[INFRA] Import NVIDIA/CCCL upstream as optimization reference library
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
This commit is contained in:
EngineX CI
219
cccl_upstream/libcudacxx/codegen/generators/fetch_ops.h
Normal file
219
cccl_upstream/libcudacxx/codegen/generators/fetch_ops.h
Normal file
@@ -0,0 +1,219 @@
|
||||
//===----------------------------------------------------------------------===//
|
||||
//
|
||||
// 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) 2024 NVIDIA CORPORATION & AFFILIATES.
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#ifndef FETCH_OPS_H
|
||||
#define FETCH_OPS_H
|
||||
|
||||
#include <array>
|
||||
#include <format>
|
||||
#include <string>
|
||||
|
||||
#include "definitions.h"
|
||||
|
||||
inline std::string fetch_op_skip_v(std::string fetch_op)
|
||||
{
|
||||
if (fetch_op == "add")
|
||||
{
|
||||
return "constexpr auto __skip_v = __atomic_ptr_skip_t<_Type>::__skip;";
|
||||
}
|
||||
return "constexpr auto __skip_v = 1;";
|
||||
}
|
||||
|
||||
inline void FormatFetchOps(std::ostream& out)
|
||||
{
|
||||
const std::vector arithmetic_types = {
|
||||
Operand::Floating,
|
||||
Operand::Unsigned,
|
||||
Operand::Signed,
|
||||
};
|
||||
|
||||
const std::vector minmax_types = {
|
||||
Operand::Unsigned,
|
||||
Operand::Signed,
|
||||
};
|
||||
|
||||
const std::vector bitwise_types = {Operand::Bit};
|
||||
|
||||
const std::map op_support_map{
|
||||
std::pair{std::string{"add"}, std::pair{arithmetic_types, std::string{"arithmetic"}}},
|
||||
std::pair{std::string{"min"}, std::pair{minmax_types, std::string{"minmax"}}},
|
||||
std::pair{std::string{"max"}, std::pair{minmax_types, std::string{"minmax"}}},
|
||||
std::pair{std::string{"or"}, std::pair{bitwise_types, std::string{"bitwise"}}},
|
||||
std::pair{std::string{"xor"}, std::pair{bitwise_types, std::string{"bitwise"}}},
|
||||
std::pair{std::string{"and"}, std::pair{bitwise_types, std::string{"bitwise"}}},
|
||||
};
|
||||
|
||||
// Memory order dispatcher
|
||||
out << R"XXX(
|
||||
template <class _Fn, class _Sco>
|
||||
static inline _CCCL_DEVICE void __cuda_atomic_fetch_memory_order_dispatch(_Fn& __cuda_fetch, int __memorder, _Sco) {
|
||||
NV_DISPATCH_TARGET(
|
||||
NV_PROVIDES_SM_70, (
|
||||
switch (__memorder) {
|
||||
case __ATOMIC_SEQ_CST: __cuda_atomic_fence(_Sco{}, __atomic_cuda_seq_cst{}); [[fallthrough]];
|
||||
case __ATOMIC_CONSUME: [[fallthrough]];
|
||||
case __ATOMIC_ACQUIRE: __cuda_fetch(__atomic_cuda_acquire{}); break;
|
||||
case __ATOMIC_ACQ_REL: __cuda_fetch(__atomic_cuda_acq_rel{}); break;
|
||||
case __ATOMIC_RELEASE: __cuda_fetch(__atomic_cuda_release{}); break;
|
||||
case __ATOMIC_RELAXED: __cuda_fetch(__atomic_cuda_relaxed{}); break;
|
||||
default: _CCCL_ASSERT(false, "invalid memory order");
|
||||
}
|
||||
),
|
||||
NV_IS_DEVICE, (
|
||||
switch (__memorder) {
|
||||
case __ATOMIC_SEQ_CST: [[fallthrough]];
|
||||
case __ATOMIC_ACQ_REL: __cuda_atomic_membar(_Sco{}); [[fallthrough]];
|
||||
case __ATOMIC_CONSUME: [[fallthrough]];
|
||||
case __ATOMIC_ACQUIRE: __cuda_fetch(__atomic_cuda_volatile{}); __cuda_atomic_membar(_Sco{}); break;
|
||||
case __ATOMIC_RELEASE: __cuda_atomic_membar(_Sco{}); __cuda_fetch(__atomic_cuda_volatile{}); break;
|
||||
case __ATOMIC_RELAXED: __cuda_fetch(__atomic_cuda_volatile{}); break;
|
||||
default: _CCCL_ASSERT(false, "invalid memory order");
|
||||
}
|
||||
)
|
||||
)
|
||||
}
|
||||
)XXX";
|
||||
|
||||
// Argument ID Reference
|
||||
// 0 - Atomic Operation
|
||||
// 1 - Operand Type
|
||||
// 2 - Operand Size
|
||||
// 3 - Type Constraint
|
||||
// 4 - Memory Order
|
||||
// 5 - Memory Order function tag
|
||||
// 6 - Scope Constraint
|
||||
// 7 - Scope function tag
|
||||
constexpr auto asm_intrinsic_format = R"XXX(
|
||||
template <class _Type>
|
||||
static inline _CCCL_DEVICE void __cuda_atomic_fetch_{0}(
|
||||
_Type* __ptr, _Type& __dst, _Type __op, {5}, __atomic_cuda_operand_{1}{2}, {7})
|
||||
{{ asm volatile("atom.{0}{4}{6}.{1}{2} %0,[%1],%2;" : "={3}"(__dst) : "l"(__ptr), "{3}"(__op) : "memory"); }})XXX";
|
||||
|
||||
// 0 - Atomic Operation
|
||||
// 1 - Operand type constraint
|
||||
// 2 - Pointer op skip_v
|
||||
constexpr auto fetch_bind_invoke = R"XXX(
|
||||
template <typename _Type, typename _Tag, typename _Sco>
|
||||
struct __cuda_atomic_bind_fetch_{0} {{
|
||||
_Type* __ptr;
|
||||
_Type* __dst;
|
||||
_Type* __op;
|
||||
|
||||
template <typename _Atomic_Memorder>
|
||||
inline _CCCL_DEVICE void operator()(_Atomic_Memorder) {{
|
||||
__cuda_atomic_fetch_{0}(__ptr, *__dst, *__op, _Atomic_Memorder{{}}, _Tag{{}}, _Sco{{}});
|
||||
}}
|
||||
}};
|
||||
template <class _Type, class _Up, class _Sco, __atomic_enable_if_native_{1}<_Type> = 0>
|
||||
[[nodiscard]] static inline _CCCL_DEVICE _Type __atomic_fetch_{0}_cuda(_Type* __ptr, _Up __op, int __memorder, _Sco)
|
||||
{{
|
||||
{2}
|
||||
__op = __op * __skip_v;
|
||||
using __proxy_t = typename __atomic_cuda_deduce_{1}<_Type>::__type;
|
||||
using __proxy_tag = typename __atomic_cuda_deduce_{1}<_Type>::__tag;
|
||||
_Type __dst{{}};
|
||||
__proxy_t* __ptr_proxy = reinterpret_cast<__proxy_t*>(__ptr);
|
||||
__proxy_t* __dst_proxy = reinterpret_cast<__proxy_t*>(&__dst);
|
||||
__proxy_t* __op_proxy = reinterpret_cast<__proxy_t*>(&__op);
|
||||
if (__cuda_fetch_{0}_weak_if_local(__ptr_proxy, *__op_proxy, __dst_proxy)) {{return __dst;}}
|
||||
__cuda_atomic_bind_fetch_{0}<__proxy_t, __proxy_tag, _Sco> __bound_{0}{{__ptr_proxy, __dst_proxy, __op_proxy}};
|
||||
__cuda_atomic_fetch_memory_order_dispatch(__bound_{0}, __memorder, _Sco{{}});
|
||||
return __dst;
|
||||
}}
|
||||
template <class _Type, class _Up, class _Sco, __atomic_enable_if_native_{1}<_Type> = 0>
|
||||
[[nodiscard]] static inline _CCCL_DEVICE _Type __atomic_fetch_{0}_cuda(_Type volatile* __ptr, _Up __op, int __memorder, _Sco)
|
||||
{{
|
||||
{2}
|
||||
__op = __op * __skip_v;
|
||||
using __proxy_t = typename __atomic_cuda_deduce_{1}<_Type>::__type;
|
||||
using __proxy_tag = typename __atomic_cuda_deduce_{1}<_Type>::__tag;
|
||||
_Type __dst{{}};
|
||||
__proxy_t* __ptr_proxy = reinterpret_cast<__proxy_t*>(const_cast<_Type*>(__ptr));
|
||||
__proxy_t* __dst_proxy = reinterpret_cast<__proxy_t*>(&__dst);
|
||||
__proxy_t* __op_proxy = reinterpret_cast<__proxy_t*>(&__op);
|
||||
if (__cuda_fetch_{0}_weak_if_local(__ptr_proxy, *__op_proxy, __dst_proxy)) {{return __dst;}}
|
||||
__cuda_atomic_bind_fetch_{0}<__proxy_t, __proxy_tag, _Sco> __bound_{0}{{__ptr_proxy, __dst_proxy, __op_proxy}};
|
||||
__cuda_atomic_fetch_memory_order_dispatch(__bound_{0}, __memorder, _Sco{{}});
|
||||
return __dst;
|
||||
}}
|
||||
)XXX";
|
||||
|
||||
constexpr size_t supported_sizes[] = {
|
||||
32,
|
||||
64,
|
||||
};
|
||||
|
||||
constexpr Semantic supported_semantics[] = {
|
||||
Semantic::Acquire,
|
||||
Semantic::Relaxed,
|
||||
Semantic::Release,
|
||||
Semantic::Acq_Rel,
|
||||
Semantic::Volatile,
|
||||
};
|
||||
|
||||
constexpr Scope supported_scopes[] = {
|
||||
Scope::CTA,
|
||||
Scope::Cluster,
|
||||
Scope::GPU,
|
||||
Scope::System,
|
||||
};
|
||||
|
||||
for (auto& op_kp : op_support_map)
|
||||
{
|
||||
const auto& op_name = op_kp.first;
|
||||
const auto& op_type_kp = op_kp.second;
|
||||
const auto& type_list = op_type_kp.first;
|
||||
const auto& deduction = op_type_kp.second;
|
||||
for (auto type : type_list)
|
||||
{
|
||||
for (auto size : supported_sizes)
|
||||
{
|
||||
const std::string proxy_type = operand_proxy_type(type, size);
|
||||
for (auto sco : supported_scopes)
|
||||
{
|
||||
for (auto sem : supported_semantics)
|
||||
{
|
||||
// There is no atom.add.s64
|
||||
if (op_name == "add" && type == Operand::Signed && size == 64)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
out << std::format(
|
||||
asm_intrinsic_format,
|
||||
/* 0 */ op_name,
|
||||
/* 1 */ operand(type),
|
||||
/* 2 */ size,
|
||||
/* 3 */ constraints(type, size),
|
||||
/* 4 */ semantic(sem),
|
||||
/* 5 */ semantic_tag(sem),
|
||||
/* 6 */ scope(sco),
|
||||
/* 7 */ scope_tag(sco));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
out << "\n" << std::format(fetch_bind_invoke, op_name, deduction, fetch_op_skip_v(op_name));
|
||||
}
|
||||
|
||||
out << R"XXX(
|
||||
template <class _Type, class _Up, class _Sco>
|
||||
[[nodiscard]] static inline _CCCL_DEVICE _Type __atomic_fetch_sub_cuda(_Type* __ptr, _Up __op, int __memorder, _Sco)
|
||||
{
|
||||
return __atomic_fetch_add_cuda(__ptr, -__op, __memorder, _Sco{});
|
||||
}
|
||||
template <class _Type, class _Up, class _Sco>
|
||||
[[nodiscard]] static inline _CCCL_DEVICE _Type __atomic_fetch_sub_cuda(_Type volatile* __ptr, _Up __op, int __memorder, _Sco)
|
||||
{
|
||||
return __atomic_fetch_add_cuda(__ptr, -__op, __memorder, _Sco{});
|
||||
}
|
||||
)XXX";
|
||||
}
|
||||
|
||||
#endif // FETCH_OPS_H
|
||||
Reference in New Issue
Block a user