Files
project_6/qwen3_6_scripts/cccl_preload/include/nv/target
dylanyunlon 8d6f9eaeb0 feat: CCCL CachingDeviceAllocator preload — 完整依赖链 288 files
从 cccl_upstream 递归追踪 cub/util_allocator.cuh 的全部 include 依赖:
  cub/         9 files (config, util_*, version, detect_cuda_runtime)
  cuda/        libcudacxx type_traits, concepts, algorithm, iterator...
  nv/          target macros, preprocessor

总计 288 个头文件 (1.4MB),打包到 include/ 目录,编译时 -I include
即可完全脱离 CCCL 原始目录结构。

.cu 文件直接 #include <cub/util_allocator.cuh>,
走原版 CUB CachingDeviceAllocator,零 mock。

BI-V100 参数: growth=2 bins=[8..32] max_cached=8GB/device
2026-08-13 09:26:41 +00:00

251 lines
8.0 KiB
Plaintext

//===----------------------------------------------------------------------===//
//
// 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) 2023 NVIDIA CORPORATION & AFFILIATES.
//
//===----------------------------------------------------------------------===//
// This header contains a preview of a portability system that enables
// CUDA C++ development with NVC++, NVCC, and supported host compilers.
// These interfaces are not guaranteed to be stable.
#ifndef __NV_TARGET_H
#define __NV_TARGET_H
#if defined(__NVCC__) || defined(__CUDACC_RTC__)
# define _NV_COMPILER_NVCC
#elif defined(__NVCOMPILER) && __cplusplus >= 201103L
# define _NV_COMPILER_NVCXX
#elif defined(__clang__) && defined(__CUDA__) && defined(__CUDA_ARCH__)
// clang compiling CUDA code, device mode.
# define _NV_COMPILER_CLANG_CUDA
#endif
// Hide `if target` support from NVRTC and C
// Some toolkit headers use <nv/target> in true C contexts
#if !defined(__CUDACC_RTC__) && defined(__cplusplus)
# if defined(_NV_COMPILER_NVCXX)
# define _NV_BITSET_ATTRIBUTE [[nv::__target_bitset]]
# else
# define _NV_BITSET_ATTRIBUTE
# endif
// This must be a nested namespace as <nv/target> works with older dialects.
namespace nv // NOLINT(modernize-concat-nested-namespaces)
{
namespace target // NOLINT(modernize-concat-nested-namespaces)
{
namespace detail
{
using base_int_t = unsigned long long;
// No host specialization
constexpr base_int_t all_hosts = 1;
// NVIDIA GPUs
constexpr base_int_t sm_35_bit = 1 << 1;
constexpr base_int_t sm_37_bit = 1 << 2;
constexpr base_int_t sm_50_bit = 1 << 3;
constexpr base_int_t sm_52_bit = 1 << 4;
constexpr base_int_t sm_53_bit = 1 << 5;
constexpr base_int_t sm_60_bit = 1 << 6;
constexpr base_int_t sm_61_bit = 1 << 7;
constexpr base_int_t sm_62_bit = 1 << 8;
constexpr base_int_t sm_70_bit = 1 << 9;
constexpr base_int_t sm_72_bit = 1 << 10;
constexpr base_int_t sm_75_bit = 1 << 11;
constexpr base_int_t sm_80_bit = 1 << 12;
constexpr base_int_t sm_86_bit = 1 << 13;
constexpr base_int_t sm_87_bit = 1 << 14;
constexpr base_int_t sm_88_bit = 1 << 15;
constexpr base_int_t sm_89_bit = 1 << 16;
constexpr base_int_t sm_90_bit = 1 << 17;
constexpr base_int_t sm_100_bit = 1 << 18;
constexpr base_int_t sm_103_bit = 1 << 19;
constexpr base_int_t sm_107_bit = 1 << 20;
constexpr base_int_t sm_110_bit = 1 << 21;
constexpr base_int_t sm_120_bit = 1 << 22;
constexpr base_int_t sm_121_bit = 1 << 23;
constexpr base_int_t all_devices =
sm_35_bit | sm_37_bit | sm_50_bit | sm_52_bit | sm_53_bit | sm_60_bit | sm_61_bit | sm_62_bit | sm_70_bit | sm_72_bit
| sm_75_bit | sm_80_bit | sm_86_bit | sm_87_bit | sm_88_bit | sm_89_bit | sm_90_bit | sm_100_bit | sm_103_bit
| sm_107_bit | sm_110_bit | sm_120_bit | sm_121_bit;
// Store a set of targets as a set of bits
struct _NV_BITSET_ATTRIBUTE target_description
{
base_int_t targets;
constexpr target_description(base_int_t a)
: targets(a)
{}
};
// The type of the user-visible names of the NVIDIA GPU targets
enum class sm_selector : base_int_t
{
sm_35 = 35,
sm_37 = 37,
sm_50 = 50,
sm_52 = 52,
sm_53 = 53,
sm_60 = 60,
sm_61 = 61,
sm_62 = 62,
sm_70 = 70,
sm_72 = 72,
sm_75 = 75,
sm_80 = 80,
sm_86 = 86,
sm_87 = 87,
sm_88 = 88,
sm_89 = 89,
sm_90 = 90,
sm_100 = 100,
sm_103 = 103,
sm_107 = 107,
sm_110 = 110,
sm_120 = 120,
sm_121 = 121,
};
constexpr base_int_t toint(sm_selector a)
{
return static_cast<base_int_t>(a);
}
constexpr base_int_t bitexact(sm_selector a)
{
return toint(a) == 35 ? sm_35_bit
: toint(a) == 37 ? sm_37_bit
: toint(a) == 50 ? sm_50_bit
: toint(a) == 52 ? sm_52_bit
: toint(a) == 53 ? sm_53_bit
: toint(a) == 60 ? sm_60_bit
: toint(a) == 61 ? sm_61_bit
: toint(a) == 62 ? sm_62_bit
: toint(a) == 70 ? sm_70_bit
: toint(a) == 72 ? sm_72_bit
: toint(a) == 75 ? sm_75_bit
: toint(a) == 80 ? sm_80_bit
: toint(a) == 86 ? sm_86_bit
: toint(a) == 87 ? sm_87_bit
: toint(a) == 88 ? sm_88_bit
: toint(a) == 89 ? sm_89_bit
: toint(a) == 90 ? sm_90_bit
: toint(a) == 100 ? sm_100_bit
: toint(a) == 103 ? sm_103_bit
: toint(a) == 107 ? sm_107_bit
: toint(a) == 110 ? sm_110_bit
: toint(a) == 120 ? sm_120_bit
: toint(a) == 121 ? sm_121_bit
: 0;
}
constexpr base_int_t bitrounddown(sm_selector a)
{
return toint(a) >= 121 ? sm_121_bit
: toint(a) >= 120 ? sm_120_bit
: toint(a) >= 110 ? sm_110_bit
: toint(a) >= 107 ? sm_107_bit
: toint(a) >= 103 ? sm_103_bit
: toint(a) >= 100 ? sm_100_bit
: toint(a) >= 90 ? sm_90_bit
: toint(a) >= 89 ? sm_89_bit
: toint(a) >= 88 ? sm_88_bit
: toint(a) >= 87 ? sm_87_bit
: toint(a) >= 86 ? sm_86_bit
: toint(a) >= 80 ? sm_80_bit
: toint(a) >= 75 ? sm_75_bit
: toint(a) >= 72 ? sm_72_bit
: toint(a) >= 70 ? sm_70_bit
: toint(a) >= 62 ? sm_62_bit
: toint(a) >= 61 ? sm_61_bit
: toint(a) >= 60 ? sm_60_bit
: toint(a) >= 53 ? sm_53_bit
: toint(a) >= 52 ? sm_52_bit
: toint(a) >= 50 ? sm_50_bit
: toint(a) >= 37 ? sm_37_bit
: toint(a) >= 35 ? sm_35_bit
: 0;
}
// Public API for NVIDIA GPUs
constexpr target_description is_exactly(sm_selector a)
{
return target_description(bitexact(a));
}
constexpr target_description provides(sm_selector a)
{
return target_description(~(bitrounddown(a) - 1) & all_devices);
}
// Boolean operations on target sets
constexpr target_description operator&&(target_description a, target_description b)
{
return target_description(a.targets & b.targets);
}
constexpr target_description operator||(target_description a, target_description b)
{
return target_description(a.targets | b.targets);
}
constexpr target_description operator!(target_description a)
{
return target_description(~a.targets & (all_devices | all_hosts));
}
} // namespace detail
using detail::sm_selector;
using detail::target_description;
// The predicates for basic host/device selection
constexpr target_description is_host = target_description(detail::all_hosts);
constexpr target_description is_device = target_description(detail::all_devices);
constexpr target_description any_target = target_description(detail::all_hosts | detail::all_devices);
constexpr target_description no_target = target_description(0);
// The public names for NVIDIA GPU architectures
constexpr sm_selector sm_35 = sm_selector::sm_35;
constexpr sm_selector sm_37 = sm_selector::sm_37;
constexpr sm_selector sm_50 = sm_selector::sm_50;
constexpr sm_selector sm_52 = sm_selector::sm_52;
constexpr sm_selector sm_53 = sm_selector::sm_53;
constexpr sm_selector sm_60 = sm_selector::sm_60;
constexpr sm_selector sm_61 = sm_selector::sm_61;
constexpr sm_selector sm_62 = sm_selector::sm_62;
constexpr sm_selector sm_70 = sm_selector::sm_70;
constexpr sm_selector sm_72 = sm_selector::sm_72;
constexpr sm_selector sm_75 = sm_selector::sm_75;
constexpr sm_selector sm_80 = sm_selector::sm_80;
constexpr sm_selector sm_86 = sm_selector::sm_86;
constexpr sm_selector sm_87 = sm_selector::sm_87;
constexpr sm_selector sm_88 = sm_selector::sm_88;
constexpr sm_selector sm_89 = sm_selector::sm_89;
constexpr sm_selector sm_90 = sm_selector::sm_90;
constexpr sm_selector sm_100 = sm_selector::sm_100;
constexpr sm_selector sm_103 = sm_selector::sm_103;
constexpr sm_selector sm_107 = sm_selector::sm_107;
constexpr sm_selector sm_110 = sm_selector::sm_110;
constexpr sm_selector sm_120 = sm_selector::sm_120;
constexpr sm_selector sm_121 = sm_selector::sm_121;
using detail::is_exactly;
using detail::provides;
} // namespace target
} // namespace nv
#endif // C++ && !defined(__CUDACC_RTC__)
#include <nv/detail/__target_macros>
#endif // __NV_TARGET_H