Compiler:
- Add `Forward` and `Reverse` to `DifferentiabilityKind`.
- Expand `DifferentiabilityMask` in `ExtInfo` to 3 bits so that it now holds all 4 cases of `DifferentiabilityKind`.
- Parse `@differentiable(reverse)` and `@differentiable(_forward)` declaration attributes and type attributes.
- Emit a warning for `@differentiable` without `reverse`.
- Emit an error for `@differentiable(_forward)`.
- Rename `@differentiable(linear)` to `@differentiable(_linear)`.
- Make `@differentiable(reverse)` type lowering go through today's `@differentiable` code path. We will specialize it to reverse-mode in a follow-up patch.
ABI:
- Add `Forward` and `Reverse` to `FunctionMetadataDifferentiabilityKind`.
- Extend `TargetFunctionTypeFlags` by 1 bit to store the highest bit of differentiability kind (linear). Note that there is a 2-bit gap in `DifferentiabilityMask` which is reserved for `AsyncMask` and `ConcurrentMask`; `AsyncMask` is ABI-stable so we cannot change that.
_Differentiation module:
- Replace all occurrences of `@differentiable` with `@differentiable(reverse)`.
- Delete `_transpose(of:)`.
Resolves rdar://69980056.
`valueWithDerivative(of:)` and `valueWithGradient(of:)` are the curried version of `valueWithDerivative(at:in:)` and `valueWithGradient(at:in:)`, respectively. They are not included the proposal and are not being used by any client that I know of. This patch removes them.
* Add differentiation benchmarks.
* Make install name of _Differentiation be @rpath/libswift_Differentiation.dylib.
Co-authored-by: Marc Rasi <marcrasi@google.com>
In derivatives of loops, no longer allocate boxes for indirect case payloads. Instead, use a custom pullback context in the runtime which contains a bump-pointer allocator.
When a function contains a differentiated loop, the closure context is a `Builtin.NativeObject`, which contains a `swift::AutoDiffLinearMapContext` and a tail-allocated top-level linear map struct (which represents the linear map struct that was previously directly partial-applied into the pullback). In branching trace enums, the payloads of previously indirect cases will be allocated by `swift::AutoDiffLinearMapContext::allocate` and stored as a `Builtin.RawPointer`.
This replaces swiftMSVCRT with swiftCRT. The big difference here is
that the `visualc` module is no longer imported nor exported. The
`visualc` module remains in use for a singular test wrt availability,
but this should effectively remove the need for the `visualc` module.
The difference between the MSVCRT and ucrt module was not well
understood by most. MSVCRT provided ucrt AND visualc, combining pieces
of the old MSVCRT and the newer ucrt. The ucrt module is what you
really wanted most of the time, however, would need to use MSVCRT for
the convenience aliases for type-generic math and the deprecated math
constants.
Unfortunately, we cannot shadow the `ucrt` module and create a Swift SDK
overlay for ucrt as that seems to result in circular dependencies when
processing the `_Concurrency` module.
Although this makes using the C library easier for most people, it has a
more important subtle change: it cleaves the dependency on visualc.
This means that this enables use of Swift without Visual Studio for the
singular purpose of providing 3 header files. Additionally, it removes
the need for the installation of 2 of the 4 support files. This greatly
simplifies the deployment process on Windows.
This effectively reverts #31183 -- we need to match the install name convention of the other stdlib libraries.
From the review feedback:
> The right way to load the stdlib & runtime libraries from a custom toolchain is to set `DYLD_LIBRARY_PATH` when executing the generated binary. This is how we run tests against the just-built libraries and this is how downloadable toolchain snapshots are currently configured in Xcode -- see #33178
Adds forward mode support for `apply` instruction with `inout` arguments.
Example of supported code:
```
func add(_ x: inout Float, _ y: inout Float) -> Float {
var result = x
result += y
return result
}
print(differential(at: 1, 1, in: add)(1, 1)) // prints "2"
```
Make `Optional` conditionally conform to `Differentiable` when the `Wrapped` type does.
`Optional.TangentVector` is a wrapper around `Wrapped.TangentVector?`.
Also, fix `Array.TangentVector.zeroTangentVectorInitializer`.
Resolves TF-1301.
LLVM doesn't have a stable ABI for Float16 on x86 yet; we're working with Intel to get that fixed, but we don't want to make the type available on macOS until a stable ABI is actually available, because we'd break binaries compiled before any calling convention changes if we do.
Clean up a few general patterns that are now obviated by canImport
This aligns more generally with the cleanup that the Swift Package
Manager has already done in their automated XCTest-plumbing tool in
apple/swift-package-manager#1826.
`Differentiable` conformance derivation now supports
`Differentiable.zeroTangentVectorInitializer`.
There are two potential cases:
1. Memberwise derivation: done when `TangentVector` can be initialized memberwise.
2. `{ TangentVector.zero }` derivation: done as a fallback.
`zeroTangentVectorInitializer` is a closure that produces a zero tangent vector,
capturing minimal necessary information from `self`.
It is an instance property, unlike the static property `AdditiveArithmetic.zero`,
and should be used by the differentiation transform for correctness.
Remove `Differentiable.zeroTangentVectorInitializer` dummy default implementation.
Update stdlib `Differentiable` conformances and tests.
Clean up DerivedConformanceDifferentiable.cpp cruft.
Resolves TF-1007.
Progress towards TF-1008: differentiation correctness for projection operations.
The derivative wrt `self` should drop the last element from the incoming seed.
Example:
- Incoming seed: [1, 2, 3, 4]
- Derivative wrt `self`: [1, 2, 3]
- Derivative wrt appended element: 4
New files were added in #30875 which did not include os(OpenBSD), so add
this.
add_swift_target_library in AddSwiftStdlib subsequently required
modification. _add_target_variant_link_flags likely needs modification as
well, but this is better suited to a separate PR.
Add `Differentiable.withDerivative(_:)`, a "derivative surgery" API.
`Differentiable.withDerivative(_:)` is an identity function returning `self`.
It takes a closure and applies it to the derivative of the return value, in
contexts where the return value is differentiated with respect to.
JVP functions are forward-mode derivative functions. They take original
arguments and return original results and a differential function. Differential
functions take derivatives wrt arguments and return derivatives wrt results.
`JVPEmitter` is a cloner that emits JVP and differential functions at the same
time. In JVP functions, function applications are replaced with JVP function
applications. In differential functions, function applications are replaced
with differential function applications.
In JVP functions, each basic block takes a differential struct containing callee
differentials. These structs are consumed by differential functions.
The differentiation transform does the following:
- Canonicalizes differentiability witnesses by filling in missing derivative
function entries.
- Canonicalizes `differentiable_function` instructions by filling in missing
derivative function operands.
- If necessary, performs automatic differentiation: generating derivative
functions for original functions.
- When encountering non-differentiability code, produces a diagnostic and
errors out.
Partially resolves TF-1211: add the main canonicalization loop.
To incrementally stage changes, derivative functions are currently created
with empty bodies that fatal error with a nice message.
Derivative emitters will be upstreamed separately.
Previously, two conditions were necessary to enable differentiable programming:
- Using the `-enable-experimental-differentiable-programming` frontend flag.
- Importing the `_Differentiation` module.
Importing the `_Differentiation` module is the true condition because it
contains the required compiler-known `Differentiable` protocol. The frontend
flag is redundant and cumbersome.
Now, the frontend flag is removed.
Importing `_Differentiation` is the only condition.
* Add all [differential operators](https://github.com/apple/swift/blob/master/docs/DifferentiableProgramming.md#list-of-differential-operators).
* Add `withoutDerivative(at:)`, used for efficiently stopping the derivative propagation at a value and causing the derivative at the value to be zero.
* Add utility `differentiableFunction(from:)`, used for creating a `@differentiable` function from an original function and a derivative function.
Mostly work done by @marcrasi and @dan-zheng.
Partially resolves TF-843.
TODO:
* Add `AnyDerivative`.
* Add `Array.differentiableMap(_:)` and `differentiableReduce(_:_:)`.
Add `Differentiable` conformances for floating-point types to the
`_Differentiation` module. The `TangentVector` associated type for
floating-point types is `Self`.
This design adheres to the differentiable programming manifesto:
docs/DifferentiableProgramming.md.
Partially resolves TF-1052.
The `_Differentiation` module is the experimental support library for
differentiable programming. It is built when the build-script flag
`--enable-experimental-differentiable-programming` is enabled.
The `Differentiable` protocol generalizes all types that work with
differentiation. It is a core piece of the differentiable programming
project. Other parts depending on the `Differentiable` protocol will
be upstreamed piece by piece.
The `Differentiable` protocol is compiler-known and will be used during
type-checking, SILGen, and the SIL differentiation transform.
