[gsoc] Mir project

[Seb]

So here's another discussion thread for this year's GSoC. This 
time it's about the Mir project.
The wiki already contains a few infos:

https://wiki.dlang.org/GSOC_2019_Ideas#Mir_Project

@community: what do you miss the most in the Mir project?

Also, Ilya recently sent my a long email to one student with more 
details on the DataFrame project and I wanted to make it 
available to all students:

---

You can choose almost any project you can finish during GSoC.

Common requirements if you choose me as your mentor:
1. You should care about time limits, reports, and GSoC 
formalities yourself.
2. I wouldn't spend a lot of time on GSoC. Almost all the things 
you would need to understand yourself. I will only formulate a 
final goal, API and implementation requirement, intermediate 
goals. If you would do cool things I would help to do them even 
better.
3. The final result of the project should have a sensible 
positive impact on Mir and D in general. A project should be 
completely ready to be accepted.
4. A GSoC project should have professional quality. You would 
need to become a professional in the field you choose a GSoC 
project, this is a mandatory requirement.

For example, if you choose to implement basic matrix operations 
in D, then the two links to start would be:
- Anatomy of High-Performance Matrix Multiplication 
(https://www.cs.utexas.edu/users/flame/pubs/GotoTOMS_final.pdf)
- [Experimental] LLVM-accelerated Generic Linear Algebra 
Subprograms (https://github.com/libmir/mir-glas)

To work on GLAS you would need to understand well Goto's paper, 
LLVM IR, SIMD programming with LDC, GLAS source code.

DataFrame project
=================

mir-algorithm package (https://github.com/libmir/mir-random) has 
Slice/ndslice (numpy.ndarray analog) and Series (pandas.Series 
analog). Series should be fused into Slice, Slice would be a 
generalized multidimensional DataFrame analog. Labels (indexes) 
will be optional, the current Slice API and speed will be 
preserved. However, this would make the development of generic 
libraries hard. To make it simpler, we need to improve D language 
and DMD compiler. This can be split into two parts: language 
change (DIP)  and pull request with required changes in DMD.

The DataFrame GSoC project results will be accepted if you write 
the 'clever alias' DIP AND the DIP is approved by Andrei 
Alexandrescu and Walter Bright before the end of the GSoC AND you 
will also do at least one of the following:
1. Implement the DIP for DMD compiler. (DMD is written in D, but 
I have no idea about its internals) OR
2. Add Labels(Indexes) support to ndslices package to make Slice 
a generalization of DataFrame

It is quite a risky project, comparing with GLAS and FFT  the 
DataFrame project also requires very well communication skills, a 
lot of patience and some luck.

Links to start with for DataFrame:

https://issues.dlang.org/show_bug.cgi?id=16486
https://issues.dlang.org/show_bug.cgi?id=16465

The brief DIP idea is that the code like below should work:

alias PackedUpperTriangularMatrix(T) = Slice!(StairsIterator!(T*, 
"-"));

// fails, issue 16486
auto foo(T)(PackedUpperTriangularMatrix!T m)
{
}

// Current workaround: it is too crazy for users to
// know what is StairsIterator!(T*, "-")).
auto foo(T)(Slice!(StairsIterator!(T*, "-")) m)
{
}

Currently used Slice types in Lubeck / Production code
Slice!(double*) - D slice analog
Slice!(double*, 1, Universal) - BLAS vector, used in mir-lapack 
and mir-blas.
Slice!(double*, 2) - Contiguous matrix, that has an efficient 
loop for iteration over elements, see mir.algorithm.iteration 
sources.
Slice!(double*, 2, Canonical) - BLAS/LAPACK matrix 
representation, used in mir-lapack and mir.blas
Slice!(double*, N, Universal) - zero copy view to work with 
ndarray in numpy, see also low level API bindgins, and high level 
bindings
Slice!(StairsIterator!(double*, "+")) and ...
Slice!(StairsIterator!(double*, "-")) - packed storage for 
triangular matrixes, for BLAS/LAPACK
Slice!(ChopIterator!(size_t*, uint*)); - Memory efficient graph 
representation without labels.
Possible future Slice types (2019?):
Slice!(double*, 1, Contiguous, string*) - like Pandas Series
Slice!(double*, 2, Contiguous, LabelT1*, LabelT2*) - like Pandas 
DataFrame
Slice!(double*, 2, Contiguous, LabelT1*, LabelT2*, LabelT3*) - 
like Pandas Panel
Slice!(ChopIterator!(size_t*, uint*), 1, Contiguous, string*); - 
Memory efficient graph representation with labels.
Slice!(ChopIterator!(size_t*, Slice!(double*, 1, Contiguous, 
uint*))) - Sparse Matrix representation that can be used to 
interact with existing C/C++/Fortran libraries

If you would be able to write a good DIP and create a pull 
request with its implementation it would be awesome. I can pay 
400$ as a bonus if the DIP implementation is merged to DMD.
---