Implementing #57: Real coefficients

[Wikunia]

First go of implementing real coefficients.

• Check everything again
• Check that there is no performance regression
• Is it reasonable to support Rational weights in a different way?

[codecov]

Codecov Report

Merging #58 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master      #58   +/-   ##
=======================================
  Coverage   98.93%   98.93%           
=======================================
  Files          20       20           
  Lines        1407     1407           
=======================================
  Hits         1392     1392           
  Misses         15       15

---

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 db86e58...cc899fd. Read the comment docs.

[matbesancon]

To stay efficient when using non-concrete types, better parameterize with a generic:

mutable struct LinearVariables{T <: Real}
    indices             :: Vector{Int}
    coeffs              ::  Vector{T}
end 

Instead of Vector{Real}

[matbesancon]

When creating a LinearConstraint, either promote every real to the same type, or use different types for the different field types:

    mins                :: Vector{TM1}
    maxs                :: Vector{TM2}
    pre_mins            :: Vector{TP1}
    pre_maxs            :: Vector{TP1l} 
  # etc

[Wikunia]

Thanks for having a look:
I only need the linear variables if JuMP isn't used as the modelling layer, therefore I'm not sure whether it's reasonable to longer support that anyway. Maybe you have some thoughts on that. If JuMP is used all variables in a constraint will have the same type which is currently: Float64 anyway.

[matbesancon]

LinearVariables might be a tricky name, maybe LinearCombination, or LinearFunction? I think you should not hard-code Float64, JuMP will eventually get generic numerical types: jump-dev/JuMP.jl#2025
Although it's not on the immediate roadmap I believe

[matbesancon]

I only need the linear variables if JuMP isn't used as the modelling layer

I think it is worth keeping, given the mention above that JuMP does not yet support generic numeric types

[Wikunia]

When creating a LinearConstraint, either promote every real to the same type, or use different types for the different field types:

What is the easiest way to get the common type of values?
Thought Julia would do that automatically but looks like it doesn't.

[matbesancon]

What is the easiest way to get the common type of values?

# on values
julia> promote(3, false)
(3, 0)

# on types
julia> promote_type(Float32, Float64)
Float64

[Wikunia]

Implemented:

• LinearVariables -> LinearCombination
• Generic type for LinearCombination

[codecov]

Codecov Report

Merging #58 into master will not change coverage.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master      #58   +/-   ##
=======================================
  Coverage   98.93%   98.93%           
=======================================
  Files          20       20           
  Lines        1407     1407           
=======================================
  Hits         1392     1392           
  Misses         15       15

Impacted Files	Coverage Δ
src/ConstraintSolver.jl	99.12% <100%> (ø)	⬆️

---

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 db86e58...7c32e07. Read the comment docs.

[Wikunia]

Checked that there are no changes compared to master for killer sudoku regarding number of steps. The runtime might be slightly worse up but not significantly.

[Wikunia]

If there are no other things @matbesancon I would merge this.

[matbesancon]

Yup let's go with it :)

[matbesancon]

anyway tests are passing, future things to do can be in other PRs