[ft-devel] Revisiting LSB [4]

[Ulrich]

George Williams writes: 

> On Fri, 2005-10-28 at 03:54, Ulrich wrote:

> The algorithms used by FreeType change over time. 
>
> Sorry. I mean that using the bytecode interpreter it might be possible
> to define a "correct" result for a glyph. Using the autohinter it is not
> possible because today's autohinter is different from last year's. If
> the autohinter chooses to move a spline a little bit more to the left
> because we think it looks better today that will invalidate all test
> suites based on the previous behavior.

It seems that we are aware of this. 

> That's why I believe the testsuite is useless. I think this will happen
> far too frequently, that the overhead involved in checking minor changes
> will not be worth the results gained.  
>
> It is not like a compiler testsuite where there is only one right answer
> (except possibly in the case above). So any change to the autohinter
> will probably introduce irrelevant changes in the results which will
> require much checking and correction.

I'm not sure whether these changes are really irrelevant. 

> It's more than just shape. It's the difference between a pixel having
> 0xfe and 0xfd.

Sorry for using the wrong term "shape". It should have been "result". I 
assume that every single change in the result has to be checked as one pixel 
being on or off might have a servere impact on the result's quality. 

> > I would appreciate it if one of the experts shed some light on that. 
> Anyway that remains my opinion. Your view keeps being repeated, so
> obviously I am not convincing. Perhaps if you write the testsuite and
> show that it is useful you can confound me.

We, at least, found that the results are deterministic and that feasibility 
of a test suite heavily depends on how often the results of the 
rasterisation will change. This shows that technically we have the same 
situation as with every other test suite like e.g. for a compiler; of 
course, only as long as we are interested in assuring the quality of the 
results. 

What I still miss to decide whether collecting a set of rasterisation 
results should be considered are some data, some numbers of how often 
rasterisation results have changed in previous releases of FreeType 2. 

It might be a good starting point to collect the rasterisation results of 
only a subset from some fonts that contains the most critical glyphs and a 
carefully choosen set of characters that exposes all operations. It could be 
sufficient to only include the result for "F" or "E" but not both; "I" and 
"L" might be left out entirely a.s.o.; of course, depending on the font, 
style, size ... 

Anyway, sooner or later I will start collecting those data - thanks for the 
hint.