From 697755e6e412b3a041df87718bb5e79d441be910 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 11:15:34 -0400
Subject: [PATCH 01/35] Add NVRTC support to precision tools

---
 include/boost/math/tools/precision.hpp | 159 ++++++++++---------------
 1 file changed, 65 insertions(+), 94 deletions(-)

diff --git a/include/boost/math/tools/precision.hpp b/include/boost/math/tools/precision.hpp
index b181a2c4a..0055d9028 100644
--- a/include/boost/math/tools/precision.hpp
+++ b/include/boost/math/tools/precision.hpp
@@ -15,12 +15,15 @@
 #include <boost/math/tools/type_traits.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
 #include <boost/math/policies/policy.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
 #include <type_traits>
 #include <limits>
 #include <climits>
 #include <cmath>
 #include <cstdint>
 #include <cfloat> // LDBL_MANT_DIG
+#endif
 
 namespace boost{ namespace math
 {
@@ -41,62 +44,30 @@ namespace tools
 template <class T>
 inline constexpr int digits(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(T)) noexcept
 {
-   static_assert( ::std::numeric_limits<T>::is_specialized, "Type T must be specialized");
-   static_assert( ::std::numeric_limits<T>::radix == 2 || ::std::numeric_limits<T>::radix == 10, "Type T must have a radix of 2 or 10");
+   static_assert( ::boost::math::numeric_limits<T>::is_specialized, "Type T must be specialized");
+   static_assert( ::boost::math::numeric_limits<T>::radix == 2 || ::boost::math::numeric_limits<T>::radix == 10, "Type T must have a radix of 2 or 10");
 
-   return std::numeric_limits<T>::radix == 2
-      ? std::numeric_limits<T>::digits
-      : ((std::numeric_limits<T>::digits + 1) * 1000L) / 301L;
+   return boost::math::numeric_limits<T>::radix == 2
+      ? boost::math::numeric_limits<T>::digits
+      : ((boost::math::numeric_limits<T>::digits + 1) * 1000L) / 301L;
 }
 
 template <class T>
-inline constexpr T max_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T))  noexcept(std::is_floating_point<T>::value)
+inline constexpr T max_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T))  noexcept(boost::math::is_floating_point<T>::value)
 {
-   static_assert( ::std::numeric_limits<T>::is_specialized, "Type T must be specialized");
-   return (std::numeric_limits<T>::max)();
+   static_assert( ::boost::math::numeric_limits<T>::is_specialized, "Type T must be specialized");
+   return (boost::math::numeric_limits<T>::max)();
 } // Also used as a finite 'infinite' value for - and +infinity, for example:
 // -max_value<double> = -1.79769e+308, max_value<double> = 1.79769e+308.
 
-#ifdef BOOST_MATH_HAS_GPU_SUPPORT
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr float max_value<float>() noexcept
-{
-   return FLT_MAX;
-}
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr double max_value<double>() noexcept
-{
-   return DBL_MAX;
-}
-
-#endif
-
 template <class T>
-inline constexpr T min_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+inline constexpr T min_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
-   static_assert( ::std::numeric_limits<T>::is_specialized, "Type T must be specialized");
-
-   return (std::numeric_limits<T>::min)();
-}
+   static_assert( ::boost::math::numeric_limits<T>::is_specialized, "Type T must be specialized");
 
-#ifdef BOOST_MATH_HAS_GPU_SUPPORT
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr float min_value<float>() noexcept
-{
-   return FLT_MIN;
+   return (boost::math::numeric_limits<T>::min)();
 }
 
-template <>
-BOOST_MATH_GPU_ENABLED constexpr double min_value<double>() noexcept
-{
-   return DBL_MIN;
-}
-
-#endif
-
 namespace detail{
 //
 // Logarithmic limits come next, note that although
@@ -107,13 +78,13 @@ namespace detail{
 // For type float first:
 //
 template <class T>
-BOOST_MATH_GPU_ENABLED constexpr T log_max_value(const std::integral_constant<int, 128>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED constexpr T log_max_value(const boost::math::integral_constant<int, 128>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    return 88.0f;
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED constexpr T log_min_value(const std::integral_constant<int, 128>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED constexpr T log_min_value(const boost::math::integral_constant<int, 128>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    return -87.0f;
 }
@@ -121,13 +92,13 @@ BOOST_MATH_GPU_ENABLED constexpr T log_min_value(const std::integral_constant<in
 // Now double:
 //
 template <class T>
-BOOST_MATH_GPU_ENABLED constexpr T log_max_value(const std::integral_constant<int, 1024>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED constexpr T log_max_value(const boost::math::integral_constant<int, 1024>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    return 709.0;
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED constexpr T log_min_value(const std::integral_constant<int, 1024>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED constexpr T log_min_value(const boost::math::integral_constant<int, 1024>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    return -708.0;
 }
@@ -135,19 +106,19 @@ BOOST_MATH_GPU_ENABLED constexpr T log_min_value(const std::integral_constant<in
 // 80 and 128-bit long doubles:
 //
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T log_max_value(const std::integral_constant<int, 16384>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T log_max_value(const boost::math::integral_constant<int, 16384>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    return 11356.0L;
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T log_min_value(const std::integral_constant<int, 16384>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T log_min_value(const boost::math::integral_constant<int, 16384>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    return -11355.0L;
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline T log_max_value(const std::integral_constant<int, 0>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
+BOOST_MATH_GPU_ENABLED inline T log_max_value(const boost::math::integral_constant<int, 0>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
 {
    BOOST_MATH_STD_USING
 #ifdef __SUNPRO_CC
@@ -160,7 +131,7 @@ BOOST_MATH_GPU_ENABLED inline T log_max_value(const std::integral_constant<int,
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline T log_min_value(const std::integral_constant<int, 0>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
+BOOST_MATH_GPU_ENABLED inline T log_min_value(const boost::math::integral_constant<int, 0>& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
 {
    BOOST_MATH_STD_USING
 #ifdef __SUNPRO_CC
@@ -173,21 +144,21 @@ BOOST_MATH_GPU_ENABLED inline T log_min_value(const std::integral_constant<int,
 }
 
 template <class T>
-constexpr T epsilon(const std::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(std::is_floating_point<T>::value)
+constexpr T epsilon(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
-   return std::numeric_limits<T>::epsilon();
+   return boost::math::numeric_limits<T>::epsilon();
 }
 
 #ifdef BOOST_MATH_HAS_GPU_SUPPORT
 
 template <>
-BOOST_MATH_GPU_ENABLED constexpr float epsilon(const std::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(float)) noexcept(true)
+BOOST_MATH_GPU_ENABLED constexpr float epsilon(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(float)) noexcept(true)
 {
    return FLT_EPSILON;
 }
 
 template <>
-BOOST_MATH_GPU_ENABLED constexpr double epsilon(const std::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(double)) noexcept(true)
+BOOST_MATH_GPU_ENABLED constexpr double epsilon(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(double)) noexcept(true)
 {
    return DBL_EPSILON;
 }
@@ -196,7 +167,7 @@ BOOST_MATH_GPU_ENABLED constexpr double epsilon(const std::true_type& BOOST_MATH
 
 #if defined(__GNUC__) && ((LDBL_MANT_DIG == 106) || (__LDBL_MANT_DIG__ == 106))
 template <>
-inline constexpr long double epsilon<long double>(const std::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(long double)) noexcept(std::is_floating_point<long double>::value)
+inline constexpr long double epsilon<long double>(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(long double)) noexcept(boost::math::is_floating_point<long double>::value)
 {
    // numeric_limits on Darwin (and elsewhere) tells lies here:
    // the issue is that long double on a few platforms is
@@ -215,7 +186,7 @@ inline constexpr long double epsilon<long double>(const std::true_type& BOOST_MA
 #endif
 
 template <class T>
-inline T epsilon(const std::false_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
+inline T epsilon(const boost::math::false_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
 {
    // Note: don't cache result as precision may vary at runtime:
    BOOST_MATH_STD_USING  // for ADL of std names
@@ -225,23 +196,23 @@ inline T epsilon(const std::false_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE
 template <class T>
 struct log_limit_traits
 {
-   typedef typename std::conditional<
-      (std::numeric_limits<T>::radix == 2) &&
-      (std::numeric_limits<T>::max_exponent == 128
-         || std::numeric_limits<T>::max_exponent == 1024
-         || std::numeric_limits<T>::max_exponent == 16384),
-      std::integral_constant<int, (std::numeric_limits<T>::max_exponent > INT_MAX ? INT_MAX : static_cast<int>(std::numeric_limits<T>::max_exponent))>,
-      std::integral_constant<int, 0>
+   typedef typename boost::math::conditional<
+      (boost::math::numeric_limits<T>::radix == 2) &&
+      (boost::math::numeric_limits<T>::max_exponent == 128
+         || boost::math::numeric_limits<T>::max_exponent == 1024
+         || boost::math::numeric_limits<T>::max_exponent == 16384),
+      boost::math::integral_constant<int, (boost::math::numeric_limits<T>::max_exponent > (boost::math::numeric_limits<int>::max)() ? (boost::math::numeric_limits<int>::max)() : static_cast<int>(boost::math::numeric_limits<T>::max_exponent))>,
+      boost::math::integral_constant<int, 0>
    >::type tag_type;
    static constexpr bool value = (tag_type::value != 0);
-   static_assert(::std::numeric_limits<T>::is_specialized || !value, "Type T must be specialized or equal to 0");
+   static_assert(::boost::math::numeric_limits<T>::is_specialized || !value, "Type T must be specialized or equal to 0");
 };
 
 template <class T, bool b> struct log_limit_noexcept_traits_imp : public log_limit_traits<T> {};
-template <class T> struct log_limit_noexcept_traits_imp<T, false> : public std::integral_constant<bool, false> {};
+template <class T> struct log_limit_noexcept_traits_imp<T, false> : public boost::math::integral_constant<bool, false> {};
 
 template <class T>
-struct log_limit_noexcept_traits : public log_limit_noexcept_traits_imp<T, std::is_floating_point<T>::value> {};
+struct log_limit_noexcept_traits : public log_limit_noexcept_traits_imp<T, boost::math::is_floating_point<T>::value> {};
 
 } // namespace detail
 
@@ -256,9 +227,9 @@ BOOST_MATH_GPU_ENABLED inline constexpr T log_max_value(BOOST_MATH_EXPLICIT_TEMP
 #ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
    return detail::log_max_value<T>(typename detail::log_limit_traits<T>::tag_type());
 #else
-   BOOST_MATH_ASSERT(::std::numeric_limits<T>::is_specialized);
+   BOOST_MATH_ASSERT(::boost::math::numeric_limits<T>::is_specialized);
    BOOST_MATH_STD_USING
-   static const T val = log((std::numeric_limits<T>::max)());
+   static const T val = log((boost::math::numeric_limits<T>::max)());
    return val;
 #endif
 }
@@ -269,9 +240,9 @@ BOOST_MATH_GPU_ENABLED inline constexpr T log_min_value(BOOST_MATH_EXPLICIT_TEMP
 #ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
    return detail::log_min_value<T>(typename detail::log_limit_traits<T>::tag_type());
 #else
-   BOOST_MATH_ASSERT(::std::numeric_limits<T>::is_specialized);
+   BOOST_MATH_ASSERT(::boost::math::numeric_limits<T>::is_specialized);
    BOOST_MATH_STD_USING
-   static const T val = log((std::numeric_limits<T>::min)());
+   static const T val = log((boost::math::numeric_limits<T>::min)());
    return val;
 #endif
 }
@@ -284,11 +255,11 @@ template <class T>
 BOOST_MATH_GPU_ENABLED constexpr T epsilon(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
 #ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
-   return detail::epsilon<T>(std::integral_constant<bool, ::std::numeric_limits<T>::is_specialized>());
+   return detail::epsilon<T>(boost::math::integral_constant<bool, ::boost::math::numeric_limits<T>::is_specialized>());
 #else
-   return ::std::numeric_limits<T>::is_specialized ?
-      detail::epsilon<T>(std::true_type()) :
-      detail::epsilon<T>(std::false_type());
+   return ::boost::math::numeric_limits<T>::is_specialized ?
+      detail::epsilon<T>(boost::math::true_type()) :
+      detail::epsilon<T>(boost::math::false_type());
 #endif
 }
 
@@ -318,25 +289,25 @@ BOOST_MATH_GPU_ENABLED constexpr long double epsilon(BOOST_MATH_EXPLICIT_TEMPLAT
 namespace detail{
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon_imp(const std::integral_constant<int, 24>&) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon_imp(const boost::math::integral_constant<int, 24>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.00034526698300124390839884978618400831996329879769945L);
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon_imp(const T*, const std::integral_constant<int, 53>&) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 53>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.1490116119384765625e-7L);
 }
 
 template <class T>
-inline constexpr T root_epsilon_imp(const T*, const std::integral_constant<int, 64>&) noexcept(std::is_floating_point<T>::value)
+inline constexpr T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.32927225399135962333569506281281311031656150598474e-9L);
 }
 
 template <class T>
-inline constexpr T root_epsilon_imp(const T*, const std::integral_constant<int, 113>&) noexcept(std::is_floating_point<T>::value)
+inline constexpr T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.1387778780781445675529539585113525390625e-16L);
 }
@@ -350,32 +321,32 @@ inline T root_epsilon_imp(const T*, const Tag&)
 }
 
 template <class T>
-inline T root_epsilon_imp(const T*, const std::integral_constant<int, 0>&)
+inline T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
 {
    BOOST_MATH_STD_USING
    return sqrt(tools::epsilon<T>());
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon_imp(const std::integral_constant<int, 24>&) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon_imp(const boost::math::integral_constant<int, 24>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.0049215666011518482998719164346805794944150447839903L);
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon_imp(const T*, const std::integral_constant<int, 53>&) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 53>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(6.05545445239333906078989272793696693569753008995e-6L);
 }
 
 template <class T>
-inline constexpr T cbrt_epsilon_imp(const T*, const std::integral_constant<int, 64>&) noexcept(std::is_floating_point<T>::value)
+inline constexpr T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(4.76837158203125e-7L);
 }
 
 template <class T>
-inline constexpr T cbrt_epsilon_imp(const T*, const std::integral_constant<int, 113>&) noexcept(std::is_floating_point<T>::value)
+inline constexpr T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(5.7749313854154005630396773604745549542403508090496e-12L);
 }
@@ -389,32 +360,32 @@ inline T cbrt_epsilon_imp(const T*, const Tag&)
 }
 
 template <class T>
-inline T cbrt_epsilon_imp(const T*, const std::integral_constant<int, 0>&)
+inline T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
 {
    BOOST_MATH_STD_USING;
    return pow(tools::epsilon<T>(), T(1) / 3);
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon_imp(const T*, const std::integral_constant<int, 24>&) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 24>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.018581361171917516667460937040007436176452688944747L);
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon_imp(const T*, const std::integral_constant<int, 53>&) noexcept(std::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 53>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.0001220703125L);
 }
 
 template <class T>
-inline constexpr T forth_root_epsilon_imp(const T*, const std::integral_constant<int, 64>&) noexcept(std::is_floating_point<T>::value)
+inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.18145860519450699870567321328132261891067079047605e-4L);
 }
 
 template <class T>
-inline constexpr T forth_root_epsilon_imp(const T*, const std::integral_constant<int, 113>&) noexcept(std::is_floating_point<T>::value)
+inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.37252902984619140625e-8L);
 }
@@ -428,7 +399,7 @@ inline T forth_root_epsilon_imp(const T*, const Tag&)
 }
 
 template <class T>
-inline T forth_root_epsilon_imp(const T*, const std::integral_constant<int, 0>&)
+inline T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
 {
    BOOST_MATH_STD_USING
    return sqrt(sqrt(tools::epsilon<T>()));
@@ -437,26 +408,26 @@ inline T forth_root_epsilon_imp(const T*, const std::integral_constant<int, 0>&)
 template <class T>
 struct root_epsilon_traits
 {
-   typedef std::integral_constant<int, (::std::numeric_limits<T>::radix == 2) && (::std::numeric_limits<T>::digits != INT_MAX) ? std::numeric_limits<T>::digits : 0> tag_type;
+   typedef boost::math::integral_constant<int, (::boost::math::numeric_limits<T>::radix == 2) && (::boost::math::numeric_limits<T>::digits != (boost::math::numeric_limits<int>::max)()) ? boost::math::numeric_limits<T>::digits : 0> tag_type;
    static constexpr bool has_noexcept = (tag_type::value == 113) || (tag_type::value == 64) || (tag_type::value == 53) || (tag_type::value == 24);
 };
 
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon() noexcept(std::is_floating_point<T>::value && detail::root_epsilon_traits<T>::has_noexcept)
+BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon() noexcept(boost::math::is_floating_point<T>::value && detail::root_epsilon_traits<T>::has_noexcept)
 {
    return detail::root_epsilon_imp(static_cast<T const*>(nullptr), typename detail::root_epsilon_traits<T>::tag_type());
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon() noexcept(std::is_floating_point<T>::value && detail::root_epsilon_traits<T>::has_noexcept)
+BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon() noexcept(boost::math::is_floating_point<T>::value && detail::root_epsilon_traits<T>::has_noexcept)
 {
    return detail::cbrt_epsilon_imp(static_cast<T const*>(nullptr), typename detail::root_epsilon_traits<T>::tag_type());
 }
 
 template <class T>
-BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon() noexcept(std::is_floating_point<T>::value && detail::root_epsilon_traits<T>::has_noexcept)
+BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon() noexcept(boost::math::is_floating_point<T>::value && detail::root_epsilon_traits<T>::has_noexcept)
 {
    return detail::forth_root_epsilon_imp(static_cast<T const*>(nullptr), typename detail::root_epsilon_traits<T>::tag_type());
 }

From 4aeb9c1b8c8846bddb915bb602d5a8a39bba935e Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 11:51:23 -0400
Subject: [PATCH 02/35] Fix missing header in error handling

---
 include/boost/math/policies/error_handling.hpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/boost/math/policies/error_handling.hpp b/include/boost/math/policies/error_handling.hpp
index d1d74c534..07e112721 100644
--- a/include/boost/math/policies/error_handling.hpp
+++ b/include/boost/math/policies/error_handling.hpp
@@ -10,6 +10,7 @@
 
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
+#include <boost/math/tools/type_traits.hpp>
 
 #ifndef BOOST_MATH_HAS_NVRTC
 

From 14694c42264ea49b2b32b7e0b2dca44acedc4d24 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 11:52:00 -0400
Subject: [PATCH 03/35] Add minimal implantation of pair for NVRTC platform

---
 include/boost/math/tools/tuple.hpp | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

diff --git a/include/boost/math/tools/tuple.hpp b/include/boost/math/tools/tuple.hpp
index f132b97f5..fba37fb4f 100644
--- a/include/boost/math/tools/tuple.hpp
+++ b/include/boost/math/tools/tuple.hpp
@@ -9,7 +9,32 @@
 
 #include <boost/math/tools/config.hpp>
 
-#ifdef BOOST_MATH_ENABLE_CUDA
+#ifdef BOOST_MATH_HAS_NVRTC
+
+namespace boost {
+namespace math {
+
+template <typename T, typename U>
+class pair
+{
+public:
+    T first;
+    U second;
+
+    BOOST_MATH_GPU_ENABLED pair(T x, U y) noexcept : first {x}, second {y} {}
+};
+
+template <typename T, typename U>
+BOOST_MATH_GPU_ENABLED auto make_pair(T x, U y) noexcept
+{
+    return pair(x, y);
+}
+
+} // namespace math
+} // namespace boost
+
+
+#elif defined(BOOST_MATH_ENABLE_CUDA)
 
 #include <thrust/pair.h>
 #include <thrust/tuple.h>

From ec5b2a96df35f89de0f507207609d9f2dfe66e43 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 11:52:15 -0400
Subject: [PATCH 04/35] Add NVRTC support to derived accessors

---
 .../boost/math/distributions/detail/derived_accessors.hpp  | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/include/boost/math/distributions/detail/derived_accessors.hpp b/include/boost/math/distributions/detail/derived_accessors.hpp
index 9d39f2d59..90679ef21 100644
--- a/include/boost/math/distributions/detail/derived_accessors.hpp
+++ b/include/boost/math/distributions/detail/derived_accessors.hpp
@@ -28,9 +28,12 @@
 // can find the definitions referred to herein.
 //
 
-#include <cmath>
-#include <boost/math/tools/assert.hpp>
 #include <boost/math/tools/config.hpp>
+#include <boost/math/tools/assert.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <cmath>
+#endif
 
 #ifdef _MSC_VER
 # pragma warning(push)

From a6d14dd6e2dcf134f51702cb6bd8ee33c0befc81 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 11:52:32 -0400
Subject: [PATCH 05/35] Add NVRTC support to the arcsine dist

---
 include/boost/math/distributions/arcsine.hpp | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

diff --git a/include/boost/math/distributions/arcsine.hpp b/include/boost/math/distributions/arcsine.hpp
index d921e1145..899bfb1b2 100644
--- a/include/boost/math/distributions/arcsine.hpp
+++ b/include/boost/math/distributions/arcsine.hpp
@@ -30,15 +30,21 @@
 #ifndef BOOST_MATH_DIST_ARCSINE_HPP
 #define BOOST_MATH_DIST_ARCSINE_HPP
 
-#include <cmath>
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/tuple.hpp>
-#include <boost/math/distributions/fwd.hpp>
 #include <boost/math/distributions/complement.hpp> // complements.
 #include <boost/math/distributions/detail/common_error_handling.hpp> // error checks.
 #include <boost/math/constants/constants.hpp>
-
 #include <boost/math/special_functions/fpclassify.hpp> // isnan.
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
+#include <cmath>
+#include <utility>
+#include <exception>  // For std::domain_error.
+#endif
 
 #if defined (BOOST_MSVC)
 #  pragma warning(push)
@@ -46,9 +52,6 @@
 // in domain_error_imp in error_handling.
 #endif
 
-#include <utility>
-#include <exception>  // For std::domain_error.
-
 namespace boost
 {
   namespace math

From 2356c73759d000a1baff1fbbefa6b7c6876522b9 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 12:01:23 -0400
Subject: [PATCH 06/35] Undef BOOST_FPU_EXCEPTION_GUARD if exists and give a
 new empty def

---
 include/boost/math/tools/config.hpp | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/include/boost/math/tools/config.hpp b/include/boost/math/tools/config.hpp
index 4c8e3dcf3..d886e420d 100644
--- a/include/boost/math/tools/config.hpp
+++ b/include/boost/math/tools/config.hpp
@@ -799,6 +799,13 @@ BOOST_MATH_GPU_ENABLED constexpr T cuda_safe_max(const T& a, const T& b) { retur
 #define BOOST_MATH_FORCEINLINE __forceinline__
 #define BOOST_MATH_STD_USING  
 
+// This should be defined to nothing but since it is not specifically a math macro
+// we need to undef before proceeding
+#ifdef BOOST_FPU_EXCEPTION_GUARD
+#  undef BOOST_FPU_EXCEPTION_GUARD
+#  define BOOST_FPU_EXCEPTION_GUARD
+#endif
+
 template <class T>
 BOOST_MATH_GPU_ENABLED constexpr void gpu_safe_swap(T& a, T& b) { T t(a); a = b; b = t; }
 

From 791e9af4f3ca563bbffdc6d6d14d81f235dde554 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 12:01:34 -0400
Subject: [PATCH 07/35] Add testing of the arcsine dist

---
 test/nvrtc_jamfile                          |   8 +
 test/test_arcsine_cdf_nvrtc_double.cpp      | 191 ++++++++++++++++++++
 test/test_arcsine_cdf_nvrtc_float.cpp       | 191 ++++++++++++++++++++
 test/test_arcsine_pdf_nvrtc_double.cpp      | 191 ++++++++++++++++++++
 test/test_arcsine_pdf_nvrtc_float.cpp       | 191 ++++++++++++++++++++
 test/test_arcsine_quantile_nvrtc_double.cpp | 191 ++++++++++++++++++++
 test/test_arcsine_quantile_nvrtc_float.cpp  | 191 ++++++++++++++++++++
 7 files changed, 1154 insertions(+)
 create mode 100644 test/test_arcsine_cdf_nvrtc_double.cpp
 create mode 100644 test/test_arcsine_cdf_nvrtc_float.cpp
 create mode 100644 test/test_arcsine_pdf_nvrtc_double.cpp
 create mode 100644 test/test_arcsine_pdf_nvrtc_float.cpp
 create mode 100644 test/test_arcsine_quantile_nvrtc_double.cpp
 create mode 100644 test/test_arcsine_quantile_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 434ca002b..325dcfb14 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -9,6 +9,14 @@ project : requirements
     [ requires cxx14_decltype_auto cxx14_generic_lambdas cxx14_return_type_deduction cxx14_variable_templates cxx14_constexpr ]
     ;
 
+# Distributions
+run test_arcsine_cdf_nvrtc_double.cpp ;
+run test_arcsine_cdf_nvrtc_float.cpp ;
+run test_arcsine_pdf_nvrtc_double.cpp ;
+run test_arcsine_pdf_nvrtc_float.cpp ;
+run test_arcsine_quantile_nvrtc_double.cpp ;
+run test_arcsine_quantile_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_arcsine_cdf_nvrtc_double.cpp b/test/test_arcsine_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..814d7491a
--- /dev/null
+++ b/test/test_arcsine_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/arcsine.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/arcsine.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::arcsine_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::arcsine_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_arcsine_cdf_nvrtc_float.cpp b/test/test_arcsine_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..a159e608c
--- /dev/null
+++ b/test/test_arcsine_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/arcsine.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/arcsine.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::arcsine_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::arcsine_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_arcsine_pdf_nvrtc_double.cpp b/test/test_arcsine_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..dde875629
--- /dev/null
+++ b/test/test_arcsine_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/arcsine.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/arcsine.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::arcsine_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::arcsine_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_arcsine_pdf_nvrtc_float.cpp b/test/test_arcsine_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..82b39c3cf
--- /dev/null
+++ b/test/test_arcsine_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/arcsine.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/arcsine.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::arcsine_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::arcsine_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_arcsine_quantile_nvrtc_double.cpp b/test/test_arcsine_quantile_nvrtc_double.cpp
new file mode 100644
index 000000000..8a2d39ed5
--- /dev/null
+++ b/test/test_arcsine_quantile_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/arcsine.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/arcsine.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::arcsine_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::arcsine_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_arcsine_quantile_nvrtc_float.cpp b/test/test_arcsine_quantile_nvrtc_float.cpp
new file mode 100644
index 000000000..7491b7e11
--- /dev/null
+++ b/test/test_arcsine_quantile_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/arcsine.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/arcsine.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::arcsine_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::arcsine_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 59c1dd4d8aa0122e5f5f8b5bea96955c3b7d08c3 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 13:42:13 -0400
Subject: [PATCH 08/35] Fix FPU Guard macro

---
 include/boost/math/tools/config.hpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/boost/math/tools/config.hpp b/include/boost/math/tools/config.hpp
index d886e420d..a5c8b3fc9 100644
--- a/include/boost/math/tools/config.hpp
+++ b/include/boost/math/tools/config.hpp
@@ -803,9 +803,10 @@ BOOST_MATH_GPU_ENABLED constexpr T cuda_safe_max(const T& a, const T& b) { retur
 // we need to undef before proceeding
 #ifdef BOOST_FPU_EXCEPTION_GUARD
 #  undef BOOST_FPU_EXCEPTION_GUARD
-#  define BOOST_FPU_EXCEPTION_GUARD
 #endif
 
+#define BOOST_FPU_EXCEPTION_GUARD
+
 template <class T>
 BOOST_MATH_GPU_ENABLED constexpr void gpu_safe_swap(T& a, T& b) { T t(a); a = b; b = t; }
 

From dfcda439be3a289ffd2d158ec06badb98422380a Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 13:42:30 -0400
Subject: [PATCH 09/35] Add NVRTC support to bernoulli dist

---
 include/boost/math/distributions/bernoulli.hpp | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/include/boost/math/distributions/bernoulli.hpp b/include/boost/math/distributions/bernoulli.hpp
index 639f9119c..f1c693f7f 100644
--- a/include/boost/math/distributions/bernoulli.hpp
+++ b/include/boost/math/distributions/bernoulli.hpp
@@ -28,15 +28,19 @@
 #ifndef BOOST_MATH_SPECIAL_BERNOULLI_HPP
 #define BOOST_MATH_SPECIAL_BERNOULLI_HPP
 
-#include <boost/math/distributions/fwd.hpp>
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/tuple.hpp>
 #include <boost/math/tools/type_traits.hpp>
 #include <boost/math/distributions/complement.hpp> // complements
 #include <boost/math/distributions/detail/common_error_handling.hpp> // error checks
 #include <boost/math/special_functions/fpclassify.hpp> // isnan.
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
 
+#ifndef BOOST_MATH_HAS_NVRTC
 #include <utility>
+#include <boost/math/distributions/fwd.hpp>
+#endif
 
 namespace boost
 {

From ba549c51349dec9dd09700a3232d73fb0169797a Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 13:42:45 -0400
Subject: [PATCH 10/35] Add bernoulli dist NVRTC testing

---
 test/nvrtc_jamfile                        |   7 +
 test/test_bernoulli_cdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++
 test/test_bernoulli_cdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++
 test/test_bernoulli_pdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++
 test/test_bernoulli_pdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++
 test/test_bernoulli_quan_nvrtc_double.cpp | 191 ++++++++++++++++++++++
 test/test_bernoulli_quan_nvrtc_float.cpp  | 191 ++++++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_bernoulli_cdf_nvrtc_double.cpp
 create mode 100644 test/test_bernoulli_cdf_nvrtc_float.cpp
 create mode 100644 test/test_bernoulli_pdf_nvrtc_double.cpp
 create mode 100644 test/test_bernoulli_pdf_nvrtc_float.cpp
 create mode 100644 test/test_bernoulli_quan_nvrtc_double.cpp
 create mode 100644 test/test_bernoulli_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 325dcfb14..10e7077d1 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -17,6 +17,13 @@ run test_arcsine_pdf_nvrtc_float.cpp ;
 run test_arcsine_quantile_nvrtc_double.cpp ;
 run test_arcsine_quantile_nvrtc_float.cpp ;
 
+run test_bernoulli_cdf_nvrtc_double.cpp ;
+run test_bernoulli_cdf_nvrtc_float.cpp ;
+run test_bernoulli_pdf_nvrtc_double.cpp ;
+run test_bernoulli_pdf_nvrtc_float.cpp ;
+run test_bernoulli_quan_nvrtc_double.cpp ;
+run test_bernoulli_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_bernoulli_cdf_nvrtc_double.cpp b/test/test_bernoulli_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..fc21b9ef6
--- /dev/null
+++ b/test/test_bernoulli_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/bernoulli.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/bernoulli.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::bernoulli_distribution<float_type>(), round(in1[i]));
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::bernoulli_distribution<float_type>(), round(h_in1[i]));
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_bernoulli_cdf_nvrtc_float.cpp b/test/test_bernoulli_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..d5112b6be
--- /dev/null
+++ b/test/test_bernoulli_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/bernoulli.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/bernoulli.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::bernoulli_distribution<float_type>(), round(in1[i]));
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::bernoulli_distribution<float_type>(), round(h_in1[i]));
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_bernoulli_pdf_nvrtc_double.cpp b/test/test_bernoulli_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..f1b4bdf99
--- /dev/null
+++ b/test/test_bernoulli_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/bernoulli.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/bernoulli.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::bernoulli_distribution<float_type>(), round(in1[i]));
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::bernoulli_distribution<float_type>(), round(h_in1[i]));
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_bernoulli_pdf_nvrtc_float.cpp b/test/test_bernoulli_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..245945cde
--- /dev/null
+++ b/test/test_bernoulli_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/bernoulli.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/bernoulli.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::bernoulli_distribution<float_type>(), round(in1[i]));
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::bernoulli_distribution<float_type>(), round(h_in1[i]));
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_bernoulli_quan_nvrtc_double.cpp b/test/test_bernoulli_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..63c8ae7fb
--- /dev/null
+++ b/test/test_bernoulli_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/bernoulli.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/bernoulli.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::bernoulli_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::bernoulli_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_bernoulli_quan_nvrtc_float.cpp b/test/test_bernoulli_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..526438867
--- /dev/null
+++ b/test/test_bernoulli_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/bernoulli.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/bernoulli.hpp>
+extern "C" __global__ 
+void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::bernoulli_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::bernoulli_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From ed65c6e91bc4dbe0bfae624e47e865fa44610f69 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 14:39:14 -0400
Subject: [PATCH 11/35] Fix kernel naming

---
 test/test_arcsine_cdf_nvrtc_double.cpp      | 8 ++++----
 test/test_arcsine_cdf_nvrtc_float.cpp       | 8 ++++----
 test/test_arcsine_pdf_nvrtc_double.cpp      | 8 ++++----
 test/test_arcsine_pdf_nvrtc_float.cpp       | 8 ++++----
 test/test_arcsine_quantile_nvrtc_double.cpp | 8 ++++----
 test/test_arcsine_quantile_nvrtc_float.cpp  | 8 ++++----
 test/test_bernoulli_cdf_nvrtc_double.cpp    | 8 ++++----
 test/test_bernoulli_cdf_nvrtc_float.cpp     | 8 ++++----
 test/test_bernoulli_pdf_nvrtc_double.cpp    | 8 ++++----
 test/test_bernoulli_pdf_nvrtc_float.cpp     | 8 ++++----
 test/test_bernoulli_quan_nvrtc_double.cpp   | 8 ++++----
 test/test_bernoulli_quan_nvrtc_float.cpp    | 8 ++++----
 12 files changed, 48 insertions(+), 48 deletions(-)

diff --git a/test/test_arcsine_cdf_nvrtc_double.cpp b/test/test_arcsine_cdf_nvrtc_double.cpp
index 814d7491a..58b6b8297 100644
--- a/test/test_arcsine_cdf_nvrtc_double.cpp
+++ b/test/test_arcsine_cdf_nvrtc_double.cpp
@@ -28,7 +28,7 @@ typedef double float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/arcsine.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_arcsine_cdf_nvrtc_float.cpp b/test/test_arcsine_cdf_nvrtc_float.cpp
index a159e608c..3f8b04bb0 100644
--- a/test/test_arcsine_cdf_nvrtc_float.cpp
+++ b/test/test_arcsine_cdf_nvrtc_float.cpp
@@ -28,7 +28,7 @@ typedef float float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/arcsine.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_arcsine_pdf_nvrtc_double.cpp b/test/test_arcsine_pdf_nvrtc_double.cpp
index dde875629..c76b47f88 100644
--- a/test/test_arcsine_pdf_nvrtc_double.cpp
+++ b/test/test_arcsine_pdf_nvrtc_double.cpp
@@ -28,7 +28,7 @@ typedef double float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/arcsine.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_arcsine_pdf_nvrtc_float.cpp b/test/test_arcsine_pdf_nvrtc_float.cpp
index 82b39c3cf..e9145a162 100644
--- a/test/test_arcsine_pdf_nvrtc_float.cpp
+++ b/test/test_arcsine_pdf_nvrtc_float.cpp
@@ -28,7 +28,7 @@ typedef float float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/arcsine.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_arcsine_quantile_nvrtc_double.cpp b/test/test_arcsine_quantile_nvrtc_double.cpp
index 8a2d39ed5..ba8e2e5df 100644
--- a/test/test_arcsine_quantile_nvrtc_double.cpp
+++ b/test/test_arcsine_quantile_nvrtc_double.cpp
@@ -28,7 +28,7 @@ typedef double float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/arcsine.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_arcsine_quantile_nvrtc_float.cpp b/test/test_arcsine_quantile_nvrtc_float.cpp
index 7491b7e11..1fd2e4884 100644
--- a/test/test_arcsine_quantile_nvrtc_float.cpp
+++ b/test/test_arcsine_quantile_nvrtc_float.cpp
@@ -28,7 +28,7 @@ typedef float float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/arcsine.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_bernoulli_cdf_nvrtc_double.cpp b/test/test_bernoulli_cdf_nvrtc_double.cpp
index fc21b9ef6..887c3430a 100644
--- a/test/test_bernoulli_cdf_nvrtc_double.cpp
+++ b/test/test_bernoulli_cdf_nvrtc_double.cpp
@@ -28,7 +28,7 @@ typedef double float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/bernoulli.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_bernoulli_cdf_nvrtc_float.cpp b/test/test_bernoulli_cdf_nvrtc_float.cpp
index d5112b6be..1a38c4348 100644
--- a/test/test_bernoulli_cdf_nvrtc_float.cpp
+++ b/test/test_bernoulli_cdf_nvrtc_float.cpp
@@ -28,7 +28,7 @@ typedef float float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/bernoulli.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_bernoulli_pdf_nvrtc_double.cpp b/test/test_bernoulli_pdf_nvrtc_double.cpp
index f1b4bdf99..f3e21a095 100644
--- a/test/test_bernoulli_pdf_nvrtc_double.cpp
+++ b/test/test_bernoulli_pdf_nvrtc_double.cpp
@@ -28,7 +28,7 @@ typedef double float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/bernoulli.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_bernoulli_pdf_nvrtc_float.cpp b/test/test_bernoulli_pdf_nvrtc_float.cpp
index 245945cde..bf9b76016 100644
--- a/test/test_bernoulli_pdf_nvrtc_float.cpp
+++ b/test/test_bernoulli_pdf_nvrtc_float.cpp
@@ -28,7 +28,7 @@ typedef float float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/bernoulli.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_bernoulli_quan_nvrtc_double.cpp b/test/test_bernoulli_quan_nvrtc_double.cpp
index 63c8ae7fb..dcdd0e1f1 100644
--- a/test/test_bernoulli_quan_nvrtc_double.cpp
+++ b/test/test_bernoulli_quan_nvrtc_double.cpp
@@ -28,7 +28,7 @@ typedef double float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/bernoulli.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;
diff --git a/test/test_bernoulli_quan_nvrtc_float.cpp b/test/test_bernoulli_quan_nvrtc_float.cpp
index 526438867..24c492381 100644
--- a/test/test_bernoulli_quan_nvrtc_float.cpp
+++ b/test/test_bernoulli_quan_nvrtc_float.cpp
@@ -28,7 +28,7 @@ typedef float float_type;
 #include <cuda/std/type_traits>
 #include <boost/math/distributions/bernoulli.hpp>
 extern "C" __global__ 
-void test_digamma_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
 {
     int i = blockDim.x * blockIdx.x + threadIdx.x;
     if (i < numElements)
@@ -83,10 +83,10 @@ int main()
         nvrtcProgram prog;
         nvrtcResult res;
 
-        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_digamma_kernel.cu", 0, nullptr, nullptr);
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
         checkNVRTCError(res, "Failed to create NVRTC program");
 
-        nvrtcAddNameExpression(prog, "test_digamma_kernel");
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
 
         #ifdef BOOST_MATH_NVRTC_CI_RUN
         const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
@@ -117,7 +117,7 @@ int main()
         CUmodule module;
         CUfunction kernel;
         checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
-        checkCUError(cuModuleGetFunction(&kernel, module, "test_digamma_kernel"), "Failed to get kernel function");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
 
         int numElements = 5000;
         float_type *h_in1, *h_in2, *h_out;

From 13ba4111de9fca43bee6c247f7935fc52eda0d77 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 14:39:38 -0400
Subject: [PATCH 12/35] Remove workarounds and markup functions that will never
 be taken anyway

---
 include/boost/math/tools/config.hpp    |  1 +
 include/boost/math/tools/precision.hpp | 90 ++++++++------------------
 2 files changed, 29 insertions(+), 62 deletions(-)

diff --git a/include/boost/math/tools/config.hpp b/include/boost/math/tools/config.hpp
index a5c8b3fc9..b67ce0ae9 100644
--- a/include/boost/math/tools/config.hpp
+++ b/include/boost/math/tools/config.hpp
@@ -795,6 +795,7 @@ BOOST_MATH_GPU_ENABLED constexpr T cuda_safe_max(const T& a, const T& b) { retur
 
 #define BOOST_MATH_NOEXCEPT(T) noexcept(boost::math::is_floating_point_v<T>)
 #define BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T) 
+#define BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(T) 
 #define BOOST_MATH_BIG_CONSTANT(T, N, V) static_cast<T>(V)
 #define BOOST_MATH_FORCEINLINE __forceinline__
 #define BOOST_MATH_STD_USING  
diff --git a/include/boost/math/tools/precision.hpp b/include/boost/math/tools/precision.hpp
index 0055d9028..096a3265a 100644
--- a/include/boost/math/tools/precision.hpp
+++ b/include/boost/math/tools/precision.hpp
@@ -42,7 +42,7 @@ namespace tools
 // See  Conceptual Requirements for Real Number Types.
 
 template <class T>
-inline constexpr int digits(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(T)) noexcept
+BOOST_MATH_GPU_ENABLED inline constexpr int digits(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(T)) noexcept
 {
    static_assert( ::boost::math::numeric_limits<T>::is_specialized, "Type T must be specialized");
    static_assert( ::boost::math::numeric_limits<T>::radix == 2 || ::boost::math::numeric_limits<T>::radix == 10, "Type T must have a radix of 2 or 10");
@@ -53,7 +53,7 @@ inline constexpr int digits(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(T)) noexcept
 }
 
 template <class T>
-inline constexpr T max_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T))  noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T max_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T))  noexcept(boost::math::is_floating_point<T>::value)
 {
    static_assert( ::boost::math::numeric_limits<T>::is_specialized, "Type T must be specialized");
    return (boost::math::numeric_limits<T>::max)();
@@ -61,7 +61,7 @@ inline constexpr T max_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T))  noexcept(boo
 // -max_value<double> = -1.79769e+308, max_value<double> = 1.79769e+308.
 
 template <class T>
-inline constexpr T min_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T min_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    static_assert( ::boost::math::numeric_limits<T>::is_specialized, "Type T must be specialized");
 
@@ -144,30 +144,14 @@ BOOST_MATH_GPU_ENABLED inline T log_min_value(const boost::math::integral_consta
 }
 
 template <class T>
-constexpr T epsilon(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED constexpr T epsilon(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
    return boost::math::numeric_limits<T>::epsilon();
 }
 
-#ifdef BOOST_MATH_HAS_GPU_SUPPORT
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr float epsilon(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(float)) noexcept(true)
-{
-   return FLT_EPSILON;
-}
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr double epsilon(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(double)) noexcept(true)
-{
-   return DBL_EPSILON;
-}
-
-#endif
-
 #if defined(__GNUC__) && ((LDBL_MANT_DIG == 106) || (__LDBL_MANT_DIG__ == 106))
 template <>
-inline constexpr long double epsilon<long double>(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(long double)) noexcept(boost::math::is_floating_point<long double>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr long double epsilon<long double>(const boost::math::true_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(long double)) noexcept(boost::math::is_floating_point<long double>::value)
 {
    // numeric_limits on Darwin (and elsewhere) tells lies here:
    // the issue is that long double on a few platforms is
@@ -186,7 +170,7 @@ inline constexpr long double epsilon<long double>(const boost::math::true_type&
 #endif
 
 template <class T>
-inline T epsilon(const boost::math::false_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
+BOOST_MATH_GPU_ENABLED inline T epsilon(const boost::math::false_type& BOOST_MATH_APPEND_EXPLICIT_TEMPLATE_TYPE(T))
 {
    // Note: don't cache result as precision may vary at runtime:
    BOOST_MATH_STD_USING  // for ADL of std names
@@ -254,38 +238,20 @@ BOOST_MATH_GPU_ENABLED inline constexpr T log_min_value(BOOST_MATH_EXPLICIT_TEMP
 template <class T>
 BOOST_MATH_GPU_ENABLED constexpr T epsilon(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(T)) noexcept(boost::math::is_floating_point<T>::value)
 {
-#ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
-   return detail::epsilon<T>(boost::math::integral_constant<bool, ::boost::math::numeric_limits<T>::is_specialized>());
+   // NVRTC does not like this dispatching method so we just skip to where we want to go
+#ifndef BOOST_MATH_HAS_NVRTC
+   #ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
+      return detail::epsilon<T>(boost::math::integral_constant<bool, ::boost::math::numeric_limits<T>::is_specialized>());
+   #else
+      return ::boost::math::numeric_limits<T>::is_specialized ?
+         detail::epsilon<T>(boost::math::true_type()) :
+         detail::epsilon<T>(boost::math::false_type());
+   #endif
 #else
-   return ::boost::math::numeric_limits<T>::is_specialized ?
-      detail::epsilon<T>(boost::math::true_type()) :
-      detail::epsilon<T>(boost::math::false_type());
+   return boost::math::numeric_limits<T>::epsilon();
 #endif
 }
 
-#ifdef BOOST_MATH_HAS_GPU_SUPPORT
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr float epsilon(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(float)) noexcept(true)
-{
-   return FLT_EPSILON;
-}
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr double epsilon(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(double)) noexcept(true)
-{
-   return DBL_EPSILON;
-}
-
-template <>
-BOOST_MATH_GPU_ENABLED constexpr long double epsilon(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE_SPEC(long double)) noexcept(true)
-{
-   return LDBL_EPSILON;
-}
-
-
-#endif
-
 namespace detail{
 
 template <class T>
@@ -301,19 +267,19 @@ BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon_imp(const T*, const boost
 }
 
 template <class T>
-inline constexpr T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.32927225399135962333569506281281311031656150598474e-9L);
 }
 
 template <class T>
-inline constexpr T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.1387778780781445675529539585113525390625e-16L);
 }
 
 template <class T, class Tag>
-inline T root_epsilon_imp(const T*, const Tag&)
+BOOST_MATH_GPU_ENABLED inline T root_epsilon_imp(const T*, const Tag&)
 {
    BOOST_MATH_STD_USING
    static const T r_eps = sqrt(tools::epsilon<T>());
@@ -321,7 +287,7 @@ inline T root_epsilon_imp(const T*, const Tag&)
 }
 
 template <class T>
-inline T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
+BOOST_MATH_GPU_ENABLED inline T root_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
 {
    BOOST_MATH_STD_USING
    return sqrt(tools::epsilon<T>());
@@ -340,19 +306,19 @@ BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon_imp(const T*, const boost
 }
 
 template <class T>
-inline constexpr T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(4.76837158203125e-7L);
 }
 
 template <class T>
-inline constexpr T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(5.7749313854154005630396773604745549542403508090496e-12L);
 }
 
 template <class T, class Tag>
-inline T cbrt_epsilon_imp(const T*, const Tag&)
+BOOST_MATH_GPU_ENABLED inline T cbrt_epsilon_imp(const T*, const Tag&)
 {
    BOOST_MATH_STD_USING;
    static const T cbrt_eps = pow(tools::epsilon<T>(), T(1) / 3);
@@ -360,7 +326,7 @@ inline T cbrt_epsilon_imp(const T*, const Tag&)
 }
 
 template <class T>
-inline T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
+BOOST_MATH_GPU_ENABLED inline T cbrt_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
 {
    BOOST_MATH_STD_USING;
    return pow(tools::epsilon<T>(), T(1) / 3);
@@ -379,19 +345,19 @@ BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon_imp(const T*, const
 }
 
 template <class T>
-inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 64>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.18145860519450699870567321328132261891067079047605e-4L);
 }
 
 template <class T>
-inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
+BOOST_MATH_GPU_ENABLED inline constexpr T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 113>&) noexcept(boost::math::is_floating_point<T>::value)
 {
    return static_cast<T>(0.37252902984619140625e-8L);
 }
 
 template <class T, class Tag>
-inline T forth_root_epsilon_imp(const T*, const Tag&)
+BOOST_MATH_GPU_ENABLED inline T forth_root_epsilon_imp(const T*, const Tag&)
 {
    BOOST_MATH_STD_USING
    static const T r_eps = sqrt(sqrt(tools::epsilon<T>()));
@@ -399,7 +365,7 @@ inline T forth_root_epsilon_imp(const T*, const Tag&)
 }
 
 template <class T>
-inline T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
+BOOST_MATH_GPU_ENABLED inline T forth_root_epsilon_imp(const T*, const boost::math::integral_constant<int, 0>&)
 {
    BOOST_MATH_STD_USING
    return sqrt(sqrt(tools::epsilon<T>()));

From 0805e5e14dc386a3b4ce6bc3a74a4f4a210ca14b Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 14:39:59 -0400
Subject: [PATCH 13/35] Add NVRTC support to cauchy dist

---
 include/boost/math/distributions/cauchy.hpp | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/include/boost/math/distributions/cauchy.hpp b/include/boost/math/distributions/cauchy.hpp
index 91c848c91..0992b389e 100644
--- a/include/boost/math/distributions/cauchy.hpp
+++ b/include/boost/math/distributions/cauchy.hpp
@@ -17,12 +17,18 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/tuple.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
-#include <boost/math/distributions/fwd.hpp>
+#include <boost/math/tools/precision.hpp>
 #include <boost/math/constants/constants.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/distributions/detail/common_error_handling.hpp>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
 #include <utility>
 #include <cmath>
+#endif
 
 namespace boost{ namespace math
 {

From 5171e837be11b7ae2cd5e0a5b80ed642aeed35d0 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 14:40:09 -0400
Subject: [PATCH 14/35] Add cauchy dist NVRTC testing

---
 test/nvrtc_jamfile                     |   7 +
 test/test_cauchy_cdf_nvrtc_double.cpp  | 191 +++++++++++++++++++++++++
 test/test_cauchy_cdf_nvrtc_float.cpp   | 191 +++++++++++++++++++++++++
 test/test_cauchy_pdf_nvrtc_double.cpp  | 191 +++++++++++++++++++++++++
 test/test_cauchy_pdf_nvrtc_float.cpp   | 191 +++++++++++++++++++++++++
 test/test_cauchy_quan_nvrtc_double.cpp | 191 +++++++++++++++++++++++++
 test/test_cauchy_quan_nvrtc_float.cpp  | 191 +++++++++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_cauchy_cdf_nvrtc_double.cpp
 create mode 100644 test/test_cauchy_cdf_nvrtc_float.cpp
 create mode 100644 test/test_cauchy_pdf_nvrtc_double.cpp
 create mode 100644 test/test_cauchy_pdf_nvrtc_float.cpp
 create mode 100644 test/test_cauchy_quan_nvrtc_double.cpp
 create mode 100644 test/test_cauchy_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 10e7077d1..2825fb969 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -24,6 +24,13 @@ run test_bernoulli_pdf_nvrtc_float.cpp ;
 run test_bernoulli_quan_nvrtc_double.cpp ;
 run test_bernoulli_quan_nvrtc_float.cpp ;
 
+run test_cauchy_cdf_nvrtc_double.cpp ;
+run test_cauchy_cdf_nvrtc_float.cpp ;
+run test_cauchy_pdf_nvrtc_double.cpp ;
+run test_cauchy_pdf_nvrtc_float.cpp ;
+run test_cauchy_quan_nvrtc_double.cpp ;
+run test_cauchy_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_cauchy_cdf_nvrtc_double.cpp b/test/test_cauchy_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..f9b16b6e0
--- /dev/null
+++ b/test/test_cauchy_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/cauchy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/cauchy.hpp>
+extern "C" __global__ 
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::cauchy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::cauchy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_cauchy_cdf_nvrtc_float.cpp b/test/test_cauchy_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..0870900ae
--- /dev/null
+++ b/test/test_cauchy_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/cauchy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/cauchy.hpp>
+extern "C" __global__ 
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::cauchy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::cauchy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_cauchy_pdf_nvrtc_double.cpp b/test/test_cauchy_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..24e2808c9
--- /dev/null
+++ b/test/test_cauchy_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/cauchy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/cauchy.hpp>
+extern "C" __global__ 
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::cauchy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::cauchy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_cauchy_pdf_nvrtc_float.cpp b/test/test_cauchy_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..605d25783
--- /dev/null
+++ b/test/test_cauchy_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/cauchy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/cauchy.hpp>
+extern "C" __global__ 
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::cauchy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::cauchy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_cauchy_quan_nvrtc_double.cpp b/test/test_cauchy_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..08fa01c4f
--- /dev/null
+++ b/test/test_cauchy_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/cauchy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/cauchy.hpp>
+extern "C" __global__ 
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::cauchy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::cauchy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_cauchy_quan_nvrtc_float.cpp b/test/test_cauchy_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..6d63d1c2d
--- /dev/null
+++ b/test/test_cauchy_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/cauchy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/cauchy.hpp>
+extern "C" __global__ 
+void test_cauchy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::cauchy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_cauchy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_cauchy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_cauchy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::cauchy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 9b3c13985dd656b33ab38001c6a22e1274913ba3 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 15:54:15 -0400
Subject: [PATCH 15/35] Add NVRTC support to exponential dist

---
 include/boost/math/distributions/exponential.hpp | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/include/boost/math/distributions/exponential.hpp b/include/boost/math/distributions/exponential.hpp
index 01c442fa8..9d45ac493 100644
--- a/include/boost/math/distributions/exponential.hpp
+++ b/include/boost/math/distributions/exponential.hpp
@@ -10,12 +10,13 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/tuple.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
-#include <boost/math/distributions/fwd.hpp>
 #include <boost/math/constants/constants.hpp>
 #include <boost/math/special_functions/log1p.hpp>
 #include <boost/math/special_functions/expm1.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/distributions/detail/common_error_handling.hpp>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
 
 #ifdef _MSC_VER
 # pragma warning(push)
@@ -23,8 +24,11 @@
 # pragma warning(disable: 4702) // unreachable code (return after domain_error throw).
 #endif
 
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
 #include <utility>
 #include <cmath>
+#endif
 
 namespace boost{ namespace math{
 

From 363f2df6e08e222db7bd07e5c0b3e3bb6eb69056 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 15:54:28 -0400
Subject: [PATCH 16/35] Add exponential dist NVRTC testing

---
 test/nvrtc_jamfile                          |   7 +
 test/test_exponential_cdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++
 test/test_exponential_cdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++
 test/test_exponential_pdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++
 test/test_exponential_pdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++
 test/test_exponential_quan_nvrtc_double.cpp | 191 ++++++++++++++++++++
 test/test_exponential_quan_nvrtc_float.cpp  | 191 ++++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_exponential_cdf_nvrtc_double.cpp
 create mode 100644 test/test_exponential_cdf_nvrtc_float.cpp
 create mode 100644 test/test_exponential_pdf_nvrtc_double.cpp
 create mode 100644 test/test_exponential_pdf_nvrtc_float.cpp
 create mode 100644 test/test_exponential_quan_nvrtc_double.cpp
 create mode 100644 test/test_exponential_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 2825fb969..da734ce07 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -31,6 +31,13 @@ run test_cauchy_pdf_nvrtc_float.cpp ;
 run test_cauchy_quan_nvrtc_double.cpp ;
 run test_cauchy_quan_nvrtc_float.cpp ;
 
+run test_exponential_cdf_nvrtc_double.cpp ;
+run test_exponential_cdf_nvrtc_float.cpp ;
+run test_exponential_pdf_nvrtc_double.cpp ;
+run test_exponential_pdf_nvrtc_float.cpp ;
+run test_exponential_quan_nvrtc_double.cpp ;
+run test_exponential_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_exponential_cdf_nvrtc_double.cpp b/test/test_exponential_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..42849f8ab
--- /dev/null
+++ b/test/test_exponential_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/exponential.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/exponential.hpp>
+extern "C" __global__ 
+void test_exponential_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::exponential_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_exponential_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_exponential_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_exponential_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::exponential_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_exponential_cdf_nvrtc_float.cpp b/test/test_exponential_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..9417abf02
--- /dev/null
+++ b/test/test_exponential_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/exponential.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/exponential.hpp>
+extern "C" __global__ 
+void test_exponential_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::exponential_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_exponential_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_exponential_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_exponential_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::exponential_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_exponential_pdf_nvrtc_double.cpp b/test/test_exponential_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..2c52a4b38
--- /dev/null
+++ b/test/test_exponential_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/exponential.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/exponential.hpp>
+extern "C" __global__ 
+void test_exponential_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::exponential_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_exponential_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_exponential_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_exponential_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::exponential_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_exponential_pdf_nvrtc_float.cpp b/test/test_exponential_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..ef1aff3f3
--- /dev/null
+++ b/test/test_exponential_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/exponential.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/exponential.hpp>
+extern "C" __global__ 
+void test_exponential_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::exponential_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_exponential_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_exponential_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_exponential_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::exponential_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_exponential_quan_nvrtc_double.cpp b/test/test_exponential_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..b05f77ffe
--- /dev/null
+++ b/test/test_exponential_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/exponential.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/exponential.hpp>
+extern "C" __global__ 
+void test_exponential_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::exponential_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_exponential_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_exponential_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_exponential_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::exponential_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_exponential_quan_nvrtc_float.cpp b/test/test_exponential_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..fb932c5d7
--- /dev/null
+++ b/test/test_exponential_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/exponential.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/exponential.hpp>
+extern "C" __global__ 
+void test_exponential_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::exponential_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_exponential_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_exponential_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_exponential_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::exponential_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From f650f77527dd01d2f5108d76aaaab6c88da88050 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 16:15:26 -0400
Subject: [PATCH 17/35] Add NVRTC support to extreme value dist

---
 include/boost/math/distributions/extreme_value.hpp | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/include/boost/math/distributions/extreme_value.hpp b/include/boost/math/distributions/extreme_value.hpp
index 0d638092b..73454d29d 100644
--- a/include/boost/math/distributions/extreme_value.hpp
+++ b/include/boost/math/distributions/extreme_value.hpp
@@ -10,12 +10,14 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
 #include <boost/math/tools/tuple.hpp>
-#include <boost/math/distributions/fwd.hpp>
+#include <boost/math/tools/precision.hpp>
 #include <boost/math/constants/constants.hpp>
 #include <boost/math/special_functions/log1p.hpp>
 #include <boost/math/special_functions/expm1.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/distributions/detail/common_error_handling.hpp>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
 
 //
 // This is the maximum extreme value distribution, see
@@ -24,8 +26,11 @@
 // Also known as a Fisher-Tippett distribution, a log-Weibull
 // distribution or a Gumbel distribution.
 
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
 #include <utility>
 #include <cmath>
+#endif
 
 #ifdef _MSC_VER
 # pragma warning(push)
@@ -280,7 +285,7 @@ BOOST_MATH_GPU_ENABLED inline RealType logcdf(const complemented2_type<extreme_v
 }
 
 template <class RealType, class Policy>
-RealType quantile(const complemented2_type<extreme_value_distribution<RealType, Policy>, RealType>& c)
+BOOST_MATH_GPU_ENABLED RealType quantile(const complemented2_type<extreme_value_distribution<RealType, Policy>, RealType>& c)
 {
    BOOST_MATH_STD_USING // for ADL of std functions
 

From e508d498ca7447c4a5f713558363e164d9c1c4de Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 16:15:48 -0400
Subject: [PATCH 18/35] Add NVRTC paths to log_max_value and log_min_value

---
 include/boost/math/tools/precision.hpp | 32 ++++++++++++++++----------
 1 file changed, 20 insertions(+), 12 deletions(-)

diff --git a/include/boost/math/tools/precision.hpp b/include/boost/math/tools/precision.hpp
index 096a3265a..1e4bf58ac 100644
--- a/include/boost/math/tools/precision.hpp
+++ b/include/boost/math/tools/precision.hpp
@@ -208,26 +208,34 @@ struct log_limit_noexcept_traits : public log_limit_noexcept_traits_imp<T, boost
 template <class T>
 BOOST_MATH_GPU_ENABLED inline constexpr T log_max_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(detail::log_limit_noexcept_traits<T>::value)
 {
-#ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
-   return detail::log_max_value<T>(typename detail::log_limit_traits<T>::tag_type());
+#ifndef BOOST_MATH_HAS_NVRTC
+   #ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
+      return detail::log_max_value<T>(typename detail::log_limit_traits<T>::tag_type());
+   #else
+      BOOST_MATH_ASSERT(::boost::math::numeric_limits<T>::is_specialized);
+      BOOST_MATH_STD_USING
+      static const T val = log((boost::math::numeric_limits<T>::max)());
+      return val;
+   #endif
 #else
-   BOOST_MATH_ASSERT(::boost::math::numeric_limits<T>::is_specialized);
-   BOOST_MATH_STD_USING
-   static const T val = log((boost::math::numeric_limits<T>::max)());
-   return val;
+   return log((boost::math::numeric_limits<T>::max)());
 #endif
 }
 
 template <class T>
 BOOST_MATH_GPU_ENABLED inline constexpr T log_min_value(BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T)) noexcept(detail::log_limit_noexcept_traits<T>::value)
 {
-#ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
-   return detail::log_min_value<T>(typename detail::log_limit_traits<T>::tag_type());
+#ifndef BOOST_MATH_HAS_NVRTC
+   #ifndef BOOST_NO_LIMITS_COMPILE_TIME_CONSTANTS
+      return detail::log_min_value<T>(typename detail::log_limit_traits<T>::tag_type());
+   #else
+      BOOST_MATH_ASSERT(::boost::math::numeric_limits<T>::is_specialized);
+      BOOST_MATH_STD_USING
+      static const T val = log((boost::math::numeric_limits<T>::min)());
+      return val;
+   #endif
 #else
-   BOOST_MATH_ASSERT(::boost::math::numeric_limits<T>::is_specialized);
-   BOOST_MATH_STD_USING
-   static const T val = log((boost::math::numeric_limits<T>::min)());
-   return val;
+   return log((boost::math::numeric_limits<T>::min)());
 #endif
 }
 

From 4217281be5e34a76626133a247ae4196f238173a Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 16:16:03 -0400
Subject: [PATCH 19/35] Add extreme value dist NVRTC testing

---
 test/nvrtc_jamfile                            |   7 +
 test/test_extreme_value_cdf_nvrtc_double.cpp  | 191 ++++++++++++++++++
 test/test_extreme_value_cdf_nvrtc_float.cpp   | 191 ++++++++++++++++++
 test/test_extreme_value_pdf_nvrtc_double.cpp  | 191 ++++++++++++++++++
 test/test_extreme_value_pdf_nvrtc_float.cpp   | 191 ++++++++++++++++++
 test/test_extreme_value_quan_nvrtc_double.cpp | 191 ++++++++++++++++++
 test/test_extreme_value_quan_nvrtc_float.cpp  | 191 ++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_extreme_value_cdf_nvrtc_double.cpp
 create mode 100644 test/test_extreme_value_cdf_nvrtc_float.cpp
 create mode 100644 test/test_extreme_value_pdf_nvrtc_double.cpp
 create mode 100644 test/test_extreme_value_pdf_nvrtc_float.cpp
 create mode 100644 test/test_extreme_value_quan_nvrtc_double.cpp
 create mode 100644 test/test_extreme_value_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index da734ce07..9620115d8 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -38,6 +38,13 @@ run test_exponential_pdf_nvrtc_float.cpp ;
 run test_exponential_quan_nvrtc_double.cpp ;
 run test_exponential_quan_nvrtc_float.cpp ;
 
+run test_extreme_value_cdf_nvrtc_double.cpp ;
+run test_extreme_value_cdf_nvrtc_float.cpp ;
+run test_extreme_value_pdf_nvrtc_double.cpp ;
+run test_extreme_value_pdf_nvrtc_float.cpp ;
+run test_extreme_value_quan_nvrtc_double.cpp ;
+run test_extreme_value_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_extreme_value_cdf_nvrtc_double.cpp b/test/test_extreme_value_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..b3600b21b
--- /dev/null
+++ b/test/test_extreme_value_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/extreme_value.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/extreme_value.hpp>
+extern "C" __global__ 
+void test_extreme_value_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::extreme_value_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_extreme_value_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_extreme_value_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_extreme_value_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::extreme_value_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_extreme_value_cdf_nvrtc_float.cpp b/test/test_extreme_value_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..b3600b21b
--- /dev/null
+++ b/test/test_extreme_value_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/extreme_value.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/extreme_value.hpp>
+extern "C" __global__ 
+void test_extreme_value_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::extreme_value_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_extreme_value_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_extreme_value_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_extreme_value_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::extreme_value_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_extreme_value_pdf_nvrtc_double.cpp b/test/test_extreme_value_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..049f45d89
--- /dev/null
+++ b/test/test_extreme_value_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/extreme_value.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/extreme_value.hpp>
+extern "C" __global__ 
+void test_extreme_value_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::extreme_value_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_extreme_value_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_extreme_value_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_extreme_value_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::extreme_value_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_extreme_value_pdf_nvrtc_float.cpp b/test/test_extreme_value_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..adbd26348
--- /dev/null
+++ b/test/test_extreme_value_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/extreme_value.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/extreme_value.hpp>
+extern "C" __global__ 
+void test_extreme_value_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::extreme_value_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_extreme_value_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_extreme_value_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_extreme_value_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::extreme_value_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_extreme_value_quan_nvrtc_double.cpp b/test/test_extreme_value_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..239df9994
--- /dev/null
+++ b/test/test_extreme_value_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/extreme_value.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/extreme_value.hpp>
+extern "C" __global__ 
+void test_extreme_value_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::extreme_value_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_extreme_value_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_extreme_value_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_extreme_value_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::extreme_value_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_extreme_value_quan_nvrtc_float.cpp b/test/test_extreme_value_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..fc9d8c5f4
--- /dev/null
+++ b/test/test_extreme_value_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/extreme_value.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/extreme_value.hpp>
+extern "C" __global__ 
+void test_extreme_value_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::extreme_value_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_extreme_value_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_extreme_value_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_extreme_value_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::extreme_value_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 5adc939da7cae0133355aada7f6c0a3172b54545 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 17:03:02 -0400
Subject: [PATCH 20/35] Replace thurst::pair with cuda::std::pair for NVRTC
 compat

---
 include/boost/math/tools/config.hpp |  1 +
 include/boost/math/tools/tuple.hpp  | 44 ++++++-----------------------
 2 files changed, 9 insertions(+), 36 deletions(-)

diff --git a/include/boost/math/tools/config.hpp b/include/boost/math/tools/config.hpp
index b67ce0ae9..3fa05b351 100644
--- a/include/boost/math/tools/config.hpp
+++ b/include/boost/math/tools/config.hpp
@@ -674,6 +674,7 @@ namespace boost{ namespace math{
 #include <cuda.h>
 #include <cuda_runtime.h>
 #include <cuda/std/type_traits>
+#include <cuda/std/utility>
 
 #  define BOOST_MATH_CUDA_ENABLED __host__ __device__
 #  define BOOST_MATH_HAS_GPU_SUPPORT
diff --git a/include/boost/math/tools/tuple.hpp b/include/boost/math/tools/tuple.hpp
index fba37fb4f..26f81042f 100644
--- a/include/boost/math/tools/tuple.hpp
+++ b/include/boost/math/tools/tuple.hpp
@@ -9,50 +9,22 @@
 
 #include <boost/math/tools/config.hpp>
 
-#ifdef BOOST_MATH_HAS_NVRTC
+#ifdef BOOST_MATH_ENABLE_CUDA
 
-namespace boost {
-namespace math {
-
-template <typename T, typename U>
-class pair
-{
-public:
-    T first;
-    U second;
-
-    BOOST_MATH_GPU_ENABLED pair(T x, U y) noexcept : first {x}, second {y} {}
-};
-
-template <typename T, typename U>
-BOOST_MATH_GPU_ENABLED auto make_pair(T x, U y) noexcept
-{
-    return pair(x, y);
-}
-
-} // namespace math
-} // namespace boost
-
-
-#elif defined(BOOST_MATH_ENABLE_CUDA)
-
-#include <thrust/pair.h>
-#include <thrust/tuple.h>
+#include <cuda/std/utility>
 
 namespace boost { 
 namespace math {
 
-using thrust::pair;
-using thrust::tuple;
+using cuda::std::pair;
+using cuda::std::tuple;
 
-using thrust::make_pair;
-using thrust::make_tuple;
+using cuda::std::make_pair;
 
-using thrust::tie;
-using thrust::get;
+using cuda::std::get;
 
-using thrust::tuple_size;
-using thrust::tuple_element;
+using cuda::std::tuple_size;
+using cuda::std::tuple_element;
 
 } // namespace math
 } // namespace boost

From 4dc7070722aa0d7cf869ae958a0a9a69ad05c716 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Wed, 14 Aug 2024 17:13:29 -0400
Subject: [PATCH 21/35] Implement make_tuple since it's missing from cuda/std

---
 include/boost/math/tools/tuple.hpp | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)

diff --git a/include/boost/math/tools/tuple.hpp b/include/boost/math/tools/tuple.hpp
index 26f81042f..6a949d5e5 100644
--- a/include/boost/math/tools/tuple.hpp
+++ b/include/boost/math/tools/tuple.hpp
@@ -11,6 +11,7 @@
 
 #ifdef BOOST_MATH_ENABLE_CUDA
 
+#include <boost/math/tools/type_traits.hpp>
 #include <cuda/std/utility>
 
 namespace boost { 
@@ -26,6 +27,31 @@ using cuda::std::get;
 using cuda::std::tuple_size;
 using cuda::std::tuple_element;
 
+namespace detail {
+
+template <typename T>
+BOOST_MATH_GPU_ENABLED T&& forward(boost::math::remove_reference_t<T>& arg) noexcept
+{
+    return static_cast<T&&>(arg);
+}
+
+template <typename T>
+BOOST_MATH_GPU_ENABLED T&& forward(boost::math::remove_reference_t<T>&& arg) noexcept
+{
+    static_assert(!boost::math::is_lvalue_reference<T>::value, "Cannot forward an rvalue as an lvalue.");
+    return static_cast<T&&>(arg);
+}
+
+} // namespace detail
+
+template <typename T, typename... Ts>
+BOOST_MATH_GPU_ENABLED auto make_tuple(T&& t, Ts&&... ts) 
+{
+    return cuda::std::tuple<boost::math::decay_t<T>, boost::math::decay_t<Ts>...>(
+        boost::math::detail::forward<T>(t), boost::math::detail::forward<Ts>(ts)...
+    );
+}
+
 } // namespace math
 } // namespace boost
 

From 1a8bfbb713d44a7fd896c61e3d745fe8eb55b22e Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 09:42:17 -0400
Subject: [PATCH 22/35] Add more macro definitions

---
 include/boost/math/tools/config.hpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/include/boost/math/tools/config.hpp b/include/boost/math/tools/config.hpp
index 3fa05b351..364aa8f54 100644
--- a/include/boost/math/tools/config.hpp
+++ b/include/boost/math/tools/config.hpp
@@ -793,6 +793,7 @@ BOOST_MATH_GPU_ENABLED constexpr T cuda_safe_max(const T& a, const T& b) { retur
 #define BOOST_MATH_GPU_ENABLED __host__ __device__
 
 #define BOOST_MATH_STATIC static
+#define BOOST_MATH_STATIC_LOCAL_VARIABLE
 
 #define BOOST_MATH_NOEXCEPT(T) noexcept(boost::math::is_floating_point_v<T>)
 #define BOOST_MATH_EXPLICIT_TEMPLATE_TYPE(T) 
@@ -800,6 +801,7 @@ BOOST_MATH_GPU_ENABLED constexpr T cuda_safe_max(const T& a, const T& b) { retur
 #define BOOST_MATH_BIG_CONSTANT(T, N, V) static_cast<T>(V)
 #define BOOST_MATH_FORCEINLINE __forceinline__
 #define BOOST_MATH_STD_USING  
+#define BOOST_MATH_IF_CONSTEXPR if constexpr
 
 // This should be defined to nothing but since it is not specifically a math macro
 // we need to undef before proceeding

From 4518351ff9b3dfd0417f6f6ca761cb37ef9f8da4 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 09:42:34 -0400
Subject: [PATCH 23/35] Add NVRTC support to holtsmark dist

---
 .../boost/math/distributions/holtsmark.hpp    | 47 +++++++++++--------
 1 file changed, 27 insertions(+), 20 deletions(-)

diff --git a/include/boost/math/distributions/holtsmark.hpp b/include/boost/math/distributions/holtsmark.hpp
index 8c976b8c8..04f5484f4 100644
--- a/include/boost/math/distributions/holtsmark.hpp
+++ b/include/boost/math/distributions/holtsmark.hpp
@@ -15,16 +15,23 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
 #include <boost/math/tools/tuple.hpp>
-#include <boost/math/distributions/fwd.hpp>
+#include <boost/math/tools/type_traits.hpp>
 #include <boost/math/constants/constants.hpp>
-#include <boost/math/tools/big_constant.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/distributions/detail/common_error_handling.hpp>
 #include <boost/math/distributions/detail/derived_accessors.hpp>
 #include <boost/math/tools/rational.hpp>
 #include <boost/math/special_functions/cbrt.hpp>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
+#include <boost/math/tools/promotion.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
+#include <boost/math/tools/big_constant.hpp>
 #include <utility>
 #include <cmath>
+#endif
 
 namespace boost { namespace math {
 template <class RealType, class Policy>
@@ -33,7 +40,7 @@ class holtsmark_distribution;
 namespace detail {
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -263,7 +270,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_plus_imp_prec(const RealTyp
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -615,14 +622,14 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_plus_imp_prec(const RealTyp
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 53> &tag) {
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53> &tag) {
     BOOST_MATH_STD_USING // for ADL of std functions
 
     return holtsmark_pdf_plus_imp_prec<RealType>(abs(x), tag);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>& tag) {
     BOOST_MATH_STD_USING // for ADL of std functions
 
     return holtsmark_pdf_plus_imp_prec<RealType>(abs(x), tag);
@@ -655,7 +662,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_imp(const holtsmark_distrib
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -671,7 +678,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_pdf_imp(const holtsmark_distrib
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -920,7 +927,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_plus_imp_prec(const RealTyp
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1303,7 +1310,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_plus_imp_prec(const RealTyp
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 53>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 53>& tag) {
     if (x >= 0) {
         return complement ? holtsmark_cdf_plus_imp_prec(x, tag) : 1 - holtsmark_cdf_plus_imp_prec(x, tag);
     }
@@ -1316,7 +1323,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_imp_prec(const RealType& x,
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 113>& tag) {
     if (x >= 0) {
         return complement ? holtsmark_cdf_plus_imp_prec(x, tag) : 1 - holtsmark_cdf_plus_imp_prec(x, tag);
     }
@@ -1355,7 +1362,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_imp(const holtsmark_distrib
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -1371,7 +1378,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_cdf_imp(const holtsmark_distrib
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1634,7 +1641,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_upper_imp_prec(const R
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2230,7 +2237,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_upper_imp_prec(const R
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 53>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 53>& tag)
 {
     if (p > 0.5) {
         return !complement ? holtsmark_quantile_upper_imp_prec(1 - p, tag) : -holtsmark_quantile_upper_imp_prec(1 - p, tag);
@@ -2240,7 +2247,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_imp_prec(const RealTyp
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 113>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 113>& tag)
 {
     if (p > 0.5) {
         return !complement ? holtsmark_quantile_upper_imp_prec(1 - p, tag) : -holtsmark_quantile_upper_imp_prec(1 - p, tag);
@@ -2277,7 +2284,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_imp(const holtsmark_di
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -2291,13 +2298,13 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_quantile_imp(const holtsmark_di
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_entropy_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_entropy_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(2.06944850513462440032);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType holtsmark_entropy_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType holtsmark_entropy_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, 2.0694485051346244003155800384542166381);
 }
@@ -2320,7 +2327,7 @@ BOOST_MATH_GPU_ENABLED inline RealType holtsmark_entropy_imp(const holtsmark_dis
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0

From b123d38e37fb052f98526bdef7756ddfe1e50d66 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 09:42:48 -0400
Subject: [PATCH 24/35] Add holtsmark dist NVRTC testing

---
 test/nvrtc_jamfile                        |   7 +
 test/test_holtsmark_cdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++
 test/test_holtsmark_cdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++
 test/test_holtsmark_pdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++
 test/test_holtsmark_pdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++
 test/test_holtsmark_quan_nvrtc_double.cpp | 191 ++++++++++++++++++++++
 test/test_holtsmark_quan_nvrtc_float.cpp  | 191 ++++++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_holtsmark_cdf_nvrtc_double.cpp
 create mode 100644 test/test_holtsmark_cdf_nvrtc_float.cpp
 create mode 100644 test/test_holtsmark_pdf_nvrtc_double.cpp
 create mode 100644 test/test_holtsmark_pdf_nvrtc_float.cpp
 create mode 100644 test/test_holtsmark_quan_nvrtc_double.cpp
 create mode 100644 test/test_holtsmark_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 9620115d8..93e6deddd 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -45,6 +45,13 @@ run test_extreme_value_pdf_nvrtc_float.cpp ;
 run test_extreme_value_quan_nvrtc_double.cpp ;
 run test_extreme_value_quan_nvrtc_float.cpp ;
 
+run test_holtsmark_cdf_nvrtc_double.cpp ;
+run test_holtsmark_cdf_nvrtc_float.cpp ;
+run test_holtsmark_pdf_nvrtc_double.cpp ;
+run test_holtsmark_pdf_nvrtc_float.cpp ;
+run test_holtsmark_quan_nvrtc_double.cpp ;
+run test_holtsmark_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_holtsmark_cdf_nvrtc_double.cpp b/test/test_holtsmark_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..3b8c9ba94
--- /dev/null
+++ b/test/test_holtsmark_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/holtsmark.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/holtsmark.hpp>
+extern "C" __global__ 
+void test_holtsmark_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::holtsmark_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_holtsmark_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_holtsmark_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_holtsmark_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::holtsmark_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_holtsmark_cdf_nvrtc_float.cpp b/test/test_holtsmark_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..a3ffec594
--- /dev/null
+++ b/test/test_holtsmark_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/holtsmark.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/holtsmark.hpp>
+extern "C" __global__ 
+void test_holtsmark_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::holtsmark_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_holtsmark_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_holtsmark_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_holtsmark_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::holtsmark_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_holtsmark_pdf_nvrtc_double.cpp b/test/test_holtsmark_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..a8355d368
--- /dev/null
+++ b/test/test_holtsmark_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/holtsmark.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/holtsmark.hpp>
+extern "C" __global__ 
+void test_holtsmark_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::holtsmark_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_holtsmark_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_holtsmark_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_holtsmark_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::holtsmark_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_holtsmark_pdf_nvrtc_float.cpp b/test/test_holtsmark_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..7a02d1983
--- /dev/null
+++ b/test/test_holtsmark_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/holtsmark.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/holtsmark.hpp>
+extern "C" __global__ 
+void test_holtsmark_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::holtsmark_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_holtsmark_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_holtsmark_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_holtsmark_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::holtsmark_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_holtsmark_quan_nvrtc_double.cpp b/test/test_holtsmark_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..1c2cc61fa
--- /dev/null
+++ b/test/test_holtsmark_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/holtsmark.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/holtsmark.hpp>
+extern "C" __global__ 
+void test_holtsmark_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::holtsmark_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_holtsmark_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_holtsmark_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_holtsmark_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::holtsmark_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_holtsmark_quan_nvrtc_float.cpp b/test/test_holtsmark_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..a343f232e
--- /dev/null
+++ b/test/test_holtsmark_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/holtsmark.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/holtsmark.hpp>
+extern "C" __global__ 
+void test_holtsmark_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::holtsmark_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_holtsmark_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_holtsmark_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_holtsmark_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::holtsmark_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 6ba8cbbbee7c281536219b3534758d92684a472a Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 09:59:34 -0400
Subject: [PATCH 25/35] Add NVRTC support to landau dist

---
 include/boost/math/distributions/landau.hpp | 74 ++++++++++++---------
 1 file changed, 41 insertions(+), 33 deletions(-)

diff --git a/include/boost/math/distributions/landau.hpp b/include/boost/math/distributions/landau.hpp
index 424931c27..129eca287 100644
--- a/include/boost/math/distributions/landau.hpp
+++ b/include/boost/math/distributions/landau.hpp
@@ -15,15 +15,22 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
 #include <boost/math/tools/tuple.hpp>
-#include <boost/math/distributions/fwd.hpp>
+#include <boost/math/tools/type_traits.hpp>
 #include <boost/math/constants/constants.hpp>
-#include <boost/math/tools/big_constant.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/distributions/detail/common_error_handling.hpp>
 #include <boost/math/distributions/detail/derived_accessors.hpp>
 #include <boost/math/tools/rational.hpp>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
+#include <boost/math/tools/promotion.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
+#include <boost/math/tools/big_constant.hpp>
 #include <utility>
 #include <cmath>
+#endif
 
 namespace boost { namespace math {
 template <class RealType, class Policy>
@@ -32,7 +39,7 @@ class landau_distribution;
 namespace detail {
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -345,7 +352,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_plus_imp_prec(const RealType&
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -873,7 +880,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_plus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1013,7 +1020,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_minus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1224,7 +1231,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_minus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 53> &tag) {
+BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53> &tag) {
     if (x >= 0) {
         return landau_pdf_plus_imp_prec<RealType>(x, tag);
     }
@@ -1237,7 +1244,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_imp_prec(const RealType& x, co
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>& tag) {
     if (x >= 0) {
         return landau_pdf_plus_imp_prec<RealType>(x, tag);
     }
@@ -1277,7 +1284,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_imp(const landau_distribution<
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -1293,7 +1300,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_pdf_imp(const landau_distribution<
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1600,7 +1607,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_plus_imp_prec(const RealType&
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2123,7 +2130,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_plus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2266,7 +2273,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_minus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2482,7 +2489,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_minus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 53>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 53>& tag) {
     if (x >= 0) {
         return complement ? landau_cdf_plus_imp_prec(x, tag) : 1 - landau_cdf_plus_imp_prec(x, tag);
     }
@@ -2495,7 +2502,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_imp_prec(const RealType& x, bo
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 113>& tag) {
     if (x >= 0) {
         return complement ? landau_cdf_plus_imp_prec(x, tag) : 1 - landau_cdf_plus_imp_prec(x, tag);
     }
@@ -2535,7 +2542,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_imp(const landau_distribution<
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -2551,7 +2558,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_cdf_imp(const landau_distribution<
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_lower_imp_prec(const RealType& p, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_lower_imp_prec(const RealType& p, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2882,7 +2889,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_lower_imp_prec(const Real
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_lower_imp_prec(const RealType& p, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_lower_imp_prec(const RealType& p, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -3542,7 +3549,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_lower_imp_prec(const Real
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -3782,7 +3789,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_upper_imp_prec(const Real
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -4229,7 +4236,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_upper_imp_prec(const Real
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 53>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 53>& tag)
 {
     if (p > 0.5) 
     {
@@ -4240,7 +4247,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 113>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 113>& tag)
 {
     if (p > 0.5) 
     {
@@ -4279,7 +4286,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_imp(const landau_distribu
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -4293,13 +4300,13 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_quantile_imp(const landau_distribu
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_mode_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_mode_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(-0.42931452986133525017);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_mode_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_mode_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, -0.42931452986133525016556463510885028346);
 }
@@ -4328,7 +4335,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_mode_imp(const landau_distribution
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -4342,13 +4349,13 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_mode_imp(const landau_distribution
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_median_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_median_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(0.57563014394507821440);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_median_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_median_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, 0.57563014394507821439627930892257517269);
 }
@@ -4377,7 +4384,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_median_imp(const landau_distributi
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -4391,13 +4398,13 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_median_imp(const landau_distributi
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_entropy_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_entropy_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(2.37263644000448182448);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType landau_entropy_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType landau_entropy_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, 2.3726364400044818244844049010588577710);
 }
@@ -4420,7 +4427,7 @@ BOOST_MATH_GPU_ENABLED inline RealType landau_entropy_imp(const landau_distribut
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -4445,12 +4452,13 @@ class landau_distribution
     BOOST_MATH_GPU_ENABLED landau_distribution(RealType l_location = 0, RealType l_scale = 1)
         : mu(l_location), c(l_scale)
     {
+        BOOST_MATH_STD_USING
+        
         constexpr auto function = "boost::math::landau_distribution<%1%>::landau_distribution";
         RealType result = 0;
         detail::check_location(function, l_location, &result, Policy());
         detail::check_scale(function, l_scale, &result, Policy());
 
-        using std::log;
         location_bias = -2 / constants::pi<RealType>() * log(l_scale);
     } // landau_distribution
 

From 3eb59a6c228d7090d2b92d13ca1d1366a6899d60 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 09:59:46 -0400
Subject: [PATCH 26/35] Add landau dist NVRTC testing

---
 test/nvrtc_jamfile                     |   7 +
 test/test_landau_cdf_nvrtc_double.cpp  | 191 +++++++++++++++++++++++++
 test/test_landau_cdf_nvrtc_float.cpp   | 191 +++++++++++++++++++++++++
 test/test_landau_pdf_nvrtc_double.cpp  | 191 +++++++++++++++++++++++++
 test/test_landau_pdf_nvrtc_float.cpp   | 191 +++++++++++++++++++++++++
 test/test_landau_quan_nvrtc_double.cpp | 191 +++++++++++++++++++++++++
 test/test_landau_quan_nvrtc_float.cpp  | 191 +++++++++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_landau_cdf_nvrtc_double.cpp
 create mode 100644 test/test_landau_cdf_nvrtc_float.cpp
 create mode 100644 test/test_landau_pdf_nvrtc_double.cpp
 create mode 100644 test/test_landau_pdf_nvrtc_float.cpp
 create mode 100644 test/test_landau_quan_nvrtc_double.cpp
 create mode 100644 test/test_landau_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 93e6deddd..0ed1c9de6 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -52,6 +52,13 @@ run test_holtsmark_pdf_nvrtc_float.cpp ;
 run test_holtsmark_quan_nvrtc_double.cpp ;
 run test_holtsmark_quan_nvrtc_float.cpp ;
 
+run test_landau_cdf_nvrtc_double.cpp ;
+run test_landau_cdf_nvrtc_float.cpp ;
+run test_landau_pdf_nvrtc_double.cpp ;
+run test_landau_pdf_nvrtc_float.cpp ;
+run test_landau_quan_nvrtc_double.cpp ;
+run test_landau_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_landau_cdf_nvrtc_double.cpp b/test/test_landau_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..f23a758e1
--- /dev/null
+++ b/test/test_landau_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/landau.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/landau.hpp>
+extern "C" __global__ 
+void test_landau_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::landau_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_landau_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_landau_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_landau_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::landau_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_landau_cdf_nvrtc_float.cpp b/test/test_landau_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..1a339724e
--- /dev/null
+++ b/test/test_landau_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/landau.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/landau.hpp>
+extern "C" __global__ 
+void test_landau_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::landau_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_landau_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_landau_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_landau_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::landau_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_landau_pdf_nvrtc_double.cpp b/test/test_landau_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..1e8df7d07
--- /dev/null
+++ b/test/test_landau_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/landau.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/landau.hpp>
+extern "C" __global__ 
+void test_landau_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::landau_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_landau_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_landau_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_landau_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::landau_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_landau_pdf_nvrtc_float.cpp b/test/test_landau_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..1e8df7d07
--- /dev/null
+++ b/test/test_landau_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/landau.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/landau.hpp>
+extern "C" __global__ 
+void test_landau_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::landau_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_landau_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_landau_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_landau_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::landau_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_landau_quan_nvrtc_double.cpp b/test/test_landau_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..f4a5b9553
--- /dev/null
+++ b/test/test_landau_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/landau.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/landau.hpp>
+extern "C" __global__ 
+void test_landau_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::landau_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_landau_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_landau_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_landau_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::landau_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_landau_quan_nvrtc_float.cpp b/test/test_landau_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..4a10b497c
--- /dev/null
+++ b/test/test_landau_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/landau.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/landau.hpp>
+extern "C" __global__ 
+void test_landau_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::landau_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_landau_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_landau_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_landau_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::landau_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 4a10ab55d71e556a055ac055f1f3542308b6e2d9 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:15:21 -0400
Subject: [PATCH 27/35] Add NVRTC support to laplace dist

---
 include/boost/math/distributions/laplace.hpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/boost/math/distributions/laplace.hpp b/include/boost/math/distributions/laplace.hpp
index 55e2236f9..912a15712 100644
--- a/include/boost/math/distributions/laplace.hpp
+++ b/include/boost/math/distributions/laplace.hpp
@@ -25,7 +25,8 @@
 #include <boost/math/distributions/detail/common_error_handling.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/constants/constants.hpp>
-#include <limits>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
 
 namespace boost{ namespace math{
 

From 6ef1632f18c4f5b0f158c7e3116628ba2d2b792d Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:15:34 -0400
Subject: [PATCH 28/35] Add laplace dist NVRTC testing

---
 test/nvrtc_jamfile                      |   8 +
 test/test_laplace_cdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++++
 test/test_laplace_cdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++++
 test/test_laplace_pdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++++
 test/test_laplace_pdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++++
 test/test_laplace_quan_nvrtc_double.cpp | 191 ++++++++++++++++++++++++
 test/test_laplace_quan_nvrtc_float.cpp  | 191 ++++++++++++++++++++++++
 7 files changed, 1154 insertions(+)
 create mode 100644 test/test_laplace_cdf_nvrtc_double.cpp
 create mode 100644 test/test_laplace_cdf_nvrtc_float.cpp
 create mode 100644 test/test_laplace_pdf_nvrtc_double.cpp
 create mode 100644 test/test_laplace_pdf_nvrtc_float.cpp
 create mode 100644 test/test_laplace_quan_nvrtc_double.cpp
 create mode 100644 test/test_laplace_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 0ed1c9de6..3a9ecb739 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -59,6 +59,14 @@ run test_landau_pdf_nvrtc_float.cpp ;
 run test_landau_quan_nvrtc_double.cpp ;
 run test_landau_quan_nvrtc_float.cpp ;
 
+run test_laplace_cdf_nvrtc_double.cpp ;
+run test_laplace_cdf_nvrtc_float.cpp ;
+run test_laplace_pdf_nvrtc_double.cpp ;
+run test_laplace_pdf_nvrtc_float.cpp ;
+run test_laplace_quan_nvrtc_double.cpp ;
+run test_laplace_quan_nvrtc_float.cpp ;
+
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_laplace_cdf_nvrtc_double.cpp b/test/test_laplace_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..9d35b5862
--- /dev/null
+++ b/test/test_laplace_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/laplace.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/laplace.hpp>
+extern "C" __global__ 
+void test_laplace_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::laplace_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_laplace_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_laplace_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_laplace_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::laplace_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_laplace_cdf_nvrtc_float.cpp b/test/test_laplace_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..aacf59cb0
--- /dev/null
+++ b/test/test_laplace_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/laplace.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/laplace.hpp>
+extern "C" __global__ 
+void test_laplace_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::laplace_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_laplace_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_laplace_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_laplace_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::laplace_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_laplace_pdf_nvrtc_double.cpp b/test/test_laplace_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..49c386415
--- /dev/null
+++ b/test/test_laplace_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/laplace.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/laplace.hpp>
+extern "C" __global__ 
+void test_laplace_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::laplace_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_laplace_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_laplace_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_laplace_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::laplace_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_laplace_pdf_nvrtc_float.cpp b/test/test_laplace_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..2100e2cdc
--- /dev/null
+++ b/test/test_laplace_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/laplace.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/laplace.hpp>
+extern "C" __global__ 
+void test_laplace_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::laplace_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_laplace_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_laplace_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_laplace_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::laplace_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_laplace_quan_nvrtc_double.cpp b/test/test_laplace_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..cf02db5a5
--- /dev/null
+++ b/test/test_laplace_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/laplace.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/laplace.hpp>
+extern "C" __global__ 
+void test_laplace_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::laplace_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_laplace_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_laplace_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_laplace_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::laplace_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_laplace_quan_nvrtc_float.cpp b/test/test_laplace_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..36472aaa2
--- /dev/null
+++ b/test/test_laplace_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/laplace.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/laplace.hpp>
+extern "C" __global__ 
+void test_laplace_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::laplace_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_laplace_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_laplace_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_laplace_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::laplace_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 3b9863d99e082bb77ed54ae5ce043738d71bdc2a Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:22:06 -0400
Subject: [PATCH 29/35] Add NVRTC support to logistic dist

---
 include/boost/math/distributions/logistic.hpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/include/boost/math/distributions/logistic.hpp b/include/boost/math/distributions/logistic.hpp
index 4f94bd8a4..56dc6e9f2 100644
--- a/include/boost/math/distributions/logistic.hpp
+++ b/include/boost/math/distributions/logistic.hpp
@@ -11,12 +11,14 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
 #include <boost/math/tools/tuple.hpp>
+#include <boost/math/tools/precision.hpp>
 #include <boost/math/distributions/fwd.hpp>
 #include <boost/math/distributions/detail/common_error_handling.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/special_functions/log1p.hpp>
 #include <boost/math/constants/constants.hpp>
-#include <utility>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
 
 namespace boost { namespace math { 
 

From e165bfafa42fc05fefd26c5a0d8971c1752eb48f Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:22:19 -0400
Subject: [PATCH 30/35] Add logistic dist NVRTC testing

---
 test/nvrtc_jamfile                       |   6 +
 test/test_logistic_cdf_nvrtc_double.cpp  | 191 +++++++++++++++++++++++
 test/test_logistic_cdf_nvrtc_float.cpp   | 191 +++++++++++++++++++++++
 test/test_logistic_pdf_nvrtc_double.cpp  | 191 +++++++++++++++++++++++
 test/test_logistic_pdf_nvrtc_float.cpp   | 191 +++++++++++++++++++++++
 test/test_logistic_quan_nvrtc_double.cpp | 191 +++++++++++++++++++++++
 test/test_logistic_quan_nvrtc_float.cpp  | 191 +++++++++++++++++++++++
 7 files changed, 1152 insertions(+)
 create mode 100644 test/test_logistic_cdf_nvrtc_double.cpp
 create mode 100644 test/test_logistic_cdf_nvrtc_float.cpp
 create mode 100644 test/test_logistic_pdf_nvrtc_double.cpp
 create mode 100644 test/test_logistic_pdf_nvrtc_float.cpp
 create mode 100644 test/test_logistic_quan_nvrtc_double.cpp
 create mode 100644 test/test_logistic_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index 3a9ecb739..c0af6e7fa 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -66,6 +66,12 @@ run test_laplace_pdf_nvrtc_float.cpp ;
 run test_laplace_quan_nvrtc_double.cpp ;
 run test_laplace_quan_nvrtc_float.cpp ;
 
+run test_logistic_cdf_nvrtc_double.cpp ;
+run test_logistic_cdf_nvrtc_float.cpp ;
+run test_logistic_pdf_nvrtc_double.cpp ;
+run test_logistic_pdf_nvrtc_float.cpp ;
+run test_logistic_quan_nvrtc_double.cpp ;
+run test_logistic_quan_nvrtc_float.cpp ;
 
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
diff --git a/test/test_logistic_cdf_nvrtc_double.cpp b/test/test_logistic_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..5548dc412
--- /dev/null
+++ b/test/test_logistic_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/logistic.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/logistic.hpp>
+extern "C" __global__ 
+void test_logistic_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::logistic_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_logistic_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_logistic_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_logistic_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::logistic_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_logistic_cdf_nvrtc_float.cpp b/test/test_logistic_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..ea65cc97c
--- /dev/null
+++ b/test/test_logistic_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/logistic.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/logistic.hpp>
+extern "C" __global__ 
+void test_logistic_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::logistic_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_logistic_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_logistic_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_logistic_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::logistic_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_logistic_pdf_nvrtc_double.cpp b/test/test_logistic_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..d287225cf
--- /dev/null
+++ b/test/test_logistic_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/logistic.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/logistic.hpp>
+extern "C" __global__ 
+void test_logistic_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::logistic_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_logistic_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_logistic_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_logistic_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::logistic_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_logistic_pdf_nvrtc_float.cpp b/test/test_logistic_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..9339a6db3
--- /dev/null
+++ b/test/test_logistic_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/logistic.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/logistic.hpp>
+extern "C" __global__ 
+void test_logistic_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::logistic_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_logistic_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_logistic_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_logistic_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::logistic_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_logistic_quan_nvrtc_double.cpp b/test/test_logistic_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..f763db82c
--- /dev/null
+++ b/test/test_logistic_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/logistic.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/logistic.hpp>
+extern "C" __global__ 
+void test_logistic_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::logistic_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_logistic_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_logistic_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_logistic_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::logistic_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_logistic_quan_nvrtc_float.cpp b/test/test_logistic_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..b14c3c5da
--- /dev/null
+++ b/test/test_logistic_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/logistic.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/logistic.hpp>
+extern "C" __global__ 
+void test_logistic_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::logistic_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_logistic_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_logistic_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_logistic_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::logistic_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 4f2eb0218f259c59fb15b3fb6f9d5e3e7911f8bd Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:40:15 -0400
Subject: [PATCH 31/35] Disable whole file on NVRTC

---
 include/boost/math/tools/big_constant.hpp | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/include/boost/math/tools/big_constant.hpp b/include/boost/math/tools/big_constant.hpp
index 2fdfec782..0d54976bc 100644
--- a/include/boost/math/tools/big_constant.hpp
+++ b/include/boost/math/tools/big_constant.hpp
@@ -8,6 +8,12 @@
 #define BOOST_MATH_TOOLS_BIG_CONSTANT_HPP
 
 #include <boost/math/tools/config.hpp>
+
+// On NVRTC we don't need any of this
+// We just have a simple definition of the macro since the largest float
+// type on the platform is a 64-bit double
+#ifndef BOOST_MATH_HAS_NVRTC 
+
 #ifndef BOOST_MATH_STANDALONE
 #include <boost/lexical_cast.hpp>
 #endif
@@ -94,5 +100,7 @@ inline constexpr T make_big_value(largest_float, const char* s, std::false_type
 
 }}} // namespaces
 
+#endif // BOOST_MATH_HAS_NVRTC
+
 #endif
 

From 20e3a7ea87c7895243c8d7aec0876f78b93d7149 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:40:28 -0400
Subject: [PATCH 32/35] Add NVRTC support to mapairy dist

---
 include/boost/math/distributions/mapairy.hpp | 69 +++++++++++---------
 1 file changed, 37 insertions(+), 32 deletions(-)

diff --git a/include/boost/math/distributions/mapairy.hpp b/include/boost/math/distributions/mapairy.hpp
index 542cf6da3..8bf1f990c 100644
--- a/include/boost/math/distributions/mapairy.hpp
+++ b/include/boost/math/distributions/mapairy.hpp
@@ -15,7 +15,6 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
 #include <boost/math/tools/tuple.hpp>
-#include <boost/math/distributions/fwd.hpp>
 #include <boost/math/constants/constants.hpp>
 #include <boost/math/tools/big_constant.hpp>
 #include <boost/math/distributions/complement.hpp>
@@ -23,8 +22,14 @@
 #include <boost/math/distributions/detail/derived_accessors.hpp>
 #include <boost/math/tools/rational.hpp>
 #include <boost/math/special_functions/cbrt.hpp>
-#include <utility>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
+#include <boost/math/tools/promotion.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
 #include <cmath>
+#endif
 
 namespace boost { namespace math {
 template <class RealType, class Policy>
@@ -33,7 +38,7 @@ class mapairy_distribution;
 namespace detail {
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -255,7 +260,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_plus_imp_prec(const RealType&
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -584,7 +589,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_plus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -757,7 +762,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_minus_imp_prec(const RealType
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1048,7 +1053,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_minus_imp_prec(const RealType
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 53> &tag) {
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53> &tag) {
     if (x >= 0) {
         return mapairy_pdf_plus_imp_prec<RealType>(x, tag);
     }
@@ -1061,7 +1066,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_imp_prec(const RealType& x, c
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>& tag) {
     if (x >= 0) {
         return mapairy_pdf_plus_imp_prec<RealType>(x, tag);
     }
@@ -1100,7 +1105,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_imp(const mapairy_distributio
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -1116,7 +1121,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_pdf_imp(const mapairy_distributio
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1331,7 +1336,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_plus_imp_prec(const RealType&
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1656,7 +1661,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_plus_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1833,7 +1838,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_minus_imp_prec(const RealType
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_minus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_minus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2133,7 +2138,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_minus_imp_prec(const RealType
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 53>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 53>& tag) {
     if (x >= 0) {
         return complement ? mapairy_cdf_plus_imp_prec(x, tag) : 1 - mapairy_cdf_plus_imp_prec(x, tag);
     }
@@ -2146,7 +2151,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_imp_prec(const RealType& x, b
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 113>& tag) {
     if (x >= 0) {
         return complement ? mapairy_cdf_plus_imp_prec(x, tag) : 1 - mapairy_cdf_plus_imp_prec(x, tag);
     }
@@ -2185,7 +2190,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_imp(const mapairy_distributio
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -2201,7 +2206,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_cdf_imp(const mapairy_distributio
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_lower_imp_prec(const RealType& p, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_lower_imp_prec(const RealType& p, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2534,7 +2539,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_lower_imp_prec(const Rea
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_lower_imp_prec(const RealType& p, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_lower_imp_prec(const RealType& p, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -3236,7 +3241,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_lower_imp_prec(const Rea
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -3440,7 +3445,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_upper_imp_prec(const Rea
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -3838,7 +3843,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_upper_imp_prec(const Rea
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 53>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 53>& tag)
 {
     if (p > 0.5) {
         return !complement ? mapairy_quantile_upper_imp_prec(1 - p, tag) : mapairy_quantile_lower_imp_prec(1 - p, tag);
@@ -3848,7 +3853,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_imp_prec(const RealType&
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 113>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 113>& tag)
 {
     if (p > 0.5) {
         return !complement ? mapairy_quantile_upper_imp_prec(1 - p, tag) : mapairy_quantile_lower_imp_prec(1 - p, tag);
@@ -3885,7 +3890,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_imp(const mapairy_distri
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -3899,13 +3904,13 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_quantile_imp(const mapairy_distri
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_mode_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_mode_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(-1.16158727113597068525);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_mode_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_mode_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, -1.1615872711359706852500000803029112987);
 }
@@ -3933,7 +3938,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_mode_imp(const mapairy_distributi
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -3947,13 +3952,13 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_mode_imp(const mapairy_distributi
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_median_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_median_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(-0.71671068545502205332);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_median_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_median_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, -0.71671068545502205331700196278067230944440);
 }
@@ -3981,7 +3986,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_median_imp(const mapairy_distribu
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -3995,13 +4000,13 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_median_imp(const mapairy_distribu
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_entropy_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_entropy_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(2.00727681841065634600);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType mapairy_entropy_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType mapairy_entropy_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, 2.0072768184106563460003025875575283708);
 }
@@ -4024,7 +4029,7 @@ BOOST_MATH_GPU_ENABLED inline RealType mapairy_entropy_imp(const mapairy_distrib
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0

From cc1a9025c05ce2a924937129548309b5e8846b5a Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:40:47 -0400
Subject: [PATCH 33/35] Add mapairy dist NVRTC testing

---
 test/nvrtc_jamfile                      |   7 +
 test/test_mapairy_cdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++++
 test/test_mapairy_cdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++++
 test/test_mapairy_pdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++++
 test/test_mapairy_pdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++++
 test/test_mapairy_quan_nvrtc_double.cpp | 191 ++++++++++++++++++++++++
 test/test_mapairy_quan_nvrtc_float.cpp  | 191 ++++++++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_mapairy_cdf_nvrtc_double.cpp
 create mode 100644 test/test_mapairy_cdf_nvrtc_float.cpp
 create mode 100644 test/test_mapairy_pdf_nvrtc_double.cpp
 create mode 100644 test/test_mapairy_pdf_nvrtc_float.cpp
 create mode 100644 test/test_mapairy_quan_nvrtc_double.cpp
 create mode 100644 test/test_mapairy_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index c0af6e7fa..fa25d0226 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -73,6 +73,13 @@ run test_logistic_pdf_nvrtc_float.cpp ;
 run test_logistic_quan_nvrtc_double.cpp ;
 run test_logistic_quan_nvrtc_float.cpp ;
 
+run test_mapairy_cdf_nvrtc_double.cpp ;
+run test_mapairy_cdf_nvrtc_float.cpp ;
+run test_mapairy_pdf_nvrtc_double.cpp ;
+run test_mapairy_pdf_nvrtc_float.cpp ;
+run test_mapairy_quan_nvrtc_double.cpp ;
+run test_mapairy_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_mapairy_cdf_nvrtc_double.cpp b/test/test_mapairy_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..87e7948b7
--- /dev/null
+++ b/test/test_mapairy_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/mapairy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/mapairy.hpp>
+extern "C" __global__ 
+void test_mapairy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::mapairy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_mapairy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_mapairy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_mapairy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::mapairy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_mapairy_cdf_nvrtc_float.cpp b/test/test_mapairy_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..d84404c5e
--- /dev/null
+++ b/test/test_mapairy_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/mapairy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/mapairy.hpp>
+extern "C" __global__ 
+void test_mapairy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::mapairy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_mapairy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_mapairy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_mapairy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::mapairy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_mapairy_pdf_nvrtc_double.cpp b/test/test_mapairy_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..5f461d6d1
--- /dev/null
+++ b/test/test_mapairy_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/mapairy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/mapairy.hpp>
+extern "C" __global__ 
+void test_mapairy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::mapairy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_mapairy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_mapairy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_mapairy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::mapairy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_mapairy_pdf_nvrtc_float.cpp b/test/test_mapairy_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..39eb4152d
--- /dev/null
+++ b/test/test_mapairy_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/mapairy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/mapairy.hpp>
+extern "C" __global__ 
+void test_mapairy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::mapairy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_mapairy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_mapairy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_mapairy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::mapairy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_mapairy_quan_nvrtc_double.cpp b/test/test_mapairy_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..43ac17c84
--- /dev/null
+++ b/test/test_mapairy_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/mapairy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/mapairy.hpp>
+extern "C" __global__ 
+void test_mapairy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::mapairy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_mapairy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_mapairy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_mapairy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::mapairy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_mapairy_quan_nvrtc_float.cpp b/test/test_mapairy_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..a127843e9
--- /dev/null
+++ b/test/test_mapairy_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/mapairy.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/mapairy.hpp>
+extern "C" __global__ 
+void test_mapairy_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::mapairy_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_mapairy_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_mapairy_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_mapairy_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::mapairy_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}

From 64a92dd527c2bce76201f8a068b2ca7ab3328ff9 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:50:38 -0400
Subject: [PATCH 34/35] Add NVRTC support to saspoint5 dist

---
 .../boost/math/distributions/saspoint5.hpp    | 46 +++++++++++--------
 1 file changed, 26 insertions(+), 20 deletions(-)

diff --git a/include/boost/math/distributions/saspoint5.hpp b/include/boost/math/distributions/saspoint5.hpp
index 3e78fca97..7846b9956 100644
--- a/include/boost/math/distributions/saspoint5.hpp
+++ b/include/boost/math/distributions/saspoint5.hpp
@@ -15,15 +15,21 @@
 #include <boost/math/tools/config.hpp>
 #include <boost/math/tools/numeric_limits.hpp>
 #include <boost/math/tools/tuple.hpp>
-#include <boost/math/distributions/fwd.hpp>
 #include <boost/math/constants/constants.hpp>
 #include <boost/math/tools/big_constant.hpp>
 #include <boost/math/distributions/complement.hpp>
 #include <boost/math/distributions/detail/common_error_handling.hpp>
 #include <boost/math/distributions/detail/derived_accessors.hpp>
 #include <boost/math/tools/rational.hpp>
-#include <utility>
+#include <boost/math/tools/promotion.hpp>
+#include <boost/math/tools/type_traits.hpp>
+#include <boost/math/policies/policy.hpp>
+#include <boost/math/policies/error_handling.hpp>
+
+#ifndef BOOST_MATH_HAS_NVRTC
+#include <boost/math/distributions/fwd.hpp>
 #include <cmath>
+#endif
 
 namespace boost { namespace math {
 template <class RealType, class Policy>
@@ -32,7 +38,7 @@ class saspoint5_distribution;
 namespace detail {
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -337,7 +343,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_plus_imp_prec(const RealTyp
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -849,14 +855,14 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_plus_imp_prec(const RealTyp
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 53> &tag) {
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53> &tag) {
     BOOST_MATH_STD_USING // for ADL of std functions
 
     return saspoint5_pdf_plus_imp_prec<RealType>(abs(x), tag);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_imp_prec(const RealType& x, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>& tag) {
     BOOST_MATH_STD_USING // for ADL of std functions
 
     return saspoint5_pdf_plus_imp_prec<RealType>(abs(x), tag);
@@ -889,7 +895,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_imp(const saspoint5_distrib
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -905,7 +911,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_pdf_imp(const saspoint5_distrib
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1163,7 +1169,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_plus_imp_prec(const RealTyp
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_plus_imp_prec(const RealType& x, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_plus_imp_prec(const RealType& x, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1642,7 +1648,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_plus_imp_prec(const RealTyp
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 53>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 53>& tag) {
     if (x >= 0) {
         return complement ? saspoint5_cdf_plus_imp_prec(x, tag) : 1 - saspoint5_cdf_plus_imp_prec(x, tag);
     }
@@ -1655,7 +1661,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_imp_prec(const RealType& x,
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_imp_prec(const RealType& x, bool complement, const std::integral_constant<int, 113>& tag) {
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_imp_prec(const RealType& x, bool complement, const boost::math::integral_constant<int, 113>& tag) {
     if (x >= 0) {
         return complement ? saspoint5_cdf_plus_imp_prec(x, tag) : 1 - saspoint5_cdf_plus_imp_prec(x, tag);
     }
@@ -1694,7 +1700,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_imp(const saspoint5_distrib
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -1710,7 +1716,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_cdf_imp(const saspoint5_distrib
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 53>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -1994,7 +2000,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_upper_imp_prec(const R
 
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_upper_imp_prec(const RealType& p, const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_upper_imp_prec(const RealType& p, const boost::math::integral_constant<int, 113>&)
 {
     BOOST_MATH_STD_USING
     RealType result;
@@ -2493,7 +2499,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_upper_imp_prec(const R
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 53>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 53>& tag)
 {
     if (p > 0.5) {
         return !complement ? saspoint5_quantile_upper_imp_prec(1 - p, tag) : -saspoint5_quantile_upper_imp_prec(1 - p, tag);
@@ -2503,7 +2509,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_imp_prec(const RealTyp
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_imp_prec(const RealType& p, bool complement, const std::integral_constant<int, 113>& tag)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_imp_prec(const RealType& p, bool complement, const boost::math::integral_constant<int, 113>& tag)
 {
     if (p > 0.5) {
         return !complement ? saspoint5_quantile_upper_imp_prec(1 - p, tag) : -saspoint5_quantile_upper_imp_prec(1 - p, tag);
@@ -2540,7 +2546,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_imp(const saspoint5_di
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0
@@ -2554,13 +2560,13 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_quantile_imp(const saspoint5_di
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_entropy_imp_prec(const std::integral_constant<int, 53>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_entropy_imp_prec(const boost::math::integral_constant<int, 53>&)
 {
     return static_cast<RealType>(3.63992444568030649573);
 }
 
 template <class RealType>
-BOOST_MATH_GPU_ENABLED inline RealType saspoint5_entropy_imp_prec(const std::integral_constant<int, 113>&)
+BOOST_MATH_GPU_ENABLED inline RealType saspoint5_entropy_imp_prec(const boost::math::integral_constant<int, 113>&)
 {
     return BOOST_MATH_BIG_CONSTANT(RealType, 113, 3.6399244456803064957308496039071853510);
 }
@@ -2583,7 +2589,7 @@ BOOST_MATH_GPU_ENABLED inline RealType saspoint5_entropy_imp(const saspoint5_dis
 
     typedef typename tools::promote_args<RealType>::type result_type;
     typedef typename policies::precision<result_type, Policy>::type precision_type;
-    typedef std::integral_constant<int,
+    typedef boost::math::integral_constant<int,
         precision_type::value <= 0 ? 0 :
         precision_type::value <= 53 ? 53 :
         precision_type::value <= 113 ? 113 : 0

From c3bab84241709b2a2ae7e34115892f88186979c8 Mon Sep 17 00:00:00 2001
From: Matt Borland <matt@mattborland.com>
Date: Thu, 15 Aug 2024 10:50:50 -0400
Subject: [PATCH 35/35] Add saspoint5 NVRTC testing

---
 test/nvrtc_jamfile                        |   7 +
 test/test_saspoint5_cdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++
 test/test_saspoint5_cdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++
 test/test_saspoint5_pdf_nvrtc_double.cpp  | 191 ++++++++++++++++++++++
 test/test_saspoint5_pdf_nvrtc_float.cpp   | 191 ++++++++++++++++++++++
 test/test_saspoint5_quan_nvrtc_double.cpp | 191 ++++++++++++++++++++++
 test/test_saspoint5_quan_nvrtc_float.cpp  | 191 ++++++++++++++++++++++
 7 files changed, 1153 insertions(+)
 create mode 100644 test/test_saspoint5_cdf_nvrtc_double.cpp
 create mode 100644 test/test_saspoint5_cdf_nvrtc_float.cpp
 create mode 100644 test/test_saspoint5_pdf_nvrtc_double.cpp
 create mode 100644 test/test_saspoint5_pdf_nvrtc_float.cpp
 create mode 100644 test/test_saspoint5_quan_nvrtc_double.cpp
 create mode 100644 test/test_saspoint5_quan_nvrtc_float.cpp

diff --git a/test/nvrtc_jamfile b/test/nvrtc_jamfile
index fa25d0226..8d33eabb5 100644
--- a/test/nvrtc_jamfile
+++ b/test/nvrtc_jamfile
@@ -80,6 +80,13 @@ run test_mapairy_pdf_nvrtc_float.cpp ;
 run test_mapairy_quan_nvrtc_double.cpp ;
 run test_mapairy_quan_nvrtc_float.cpp ;
 
+run test_saspoint5_cdf_nvrtc_double.cpp ;
+run test_saspoint5_cdf_nvrtc_float.cpp ;
+run test_saspoint5_pdf_nvrtc_double.cpp ;
+run test_saspoint5_pdf_nvrtc_float.cpp ;
+run test_saspoint5_quan_nvrtc_double.cpp ;
+run test_saspoint5_quan_nvrtc_float.cpp ;
+
 # Special Functions
 run test_cbrt_nvrtc_double.cpp ;
 run test_cbrt_nvrtc_float.cpp ;
diff --git a/test/test_saspoint5_cdf_nvrtc_double.cpp b/test/test_saspoint5_cdf_nvrtc_double.cpp
new file mode 100644
index 000000000..ff2067fa0
--- /dev/null
+++ b/test/test_saspoint5_cdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/saspoint5.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/saspoint5.hpp>
+extern "C" __global__ 
+void test_saspoint5_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::saspoint5_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_saspoint5_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_saspoint5_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_saspoint5_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::saspoint5_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_saspoint5_cdf_nvrtc_float.cpp b/test/test_saspoint5_cdf_nvrtc_float.cpp
new file mode 100644
index 000000000..11c816da1
--- /dev/null
+++ b/test/test_saspoint5_cdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/saspoint5.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/saspoint5.hpp>
+extern "C" __global__ 
+void test_saspoint5_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = cdf(boost::math::saspoint5_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_saspoint5_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_saspoint5_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_saspoint5_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = cdf(boost::math::saspoint5_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_saspoint5_pdf_nvrtc_double.cpp b/test/test_saspoint5_pdf_nvrtc_double.cpp
new file mode 100644
index 000000000..4c7444383
--- /dev/null
+++ b/test/test_saspoint5_pdf_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/saspoint5.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/saspoint5.hpp>
+extern "C" __global__ 
+void test_saspoint5_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::saspoint5_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_saspoint5_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_saspoint5_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_saspoint5_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::saspoint5_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_saspoint5_pdf_nvrtc_float.cpp b/test/test_saspoint5_pdf_nvrtc_float.cpp
new file mode 100644
index 000000000..8cd93aaa9
--- /dev/null
+++ b/test/test_saspoint5_pdf_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/saspoint5.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/saspoint5.hpp>
+extern "C" __global__ 
+void test_saspoint5_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = pdf(boost::math::saspoint5_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_saspoint5_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_saspoint5_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_saspoint5_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = pdf(boost::math::saspoint5_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_saspoint5_quan_nvrtc_double.cpp b/test/test_saspoint5_quan_nvrtc_double.cpp
new file mode 100644
index 000000000..631ac4b24
--- /dev/null
+++ b/test/test_saspoint5_quan_nvrtc_double.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/saspoint5.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef double float_type;
+
+const char* cuda_kernel = R"(
+typedef double float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/saspoint5.hpp>
+extern "C" __global__ 
+void test_saspoint5_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::saspoint5_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_saspoint5_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_saspoint5_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_saspoint5_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::saspoint5_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/test/test_saspoint5_quan_nvrtc_float.cpp b/test/test_saspoint5_quan_nvrtc_float.cpp
new file mode 100644
index 000000000..fa152622e
--- /dev/null
+++ b/test/test_saspoint5_quan_nvrtc_float.cpp
@@ -0,0 +1,191 @@
+//  Copyright John Maddock 2016.
+//  Copyright Matt Borland 2024.
+//  Use, modification and distribution are subject to the
+//  Boost Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+#define BOOST_MATH_OVERFLOW_ERROR_POLICY ignore_error
+#define BOOST_MATH_PROMOTE_DOUBLE_POLICY false
+
+// Must be included first
+#include <nvrtc.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#include <iostream>
+#include <iomanip>
+#include <vector>
+#include <random>
+#include <exception>
+#include <boost/math/distributions/saspoint5.hpp>
+#include <boost/math/special_functions/fpclassify.hpp>
+#include <boost/math/special_functions/relative_difference.hpp>
+
+typedef float float_type;
+
+const char* cuda_kernel = R"(
+typedef float float_type;
+#include <cuda/std/type_traits>
+#include <boost/math/distributions/saspoint5.hpp>
+extern "C" __global__ 
+void test_saspoint5_kernel(const float_type *in1, const float_type*, float_type *out, int numElements)
+{
+    int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i < numElements)
+    {
+        out[i] = quantile(boost::math::saspoint5_distribution<float_type>(), in1[i]);
+    }
+}
+)";
+
+void checkCUDAError(cudaError_t result, const char* msg)
+{
+    if (result != cudaSuccess)
+    {
+        std::cerr << msg << ": " << cudaGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkCUError(CUresult result, const char* msg)
+{
+    if (result != CUDA_SUCCESS)
+    {
+        const char* errorStr;
+        cuGetErrorString(result, &errorStr);
+        std::cerr << msg << ": " << errorStr << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+void checkNVRTCError(nvrtcResult result, const char* msg)
+{
+    if (result != NVRTC_SUCCESS)
+    {
+        std::cerr << msg << ": " << nvrtcGetErrorString(result) << std::endl;
+        exit(EXIT_FAILURE);
+    }
+}
+
+int main() 
+{
+    try
+    {
+        // Initialize CUDA driver API
+        checkCUError(cuInit(0), "Failed to initialize CUDA");
+
+        // Create CUDA context
+        CUcontext context;
+        CUdevice device;
+        checkCUError(cuDeviceGet(&device, 0), "Failed to get CUDA device");
+        checkCUError(cuCtxCreate(&context, 0, device), "Failed to create CUDA context");
+
+        nvrtcProgram prog;
+        nvrtcResult res;
+
+        res = nvrtcCreateProgram(&prog, cuda_kernel, "test_saspoint5_kernel.cu", 0, nullptr, nullptr);
+        checkNVRTCError(res, "Failed to create NVRTC program");
+
+        nvrtcAddNameExpression(prog, "test_saspoint5_kernel");
+
+        #ifdef BOOST_MATH_NVRTC_CI_RUN
+        const char* opts[] = {"--std=c++14", "--gpu-architecture=compute_75", "--include-path=/home/runner/work/cuda-math/boost-root/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #else
+        const char* opts[] = {"--std=c++14", "--include-path=/home/mborland/Documents/boost/libs/cuda-math/include/", "-I/usr/local/cuda/include"};
+        #endif
+
+        // Compile the program
+        res = nvrtcCompileProgram(prog, sizeof(opts) / sizeof(const char*), opts);
+        if (res != NVRTC_SUCCESS) 
+        {
+            size_t log_size;
+            nvrtcGetProgramLogSize(prog, &log_size);
+            char* log = new char[log_size];
+            nvrtcGetProgramLog(prog, log);
+            std::cerr << "Compilation failed:\n" << log << std::endl;
+            delete[] log;
+            exit(EXIT_FAILURE);
+        }
+
+        // Get PTX from the program
+        size_t ptx_size;
+        nvrtcGetPTXSize(prog, &ptx_size);
+        char* ptx = new char[ptx_size];
+        nvrtcGetPTX(prog, ptx);
+
+        // Load PTX into CUDA module
+        CUmodule module;
+        CUfunction kernel;
+        checkCUError(cuModuleLoadDataEx(&module, ptx, 0, 0, 0), "Failed to load module");
+        checkCUError(cuModuleGetFunction(&kernel, module, "test_saspoint5_kernel"), "Failed to get kernel function");
+
+        int numElements = 5000;
+        float_type *h_in1, *h_in2, *h_out;
+        float_type *d_in1, *d_in2, *d_out;
+
+        // Allocate memory on the host
+        h_in1 = new float_type[numElements];
+        h_in2 = new float_type[numElements];
+        h_out = new float_type[numElements];
+
+        // Initialize input arrays
+        std::mt19937_64 rng(42);
+        std::uniform_real_distribution<float_type> dist(0.0f, 1.0f);
+        for (int i = 0; i < numElements; ++i) 
+        {
+            h_in1[i] = static_cast<float_type>(dist(rng));
+            h_in2[i] = static_cast<float_type>(dist(rng));
+        }
+
+        checkCUDAError(cudaMalloc(&d_in1, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in1");
+        checkCUDAError(cudaMalloc(&d_in2, numElements * sizeof(float_type)), "Failed to allocate device memory for d_in2");
+        checkCUDAError(cudaMalloc(&d_out, numElements * sizeof(float_type)), "Failed to allocate device memory for d_out");
+
+        checkCUDAError(cudaMemcpy(d_in1, h_in1, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in1");
+        checkCUDAError(cudaMemcpy(d_in2, h_in2, numElements * sizeof(float_type), cudaMemcpyHostToDevice), "Failed to copy data to device for d_in2");
+
+        int blockSize = 256;
+        int numBlocks = (numElements + blockSize - 1) / blockSize;
+        void* args[] = { &d_in1, &d_in2, &d_out, &numElements };
+        checkCUError(cuLaunchKernel(kernel, numBlocks, 1, 1, blockSize, 1, 1, 0, 0, args, 0), "Kernel launch failed");
+
+        checkCUDAError(cudaMemcpy(h_out, d_out, numElements * sizeof(float_type), cudaMemcpyDeviceToHost), "Failed to copy data back to host for h_out");
+
+        // Verify Result
+        for (int i = 0; i < numElements; ++i) 
+        {
+            auto res = quantile(boost::math::saspoint5_distribution<float_type>(), h_in1[i]);
+            
+            if (boost::math::isfinite(res))
+            {
+                if (boost::math::epsilon_difference(res, h_out[i]) > 300)
+                {
+                    std::cout << "error at line: " << i
+                            << "\nParallel: " << h_out[i]
+                            << "\n  Serial: " << res
+                            << "\n    Dist: " << boost::math::epsilon_difference(res, h_out[i]) << std::endl;
+                }
+            }
+        }
+
+        cudaFree(d_in1);
+        cudaFree(d_in2);
+        cudaFree(d_out);
+        delete[] h_in1;
+        delete[] h_in2;
+        delete[] h_out;
+
+        nvrtcDestroyProgram(&prog);
+        delete[] ptx;
+
+        cuCtxDestroy(context);
+
+        std::cout << "Kernel executed successfully." << std::endl;
+        return 0;
+    }
+    catch(const std::exception& e)
+    {
+        std::cerr << "Stopped with exception: " << e.what() << std::endl;
+        return EXIT_FAILURE;
+    }
+}