Skip to content

Commit 88bcc04

Browse files
author
Requiem
committed
2 parents ff21f74 + c50a8d7 commit 88bcc04

2 files changed

Lines changed: 124 additions & 59 deletions

File tree

‎CMakeLists.txt‎

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -92,6 +92,9 @@ elseif(LINUX)
9292
if(CMAKE_BUILD_TYPE MATCHES "Debug")
9393
message(STATUS "Build set to debug mode")
9494
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -DDEBUG -O0 -fsanitize=address,leak")
95+
elseif(CMAKE_BUILD_TYPE MATCHES "Dev_Release")
96+
message(STATUS "Build set to development release mode")
97+
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g0 -O2 -mno-avx512f -mno-avx512bw -mno-avx512vl")
9598
elseif(CMAKE_BUILD_TYPE MATCHES "Release")
9699
message(STATUS "Build set to release mode")
97100
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g0 -O2")
@@ -100,9 +103,6 @@ elseif(LINUX)
100103
elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm")
101104
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mcpu=native")
102105
endif()
103-
elseif(CMAKE_BUILD_TYPE MATCHES "Dev_Release")
104-
message(STATUS "Build set to development release mode")
105-
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g0 -O2 -mno-avx512f -mno-avx512bw -mno-avx512vl")
106106
endif()
107107
else()
108108
message(STATUS "Build set to release mode")

‎src/vmaware.hpp‎

Lines changed: 121 additions & 56 deletions
Original file line numberDiff line numberDiff line change
@@ -3473,6 +3473,25 @@ struct VM {
34733473
return (base_str.find(keyword) != std::string::npos);
34743474
};
34753475

3476+
[[nodiscard]] static i32 popcount(u64 v) {
3477+
#if (GCC) || (CLANG)
3478+
return __builtin_popcountll(v);
3479+
#elif (MSVC)
3480+
#if (x86_32)
3481+
return static_cast<int>(
3482+
__popcnt(static_cast<unsigned int>(v)) +
3483+
__popcnt(static_cast<unsigned int>(v >> 32))
3484+
);
3485+
#else
3486+
return static_cast<int>(__popcnt64(static_cast<unsigned long long>(v)));
3487+
#endif
3488+
#else
3489+
int c = 0;
3490+
while (v) { c += static_cast<int>(v & 1ull); v >>= 1; }
3491+
return c;
3492+
#endif
3493+
};
3494+
34763495
static std::string narrow_wide(const wchar_t* wstr) {
34773496
if (!wstr) return std::string{};
34783497
std::wstring ws(wstr);
@@ -4005,10 +4024,13 @@ struct VM {
40054024

40064025
while (true) {
40074026
char k = str[j];
4008-
const bool is_valid = (k >= '0' && k <= '9') ||
4027+
const bool is_valid = (
4028+
(k >= '0' && k <= '9') ||
40094029
(k >= 'A' && k <= 'Z') ||
40104030
(k >= 'a' && k <= 'z') ||
4011-
(k == '-'); // models have hyphen
4031+
(k == '-') // models have hyphen
4032+
);
4033+
40124034
if (!is_valid) break;
40134035

40144036
if (current_len >= max_model_len) {
@@ -4046,16 +4068,18 @@ struct VM {
40464068

40474069
// since it's a contiguous block of integers in .rodata/.rdata, this is extremely fast
40484070
for (size_t idx = 0; idx < db_size; ++idx) {
4049-
if (db[idx].hash == current_hash) {
4050-
if (current_len > best_len) {
4051-
best_len = current_len;
4052-
expected_threads = db[idx].threads;
4053-
found = true;
4054-
}
4055-
// since hashing implies uniqueness in this dataset, you might say we could break here,
4056-
// but we continue to ensure we find the longest substring match if overlaps exist,
4057-
// so like it finds both "i9-11900" and "i9-11900K" i.e.
4071+
if (db[idx].hash != current_hash) {
4072+
continue;
4073+
}
4074+
4075+
if (current_len > best_len) {
4076+
best_len = current_len;
4077+
expected_threads = db[idx].threads;
4078+
found = true;
40584079
}
4080+
// since hashing implies uniqueness in this dataset, you might say we could break here,
4081+
// but we continue to ensure we find the longest substring match if overlaps exist,
4082+
// so like it finds both "i9-11900" and "i9-11900K" i.e.
40594083
}
40604084
}
40614085
}
@@ -5019,105 +5043,145 @@ struct VM {
50195043
return false;
50205044
#else
50215045
auto is_smt_enabled = []() noexcept -> bool {
5022-
auto popcount = [](uint64_t v) noexcept -> int {
5023-
#if (GCC) || (CLANG)
5024-
return __builtin_popcountll(v);
5025-
#elif (MSVC)
5026-
#if (x86_32)
5027-
return static_cast<int>(
5028-
__popcnt(static_cast<unsigned int>(v)) +
5029-
__popcnt(static_cast<unsigned int>(v >> 32))
5030-
);
5031-
#else
5032-
return static_cast<int>(__popcnt64(static_cast<unsigned long long>(v)));
5033-
#endif
5034-
#else
5035-
int c = 0;
5036-
while (v) { c += static_cast<int>(v & 1ull); v >>= 1; }
5037-
return c;
5038-
#endif
5039-
};
50405046
#if (WINDOWS)
50415047
DWORD len = 0;
50425048
if (GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &len) ||
50435049
GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
50445050
return false;
50455051
}
5052+
50465053
std::vector<char> buf(static_cast<size_t>(len));
50475054
if (!GetLogicalProcessorInformationEx(RelationProcessorCore,
50485055
reinterpret_cast<SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*>(buf.data()), &len)) {
50495056
return false;
50505057
}
5058+
50515059
// first RelationProcessorCore record encountered, basically if two logical processors maps to the same core, SMT is enabled to the OS point of view
50525060
size_t offset = 0;
50535061
while (offset + sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX) <= static_cast<size_t>(len)) {
50545062
auto rec = reinterpret_cast<SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*>(buf.data() + offset);
5063+
50555064
if (rec->Relationship == RelationProcessorCore) {
50565065
const PROCESSOR_RELATIONSHIP& pr = rec->Processor;
50575066
unsigned total = 0;
5067+
50585068
for (WORD i = 0; i < pr.GroupCount; ++i) {
5059-
total += popcount(static_cast<uint64_t>(pr.GroupMask[i].Mask));
5069+
total += util::popcount(static_cast<uint64_t>(pr.GroupMask[i].Mask));
50605070
}
5071+
50615072
return total > 1;
50625073
}
5074+
50635075
if (rec->Size == 0) break;
50645076
offset += rec->Size;
50655077
}
50665078
return false;
50675079
#elif (APPLE)
50685080
int logical = 0, physical = 0;
50695081
size_t sz = sizeof(logical);
5070-
if (sysctlbyname("hw.logicalcpu", &logical, &sz, nullptr, 0) != 0) logical = 0;
5082+
5083+
if (sysctlbyname("hw.logicalcpu", &logical, &sz, nullptr, 0) != 0) {
5084+
logical = 0;
5085+
}
5086+
50715087
sz = sizeof(physical);
5072-
if (sysctlbyname("hw.physicalcpu", &physical, &sz, nullptr, 0) != 0) physical = 0;
5073-
if (logical > 0 && physical > 0) return logical > physical;
5088+
5089+
if (sysctlbyname("hw.physicalcpu", &physical, &sz, nullptr, 0) != 0) {
5090+
physical = 0;
5091+
}
5092+
5093+
if (logical > 0 && physical > 0) {
5094+
return logical > physical;
5095+
}
5096+
50745097
return false;
50755098
#else
50765099
// check cpu0 thread_siblings_list
5077-
{
5078-
std::ifstream f("/sys/devices/system/cpu/cpu0/topology/thread_siblings_list");
5079-
if (f) {
5080-
std::string s;
5081-
if (std::getline(f, s)) {
5082-
// trim
5083-
size_t a = 0; while (a < s.size() && std::isspace(static_cast<unsigned char>(s[a]))) ++a;
5084-
size_t b = s.size(); while (b > a && std::isspace(static_cast<unsigned char>(s[b - 1]))) --b;
5085-
if (b > a) {
5086-
for (size_t k = a; k < b; ++k) {
5087-
if (s[k] == ',' || s[k] == '-') return true;
5088-
}
5089-
return false;
5100+
std::ifstream f("/sys/devices/system/cpu/cpu0/topology/thread_siblings_list");
5101+
if (f) {
5102+
std::string s;
5103+
if (std::getline(f, s)) {
5104+
// trim
5105+
size_t a = 0;
5106+
5107+
while (a < s.size() && std::isspace(static_cast<u8>(s[a]))) {
5108+
++a;
5109+
}
5110+
5111+
size_t b = s.size();
5112+
5113+
while (b > a && std::isspace(static_cast<u8>(s[b - 1]))) {
5114+
--b;
5115+
}
5116+
5117+
if (b > a) {
5118+
for (size_t k = a; k < b; ++k) {
5119+
if (s[k] == ',' || s[k] == '-') return true;
50905120
}
5121+
return false;
50915122
}
50925123
}
50935124
}
5125+
50945126
// /proc/cpuinfo for unique (physical id, core id) pairs vs processors
50955127
std::ifstream cpuinfo("/proc/cpuinfo");
5096-
if (!cpuinfo) return false;
5128+
5129+
if (!cpuinfo) {
5130+
return false;
5131+
}
5132+
50975133
std::string line;
50985134
int processors = 0;
50995135
bool in_section = false;
51005136
int cur_phys = -1, cur_core = -1;
51015137
std::vector<std::pair<int, int>> cores;
5138+
51025139
while (std::getline(cpuinfo, line)) {
51035140
if (line.empty()) {
5104-
if (cur_phys != -1 && cur_core != -1) cores.emplace_back(cur_phys, cur_core);
5141+
if (cur_phys != -1 && cur_core != -1) {
5142+
cores.emplace_back(cur_phys, cur_core);
5143+
}
5144+
51055145
cur_phys = cur_core = -1;
51065146
in_section = false;
51075147
continue;
51085148
}
5149+
51095150
auto pos = line.find(':');
51105151
if (pos == std::string::npos) continue;
51115152
std::string key = line.substr(0, pos);
51125153
std::string val = line.substr(pos + 1);
5154+
51135155
// trim
5114-
while (!key.empty() && std::isspace(static_cast<unsigned char>(key.back()))) key.pop_back();
5115-
while (!val.empty() && std::isspace(static_cast<unsigned char>(val.front()))) val.erase(val.begin());
5116-
if (key == "processor") ++processors;
5117-
else if (key == "physical id") { try { cur_phys = std::stoi(val); } catch (...) { cur_phys = -1; } }
5118-
else if (key == "core id") { try { cur_core = std::stoi(val); } catch (...) { cur_core = -1; } }
5156+
while (!key.empty() && std::isspace(static_cast<u8>(key.back()))) {
5157+
key.pop_back();
5158+
}
5159+
5160+
while (!val.empty() && std::isspace(static_cast<u8>(val.front()))) {
5161+
val.erase(val.begin());
5162+
}
5163+
5164+
if (key == "processor") {
5165+
processors++;
5166+
} else if (key == "physical id") {
5167+
try {
5168+
cur_phys = std::stoi(val);
5169+
} catch (...) {
5170+
cur_phys = -1;
5171+
}
5172+
} else if (key == "core id") {
5173+
try {
5174+
cur_core = std::stoi(val);
5175+
} catch (...) {
5176+
cur_core = -1;
5177+
}
5178+
}
51195179
}
5120-
if (cur_phys != -1 && cur_core != -1) cores.emplace_back(cur_phys, cur_core);
5180+
5181+
if (cur_phys != -1 && cur_core != -1) {
5182+
cores.emplace_back(cur_phys, cur_core);
5183+
}
5184+
51215185
if (!cores.empty() && processors > 0) {
51225186
std::sort(cores.begin(), cores.end());
51235187
cores.erase(std::unique(cores.begin(), cores.end()), cores.end());
@@ -5134,14 +5198,15 @@ struct VM {
51345198
debug(info.debug_tag, ": CPU model = ", info.model_name);
51355199

51365200
const u32 actual = memo::threadcount::fetch();
5201+
51375202
if (actual != info.expected_threads) {
51385203
debug(info.debug_tag, ": Current threads -> ", actual);
51395204
const bool smt = is_smt_enabled();
5205+
51405206
if (smt) {
51415207
debug(info.debug_tag, ": Expected ", info.expected_threads, " threads");
51425208
return true;
5143-
}
5144-
else {
5209+
} else {
51455210
debug(info.debug_tag, ": Expected ", info.expected_threads, " threads, but found SMT disabled");
51465211
return false;
51475212
}

0 commit comments

Comments
 (0)