replace splat with Tuple

[wheeheee]

For a random matrix M of size (500, 500), allocations measured for unwrap(M; dims=1:2) drops from 2.2M+ to 250k, with a small speedup of about 30% on Windows, Julia 1.9.

This is achieved by changing every instance of splatting a vector into CartesianIndex{N} to Tuple.

[martinholters]

Yes, most of these look like optimizations that should be done. IIUC, the 30% performance improvement you've reported was only for the first step of replacing the splats with Tuple calls? What extra benefit to you get from the other optimizations?

[wheeheee]

I get a small (~3ms) speedup on a 50x50x50 array for using sum instead of the manually unrolled version.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (3e3524f) 97.29% compared to head (32a8298) 97.46%.
Report is 7 commits behind head on master.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #513      +/-   ##
==========================================
+ Coverage   97.29%   97.46%   +0.17%     
==========================================
  Files          18       18              
  Lines        3105     3080      -25     
==========================================
- Hits         3021     3002      -19     
+ Misses         84       78       -6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[martinholters]

I did a bit of benchmarking:

# Julia 1.0.5
julia> @btime unwrap(A; dims=1:2) setup=A=rand(500,500);
  154.448 ms (250021 allocations: 53.34 MiB) #master
  153.964 ms (250014 allocations: 53.34 MiB) #PR

julia> @btime unwrap(A; dims=1:3) setup=A=rand(50,50,50);
  118.314 ms (125025 allocations: 43.18 MiB) #master
  104.837 ms (235608 allocations: 49.93 MiB) #PR

# Julia 1.6.7
julia> @btime unwrap(A; dims=1:2) setup=A=rand(500,500);
  216.799 ms (2246030 allocations: 121.87 MiB) #master
  148.143 ms (250024 allocations: 53.34 MiB) #PR

julia> @btime unwrap(A; dims=1:3) setup=A=rand(50,50,50);
  111.778 ms (125034 allocations: 43.18 MiB) #master
  96.549 ms (125026 allocations: 43.18 MiB) #PR

# Julia 1.9.4
julia> @btime unwrap(A; dims=1:2) setup=A=rand(500,500);
  208.574 ms (2246032 allocations: 121.87 MiB) #master
  143.879 ms (250026 allocations: 53.34 MiB) #PR

julia> @btime unwrap(A; dims=1:3) setup=A=rand(50,50,50);
  110.831 ms (125036 allocations: 43.18 MiB) #master
  97.857 ms (125028 allocations: 43.18 MiB) #PR

So interestingly, between Julia v1.0.5 and v1.6.7, we've lost quite some performance for the 2d case which this PR successfully recovers almost without negative side-effects. There is a significant increase in the allocations for the 3d case on Julia v1.0.5, but a) it's a bit faster nevertheless and b) I don't think we should worry too much about Julia older than 1.6.