Skip to content

Defines derivatives for remaining tgmath math functions. #27559

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
Oct 9, 2019
Merged

Defines derivatives for remaining tgmath math functions. #27559

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
79e0fea
Defines derivatives for remaning tgmath math functions.
vguerra Oct 6, 2019
c92e34b
addressing feedback on formating
vguerra Oct 7, 2019
62c019b
Adding derivatives for generic functions: sqrt and fma.
vguerra Oct 7, 2019
27ca6a2
adding SWIFT_ENABLE_TENSORFLOW comments where needed
vguerra Oct 7, 2019
6b1ebf2
addressing formating issues
vguerra Oct 7, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
120 changes: 119 additions & 1 deletion stdlib/public/Platform/tgmath.swift.gyb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,11 +20,22 @@ public func fabs<T: FloatingPoint>(_ x: T) -> T {
}

@_transparent
// SWIFT_ENABLE_TENSORFLOW
@differentiable(
Copy link
Contributor

@rxwei rxwei Oct 7, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
@differentiable(
// SWIFT_ENABLE_TENSORFLOW
@differentiable(

Copy link
Contributor Author

@vguerra vguerra Oct 7, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done in 27ca6a2

vjp: _vjpSqrt
where T : Differentiable & FloatingPoint, T == T.TangentVector
)
public func sqrt<T: FloatingPoint>(_ x: T) -> T {
return x.squareRoot()
}

@_transparent
// SWIFT_ENABLE_TENSORFLOW
@differentiable(
Copy link
Contributor

@rxwei rxwei Oct 7, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
@differentiable(
// SWIFT_ENABLE_TENSORFLOW
@differentiable(

Copy link
Contributor Author

@vguerra vguerra Oct 7, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done in 27ca6a2

wrt: (x, y, z),
vjp: _vjpFma
where T : Differentiable & FloatingPoint, T == T.TangentVector
)
public func fma<T: FloatingPoint>(_ x: T, _ y: T, _ z: T) -> T {
return z.addingProduct(x, y)
}
Expand Down Expand Up @@ -82,6 +93,24 @@ public func frexp<T: BinaryFloatingPoint>(_ x: T) -> (T, Int) {
return (x.significand / 2, Int(x.exponent + 1))
}

// SWIFT_ENABLE_TENSORFLOW
@usableFromInline
func _vjpSqrt<T: FloatingPoint & Differentiable> (
_ x: T
) -> (T, (T) -> T) where T == T.TangentVector {
let value = x.squareRoot()
return (value, { v in v / (2 * value) })
}

@usableFromInline
func _vjpFma<T: FloatingPoint & Differentiable> (
_ x: T,
_ y: T,
_ z: T
) -> (T, (T) -> (T, T, T)) where T == T.TangentVector {
return (fma(x, y, z), { v in (v * y, v * x, v) })
}

%for T in ['Float','Double']:
@available(swift, deprecated: 4.2, renamed: "scalbn")
@_transparent
Expand All @@ -102,11 +131,27 @@ func _vjpExp(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (value, { v in value * v })
}

@usableFromInline
func _vjpExp2(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
let value = exp2(x)
return (value, { v in v * ${T}(M_LN2) * value })
}

@usableFromInline
func _vjpLog(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (log(x), { v in v / x })
}

@usableFromInline
func _vjpLog10(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (log10(x), { v in v * ${T}(M_LOG10E) / x })
}

@usableFromInline
func _vjpLog2(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (log2(x), { v in v / (${T}(M_LN2) * x) })
}

@usableFromInline
func _vjpSin(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (sin(x), { v in v * cos(x) })
Expand All @@ -122,6 +167,72 @@ func _vjpTan(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
let value = tan(x)
return (value, { v in v * (1 + value * value) })
}

@usableFromInline
func _vjpAsin(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (asin(x), { v in v / sqrt(1 - x * x) })
}

@usableFromInline
func _vjpAcos(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (acos(x), { v in -v / sqrt(1 - x * x) })
}

@usableFromInline
func _vjpAtan(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (atan(x), { v in v / (1 + x * x) })
}

@usableFromInline
func _vjpSinh(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (sinh(x), { v in v * cosh(x) })
}

@usableFromInline
func _vjpCosh(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (cosh(x), { v in v * sinh(x) })
}

@usableFromInline
func _vjpTanh(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
let value = tanh(x)
return (value, { v in v * (1 - value * value) })
}

@usableFromInline
func _vjpAsinh(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (asinh(x), { v in v / sqrt(1 + x * x) })
}

@usableFromInline
func _vjpAcosh(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (acosh(x), { v in v / sqrt(x * x - 1) })
}

@usableFromInline
func _vjpAtanh(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (atanh(x), { v in v / (1 - x * x) })
}

@usableFromInline
func _vjpExpm1(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (expm1(x), { v in exp(x) * v })
}

@usableFromInline
func _vjpLog1p(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (log1p(x), { v in v / (x + 1) })
}

@usableFromInline
func _vjpErf(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (erf(x), { v in v * ${T}(M_2_SQRTPI) * exp(-x * x) })
}

@usableFromInline
func _vjpErfc(_ x: ${T}) -> (${T}, (${T}) -> ${T}) {
return (erfc(x), { v in v * -${T}(M_2_SQRTPI) * exp(-x * x) })
}
% if T == 'Float80':
#endif
% end
Expand Down Expand Up @@ -201,7 +312,14 @@ UnaryIntrinsicFunctions = [
]

# SWIFT_ENABLE_TENSORFLOW
HasVJP = ["exp", "log", "tan", "cos", "sin"]
HasVJP = [
'acos', 'asin', 'atan', 'tan',
'acosh', 'asinh', 'atanh', 'cosh', 'sinh', 'tanh',
'expm1',
'log1p',
'erf', 'erfc',
'cos', 'sin', 'exp', 'exp2', 'log', 'log10', 'log2'
]

def AllFloatTypes():
for bits in allFloatBits:
Expand Down
31 changes: 26 additions & 5 deletions test/stdlib/tgmath.swift.gyb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -247,11 +247,32 @@ MathTests.test("${T}") {
// SWIFT_ENABLE_TENSORFLOW
% for T in ['Float', 'Float80']:
MathTests.test("gradient_${T}") {
expectEqualWithTolerance(7.3890560989306502274, gradient(at: 2.0 as ${T}, in: exp), ulps:16)
expectEqualWithTolerance(0.5, gradient(at: 2.0 as ${T}, in: log), ulps:16)
expectEqualWithTolerance(5.774399204041917612, gradient(at: 2.0 as ${T}, in: tan), ulps:16)
expectEqualWithTolerance(-0.416146836547142387, gradient(at: 2.0 as ${T}, in: sin), ulps:16)
expectEqualWithTolerance(-0.9092974268256816954, gradient(at: 2.0 as ${T}, in: cos), ulps:16)
expectEqualWithTolerance(7.3890560989306502274, gradient(at: 2.0 as ${T}, in: exp), ulps: 16)
expectEqualWithTolerance(2.772588722239781145, gradient(at: 2.0 as ${T}, in: exp2), ulps: 16)
expectEqualWithTolerance(7.3890560989306502274, gradient(at: 2.0 as ${T}, in: expm1), ulps: 16)
expectEqualWithTolerance(0.5, gradient(at: 2.0 as ${T}, in: log), ulps: 16)
expectEqualWithTolerance(0.21714724095162590833, gradient(at: 2.0 as ${T}, in: log10), ulps: 16)
expectEqualWithTolerance(0.7213475204444817278, gradient(at: 2.0 as ${T}, in: log2), ulps: 16)
expectEqualWithTolerance(0.33333333333333333334, gradient(at: 2.0 as ${T}, in: log1p), ulps: 16)
expectEqualWithTolerance(5.774399204041917612, gradient(at: 2.0 as ${T}, in: tan), ulps: 16)
expectEqualWithTolerance(-0.9092974268256816954, gradient(at: 2.0 as ${T}, in: cos), ulps: 16)
expectEqualWithTolerance(-0.416146836547142387, gradient(at: 2.0 as ${T}, in: sin), ulps: 16)
expectEqualWithTolerance(1.154700538379251529, gradient(at: 0.5 as ${T}, in: asin), ulps: 16)
expectEqualWithTolerance(-1.154700538379251529, gradient(at: 0.5 as ${T}, in: acos), ulps: 16)
expectEqualWithTolerance(0.8, gradient(at: 0.5 as ${T}, in: atan), ulps: 16)
expectEqualWithTolerance(3.7621956910836314597, gradient(at: 2.0 as ${T}, in: sinh), ulps: 16)
expectEqualWithTolerance(3.6268604078470187677, gradient(at: 2.0 as ${T}, in: cosh), ulps: 16)
expectEqualWithTolerance(0.07065082485316446565, gradient(at: 2.0 as ${T}, in: tanh), ulps: 16)
expectEqualWithTolerance(0.44721359549995793928, gradient(at: 2.0 as ${T}, in: asinh), ulps: 16)
expectEqualWithTolerance(0.5773502691896257645, gradient(at: 2.0 as ${T}, in: acosh), ulps: 16)
expectEqualWithTolerance(1.3333333333333333334, gradient(at: 0.5 as ${T}, in: atanh), ulps: 16)
expectEqualWithTolerance(0.020666985354092053575, gradient(at: 2.0 as ${T}, in: erf), ulps: 16)
expectEqualWithTolerance(-0.020666985354092053575, gradient(at: 2.0 as ${T}, in: erfc), ulps: 16)
expectEqualWithTolerance(0.35355339059327376222, gradient(at: 2.0 as ${T}, in: { sqrt($0) }), ulps: 16)
let fmaGrad = gradient(at: 4.0 as ${T}, 5.0 as ${T}, 6.0 as ${T}, in: { x, y, z in fma(x, y, z) })
expectEqualWithTolerance(5.0, fmaGrad.0, ulps: 16)
expectEqualWithTolerance(4.0, fmaGrad.1, ulps: 16)
expectEqualWithTolerance(1.0, fmaGrad.2, ulps: 16)
}
%end

Expand Down