Speeding up my logsumexp function

General Usage Performance
, , ,
7

The problem with lsexp_mat is it allocates a lot of un-necessary matrices. If you re-write it as

function lsexp_mat1(mat; dims=1)
   max_ = maximum(mat, dims=1)
   sum_exp_ = sum(exp.(mat .- max_) .- mat .== max, dims=dims)
   log1p.(sum_exp_) .+ max_
end

it’s about as fast as the vector version.

2 Likes