Blog/Wiki/Documentation on your approach to writing glas and gemm algorithm

[dataPulverizer]

Blog/Wiki/Documentation on your approach to writing glas and gemm algorithm

[9il]

Hi

The only D specific features was used is static if, static foreach, and pragma inline.

Other features is blocking, GLAS blocking is improved version of https://www.cs.utexas.edu/users/pingali/CS378/2008sp/papers/gotoPaper.pdf. LDC LLVM IR allows to make it generic.

If you are interesting to contribute or write an article we can chat in more details

[dataPulverizer]

Thanks for the paper, I shall read it today! I would be happy to begin by writing an article on the implementation details of gemm with you. I will start with me asking lots of questions! The best way might be for me to make some commented questions on a forked copy of the glas library and you can write your answers there. Once I have a good idea of how the code works for instance how it uses llvm to improve performance, how it does pre-fetching, and cpu specific targeting and the role D's template/library tools play in composing all these together, I would be happy to write the article with you.

[9il]

Looks good

[dataPulverizer]

I have started looking at the code in your glas library and decided to start with a small module with fewest dependencies to begin to understand the library. I have forked the library and made some comment questions on the glas.ndslice.d module:

https://github.com/dataPulverizer/mir-glas-comments/commit/20ef2b8a0d3fd621d2844103fc42081c61b71851

I await your comments. Cheers

[RoyiAvital]

Is MIR Multi Threaded (Like Eigen) or only Single Thread?

[9il]

1. Mir GLAS is single thread and experimental.
2. For other mir parts can be easily parallelised with https://dlang.org/phobos/std_parallelism.html as well as with some more efforts with C's alternatives