Merge pull request #305 from stephentyrone/remove-stochastic-rounding

Remove stochastic rounding mode
This commit is contained in:
Stephen Canon authored and GitHub committed 2025-04-10 11:51:13 -04:00
commit 2b458e8aeb
9 files changed
+22 -510

No files matched your search

+2 -15
View File
@@ -1,9 +1,9 @@
// swift-tools-version:5.5
// swift-tools-version:5.9
//===--- Package.swift ----------------------------------------*- swift -*-===//
//
// This source file is part of the Swift Numerics open source project
//
// Copyright (c) 2019-2021 Apple Inc. and the Swift Numerics project authors
// Copyright (c) 2019-2025 Apple Inc. and the Swift Numerics project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
@@ -81,19 +81,6 @@ let package = Package(
name: "RealTests",
dependencies: ["_TestSupport"],
exclude: ["CMakeLists.txt"]
),
// MARK: - Test executables
.executableTarget(
name: "ComplexLog",
dependencies: ["Numerics", "_TestSupport"],
path: "Tests/Executable/ComplexLog"
),
.executableTarget(
name: "ComplexLog1p",
dependencies: ["Numerics", "_TestSupport"],
path: "Tests/Executable/ComplexLog1p"
)
]
)
@@ -116,22 +116,6 @@ extension BinaryInteger {
// If q is already odd, we have the correct result.
if q._lowWord & 1 == 1 { return q }
case .stochastically:
let bmag = other.magnitude
let rmag = r.magnitude
var bhi: UInt64
var rhi: UInt64
if other.magnitude <= UInt64.max {
bhi = UInt64(bmag)
rhi = UInt64(rmag)
} else {
let shift = bmag._msb - 63
bhi = UInt64(truncatingIfNeeded: bmag >> shift)
rhi = UInt64(truncatingIfNeeded: rmag >> shift)
}
let (sum, car) = rhi.addingReportingOverflow(.random(in: 0 ..< bhi))
if sum < bhi && !car { return q }
case .requireExact:
preconditionFailure("Division was not exact.")
}
@@ -299,22 +283,6 @@ extension SignedInteger {
// If q is already odd, we have the correct result.
if q._lowWord & 1 == 1 { return (q, r) }
case .stochastically:
let bmag = other.magnitude
let rmag = r.magnitude
var bhi: UInt64
var rhi: UInt64
if other.magnitude <= UInt64.max {
bhi = UInt64(bmag)
rhi = UInt64(rmag)
} else {
let shift = bmag._msb - 63
bhi = UInt64(truncatingIfNeeded: bmag >> shift)
rhi = UInt64(truncatingIfNeeded: rmag >> shift)
}
let (sum, car) = rhi.addingReportingOverflow(.random(in: 0 ..< bhi))
if sum < bhi && !car { return (q, r) }
case .requireExact:
preconditionFailure("Division was not exact.")
}
@@ -347,7 +315,7 @@ extension SignedInteger {
/// is not representable.
///
/// - Returns: `(quotient, remainder)`, with `0 <= remainder < b.magnitude`.
func euclideanDivision<T>(_ a: T, _ b: T) -> (quotient: T, remainder: T)
public func euclideanDivision<T>(_ a: T, _ b: T) -> (quotient: T, remainder: T)
where T: SignedInteger
{
a.divided(by: b, rounding: a >= 0 ? .towardZero : .awayFromZero)
+16 -72
View File
@@ -2,21 +2,13 @@
//
// This source file is part of the Swift Numerics open source project
//
// Copyright (c) 2021-2024 Apple Inc. and the Swift Numerics project authors
// Copyright (c) 2021-2025 Apple Inc. and the Swift Numerics project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
//
//===----------------------------------------------------------------------===//
// TODO: it's unfortunate that we can't specify a custom random source
// for the stochastic rounding rule, but I don't see a nice way to have
// that share the API with the other rounding rules, because we'd then
// have to take either the rule in-out or have an additional RNG/state
// parameter. The same problem applies to rounding with dithering and
// any other stateful rounding method. We should consider adding a
// stateful rounding API down the road to support those use cases.
/// A rule that defines how to select one of the two representable results
/// closest to a given value.
///
@@ -59,22 +51,22 @@
/// 2.0 | 2 | 2 | 2 | 2 | 2 |
/// -------+----------+----------+----------+----------+----------+
///
/// Specialized rounding rules
/// Specialized rounding rules
///
/// value | toOdd | stochastically | requireExact |
/// =======+==============+=======================+================+
/// -1.5 | -1 | 50% -2, 50% -1 | trap |
/// -------+--------------+-----------------------+----------------+
/// -0.5 | -1 | 50% -1, 50% 0 | trap |
/// -------+--------------+-----------------------+----------------+
/// 0.5 | 1 | 50% 0, 50% 1 | trap |
/// -------+--------------+-----------------------+----------------+
/// 0.7 | 1 | 30% 0, 70% 1 | trap |
/// -------+--------------+-----------------------+----------------+
/// 1.2 | 1 | 80% 1, 20% 2 | trap |
/// -------+--------------+-----------------------+----------------+
/// 2.0 | 2 | 2 | 2 |
/// -------+--------------+-----------------------+----------------+
/// value | toOdd | requireExact |
/// =======+==============+================+
/// -1.5 | -1 | trap |
/// -------+--------------+----------------+
/// -0.5 | -1 | trap |
/// -------+--------------+----------------+
/// 0.5 | 1 | trap |
/// -------+--------------+----------------+
/// 0.7 | 1 | trap |
/// -------+--------------+----------------+
/// 1.2 | 1 | trap |
/// -------+--------------+----------------+
/// 2.0 | 2 | 2 |
/// -------+--------------+----------------+
/// ```
public enum RoundingRule {
/// Produces the closest representable value that is less than or equal
@@ -200,45 +192,6 @@ public enum RoundingRule {
/// because 5/2 = 2.5 is equally close to 2 and 3, and 2 is even.
case toNearestOrEven
/// Adds a uniform random value from [0, d) to the value being rounded,
/// where d is the distance between the two closest representable values,
/// then rounds the sum downwards.
///
/// Unlike all the other rounding modes, this mode is _not deterministic_;
/// repeated calls to rounding operations with this mode will generally
/// produce different results. There is a tradeoff implicit in using this
/// mode: you can sacrifice _reproducible_ results to get _more accurate_
/// results in aggregate. For a contrived but illustrative example, consider
/// the following:
/// ```
/// let data = Array(repeating: 1, count: 100)
/// let result = data.reduce(0) {
/// $0 + $1.divided(by: 3, rounding: rule)
/// }
/// ```
/// because 1/3 is always the same value between 0 and 1, any
/// deterministic rounding rule must produce either 0 or 100 for
/// this computation. But rounding `stochastically` will
/// produce a value close to 33. The _error_ of the computation
/// is smaller, but the result will now change between runs of the
/// program.
///
/// For this simple case a better solution would be to add the
/// values first, and then divide. This gives a result that is both
/// reproducible _and_ accurate:
/// ```
/// let result = data.reduce(0, +)/3
/// ```
/// but this isn't always possible in more sophisticated scenarios,
/// and in those cases this rounding rule may be useful.
///
/// Examples:
/// - `(-4).divided(by: 3, rounding: .stochastically)`
/// will be –1 with probability 2/3 and –2 with probability 1/3.
/// - `5.shifted(rightBy: 1, rounding: .stochastically)`
/// will be 2 with probability 1/2 and 3 with probability 1/2.
case stochastically
/// If the value being rounded is representable, that value is returned.
/// Otherwise, a precondition failure occurs.
///
@@ -304,15 +257,6 @@ extension FloatingPoint {
// which way that rounds, then select the other value.
let even = (trunc + one/2).rounded(.toNearestOrEven)
return trunc == even ? trunc + one : trunc
case .stochastically:
let trunc = rounded(.towardZero)
if trunc == self { return trunc }
// We have eliminated all large values at this point; add dither in
// ±[0,1) and then truncate.
let bits = Swift.min(-Self.ulpOfOne.exponent, 32)
let random = Self(UInt32.random(in: 0 ... (1 << bits &- 1)))
let dither = Self(sign: sign, exponent: -bits, significand: random)
return (self + dither).rounded(.towardZero)
case .requireExact:
let trunc = rounded(.towardZero)
precondition(isInfinite || trunc == self, "\(self) is not an exact integer.")
@@ -130,27 +130,6 @@ extension BinaryInteger {
return floor + Self((round + lost) >> count)
case .toOdd:
return floor | (lost == 0 ? 0 : 1)
case .stochastically:
// In theory, u01 should be Self.random(in: 0 ..< onesBit), but the
// random(in:) method does not exist on BinaryInteger. This is
// (arguably) good, though, because there's actually no reason to
// generate large amounts of randomness just to implement stochastic
// rounding; 32b suffices for almost all purposes, and 64b is more
// than enough.
var g = SystemRandomNumberGenerator()
let u01 = g.next()
if count < 64 {
// count is small, so mask and lost are representable as both
// UInt64 and Self, regardless of what type Self actually is.
return floor + Self(((u01 & UInt64(mask)) + UInt64(lost)) >> count)
} else {
// count is large, so lost may not be representable as UInt64; pre-
// shift by count-64 to isolate the high 64b of the fraction, then
// add u01 and carry-out to round.
let highWord = UInt64(truncatingIfNeeded: lost >> (Int(count) - 64))
let (_, carry) = highWord.addingReportingOverflow(u01)
return floor + (carry ? 1 : 0)
}
case .requireExact:
precondition(lost == 0, "shift was not exact.")
return floor
-108
View File
@@ -1,108 +0,0 @@
//===--- main.swift -------------------------------------------*- swift -*-===//
//
// This source file is part of the Swift Numerics open source project
//
// Copyright (c) 2020 Apple Inc. and the Swift Numerics project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
import Numerics
import _TestSupport
#if DEBUG
fatalError("Run this test in Release configuration")
#else
var componentError = Double(Float.ulpOfOne/2)
var complexError = Double(Float.ulpOfOne/2)
var componentMaxInput = Complex<Float>.zero
var complexMaxInput = Complex<Float>.zero
func test(_ z: Complex<Float>) {
let tst = Complex.log(z)
// TODO: we _should_ be able to say Complex<Double>(z), but that goes through
// the slow, generic path for BinaryFloatingPoint conversion; once we get
// that resolved in the standard library, we can replace the conversion on
// this and the next line. Currently it dominate testing time if we do that.
let ref = Complex.log(Complex(Double(z.real), Double(z.imaginary)))
if tst == Complex(Float(ref.real), Float(ref.imaginary)) { return }
let thisError = relativeError(tst, ref)
if thisError > complexError {
complexMaxInput = z
complexError = thisError
}
let thisComponentError = componentwiseError(tst, ref)
if thisComponentError > componentError {
componentMaxInput = z
componentError = thisComponentError
}
}
func testWithSymmetries(_ x: Float, _ y: Float) {
test(Complex( x, y))
test(Complex( y, x))
test(Complex(-y, x))
test(Complex(-x, y))
test(Complex(-x,-y))
test(Complex(-y,-x))
test(Complex( y,-x))
test(Complex( x,-y))
}
// The hardest to evaluate cases for log are those close to the unit circle,
// where log(z) is nearly zero. We want to have plausibly dense test coverage
// close to the circle.
let radii: [Float] = [0.9,
0.95,
0.99,
0.999,
0.9999,
1 - 10 * .ulpOfOne,
1,
1 + 10 * .ulpOfOne,
1.0001,
1.001,
1.01,
1.05,
1.1]
for r in radii {
for x in Interval(from: 1/Float.sqrt(2), to: r) {
// Generate the two y values that put us closest to the circle of radius r
let base = Double.sqrt(Double(r).addingProduct(Double(-x), Double(x)))
let a, b: Float
if Double(Float(base)) < base { a = Float(base); b = a.nextUp }
else { b = Float(base); a = b.nextDown }
testWithSymmetries(x, a)
testWithSymmetries(x, b)
}
}
// Away from the unit circle is "easy" but we still want to get some coverage.
// Generate 1 million uniform random points inside the circle, and then test
// both z and 1/z.
var g = SystemRandomNumberGenerator()
var count = 0
while count < 1_000_000 {
let z = Complex<Float>(.random(in: -1 ... 1, using: &g), .random(in: -1 ... 1, using: &g))
if z.length > 1 { continue }
count += 1
test(z)
test(1/z)
}
print("Worst complex norm error seen for log was \(complexError)")
print("For input \(complexMaxInput).")
print("Reference result: \(Complex.log(Complex<Double>(complexMaxInput)))")
print(" Observed result: \(Complex.log(complexMaxInput))")
print("Worst componentwise error seen for log was \(componentError)")
print("For input \(componentMaxInput).")
print("Reference result: \(Complex.log(Complex<Double>(componentMaxInput)))")
print(" Observed result: \(Complex.log(componentMaxInput))")
#endif
-108
View File
@@ -1,108 +0,0 @@
//===--- main.swift -------------------------------------------*- swift -*-===//
//
// This source file is part of the Swift Numerics open source project
//
// Copyright (c) 2020 Apple Inc. and the Swift Numerics project authors
// Licensed under Apache License v2.0 with Runtime Library Exception
//
// See https://swift.org/LICENSE.txt for license information
// See https://swift.org/CONTRIBUTORS.txt for the list of Swift project authors
//
//===----------------------------------------------------------------------===//
import Numerics
import _TestSupport
#if DEBUG
fatalError("Run this test in Release configuration")
#else
var componentError = Double(Float.ulpOfOne/2)
var complexError = Double(Float.ulpOfOne/2)
var componentMaxInput = Complex<Float>.zero
var complexMaxInput = Complex<Float>.zero
func test(_ z: Complex<Float>) {
let tst = Complex.log(onePlus: z)
// TODO: we _should_ be able to say Complex<Double>(z), but that goes through
// the slow, generic path for BinaryFloatingPoint conversion; once we get
// that resolved in the standard library, we can replace the conversion on
// this and the next line. Currently it dominate testing time if we do that.
let ref = Complex.log(onePlus: Complex(Double(z.real), Double(z.imaginary)))
if tst == Complex(Float(ref.real), Float(ref.imaginary)) { return }
let thisError = relativeError(tst, ref)
if thisError > complexError {
complexMaxInput = z
complexError = thisError
}
let thisComponentError = componentwiseError(tst, ref)
if thisComponentError > componentError {
componentMaxInput = z
componentError = thisComponentError
}
}
func testWithSymmetry(_ x: Float, _ y: Float) {
test(Complex( x-1, y))
test(Complex( y-1, x))
test(Complex(-y-1, x))
test(Complex(-x-1, y))
test(Complex(-x-1,-y))
test(Complex(-y-1,-x))
test(Complex( y-1,-x))
test(Complex( x-1,-y))
}
// The hardest to evaluate cases for log are those close to the unit circle
// centered at -1, where log(1+z) is nearly zero. We want to have plausibly
// dense test coverage close to the circle.
let radii: [Float] = [0.9,
0.95,
0.99,
0.999,
0.9999,
1 - 10 * .ulpOfOne,
1,
1 + 10 * .ulpOfOne,
1.0001,
1.001,
1.01,
1.05,
1.1]
for r in radii {
for x in Interval(from: 1/Float.sqrt(2), to: r) {
// Generate the two y values that put us closest to the circle of radius r
let base = Double.sqrt(Double(r).addingProduct(Double(-x), Double(x)))
let a, b: Float
if Double(Float(base)) < base { a = Float(base); b = a.nextUp }
else { b = Float(base); a = b.nextDown }
testWithSymmetry(x, a)
testWithSymmetry(x, b)
}
}
// Away from the unit circle is "easy" but we still want to get some coverage.
// Generate 1 million uniform random points inside the circle, and then test
// both z and 1/z.
var g = SystemRandomNumberGenerator()
var count = 0
while count < 1_000_000 {
let z = Complex<Float>(.random(in: -1 ... 1, using: &g), .random(in: -1 ... 1, using: &g))
if z.length > 1 { continue }
count += 1
test(z - 1)
test(1/z - 1)
}
print("Worst complex norm error seen for log(onePlus:) was \(complexError)")
print("For input \(complexMaxInput).")
print("Reference result: \(Complex.log(onePlus: Complex<Double>(complexMaxInput)))")
print(" Observed result: \(Complex.log(onePlus: complexMaxInput))")
print("Worst componentwise error seen for log(onePlus:) was \(componentError)")
print("For input \(componentMaxInput).")
print("Reference result: \(Complex.log(onePlus: Complex<Double>(componentMaxInput)))")
print(" Observed result: \(Complex.log(onePlus: componentMaxInput))")
#endif
@@ -374,17 +374,6 @@ final class IntegerUtilitiesDivideTests: XCTestCase {
}
}
func testDivideStochastic<T: SignedInteger & FixedWidthInteger>(_ values: [T]) {
for a in values {
for b in values where b != 0 {
// Skip any SignedInt.min / -1 cases, because those will trap.
if a == .min && b == -1 { continue }
let (q, r) = a.divided(by: b, rounding: .stochastically)
let _ = divisionRuleHolds(a, b, q, r)
}
}
}
func testDivideExact<T: SignedInteger & FixedWidthInteger>(_ values: [T]) {
for a in values {
for b in values where b != 0 {
@@ -412,7 +401,6 @@ final class IntegerUtilitiesDivideTests: XCTestCase {
testDivideToNearestOrAway(values)
testDivideToNearestOrEven(values)
testDivideToOdd(values)
testDivideStochastic(values)
testDivideExact(values)
}
@@ -433,7 +421,6 @@ final class IntegerUtilitiesDivideTests: XCTestCase {
testDivideToNearestOrAway(values)
testDivideToNearestOrEven(values)
testDivideToOdd(values)
testDivideStochastic(values)
testDivideExact(values)
}
@@ -454,7 +441,6 @@ final class IntegerUtilitiesDivideTests: XCTestCase {
testDivideToNearestOrAway(values)
testDivideToNearestOrEven(values)
testDivideToOdd(values)
testDivideStochastic(values)
testDivideExact(values)
}
@@ -489,74 +475,10 @@ final class IntegerUtilitiesDivideTests: XCTestCase {
}
}
// stochastically rounding doesn't have a deterministic "expected" answer,
// but we know that the result must be either the floor or the ceiling.
// The above tests ensure that, but that's not a very strong guarantee;
// an implementation could just implement it as self / other and pass
// that test.
//
// Here we round the _same_ value many times, compute the average, and
// check that it is acceptably close to the exact expected value; simple
// use of any deterministic rounding rule will not achieve this.
func testStochasticDivide<T: FixedWidthInteger>(_ a: T, _ b: T) -> Bool {
let trunc = a/b
let sum = (0..<1024).reduce(into: 0.0) { sum, _ in
let rounding = a.divided(by: b, rounding: .stochastically) - trunc
sum += Double(rounding)
}
let expected = 1024*Double(a%b)/Double(b)
let difference = abs(sum - expected)
// Waving my hands slightly instead of giving a precise explanation
// here, the expectation is that difference should be about
// 1/2 sqrt(1024). If we're more than a couple standard deviations
// off, we should flag that. This isn't _necessarily_ a problem,
// but if you see a repeated failure, that's almost surely a real bug.
//
// TODO: precise justification of thresholds
XCTAssertLessThanOrEqual(difference, 64,
"Accumulated error (\(difference)) was unexpectedly large in \(a).divided(by: \(b))"
)
return difference > 16
}
func testDivideStochasticInt8() {
var values = [Int8](repeating: 0, count: 32)
for i in 0 ..< values.count {
while values[i] == 0 {
values[i] = .random(in: .min ... .max)
}
}
var fails = 0
for a in values {
for b in values {
if a == .min && b == -1 { continue }
fails += testStochasticDivide(a, b) ? 1 : 0
}
}
XCTAssertLessThanOrEqual(fails, 32*16)
}
func testDivideStochasticUInt32() {
var values = [UInt32](repeating: 0, count: 32)
for i in 0 ..< values.count {
while values[i] == 0 {
values[i] = .random(in: .min ... .max)
}
}
var fails = 0
for a in values {
for b in values {
fails += testStochasticDivide(a, b) ? 1 : 0
}
}
XCTAssertLessThanOrEqual(fails, 32*16)
}
func testRemainderByMinusOne() {
// These would trap if implemented as a - bq or similar, even though
// the remainder is well-defined.
XCTAssertEqual(0, Int.min.remainder(dividingBy: -1))
XCTAssertEqual(0, Int.min.remainder(dividingBy: -1, rounding: .up))
XCTAssertEqual(0, Int.min.remainder(dividingBy: -1, rounding: .stochastically))
}
}
@@ -77,18 +77,6 @@ final class IntegerUtilitiesShiftTests: XCTestCase {
case 0b11: expected = ceiling
default: preconditionFailure()
}
case .stochastically:
// Just test that it's floor if exact, otherwise either floor
// or ceiling.
let observed = value.shifted(rightBy: count, rounding: rule)
if observed != floor && observed != ceiling {
print("Error found in \(T.self).shifted(rightBy: \(count), rounding: \(rule)).")
print(" Value: \(String(value, radix: 2))")
print("Expected: \(String(floor, radix: 2)) or \(String(ceiling, radix: 2))")
print("Observed: \(String(observed, radix: 2))")
XCTFail()
}
return
case .requireExact:
preconditionFailure()
}
@@ -130,7 +118,6 @@ final class IntegerUtilitiesShiftTests: XCTestCase {
testRoundingShift(Int8.self, rounding: .toNearestOrAway)
testRoundingShift(Int8.self, rounding: .toNearestOrEven)
testRoundingShift(Int8.self, rounding: .toOdd)
testRoundingShift(Int8.self, rounding: .stochastically)
testRoundingShift(UInt8.self, rounding: .down)
testRoundingShift(UInt8.self, rounding: .up)
@@ -142,7 +129,6 @@ final class IntegerUtilitiesShiftTests: XCTestCase {
testRoundingShift(UInt8.self, rounding: .toNearestOrAway)
testRoundingShift(UInt8.self, rounding: .toNearestOrEven)
testRoundingShift(UInt8.self, rounding: .toOdd)
testRoundingShift(UInt8.self, rounding: .stochastically)
testRoundingShift(Int.self, rounding: .down)
testRoundingShift(Int.self, rounding: .up)
@@ -154,7 +140,6 @@ final class IntegerUtilitiesShiftTests: XCTestCase {
testRoundingShift(Int.self, rounding: .toNearestOrAway)
testRoundingShift(Int.self, rounding: .toNearestOrEven)
testRoundingShift(Int.self, rounding: .toOdd)
testRoundingShift(Int.self, rounding: .stochastically)
testRoundingShift(UInt.self, rounding: .down)
testRoundingShift(UInt.self, rounding: .up)
@@ -166,62 +151,5 @@ final class IntegerUtilitiesShiftTests: XCTestCase {
testRoundingShift(UInt.self, rounding: .toNearestOrAway)
testRoundingShift(UInt.self, rounding: .toNearestOrEven)
testRoundingShift(UInt.self, rounding: .toOdd)
testRoundingShift(UInt.self, rounding: .stochastically)
}
// Stochastic rounding doesn't have a deterministic "expected" answer,
// but we know that the result must be either the floor or the ceiling.
// The above tests ensure that, but that's not a very strong guarantee;
// an implementation could just implement it as self >> count and pass
// that test.
//
// Here we round the _same_ value many times, compute the average, and
// check that it is acceptably close to the exact expected value; simple
// use of any deterministic rounding rule will not achieve this.
func testStochasticAverage<T: FixedWidthInteger>(_ value: T) {
var fails = 0
for count in 1 ... T.bitWidth {
let sum = (0..<256).reduce(into: DoubleWidth<T>.zero) { sum, _ in
let rounded = value.shifted(rightBy: count, rounding: .stochastically)
sum += DoubleWidth(rounded)
}
let expected = DoubleWidth<T>(value) << (8 - count)
let difference = sum >= expected ? sum - expected : expected - sum
// Waving my hands slightly instead of giving a precise explanation
// here, the expectation is that difference should be about
// 1/2 sqrt(256). If it's repeatedly bigger than that, we _may_
// have a problem, but it's OK for this to fail occasionally.
//
// TODO: precise justification of thresholds
if difference > 8 { fails += 1 }
// On the other hand, if we're more than a couple standard deviations
// off, we should flag that. This still isn't _necessarily_ a problem,
// but if you see a repeated failure for a given shift count, that's
// almost surely a real bug.
XCTAssertLessThanOrEqual(difference, 32,
"Accumulated error (\(difference)) was unexpectedly large in \(value).shifted(rightBy: \(count))"
)
}
// Threshold chosen so that this is expected to _usually_ pass, but
// it will fail sporadically even with a correct implementation. This is
// not a great fit for CI workflows, sorry. Basically ignore one-off
// failures, but a repeated failure here is an indication that a bug
// exists.
XCTAssertLessThanOrEqual(fails, T.bitWidth/2,
"Accumulated error was large more often than expected for \(value).shifted(rightBy:)"
)
}
func testStochasticShifts() {
testStochasticAverage(Int8.random(in: .min ... .max))
testStochasticAverage(Int16.random(in: .min ... .max))
testStochasticAverage(Int32.random(in: .min ... .max))
testStochasticAverage(UInt8.random(in: .min ... .max))
testStochasticAverage(UInt16.random(in: .min ... .max))
testStochasticAverage(UInt32.random(in: .min ... .max))
testStochasticAverage(Int64.random(in: .min ... .max))
testStochasticAverage(UInt64.random(in: .min ... .max))
testStochasticAverage(DoubleWidth<Int64>.random(in: .min ... .max))
testStochasticAverage(DoubleWidth<UInt64>.random(in: .min ... .max))
}
}
+3 -3
View File
@@ -157,10 +157,10 @@ extension Double {
#endif
assertClose( 1.7976931348623151738531864721534215e308, Double.pow(-u, 3196577161300664268), allowedError: tol)
assertClose(-1.7976931348623155730212483790972209e308, Double.pow(-u, 3196577161300664269), allowedError: tol)
assertClose( 1.7976931348623159721893102860411089e308, Double.pow(-u, 3196577161300664270), allowedError: tol)
assertClose( 1.7976931348623159721893102860411089e308, Double.pow(-u, 3196577161300664270), allowedError: tol) // warning expected on non-x86
assertClose( 1.7976931348623157075547244136070910e308, Double.pow(-d, -6393154322601327474), allowedError: tol)
assertClose(-1.7976931348623159071387553670790721e308, Double.pow(-d, -6393154322601327475), allowedError: tol)
assertClose( 1.7976931348623161067227863205510754e308, Double.pow(-d, -6393154322601327476), allowedError: tol)
assertClose(-1.7976931348623159071387553670790721e308, Double.pow(-d, -6393154322601327475), allowedError: tol) // warning expected on non-x86
assertClose( 1.7976931348623161067227863205510754e308, Double.pow(-d, -6393154322601327476), allowedError: tol) // warning expected on non-x86
// Exponents close to underflow boundary.
assertClose( 2.4703282292062334560337346683707907e-324, Double.pow(-u, -3355781687888880946))
assertClose(-2.4703282292062329075106789791206172e-324, Double.pow(-u, -3355781687888880947))