search

matrixfunctions / GraphMatFun.jl

Computation graphs for matrix functions
MIT License
13 stars 0 forks source link

JuliaLang code generation slow I-operator #18

Closed jarlebring closed 3 years ago

jarlebring commented 3 years ago

Documentation of an upcoming change.

Currently, addition with identity in the generated Julia code becomes code like

    # Computing Ba5_2 with operation: lincomb
    coeff1=-9.932061741024752e-6
    coeff2=0.00035453728444091774
    memslots[6][:]=coeff1*I + coeff2*memslots[2]

This is slow.

In-place addition with identity operator can be done like this:

  copy!(memslots[6],memslots[2]);
  memslots[6] .*= coeff2;
  memslots[6] .+= coeff1*I

However, for some reason (I think it's because it reduces to a blas-call similar to our blas code) this is faster

  copy!(memslots[6],memslots[2]);
  memslots[6] .*= coeff2;
  D=view(memslots[6], diagind(memslots[6], 0)); # Create a view of the diagonal
  D .+= coeff1;
jarlebring commented 3 years ago

Tag 0126ea62f77cb27aeddf85bdd6d8a08803fe3bb9

jarlebring commented 3 years ago

Updated code generation works efficiently with identity op.