public void Subtract(DiyFp other) { Debug.Assert(Exponent == other.Exponent); Debug.Assert(Significand >= other.Significand); Significand -= other.Significand; }
public static DiyFp Minus(DiyFp a, DiyFp b) { var result = a; result.Subtract(b); return(result); }
public static DiyFp Normalize(DiyFp a) { var result = a; result.Normalize(); return(result); }
public static DiyFp Times(DiyFp a, DiyFp b) { var result = a; result.Multiply(b); return(result); }
public void NormalizedBoundaries(out DiyFp minus, out DiyFp plus) { Debug.Assert(Value() > 0.0); var v = AsDiyFp(); plus = DiyFp.Normalize(new DiyFp((v.Significand << 1) + 1, v.Exponent - 1)); minus = IsLowerBoundaryCloser() ? new DiyFp((v.Significand << 2) - 1, v.Exponent - 2) : new DiyFp((v.Significand << 1) - 1, v.Exponent - 1); minus.Significand = minus.Significand << (minus.Exponent - plus.Exponent); minus.Exponent = plus.Exponent; }
// Returns a cached power-of-ten with a binary exponent in the range // [min_exponent; max_exponent] (boundaries included). private static void GetCachedPowerForBinaryExponentRange(Int32 min_exponent, Int32 max_exponent, out DiyFp power, out Int32 decimal_exponent) { const Double D_1_LOG2_10 = 0.30102999566398114; // 1 / lg(10) var kQ = DiyFp.SignificantSize; var k = Math.Ceiling((min_exponent + kQ - 1) * D_1_LOG2_10); var index = (CACHED_POWERS_OFFSET + (Int32)(k) - 1) / DECIMAL_EXPONENT_DISTANCE + 1; var cached_power = CACHED_POWERS[index]; Debug.Assert(min_exponent <= cached_power.binary_exponent); Debug.Assert(cached_power.binary_exponent <= max_exponent); decimal_exponent = cached_power.decimal_exponent; power = new DiyFp(cached_power.significand, cached_power.binary_exponent); }
public void Multiply(DiyFp other) { // Simply "emulates" a 128 bit multiplication. // However: the resulting number only contains 64 bits. The least // significant 64 bits are only used for rounding the most significant 64 bits. const UInt64 M32 = 0xFFFFFFFFU; var a = Significand >> 32; var b = Significand & M32; var c = other.Significand >> 32; var d = other.Significand & M32; var ac = a * c; var bc = b * c; var ad = a * d; var bd = b * d; var tmp = (bd >> 32) + (ad & M32) + (bc & M32); // By adding 1U << 31 to tmp we round the final result. // Halfway cases will be round up. tmp += 1U << 31; Exponent += other.Exponent + 64; Significand = ac + (ad >> 32) + (bc >> 32) + (tmp >> 32); }
private static Boolean DigitGen(DiyFp low, DiyFp w, DiyFp high, Span <Byte> buffer, out Int32 length, out Int32 kappa) { Debug.Assert(low.Exponent == w.Exponent && w.Exponent == high.Exponent); Debug.Assert(low.Significand + 1 <= high.Significand - 1); Debug.Assert(MINIMAL_TARGET_EXPONENT <= w.Exponent && w.Exponent <= MAXIMAL_TARGET_EXPONENT); // low, w and high are imprecise, but by less than one ulp (unit in the last place). // If we remove (resp. add) 1 ulp from low (resp. high) we are certain that // the new numbers are outside of the interval we want the final representation to lie in. // Inversely adding (resp. removing) 1 ulp from low (resp. high) would yield // numbers that are certain to lie in the interval. We will use this fact later on. // We will now start by generating the digits within the uncertain // interval. Later we will weed out representations that lie outside the safe // interval and thus _might_ lie outside the correct interval. UInt64 unit = 1; var too_low = new DiyFp(low.Significand - unit, low.Exponent); var too_high = new DiyFp(high.Significand + unit, high.Exponent); // too_low and too_high are guaranteed to lie outside the interval we want the generated number in. var unsafe_interval = DiyFp.Minus(too_high, too_low); // We now cut the input number into two parts: the integral digits and the // fractionals. We will not write any decimal separator though, but adapt // kappa instead. // Reminder: we are currently computing the digits (stored inside the buffer) // such that: too_low < buffer * 10^kappa < too_high // We use too_high for the digit_generation and stop as soon as possible. // If we stop early we effectively round down. var one = new DiyFp(1UL << -w.Exponent, w.Exponent); // Division by one is a shift. var integrals = (UInt32)(too_high.Significand >> -one.Exponent); // Modulo by one is an and. var fractionals = too_high.Significand & (one.Significand - 1); BiggestPowerTen(integrals, DiyFp.SignificantSize - (-one.Exponent), out var divisor, out var divisor_exponent_plus_one); kappa = divisor_exponent_plus_one; length = 0; // Loop invariant: buffer = too_high / 10^kappa (integer division) // The invariant holds for the first iteration: kappa has been initialized // with the divisor exponent + 1. And the divisor is the biggest power of ten // that is smaller than integrals. while (kappa > 0) { var digit = integrals / divisor; Debug.Assert(digit <= 9); buffer[length] = (Byte)('0' + digit); ++length; integrals %= divisor; --kappa; // Note that kappa now equals the exponent of the divisor and that the // invariant thus holds again. var rest = ((UInt64)(integrals) << -one.Exponent) + fractionals; // Invariant: too_high = buffer * 10^kappa + DiyFp(rest, one.e()) // Reminder: unsafe_interval.e() == one.e() if (rest < unsafe_interval.Significand) { // Rounding down (by not emitting the remaining digits) yields a number // that lies within the unsafe interval. return(RoundWeed(buffer, length, DiyFp.Minus(too_high, w).Significand, unsafe_interval.Significand, rest, (UInt64)(divisor) << -one.Exponent, unit)); } divisor /= 10; } // The integrals have been generated. We are at the point of the decimal // separator. In the following loop we simply multiply the remaining digits by // 10 and divide by one. We just need to pay attention to multiply associated // data (like the interval or 'unit'), too. // Note that the multiplication by 10 does not overflow, because w.e >= -60 // and thus one.e >= -60. Debug.Assert(one.Exponent >= -60); Debug.Assert(fractionals < one.Significand); Debug.Assert(0xFFFFFFFF_FFFFFFFF / 10 >= one.Significand); while (true) { fractionals *= 10; unit *= 10; unsafe_interval.Significand = unsafe_interval.Significand * 10; // Integer division by one. var digit = (Int32)(fractionals >> -one.Exponent); Debug.Assert(digit <= 9); buffer[length] = (Byte)('0' + digit); ++length; fractionals &= one.Significand - 1; // Modulo by one. --kappa; if (fractionals < unsafe_interval.Significand) { return(RoundWeed(buffer, length, DiyFp.Minus(too_high, w).Significand *unit, unsafe_interval.Significand, fractionals, one.Significand, unit)); } } }
private static Boolean Grisu3(Double v, Boolean single_precision, Span <Byte> buffer, out Int32 length, out Int32 decimal_exponent) { var w = new DiyDouble(v).AsNormalizedDiyFp(); // boundary_minus and boundary_plus are the boundaries between v and its // closest floating-point neighbors. Any number strictly between // boundary_minus and boundary_plus will round to v when convert to a double. // Grisu3 will never output representations that lie exactly on a boundary. DiyFp boundary_minus, boundary_plus; if (single_precision) { new DiySingle((Single)v).NormalizedBoundaries(out boundary_minus, out boundary_plus); } else { new DiyDouble(v).NormalizedBoundaries(out boundary_minus, out boundary_plus); } Debug.Assert(boundary_plus.Exponent == w.Exponent); var ten_mk_minimal_binary_exponent = MINIMAL_TARGET_EXPONENT - (w.Exponent + DiyFp.SignificantSize); var ten_mk_maximal_binary_exponent = MAXIMAL_TARGET_EXPONENT - (w.Exponent + DiyFp.SignificantSize); // Note that ten_mk is only an approximation of 10^-k. A DiyFp only contains a // 64 bit significand and ten_mk is thus only precise up to 64 bits. GetCachedPowerForBinaryExponentRange(ten_mk_minimal_binary_exponent, ten_mk_maximal_binary_exponent, out var ten_mk, out var mk); Debug.Assert(MINIMAL_TARGET_EXPONENT <= w.Exponent + ten_mk.Exponent + DiyFp.SignificantSize); Debug.Assert(MAXIMAL_TARGET_EXPONENT >= w.Exponent + ten_mk.Exponent + DiyFp.SignificantSize); // The DiyFp.Times procedure rounds its result, and ten_mk is approximated // too. The variable scaled_w (as well as scaled_boundary_minus/plus) are now // off by a small amount. // In fact: scaled_w - w*10^k < 1ulp (unit in the last place) of scaled_w. // In other words: let f = scaled_w.f() and e = scaled_w.e(), then (f-1) * 2^e < w*10^k < (f+1) * 2^e var scaled_w = DiyFp.Times(w, ten_mk); Debug.Assert(scaled_w.Exponent == boundary_plus.Exponent + ten_mk.Exponent + DiyFp.SignificantSize); // In theory it would be possible to avoid some recomputations by computing // the difference between w and boundary_minus/plus (a power of 2) and to // compute scaled_boundary_minus/plus by subtracting/adding from // scaled_w. However the code becomes much less readable and the speed enhancements are not terriffic. var scaled_boundary_minus = DiyFp.Times(boundary_minus, ten_mk); var scaled_boundary_plus = DiyFp.Times(boundary_plus, ten_mk); // DigitGen will generate the digits of scaled_w. Therefore we have // v == (double) (scaled_w * 10^-mk). // Set decimal_exponent == -mk and pass it to DigitGen. If scaled_w is not an // integer than it will be updated. For instance if scaled_w == 1.23 then // the buffer will be filled with "123" und the decimal_exponent will be decreased by 2. var result = DigitGen(scaled_boundary_minus, scaled_w, scaled_boundary_plus, buffer, out length, out var kappa); decimal_exponent = -mk + kappa; return(result); }