cuda-cccl 0.1.3.2.0.dev438__cp313-cp313-manylinux_2_24_aarch64.whl → 0.3.1__cp313-cp313-manylinux_2_24_aarch64.whl

cuda/cccl/headers/include/cuda/__stream/stream.h

@@ -24,9 +24,9 @@
 #if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
 
 #  include <cuda/__device/device_ref.h>
+#  include <cuda/__driver/driver_api.h>
 #  include <cuda/__runtime/ensure_current_context.h>
 #  include <cuda/__stream/stream_ref.h> // IWYU pragma: export
-#  include <cuda/std/__cuda/api_wrapper.h>
 
 #  include <cuda/std/__cccl/prologue.h>
 
@@ -47,8 +47,7 @@ struct stream : stream_ref
       : stream_ref(__detail::__invalid_stream)
   {
     [[maybe_unused]] __ensure_current_context __ctx_setter(__dev);
-    _CCCL_TRY_CUDA_API(
-      ::cudaStreamCreateWithPriority, "Failed to create a stream", &__stream, cudaStreamNonBlocking, __priority);
+    __stream = ::cuda::__driver::__streamCreateWithPriority(cudaStreamNonBlocking, __priority);
   }
 
   //! @brief Construct a new `stream` object into the moved-from state.

cuda/cccl/headers/include/cuda/__stream/stream_ref.h

@@ -23,11 +23,12 @@
 
 #if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
 
+#  include <cuda/__device/device_ref.h>
+#  include <cuda/__driver/driver_api.h>
 #  include <cuda/__event/timed_event.h>
 #  include <cuda/__fwd/get_stream.h>
 #  include <cuda/__runtime/ensure_current_context.h>
 #  include <cuda/__utility/no_init.h>
-#  include <cuda/std/__cuda/api_wrapper.h>
 #  include <cuda/std/__exception/cuda_error.h>
 #  include <cuda/std/cstddef>
 
@@ -39,7 +40,7 @@ namespace __detail
 {
 // 0 is a valid stream in CUDA, so we need some other invalid stream representation
 // Can't make it constexpr, because cudaStream_t is a pointer type
-static const ::cudaStream_t __invalid_stream = reinterpret_cast<cudaStream_t>(~0ULL);
+static const ::cudaStream_t __invalid_stream = reinterpret_cast<::cudaStream_t>(~0ULL);
 } // namespace __detail
 
 //! @brief A type representing a stream ID.
@@ -238,11 +239,17 @@ public:
   //! @throws cuda_error if device check fails
   _CCCL_HOST_API device_ref device() const
   {
-    CUcontext __stream_ctx = ::cuda::__driver::__streamGetCtx(__stream);
-    __ensure_current_context __setter(__stream_ctx);
-    int __id;
-    _CCCL_TRY_CUDA_API(cudaGetDevice, "Could not get device from a stream", &__id);
-    return device_ref{__id};
+    ::CUdevice __device{};
+#  if _CCCL_CTK_AT_LEAST(13, 0)
+    __device = ::cuda::__driver::__streamGetDevice(__stream);
+#  else // ^^^ _CCCL_CTK_AT_LEAST(13, 0) ^^^ / vvv _CCCL_CTK_BELOW(13, 0) vvv
+    {
+      ::CUcontext __stream_ctx = ::cuda::__driver::__streamGetCtx(__stream);
+      __ensure_current_context __setter(__stream_ctx);
+      __device = ::cuda::__driver::__ctxGetDevice();
+    }
+#  endif // ^^^ _CCCL_CTK_BELOW(13, 0) ^^^
+    return device_ref{::cuda::__driver::__cudevice_to_ordinal(__device)};
   }
 
   //! @brief Queries the \c stream_ref for itself. This makes \c stream_ref usable in places where we expect an
@@ -262,21 +269,20 @@ inline void event_ref::record(stream_ref __stream) const
 }
 
 inline event::event(stream_ref __stream, event::flags __flags)
-    : event(__stream, static_cast<unsigned int>(__flags) | cudaEventDisableTiming)
+    : event(__stream, static_cast<unsigned>(__flags) | cudaEventDisableTiming)
 {
   record(__stream);
 }
 
-inline event::event(stream_ref __stream, unsigned int __flags)
+inline event::event(stream_ref __stream, unsigned __flags)
     : event_ref(::cudaEvent_t{})
 {
   [[maybe_unused]] __ensure_current_context __ctx_setter(__stream);
-  _CCCL_TRY_CUDA_API(
-    ::cudaEventCreateWithFlags, "Failed to create CUDA event", &__event_, static_cast<unsigned int>(__flags));
+  __event_ = ::cuda::__driver::__eventCreate(static_cast<unsigned>(__flags));
 }
 
 inline timed_event::timed_event(stream_ref __stream, event::flags __flags)
-    : event(__stream, static_cast<unsigned int>(__flags))
+    : event(__stream, static_cast<unsigned>(__flags))
 {
   record(__stream);
 }

cuda/cccl/headers/include/cuda/__utility/__basic_any/virtual_tables.h

@@ -42,7 +42,7 @@ using __vtable_for _CCCL_NODEBUG_ALIAS = typename __overrides_for_t<_Interface>:
 //! __basic_vtable
 //!
 template <class _Interface, auto... _Mbrs>
-struct _CCCL_DECLSPEC_EMPTY_BASES __basic_vtable
+struct _CCCL_DECLSPEC_EMPTY_BASES _CCCL_TYPE_VISIBILITY_DEFAULT __basic_vtable
     : __rtti_base
     , __virtual_fn<_Mbrs>...
 {
@@ -105,7 +105,7 @@ struct _CCCL_DECLSPEC_EMPTY_BASES __basic_vtable
 //!
 
 template <class... _Interfaces>
-struct _CCCL_DECLSPEC_EMPTY_BASES __vtable_tuple
+struct _CCCL_DECLSPEC_EMPTY_BASES _CCCL_TYPE_VISIBILITY_DEFAULT __vtable_tuple
     : __rtti_ex<sizeof...(_Interfaces)>
     , __vtable_for<_Interfaces>...
 {

cuda/cccl/headers/include/cuda/__utility/basic_any.h

@@ -11,7 +11,7 @@
 #ifndef _CUDA___UTILITY_BASIC_ANY_H
 #define _CUDA___UTILITY_BASIC_ANY_H
 
-#include <cuda/__cccl_config>
+#include <cuda/std/detail/__config>
 
 #if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
 #  pragma GCC system_header

cuda/cccl/headers/include/cuda/algorithm

@@ -11,7 +11,7 @@
 #ifndef _CUDA_ALGORITHM
 #define _CUDA_ALGORITHM
 
-#include <cuda/__cccl_config>
+#include <cuda/std/detail/__config>
 
 #if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
 #  pragma GCC system_header

cuda/cccl/headers/include/cuda/devices

@@ -11,6 +11,16 @@
 #ifndef _CUDA_DEVICES
 #define _CUDA_DEVICES
 
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
 #include <cuda/__device/all_devices.h>
 #include <cuda/__device/arch_traits.h>
 #include <cuda/__device/attributes.h>

cuda/cccl/headers/include/cuda/iterator

@@ -33,6 +33,7 @@
 #include <cuda/__iterator/transform_output_iterator.h>
 #include <cuda/__iterator/zip_function.h>
 #include <cuda/__iterator/zip_iterator.h>
+#include <cuda/__iterator/zip_transform_iterator.h>
 #include <cuda/std/iterator>
 
 #endif // _CUDA_ITERATOR

cuda/cccl/headers/include/cuda/std/__bit/countl.h

@@ -100,7 +100,14 @@ template <typename _Tp>
 template <typename _Tp>
 [[nodiscard]] _CCCL_HIDE_FROM_ABI _CCCL_DEVICE int __cccl_countl_zero_impl_device(_Tp __v) noexcept
 {
-  return (sizeof(_Tp) == sizeof(uint32_t)) ? ::__clz(static_cast<int>(__v)) : ::__clzll(static_cast<long long>(__v));
+  if constexpr (sizeof(_Tp) == sizeof(uint32_t))
+  {
+    return static_cast<int>(::__clz(static_cast<int>(__v)));
+  }
+  else
+  {
+    return static_cast<int>(::__clzll(static_cast<long long>(__v)));
+  }
 }
 #endif // _CCCL_CUDA_COMPILATION()
 

cuda/cccl/headers/include/cuda/std/__bit/countr.h

@@ -114,11 +114,11 @@ template <typename _Tp>
 {
   if constexpr (sizeof(_Tp) == sizeof(uint32_t))
   {
-    return ::__clz(static_cast<int>(::__brev(__v)));
+    return static_cast<int>(::__clz(static_cast<int>(::__brev(__v))));
   }
   else
   {
-    return ::__clzll(static_cast<long long>(::__brevll(__v)));
+    return static_cast<int>(::__clzll(static_cast<long long>(::__brevll(__v))));
   }
 }
 #endif // _CCCL_CUDA_COMPILATION()

cuda/cccl/headers/include/cuda/std/__bit/reference.h

@@ -275,10 +275,10 @@ _CCCL_API constexpr __bit_iterator<_Cp, false> __copy_aligned(
     // do first word
     if (__first.__ctz_ != 0)
     {
-      unsigned __clz       = __bits_per_word - __first.__ctz_;
-      difference_type __dn = ::cuda::std::min(static_cast<difference_type>(__clz), __n);
+      unsigned __clz_f     = __bits_per_word - __first.__ctz_;
+      difference_type __dn = ::cuda::std::min(static_cast<difference_type>(__clz_f), __n);
       __n -= __dn;
-      __storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz - __dn));
+      __storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
       __storage_type __b = *__first.__seg_ & __m;
       *__result.__seg_ &= ~__m;
       *__result.__seg_ |= __b;
@@ -420,8 +420,8 @@ _CCCL_API constexpr __bit_iterator<_Cp, false> __copy_backward_aligned(
     {
       difference_type __dn = ::cuda::std::min(static_cast<difference_type>(__last.__ctz_), __n);
       __n -= __dn;
-      unsigned __clz     = __bits_per_word - __last.__ctz_;
-      __storage_type __m = (~__storage_type(0) << (__last.__ctz_ - __dn)) & (~__storage_type(0) >> __clz);
+      unsigned __clz_f   = __bits_per_word - __last.__ctz_;
+      __storage_type __m = (~__storage_type(0) << (__last.__ctz_ - __dn)) & (~__storage_type(0) >> __clz_f);
       __storage_type __b = *__last.__seg_ & __m;
       *__result.__seg_ &= ~__m;
       *__result.__seg_ |= __b;
@@ -635,10 +635,10 @@ _CCCL_API inline __bit_iterator<_Cr, false> __swap_ranges_aligned(
     // do first word
     if (__first.__ctz_ != 0)
     {
-      unsigned __clz       = __bits_per_word - __first.__ctz_;
-      difference_type __dn = ::cuda::std::min(static_cast<difference_type>(__clz), __n);
+      unsigned __clz_f     = __bits_per_word - __first.__ctz_;
+      difference_type __dn = ::cuda::std::min(static_cast<difference_type>(__clz_f), __n);
       __n -= __dn;
-      __storage_type __m  = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz - __dn));
+      __storage_type __m  = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
       __storage_type __b1 = *__first.__seg_ & __m;
       *__first.__seg_ &= ~__m;
       __storage_type __b2 = *__result.__seg_ & __m;
@@ -988,10 +988,10 @@ _CCCL_API constexpr bool __equal_aligned(
     // do first word
     if (__first1.__ctz_ != 0)
     {
-      unsigned __clz       = __bits_per_word - __first1.__ctz_;
-      difference_type __dn = ::cuda::std::min(static_cast<difference_type>(__clz), __n);
+      unsigned __clz_f     = __bits_per_word - __first1.__ctz_;
+      difference_type __dn = ::cuda::std::min(static_cast<difference_type>(__clz_f), __n);
       __n -= __dn;
-      __storage_type __m = (~__storage_type(0) << __first1.__ctz_) & (~__storage_type(0) >> (__clz - __dn));
+      __storage_type __m = (~__storage_type(0) << __first1.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
       if ((*__first2.__seg_ & __m) != (*__first1.__seg_ & __m))
       {
         return false;

cuda/cccl/headers/include/cuda/std/__cccl/cuda_capabilities.h

@@ -39,10 +39,10 @@
 #if _CCCL_HAS_PDL()
 // Waits for the previous kernel to complete (when it reaches its final membar). Should be put before the first global
 // memory access in a kernel.
-#  define _CCCL_PDL_GRID_DEPENDENCY_SYNC() NV_IF_TARGET(NV_PROVIDES_SM_90, cudaGridDependencySynchronize();)
+#  define _CCCL_PDL_GRID_DEPENDENCY_SYNC() NV_IF_TARGET(NV_PROVIDES_SM_90, ::cudaGridDependencySynchronize();)
 // Allows the subsequent kernel in the same stream to launch. Can be put anywhere in a kernel.
 // Heuristic(ahendriksen): put it after the last load.
-#  define _CCCL_PDL_TRIGGER_NEXT_LAUNCH() NV_IF_TARGET(NV_PROVIDES_SM_90, cudaTriggerProgrammaticLaunchCompletion();)
+#  define _CCCL_PDL_TRIGGER_NEXT_LAUNCH() NV_IF_TARGET(NV_PROVIDES_SM_90, ::cudaTriggerProgrammaticLaunchCompletion();)
 #else // _CCCL_HAS_PDL()
 #  define _CCCL_PDL_GRID_DEPENDENCY_SYNC()
 #  define _CCCL_PDL_TRIGGER_NEXT_LAUNCH()

cuda/cccl/headers/include/cuda/std/__cccl/preprocessor.h

@@ -107,6 +107,8 @@
 #define _CCCL_PP_FOR_EACH_7(_Mp, _1, _2, _3, _4, _5, _6, _7) _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7)
 #define _CCCL_PP_FOR_EACH_8(_Mp, _1, _2, _3, _4, _5, _6, _7, _8) \
   _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8)
+#define _CCCL_PP_FOR_EACH_9(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9)
 
 #define _CCCL_PP_PROBE_EMPTY_PROBE__CCCL_PP_PROBE_EMPTY _CCCL_PP_PROBE(~)
 

cuda/cccl/headers/include/cuda/std/__chrono/duration.h

@@ -43,19 +43,19 @@ template <class _Rep, class _Period = ratio<1>>
 class _CCCL_TYPE_VISIBILITY_DEFAULT duration;
 
 template <class _Tp>
-inline const bool __is_duration_v = false;
+inline constexpr bool __is_duration_v = false;
 
 template <class _Rep, class _Period>
-inline const bool __is_duration_v<duration<_Rep, _Period>> = true;
+inline constexpr bool __is_duration_v<duration<_Rep, _Period>> = true;
 
 template <class _Rep, class _Period>
-inline const bool __is_duration_v<const duration<_Rep, _Period>> = true;
+inline constexpr bool __is_duration_v<const duration<_Rep, _Period>> = true;
 
 template <class _Rep, class _Period>
-inline const bool __is_duration_v<volatile duration<_Rep, _Period>> = true;
+inline constexpr bool __is_duration_v<volatile duration<_Rep, _Period>> = true;
 
 template <class _Rep, class _Period>
-inline const bool __is_duration_v<const volatile duration<_Rep, _Period>> = true;
+inline constexpr bool __is_duration_v<const volatile duration<_Rep, _Period>> = true;
 
 } // namespace chrono
 
@@ -190,29 +190,29 @@ class _CCCL_TYPE_VISIBILITY_DEFAULT duration
   struct __no_overflow
   {
   private:
-    static const intmax_t __gcd_n1_n2 = __static_gcd<_R1::num, _R2::num>::value;
-    static const intmax_t __gcd_d1_d2 = __static_gcd<_R1::den, _R2::den>::value;
-    static const intmax_t __n1        = _R1::num / __gcd_n1_n2;
-    static const intmax_t __d1        = _R1::den / __gcd_d1_d2;
-    static const intmax_t __n2        = _R2::num / __gcd_n1_n2;
-    static const intmax_t __d2        = _R2::den / __gcd_d1_d2;
-    static const intmax_t max         = -((intmax_t(1) << (sizeof(intmax_t) * CHAR_BIT - 1)) + 1);
+    static constexpr intmax_t __gcd_n1_n2 = __static_gcd<_R1::num, _R2::num>::value;
+    static constexpr intmax_t __gcd_d1_d2 = __static_gcd<_R1::den, _R2::den>::value;
+    static constexpr intmax_t __n1        = _R1::num / __gcd_n1_n2;
+    static constexpr intmax_t __d1        = _R1::den / __gcd_d1_d2;
+    static constexpr intmax_t __n2        = _R2::num / __gcd_n1_n2;
+    static constexpr intmax_t __d2        = _R2::den / __gcd_d1_d2;
+    static constexpr intmax_t max         = -((intmax_t(1) << (sizeof(intmax_t) * CHAR_BIT - 1)) + 1);
 
     template <intmax_t _Xp, intmax_t _Yp, bool __overflow>
     struct __mul // __overflow == false
     {
-      static const intmax_t value = _Xp * _Yp;
+      static constexpr intmax_t value = _Xp * _Yp;
     };
 
     template <intmax_t _Xp, intmax_t _Yp>
     struct __mul<_Xp, _Yp, true>
     {
-      static const intmax_t value = 1;
+      static constexpr intmax_t value = 1;
     };
 
   public:
-    static const bool value = (__n1 <= max / __d2) && (__n2 <= max / __d1);
-    using type              = ratio<__mul<__n1, __d2, !value>::value, __mul<__n2, __d1, !value>::value>;
+    static constexpr bool value = (__n1 <= max / __d2) && (__n2 <= max / __d1);
+    using type                  = ratio<__mul<__n1, __d2, !value>::value, __mul<__n2, __d1, !value>::value>;
   };
 
 public:

cuda/cccl/headers/include/cuda/std/__chrono/steady_clock.h

@@ -40,11 +40,11 @@ namespace chrono
 class _CCCL_TYPE_VISIBILITY_DEFAULT steady_clock
 {
 public:
-  using duration                        = nanoseconds;
-  using rep                             = duration::rep;
-  using period                          = duration::period;
-  using time_point                      = ::cuda::std::chrono::time_point<steady_clock, duration>;
-  static constexpr const bool is_steady = true;
+  using duration                  = nanoseconds;
+  using rep                       = duration::rep;
+  using period                    = duration::period;
+  using time_point                = ::cuda::std::chrono::time_point<steady_clock, duration>;
+  static constexpr bool is_steady = true;
 
   [[nodiscard]] _CCCL_API static time_point now() noexcept;
 };

cuda/cccl/headers/include/cuda/std/__chrono/system_clock.h

@@ -39,11 +39,11 @@ namespace chrono
 class _CCCL_TYPE_VISIBILITY_DEFAULT system_clock
 {
 public:
-  using duration                        = ::cuda::std::chrono::nanoseconds;
-  using rep                             = duration::rep;
-  using period                          = duration::period;
-  using time_point                      = ::cuda::std::chrono::time_point<system_clock>;
-  static constexpr const bool is_steady = false;
+  using duration                  = ::cuda::std::chrono::nanoseconds;
+  using rep                       = duration::rep;
+  using period                    = duration::period;
+  using time_point                = ::cuda::std::chrono::time_point<system_clock>;
+  static constexpr bool is_steady = false;
 
   [[nodiscard]] _CCCL_API inline static time_point now() noexcept
   {

cuda/cccl/headers/include/cuda/std/__cmath/isnan.h

@@ -21,16 +21,15 @@
 #  pragma system_header
 #endif // no system header
 
-#include <cuda/std/__bit/popcount.h>
 #include <cuda/std/__concepts/concept_macros.h>
 #include <cuda/std/__floating_point/fp.h>
 #include <cuda/std/__type_traits/is_constant_evaluated.h>
+#include <cuda/std/__type_traits/is_floating_point.h>
 #include <cuda/std/__type_traits/is_integral.h>
 
-// MSVC and clang cuda need the host side functions included
-#if _CCCL_COMPILER(MSVC) || _CCCL_CUDA_COMPILER(CLANG)
+#if _CCCL_COMPILER(MSVC) || _CCCL_CUDA_COMPILER(CLANG) || !_CCCL_CUDA_COMPILATION()
 #  include <math.h>
-#endif // _CCCL_COMPILER(MSVC) || _CCCL_CUDA_COMPILER(CLANG)
+#endif // _CCCL_COMPILER(MSVC) || _CCCL_CUDA_COMPILER(CLANG) || !_CCCL_CUDA_COMPILATION()
 
 #include <cuda/std/__cccl/prologue.h>
 
@@ -158,10 +157,16 @@ template <class _Tp>
 #if _CCCL_HAS_FLOAT128()
 [[nodiscard]] _CCCL_API constexpr bool isnan(__float128 __x) noexcept
 {
+  // __builtin_isnan is not efficient for __float128, prefer __nv_fp128_isnan at run-time
+  if (!::cuda::std::__cccl_default_is_constant_evaluated())
+  {
+    NV_IF_TARGET(NV_PROVIDES_SM_100, (return ::__nv_fp128_isnan(__x);)) // preserve NaN behavior even with optimization
+                                                                        // flags
+  }
 #  if defined(_CCCL_BUILTIN_ISNAN)
   return _CCCL_BUILTIN_ISNAN(__x);
 #  else // ^^^ _CCCL_BUILTIN_ISNAN ^^^ / vvv !_CCCL_BUILTIN_ISNAN vvv
-  return ::cuda::std::__isnan_impl(__x);
+  return __x != __x;
 #  endif // ^^^ !_CCCL_BUILTIN_ISNAN ^^^
 }
 #endif // _CCCL_HAS_FLOAT128()

cuda/cccl/headers/include/cuda/std/__cmath/min_max.h

@@ -24,11 +24,11 @@
 #include <cuda/__type_traits/is_floating_point.h>
 #include <cuda/std/__cmath/isnan.h>
 #include <cuda/std/__concepts/concept_macros.h>
-#include <cuda/std/__floating_point/fp.h>
+#include <cuda/std/__type_traits/conditional.h>
 #include <cuda/std/__type_traits/is_extended_arithmetic.h>
 #include <cuda/std/__type_traits/is_integral.h>
+#include <cuda/std/__type_traits/is_same.h>
 #include <cuda/std/__type_traits/promote.h>
-#include <cuda/std/limits>
 
 #include <nv/target>
 
@@ -36,6 +36,10 @@
 
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
+/***********************************************************************************************************************
+ * fmax
+ **********************************************************************************************************************/
+
 // We do explicitly also enable GCC here, because that makes the condition below simpler
 #if _CCCL_CHECK_BUILTIN(builtin_fmax) || _CCCL_COMPILER(GCC)
 _CCCL_TEMPLATE(class _Tp)
@@ -63,13 +67,12 @@ _CCCL_REQUIRES(is_floating_point_v<_Tp>)
 #  define _CCCL_USE_BUILTIN_FMAX() 0
 #endif // _CCCL_BUILTIN_FABSF
 
-// fmax
 _CCCL_TEMPLATE(class _Tp)
 _CCCL_REQUIRES(__is_extended_arithmetic_v<_Tp>)
 [[nodiscard]] _CCCL_API constexpr conditional_t<is_integral_v<_Tp>, double, _Tp> fmax(_Tp __x, _Tp __y) noexcept
 {
 #if _CCCL_HAS_NVFP16()
-  if constexpr (is_same_v<_Tp, __half>)
+  if constexpr (is_same_v<_Tp, ::__half>)
   {
 #  if _CCCL_CTK_AT_LEAST(12, 2)
     return ::__hmax(__x, __y);
@@ -82,7 +85,7 @@ _CCCL_REQUIRES(__is_extended_arithmetic_v<_Tp>)
   else
 #endif // _CCCL_HAS_NVFP16()
 #if _CCCL_HAS_NVBF16()
-    if constexpr (is_same_v<_Tp, __nv_bfloat16>)
+    if constexpr (is_same_v<_Tp, ::__nv_bfloat16>)
   {
 #  if _CCCL_CTK_AT_LEAST(12, 2)
     return ::__hmax(__x, __y);
@@ -100,17 +103,27 @@ _CCCL_REQUIRES(__is_extended_arithmetic_v<_Tp>)
     }
     else
     {
-#if _CCCL_USE_BUILTIN_FMAX()
       if (!::cuda::std::__cccl_default_is_constant_evaluated())
       {
+#if _CCCL_HAS_FLOAT128()
+        if constexpr (is_same_v<_Tp, __float128>)
+        {
+          NV_IF_TARGET(NV_PROVIDES_SM_100, (return ::__nv_fp128_fmax(__x, __y);))
+        }
+        else
+#endif // _CCCL_HAS_FLOAT128()
+#if _CCCL_USE_BUILTIN_FMAX()
+          if constexpr (is_floating_point_v<_Tp>)
+        {
 // GCC builtins do not treat NaN properly
 #  if _CCCL_COMPILER(GCC)
-        NV_IF_TARGET(NV_IS_DEVICE, (return ::cuda::std::__with_builtin_fmax(__x, __y);))
+          NV_IF_TARGET(NV_IS_DEVICE, (return ::cuda::std::__with_builtin_fmax(__x, __y);))
 #  else // ^^^ _CCCL_COMPILER(GCC) ^^^ / vvv !_CCCL_COMPILER(GCC)
-        return ::cuda::std::__with_builtin_fmax(__x, __y);
+          return ::cuda::std::__with_builtin_fmax(__x, __y);
 #  endif // !_CCCL_COMPILER(GCC)
-      }
+        }
 #endif // _CCCL_USE_BUILTIN_FMAX
+      }
       if (::cuda::std::isnan(__x))
       {
         return __y;
@@ -119,7 +132,10 @@ _CCCL_REQUIRES(__is_extended_arithmetic_v<_Tp>)
       {
         return __x;
       }
-      return __x < __y ? __y : __x;
+      else
+      {
+        return __x < __y ? __y : __x;
+      }
     }
 }
 
@@ -144,7 +160,9 @@ _CCCL_REQUIRES(::cuda::is_floating_point_v<_Tp> _CCCL_AND ::cuda::is_floating_po
   return ::cuda::std::fmax(static_cast<__result_type>(__x), static_cast<__result_type>(__y));
 }
 
-// fmin
+/***********************************************************************************************************************
+ * fmin
+ **********************************************************************************************************************/
 
 // We do explicitly also enable GCC here, because that makes the condition below simpler
 #if _CCCL_CHECK_BUILTIN(builtin_fmin) || _CCCL_COMPILER(GCC)
@@ -178,7 +196,7 @@ _CCCL_REQUIRES(__is_extended_arithmetic_v<_Tp>)
 [[nodiscard]] _CCCL_API constexpr conditional_t<is_integral_v<_Tp>, double, _Tp> fmin(_Tp __x, _Tp __y) noexcept
 {
 #if _CCCL_HAS_NVFP16()
-  if constexpr (is_same_v<_Tp, __half>)
+  if constexpr (is_same_v<_Tp, ::__half>)
   {
 #  if _CCCL_CTK_AT_LEAST(12, 2)
     return ::__hmin(__x, __y);
@@ -191,7 +209,7 @@ _CCCL_REQUIRES(__is_extended_arithmetic_v<_Tp>)
   else
 #endif // _CCCL_HAS_NVFP16()
 #if _CCCL_HAS_NVBF16()
-    if constexpr (is_same_v<_Tp, __nv_bfloat16>)
+    if constexpr (is_same_v<_Tp, ::__nv_bfloat16>)
   {
 #  if _CCCL_CTK_AT_LEAST(12, 2)
     return ::__hmin(__x, __y);
@@ -209,17 +227,26 @@ _CCCL_REQUIRES(__is_extended_arithmetic_v<_Tp>)
     }
     else
     {
-#if _CCCL_USE_BUILTIN_FMAX()
       if (!::cuda::std::__cccl_default_is_constant_evaluated())
       {
+#if _CCCL_HAS_FLOAT128()
+        if constexpr (is_same_v<_Tp, __float128>)
+        {
+          NV_IF_TARGET(NV_PROVIDES_SM_100, (return ::__nv_fp128_fmin(__x, __y);))
+        }
+#endif // _CCCL_HAS_FLOAT128()
+#if _CCCL_USE_BUILTIN_FMAX()
+        if constexpr (is_floating_point_v<_Tp>)
+        {
 // GCC builtins do not treat NaN properly
 #  if _CCCL_COMPILER(GCC)
-        NV_IF_TARGET(NV_IS_DEVICE, (return ::cuda::std::__with_builtin_fmin(__x, __y);))
+          NV_IF_TARGET(NV_IS_DEVICE, (return ::cuda::std::__with_builtin_fmin(__x, __y);))
 #  else // ^^^ _CCCL_COMPILER(GCC) ^^^ / vvv !_CCCL_COMPILER(GCC)
-        return ::cuda::std::__with_builtin_fmin(__x, __y);
+          return ::cuda::std::__with_builtin_fmin(__x, __y);
 #  endif // !_CCCL_COMPILER(GCC)
-      }
+        }
 #endif // _CCCL_USE_BUILTIN_FMAX
+      }
       if (::cuda::std::isnan(__x))
       {
         return __y;

cuda/cccl/headers/include/cuda/std/__concepts/constructible.h

@@ -138,7 +138,7 @@ _CCCL_CONCEPT __nothrow_initializable_from =
         ? ::cuda::std::is_nothrow_constructible_v<_Tp, _Args...>
         : __nothrow_list_initializable_from<_Tp, _Args...>);
 
-#if !_CCCL_COMPILER(MSVC)
+#if !_CCCL_COMPILER(MSVC) && !_CCCL_CUDA_COMPILER(NVCC, <, 12, 9)
 
 //! Constructible with direct non-list initialization syntax from the result of
 //! a function call expression (often useful for immovable types).

cuda/cccl/headers/include/cuda/std/__cuda/api_wrapper.h

@@ -23,18 +23,18 @@
 
 #include <cuda/std/__exception/cuda_error.h>
 
-#define _CCCL_TRY_CUDA_API(_NAME, _MSG, ...)                        \
-  do                                                                \
-  {                                                                 \
-    const ::cudaError_t __status = _NAME(__VA_ARGS__);              \
-    switch (__status)                                               \
-    {                                                               \
-      case ::cudaSuccess:                                           \
-        break;                                                      \
-      default:                                                      \
-        /* CUDA error state is cleared inside __throw_cuda_error */ \
-        ::cuda::__throw_cuda_error(__status, _MSG, #_NAME);         \
-    }                                                               \
+#define _CCCL_TRY_CUDA_API(_NAME, _MSG, ...)                \
+  do                                                        \
+  {                                                         \
+    const ::cudaError_t __status = _NAME(__VA_ARGS__);      \
+    switch (__status)                                       \
+    {                                                       \
+      case ::cudaSuccess:                                   \
+        break;                                              \
+      default:                                              \
+        ::cudaGetLastError(); /* clear CUDA error state */  \
+        ::cuda::__throw_cuda_error(__status, _MSG, #_NAME); \
+    }                                                       \
   } while (0)
 
 #define _CCCL_ASSERT_CUDA_API(_NAME, _MSG, ...)                         \

cuda/cccl/headers/include/cuda/std/__exception/cuda_error.h

@@ -109,14 +109,7 @@ private:
   [[maybe_unused]] const char* __api                  = nullptr,
   [[maybe_unused]] ::cuda::std::source_location __loc = ::cuda::std::source_location::current())
 {
-#  if _CCCL_CUDA_COMPILATION()
-  NV_IF_ELSE_TARGET(NV_IS_HOST,
-                    (::cudaGetLastError(); // clear CUDA error state
-                     throw ::cuda::cuda_error(__status, __msg, __api, __loc);), //
-                    (::cuda::std::terminate();))
-#  else // ^^^ _CCCL_CUDA_COMPILATION() ^^^ / vvv !_CCCL_CUDA_COMPILATION() vvv
-  throw ::cuda::cuda_error(__status, __msg, __api, __loc);
-#  endif // !_CCCL_CUDA_COMPILATION()
+  NV_IF_TARGET(NV_IS_HOST, (throw ::cuda::cuda_error(__status, __msg, __api, __loc);), (::cuda::std::terminate();))
 }
 #else // ^^^ _CCCL_HAS_EXCEPTIONS() ^^^ / vvv !_CCCL_HAS_EXCEPTIONS() vvv
 class cuda_error