#pragma once
#endif
+#include <boost/math/special_functions/math_fwd.hpp>
#include <boost/math/policies/error_handling.hpp>
#include <boost/math/special_functions/fpclassify.hpp>
#include <boost/math/special_functions/sign.hpp>
#include <boost/math/special_functions/trunc.hpp>
-#ifdef BOOST_MSVC
#include <float.h>
+
+#if !defined(_CRAYC) && !defined(__CUDACC__) && (!defined(__GNUC__) || (__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ > 3)))
+#if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || defined(__SSE2__)
+#include "xmmintrin.h"
+#define BOOST_MATH_CHECK_SSE2
+#endif
#endif
namespace boost{ namespace math{
template <class T>
inline T get_smallest_value(mpl::true_ const&)
{
- return std::numeric_limits<T>::denorm_min();
+ //
+ // numeric_limits lies about denorms being present - particularly
+ // when this can be turned on or off at runtime, as is the case
+ // when using the SSE2 registers in DAZ or FTZ mode.
+ //
+ static const T m = std::numeric_limits<T>::denorm_min();
+#ifdef BOOST_MATH_CHECK_SSE2
+ return (_mm_getcsr() & (_MM_FLUSH_ZERO_ON | 0x40)) ? tools::min_value<T>() : m;;
+#else
+ return ((tools::min_value<T>() / 2) == 0) ? tools::min_value<T>() : m;
+#endif
}
template <class T>
#endif
}
+//
+// Returns the smallest value that won't generate denorms when
+// we calculate the value of the least-significant-bit:
+//
+template <class T>
+T get_min_shift_value();
+
+template <class T>
+struct min_shift_initializer
+{
+ struct init
+ {
+ init()
+ {
+ do_init();
+ }
+ static void do_init()
+ {
+ get_min_shift_value<T>();
+ }
+ void force_instantiate()const{}
+ };
+ static const init initializer;
+ static void force_instantiate()
+ {
+ initializer.force_instantiate();
+ }
+};
+
+template <class T>
+const typename min_shift_initializer<T>::init min_shift_initializer<T>::initializer;
+
+
+template <class T>
+inline T get_min_shift_value()
+{
+ BOOST_MATH_STD_USING
+ static const T val = ldexp(tools::min_value<T>(), tools::digits<T>() + 1);
+ min_shift_initializer<T>::force_instantiate();
+
+ return val;
}
template <class T, class Policy>
-T float_next(const T& val, const Policy& pol)
+T float_next_imp(const T& val, const Policy& pol)
{
BOOST_MATH_STD_USING
int expon;
static const char* function = "float_next<%1%>(%1%)";
- if(!(boost::math::isfinite)(val))
+ int fpclass = (boost::math::fpclassify)(val);
+
+ if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
{
if(val < 0)
return -tools::max_value<T>();
if(val == 0)
return detail::get_smallest_value<T>();
+ if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != -tools::min_value<T>()))
+ {
+ //
+ // Special case: if the value of the least significant bit is a denorm, and the result
+ // would not be a denorm, then shift the input, increment, and shift back.
+ // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
+ //
+ return ldexp(float_next(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>());
+ }
+
if(-0.5f == frexp(val, &expon))
--expon; // reduce exponent when val is a power of two, and negative.
T diff = ldexp(T(1), expon - tools::digits<T>());
return val + diff;
}
-#ifdef BOOST_MSVC
+}
+
+template <class T, class Policy>
+inline typename tools::promote_args<T>::type float_next(const T& val, const Policy& pol)
+{
+ typedef typename tools::promote_args<T>::type result_type;
+ return detail::float_next_imp(static_cast<result_type>(val), pol);
+}
+
+#if 0 //def BOOST_MSVC
+//
+// We used to use ::_nextafter here, but doing so fails when using
+// the SSE2 registers if the FTZ or DAZ flags are set, so use our own
+// - albeit slower - code instead as at least that gives the correct answer.
+//
template <class Policy>
inline double float_next(const double& val, const Policy& pol)
{
#endif
template <class T>
-inline T float_next(const T& val)
+inline typename tools::promote_args<T>::type float_next(const T& val)
{
return float_next(val, policies::policy<>());
}
+namespace detail{
+
template <class T, class Policy>
-T float_prior(const T& val, const Policy& pol)
+T float_prior_imp(const T& val, const Policy& pol)
{
BOOST_MATH_STD_USING
int expon;
static const char* function = "float_prior<%1%>(%1%)";
- if(!(boost::math::isfinite)(val))
+ int fpclass = (boost::math::fpclassify)(val);
+
+ if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
{
if(val > 0)
return tools::max_value<T>();
if(val == 0)
return -detail::get_smallest_value<T>();
+ if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != tools::min_value<T>()))
+ {
+ //
+ // Special case: if the value of the least significant bit is a denorm, and the result
+ // would not be a denorm, then shift the input, increment, and shift back.
+ // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
+ //
+ return ldexp(float_prior(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>());
+ }
+
T remain = frexp(val, &expon);
if(remain == 0.5)
--expon; // when val is a power of two we must reduce the exponent
return val - diff;
}
-#ifdef BOOST_MSVC
+}
+
+template <class T, class Policy>
+inline typename tools::promote_args<T>::type float_prior(const T& val, const Policy& pol)
+{
+ typedef typename tools::promote_args<T>::type result_type;
+ return detail::float_prior_imp(static_cast<result_type>(val), pol);
+}
+
+#if 0 //def BOOST_MSVC
+//
+// We used to use ::_nextafter here, but doing so fails when using
+// the SSE2 registers if the FTZ or DAZ flags are set, so use our own
+// - albeit slower - code instead as at least that gives the correct answer.
+//
template <class Policy>
inline double float_prior(const double& val, const Policy& pol)
{
#endif
template <class T>
-inline T float_prior(const T& val)
+inline typename tools::promote_args<T>::type float_prior(const T& val)
{
return float_prior(val, policies::policy<>());
}
-template <class T, class Policy>
-inline T nextafter(const T& val, const T& direction, const Policy& pol)
+template <class T, class U, class Policy>
+inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction, const Policy& pol)
{
- return val < direction ? boost::math::float_next(val, pol) : val == direction ? val : boost::math::float_prior(val, pol);
+ typedef typename tools::promote_args<T, U>::type result_type;
+ return val < direction ? boost::math::float_next<result_type>(val, pol) : val == direction ? val : boost::math::float_prior<result_type>(val, pol);
}
-template <class T>
-inline T nextafter(const T& val, const T& direction)
+template <class T, class U>
+inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction)
{
return nextafter(val, direction, policies::policy<>());
}
+namespace detail{
+
template <class T, class Policy>
-T float_distance(const T& a, const T& b, const Policy& pol)
+T float_distance_imp(const T& a, const T& b, const Policy& pol)
{
BOOST_MATH_STD_USING
//
// Special cases:
//
if(a > b)
- return -float_distance(b, a);
+ return -float_distance(b, a, pol);
if(a == b)
return 0;
if(a == 0)
- return 1 + fabs(float_distance(static_cast<T>(boost::math::sign(b) * detail::get_smallest_value<T>()), b, pol));
+ return 1 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol));
if(b == 0)
- return 1 + fabs(float_distance(static_cast<T>(boost::math::sign(a) * detail::get_smallest_value<T>()), a, pol));
+ return 1 + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol));
if(boost::math::sign(a) != boost::math::sign(b))
- return 2 + fabs(float_distance(static_cast<T>(boost::math::sign(b) * detail::get_smallest_value<T>()), b, pol))
- + fabs(float_distance(static_cast<T>(boost::math::sign(a) * detail::get_smallest_value<T>()), a, pol));
+ return 2 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol))
+ + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol));
//
// By the time we get here, both a and b must have the same sign, we want
// b > a and both postive for the following logic:
//
if(a < 0)
- return float_distance(static_cast<T>(-b), static_cast<T>(-a));
+ return float_distance(static_cast<T>(-b), static_cast<T>(-a), pol);
BOOST_ASSERT(a >= 0);
BOOST_ASSERT(b >= a);
// because we actually have fewer than tools::digits<T>()
// significant bits in the representation:
//
- frexp(((boost::math::fpclassify)(a) == FP_SUBNORMAL) ? tools::min_value<T>() : a, &expon);
+ frexp(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) ? tools::min_value<T>() : a, &expon);
T upper = ldexp(T(1), expon);
T result = 0;
expon = tools::digits<T>() - expon;
result = float_distance(upper, b);
}
//
- // Use compensated double-double addition to avoid rounding
+ // Use compensated double-double addition to avoid rounding
// errors in the subtraction:
//
- T mb = -(std::min)(upper, b);
- T x = a + mb;
- T z = x - a;
- T y = (a - (x - z)) + (mb - z);
+ T mb, x, y, z;
+ if(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) || (b - a < tools::min_value<T>()))
+ {
+ //
+ // Special case - either one end of the range is a denormal, or else the difference is.
+ // The regular code will fail if we're using the SSE2 registers on Intel and either
+ // the FTZ or DAZ flags are set.
+ //
+ T a2 = ldexp(a, tools::digits<T>());
+ T b2 = ldexp(b, tools::digits<T>());
+ mb = -(std::min)(T(ldexp(upper, tools::digits<T>())), b2);
+ x = a2 + mb;
+ z = x - a2;
+ y = (a2 - (x - z)) + (mb - z);
+
+ expon -= tools::digits<T>();
+ }
+ else
+ {
+ mb = -(std::min)(upper, b);
+ x = a + mb;
+ z = x - a;
+ y = (a - (x - z)) + (mb - z);
+ }
if(x < 0)
{
x = -x;
return result;
}
-template <class T>
-T float_distance(const T& a, const T& b)
+}
+
+template <class T, class U, class Policy>
+inline typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b, const Policy& pol)
+{
+ typedef typename tools::promote_args<T, U>::type result_type;
+ return detail::float_distance_imp(static_cast<result_type>(a), static_cast<result_type>(b), pol);
+}
+
+template <class T, class U>
+typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b)
{
return boost::math::float_distance(a, b, policies::policy<>());
}
+namespace detail{
+
template <class T, class Policy>
-T float_advance(T val, int distance, const Policy& pol)
+T float_advance_imp(T val, int distance, const Policy& pol)
{
+ BOOST_MATH_STD_USING
//
// Error handling:
//
static const char* function = "float_advance<%1%>(%1%, int)";
- if(!(boost::math::isfinite)(val))
+
+ int fpclass = (boost::math::fpclassify)(val);
+
+ if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
return policies::raise_domain_error<T>(
function,
"Argument val must be finite, but got %1%", val, pol);
return float_next(val, pol);
if(distance == -1)
return float_prior(val, pol);
- BOOST_MATH_STD_USING
+
+ if(fabs(val) < detail::get_min_shift_value<T>())
+ {
+ //
+ // Special case: if the value of the least significant bit is a denorm,
+ // implement in terms of float_next/float_prior.
+ // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
+ //
+ if(distance > 0)
+ {
+ do{ val = float_next(val, pol); } while(--distance);
+ }
+ else
+ {
+ do{ val = float_prior(val, pol); } while(++distance);
+ }
+ return val;
+ }
+
int expon;
frexp(val, &expon);
T limit = ldexp((distance < 0 ? T(0.5f) : T(1)), expon);
{
distance -= itrunc(limit_distance);
val = limit;
- if(distance < 0)
+ if(distance < 0)
{
limit /= 2;
expon--;
expon++;
}
limit_distance = float_distance(val, limit);
+ if(distance && (limit_distance == 0))
+ {
+ return policies::raise_evaluation_error<T>(function, "Internal logic failed while trying to increment floating point value %1%: most likely your FPU is in non-IEEE conforming mode.", val, pol);
+ }
}
if((0.5f == frexp(val, &expon)) && (distance < 0))
--expon;
return val += diff;
}
+}
+
+template <class T, class Policy>
+inline typename tools::promote_args<T>::type float_advance(T val, int distance, const Policy& pol)
+{
+ typedef typename tools::promote_args<T>::type result_type;
+ return detail::float_advance_imp(static_cast<result_type>(val), distance, pol);
+}
+
template <class T>
-inline T float_advance(const T& val, int distance)
+inline typename tools::promote_args<T>::type float_advance(const T& val, int distance)
{
return boost::math::float_advance(val, distance, policies::policy<>());
}