2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
5 Stockfish is free software: you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation, either version 3 of the License, or
8 (at your option) any later version.
10 Stockfish is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with this program. If not, see <http://www.gnu.org/licenses/>.
20 #if _WIN32_WINNT < 0x0601
22 #define _WIN32_WINNT 0x0601 // Force to include needed API prototypes
30 // The needed Windows API for processor groups could be missed from old Windows
31 // versions, so instead of calling them directly (forcing the linker to resolve
32 // the calls at compile time), try to load them at runtime. To do this we need
33 // first to define the corresponding function pointers.
35 using fun1_t = bool(*)(LOGICAL_PROCESSOR_RELATIONSHIP,
36 PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);
37 using fun2_t = bool(*)(USHORT, PGROUP_AFFINITY);
38 using fun3_t = bool(*)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
39 using fun4_t = bool(*)(USHORT, PGROUP_AFFINITY, USHORT, PUSHORT);
40 using fun5_t = WORD(*)();
50 #include <string_view>
53 #if defined(__linux__) && !defined(__ANDROID__)
58 #if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) || defined(__e2k__)
59 #define POSIXALIGNEDALLOC
72 /// Version number or dev.
73 constexpr string_view version = "dev";
75 /// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
76 /// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
77 /// can toggle the logging of std::cout and std:cin at runtime whilst preserving
78 /// usual I/O functionality, all without changing a single line of code!
79 /// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
81 struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout
83 Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {}
85 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
86 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
87 int underflow() override { return buf->sgetc(); }
88 int uflow() override { return log(buf->sbumpc(), ">> "); }
90 streambuf *buf, *logBuf;
92 int log(int c, const char* prefix) {
94 static int last = '\n'; // Single log file
97 logBuf->sputn(prefix, 3);
99 return last = logBuf->sputc((char)c);
105 Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {}
106 ~Logger() { start(""); }
112 static void start(const std::string& fname) {
116 if (l.file.is_open())
118 cout.rdbuf(l.out.buf);
125 l.file.open(fname, ifstream::out);
127 if (!l.file.is_open())
129 cerr << "Unable to open debug log file " << fname << endl;
142 /// engine_info() returns the full name of the current Stockfish version.
143 /// For local dev compiles we try to append the commit sha and commit date
144 /// from git if that fails only the local compilation date is set and "nogit" is specified:
145 /// Stockfish dev-YYYYMMDD-SHA
147 /// Stockfish dev-YYYYMMDD-nogit
149 /// For releases (non dev builds) we only include the version number:
150 /// Stockfish version
152 string engine_info(bool to_uci) {
154 ss << "Stockfish " << version << setfill('0');
156 if constexpr (version == "dev")
162 constexpr string_view months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
163 string month, day, year;
164 stringstream date(__DATE__); // From compiler, format is "Sep 21 2008"
166 date >> month >> day >> year;
167 ss << year << setw(2) << setfill('0') << (1 + months.find(month) / 4) << setw(2) << setfill('0') << day;
179 ss << (to_uci ? "\nid author ": " by ")
180 << "the Stockfish developers (see AUTHORS file)";
186 /// compiler_info() returns a string trying to describe the compiler we use
188 std::string compiler_info() {
190 #define stringify2(x) #x
191 #define stringify(x) stringify2(x)
192 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
194 /// Predefined macros hell:
196 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
197 /// __INTEL_COMPILER Compiler is Intel
198 /// _MSC_VER Compiler is MSVC or Intel on Windows
199 /// _WIN32 Building on Windows (any)
200 /// _WIN64 Building on Windows 64 bit
202 std::string compiler = "\nCompiled by ";
205 compiler += "clang++ ";
206 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
207 #elif __INTEL_COMPILER
208 compiler += "Intel compiler ";
209 compiler += "(version ";
210 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
214 compiler += "(version ";
215 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
217 #elif defined(__e2k__) && defined(__LCC__)
218 #define dot_ver2(n) \
219 compiler += (char)'.'; \
220 compiler += (char)('0' + (n) / 10); \
221 compiler += (char)('0' + (n) % 10);
223 compiler += "MCST LCC ";
224 compiler += "(version ";
225 compiler += std::to_string(__LCC__ / 100);
226 dot_ver2(__LCC__ % 100)
227 dot_ver2(__LCC_MINOR__)
230 compiler += "g++ (GNUC) ";
231 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
233 compiler += "Unknown compiler ";
234 compiler += "(unknown version)";
237 #if defined(__APPLE__)
238 compiler += " on Apple";
239 #elif defined(__CYGWIN__)
240 compiler += " on Cygwin";
241 #elif defined(__MINGW64__)
242 compiler += " on MinGW64";
243 #elif defined(__MINGW32__)
244 compiler += " on MinGW32";
245 #elif defined(__ANDROID__)
246 compiler += " on Android";
247 #elif defined(__linux__)
248 compiler += " on Linux";
249 #elif defined(_WIN64)
250 compiler += " on Microsoft Windows 64-bit";
251 #elif defined(_WIN32)
252 compiler += " on Microsoft Windows 32-bit";
254 compiler += " on unknown system";
257 compiler += "\nCompilation settings include: ";
258 compiler += (Is64Bit ? " 64bit" : " 32bit");
259 #if defined(USE_VNNI)
262 #if defined(USE_AVX512)
263 compiler += " AVX512";
265 compiler += (HasPext ? " BMI2" : "");
266 #if defined(USE_AVX2)
269 #if defined(USE_SSE41)
270 compiler += " SSE41";
272 #if defined(USE_SSSE3)
273 compiler += " SSSE3";
275 #if defined(USE_SSE2)
278 compiler += (HasPopCnt ? " POPCNT" : "");
282 #if defined(USE_NEON)
287 compiler += " DEBUG";
290 compiler += "\n__VERSION__ macro expands to: ";
292 compiler += __VERSION__;
294 compiler += "(undefined macro)";
302 /// Debug functions used mainly to collect run-time statistics
303 constexpr int MaxDebugSlots = 32;
309 std::atomic<int64_t> data[N] = { 0 };
311 constexpr inline std::atomic<int64_t>& operator[](int index) { return data[index]; }
314 DebugInfo<2> hit[MaxDebugSlots];
315 DebugInfo<2> mean[MaxDebugSlots];
316 DebugInfo<3> stdev[MaxDebugSlots];
317 DebugInfo<6> correl[MaxDebugSlots];
321 void dbg_hit_on(bool cond, int slot) {
328 void dbg_mean_of(int64_t value, int slot) {
331 mean[slot][1] += value;
334 void dbg_stdev_of(int64_t value, int slot) {
337 stdev[slot][1] += value;
338 stdev[slot][2] += value * value;
341 void dbg_correl_of(int64_t value1, int64_t value2, int slot) {
344 correl[slot][1] += value1;
345 correl[slot][2] += value1 * value1;
346 correl[slot][3] += value2;
347 correl[slot][4] += value2 * value2;
348 correl[slot][5] += value1 * value2;
354 auto E = [&n](int64_t x) { return double(x) / n; };
355 auto sqr = [](double x) { return x * x; };
357 for (int i = 0; i < MaxDebugSlots; ++i)
359 std::cerr << "Hit #" << i
360 << ": Total " << n << " Hits " << hit[i][1]
361 << " Hit Rate (%) " << 100.0 * E(hit[i][1])
364 for (int i = 0; i < MaxDebugSlots; ++i)
365 if ((n = mean[i][0]))
367 std::cerr << "Mean #" << i
368 << ": Total " << n << " Mean " << E(mean[i][1])
372 for (int i = 0; i < MaxDebugSlots; ++i)
373 if ((n = stdev[i][0]))
375 double r = sqrtl(E(stdev[i][2]) - sqr(E(stdev[i][1])));
376 std::cerr << "Stdev #" << i
377 << ": Total " << n << " Stdev " << r
381 for (int i = 0; i < MaxDebugSlots; ++i)
382 if ((n = correl[i][0]))
384 double r = (E(correl[i][5]) - E(correl[i][1]) * E(correl[i][3]))
385 / ( sqrtl(E(correl[i][2]) - sqr(E(correl[i][1])))
386 * sqrtl(E(correl[i][4]) - sqr(E(correl[i][3]))));
387 std::cerr << "Correl. #" << i
388 << ": Total " << n << " Coefficient " << r
394 /// Used to serialize access to std::cout to avoid multiple threads writing at
397 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
411 /// Trampoline helper to avoid moving Logger to misc.h
412 void start_logger(const std::string& fname) { Logger::start(fname); }
415 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
416 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
417 /// which can be quite slow.
420 void prefetch(void*) {}
424 void prefetch(void* addr) {
426 # if defined(__INTEL_COMPILER)
427 // This hack prevents prefetches from being optimized away by
428 // Intel compiler. Both MSVC and gcc seem not be affected by this.
432 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
433 _mm_prefetch((char*)addr, _MM_HINT_T0);
435 __builtin_prefetch(addr);
442 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
443 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
444 /// std_aligned_alloc() must be freed with std_aligned_free().
446 void* std_aligned_alloc(size_t alignment, size_t size) {
448 #if defined(POSIXALIGNEDALLOC)
450 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
451 #elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
452 return _mm_malloc(size, alignment);
453 #elif defined(_WIN32)
454 return _aligned_malloc(size, alignment);
456 return std::aligned_alloc(alignment, size);
460 void std_aligned_free(void* ptr) {
462 #if defined(POSIXALIGNEDALLOC)
464 #elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
466 #elif defined(_WIN32)
473 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
477 static void* aligned_large_pages_alloc_windows([[maybe_unused]] size_t allocSize) {
483 HANDLE hProcessToken { };
487 const size_t largePageSize = GetLargePageMinimum();
491 // We need SeLockMemoryPrivilege, so try to enable it for the process
492 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
495 if (LookupPrivilegeValue(nullptr, SE_LOCK_MEMORY_NAME, &luid))
497 TOKEN_PRIVILEGES tp { };
498 TOKEN_PRIVILEGES prevTp { };
501 tp.PrivilegeCount = 1;
502 tp.Privileges[0].Luid = luid;
503 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
505 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
506 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
507 if (AdjustTokenPrivileges(
508 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
509 GetLastError() == ERROR_SUCCESS)
511 // Round up size to full pages and allocate
512 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
514 nullptr, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
516 // Privilege no longer needed, restore previous state
517 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, nullptr, nullptr);
521 CloseHandle(hProcessToken);
528 void* aligned_large_pages_alloc(size_t allocSize) {
530 // Try to allocate large pages
531 void* mem = aligned_large_pages_alloc_windows(allocSize);
533 // Fall back to regular, page aligned, allocation if necessary
535 mem = VirtualAlloc(nullptr, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
542 void* aligned_large_pages_alloc(size_t allocSize) {
544 #if defined(__linux__)
545 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
547 constexpr size_t alignment = 4096; // assumed small page size
550 // round up to multiples of alignment
551 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
552 void *mem = std_aligned_alloc(alignment, size);
553 #if defined(MADV_HUGEPAGE)
554 madvise(mem, size, MADV_HUGEPAGE);
562 /// aligned_large_pages_free() will free the previously allocated ttmem
566 void aligned_large_pages_free(void* mem) {
568 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
570 DWORD err = GetLastError();
571 std::cerr << "Failed to free large page memory. Error code: 0x"
573 << std::dec << std::endl;
580 void aligned_large_pages_free(void *mem) {
581 std_aligned_free(mem);
587 namespace WinProcGroup {
591 void bindThisThread(size_t) {}
595 /// best_node() retrieves logical processor information using Windows specific
596 /// API and returns the best node id for the thread with index idx. Original
597 /// code from Texel by Peter Ă–sterlund.
599 static int best_node(size_t idx) {
604 DWORD returnLength = 0;
605 DWORD byteOffset = 0;
607 // Early exit if the needed API is not available at runtime
608 HMODULE k32 = GetModuleHandle("Kernel32.dll");
609 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
613 // First call to GetLogicalProcessorInformationEx() to get returnLength.
614 // We expect the call to fail due to null buffer.
615 if (fun1(RelationAll, nullptr, &returnLength))
618 // Once we know returnLength, allocate the buffer
619 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
620 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
622 // Second call to GetLogicalProcessorInformationEx(), now we expect to succeed
623 if (!fun1(RelationAll, buffer, &returnLength))
629 while (byteOffset < returnLength)
631 if (ptr->Relationship == RelationNumaNode)
634 else if (ptr->Relationship == RelationProcessorCore)
637 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
641 byteOffset += ptr->Size;
642 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
647 std::vector<int> groups;
649 // Run as many threads as possible on the same node until core limit is
650 // reached, then move on filling the next node.
651 for (int n = 0; n < nodes; n++)
652 for (int i = 0; i < cores / nodes; i++)
655 // In case a core has more than one logical processor (we assume 2) and we
656 // have still threads to allocate, then spread them evenly across available
658 for (int t = 0; t < threads - cores; t++)
659 groups.push_back(t % nodes);
661 // If we still have more threads than the total number of logical processors
662 // then return -1 and let the OS to decide what to do.
663 return idx < groups.size() ? groups[idx] : -1;
667 /// bindThisThread() set the group affinity of the current thread
669 void bindThisThread(size_t idx) {
671 // Use only local variables to be thread-safe
672 int node = best_node(idx);
677 // Early exit if the needed API are not available at runtime
678 HMODULE k32 = GetModuleHandle("Kernel32.dll");
679 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
680 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
681 auto fun4 = (fun4_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMask2");
682 auto fun5 = (fun5_t)(void(*)())GetProcAddress(k32, "GetMaximumProcessorGroupCount");
689 GROUP_AFFINITY affinity;
690 if (fun2(node, &affinity)) // GetNumaNodeProcessorMaskEx
691 fun3(GetCurrentThread(), &affinity, nullptr); // SetThreadGroupAffinity
695 // If a numa node has more than one processor group, we assume they are
696 // sized equal and we spread threads evenly across the groups.
697 USHORT elements, returnedElements;
698 elements = fun5(); // GetMaximumProcessorGroupCount
699 GROUP_AFFINITY *affinity = (GROUP_AFFINITY*)malloc(elements * sizeof(GROUP_AFFINITY));
700 if (fun4(node, affinity, elements, &returnedElements)) // GetNumaNodeProcessorMask2
701 fun3(GetCurrentThread(), &affinity[idx % returnedElements], nullptr); // SetThreadGroupAffinity
708 } // namespace WinProcGroup
712 #define GETCWD _getcwd
715 #define GETCWD getcwd
718 namespace CommandLine {
720 string argv0; // path+name of the executable binary, as given by argv[0]
721 string binaryDirectory; // path of the executable directory
722 string workingDirectory; // path of the working directory
724 void init([[maybe_unused]] int argc, char* argv[]) {
725 string pathSeparator;
727 // extract the path+name of the executable binary
731 pathSeparator = "\\";
733 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
734 // issues in some windows 10 versions, so check returned values carefully.
735 char* pgmptr = nullptr;
736 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
743 // extract the working directory
744 workingDirectory = "";
746 char* cwd = GETCWD(buff, 40000);
748 workingDirectory = cwd;
750 // extract the binary directory path from argv0
751 binaryDirectory = argv0;
752 size_t pos = binaryDirectory.find_last_of("\\/");
753 if (pos == std::string::npos)
754 binaryDirectory = "." + pathSeparator;
756 binaryDirectory.resize(pos + 1);
758 // pattern replacement: "./" at the start of path is replaced by the working directory
759 if (binaryDirectory.find("." + pathSeparator) == 0)
760 binaryDirectory.replace(0, 1, workingDirectory);
764 } // namespace CommandLine
766 } // namespace Stockfish