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
86 lines
3.1 KiB
C++
86 lines
3.1 KiB
C++
// SPDX-FileCopyrightText: Copyright (c) 2018, NVIDIA Corporation. All rights reserved.
|
|
// SPDX-License-Identifier: Apache-2.0
|
|
|
|
/*! \file
|
|
* \brief Global operator new-based memory resource.
|
|
*/
|
|
|
|
#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
|
|
|
|
#include <thrust/mr/memory_resource.h>
|
|
|
|
THRUST_NAMESPACE_BEGIN
|
|
namespace mr
|
|
{
|
|
/** \addtogroup memory_resources Memory Resources
|
|
* \ingroup memory_management
|
|
* \{
|
|
*/
|
|
|
|
class new_delete_resource_base : public memory_resource<>
|
|
{
|
|
public:
|
|
void* do_allocate(std::size_t bytes, std::size_t alignment = THRUST_MR_DEFAULT_ALIGNMENT) override
|
|
{
|
|
#if __cpp_aligned_new >= 201606L
|
|
return ::operator new(bytes, std::align_val_t(alignment));
|
|
#else // ^^^ __cpp_aligned_new >= 201606L ^^^ / vvv __cpp_aligned_new < 201606L vvv
|
|
// allocate memory for bytes, plus potential alignment correction,
|
|
// plus store of the correction offset
|
|
void* p = ::operator new(bytes + alignment + sizeof(std::size_t));
|
|
std::size_t ptr_int = reinterpret_cast<std::size_t>(p);
|
|
// calculate the offset, i.e. how many bytes of correction was necessary
|
|
// to get an aligned pointer
|
|
std::size_t offset = (ptr_int % alignment) ? (alignment - ptr_int % alignment) : 0;
|
|
// calculate the return pointer
|
|
char* ptr = static_cast<char*>(p) + offset;
|
|
// store the offset right after the actually returned value
|
|
std::size_t* offset_store = reinterpret_cast<std::size_t*>(ptr + bytes);
|
|
*offset_store = offset;
|
|
return static_cast<void*>(ptr);
|
|
#endif // ^^^ __cpp_aligned_new < 201606L ^^^
|
|
}
|
|
|
|
void do_deallocate(void* p,
|
|
[[maybe_unused]] std::size_t bytes,
|
|
[[maybe_unused]] std::size_t alignment = THRUST_MR_DEFAULT_ALIGNMENT) override
|
|
{
|
|
#if __cpp_aligned_new >= 201606L
|
|
# if __cpp_sized_deallocation >= 201309L
|
|
::operator delete(p, bytes, std::align_val_t(alignment));
|
|
# else // ^^^ __cpp_sized_deallocation >= 201309L ^^^ / vvv __cpp_sized_deallocation < 201309L vvv
|
|
::operator delete(p, std::align_val_t(alignment));
|
|
# endif // ^^^ __cpp_sized_deallocation < 201309L ^^^
|
|
#else // ^^^ __cpp_aligned_new >= 201606L ^^^ / vvv __cpp_aligned_new < 201606L vvv
|
|
char* ptr = static_cast<char*>(p);
|
|
// calculate where the offset is stored
|
|
std::size_t* offset = reinterpret_cast<std::size_t*>(ptr + bytes);
|
|
// calculate the original pointer
|
|
p = static_cast<void*>(ptr - *offset);
|
|
::operator delete(p);
|
|
#endif // ^^^ __cpp_aligned_new < 201606L ^^^
|
|
}
|
|
};
|
|
|
|
/*! A memory resource that uses global operators new and delete to allocate and deallocate memory. Uses
|
|
* alignment-enabled overloads when available, otherwise uses regular overloads and implements alignment requirements by
|
|
* itself.
|
|
*/
|
|
class new_delete_resource final : public new_delete_resource_base
|
|
{};
|
|
|
|
/*! \} // memory_resources
|
|
*/
|
|
} // namespace mr
|
|
THRUST_NAMESPACE_END
|