This could possibly increase performance. With combination of the la

<div class="highlight highlight-source-julia notranslate position-relative overflow-auto" dir="auto"

Make billiard tables be `Tuple` instead of `Vector` about dynamicalbilliards.jl HOT 2 CLOSED

Datseris commented on August 23, 2024

Make billiard tables be `Tuple` instead of `Vector`

from dynamicalbilliards.jl.

Comments (2)

Datseris commented on August 23, 2024

The package to use here is Unrolled.jl.

But a benchmark on whether this makes the code slower or not should be made.

The benchmark can be done using the current implementation of:

function next_collision(
    p::AbstractParticle{T}, bt::Vector{<:Obstacle{T}})::Tuple{T,Int} where {T}
    tmin::T = T(Inf)
    ind::Int = 0
    for i in eachindex(bt)
        tcol::T = collisiontime(p, bt[i])
        # Set minimum time:
        if tcol < tmin
            tmin = tcol
            ind = i
        end
    end#obstacle loop
    return tmin, ind
end

from dynamicalbilliards.jl.

Datseris commented on August 23, 2024

@btime typeof($bt2[2]) <: PeriodicWall;

  1.706 ns (0 allocations: 0 bytes)
@btime typeof($bt[2]) <: PeriodicWall;

  90.531 ns (0 allocations: 0 bytes)

bt = billiard_buminovich()
p = randominside()
bt2 = (bt...)
@btime next_collision($p, $bt);

  1.237 μs (8 allocations: 128 bytes)
@btime next_collision($p, $bt2);

  36.266 ns (0 allocations: 0 bytes)

the only difference is that bt is vector while bt2 is tuple.

We may not need unrolled at all, although because unrolled uses generated functions, the result specialized on the exact structure of the billiard table, which will make it suuuuuper efficient!

from dynamicalbilliards.jl.

Recommend Projects