Improving Cantabile's Text Rendering


#41

I have use cases where support for some simple tables would be nice. Or at least some sort of column layout. I.e. each line of a lyric may not be very long and I could get more on one screen if I could put the 1st verse on left and 2nd verse on the right.

Other use cases would be to have the lyrics/chords on left and some other notes on the right.


#42

Interesting - I hadn’t considered that kind of layout. Let me think about it. In any case it’s a higher level construct than what I’m trying to solve right now.


#43

I could. :slight_smile:

That’s just a function of my test program where I can increase the size of the font’s being used so it’s technically not a zoom. Still it’s an interesting point but makes me wonder what you might really want here as far as show notes is concerned:

  1. The ability to zoom the entire show notes panel. or
  2. The ability to quickly increase the size of a particular element in the show notes?

Brad


#44

So today was a bit of a set back…

My plan was to integrate this new text rendering support into Cantabile with the goal of just fixing the current issues with text rendering.

However I was a little concerned about the performance and memory usage of some of the code I ported from other source. On closer inspection, I’m really not happy with the Bidi algorithm that I ported from the Unicode’s reference implementation. It does the job, it seems fast enough, but it’s terrible as far as memory management is concerned. Given how much and how often Cantabile needs to draw text it might have an impact on memory usage and put extra pressure the garbage collector.

So I need a better implementation and after spending all morning looking for one, there just doesn’t seem to be that fits all the requirements. It looks like I might need to write it myself which isn’t a trivial task - probably about a week’s work - the spec is over 40 printed A4 pages.

I don’t know if this is a path I want to take. I’ll read the spec again before making a decision . :roll_eyes:

Brad


#45

That sounds like a school punishment! :flushed:

I think it must zoom the whole text with keeping the format and indent.
Most important when you have chords above text lines.


#46

I stayed up too late last night working on this but this morning finished a rewrite of the bidi algorithm.

First run passed 670k out of 770k tests with no crashes. Not too bad.

(I’m using a massive set of test cases that the Unicode consortium provide).

Unfortunately it was excruciatingly slow. Like really, really, really slow. That 34 minutes should be more like 2 seconds.

Probably a stupid mistake on my part…


#47

Yep, stupid mistake… was forgetting to clear one of the working buffers used by the algorithm. 0.78 seconds is more like it :slight_smile:

Next step is to get it passing all the tests.


#48

0.78 is :rocket: fast !
:zap::zap::zap::zap::zap::zap:


#49

Another hundred thousand tests passing. So close!


#50
    🌕

:rocket:

Almost there :blush:


#51

All tests passing!

Also comparing the new version (on the right) to the old version (on the left):

  • It’s faster (.79 seconds vs 1.42 seconds),
  • uses about a quarter of the memory and
  • resulted in far fewer garbage collections!

Tomorrow’s job: there’s another set of test cases I need to test with, and then more performance and memory tuning.


#52

Come on man you’re killing us :joy::joy::joy:


#53

the trials and tribulations of programming. I am glad that for my Java librarians, I have stuck with the good old Queen’s English!

I can appreciate what Brad is doing for a product with far wider reach. :slight_smile:


#54


#55

Brad, you’re amazing! I have no idea what any of this means…

I assume you’re going to sell this back to the Bidi/Unicode guys for a vast sum of money?


#56

Fixed a couple of minor bugs and the new implementation now passes the final 90k test cases. Also, it’s more than twice as fast as the old routine for the same data.

Also, a bit more perf tuning got the first set of test data down to 0.58 seconds.

PS: I’m not normally this paranoid with performance tuning UI related code, but this is big chunk of fairly complicated code that will be used for most of the text that Cantabile needs to display on screen. It needs to be rock solid, fast and not put pressure on the garbage collector.


#57

Tracked down that last 3mb of memory usage, making the new implementation almost 4x faster than the old one (for the second set of test data) and uses 8Kb instead of 240Mb.

I’ve got one more performance experiment I want to try then I’ll be taking a closer look at the line breaking algorithm (but I think it might be ok as is).


#58

Amazing!
Kinda like programmers porn!


#59

Does that stuff still makes fun? :smiley:


#60

Ah, the joys of code optimisation. I still remember the heady days of the late 80s when you really had to squeeze very last drop out of processor performance. Often that meant dropping to assembly language (as high level language compilers were still not brilliant in terms of code efficiency), overlapping integer and floating point unit operations (a floating point multiply might take, for example, 20 clock cycles, during which you could do several integer operations before needing the floating point result), keeping variables in processor registers to avoid unnecessary memory writes, etc. , etc.

I remember demonstrating to a sceptical former boss a “Sobel” convolution edge detection image processing algorithm running over 20 times faster in assembly language compared to the PDP11 Fortran equivalent.

These days, I’ll just buy a faster processor or more memory :slight_smile: but there was a certain satisfaction when you got something running on steroids in the old days, so hopefully Brad is feeling the same way right now over this! :slight_smile:

So, huge Kudos to what Brad is doing on this area. Sounds like real progress that will benefit Cantabile as a multi-national/language platform (just look at the spread of users on the “where are you from” thread). :slight_smile: