diff options
Diffstat (limited to 'fpm/math.hpp')
-rw-r--r-- | fpm/math.hpp | 684 |
1 files changed, 684 insertions, 0 deletions
diff --git a/fpm/math.hpp b/fpm/math.hpp new file mode 100644 index 0000000..7a76349 --- /dev/null +++ b/fpm/math.hpp @@ -0,0 +1,684 @@ +#ifndef FPM_MATH_HPP +#define FPM_MATH_HPP + +#include "fixed.hpp" +#include <cmath> + +#ifdef _MSC_VER +#include <intrin.h> +#endif + +namespace fpm +{ + +// +// Helper functions +// +namespace detail +{ + +// Returns the index of the most-signifcant set bit +inline long find_highest_bit(unsigned long long value) noexcept +{ + assert(value != 0); +#if defined(_MSC_VER) + unsigned long index; +#if defined(_WIN64) + _BitScanReverse64(&index, value); +#else + if (_BitScanReverse(&index, static_cast<unsigned long>(value >> 32)) != 0) { + index += 32; + } else { + _BitScanReverse(&index, static_cast<unsigned long>(value & 0xfffffffflu)); + } +#endif + return index; +#elif defined(__GNUC__) || defined(__clang__) + return sizeof(value) * 8 - 1 - __builtin_clzll(value); +#else +# error "your platform does not support find_highest_bit()" +#endif +} + +} + +// +// Classification methods +// + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline int fpclassify(fixed<B, I, F, R> x) noexcept +{ + return (x.raw_value() == 0) ? FP_ZERO : FP_NORMAL; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isfinite(fixed<B, I, F, R>) noexcept +{ + return true; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isinf(fixed<B, I, F, R>) noexcept +{ + return false; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isnan(fixed<B, I, F, R>) noexcept +{ + return false; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isnormal(fixed<B, I, F, R> x) noexcept +{ + return x.raw_value() != 0; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool signbit(fixed<B, I, F, R> x) noexcept +{ + return x.raw_value() < 0; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isgreater(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return x > y; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isgreaterequal(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return x >= y; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isless(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return x < y; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool islessequal(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return x <= y; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool islessgreater(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return x != y; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline bool isunordered(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return false; +} + +// +// Nearest integer operations +// +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> ceil(fixed<B, I, F, R> x) noexcept +{ + constexpr auto FRAC = B(1) << F; + auto value = x.raw_value(); + if (value > 0) value += FRAC - 1; + return fixed<B, I, F, R>::from_raw_value(value / FRAC * FRAC); +} + +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> floor(fixed<B, I, F, R> x) noexcept +{ + constexpr auto FRAC = B(1) << F; + auto value = x.raw_value(); + if (value < 0) value -= FRAC - 1; + return fixed<B, I, F, R>::from_raw_value(value / FRAC * FRAC); +} + +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> trunc(fixed<B, I, F, R> x) noexcept +{ + constexpr auto FRAC = B(1) << F; + return fixed<B, I, F, R>::from_raw_value(x.raw_value() / FRAC * FRAC); +} + +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> round(fixed<B, I, F, R> x) noexcept +{ + constexpr auto FRAC = B(1) << F; + auto value = x.raw_value() / (FRAC / 2); + return fixed<B, I, F, R>::from_raw_value(((value / 2) + (value % 2)) * FRAC); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> nearbyint(fixed<B, I, F, R> x) noexcept +{ + // Rounding mode is assumed to be FE_TONEAREST + constexpr auto FRAC = B(1) << F; + auto value = x.raw_value(); + const bool is_half = std::abs(value % FRAC) == FRAC / 2; + value /= FRAC / 2; + value = (value / 2) + (value % 2); + value -= (value % 2) * is_half; + return fixed<B, I, F, R>::from_raw_value(value * FRAC); +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline fixed<B, I, F, R> rint(fixed<B, I, F, R> x) noexcept +{ + // Rounding mode is assumed to be FE_TONEAREST + return nearbyint(x); +} + +// +// Mathematical functions +// +template <typename B, typename I, unsigned int F, bool R> +constexpr inline fixed<B, I, F, R> abs(fixed<B, I, F, R> x) noexcept +{ + return (x >= fixed<B, I, F, R>{0}) ? x : -x; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline fixed<B, I, F, R> fmod(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return + assert(y.raw_value() != 0), + fixed<B, I, F, R>::from_raw_value(x.raw_value() % y.raw_value()); +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline fixed<B, I, F, R> remainder(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + return + assert(y.raw_value() != 0), + x - nearbyint(x / y) * y; +} + +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> remquo(fixed<B, I, F, R> x, fixed<B, I, F, R> y, int* quo) noexcept +{ + assert(y.raw_value() != 0); + assert(quo != nullptr); + *quo = x.raw_value() / y.raw_value(); + return fixed<B, I, F, R>::from_raw_value(x.raw_value() % y.raw_value()); +} + +// +// Manipulation functions +// + +template <typename B, typename I, unsigned int F, bool R, typename C, typename J, unsigned int G, bool S> +constexpr inline fixed<B, I, F, R> copysign(fixed<B, I, F, R> x, fixed<C, J, G, S> y) noexcept +{ + return + x = abs(x), + (y >= fixed<C, J, G, S>{0}) ? x : -x; +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline fixed<B, I, F, R> nextafter(fixed<B, I, F, R> from, fixed<B, I, F, R> to) noexcept +{ + return from == to ? to : + to > from ? fixed<B, I, F, R>::from_raw_value(from.raw_value() + 1) + : fixed<B, I, F, R>::from_raw_value(from.raw_value() - 1); +} + +template <typename B, typename I, unsigned int F, bool R> +constexpr inline fixed<B, I, F, R> nexttoward(fixed<B, I, F, R> from, fixed<B, I, F, R> to) noexcept +{ + return nextafter(from, to); +} + +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> modf(fixed<B, I, F, R> x, fixed<B, I, F, R>* iptr) noexcept +{ + const auto raw = x.raw_value(); + constexpr auto FRAC = B{1} << F; + *iptr = fixed<B, I, F, R>::from_raw_value(raw / FRAC * FRAC); + return fixed<B, I, F, R>::from_raw_value(raw % FRAC); +} + + +// +// Power functions +// + +template <typename B, typename I, unsigned int F, bool R, typename T, typename std::enable_if<std::is_integral<T>::value>::type* = nullptr> +fixed<B, I, F, R> pow(fixed<B, I, F, R> base, T exp) noexcept +{ + using Fixed = fixed<B, I, F, R>; + + if (base == Fixed(0)) { + assert(exp > 0); + return Fixed(0); + } + + Fixed result {1}; + if (exp < 0) + { + for (Fixed intermediate = base; exp != 0; exp /= 2, intermediate *= intermediate) + { + if ((exp % 2) != 0) + { + result /= intermediate; + } + } + } + else + { + for (Fixed intermediate = base; exp != 0; exp /= 2, intermediate *= intermediate) + { + if ((exp % 2) != 0) + { + result *= intermediate; + } + } + } + return result; +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> pow(fixed<B, I, F, R> base, fixed<B, I, F, R> exp) noexcept +{ + using Fixed = fixed<B, I, F, R>; + + if (base == Fixed(0)) { + assert(exp > Fixed(0)); + return Fixed(0); + } + + if (exp < Fixed(0)) + { + return 1 / pow(base, -exp); + } + + constexpr auto FRAC = B(1) << F; + if (exp.raw_value() % FRAC == 0) + { + // Non-fractional exponents are easier to calculate + return pow(base, exp.raw_value() / FRAC); + } + + // For negative bases we do not support fractional exponents. + // Technically fractions with odd denominators could work, + // but that's too much work to figure out. + assert(base > Fixed(0)); + return exp2(log2(base) * exp); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> exp(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + if (x < Fixed(0)) { + return 1 / exp(-x); + } + constexpr auto FRAC = B(1) << F; + const B x_int = x.raw_value() / FRAC; + x -= x_int; + assert(x >= Fixed(0) && x < Fixed(1)); + + constexpr auto fA = Fixed::template from_fixed_point<63>( 128239257017632854ll); // 1.3903728105644451e-2 + constexpr auto fB = Fixed::template from_fixed_point<63>( 320978614890280666ll); // 3.4800571158543038e-2 + constexpr auto fC = Fixed::template from_fixed_point<63>(1571680799599592947ll); // 1.7040197373796334e-1 + constexpr auto fD = Fixed::template from_fixed_point<63>(4603349000587966862ll); // 4.9909609871464493e-1 + constexpr auto fE = Fixed::template from_fixed_point<62>(4612052447974689712ll); // 1.0000794567422495 + constexpr auto fF = Fixed::template from_fixed_point<63>(9223361618412247875ll); // 9.9999887043019773e-1 + return pow(Fixed::e(), x_int) * (((((fA * x + fB) * x + fC) * x + fD) * x + fE) * x + fF); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> exp2(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + if (x < Fixed(0)) { + return 1 / exp2(-x); + } + constexpr auto FRAC = B(1) << F; + const B x_int = x.raw_value() / FRAC; + x -= x_int; + assert(x >= Fixed(0) && x < Fixed(1)); + + constexpr auto fA = Fixed::template from_fixed_point<63>( 17491766697771214ll); // 1.8964611454333148e-3 + constexpr auto fB = Fixed::template from_fixed_point<63>( 82483038782406547ll); // 8.9428289841091295e-3 + constexpr auto fC = Fixed::template from_fixed_point<63>( 515275173969157690ll); // 5.5866246304520701e-2 + constexpr auto fD = Fixed::template from_fixed_point<63>(2214897896212987987ll); // 2.4013971109076949e-1 + constexpr auto fE = Fixed::template from_fixed_point<63>(6393224161192452326ll); // 6.9315475247516736e-1 + constexpr auto fF = Fixed::template from_fixed_point<63>(9223371050976163566ll); // 9.9999989311082668e-1 + return Fixed(1 << x_int) * (((((fA * x + fB) * x + fC) * x + fD) * x + fE) * x + fF); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> expm1(fixed<B, I, F, R> x) noexcept +{ + return exp(x) - 1; +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> log2(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + assert(x > Fixed(0)); + + // Normalize input to the [1:2] domain + B value = x.raw_value(); + const long highest = detail::find_highest_bit(value); + if (highest >= F) { + value >>= (highest - F); + } else { + value <<= (F - highest); + } + x = Fixed::from_raw_value(value); + assert(x >= Fixed(1) && x < Fixed(2)); + + constexpr auto fA = Fixed::template from_fixed_point<63>( 413886001457275979ll); // 4.4873610194131727e-2 + constexpr auto fB = Fixed::template from_fixed_point<63>(-3842121857793256941ll); // -4.1656368651734915e-1 + constexpr auto fC = Fixed::template from_fixed_point<62>( 7522345947206307744ll); // 1.6311487636297217 + constexpr auto fD = Fixed::template from_fixed_point<61>(-8187571043052183818ll); // -3.5507929249026341 + constexpr auto fE = Fixed::template from_fixed_point<60>( 5870342889289496598ll); // 5.0917108110420042 + constexpr auto fF = Fixed::template from_fixed_point<61>(-6457199832668582866ll); // -2.8003640347009253 + return Fixed(highest - F) + (((((fA * x + fB) * x + fC) * x + fD) * x + fE) * x + fF); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> log(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + return log2(x) / log2(Fixed::e()); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> log10(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + return log2(x) / log2(Fixed(10)); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> log1p(fixed<B, I, F, R> x) noexcept +{ + return log(1 + x); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> cbrt(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + + if (x == Fixed(0)) + { + return x; + } + if (x < Fixed(0)) + { + return -cbrt(-x); + } + assert(x >= Fixed(0)); + + // Finding the cube root of an integer, taken from Hacker's Delight, + // based on the square root algorithm. + + // We start at the greatest power of eight that's less than the argument. + int ofs = ((detail::find_highest_bit(x.raw_value()) + 2*F) / 3 * 3); + I num = I{x.raw_value()}; + I res = 0; + + const auto do_round = [&] + { + for (; ofs >= 0; ofs -= 3) + { + res += res; + const I val = (3*res*(res + 1) + 1) << ofs; + if (num >= val) + { + num -= val; + res++; + } + } + }; + + // We should shift by 2*F (since there are two multiplications), but that + // could overflow even the intermediate type, so we have to split the + // algorithm up in two rounds of F bits each. Each round will deplete + // 'num' digit by digit, so after a round we can shift it again. + num <<= F; + ofs -= F; + do_round(); + + num <<= F; + ofs += F; + do_round(); + + return Fixed::from_raw_value(static_cast<B>(res)); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> sqrt(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + + assert(x >= Fixed(0)); + if (x == Fixed(0)) + { + return x; + } + + // Finding the square root of an integer in base-2, from: + // https://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Binary_numeral_system_.28base_2.29 + + // Shift by F first because it's fixed-point. + I num = I{x.raw_value()} << F; + I res = 0; + + // "bit" starts at the greatest power of four that's less than the argument. + for (I bit = I{1} << ((detail::find_highest_bit(x.raw_value()) + F) / 2 * 2); bit != 0; bit >>= 2) + { + const I val = res + bit; + res >>= 1; + if (num >= val) + { + num -= val; + res += bit; + } + } + + // Round the last digit up if necessary + if (num > res) + { + res++; + } + + return Fixed::from_raw_value(static_cast<B>(res)); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> hypot(fixed<B, I, F, R> x, fixed<B, I, F, R> y) noexcept +{ + assert(x != 0 || y != 0); + return sqrt(x*x + y*y); +} + +// +// Trigonometry functions +// + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> sin(fixed<B, I, F, R> x) noexcept +{ + // This sine uses a fifth-order curve-fitting approximation originally + // described by Jasper Vijn on coranac.com which has a worst-case + // relative error of 0.07% (over [-pi:pi]). + using Fixed = fixed<B, I, F, R>; + + // Turn x from [0..2*PI] domain into [0..4] domain + x = fmod(x, Fixed::two_pi()); + x = x / Fixed::half_pi(); + + // Take x modulo one rotation, so [-4..+4]. + if (x < Fixed(0)) { + x += Fixed(4); + } + + int sign = +1; + if (x > Fixed(2)) { + // Reduce domain to [0..2]. + sign = -1; + x -= Fixed(2); + } + + if (x > Fixed(1)) { + // Reduce domain to [0..1]. + x = Fixed(2) - x; + } + + const Fixed x2 = x*x; + return sign * x * (Fixed::pi() - x2*(Fixed::two_pi() - 5 - x2*(Fixed::pi() - 3)))/2; +} + +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> cos(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + if (x > Fixed(0)) { // Prevent an overflow due to the addition of π/2 + return sin(x - (Fixed::two_pi() - Fixed::half_pi())); + } else { + return sin(Fixed::half_pi() + x); + } +} + +template <typename B, typename I, unsigned int F, bool R> +inline fixed<B, I, F, R> tan(fixed<B, I, F, R> x) noexcept +{ + auto cx = cos(x); + + // Tangent goes to infinity at 90 and -90 degrees. + // We can't represent that with fixed-point maths. + assert(abs(cx).raw_value() > 1); + + return sin(x) / cx; +} + +namespace detail { + +// Calculates atan(x) assuming that x is in the range [0,1] +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> atan_sanitized(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + assert(x >= Fixed(0) && x <= Fixed(1)); + + constexpr auto fA = Fixed::template from_fixed_point<63>( 716203666280654660ll); // 0.0776509570923569 + constexpr auto fB = Fixed::template from_fixed_point<63>(-2651115102768076601ll); // -0.287434475393028 + constexpr auto fC = Fixed::template from_fixed_point<63>( 9178930894564541004ll); // 0.995181681698119 (PI/4 - A - B) + + const auto xx = x * x; + return ((fA*xx + fB)*xx + fC)*x; +} + +// Calculate atan(y / x), assuming x != 0. +// +// If x is very, very small, y/x can easily overflow the fixed-point range. +// If q = y/x and q > 1, atan(q) would calculate atan(1/q) as intermediate step +// anyway. We can shortcut that here and avoid the loss of information, thus +// improving the accuracy of atan(y/x) for very small x. +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> atan_div(fixed<B, I, F, R> y, fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + assert(x != Fixed(0)); + + // Make sure y and x are positive. + // If y / x is negative (when y or x, but not both, are negative), negate the result to + // keep the correct outcome. + if (y < Fixed(0)) { + if (x < Fixed(0)) { + return atan_div(-y, -x); + } + return -atan_div(-y, x); + } + if (x < Fixed(0)) { + return -atan_div(y, -x); + } + assert(y >= Fixed(0)); + assert(x > Fixed(0)); + + if (y > x) { + return Fixed::half_pi() - detail::atan_sanitized(x / y); + } + return detail::atan_sanitized(y / x); +} + +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> atan(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + if (x < Fixed(0)) + { + return -atan(-x); + } + + if (x > Fixed(1)) + { + return Fixed::half_pi() - detail::atan_sanitized(Fixed(1) / x); + } + + return detail::atan_sanitized(x); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> asin(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + assert(x >= Fixed(-1) && x <= Fixed(+1)); + + const auto yy = Fixed(1) - x * x; + if (yy == Fixed(0)) + { + return copysign(Fixed::half_pi(), x); + } + return detail::atan_div(x, sqrt(yy)); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> acos(fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + assert(x >= Fixed(-1) && x <= Fixed(+1)); + + if (x == Fixed(-1)) + { + return Fixed::pi(); + } + const auto yy = Fixed(1) - x * x; + return Fixed(2)*detail::atan_div(sqrt(yy), Fixed(1) + x); +} + +template <typename B, typename I, unsigned int F, bool R> +fixed<B, I, F, R> atan2(fixed<B, I, F, R> y, fixed<B, I, F, R> x) noexcept +{ + using Fixed = fixed<B, I, F, R>; + if (x == Fixed(0)) + { + assert(y != Fixed(0)); + return (y > Fixed(0)) ? Fixed::half_pi() : -Fixed::half_pi(); + } + + auto ret = detail::atan_div(y, x); + + if (x < Fixed(0)) + { + return (y >= Fixed(0)) ? ret + Fixed::pi() : ret - Fixed::pi(); + } + return ret; +} + +} + +#endif |