Make dot more generic (and safer)

[joehuchette]

Also adds a missing method so that e.g. +(::OneIndexedArray, ::OneIndexedArray) works as expected.

[joehuchette]

SCS :'(

[mlubin]

This is slow in every case, can't we do better? At least dot(Vector{Float64},OneIndexedArray{Variable}) should be very fast.

[joehuchette]

I'd rather just get rid of multidimensional dot, to be honest, considering
it's not in Base.
On Tue, Jun 2, 2015 at 6:34 AM Miles Lubin notifications@github.com wrote:

In src/operators.jl
#459 (comment):

-Base.dot{T<:Real,N}(lhs::Array{T,N}, rhs::JuMPArray{Float64,N}) = dot(vec(lhs), vec(rhs.innerArray))
-Base.dot{T<:Real,N}(lhs::JuMPArray{Float64,N}, rhs::Array{T,N}) = dot(vec(rhs), vec(lhs.innerArray))
-Base.dot{T<:Real,N}(lhs::Array{T,N}, rhs::Array{Variable,N}) = AffExpr(vec(rhs), vec(float(lhs)), 0.0)

-Base.dot{T<:Real,N}(rhs::Array{Variable,N}, lhs::Array{T,N}) = AffExpr(vec(rhs), vec(float(lhs)), 0.0)

-function Base.dot{N}(lhs::JuMPArray{Variable,N},rhs::JuMPArray{Variable,N})

• size(lhs.innerArray) == size(rhs.innerArray) || error("Incompatible dimensions")
• return QuadExpr(vec(lhs.innerArray), vec(rhs.innerArray), ones(length(lhs.innerArray)), AffExpr())
  +Base.dot{T,S,N}(lhs::OneIndexedArray{T,N},rhs::OneIndexedArray{S,N}) = _dot(lhs.innerArray, rhs.innerArray)
  +Base.dot{T,S,N}(lhs::OneIndexedArray{T,N},rhs::Array{S,N}) = _dot(lhs.innerArray, rhs)
  +Base.dot{T,S,N}(lhs::Array{T,N},rhs::OneIndexedArray{S,N}) = _dot(lhs, rhs.innerArray)
  +Base.dot{T<:JuMPTypes,S,N}(lhs::Array{T,N},rhs::Array{S,N}) = _dot(lhs,rhs)
  +Base.dot{T<:JuMPTypes,S<:JuMPTypes,N}(lhs::Array{T,N},rhs::Array{S,N}) = _dot(lhs,rhs)
  +Base.dot{T,S<:JuMPTypes,N}(lhs::Array{T,N},rhs::Array{S,N}) = _dot(lhs,rhs)

+function _dot{T,S}(lhs::Array{T}, rhs::Array{S})

This is slow in every case, can't we do better? At least
dot(Vector{Float64},OneIndexedArray{Variable}) should be very fast.

—
Reply to this email directly or view it on GitHub
https://github.com/JuliaOpt/JuMP.jl/pull/459/files#r31513366.

[tkelman]

there is now a vecdot in 0.4

[joehuchette]

Cool, I thought I remembered something about to be added. I'll change the naming then.