Slow multiplication of reshaped matrices

[cossio]

Compare:

v = randn(728, 600)
w = randn(70, 728)
@time w * v #   0.017802 seconds (2 allocations: 328.172 KiB)

with:

v_ = reshape(reshape(v', 600, 728)', 728, 600)
w_ = reshape(reshape(w', 728, 70)', 70, 728)
@time w_ * v_   0.161401 seconds (5 allocations: 358.922 KiB)

The second one is 10x slower. Note that the underlying data in memory is exactly the same, and with the same layout.

Can the reshape / adjoint type signature be simplified here?

[mcabbott]

In this case w == w_ so the simple fast option is not to call reshape at all. For general reshapes, I don't think there's any obvious way to simplify wrappers, and the layout will be wrong for BLAS.

In most such cases, it's probably quicker to copy:

julia> @btime $w_ * $v_;
  17.546 ms (5 allocations: 358.92 KiB)

julia> @btime collect($w_) * collect($v_);
  1.717 ms (6 allocations: 4.04 MiB)

That's what * does for some mixed element type cases, maybe it could be done automatically here? Dispatching on ReshapedArray{T,2,<:Adjoint{T,<:ReshapedArray{T,2,<:Adjoint{T,<:StridedArray{T}}}}} where {T<:BlasFloat} sounds messy. Maybe it need not be so restricted?