cuda-cccl 0.3.1__cp310-cp310-manylinux_2_24_aarch64.whl → 0.3.2__cp310-cp310-manylinux_2_24_aarch64.whl

cuda/cccl/headers/include/cuda/__device/attributes.h

@@ -23,6 +23,7 @@
 
 #if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
 
+#  include <cuda/__device/compute_capability.h>
 #  include <cuda/__device/device_ref.h>
 #  include <cuda/__driver/driver_api.h>
 #  include <cuda/__fwd/devices.h>
@@ -739,12 +740,12 @@ static constexpr numa_id_t numa_id{};
 // capability in a single query
 struct compute_capability_t
 {
-  using type = int;
+  using type = ::cuda::compute_capability;
 
   [[nodiscard]] _CCCL_HOST_API type operator()(device_ref __dev_id) const
   {
-    return 10 * ::cuda::device_attributes::compute_capability_major(__dev_id)
-         + ::cuda::device_attributes::compute_capability_minor(__dev_id);
+    return type{::cuda::device_attributes::compute_capability_major(__dev_id),
+                ::cuda::device_attributes::compute_capability_minor(__dev_id)};
   }
 };
 static constexpr compute_capability_t compute_capability{};

cuda/cccl/headers/include/cuda/__device/compute_capability.h

@@ -0,0 +1,171 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___DEVICE_COMPUTE_CAPABILITY_H
+#define _CUDA___DEVICE_COMPUTE_CAPABILITY_H
+
+#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/__fwd/devices.h>
+#include <cuda/std/__utility/to_underlying.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+//! @brief Type representing the CUDA compute capability.
+class compute_capability
+{
+  int __cc_{}; //!< The stored compute capability in format 10 * major + minor.
+
+public:
+  _CCCL_HIDE_FROM_ABI constexpr compute_capability() noexcept = default;
+
+  //! @brief Constructs the object from compute capability \c __cc. The expected format is 10 * major + minor.
+  //!
+  //! @param __cc Compute capability.
+  _CCCL_API explicit constexpr compute_capability(int __cc) noexcept
+      : __cc_{__cc}
+  {}
+
+  //! @brief Constructs the object by combining the \c __major and \c __minor compute capability.
+  //!
+  //! @param __major The major compute capability.
+  //! @param __minor The minor compute capability. Must be less than 10.
+  _CCCL_API constexpr compute_capability(int __major, int __minor) noexcept
+      : __cc_{10 * __major + __minor}
+  {
+    _CCCL_ASSERT(__minor < 10, "invalid minor compute capability");
+  }
+
+  //! @brief Constructs the object from the architecture id.
+  //!
+  //! @param __arch_id The architecture id.
+  _CCCL_API explicit constexpr compute_capability(arch_id __arch_id) noexcept
+  {
+    const auto __val = ::cuda::std::to_underlying(__arch_id);
+    if (__val > __arch_specific_id_multiplier)
+    {
+      __cc_ = __val / __arch_specific_id_multiplier;
+    }
+    else
+    {
+      __cc_ = __val;
+    }
+  }
+
+  _CCCL_HIDE_FROM_ABI constexpr compute_capability(const compute_capability&) noexcept = default;
+
+  _CCCL_HIDE_FROM_ABI constexpr compute_capability& operator=(const compute_capability& __other) noexcept = default;
+
+  //! @brief Gets the stored compute capability.
+  //!
+  //! @return The stored compute capability in format 10 * major + minor.
+  [[nodiscard]] _CCCL_API constexpr int get() const noexcept
+  {
+    return __cc_;
+  }
+
+  //! @brief Gets the major compute capability.
+  //!
+  //! @return Major compute capability.
+  [[nodiscard]] _CCCL_API constexpr int major() const noexcept
+  {
+    return __cc_ / 10;
+  }
+
+  //! @brief Gets the minor compute capability.
+  //!
+  //! @return Minor compute capability. The value is always less than 10.
+  [[nodiscard]] _CCCL_API constexpr int minor() const noexcept
+  {
+    return __cc_ % 10;
+  }
+
+  //! @brief Conversion operator to \c int.
+  //!
+  //! @return The stored compute capability in format 10 * major + minor.
+  _CCCL_API explicit constexpr operator int() const noexcept
+  {
+    return __cc_;
+  }
+
+  //! @brief Equality operator.
+  [[nodiscard]] friend _CCCL_API constexpr bool operator==(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ == __rhs.__cc_;
+  }
+
+  //! @brief Inequality operator.
+  [[nodiscard]] friend _CCCL_API constexpr bool operator!=(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ != __rhs.__cc_;
+  }
+
+  //! @brief Less than operator.
+  [[nodiscard]] friend _CCCL_API constexpr bool operator<(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ < __rhs.__cc_;
+  }
+
+  //! @brief Less than or equal to operator.
+  [[nodiscard]] friend _CCCL_API constexpr bool operator<=(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ <= __rhs.__cc_;
+  }
+
+  //! @brief Greater than operator.
+  [[nodiscard]] friend _CCCL_API constexpr bool operator>(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ > __rhs.__cc_;
+  }
+
+  //! @brief Greater than or equal to operator.
+  [[nodiscard]] friend _CCCL_API constexpr bool operator>=(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ >= __rhs.__cc_;
+  }
+};
+
+_CCCL_END_NAMESPACE_CUDA
+
+#if _CCCL_CUDA_COMPILATION()
+
+_CCCL_BEGIN_NAMESPACE_CUDA_DEVICE
+
+//! @brief Returns the \c cuda::compute_capability that is currently being compiled.
+//!
+//! @note This API cannot be used in constexpr context when compiling with nvc++ in CUDA mode.
+[[nodiscard]] _CCCL_DEVICE_API _CCCL_TARGET_CONSTEXPR ::cuda::compute_capability current_compute_capability() noexcept
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC)
+  return ::cuda::compute_capability{__builtin_current_device_sm()};
+#  elif _CCCL_DEVICE_COMPILATION()
+  return ::cuda::compute_capability{__CUDA_ARCH__ / 10};
+#  else // ^^^ _CCCL_DEVICE_COMPILATION() ^^^ / vvv !_CCCL_DEVICE_COMPILATION() vvv
+  return {};
+#  endif // ^^^ !_CCCL_DEVICE_COMPILATION() ^^^
+}
+
+_CCCL_END_NAMESPACE_CUDA_DEVICE
+
+#endif // _CCCL_CUDA_COMPILATION()
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___DEVICE_COMPUTE_CAPABILITY_H

cuda/cccl/headers/include/cuda/__device/device_ref.h

@@ -133,16 +133,6 @@ public:
       ::cuda::__driver::__deviceGet(get()), ::cuda::__driver::__deviceGet(__other_dev.get()));
   }
 
-  //! @brief Retrieve architecture traits of this device.
-  //!
-  //! Architecture traits object contains information about certain traits
-  //! that are shared by all devices belonging to given architecture.
-  //!
-  //! @return A reference to `arch_traits_t` object containing architecture traits of this device
-  [[nodiscard]] _CCCL_HOST_API const arch::traits_t& arch_traits() const; // implemented in
-                                                                          // <cuda/__device/physical_device.h> to avoid
-                                                                          // circular dependency
-
   // TODO this might return some more complex type in the future
   // TODO we might want to include the calling device, depends on what we decide
   // peer access APIs

cuda/cccl/headers/include/cuda/__device/physical_device.h

@@ -23,16 +23,15 @@
 
 #if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
 
-#  include <cuda/__device/arch_traits.h>
 #  include <cuda/__device/device_ref.h>
 #  include <cuda/__driver/driver_api.h>
 #  include <cuda/__fwd/devices.h>
+#  include <cuda/std/__cccl/memory_wrapper.h>
 #  include <cuda/std/__cstddef/types.h>
 #  include <cuda/std/span>
 #  include <cuda/std/string_view>
 
 #  include <cassert>
-#  include <memory>
 #  include <mutex>
 #  include <vector>
 
@@ -53,10 +52,6 @@ class __physical_device
 
   ::CUdevice __device_{};
 
-  // TODO We should have some of the attributes just return from the arch traits
-  ::std::once_flag __traits_once_flag_{};
-  arch::traits_t __traits_{};
-
   ::std::once_flag __primary_ctx_once_flag_{};
   ::CUcontext __primary_ctx_{};
 
@@ -90,21 +85,6 @@ public:
     return __primary_ctx_;
   }
 
-  //! @brief Retrieve architecture traits of this device.
-  //!
-  //! Architecture traits object contains information about certain traits
-  //! that are shared by all devices belonging to given architecture.
-  //!
-  //! @return A reference to `arch_traits_t` object containing architecture traits of this device
-  [[nodiscard]] _CCCL_HOST_API const arch::traits_t& __arch_traits()
-  {
-    ::std::call_once(__traits_once_flag_, [this]() {
-      const auto __id = ::cuda::__driver::__cudevice_to_ordinal(__device_);
-      __traits_       = ::cuda::arch::__arch_traits_might_be_unknown(__id, device_attributes::compute_capability(__id));
-    });
-    return __traits_;
-  }
-
   [[nodiscard]] _CCCL_HOST_API ::cuda::std::string_view __name()
   {
     ::std::call_once(__name_once_flag_, [this]() {
@@ -178,11 +158,6 @@ _CCCL_HOST_API inline void device_ref::init() const
   return ::cuda::__physical_devices()[__id_].__name();
 }
 
-[[nodiscard]] _CCCL_HOST_API inline const arch::traits_t& device_ref::arch_traits() const
-{
-  return ::cuda::__physical_devices()[__id_].__arch_traits();
-}
-
 [[nodiscard]] _CCCL_HOST_API inline ::cuda::std::span<const device_ref> device_ref::peers() const
 {
   return ::cuda::__physical_devices()[__id_].__peers();

cuda/cccl/headers/include/cuda/__event/event.h

@@ -28,8 +28,8 @@
 #  include <cuda/__event/event_ref.h>
 #  include <cuda/__runtime/ensure_current_context.h>
 #  include <cuda/__utility/no_init.h>
+#  include <cuda/std/__utility/to_underlying.h>
 #  include <cuda/std/cstddef>
-#  include <cuda/std/utility>
 
 #  include <cuda/std/__cccl/prologue.h>
 
@@ -37,38 +37,43 @@ _CCCL_BEGIN_NAMESPACE_CUDA
 
 class timed_event;
 
+//! @brief Flags to use when creating the event.
+enum class event_flags : unsigned
+{
+  none          = cudaEventDefault,
+  blocking_sync = cudaEventBlockingSync,
+  interprocess  = cudaEventInterprocess,
+};
+
+[[nodiscard]] _CCCL_HOST_API constexpr event_flags operator|(event_flags __lhs, event_flags __rhs) noexcept
+{
+  return static_cast<event_flags>(::cuda::std::to_underlying(__lhs) | ::cuda::std::to_underlying(__rhs));
+}
+
 //! @brief An owning wrapper for an untimed `cudaEvent_t`.
 class event : public event_ref
 {
   friend class timed_event;
 
 public:
-  //! @brief Flags to use when creating the event.
-  enum class flags : unsigned
-  {
-    none          = cudaEventDefault,
-    blocking_sync = cudaEventBlockingSync,
-    interprocess  = cudaEventInterprocess,
-  };
-
   //! @brief Construct a new `event` object with timing disabled, and record
   //!        the event in the specified stream.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit event(stream_ref __stream, flags __flags = flags::none);
+  _CCCL_HOST_API explicit event(stream_ref __stream, event_flags __flags = event_flags::none);
 
   //! @brief Construct a new `event` object with timing disabled. The event can only be recorded on streams from the
   //! specified device.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit event(device_ref __device, flags __flags = flags::none)
-      : event(__device, static_cast<unsigned int>(__flags) | cudaEventDisableTiming)
+  _CCCL_HOST_API explicit event(device_ref __device, event_flags __flags = event_flags::none)
+      : event(__device, ::cuda::std::to_underlying(__flags) | cudaEventDisableTiming)
   {}
 
   //! @brief Construct a new `event` object into the moved-from state.
   //!
   //! @post `get()` returns `cudaEvent_t()`.
-  explicit constexpr event(no_init_t) noexcept
+  _CCCL_HOST_API explicit constexpr event(no_init_t) noexcept
       : event_ref(::cudaEvent_t{})
   {}
 
@@ -77,7 +82,7 @@ public:
   //! @param __other
   //!
   //! @post `__other` is in a moved-from state.
-  constexpr event(event&& __other) noexcept
+  _CCCL_HOST_API constexpr event(event&& __other) noexcept
       : event_ref(::cuda::std::exchange(__other.__event_, {}))
   {}
 
@@ -87,7 +92,7 @@ public:
   //! @brief Destroy the `event` object
   //!
   //! @note If the event fails to be destroyed, the error is silently ignored.
-  ~event()
+  _CCCL_HOST_API ~event()
   {
     if (__event_ != nullptr)
     {
@@ -102,7 +107,7 @@ public:
   //! @param __other
   //!
   //! @post `__other` is in a moved-from state.
-  event& operator=(event&& __other) noexcept
+  _CCCL_HOST_API event& operator=(event&& __other) noexcept
   {
     event __tmp(::cuda::std::move(__other));
     ::cuda::std::swap(__event_, __tmp.__event_);
@@ -119,7 +124,7 @@ public:
   //! @return event The constructed `event` object
   //!
   //! @note The constructed `event` object takes ownership of the native handle.
-  [[nodiscard]] static event from_native_handle(::cudaEvent_t __evnt) noexcept
+  [[nodiscard]] static _CCCL_HOST_API event from_native_handle(::cudaEvent_t __evnt) noexcept
   {
     return event(__evnt);
   }
@@ -135,26 +140,21 @@ public:
   //! @return cudaEvent_t The native handle being held by the `event` object.
   //!
   //! @post The event object is in a moved-from state.
-  [[nodiscard]] constexpr ::cudaEvent_t release() noexcept
+  [[nodiscard]] _CCCL_HOST_API constexpr ::cudaEvent_t release() noexcept
   {
     return ::cuda::std::exchange(__event_, {});
   }
 
-  [[nodiscard]] friend constexpr flags operator|(flags __lhs, flags __rhs) noexcept
-  {
-    return static_cast<flags>(static_cast<unsigned>(__lhs) | static_cast<unsigned>(__rhs));
-  }
-
 private:
   // Use `event::from_native_handle(e)` to construct an owning `event`
   // object from a `cudaEvent_t` handle.
-  explicit constexpr event(::cudaEvent_t __evnt) noexcept
+  _CCCL_HOST_API explicit constexpr event(::cudaEvent_t __evnt) noexcept
       : event_ref(__evnt)
   {}
 
-  explicit event(stream_ref __stream, unsigned __flags);
+  _CCCL_HOST_API explicit event(stream_ref __stream, unsigned __flags);
 
-  explicit event(device_ref __device, unsigned __flags)
+  _CCCL_HOST_API explicit event(device_ref __device, unsigned __flags)
       : event_ref(::cudaEvent_t{})
   {
     [[maybe_unused]] __ensure_current_context __ctx_setter(__device);

cuda/cccl/headers/include/cuda/__event/event_ref.h

@@ -56,7 +56,7 @@ public:
   //!
   //! @note: It is the callers responsibility to ensure the `event_ref` does not
   //! outlive the event denoted by the `cudaEvent_t` handle.
-  constexpr event_ref(::cudaEvent_t __evnt) noexcept
+  _CCCL_HOST_API constexpr event_ref(::cudaEvent_t __evnt) noexcept
       : __event_(__evnt)
   {}
 
@@ -108,7 +108,7 @@ public:
   //! @brief Retrieve the native `cudaEvent_t` handle.
   //!
   //! @return cudaEvent_t The native handle being held by the event_ref object.
-  [[nodiscard]] constexpr ::cudaEvent_t get() const noexcept
+  [[nodiscard]] _CCCL_HOST_API constexpr ::cudaEvent_t get() const noexcept
   {
     return __event_;
   }
@@ -116,7 +116,7 @@ public:
   //! @brief Checks if the `event_ref` is valid
   //!
   //! @return true if the `event_ref` is valid, false otherwise.
-  [[nodiscard]] explicit constexpr operator bool() const noexcept
+  [[nodiscard]] _CCCL_HOST_API explicit constexpr operator bool() const noexcept
   {
     return __event_ != nullptr;
   }
@@ -129,7 +129,7 @@ public:
   //! @param __lhs The first `event_ref` to compare
   //! @param __rhs The second `event_ref` to compare
   //! @return true if `lhs` and `rhs` refer to the same `cudaEvent_t` object.
-  [[nodiscard]] friend constexpr bool operator==(event_ref __lhs, event_ref __rhs) noexcept
+  [[nodiscard]] friend _CCCL_HOST_API constexpr bool operator==(event_ref __lhs, event_ref __rhs) noexcept
   {
     return __lhs.__event_ == __rhs.__event_;
   }
@@ -142,7 +142,7 @@ public:
   //! @param __lhs The first `event_ref` to compare
   //! @param __rhs The second `event_ref` to compare
   //! @return true if `lhs` and `rhs` refer to different `cudaEvent_t` objects.
-  [[nodiscard]] friend constexpr bool operator!=(event_ref __lhs, event_ref __rhs) noexcept
+  [[nodiscard]] friend _CCCL_HOST_API constexpr bool operator!=(event_ref __lhs, event_ref __rhs) noexcept
   {
     return __lhs.__event_ != __rhs.__event_;
   }

cuda/cccl/headers/include/cuda/__event/timed_event.h

@@ -31,6 +31,7 @@
 #  include <cuda/__event/event.h>
 #  include <cuda/__utility/no_init.h>
 #  include <cuda/std/__chrono/duration.h>
+#  include <cuda/std/__utility/to_underlying.h>
 #  include <cuda/std/cstddef>
 
 #  include <cuda/std/__cccl/prologue.h>
@@ -45,20 +46,20 @@ public:
   //!        and record the event on the specified stream.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit timed_event(stream_ref __stream, flags __flags = flags::none);
+  _CCCL_HOST_API explicit timed_event(stream_ref __stream, event_flags __flags = event_flags::none);
 
   //! @brief Construct a new `timed_event` object with the specified flags. The event can only be recorded on streams
   //! from the specified device.
   //!
   //! @throws cuda_error if the event creation fails.
-  explicit timed_event(device_ref __device, flags __flags = flags::none)
-      : event(__device, static_cast<unsigned>(__flags))
+  _CCCL_HOST_API explicit timed_event(device_ref __device, event_flags __flags = event_flags::none)
+      : event(__device, ::cuda::std::to_underlying(__flags))
   {}
 
   //! @brief Construct a new `timed_event` object into the moved-from state.
   //!
   //! @post `get()` returns `cudaEvent_t()`.
-  explicit constexpr timed_event(no_init_t) noexcept
+  _CCCL_HOST_API explicit constexpr timed_event(no_init_t) noexcept
       : event(no_init)
   {}
 
@@ -74,7 +75,7 @@ public:
   //! @return timed_event The constructed `timed_event` object
   //!
   //! @note The constructed `timed_event` object takes ownership of the native handle.
-  [[nodiscard]] static timed_event from_native_handle(::cudaEvent_t __evnt) noexcept
+  [[nodiscard]] static _CCCL_HOST_API timed_event from_native_handle(::cudaEvent_t __evnt) noexcept
   {
     return timed_event(__evnt);
   }
@@ -95,7 +96,8 @@ public:
   //! @return cuda::std::chrono::nanoseconds The elapsed time in nanoseconds.
   //!
   //! @note The elapsed time has a resolution of approximately 0.5 microseconds.
-  [[nodiscard]] friend ::cuda::std::chrono::nanoseconds operator-(const timed_event& __end, const timed_event& __start)
+  [[nodiscard]] friend _CCCL_HOST_API ::cuda::std::chrono::nanoseconds
+  operator-(const timed_event& __end, const timed_event& __start)
   {
     const auto __ms = ::cuda::__driver::__eventElapsedTime(__start.get(), __end.get());
     return ::cuda::std::chrono::nanoseconds(static_cast<::cuda::std::chrono::nanoseconds::rep>(__ms * 1'000'000.0));
@@ -104,7 +106,7 @@ public:
 private:
   // Use `timed_event::from_native_handle(e)` to construct an owning `timed_event`
   // object from a `cudaEvent_t` handle.
-  explicit constexpr timed_event(::cudaEvent_t __evnt) noexcept
+  _CCCL_HOST_API explicit constexpr timed_event(::cudaEvent_t __evnt) noexcept
       : event(__evnt)
   {}
 };

cuda/cccl/headers/include/cuda/__fwd/devices.h

@@ -31,11 +31,11 @@ class __physical_device;
 class device_ref;
 template <::cudaDeviceAttr _Attr>
 struct __dev_attr;
+struct arch_traits_t;
+class compute_capability;
+enum class arch_id : int;
 
-namespace arch
-{
-struct traits_t;
-} // namespace arch
+inline constexpr int __arch_specific_id_multiplier = 100000;
 
 _CCCL_END_NAMESPACE_CUDA