162 lines
6.9 KiB
C#
162 lines
6.9 KiB
C#
namespace Just.PreciseMath.Tests;
|
|
|
|
public class PreciseMathReciprocalTests
|
|
{
|
|
[Theory]
|
|
[InlineData(3.0)]
|
|
[InlineData(7.0)]
|
|
[InlineData(-3.0)]
|
|
[InlineData(-7.0)]
|
|
public void OrdinaryReciprocalsRetainMoreThanBinary64Precision(double high)
|
|
{
|
|
foreach (double low in new[] { 0.0, Math.ScaleB(1.0, -54), -Math.ScaleB(1.0, -54) })
|
|
{
|
|
DoubleDouble input = DoubleDouble.FromComponents(high, low);
|
|
DoubleDouble actual = DDMath.Reciprocal(input);
|
|
actual.Low.ShouldNotBe(0.0);
|
|
AssertReciprocalBound(input, actual);
|
|
}
|
|
}
|
|
|
|
[Fact]
|
|
public void SpecialValuesFollowBinary64ReciprocalAndRemainCanonical()
|
|
{
|
|
double[] values = [0.0, -0.0, double.PositiveInfinity, double.NegativeInfinity,
|
|
double.NaN, BitConverter.Int64BitsToDouble(0x7ff0000000000001L)];
|
|
foreach (double value in values)
|
|
{
|
|
DoubleDouble actual = DDMath.Reciprocal(new DoubleDouble(value));
|
|
// Binary64 is an independent oracle for special values only.
|
|
double expected = 1.0 / value;
|
|
BitConverter.DoubleToInt64Bits(actual.High).ShouldBe(
|
|
BitConverter.DoubleToInt64Bits(double.IsNaN(expected) ? double.NaN : expected));
|
|
BitConverter.DoubleToInt64Bits(actual.Low).ShouldBe(0L);
|
|
DoubleDouble.IsCanonical(actual).ShouldBeTrue();
|
|
}
|
|
}
|
|
|
|
[Fact]
|
|
public void PowersOfTwoHaveExactReciprocalsOrSignedOverflowAcrossTheRange()
|
|
{
|
|
// 1/(s*2^k) = s*2^-k. All finite results here are exactly representable,
|
|
// including subnormals. Exponents below -1023 overflow the result.
|
|
for (int exponent = -1074; exponent <= 1023; ++exponent)
|
|
{
|
|
foreach (double sign in new[] { -1.0, 1.0 })
|
|
{
|
|
DoubleDouble actual = DDMath.Reciprocal(new DoubleDouble(Math.ScaleB(sign, exponent)));
|
|
actual.High.ShouldBe(Math.ScaleB(sign, -exponent));
|
|
BitConverter.DoubleToInt64Bits(actual.Low).ShouldBe(0L);
|
|
DoubleDouble.IsCanonical(actual).ShouldBeTrue();
|
|
}
|
|
}
|
|
}
|
|
|
|
[Fact]
|
|
public void ReciprocalsMatchScalarDivisionBitsAcrossTheRange()
|
|
{
|
|
// Compatibility is bitwise, not just the same accuracy tolerance. Sample
|
|
// every exponent, both signs, dense/sparse lows and binade neighbors;
|
|
// this includes both ends of the scalar division fast-path guard.
|
|
Random random = new(314159);
|
|
for (int exponent = -1074; exponent <= 1023; ++exponent)
|
|
{
|
|
double high = Math.ScaleB(1.0 + random.NextDouble(), exponent);
|
|
double halfUlp = Math.ScaleB(1.0, exponent - 53);
|
|
double power = Math.ScaleB(1.0, exponent);
|
|
foreach (double sign in new[] { -1.0, 1.0 })
|
|
{
|
|
foreach (double low in new[] { 0.0, halfUlp, -halfUlp,
|
|
Math.BitDecrement(halfUlp), -Math.BitDecrement(halfUlp),
|
|
Math.BitIncrement(halfUlp), -Math.BitIncrement(halfUlp),
|
|
double.Epsilon, -double.Epsilon })
|
|
{
|
|
AssertMatchesDivision(DoubleDouble.FromComponents(sign * high, low));
|
|
}
|
|
foreach (double boundary in new[] { Math.BitDecrement(power), power, Math.BitIncrement(power) })
|
|
{
|
|
AssertMatchesDivision(new DoubleDouble(sign * boundary));
|
|
}
|
|
}
|
|
}
|
|
|
|
DoubleDouble[] specials = [DoubleDouble.Zero, new(-0.0), DoubleDouble.NaN,
|
|
new(double.PositiveInfinity), new(double.NegativeInfinity),
|
|
new(double.MaxValue), new(double.MinValue),
|
|
DoubleDouble.FromComponents(double.MaxValue, Math.BitDecrement(Math.ScaleB(1.0, 970))),
|
|
DoubleDouble.FromComponents(double.MinValue, -Math.BitDecrement(Math.ScaleB(1.0, 970)))];
|
|
foreach (DoubleDouble input in specials)
|
|
{
|
|
AssertMatchesDivision(input);
|
|
}
|
|
}
|
|
|
|
[Fact]
|
|
public void RepresentableSparseCorrectionsAreRetained()
|
|
{
|
|
// 1/(1+d) = 1-d+O(d^2). The omitted tail is less than half an ulp
|
|
// of d for these exact dyadics, including the minimum subnormal.
|
|
foreach (int exponent in new[] { -100, -500, -1000, -1074 })
|
|
{
|
|
foreach (double sign in new[] { -1.0, 1.0 })
|
|
{
|
|
foreach (double lowSign in new[] { -1.0, 1.0 })
|
|
{
|
|
double low = Math.ScaleB(lowSign, exponent);
|
|
DoubleDouble input = DoubleDouble.FromComponents(sign, low);
|
|
DoubleDouble actual = DDMath.Reciprocal(input);
|
|
actual.High.ShouldBe(sign);
|
|
actual.Low.ShouldBe(-low);
|
|
AssertReciprocalBound(input, actual);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
private static void AssertMatchesDivision(DoubleDouble input)
|
|
{
|
|
DoubleDouble expected = 1.0 / input;
|
|
DoubleDouble actual = DDMath.Reciprocal(input);
|
|
string context = $"Input: ({input.High:R}, {input.Low:R})";
|
|
BitConverter.DoubleToInt64Bits(actual.High).ShouldBe(BitConverter.DoubleToInt64Bits(expected.High), context);
|
|
BitConverter.DoubleToInt64Bits(actual.Low).ShouldBe(BitConverter.DoubleToInt64Bits(expected.Low), context);
|
|
DoubleDouble.IsCanonical(actual).ShouldBeTrue();
|
|
if (DoubleDouble.IsFinite(input) && input.High != 0.0 && DoubleDouble.IsFinite(actual))
|
|
{
|
|
// Division establishes compatibility, not accuracy: check the exact
|
|
// rational inequality independently for every finite matrix result.
|
|
AssertReciprocalBound(input, actual);
|
|
}
|
|
}
|
|
|
|
private static void AssertReciprocalBound(DoubleDouble input, DoubleDouble actual)
|
|
{
|
|
DoubleDouble.IsFinite(actual).ShouldBeTrue();
|
|
DoubleDouble.IsCanonical(actual).ShouldBeTrue();
|
|
Math.Sign(actual.High).ShouldBe(Math.Sign(input.High));
|
|
|
|
// Exact dyadic oracle: x = X*2^-1074, y = Y*2^-1074.
|
|
// |y - 1/x| <= |1/x|*2^-100 + 2^-1074 is equivalent to
|
|
// |X*Y - 2^2148|*2^100 <= 2^2148 + |X|*2^100.
|
|
BigInteger x = Units(input.High) + Units(input.Low);
|
|
BigInteger y = Units(actual.High) + Units(actual.Low);
|
|
BigInteger scale = BigInteger.One << 2148;
|
|
BigInteger error = BigInteger.Abs((x * y) - scale) << 100;
|
|
(error <= scale + (BigInteger.Abs(x) << 100)).ShouldBeTrue(
|
|
$"Reciprocal bound failed for ({input.High:R}, {input.Low:R}): ({actual.High:R}, {actual.Low:R})");
|
|
}
|
|
|
|
private static BigInteger Units(double value)
|
|
{
|
|
long bits = BitConverter.DoubleToInt64Bits(value);
|
|
int exponent = (int)((bits >> 52) & 0x7ff);
|
|
BigInteger significand = bits & 0xfffffffffffffL;
|
|
if (exponent != 0)
|
|
{
|
|
significand += BigInteger.One << 52;
|
|
significand <<= exponent - 1;
|
|
}
|
|
return bits < 0 ? -significand : significand;
|
|
}
|
|
}
|