[JuliaRegistrator]

• Registering package: Piste
• Repository: https://github.com/EvoArt/Piste.jl
• Created by: @EvoArt
• Version: v0.4.0
• Commit: 4655c6b96f16f26eb9206cf0bdd90920b56ad210
• Reviewed by: @EvoArt
• Reference: EvoArt/Piste.jl@4655c6b#commitcomment-208974227
• Description: Forward-mode automatic differentiation that survives juliac --trim. A piste is a groomed ski run: a slope, trimmed.
• Release notes:

This is the first registered release since v0.1.0. The v0.2.0 and v0.3.0 registrations lapsed, so these notes cover everything since v0.1.0.

## Breaking changes

- `eltype(Dual{N,T})` is now `Dual{N,T}` and `eltype(TapedReal)` is `TapedReal`, as for any `Number` type. They used to return the float type, which broke broadcasting with a scalar dual (e.g. `[1, 2] .* d`). Code that relied on `eltype` to get the underlying float type should use `typeof(value(d))` instead.
- `TapedReal` is now an isbits value that refers to its tape through a slot index, instead of holding a reference to the `Tape`. Only code that constructed `TapedReal`s or read their fields directly is affected. The public functions (`track!`, `backward!`, `rev_gradient!`, ...) are unchanged.

## New

- Reverse mode: `TapedReal`, `Tape`, `ReverseWorkspace`, `track!`, `backward!`, `reset!`, `rev_gradient!` and `rev_value_and_gradient!`.
- `rev_value_and_pullback!` and `rev_value_and_pullback` compute vector-Jacobian products for array- or scalar-valued functions. They are the basis of the upcoming DifferentiationInterface backend.
- `GradientWorkspace` and `pickchunk`, so that `gradient!` and `value_and_gradient!` can reuse their buffers across calls.
- `Float32` support, and the float-introspection methods on `Dual` that Distributions.jl's `logpdf` paths need.
- `PisteLogExpFunctionsExt`: one-node `log1pexp`, `logistic`, `logit`, `log1mexp`, `logexpm1` and `log1psq`.

## Fixes

- Reverse mode: two cases that silently returned wrong gradients, both now fixed. One was a peephole fusion that reused tape slots the caller still held. The other was a matrix-vector product whose operand was itself computed, which gave a zero gradient.
- A `Float32` data matrix is no longer copied on every reverse-mode gradient call.

## Performance

Reverse mode is 1.2-2.2x faster on normal regression models, and `sum(x)` and `sum(abs2, x)` over tracked arrays now record a single node. See the README for benchmarks.

[github-actions]

Hello, I am an automated registration bot. I help manage the registration process by checking your registration against a set of AutoMerge guidelines. If all these guidelines are met, this pull request will be merged automatically, completing your registration. It is strongly recommended to follow the guidelines, since otherwise the pull request needs to be manually reviewed and merged by a human.

1. AutoMerge Guidelines are all met! ✅

Your new version registration met all of the guidelines for auto-merging and is scheduled to be merged in the next round (~20 minutes).

2. Code changes since last version

Code changes from v0.1.0:

❯ git diff-tree --shortstat 12a280e314be2968f90847f62120f145797e8763 382867c0c7f694ae4d7f395375508d034c08f165
 19 files changed, 7966 insertions(+), 57 deletions(-)

Full diff has 295918 characters (over limit of 50k), so is omitted here.
View full patch diff on GitHub

3. To pause or stop registration

If you want to prevent this pull request from being auto-merged, simply leave a comment. If you want to post a comment without blocking auto-merging, you must include the text [noblock] in your comment.

Tip: You can edit blocking comments to add [noblock] in order to unblock auto-merging. Auto-merge runs periodically. After unblocking a comment you may have to wait up to 15 minutes.

[EvoArt]

[merge approved]

The skip is intentional. v0.2.0 (#168302) and v0.3.0 (#168430) were tagged on GitHub but their registrations lapsed for missing release notes, so v0.4.0 is the first registration since v0.1.0. The release notes cover everything since v0.1.0. A DifferentiationInterface backend (JuliaDiff/DifferentiationInterface.jl#1089) depends on this version.