Improve cache localization in default_blasmul

[jishnub]

The order of the loop was cache unfriendly, but it was written this way because the range of the row variable may depend on the column one. However, we may carry out a quick check to ascertain whether the row range is actually independent of the column one, in which case we may switch them and loop in a cache-friendly manner. This branching improves performance in banded*dense matmul. The original behavior is retained in the fallback branch.

julia> B = brand(4000, 4000, 5, 5);

julia> X = rand(4000, 4000);

julia> @btime $B * $X;
  977.570 ms (2 allocations: 122.07 MiB) # master
  518.942 ms (2 allocations: 122.07 MiB) # PR

[codecov]

Codecov Report

Patch coverage: 93.33% and project coverage change: +2.10 🎉

Comparison is base (0a9a57c) 86.24% compared to head (b0e8569) 88.34%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #130      +/-   ##
==========================================
+ Coverage   86.24%   88.34%   +2.10%     
==========================================
  Files          10       10              
  Lines        1708     1716       +8     
==========================================
+ Hits         1473     1516      +43     
+ Misses        235      200      -35     

Impacted Files	Coverage Δ
src/muladd.jl	75.29% <93.33%> (+0.81%)	⬆️

... and 6 files with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.