Stop precomputing chunk sizes

[KristofferC]

The non-concretely typed vector with chunks prevents constant propagation of the length of a static array. Alternative to #707

This code was used to workaround JuliaLang/julia#29887 but with modern Julia it just inhibits optimizations.

[ChrisRackauckas]

To get this actually released, we should do this to the release-0.10 branch?

[KristofferC]

Or at least backport it

[devmotion]

I ran the benchmark in #373, on the master branch and with this PR:

master

julia>@btime ForwardDiff.gradient(rosenbrock, x);
328.805 ns (3 allocations:1.92 KiB)
julia>@btime ForwardDiff.gradient(rosenbrock, x);
331.407 ns (3 allocations:1.92 KiB)
julia>@btime ForwardDiff.gradient(rosenbrock, x);
328.077 ns (3 allocations:1.92 KiB)

PR #708

julia>@btime ForwardDiff.gradient(rosenbrock, x);
417.503 ns (3 allocations:1.92 KiB)
julia>@btime ForwardDiff.gradient(rosenbrock, x);
406.040 ns (3 allocations:1.92 KiB)
julia>@btime ForwardDiff.gradient(rosenbrock, x);
407.500 ns (3 allocations:1.92 KiB)

Setup

julia>versioninfo()
Julia Version 1.10.4
Commit 48d4fd48430 (2024-06-0410:41 UTC)
Build Info:
Official https://julialang.org/ release
Platform Info:
OS: macOS (arm64-apple-darwin22.4.0)
CPU:10× Apple M2 Pro
WORD_SIZE:64
LIBM: libopenlibm
LLVM: libLLVM-15.0.7 (ORCJIT, apple-m1)
Threads:1 default, 0 interactive, 1 GC (on 6 virtual cores)
Environment:
JULIA_PKG_USE_CLI_GIT =true

[KristofferC]

This was an easier way too see the regression:

julia>const x =rand(10)
# master
julia>@btime ForwardDiff.Chunk($x)
3.792 ns (0 allocations:0 bytes)
ForwardDiff.Chunk{10}()
# PR
julia>@btime ForwardDiff.Chunk($x)
55.894 ns (0 allocations:0 bytes)
ForwardDiff.Chunk{10}()

I made a new commit to try work around that. Could you try again @devmotion?

[devmotion]

I reran the benchmark in #373 with the latest commit:

julia>@btime ForwardDiff.gradient(rosenbrock, x);
358.333 ns (3 allocations:1.92 KiB)
julia>@btime ForwardDiff.gradient(rosenbrock, x);
346.847 ns (3 allocations:1.92 KiB)
julia>@btime ForwardDiff.gradient(rosenbrock, x);
347.417 ns (3 allocations:1.92 KiB)

And that's what I get when I benchmark Chunk:

# master
julia>@btime ForwardDiff.Chunk($x)
3.458 ns (0 allocations:0 bytes)
# PR
julia>@btime ForwardDiff.Chunk($x)
3.625 ns (0 allocations:0 bytes)

[KristofferC]

So would you say this is ok as it is now?

[devmotion]

Personally, I think the benefits (for type inference and static arrays) are worth the ~5% regression in these benchmarks, even more so since these benchmarks only focus on the construction of Chunk and GradientConfig.

[KristofferC]

Should backport and release a new version