Update throughput and latency cost calc for shuffling reductions

[chriselrod]

Also, it heuristically doubles the estimated throughput of apple-m*. This check should work regardless of whether running Mac or Linux, as long as the CPU is recognized.

@heltonmc

julia> function dot_turbo_ls(ca::AbstractVector{Complex{T}}, cb::AbstractVector{Complex{T}}) where {T}
           a = reinterpret(reshape, T, ca)
           b = reinterpret(reshape, T, cb)
           re = zero(T)
           im = zero(T)
           LoopVectorization.@turbo_debug for i ∈ axes(a, 2)
             re += a[1, i] * b[1, i] + a[2, i] * b[2, i]
             im += a[1, i] * b[2, i] - a[2, i] * b[1, i]
           end
       end
dot_turbo_ls (generic function with 1 method)

julia> T = Complex{Float64}; n = 203;

julia> xv = rand(T, n);

julia> yv = rand(T, n);

julia> ls = dot_turbo_ls(xv,yv);

julia> LoopVectorization.choose_order_cost(ls)
rt = 4.5
sl = 12
op = var"##op#307" = LoopVectorization.vfmadd_fast(var"##op#303", var"##op#304", var"##op#308")
rt = 0.5
sl = 8
op = var"##op#309" = LoopVectorization.vfmadd_fast(var"##op#303", var"##op#306", var"##op#310")
(latency, rt) = (12.0, 5.0)
([:i], :i, Symbol("##undefined##"), :i, 4, -1, 2970.3953125000003, true)

It'll now unroll the complex dot product by 4, because 12 / 5 = 2.4.
It'll round 2.4 up to the nearest power of 2, i.e. 4.

[codecov]

Codecov Report

Patch coverage: 92.85% and project coverage change: +2.18 🎉

Comparison is base (9eb6f63) 84.58% compared to head (70622b0) 86.76%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #488      +/-   ##
==========================================
+ Coverage   84.58%   86.76%   +2.18%     
==========================================
  Files          38       39       +1     
  Lines        9489     9516      +27     
==========================================
+ Hits         8026     8257     +231     
+ Misses       1463     1259     -204     

Impacted Files	Coverage Δ
src/modeling/determinestrategy.jl	95.60% <92.85%> (+2.12%)	⬆️

... and 30 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.