mirror of
https://github.com/apple/swift.git
synced 2025-12-14 20:36:38 +01:00
Inside fragile functions, we expect function derivatives to be public, which could be achieved by either explicitly marking the functions as differentiable or having a public explicit derivative defined for them. This is obviously not possible for single and double curry thunks which are a special case of `AutoClosureExpr`. Instead of looking at the thunk itself, we unwrap it and look at the function being wrapped. While the thunk itself and its differentiability witness will not have public visibility, it's not an issue for the case where the function being wrapped (and its witness) have public visibility. Fixes #54819 Fixes #75776
370 lines
11 KiB
Swift
370 lines
11 KiB
Swift
// RUN: %target-swift-emit-sil -verify %s | %FileCheck %s -check-prefix=CHECK-SIL
|
|
|
|
import _Differentiation
|
|
|
|
@_silgen_name("identity")
|
|
func identity<T : Differentiable>(_ x: T) -> T {
|
|
return x
|
|
}
|
|
_ = gradient(at: Float(1), of: { x in identity(x) })
|
|
|
|
// Test PullbackCloner local buffer allocation.
|
|
// Verify that local buffers are immediately set to zero.
|
|
|
|
// CHECK-SIL-LABEL: sil private @identity16_Differentiation14DifferentiableRzlTJpSpSr
|
|
// CHECK-SIL: [[ORIG_COTAN:%.*]] = alloc_stack $τ_0_0.TangentVector
|
|
// CHECK-SIL-NEXT: [[ZERO_WITNESS:%.*]] = witness_method $τ_0_0.TangentVector, #AdditiveArithmetic.zero!getter
|
|
// CHECK-SIL-NEXT: [[ORIG_COTAN_METATYPE:%.*]] = metatype $@thick τ_0_0.TangentVector.Type
|
|
// CHECK-SIL-NEXT: [[EMIT_ZERO_INDIRECT:%.*]] = apply [[ZERO_WITNESS]]<τ_0_0.TangentVector>([[ORIG_COTAN]], [[ORIG_COTAN_METATYPE]])
|
|
// CHECK-SIL: }
|
|
|
|
// Test TF-201: differentiate direct references to generic function.
|
|
// This involves reabstraction thunk differentiation.
|
|
|
|
_ = gradient(at: Float(1), of: identity)
|
|
|
|
protocol DifferentiableAdditiveArithmetic: Differentiable & AdditiveArithmetic {
|
|
@differentiable(reverse)
|
|
static func + (lhs: Self, rhs: Self) -> Self
|
|
}
|
|
extension Float: DifferentiableAdditiveArithmetic {}
|
|
func generic<T: DifferentiableAdditiveArithmetic>(_ x: T) -> T {
|
|
x + x + x
|
|
}
|
|
_ = gradient(at: Float(10), of: generic)
|
|
|
|
struct Wrapper<Scalar : Differentiable> : Differentiable {
|
|
var value: Scalar
|
|
init(_ value: Scalar) { self.value = value }
|
|
}
|
|
func generic<T>(_ x: Wrapper<T>) -> T {
|
|
return x.value
|
|
}
|
|
_ = gradient(at: Wrapper<Float>(1), of: generic)
|
|
|
|
func generic2<T: Differentiable, U: Differentiable>(_ x: T, _ y: Float, _ z: U) -> T {
|
|
return x
|
|
}
|
|
func foo<T>(_ x: Wrapper<T>) {
|
|
_ = gradient(at: Float(1), 2, x, of: generic2)
|
|
}
|
|
|
|
// Test case where associated derivative function's requirements are met.
|
|
extension Wrapper where Scalar : Numeric {
|
|
@differentiable(reverse, wrt: self where Scalar : Differentiable & FloatingPoint)
|
|
func mean() -> Wrapper {
|
|
return self
|
|
}
|
|
|
|
@differentiable(reverse, wrt: self where Scalar : Differentiable & FloatingPoint)
|
|
func variance() -> Wrapper {
|
|
return mean() // ok
|
|
}
|
|
}
|
|
_ = pullback(at: Wrapper<Float>(1), of: { $0.variance() })
|
|
|
|
// Tests TF-277.
|
|
protocol Layer : Differentiable {
|
|
associatedtype Output : Differentiable
|
|
}
|
|
struct SupervisedTrainer<Model : Layer> {
|
|
var model: Model
|
|
var lossFunction: @differentiable(reverse) (Model.Output, Model.Output) -> Float
|
|
func fit(y: Model.Output) {
|
|
_ = gradient(at: y) { y in return self.lossFunction(y, y) }
|
|
}
|
|
}
|
|
|
|
// Tests TF-440.
|
|
struct TF_440_Input<Input: Differentiable, State: Differentiable>
|
|
: Differentiable {
|
|
var input: Input
|
|
var state: State
|
|
}
|
|
struct TF_440<T : Differentiable> {
|
|
@differentiable(reverse)
|
|
func applied(to input: TF_440_Input<Float, Float>) -> Float {
|
|
return input.state
|
|
}
|
|
|
|
@differentiable(reverse)
|
|
func applied(to input: TF_440_Input<T, Float>) -> Float {
|
|
return input.state
|
|
}
|
|
|
|
@differentiable(reverse)
|
|
func applied(to input: TF_440_Input<T, Float>) -> T {
|
|
return input.input
|
|
}
|
|
}
|
|
|
|
// Tests TF-508: differentiation requirements with dependent member types.
|
|
protocol TF_508_Proto {
|
|
associatedtype Scalar
|
|
}
|
|
extension TF_508_Proto where Scalar : FloatingPoint {
|
|
@differentiable(reverse
|
|
where Self : Differentiable, Scalar : Differentiable,
|
|
// Conformance requirement with dependent member type.
|
|
Self.TangentVector : TF_508_Proto
|
|
)
|
|
static func +(lhs: Self, rhs: Self) -> Self {
|
|
return lhs
|
|
}
|
|
|
|
@differentiable(reverse
|
|
where Self : Differentiable, Scalar : Differentiable,
|
|
// Same-type requirement with dependent member type.
|
|
Self.TangentVector == Float
|
|
)
|
|
static func -(lhs: Self, rhs: Self) -> Self {
|
|
return lhs
|
|
}
|
|
}
|
|
extension TF_508_Proto where Self : Differentiable,
|
|
Scalar : FloatingPoint & Differentiable,
|
|
Self.TangentVector : TF_508_Proto {
|
|
@derivative(of: +)
|
|
static func vjpAdd(lhs: Self, rhs: Self)
|
|
-> (value: Self, pullback: (TangentVector) -> (TangentVector, TangentVector)) {
|
|
return (lhs, { v in (v, v) })
|
|
}
|
|
}
|
|
extension TF_508_Proto where Self : Differentiable,
|
|
Scalar : FloatingPoint & Differentiable,
|
|
Self.TangentVector == Float {
|
|
@derivative(of: -)
|
|
static func vjpSubtract(lhs: Self, rhs: Self)
|
|
-> (value: Self, pullback: (TangentVector) -> (TangentVector, TangentVector)) {
|
|
return (lhs, { v in (v, v) })
|
|
}
|
|
}
|
|
|
|
struct TF_508_Struct<Scalar : AdditiveArithmetic>
|
|
: TF_508_Proto, AdditiveArithmetic {}
|
|
extension TF_508_Struct : Differentiable where Scalar : Differentiable {
|
|
typealias TangentVector = TF_508_Struct
|
|
}
|
|
|
|
func TF_508() {
|
|
let x = TF_508_Struct<Float>()
|
|
// Test conformance requirement with dependent member type.
|
|
_ = pullback(at: x, of: { (x: TF_508_Struct<Float>) -> TF_508_Struct<Float> in
|
|
return x + x
|
|
})
|
|
// Test same-type requirement with dependent member type.
|
|
_ = pullback(at: x, of: { (x: TF_508_Struct<Float>) -> TF_508_Struct<Float> in
|
|
return x - x
|
|
})
|
|
}
|
|
|
|
// TF-523
|
|
struct TF_523_Struct : Differentiable & AdditiveArithmetic {
|
|
var a: Float = 1
|
|
typealias TangentVector = TF_523_Struct
|
|
}
|
|
|
|
@differentiable(reverse)
|
|
func TF_523_f(_ x: TF_523_Struct) -> Float {
|
|
return x.a * 2
|
|
}
|
|
|
|
// TF-534: Thunk substitution map remapping.
|
|
protocol TF_534_Layer : Differentiable {
|
|
associatedtype Input : Differentiable
|
|
associatedtype Output : Differentiable
|
|
|
|
@differentiable(reverse)
|
|
func callAsFunction(_ input: Input) -> Output
|
|
}
|
|
struct TF_534_Tensor<Scalar> : Differentiable {}
|
|
|
|
func TF_534<Model: TF_534_Layer>(
|
|
_ model: inout Model, inputs: Model.Input
|
|
) -> TF_534_Tensor<Float> where Model.Output == TF_534_Tensor<Float> {
|
|
return valueWithPullback(at: model) { model -> Model.Output in
|
|
return model(inputs)
|
|
}.0
|
|
}
|
|
|
|
// TF-546: Test that SILGen linear map thunk performs correct reabstraction.
|
|
struct TF_546<T: FloatingPoint>: AdditiveArithmetic {
|
|
var real: T
|
|
var imaginary: T
|
|
|
|
@differentiable(reverse where T: Differentiable, T == T.TangentVector)
|
|
init(real: T = 0, imaginary: T = 0) {
|
|
self.real = real
|
|
self.imaginary = imaginary
|
|
}
|
|
}
|
|
extension TF_546: Differentiable where T: Differentiable {
|
|
typealias TangentVector = TF_546
|
|
}
|
|
extension TF_546 where T: Differentiable, T == T.TangentVector {
|
|
@derivative(of: init)
|
|
static func _vjpInit(real: T, imaginary: T) -> (value: TF_546, pullback: (TF_546) -> (T, T)) {
|
|
return (TF_546(real: real, imaginary: imaginary), { ($0.real, $0.imaginary) })
|
|
}
|
|
}
|
|
let _: @differentiable(reverse) (Float, Float) -> TF_546<Float> = { r, i in
|
|
TF_546(real: r, imaginary: i)
|
|
}
|
|
|
|
// TF-652: Test VJPCloner substitution map generic signature.
|
|
// The substitution map should have the VJP's generic signature, not the
|
|
// original function's.
|
|
struct TF_652<Scalar> {}
|
|
extension TF_652 : Differentiable where Scalar : FloatingPoint {}
|
|
|
|
@differentiable(reverse, wrt: x where Scalar: FloatingPoint)
|
|
func test<Scalar: Numeric>(x: TF_652<Scalar>) -> TF_652<Scalar> {
|
|
for _ in 0..<10 {
|
|
let _ = x
|
|
}
|
|
return x
|
|
}
|
|
|
|
// TF-682: Test that SILGen linear map thunk performs correct reabstraction.
|
|
protocol TF_682_Proto {
|
|
associatedtype Scalar
|
|
}
|
|
extension TF_682_Proto where Scalar : FloatingPoint {
|
|
@differentiable(reverse
|
|
where Self : Differentiable, Scalar : Differentiable,
|
|
// Same-type requirement with dependent member type.
|
|
Self.TangentVector == Float
|
|
)
|
|
func foo(lhs: Self) -> Self {
|
|
return lhs
|
|
}
|
|
}
|
|
extension TF_682_Proto where Self : Differentiable,
|
|
Scalar : FloatingPoint & Differentiable,
|
|
Self.TangentVector == Float {
|
|
@derivative(of: foo)
|
|
func vjpFoo(lhs: Self) -> (
|
|
value: Self, pullback: (TangentVector) -> (TangentVector, TangentVector)
|
|
) {
|
|
return (lhs, { v in (v, v) })
|
|
}
|
|
}
|
|
|
|
// TF-697: Test generic requirements of generated derivative function.
|
|
protocol TF_697_Module: Differentiable {
|
|
associatedtype Input
|
|
associatedtype Output: Differentiable
|
|
|
|
@differentiable(reverse, wrt: self)
|
|
func callModule(_ input: Input) -> Output
|
|
}
|
|
protocol TF_697_Layer: TF_697_Module where Input: Differentiable {
|
|
@differentiable(reverse)
|
|
func callLayer(_ input: Input) -> Output
|
|
}
|
|
struct TF_697_Sequential<Layer1: TF_697_Module, Layer2: TF_697_Layer>: TF_697_Module
|
|
where Layer1.Output == Layer2.Input {
|
|
var layer1: Layer1
|
|
var layer2: Layer2
|
|
|
|
@differentiable(reverse, wrt: self)
|
|
func callModule(_ input: Layer1.Input) -> Layer2.Output {
|
|
layer2.callLayer(layer1.callModule(input))
|
|
}
|
|
}
|
|
extension TF_697_Sequential: TF_697_Layer where Layer1: TF_697_Layer {
|
|
@differentiable(reverse)
|
|
func callLayer(_ input: Layer1.Input) -> Layer2.Output {
|
|
layer2.callLayer(layer1.callLayer(input))
|
|
}
|
|
}
|
|
|
|
// TF-817: Test remapping `apply` callee types in derivative function context.
|
|
struct TF_817<T> {
|
|
func foo(_ index: Int) -> T {
|
|
fatalError()
|
|
}
|
|
}
|
|
extension TF_817: Differentiable where T: Differentiable {
|
|
@derivative(of: foo)
|
|
func vjpFoo(index: Int) -> (value: T, pullback: (T.TangentVector) -> (TangentVector)) {
|
|
fatalError()
|
|
}
|
|
}
|
|
extension TF_817 {
|
|
@differentiable(reverse, wrt: self where T: Differentiable)
|
|
public func test(index: Int) -> T {
|
|
return self.foo(0) // crash happened here
|
|
}
|
|
}
|
|
|
|
// TF-886: Test `partial_apply` of linear map subset parameters thunk.
|
|
@differentiable(reverse)
|
|
func TF_886_foo<T, U: Differentiable>(_: Float, _: T, _: U) -> Float {
|
|
return 0
|
|
}
|
|
@differentiable(reverse)
|
|
func TF_886_bar<T>(x: Float, y: T) -> Float {
|
|
return TF_886_foo(x, y, 0)
|
|
}
|
|
|
|
// Test layout requirements.
|
|
|
|
// The layout requirement is "contextual": the requirement is not on `T`, the
|
|
// differentiable function parameter/result type.
|
|
struct ContextualLayoutRequirement<T: Differentiable, U: AnyObject> {
|
|
var stored: T
|
|
}
|
|
extension ContextualLayoutRequirement {
|
|
func test(_ x: T) {
|
|
let _: @differentiable(reverse) (T) -> T = { _ in self.stored }
|
|
let _: @differentiable(reverse) (T) -> T = { $0 }
|
|
}
|
|
}
|
|
// The layout requirement directly involves `T`, the differentiable function
|
|
// parameter/result type.
|
|
// TODO(TF-851): Uncomment the tests below after `@differentiable` function
|
|
// SILGen thunking is fixed.
|
|
/*
|
|
struct LayoutRequirement<T: AnyObject & Differentiable> {
|
|
var stored: T
|
|
}
|
|
extension LayoutRequirement {
|
|
func test(_ x: T) {
|
|
let _: @differentiable(reverse) (T) -> T = { _ in self.stored }
|
|
let _: @differentiable(reverse) (T) -> T = { $0 }
|
|
}
|
|
}
|
|
*/
|
|
|
|
// Test superclass requirements.
|
|
|
|
class Super: Differentiable {}
|
|
|
|
// The superclass requirement is "contextual": the requirement is not on `T`,
|
|
// the differentiable function parameter/result type.
|
|
struct ContextualSuperclassRequirement<T: Differentiable, U: Super> {
|
|
var stored: T
|
|
}
|
|
extension ContextualSuperclassRequirement {
|
|
func test(_ x: T) {
|
|
let _: @differentiable(reverse) (T) -> T = { _ in self.stored }
|
|
let _: @differentiable(reverse) (T) -> T = { $0 }
|
|
}
|
|
}
|
|
// The superclass requirement directly involves `T`, the differentiable
|
|
// function parameter/result type.
|
|
// TODO(TF-851): Uncomment the tests below after `@differentiable` function
|
|
// SILGen thunking is fixed.
|
|
/*
|
|
struct SuperclassRequirement<T: Super & Differentiable> {
|
|
var stored: T
|
|
}
|
|
extension SuperclassRequirement {
|
|
func test(_ x: T) {
|
|
let _: @differentiable(reverse) (T) -> T = { _ in self.stored }
|
|
let _: @differentiable(reverse) (T) -> T = { $0 }
|
|
}
|
|
}
|
|
*/
|