WIP: Shared arrays and GPU arrays

[mdavezac]

Replaces and extends #12

@dlfivefifty, I have some benchmark results for GPUs (K1000). I'll try and send a plot ASAP. The gist is that for relatively large blocks there is a speedup of ~4-5 compared to using a BlockArray backend. For similar matrices, there is also speedup of the BlockArray vs PseudoBlockArray backend (say of 1.x). Not sure whether it comes from the overhead from the Mul and friends setup, or whether it is tied to memory locality, or something else.

The shared array example works, though I have not benchmarked it yet.

[mdavezac]

On a different note, f3d26d5 adds a .editorconfig. If you think it's worth it, it might help harmonize coding styles across contributors until somebody writes a code formatter for Julia.

[dlfivefifty]

Awesome! Mul has no overhead as it is compiled away (unless there’s a bug). It’s not too surprising that for large bandwidths the BlockArray backend is faster: BLAS is slower for large strides.

The reason I chose PseudoBlockArray was for small blocks, but it’s possible in retrospect that it was not the best choice.

[codecov-io]

Codecov Report

Merging #19 into master will not change coverage.
The diff coverage is n/a.

@@          Coverage Diff           @@
##           master     #19   +/-   ##
======================================
  Coverage    48.4%   48.4%           
======================================
  Files           8       8           
  Lines         849     849           
======================================
  Hits          411     411           
  Misses        438     438

Impacted Files	Coverage Δ
src/linalg.jl	32.66% <ø> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 09f8364...f3d26d5. Read the comment docs.