下面是一些使用GetLogicalProcessorInformation在Windows上检测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，我希望获得共享缓存的逻辑CPU核心数量。在具有超线程的x64 CPU上，两个逻辑CPU内核通常共享一个L2缓存，所有逻辑CPU内核共享L3缓存。
通过MSDN阅读后，我认为GetLogicalProcessorInformationEx和CACHE_RELATIONSHIP以及GROUP_AFFINITY我正在寻找的数据结构，但在尝试了这些数据结构后，对我来说似乎毫无用处。 / p>
问题：
有没有办法让Windows上使用C / C ++共享缓存的逻辑CPU核心数量？ （理想情况下不直接使用cpuid）
解决方案：
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核心数。 GROUP_AFFINITY.Mask值包含为每个共享当前缓存（RelationCache）的CPU核心设置的一个位。作为大多数具有超线程{1}的英特尔CPU的示例，将包含为L2高速缓存设置的2位和为具有4个物理CPU核心和8个逻辑CPU核心的CPU设置的用于L3高速缓存的8位。
这是C ++代码：
GROUP_AFFINITY.Mask注意事项：
我发现在虚拟机内部运行Windows时，上面的代码无法正确检测共享缓存的CPU核心数量，例如：在具有2个虚拟CPU核心的VM上，上面的代码报告每个逻辑CPU核心都有一个专用的L1，L2和L3缓存。
答案 0 :(得分：1)
@RbMm：但CACHE_RELATIONSHIP包含所需的所有信息。逻辑CPU核心数= Cache-＆gt; GroupMask.Mask
中设置的位数
我在AppVeyor CI上测试了这个（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
根据MSDN文档：＆＃34; GroupMask.Mask - 指定指定组内零个或多个处理器的亲和性的位图。＆＃34;。基于此文档，我预计L3缓存会有不同的GroupMask.Mask，但上面的输出并未显示此信息。对我来说，GroupMask.Mask中的数据毫无意义！
这是link to the code，它产生上面的数据
答案 1 :(得分：0)
可能是特定处理器和Windows版本的结果，但我在2个处理器上测试（win10）并得到了正确的结果：
i5（2核，4个线程）：
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个线程）：
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);
}
答案 2 :(得分：-1)
还有一个来自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。