Move default alg choice sooner to improve inner loop stability

[ChrisRackauckas]

No description provided.

[ChrisRackauckas]

using OrdinaryDiffEq, ParameterizedFunctions, BenchmarkTools
using LinearSolve

f = @ode_def Orego begin
  dy1 = p1*(y2+y1*(1-p2*y1-y2))
  dy2 = (y3-(1+y1)*y2)/p1
  dy3 = p3*(y1-y3)
end p1 p2 p3

p = [77.27,8.375e-6,0.161]
prob = ODEProblem(f,[1.0,2.0,3.0],(0.0,30.0),p)
@btime solve(prob,Rosenbrock23(),abstol=1e-6,reltol=1e-6)
# Before:
# 1.405 ms (23476 allocations: 1.95 MiB)
# After:
# 1.423 ms (23478 allocations: 1.95 MiB)

@btime solve(prob,Rosenbrock23(linsolve=RFLUFactorization()),abstol=1e-6,reltol=1e-6)
# Before:
# 1.262 ms (10521 allocations: 952.25 KiB)
# After:
# 1.252 ms (10521 allocations: 952.25 KiB)

It's surprising this isn't helpful while making the choice is.

[ChrisRackauckas]

Fixed:

using OrdinaryDiffEq, ParameterizedFunctions, BenchmarkTools
using LinearSolve

f = @ode_def Orego begin
  dy1 = p1*(y2+y1*(1-p2*y1-y2))
  dy2 = (y3-(1+y1)*y2)/p1
  dy3 = p3*(y1-y3)
end p1 p2 p3

p = [77.27,8.375e-6,0.161]
prob = ODEProblem(f,[1.0,2.0,3.0],(0.0,30.0),p)
solve(prob,Rosenbrock23(),abstol=1e-6,reltol=1e-6)

@btime solve(prob,Rosenbrock23(),abstol=1e-6,reltol=1e-6)
# Before:
# 1.405 ms (23476 allocations: 1.95 MiB)
# After:
# 1.317 ms (16999 allocations: 1.06 MiB)

@btime solve(prob,Rosenbrock23(linsolve=RFLUFactorization(),chunk_size = Val{3}()),abstol=1e-6,reltol=1e-6)
# Before:
# 1.262 ms (10521 allocations: 952.25 KiB)
# After:
# 1.323 ms (16997 allocations: 1.06 MiB)

But we found an inference bug.

[ChrisRackauckas]

@chriselrod @YingboMa that seems to be something in RecursiveFactorization or JuliaSIMD stuff. Don't know if there's a workaround JuliaLang/julia#35800

[chriselrod]

@chriselrod @YingboMa that seems to be something in RecursiveFactorization or JuliaSIMD stuff. Don't know if there's a workaround JuliaLang/julia#35800

Yeah, that's why I removed the precompilation from TriangularSolve.jl.
With precompilation, performance was bad until you forced it to recompile with Revise.jl.
Without the precompilation, it was fast the first time. =/

I don't know inference or the compiler well enough to know what causes.
Jameson suggested it probably means the code is bad.
I define code quality mostly based on outcomes, so that seems true by definition and not helpful for finding the problem and fixing it.

[YingboMa]

The failure looks real.