Compute values for rational magnitude powers

Since this will only ever be done at compile time (as guaranteed by
using `consteval`), we can afford to prioritize precision over speed.
To compute an Nth root, we simply do a binary search over representable
floating point numbers, looking for the number whose Nth power most
closely matches the original number.

Fixes #494.  We have included a test case reproducing the original
problem exactly.  All tests use "within 4 ULPs" as the criterion, which
is (I believe) equivalent to the googletest `EXPECT_DOUBLE_EQ`
criterion.
This commit is contained in:
Chip Hogg
2024-07-24 10:31:44 -04:00
parent f9ae701b4e
commit 7e894788d7
2 changed files with 155 additions and 5 deletions

View File

@ -307,6 +307,119 @@ template<typename T>
return checked_square(int_power(base, exp / 2));
}
template <typename T>
[[nodiscard]] consteval std::optional<T> checked_int_pow(T base, std::uintmax_t exp) {
T result = T{1};
while (exp > 0u) {
if (exp % 2u == 1u) {
if (base > std::numeric_limits<T>::max() / result) {
return std::nullopt;
}
result *= base;
}
exp /= 2u;
if (base > std::numeric_limits<T>::max() / base) {
return (exp == 0u)
? std::make_optional(result)
: std::nullopt;
}
base *= base;
}
return result;
}
template <typename T>
[[nodiscard]] consteval std::optional<T> root(T x, std::uintmax_t n) {
// The "zeroth root" would be mathematically undefined.
if (n == 0) {
return std::nullopt;
}
// The "first root" is trivial.
if (n == 1) {
return x;
}
// We only support nontrivial roots of floating point types.
if (!std::is_floating_point<T>::value) {
return std::nullopt;
}
// Handle negative numbers: only odd roots are allowed.
if (x < 0) {
if (n % 2 == 0) {
return std::nullopt;
} else {
const auto negative_result = root(-x, n);
if (!negative_result.has_value()) {
return std::nullopt;
}
return static_cast<T>(-negative_result.value());
}
}
// Handle special cases of zero and one.
if (x == 0 || x == 1) {
return x;
}
// Handle numbers bewtween 0 and 1.
if (x < 1) {
const auto inverse_result = root(T{1} / x, n);
if (!inverse_result.has_value()) {
return std::nullopt;
}
return static_cast<T>(T{1} / inverse_result.value());
}
//
// At this point, error conditions are finished, and we can proceed with the "core" algorithm.
//
// Always use `long double` for intermediate computations. We don't ever expect people to be
// calling this at runtime, so we want maximum accuracy.
long double lo = 1.0;
long double hi = static_cast<long double>(x);
// Do a binary search to find the closest value such that `checked_int_pow` recovers the input.
//
// Because we know `n > 1`, and `x > 1`, and x^n is monotonically increasing, we know that
// `checked_int_pow(lo, n) < x < checked_int_pow(hi, n)`. We will preserve this as an
// invariant.
while (lo < hi) {
long double mid = lo + (hi - lo) / 2;
auto result = checked_int_pow(mid, n);
if (!result.has_value()) {
return std::nullopt;
}
// Early return if we get lucky with an exact answer.
if (result.value() == x) {
return static_cast<T>(mid);
}
// Check for stagnation.
if (mid == lo || mid == hi) {
break;
}
// Preserve the invariant that `checked_int_pow(lo, n) < x < checked_int_pow(hi, n)`.
if (result.value() < x) {
lo = mid;
} else {
hi = mid;
}
}
// Pick whichever one gets closer to the target.
const auto lo_diff = x - checked_int_pow(lo, n).value();
const auto hi_diff = checked_int_pow(hi, n).value() - x;
return static_cast<T>(lo_diff < hi_diff ? lo : hi);
}
template<typename T>
[[nodiscard]] consteval widen_t<T> compute_base_power(MagnitudeSpec auto el)
@ -317,9 +430,6 @@ template<typename T>
// Note that since this function should only be called at compile time, the point of these
// terminations is to act as "static_assert substitutes", not to actually terminate at runtime.
const auto exp = get_exponent(el);
if (exp.den != 1) {
std::abort(); // Rational powers not yet supported
}
if (exp.num < 0) {
if constexpr (std::is_integral_v<T>) {
@ -329,8 +439,19 @@ template<typename T>
}
}
auto power = exp.num;
return int_power(static_cast<widen_t<T>>(get_base_value(el)), power);
const auto pow_result = checked_int_pow(static_cast<widen_t<T>>(get_base_value(el)), static_cast<uintmax_t>(exp.num));
if (pow_result.has_value()) {
const auto final_result = (exp.den > 1) ? root(pow_result.value(), static_cast<uintmax_t>(exp.den)) : pow_result;
if (final_result.has_value()) {
return final_result.value();
}
else {
std::abort(); // Root computation failed.
}
}
else {
std::abort(); // Power computation failed.
}
}
// A converter for the value member variable of magnitude (below).

View File

@ -24,6 +24,7 @@
#include <mp-units/bits/hacks.h>
#include <mp-units/ext/fixed_string.h>
#include <mp-units/ext/type_traits.h>
#include <mp-units/math.h>
#include <mp-units/systems/isq/mechanics.h>
#include <mp-units/systems/isq/space_and_time.h>
#include <mp-units/systems/si.h>
@ -52,6 +53,24 @@ using namespace mp_units::si::unit_symbols;
// quantity class invariants
//////////////////////////////
template <typename T>
constexpr bool within_4_ulps(T a, T b) {
static_assert(std::is_floating_point_v<T>);
auto walk_ulps = [](T x, int n) {
while (n > 0) {
x = std::nextafter(x, std::numeric_limits<T>::infinity());
--n;
}
while (n < 0) {
x = std::nextafter(x, -std::numeric_limits<T>::infinity());
++n;
}
return x;
};
return (walk_ulps(a, -4) <= b) && (b <= walk_ulps(a, 4));
}
static_assert(sizeof(quantity<si::metre>) == sizeof(double));
static_assert(sizeof(quantity<isq::length[m]>) == sizeof(double));
static_assert(sizeof(quantity<si::metre, short>) == sizeof(short));
@ -199,6 +218,16 @@ static_assert(std::convertible_to<quantity<isq::length[km], int>, quantity<isq::
static_assert(std::constructible_from<quantity<isq::length[km]>, quantity<isq::length[m], int>>);
static_assert(std::convertible_to<quantity<isq::length[m], int>, quantity<isq::length[km]>>);
// conversion requiring radical magnitudes
static_assert(within_4_ulps(sqrt((1.0 * m) * (1.0 * km)).numerical_value_in(m), sqrt(1000.0)));
// Reproducing issue #494 exactly:
constexpr auto val_issue_494 = 8.0 * si::si2019::boltzmann_constant * 1000.0 * K / (std::numbers::pi * 10 * Da);
static_assert(
within_4_ulps(
sqrt(val_issue_494).numerical_value_in(m / s),
sqrt(val_issue_494.numerical_value_in(m * m / s / s))));
///////////////////////
// obtaining a number