I tried with sub/slice a while ago and I think it gave me a very large
performance decrease. I will try again when I get to a good computer to
test it.
The "`AbstractVecOrMat` with a default value for idx"-method sounds
interesting though. I will try it out.
Thank you.
On Monday, February 16, 2015 at 2:34:23 PM UTC+1, Andreas Noack wrote:
>
> You could Use `AbstractVecOrMat` and `idx=1` as default value, but you
> could also use `sub` or `slice` to avoid copying the matrix columns since
>
> *julia> **A = randn(2,2)*
>
> *2x2 Array{Float64,2}:*
>
> * -0.124018 1.3846 *
>
> * -0.259116 -1.19279*
>
>
> *julia> **isa(sub(A, :,1), AbstractVector{Float64})*
>
> *true*
>
> 2015-02-16 8:11 GMT-05:00 Kristoffer Carlsson <[email protected]
> <javascript:>>:
>
>> Actually the best thing would be if I could call a function with the
>>
>> dist(point_1::AbstractVector{T}, point_2::AbstractVector{T})
>>
>> signature because then I could just use all the stuff in the Distances.jl
>> package
>>
>>
>>
>> On Monday, February 16, 2015 at 2:05:17 PM UTC+1, Kristoffer Carlsson
>> wrote:
>>>
>>> I am writing some code where I sometimes want to calculate the distance
>>> between two vectors and sometimes the distance between a column in a matrix
>>> and another vector.
>>>
>>> This is in a very hot spot of my program so any unnecessary copying is
>>> unacceptable.
>>>
>>> What I am currently doing is basically having two versions of the
>>> distance function, one for vector-vector and one for column-vector:
>>>
>>> # Reduced euclidian distances
>>>
>>> # Vector-Vector
>>> @inline function euclidean_distance_red{T <: FloatingPoint}(point_1::
>>> AbstractVector{T},
>>> point_2::
>>> AbstractVector{T})
>>> dist = 0.0
>>> for i = 1:length(point_1)
>>> @inbounds dist += abs2(point_1[i] - point_2[i])
>>> end
>>> return dist
>>> end
>>>
>>> # Column-Vector
>>> @inline function euclidean_distance_red{T <: FloatingPoint}(data::
>>> AbstractMatrix{T},
>>> point::
>>> AbstractVector{T},
>>> idx::Int)
>>> dist = 0.0
>>> for i = 1:length(point)
>>> @inbounds dist += abs2(data[i, idx] - point[i])
>>> end
>>> return dist
>>> end
>>>
>>>
>>> This feels unnecessary since the transformation from the second to the
>>> first function is obvious. Is there anyway I can collapse these two
>>> functions into one that will come with zero performance loss?
>>>
>>> Best regards,
>>> Kristoffer Carlsson
>>>
>>
>
