, at
> > least if you're using the master branch.
> 
> I wanted to make few benchmarks that are relevant to normal use. I do some
> searches, replacements and similar in a big text buffer (Plat's dialoagues as I
> posted a day before), and I did some symbol lookups in Emacs lisp sources. SVG
> is getting a bit of uprise lately, so thought it might matter. But I'll guess
> I'll skip SVG then.

Well, don't let me put you off if you're bench-marking things that
actually matter. I assumed it was a purely intellectual exercise.

> > Image scaling in SVG is handled at the time the image is rasterized,
> > so if you ask for the image to be doubled in size, the rasterizer
> > creates a bitmap that is twice the size.
> 
> Yes, I saw the code. But I wasn't sure if it does upscaling itself, or it
> outsources the entire venture to librsvg & co. I understand the most of image
> operations are handled by 3rd party libraries. I saw some code for image
> transforms in image.c, but I haven't looked much where it is used and so, maybe
> I should have :).
> 
> Why is it so drammatic difference then when scaling up compared to scaling down?
> It takes like 5 seconds to do that loop in scale-up test, but just a fraction of
> a second when scaling down. I tested little different versions, and I see the
> effect, so scaling is done.

I have no idea. Both scaling up and down are performed by librsvg
itself. It may be some side-effect of the way we do the scaling, with
the wrapped SVG, but I don't have enough knowledge of librsvg to know
why.

Alternatively we're doing something wrong that I haven't noticed.

Is it possible to use a profiler to find out which calls are slow?

> > You could perhaps try stepping through the frames of an animated GIF.
> > Our rendering algorithm is rather... inefficient.
> 
> I am not sure how to write such test. I'll see, never really worked with gifs in
> Emacs.

If it doesn't matter to you, it's probably not worth your while. We
know why it's slow (we need to cache the intermediate images).
-- 
Alan Third