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/>.
22 #if _WIN32_WINNT < 0x0601
24 #define _WIN32_WINNT 0x0601 // Force to include needed API prototypes
32 // The needed Windows API for processor groups could be missed from old Windows
33 // versions, so instead of calling them directly (forcing the linker to resolve
34 // the calls at compile time), try to load them at runtime. To do this we need
35 // first to define the corresponding function pointers.
37 using fun1_t = bool(*)(LOGICAL_PROCESSOR_RELATIONSHIP,
38 PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);
39 using fun2_t = bool(*)(USHORT, PGROUP_AFFINITY);
40 using fun3_t = bool(*)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
41 using fun4_t = bool(*)(USHORT, PGROUP_AFFINITY, USHORT, PUSHORT);
42 using fun5_t = WORD(*)();
43 using fun6_t = bool(*)(HANDLE, DWORD, PHANDLE);
44 using fun7_t = bool(*)(LPCSTR, LPCSTR, PLUID);
45 using fun8_t = bool(*)(HANDLE, BOOL, PTOKEN_PRIVILEGES, DWORD, PTOKEN_PRIVILEGES, PDWORD);
57 #include <string_view>
61 #if defined(__linux__) && !defined(__ANDROID__)
65 #if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) || defined(__e2k__)
66 #define POSIXALIGNEDALLOC
74 /// Version number or dev.
75 constexpr std::string_view version = "dev";
77 /// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
78 /// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
79 /// can toggle the logging of std::cout and std:cin at runtime whilst preserving
80 /// usual I/O functionality, all without changing a single line of code!
81 /// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
83 struct Tie: public std::streambuf { // MSVC requires split streambuf for cin and cout
85 Tie(std::streambuf* b, std::streambuf* l) : buf(b), logBuf(l) {}
87 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
88 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
89 int underflow() override { return buf->sgetc(); }
90 int uflow() override { return log(buf->sbumpc(), ">> "); }
92 std::streambuf *buf, *logBuf;
94 int log(int c, const char* prefix) {
96 static int last = '\n'; // Single log file
99 logBuf->sputn(prefix, 3);
101 return last = logBuf->sputc((char)c);
107 Logger() : in(std::cin.rdbuf(), file.rdbuf()), out(std::cout.rdbuf(), file.rdbuf()) {}
108 ~Logger() { start(""); }
114 static void start(const std::string& fname) {
118 if (l.file.is_open())
120 std::cout.rdbuf(l.out.buf);
121 std::cin.rdbuf(l.in.buf);
127 l.file.open(fname, std::ifstream::out);
129 if (!l.file.is_open())
131 std::cerr << "Unable to open debug log file " << fname << std::endl;
135 std::cin.rdbuf(&l.in);
136 std::cout.rdbuf(&l.out);
144 /// engine_info() returns the full name of the current Stockfish version.
145 /// For local dev compiles we try to append the commit sha and commit date
146 /// from git if that fails only the local compilation date is set and "nogit" is specified:
147 /// Stockfish dev-YYYYMMDD-SHA
149 /// Stockfish dev-YYYYMMDD-nogit
151 /// For releases (non dev builds) we only include the version number:
152 /// Stockfish version
154 std::string engine_info(bool to_uci) {
155 std::stringstream ss;
156 ss << "Stockfish " << version << std::setfill('0');
158 if constexpr (version == "dev")
162 ss << stringify(GIT_DATE);
164 constexpr std::string_view months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
165 std::string month, day, year;
166 std::stringstream date(__DATE__); // From compiler, format is "Sep 21 2008"
168 date >> month >> day >> year;
169 ss << year << std::setw(2) << std::setfill('0') << (1 + months.find(month) / 4) << std::setw(2) << std::setfill('0') << day;
175 ss << stringify(GIT_SHA);
181 ss << (to_uci ? "\nid author ": " by ")
182 << "the Stockfish developers (see AUTHORS file)";
188 /// compiler_info() returns a string trying to describe the compiler we use
190 std::string compiler_info() {
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_DOTPROD)
283 compiler += " NEON_DOTPROD";
284 #elif defined(USE_NEON)
289 compiler += " DEBUG";
292 compiler += "\n__VERSION__ macro expands to: ";
294 compiler += __VERSION__;
296 compiler += "(undefined macro)";
304 /// Debug functions used mainly to collect run-time statistics
305 constexpr int MaxDebugSlots = 32;
311 std::atomic<int64_t> data[N] = { 0 };
313 constexpr inline std::atomic<int64_t>& operator[](int index) { return data[index]; }
316 DebugInfo<2> hit[MaxDebugSlots];
317 DebugInfo<2> mean[MaxDebugSlots];
318 DebugInfo<3> stdev[MaxDebugSlots];
319 DebugInfo<6> correl[MaxDebugSlots];
323 void dbg_hit_on(bool cond, int slot) {
330 void dbg_mean_of(int64_t value, int slot) {
333 mean[slot][1] += value;
336 void dbg_stdev_of(int64_t value, int slot) {
339 stdev[slot][1] += value;
340 stdev[slot][2] += value * value;
343 void dbg_correl_of(int64_t value1, int64_t value2, int slot) {
346 correl[slot][1] += value1;
347 correl[slot][2] += value1 * value1;
348 correl[slot][3] += value2;
349 correl[slot][4] += value2 * value2;
350 correl[slot][5] += value1 * value2;
356 auto E = [&n](int64_t x) { return double(x) / n; };
357 auto sqr = [](double x) { return x * x; };
359 for (int i = 0; i < MaxDebugSlots; ++i)
361 std::cerr << "Hit #" << i
362 << ": Total " << n << " Hits " << hit[i][1]
363 << " Hit Rate (%) " << 100.0 * E(hit[i][1])
366 for (int i = 0; i < MaxDebugSlots; ++i)
367 if ((n = mean[i][0]))
369 std::cerr << "Mean #" << i
370 << ": Total " << n << " Mean " << E(mean[i][1])
374 for (int i = 0; i < MaxDebugSlots; ++i)
375 if ((n = stdev[i][0]))
377 double r = sqrt(E(stdev[i][2]) - sqr(E(stdev[i][1])));
378 std::cerr << "Stdev #" << i
379 << ": Total " << n << " Stdev " << r
383 for (int i = 0; i < MaxDebugSlots; ++i)
384 if ((n = correl[i][0]))
386 double r = (E(correl[i][5]) - E(correl[i][1]) * E(correl[i][3]))
387 / ( sqrt(E(correl[i][2]) - sqr(E(correl[i][1])))
388 * sqrt(E(correl[i][4]) - sqr(E(correl[i][3]))));
389 std::cerr << "Correl. #" << i
390 << ": Total " << n << " Coefficient " << r
396 /// Used to serialize access to std::cout to avoid multiple threads writing at
399 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
413 /// Trampoline helper to avoid moving Logger to misc.h
414 void start_logger(const std::string& fname) { Logger::start(fname); }
417 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
418 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
419 /// which can be quite slow.
422 void prefetch(void*) {}
426 void prefetch(void* addr) {
428 # if defined(__INTEL_COMPILER)
429 // This hack prevents prefetches from being optimized away by
430 // Intel compiler. Both MSVC and gcc seem not be affected by this.
434 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
435 _mm_prefetch((char*)addr, _MM_HINT_T0);
437 __builtin_prefetch(addr);
444 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
445 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
446 /// std_aligned_alloc() must be freed with std_aligned_free().
448 void* std_aligned_alloc(size_t alignment, size_t size) {
450 #if defined(POSIXALIGNEDALLOC)
452 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
453 #elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
454 return _mm_malloc(size, alignment);
455 #elif defined(_WIN32)
456 return _aligned_malloc(size, alignment);
458 return std::aligned_alloc(alignment, size);
462 void std_aligned_free(void* ptr) {
464 #if defined(POSIXALIGNEDALLOC)
466 #elif defined(_WIN32) && !defined(_M_ARM) && !defined(_M_ARM64)
468 #elif defined(_WIN32)
475 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
479 static void* aligned_large_pages_alloc_windows([[maybe_unused]] size_t allocSize) {
485 HANDLE hProcessToken { };
489 const size_t largePageSize = GetLargePageMinimum();
493 // Dynamically link OpenProcessToken, LookupPrivilegeValue and AdjustTokenPrivileges
495 HMODULE hAdvapi32 = GetModuleHandle(TEXT("advapi32.dll"));
498 hAdvapi32 = LoadLibrary(TEXT("advapi32.dll"));
500 auto fun6 = (fun6_t)(void(*)())GetProcAddress(hAdvapi32, "OpenProcessToken");
503 auto fun7 = (fun7_t)(void(*)())GetProcAddress(hAdvapi32, "LookupPrivilegeValueA");
506 auto fun8 = (fun8_t)(void(*)())GetProcAddress(hAdvapi32, "AdjustTokenPrivileges");
510 // We need SeLockMemoryPrivilege, so try to enable it for the process
511 if (!fun6( // OpenProcessToken()
512 GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
515 if (fun7( // LookupPrivilegeValue(nullptr, SE_LOCK_MEMORY_NAME, &luid)
516 nullptr, "SeLockMemoryPrivilege", &luid))
518 TOKEN_PRIVILEGES tp { };
519 TOKEN_PRIVILEGES prevTp { };
522 tp.PrivilegeCount = 1;
523 tp.Privileges[0].Luid = luid;
524 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
526 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
527 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
528 if (fun8( // AdjustTokenPrivileges()
529 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
530 GetLastError() == ERROR_SUCCESS)
532 // Round up size to full pages and allocate
533 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
535 nullptr, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
537 // Privilege no longer needed, restore previous state
538 fun8( // AdjustTokenPrivileges ()
539 hProcessToken, FALSE, &prevTp, 0, nullptr, nullptr);
543 CloseHandle(hProcessToken);
550 void* aligned_large_pages_alloc(size_t allocSize) {
552 // Try to allocate large pages
553 void* mem = aligned_large_pages_alloc_windows(allocSize);
555 // Fall back to regular, page aligned, allocation if necessary
557 mem = VirtualAlloc(nullptr, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
564 void* aligned_large_pages_alloc(size_t allocSize) {
566 #if defined(__linux__)
567 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
569 constexpr size_t alignment = 4096; // assumed small page size
572 // round up to multiples of alignment
573 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
574 void *mem = std_aligned_alloc(alignment, size);
575 #if defined(MADV_HUGEPAGE)
576 madvise(mem, size, MADV_HUGEPAGE);
584 /// aligned_large_pages_free() will free the previously allocated ttmem
588 void aligned_large_pages_free(void* mem) {
590 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
592 DWORD err = GetLastError();
593 std::cerr << "Failed to free large page memory. Error code: 0x"
595 << std::dec << std::endl;
602 void aligned_large_pages_free(void *mem) {
603 std_aligned_free(mem);
609 namespace WinProcGroup {
613 void bindThisThread(size_t) {}
617 /// best_node() retrieves logical processor information using Windows specific
618 /// API and returns the best node id for the thread with index idx. Original
619 /// code from Texel by Peter Ă–sterlund.
621 static int best_node(size_t idx) {
626 DWORD returnLength = 0;
627 DWORD byteOffset = 0;
629 // Early exit if the needed API is not available at runtime
630 HMODULE k32 = GetModuleHandle(TEXT("Kernel32.dll"));
631 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
635 // First call to GetLogicalProcessorInformationEx() to get returnLength.
636 // We expect the call to fail due to null buffer.
637 if (fun1(RelationAll, nullptr, &returnLength))
640 // Once we know returnLength, allocate the buffer
641 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
642 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
644 // Second call to GetLogicalProcessorInformationEx(), now we expect to succeed
645 if (!fun1(RelationAll, buffer, &returnLength))
651 while (byteOffset < returnLength)
653 if (ptr->Relationship == RelationNumaNode)
656 else if (ptr->Relationship == RelationProcessorCore)
659 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
663 byteOffset += ptr->Size;
664 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
669 std::vector<int> groups;
671 // Run as many threads as possible on the same node until core limit is
672 // reached, then move on filling the next node.
673 for (int n = 0; n < nodes; n++)
674 for (int i = 0; i < cores / nodes; i++)
677 // In case a core has more than one logical processor (we assume 2) and we
678 // have still threads to allocate, then spread them evenly across available
680 for (int t = 0; t < threads - cores; t++)
681 groups.push_back(t % nodes);
683 // If we still have more threads than the total number of logical processors
684 // then return -1 and let the OS to decide what to do.
685 return idx < groups.size() ? groups[idx] : -1;
689 /// bindThisThread() set the group affinity of the current thread
691 void bindThisThread(size_t idx) {
693 // Use only local variables to be thread-safe
694 int node = best_node(idx);
699 // Early exit if the needed API are not available at runtime
700 HMODULE k32 = GetModuleHandle(TEXT("Kernel32.dll"));
701 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
702 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
703 auto fun4 = (fun4_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMask2");
704 auto fun5 = (fun5_t)(void(*)())GetProcAddress(k32, "GetMaximumProcessorGroupCount");
711 GROUP_AFFINITY affinity;
712 if (fun2(node, &affinity)) // GetNumaNodeProcessorMaskEx
713 fun3(GetCurrentThread(), &affinity, nullptr); // SetThreadGroupAffinity
717 // If a numa node has more than one processor group, we assume they are
718 // sized equal and we spread threads evenly across the groups.
719 USHORT elements, returnedElements;
720 elements = fun5(); // GetMaximumProcessorGroupCount
721 GROUP_AFFINITY *affinity = (GROUP_AFFINITY*)malloc(elements * sizeof(GROUP_AFFINITY));
722 if (fun4(node, affinity, elements, &returnedElements)) // GetNumaNodeProcessorMask2
723 fun3(GetCurrentThread(), &affinity[idx % returnedElements], nullptr); // SetThreadGroupAffinity
730 } // namespace WinProcGroup
734 #define GETCWD _getcwd
737 #define GETCWD getcwd
740 namespace CommandLine {
742 std::string argv0; // path+name of the executable binary, as given by argv[0]
743 std::string binaryDirectory; // path of the executable directory
744 std::string workingDirectory; // path of the working directory
746 void init([[maybe_unused]] int argc, char* argv[]) {
747 std::string pathSeparator;
749 // extract the path+name of the executable binary
753 pathSeparator = "\\";
755 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
756 // issues in some windows 10 versions, so check returned values carefully.
757 char* pgmptr = nullptr;
758 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
765 // extract the working directory
766 workingDirectory = "";
768 char* cwd = GETCWD(buff, 40000);
770 workingDirectory = cwd;
772 // extract the binary directory path from argv0
773 binaryDirectory = argv0;
774 size_t pos = binaryDirectory.find_last_of("\\/");
775 if (pos == std::string::npos)
776 binaryDirectory = "." + pathSeparator;
778 binaryDirectory.resize(pos + 1);
780 // pattern replacement: "./" at the start of path is replaced by the working directory
781 if (binaryDirectory.find("." + pathSeparator) == 0)
782 binaryDirectory.replace(0, 1, workingDirectory);
786 } // namespace CommandLine
788 } // namespace Stockfish