[英]Get the number of logical CPU cores sharing a cache (L1, L2, L3)
下面是一些使用GetLogicalProcessorInformation 检测Windows 上 L1、L2 和 L3 CPU 缓存大小的 C++ 代码:
typedef BOOL (WINAPI *LPFN_GLPI)(PSYSTEM_LOGICAL_PROCESSOR_INFORMATION, PDWORD);
LPFN_GLPI glpi = (LPFN_GLPI) GetProcAddress(
GetModuleHandle(TEXT("kernel32")), "GetLogicalProcessorInformation");
if (glpi)
{
DWORD bytes = 0;
glpi(0, &bytes);
size_t size = bytes / sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION);
vector<SYSTEM_LOGICAL_PROCESSOR_INFORMATION> info(size);
glpi(info.data(), &bytes);
for (size_t i = 0; i < size; i++)
{
if (info[i].Relationship == RelationCache)
{
if (info[i].Cache.Level == 1)
l1_cache_Size = info[i].Cache.Size;
if (info[i].Cache.Level == 2)
l2_cache_Size = info[i].Cache.Size;
if (info[i].Cache.Level == 3)
l3_cache_Size = info[i].Cache.Size;
}
}
}
作为下一步,我想获得共享缓存的逻辑 CPU 内核的数量。 在具有超线程的 x64 CPU 上,两个逻辑 CPU 内核通常共享 L2 缓存,所有逻辑 CPU 内核共享 L3 缓存。
通读 MSDN 后,我认为GetLogicalProcessorInformationEx
和CACHE_RELATIONSHIP和GROUP_AFFINITY是我正在寻找的数据结构,但在尝试之后,这些数据结构对我的目的似乎毫无用处。
问题:
有没有办法使用 C/C++ 在 Windows 上获取共享缓存的逻辑 CPU 内核的数量? (理想情况下不直接使用cpuid
)
解决方法:
可以使用GetLogicalProcessorInformationEx
以及CACHE_RELATIONSHIP和GROUP_AFFINITY数据结构获取共享缓存的逻辑 CPU 内核的数量。 GROUP_AFFINITY.Mask
值包含为共享当前缓存 ( RelationCache
) 的每个 CPU 内核设置的一位。 作为大多数具有超线程的 Intel CPU 的示例GROUP_AFFINITY.Mask
将包含为 L2 缓存设置的 2 位和为具有 4 个物理 CPU 内核和 8 个逻辑 CPU 内核的 CPU 设置的 L3 缓存的 8 位。
这是C++代码:
#include <windows.h>
#include <vector>
#include <iostream>
using namespace std;
typedef BOOL (WINAPI *LPFN_GLPI)(LOGICAL_PROCESSOR_RELATIONSHIP,
PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);
int main()
{
LPFN_GLPI glpi = (LPFN_GLPI) GetProcAddress(
GetModuleHandle(TEXT("kernel32")), "GetLogicalProcessorInformationEx");
if (!glpi)
return 1;
DWORD bytes = 0;
glpi(RelationAll, 0, &bytes);
vector<char> buffer(bytes);
SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX* info;
if (!glpi(RelationAll, (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*) &buffer[0], &bytes))
return 1;
for (size_t i = 0; i < bytes; i += info->Size)
{
info = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*) &buffer[i];
if (info->Relationship == RelationCache &&
(info->Cache.Type == CacheData ||
info->Cache.Type == CacheUnified))
{
cout << "info->Cache.Level: " << (int) info->Cache.Level << endl;
cout << "info->Cache.CacheSize: " << (int) info->Cache.CacheSize << endl;
cout << "info->Cache.GroupMask.Group: " << info->Cache.GroupMask.Group << endl;
cout << "info->Cache.GroupMask.Mask: " << info->Cache.GroupMask.Mask << endl << endl;
}
}
return 0;
}
注意事项:
我发现在虚拟机中运行 Windows 时,上面的代码无法正确检测共享缓存的 CPU 内核数,例如在具有 2 个虚拟 CPU 内核的 VM 上,上面的代码报告每个逻辑 CPU 内核都有一个私有 L1 ,L2和L3缓存。
@RbMm:但 CACHE_RELATIONSHIP 包含所需的所有信息。 逻辑 CPU 核心数 = Cache->GroupMask.Mask 中设置的位数
我已经在 AppVeyor CI 上测试过这个(甚至在发布到 stackoverflow 之前)。 这是 x64 CPU 的输出:
info->Cache.Level: 1
info->Cache.CacheSize: 32768
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 1
info->Cache.Level: 1
info->Cache.CacheSize: 32768
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 1
info->Cache.Level: 2
info->Cache.CacheSize: 262144
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 1
info->Cache.Level: 3
info->Cache.CacheSize: 31457280
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 1
info->Cache.Level: 1
info->Cache.CacheSize: 32768
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 2
info->Cache.Level: 1
info->Cache.CacheSize: 32768
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 2
info->Cache.Level: 2
info->Cache.CacheSize: 262144
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 2
info->Cache.Level: 3
info->Cache.CacheSize: 31457280
info->Cache.GroupMask.Group: 0
info->Cache.GroupMask.Mask: 2
或者:
| Cache Level | Processor 1 | Processor 2 |
|-------------|--------------------|--------------------|
| L1 | 32 KB Data | 32 KB Data |
| | 32 KB Instruction | 32 KB Instruction |
|-------------|--------------------|--------------------|
| L2 | 256 KB Unified | 256 KB Unified |
|-------------|--------------------|--------------------|
| L3 | 30 MB Unified | 30 MB Unified |
根据 MSDN 文档:
GroupMask.Mask - 一个位图,指定指定组内零个或多个处理器的关联。
基于此文档,我期望 L3 缓存使用不同的GroupMask.Mask
,但上面的输出未显示这一点。 对我来说GroupMask.Mask
的数据毫无意义!
结果可能取决于特定的处理器和 Windows 版本,但我在 2 个处理器上测试 (win10) 并得到正确的结果:
i5(2 核,4 线程):
ProcessorPackage
[G0 000000000000000F { #3, #2, #1, #0}]
ProcessorCore HP=1 0
[G0 0000000000000003 { #1, #0}]
Cache L1 8000 40 [G0 0000000000000003 { #1, #0}] A=8 Data
Cache L1 8000 40 [G0 0000000000000003 { #1, #0}] A=8 Instruction
Cache L2 40000 40 [G0 0000000000000003 { #1, #0}] A=8 Unified
Cache L3 300000 40 [G0 000000000000000F { #3, #2, #1, #0}] A=c Unified
ProcessorCore HP=1 0
[G0 000000000000000C { #3, #2}]
Cache L1 8000 40 [G0 000000000000000C { #3, #2}] A=8 Data
Cache L1 8000 40 [G0 000000000000000C { #3, #2}] A=8 Instruction
Cache L2 40000 40 [G0 000000000000000C { #3, #2}] A=8 Unified
NumaNode #0 [G0 000000000000000F { #3, #2, #1, #0}]
Group:1/1
4/4 [000000000000000F { #3, #2, #1, #0}]
i7(4 核,8 线程):
ProcessorPackage
[G0 00000000000000FF { #7, #6, #5, #4, #3, #2, #1, #0}]
ProcessorCore HP=1 0
[G0 0000000000000003 { #1, #0}]
Cache L1 8000 40 [G0 0000000000000003 { #1, #0}] A=8 Data
Cache L1 8000 40 [G0 0000000000000003 { #1, #0}] A=8 Instruction
Cache L2 40000 40 [G0 0000000000000003 { #1, #0}] A=4 Unified
Cache L3 800000 40 [G0 00000000000000FF { #7, #6, #5, #4, #3, #2, #1, #0}] A=10 Unified
ProcessorCore HP=1 0
[G0 000000000000000C { #3, #2}]
Cache L1 8000 40 [G0 000000000000000C { #3, #2}] A=8 Data
Cache L1 8000 40 [G0 000000000000000C { #3, #2}] A=8 Instruction
Cache L2 40000 40 [G0 000000000000000C { #3, #2}] A=4 Unified
ProcessorCore HP=1 0
[G0 0000000000000030 { #5, #4}]
Cache L1 8000 40 [G0 0000000000000030 { #5, #4}] A=8 Data
Cache L1 8000 40 [G0 0000000000000030 { #5, #4}] A=8 Instruction
Cache L2 40000 40 [G0 0000000000000030 { #5, #4}] A=4 Unified
ProcessorCore HP=1 0
[G0 00000000000000C0 { #7, #6}]
Cache L1 8000 40 [G0 00000000000000C0 { #7, #6}] A=8 Data
Cache L1 8000 40 [G0 00000000000000C0 { #7, #6}] A=8 Instruction
Cache L2 40000 40 [G0 00000000000000C0 { #7, #6}] A=4 Unified
NumaNode #0 [G0 00000000000000FF { #7, #6, #5, #4, #3, #2, #1, #0}]
Group:1/1
8/8 [00000000000000FF { #7, #6, #5, #4, #3, #2, #1, #0}]
代码:
void FormatMask(KAFFINITY Mask, PSTR sz)
{
sz += sprintf(sz, "%p {", (PVOID)Mask);
ULONG i = sizeof(KAFFINITY) * 8;
do
{
if (_bittest((PLONG)&Mask, --i))
{
sz += sprintf(sz, " #%u,", i);
}
} while (i);
*--sz = '}';
}
void DumpCpuInfo()
{
static PCSTR szCacheType[] = {
"Unified",
"Instruction",
"Data",
"Trace"
};
char szMask[64 * 5 + 19];
ULONG cb = 0, rcb = 0x400;
static volatile UCHAR guz;
PVOID stack = alloca(guz);
union {
PVOID Buffer;
PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX plpi;
};
do
{
if (cb < rcb) rcb = cb = RtlPointerToOffset(Buffer = alloca(rcb - cb), stack);
if (GetLogicalProcessorInformationEx(::RelationAll, plpi, &rcb))
{
DWORD Size;
do
{
Size = plpi->Size;
union {
PPROCESSOR_RELATIONSHIP Processor;
PNUMA_NODE_RELATIONSHIP NumaNode;
PCACHE_RELATIONSHIP Cache;
PGROUP_RELATIONSHIP Group;
PVOID pv;
};
pv = &plpi->Processor;
switch (plpi->Relationship)
{
case RelationProcessorPackage:
DbgPrint("ProcessorPackage\n");
goto __0;
case RelationProcessorCore:
DbgPrint("ProcessorCore HP=%x %x\n",
Processor->Flags & LTP_PC_SMT ? 1 : 0, Processor->EfficiencyClass);
__0:
if (WORD GroupCount = Processor->GroupCount)
{
PGROUP_AFFINITY GroupMask = Processor->GroupMask;
do
{
FormatMask(GroupMask->Mask, szMask);
DbgPrint("\t[G%u %s]\n", GroupMask->Group, szMask);
} while (GroupMask++, --GroupCount);
}
break;
case RelationNumaNode:
FormatMask(NumaNode->GroupMask.Mask, szMask);
DbgPrint("NumaNode #%u [G%u %s]\n",
NumaNode->NodeNumber, NumaNode->GroupMask.Group, szMask);
break;
case RelationGroup:
DbgPrint("Group:%u/%u\n", Group->ActiveGroupCount, Group->MaximumGroupCount);
if (WORD ActiveGroupCount = Group->ActiveGroupCount)
{
PPROCESSOR_GROUP_INFO GroupInfo = Group->GroupInfo;
do
{
FormatMask(GroupInfo->ActiveProcessorMask, szMask);
DbgPrint("\t%u/%u [%s]\n",
GroupInfo->ActiveProcessorCount,
GroupInfo->MaximumProcessorCount, szMask);
} while (GroupInfo, --ActiveGroupCount);
}
break;
case RelationCache:
FormatMask(Cache->GroupMask.Mask, szMask);
DbgPrint("Cache L%u %8x %2x [G%u %s] A=%x %s\n",
Cache->Level,
Cache->CacheSize, Cache->LineSize,
Cache->GroupMask.Group, szMask,
Cache->Associativity,
szCacheType[Cache->Type % RTL_NUMBER_OF(szCacheType)]
);
break;
}
Buffer = RtlOffsetToPointer(plpi, Size);
} while (rcb -= Size);
break;
}
} while (GetLastError() == ERROR_INSUFFICIENT_BUFFER);
}
有一个来自 boost 库的替代解决方案。
// number of logical cores
auto logical = boost::thread::hardware_concurrency();
// number of physical cores
auto physical = boost::thread::physical_concurrency();
但不考虑缓存。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.