56fd68e7dd
CCCL (CUDA C++ Core Libraries) provides: - CUB: device/block/warp-level GPU primitives (reduce, scan, sort, topk) - Thrust: high-level parallel algorithms (transform_reduce, sort, scan) - libcudacxx: CUDA C++ standard library (atomics, barriers, memory) - cudax: experimental features (memory resources, allocators) - Tuning policies: per-SM hardware-specific algorithm parameters Competition optimization vectors mapped to CCCL: - Output TPS (83% weight): warp_reduce, block_reduce, device_topk - Input TPS (14% weight): device_scan, block_load, prefetch - Cache TPS (3% weight): prefix caching strategy patterns - Memory (0.9 util): pooled/cached/buddy allocators Source: https://github.com/NVIDIA/cccl (shallow clone, HEAD only) License: Apache-2.0
93 lines
3.2 KiB
C++
93 lines
3.2 KiB
C++
// SPDX-FileCopyrightText: Copyright (c) 2008-2020, NVIDIA Corporation. All rights reserved.
|
|
// SPDX-License-Identifier: Apache-2.0
|
|
|
|
/*! \file universal_allocator.h
|
|
* \brief An allocator which creates new elements in memory accessible to both
|
|
* hosts and devices.
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include <thrust/detail/config.h>
|
|
|
|
#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
|
|
|
|
// Some build systems need a hint to know which files we could include
|
|
#if 0
|
|
# include <thrust/system/cpp/memory.h>
|
|
# include <thrust/system/cuda/memory.h>
|
|
# include <thrust/system/omp/memory.h>
|
|
# include <thrust/system/tbb/memory.h>
|
|
#endif
|
|
|
|
// #include the device system's vector header
|
|
#define __THRUST_DEVICE_SYSTEM_MEMORY_HEADER <__THRUST_DEVICE_SYSTEM_ROOT/memory.h>
|
|
#include __THRUST_DEVICE_SYSTEM_MEMORY_HEADER
|
|
#undef __THRUST_DEVICE_SYSTEM_MEMORY_HEADER
|
|
|
|
THRUST_NAMESPACE_BEGIN
|
|
|
|
/** \addtogroup memory_resources Memory Resources
|
|
* \ingroup memory_management_classes
|
|
* \{
|
|
*/
|
|
|
|
/*! \brief An allocator which creates new elements in memory accessible by
|
|
* both hosts and devices.
|
|
*
|
|
* \see https://en.cppreference.com/w/cpp/named_req/Allocator
|
|
*/
|
|
using thrust::system::__THRUST_DEVICE_SYSTEM_NAMESPACE::universal_allocator;
|
|
|
|
/*! \brief An allocator which creates new elements in memory accessible by both hosts and devices. Uses pinned memory
|
|
* when the system supports it.
|
|
*
|
|
* \see https://en.cppreference.com/w/cpp/named_req/Allocator
|
|
*/
|
|
using thrust::system::__THRUST_DEVICE_SYSTEM_NAMESPACE::universal_host_pinned_allocator;
|
|
|
|
/*! \p universal_ptr stores a pointer to an object allocated in memory accessible
|
|
* to both hosts and devices.
|
|
*
|
|
* Algorithms dispatched with this type of pointer will be dispatched to
|
|
* either host or device, depending on which backend you are using. Explicit
|
|
* policies (\p thrust::device, etc) can be used to specify where an algorithm
|
|
* should be run.
|
|
*
|
|
* \p universal_ptr has pointer semantics: it may be dereferenced safely from
|
|
* both hosts and devices and may be manipulated with pointer arithmetic.
|
|
*
|
|
* \p universal_ptr can be created with \p universal_allocator or by explicitly
|
|
* calling its constructor with a raw pointer.
|
|
*
|
|
* The raw pointer encapsulated by a \p universal_ptr may be obtained by
|
|
* either its <tt>get</tt> method or the \p raw_pointer_cast free function.
|
|
*
|
|
* \note \p universal_ptr is not a smart pointer; it is the programmer's
|
|
* responsibility to deallocate memory pointed to by \p universal_ptr.
|
|
*
|
|
* \see host_ptr For the documentation of the complete interface which is
|
|
* shared by \p universal_ptr.
|
|
* \see raw_pointer_cast
|
|
*
|
|
* \verbatim embed:rst:leading-asterisk
|
|
* .. versionadded:: 2.2.0
|
|
* \endverbatim
|
|
*/
|
|
template <typename T>
|
|
using universal_ptr = thrust::system::__THRUST_DEVICE_SYSTEM_NAMESPACE::universal_pointer<T>;
|
|
|
|
template <typename T>
|
|
using universal_host_pinned_ptr = thrust::system::__THRUST_DEVICE_SYSTEM_NAMESPACE::universal_host_pinned_pointer<T>;
|
|
|
|
/*! \}
|
|
*/
|
|
|
|
THRUST_NAMESPACE_END
|