Unconditional max_methods=1

[chriselrod]

@ChrisRackauckas no fear
max_methods=1 can't hurt you if you're pure of heart type stable.

[chriselrod]

Well, it doesn't seem to help either.
PR:

julia> @time @eval function AmulB!(C,A,B)
           @turbo for n ∈ indices((C,B),2), m ∈ indices((C,A),1)
               Cmn = zero(eltype(C))
               for k ∈ indices((A,B),(2,1))
                   Cmn += A[m,k] * B[k,n]
               end
               C[m,n] = Cmn
           end
       end
  0.023787 seconds (4.70 k allocations: 271.455 KiB, 76.64% compilation time)
AmulB! (generic function with 1 method)

julia> M = K = N = 71; A = rand(M,K); B = rand(K,N); C1 = A*B; C0 = zero(C1);

julia> @time @eval AmulB!(C0, A, B);
  2.113024 seconds (4.32 M allocations: 261.378 MiB, 3.93% gc time, 99.99% compilation time)

Latest release:

julia> @time @eval function AmulB!(C,A,B)
           @turbo for n ∈ indices((C,B),2), m ∈ indices((C,A),1)
               Cmn = zero(eltype(C))
               for k ∈ indices((A,B),(2,1))
                   Cmn += A[m,k] * B[k,n]
               end
               C[m,n] = Cmn
           end
       end
  0.024604 seconds (4.70 k allocations: 271.307 KiB, 76.93% compilation time)
AmulB! (generic function with 1 method)

julia> M = K = N = 71; A = rand(M,K); B = rand(K,N); C1 = A*B; C0 = zero(C1);

julia> @time @eval AmulB!(C0, A, B);
  2.093600 seconds (4.30 M allocations: 260.192 MiB, 4.07% gc time, 99.99% compilation time)

Of course, there is a ton of type unstable code in LV to process expressions.

Maybe I should merge this anyway as a symbolic gesture.

[ChrisRackauckas]

I was wondering if it could effect compilation downstream though when it's being hit behind a function barrier (due to the autoswitch solvers). That's really tedious to test though.

[chriselrod]

Of course, there is a ton of type unstable code in LV to process expressions.

Which is why I was thinking this may help.

I was wondering if it could effect compilation downstream though when it's being hit behind a function barrier (due to the autoswitch solvers). That's really tedious to test though.

I'm going to @descend through the calls in OrdinaryDiffEq's precompile statements and see if I run into instabilities.

But I think
@max_methods 1 as a default, and then special casing specific methods that need more (if we have no other way of fixing it) is a reasonable approach.

For reference: JuliaLang/julia#44426

But I think we should also see more examples where it helps before going all out.
One of us should build Julia that defaults to max_methods=1 and see if OrdinaryDiffEq compiles any faster vs default Julia master before committing to anything like PRing max_methods=1 into the entire JuliaSIMD and SciML stacks.

[chriselrod]

I'm going to @descend through the calls in OrdinaryDiffEq's precompile statements and see if I run into instabilities.

I run into type instabilities immediately.

[ChrisRackauckas]

There should be some due to the auto-switch, which does manual branching in order to get around the way it would dispatch, but I'm not sure it gives good precompilation.

[chriselrod]

Even without autoswitch

EDIT: false alarm

[codecov]

Codecov Report

Base: 87.00% // Head: 86.84% // Decreases project coverage by -0.17% ⚠️

Coverage data is based on head (99afd57) compared to base (46d3333).
Patch has no changes to coverable lines.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #462      +/-   ##
==========================================
- Coverage   87.00%   86.84%   -0.17%     
==========================================
  Files          38       38              
  Lines        9372     9372              
==========================================
- Hits         8154     8139      -15     
- Misses       1218     1233      +15     

Impacted Files	Coverage Δ
src/LoopVectorization.jl	75.00% <ø> (ø)
src/codegen/lower_constant.jl	72.76% <0.00%> (-2.82%)	⬇️
src/reconstruct_loopset.jl	91.79% <0.00%> (-1.18%)	⬇️
src/modeling/determinestrategy.jl	94.92% <0.00%> (-0.32%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.