• Vlad Levenfeld (81/81) Jan 10 2015 Hoping to gauge interest and get some advice on this:
• bearophile (5/11) Jan 10 2015 Are those bugs in Bugzilla?
• Vlad Levenfeld (6/8) Jan 10 2015 Understandable, I'll change N to Nat and R to Real or something

"Vlad Levenfeld" <vlevenfeld gmail.com> writes:

Hoping to gauge interest and get some advice on this:

Thanks to Kenji's multidimensional operators, I've managed to 
make a template library for generating multidimensional 
structures with arbitrary index types. It uses mixin templates 
that take some user-defined primitives as arguments and uses them 
to generate uniform-syntax interfaces in several layers 
(depending on how much functionality you feel like defining) from 
simple indexing to RAII-enabled buffers.

I've been using it to unify GPU data, local arrays and 
procedurally generated data under generic algorithms and I'm 
pretty happy with how it works. Its very easy to use (though as 
the writer I am definitely biased so take that with some salt). 
Its got hooks for customization and optimization (e.g. 
transferring data between GPU arrays doesn't cross the PCI bus, 
you can make it so expensive-to-read data sources precache 
themselves before copying to target arrays, things like that), 
debug-build bounds checking (though I'd prefer to tie it to the 
noboundscheck flag if possible), support for negative and 
non-integer indices, and comes complete with a set of functional 
primitives, like the ones in Phobos but obviously had to be 
reimplemented to work with multidimensional structures.

Regarding the functional bits - unittests are pretty extensive, 
map and zip and the like are capable of  nogc, capturing frame 
pointers, CTFE, some other stuff I'm forgetting, and there's a 
full suite of type-level functional primitives too (like 
staticMap, I've got Sort and Reduce and pattern matching stuff, 
just things I needed to make the multidim stuff work but turned 
out to be pretty damn useful in their own right).

Really there's a metric crapton of functionality all across the 
board, I've spent the last several months as a shut-in building 
this thing, and I really just need to get around to doccing it 
properly. The stuff I've listed just scratches the surface.

Here's a little example of the syntax it enables - I make two 
procedurally-generated checkerboard textures of different 
coarseness and copy a subregion of one into the other:

	auto tex1 = ℕ[0..100].by (ℕ[0..100]).map!((i,j) => (i+j)%2? red: 
yellow).Texture;
	auto tex2 = ℕ[0..50].by (ℕ[0..50]).map!((i,j) => (i+j)%2? blue: 
green).grid (100,100).Texture;

	tex1[50..75, 25..75] = tex2[0..25, 0..50];

where "by" is a cartesian product, "N" is the natural numbers (I 
also have R), grid is a resampling operator (so you could 
ostensibly use some subset of R by R and then cast grid on it to 
discretize, but I've also implemented editable spaces with R and 
a stack of delegates, to be a bit like Mathematica).

Which reminds me - making multidim structures doesn't necessarily 
involve using the mixins at all, plenty can be accomplished by 
functional composition. But I digress - like I said, scratching 
the surface.

But the amount of possibilities and features means doccing all 
the features would be a major effort, and I'm pretty busy with 
school and stuff... I want to know what the interest is, because, 
reading the forums, it sounds to me like other people could use 
something like this, and if I'm right, then I'll definitely 
prioritize it.

But the other problems are this -
1) there's a nontrivial amount of supporting code, some of which 
is basically reimplementations of some parts of Phobos (mostly 
std.algorithm, I don't use too many dirty hacks though, mostly 
workarounds for dmd bugs; its fairly idiomatic D although I 
haven't tested with immutable yet). I don't know how much of a 
problem this is, if it is a problem at all.
2) while the thing has been stable under my personal use, I have 
no idea how much it will need to change if/when others start 
using it (and I've heard that once you make the first release, 
you are locked into your public API, for better or worse). I've 
never tried to open source anything before or offer any kind of 
"public service" so this is sort of new world for me and I'm not 
sure what to expect. I feel I especially need some feedback in 
this area so I don't waste my efforts doing things that aren't 
useful.

So, in summary:

The point of this lib is to provide a level of standardization to 
multidimensional structures and operators while providing a suite 
of functional tools and handling the petty concerns to let the 
user focus on business logic.

What I want to know is: Is there interest? What points should I 
have covered before I throw the lib into the outside world? Or 
should I just open it up and deal with issues as they arise? Any 
advice on how to go about creating documentation for it that will 
actually be helpful?

"bearophile" <bearophileHUGS lycos.com> writes:

Vlad Levenfeld:

 	auto tex1 = ℕ[0..100].by (ℕ[0..100]).map!((i,j) => (i+j)%2? 
 red: yellow).Texture;
 	auto tex2 = ℕ[0..50].by (ℕ[0..50]).map!((i,j) => (i+j)%2? 
 blue: green).grid (100,100).Texture;

 	tex1[50..75, 25..75] = tex2[0..25, 0..50];

The library seems nice, but I don't like Unicode identifiers.


 mostly workarounds for dmd bugs;

Are those bugs in Bugzilla?

Bye,
bearophile

"Vlad Levenfeld" <vlevenfeld gmail.com> writes:

On Saturday, 10 January 2015 at 22:56:28 UTC, bearophile wrote:
 The library seems nice, but I don't like Unicode identifiers.

Understandable, I'll change N to Nat and R to Real or something 
and just make unicode aliases in my own code.

 Are those bugs in Bugzilla?

They are, I tag them with // BUG followed by the url to the 
bugzilla issue, so when I get an email about a fix I can grep for 
them and clean up.