Codecov Report

Merging #31 into master will increase coverage by 1.25%.
The diff coverage is 80%.

@@            Coverage Diff             @@
##           master      #31      +/-   ##
==========================================
+ Coverage   55.19%   56.44%   +1.25%     
==========================================
  Files          14       15       +1     
  Lines         924      946      +22     
==========================================
+ Hits          510      534      +24     
+ Misses        414      412       -2

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 1c63bd0...37e8474. Read the comment docs.

What do you think about using Applied instead of hacking Broadcasted for the BLAS MulAdd operations? That is,

Y .= @~ α .* (A * x) .+ β .* c

would become

Y .= @~ α * A * x + β * c

which is code for

Y .= applied(+, applied(*, α, A, x), applied(*, β, c))

Using ApplyStyle we could also have a much more robust way of rearranging complicated MulAdd operations to reduce to MulAdd.

@~ α * A * x + β * c sounds good to me. It looks cleaner.

In a long run, I think it would be nice if both @~ α .* (A * x) .+ β .* c and @~ α * A * x + β * c can be "normalized" to a same representation via instantiate mechanism. This way, users don't have to know when to use, e.g., + or .+ to invoke performance specializations.

Yes, I suppose both could work by an intelligent use of ApplyStyle and BroadcastStyle.

In the meantime, I'll go ahead and merge this PR.