Allocating lmul!

[rveltz]

Hi,

This is not really an issue. I am wondering if the following is expected to allocate:

julia> F = zeros(31,61);

julia> PS = FastSphericalHarmonics.plan_sph_synthesis(copy(F))
FastTransforms plan for FFTW Fourier synthesis on the sphere for 31×61-element array of Float64

julia> @time lmul!(PS,F);
  0.000356 seconds (147 allocations: 10.453 KiB)

julia> @time lmul!(PS,F);
  0.000368 seconds (147 allocations: 10.453 KiB)

I am trying to optimize FastSphericalHarmonics.sph_evaluate! by caching the plans and stumbled on this allocation.

[MikaelSlevinsky]

I'm getting 0 allocations when using FastTransforms alone:

julia> using FastTransforms, LinearAlgebra

julia> F = zeros(31, 61);

julia> PS = plan_sph_synthesis(copy(F))
FastTransforms plan for FFTW Fourier synthesis on the sphere for 31×61-element array of Float64

julia> @time lmul!(PS, F);
  0.000379 seconds

[rveltz]

I guess it is a version problem. It seems to allocate on

[057dd010] FastTransforms v0.15.16

but I cant get higher because of FastSphericalHarmonics

[rveltz]

Well:

julia> using FastTransforms, LinearAlgebra

julia> F = zeros(31, 61);

julia> PS = plan_sph_synthesis(copy(F))
FastTransforms plan for FFTW Fourier synthesis on the sphere for 31×61-element array of Float64

julia> @time lmul!(PS, F);
  0.000774 seconds (147 allocations: 10.453 KiB)

julia> @time lmul!(PS, F);
  0.000683 seconds (147 allocations: 10.453 KiB)

with:

  [7d9f7c33] Accessors v0.1.38
  [28f2ccd6] ApproxFun v0.13.27
  [b4f34e82] Distances v0.10.11
  [31c24e10] Distributions v0.25.111
  [057dd010] FastTransforms v0.16.4
  [e9467ef8] GLMakie v0.10.10
  [bdcacae8] LoopVectorization v0.12.171
  [a3a9e032] NIfTI v0.6.0
  [d96e819e] Parameters v0.12.3
  [91a5bcdd] Plots v1.40.8
  [92933f4c] ProgressMeter v1.10.2
  [2913bbd2] StatsBase v0.34.3
  [276b4fcb] WGLMakie v0.10.10

[MikaelSlevinsky]

No idea!

[rveltz]

It is on osx M2.

[jishnub]

Could you possibly profile the code to check where the allocations arise from?

[rveltz]

I am surprised. In the REPL, it does not allocate but in VS code REPL it does ...

in vscode:

julia> @which lmul!(PS, F)
lmul!(p::FastTransforms.FTPlan{Float64, 2, FastTransforms.SPHERESYNTHESIS}, x::Matrix{Float64})
     @ FastTransforms ~/.julia/packages/FastTransforms/S7huq/src/libfasttransforms.jl:771

which is a ccall so not sure what to profile