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 typedef bool(*fun1_t)(LOGICAL_PROCESSOR_RELATIONSHIP,
36 PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);
37 typedef bool(*fun2_t)(USHORT, PGROUP_AFFINITY);
38 typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
39 typedef bool(*fun4_t)(USHORT, PGROUP_AFFINITY, USHORT, PUSHORT);
40 typedef WORD(*fun5_t)();
49 #include <string_view>
52 #if defined(__linux__) && !defined(__ANDROID__)
57 #if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) || defined(__e2k__)
58 #define POSIXALIGNEDALLOC
71 /// Version number or dev.
72 constexpr string_view version = "dev";
74 /// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
75 /// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
76 /// can toggle the logging of std::cout and std:cin at runtime whilst preserving
77 /// usual I/O functionality, all without changing a single line of code!
78 /// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
80 struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout
82 Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {}
84 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
85 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
86 int underflow() override { return buf->sgetc(); }
87 int uflow() override { return log(buf->sbumpc(), ">> "); }
89 streambuf *buf, *logBuf;
91 int log(int c, const char* prefix) {
93 static int last = '\n'; // Single log file
96 logBuf->sputn(prefix, 3);
98 return last = logBuf->sputc((char)c);
104 Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {}
105 ~Logger() { start(""); }
111 static void start(const std::string& fname) {
115 if (l.file.is_open())
117 cout.rdbuf(l.out.buf);
124 l.file.open(fname, ifstream::out);
126 if (!l.file.is_open())
128 cerr << "Unable to open debug log file " << fname << endl;
141 /// engine_info() returns the full name of the current Stockfish version.
142 /// For local dev compiles we try to append the commit sha and commit date
143 /// from git if that fails only the local compilation date is set and "nogit" is specified:
144 /// Stockfish dev-YYYYMMDD-SHA
146 /// Stockfish dev-YYYYMMDD-nogit
148 /// For releases (non dev builds) we only include the version number:
149 /// Stockfish version
151 string engine_info(bool to_uci) {
153 ss << "Stockfish " << version << setfill('0');
155 if constexpr (version == "dev")
161 constexpr string_view months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
162 string month, day, year;
163 stringstream date(__DATE__); // From compiler, format is "Sep 21 2008"
165 date >> month >> day >> year;
166 ss << year << setw(2) << setfill('0') << (1 + months.find(month) / 4) << setw(2) << setfill('0') << day;
178 ss << (to_uci ? "\nid author ": " by ")
179 << "the Stockfish developers (see AUTHORS file)";
185 /// compiler_info() returns a string trying to describe the compiler we use
187 std::string compiler_info() {
189 #define stringify2(x) #x
190 #define stringify(x) stringify2(x)
191 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
193 /// Predefined macros hell:
195 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
196 /// __INTEL_COMPILER Compiler is Intel
197 /// _MSC_VER Compiler is MSVC or Intel on Windows
198 /// _WIN32 Building on Windows (any)
199 /// _WIN64 Building on Windows 64 bit
201 std::string compiler = "\nCompiled by ";
204 compiler += "clang++ ";
205 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
206 #elif __INTEL_COMPILER
207 compiler += "Intel compiler ";
208 compiler += "(version ";
209 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
213 compiler += "(version ";
214 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
216 #elif defined(__e2k__) && defined(__LCC__)
217 #define dot_ver2(n) \
218 compiler += (char)'.'; \
219 compiler += (char)('0' + (n) / 10); \
220 compiler += (char)('0' + (n) % 10);
222 compiler += "MCST LCC ";
223 compiler += "(version ";
224 compiler += std::to_string(__LCC__ / 100);
225 dot_ver2(__LCC__ % 100)
226 dot_ver2(__LCC_MINOR__)
229 compiler += "g++ (GNUC) ";
230 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
232 compiler += "Unknown compiler ";
233 compiler += "(unknown version)";
236 #if defined(__APPLE__)
237 compiler += " on Apple";
238 #elif defined(__CYGWIN__)
239 compiler += " on Cygwin";
240 #elif defined(__MINGW64__)
241 compiler += " on MinGW64";
242 #elif defined(__MINGW32__)
243 compiler += " on MinGW32";
244 #elif defined(__ANDROID__)
245 compiler += " on Android";
246 #elif defined(__linux__)
247 compiler += " on Linux";
248 #elif defined(_WIN64)
249 compiler += " on Microsoft Windows 64-bit";
250 #elif defined(_WIN32)
251 compiler += " on Microsoft Windows 32-bit";
253 compiler += " on unknown system";
256 compiler += "\nCompilation settings include: ";
257 compiler += (Is64Bit ? " 64bit" : " 32bit");
258 #if defined(USE_VNNI)
261 #if defined(USE_AVX512)
262 compiler += " AVX512";
264 compiler += (HasPext ? " BMI2" : "");
265 #if defined(USE_AVX2)
268 #if defined(USE_SSE41)
269 compiler += " SSE41";
271 #if defined(USE_SSSE3)
272 compiler += " SSSE3";
274 #if defined(USE_SSE2)
277 compiler += (HasPopCnt ? " POPCNT" : "");
281 #if defined(USE_NEON)
286 compiler += " DEBUG";
289 compiler += "\n__VERSION__ macro expands to: ";
291 compiler += __VERSION__;
293 compiler += "(undefined macro)";
301 /// Debug functions used mainly to collect run-time statistics
302 static std::atomic<int64_t> hits[2], means[2];
304 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
305 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
306 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
311 cerr << "Total " << hits[0] << " Hits " << hits[1]
312 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
315 cerr << "Total " << means[0] << " Mean "
316 << (double)means[1] / means[0] << endl;
320 /// Used to serialize access to std::cout to avoid multiple threads writing at
323 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
337 /// Trampoline helper to avoid moving Logger to misc.h
338 void start_logger(const std::string& fname) { Logger::start(fname); }
341 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
342 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
343 /// which can be quite slow.
346 void prefetch(void*) {}
350 void prefetch(void* addr) {
352 # if defined(__INTEL_COMPILER)
353 // This hack prevents prefetches from being optimized away by
354 // Intel compiler. Both MSVC and gcc seem not be affected by this.
358 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
359 _mm_prefetch((char*)addr, _MM_HINT_T0);
361 __builtin_prefetch(addr);
368 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
369 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
370 /// std_aligned_alloc() must be freed with std_aligned_free().
372 void* std_aligned_alloc(size_t alignment, size_t size) {
374 #if defined(POSIXALIGNEDALLOC)
376 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
377 #elif defined(_WIN32)
378 return _mm_malloc(size, alignment);
380 return std::aligned_alloc(alignment, size);
384 void std_aligned_free(void* ptr) {
386 #if defined(POSIXALIGNEDALLOC)
388 #elif defined(_WIN32)
395 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
399 static void* aligned_large_pages_alloc_windows([[maybe_unused]] size_t allocSize) {
405 HANDLE hProcessToken { };
409 const size_t largePageSize = GetLargePageMinimum();
413 // We need SeLockMemoryPrivilege, so try to enable it for the process
414 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
417 if (LookupPrivilegeValue(nullptr, SE_LOCK_MEMORY_NAME, &luid))
419 TOKEN_PRIVILEGES tp { };
420 TOKEN_PRIVILEGES prevTp { };
423 tp.PrivilegeCount = 1;
424 tp.Privileges[0].Luid = luid;
425 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
427 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
428 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
429 if (AdjustTokenPrivileges(
430 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
431 GetLastError() == ERROR_SUCCESS)
433 // Round up size to full pages and allocate
434 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
436 nullptr, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
438 // Privilege no longer needed, restore previous state
439 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, nullptr, nullptr);
443 CloseHandle(hProcessToken);
450 void* aligned_large_pages_alloc(size_t allocSize) {
452 // Try to allocate large pages
453 void* mem = aligned_large_pages_alloc_windows(allocSize);
455 // Fall back to regular, page aligned, allocation if necessary
457 mem = VirtualAlloc(nullptr, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
464 void* aligned_large_pages_alloc(size_t allocSize) {
466 #if defined(__linux__)
467 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
469 constexpr size_t alignment = 4096; // assumed small page size
472 // round up to multiples of alignment
473 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
474 void *mem = std_aligned_alloc(alignment, size);
475 #if defined(MADV_HUGEPAGE)
476 madvise(mem, size, MADV_HUGEPAGE);
484 /// aligned_large_pages_free() will free the previously allocated ttmem
488 void aligned_large_pages_free(void* mem) {
490 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
492 DWORD err = GetLastError();
493 std::cerr << "Failed to free large page memory. Error code: 0x"
495 << std::dec << std::endl;
502 void aligned_large_pages_free(void *mem) {
503 std_aligned_free(mem);
509 namespace WinProcGroup {
513 void bindThisThread(size_t) {}
517 /// best_node() retrieves logical processor information using Windows specific
518 /// API and returns the best node id for the thread with index idx. Original
519 /// code from Texel by Peter Ă–sterlund.
521 static int best_node(size_t idx) {
526 DWORD returnLength = 0;
527 DWORD byteOffset = 0;
529 // Early exit if the needed API is not available at runtime
530 HMODULE k32 = GetModuleHandle("Kernel32.dll");
531 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
535 // First call to GetLogicalProcessorInformationEx() to get returnLength.
536 // We expect the call to fail due to null buffer.
537 if (fun1(RelationAll, nullptr, &returnLength))
540 // Once we know returnLength, allocate the buffer
541 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
542 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
544 // Second call to GetLogicalProcessorInformationEx(), now we expect to succeed
545 if (!fun1(RelationAll, buffer, &returnLength))
551 while (byteOffset < returnLength)
553 if (ptr->Relationship == RelationNumaNode)
556 else if (ptr->Relationship == RelationProcessorCore)
559 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
563 byteOffset += ptr->Size;
564 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
569 std::vector<int> groups;
571 // Run as many threads as possible on the same node until core limit is
572 // reached, then move on filling the next node.
573 for (int n = 0; n < nodes; n++)
574 for (int i = 0; i < cores / nodes; i++)
577 // In case a core has more than one logical processor (we assume 2) and we
578 // have still threads to allocate, then spread them evenly across available
580 for (int t = 0; t < threads - cores; t++)
581 groups.push_back(t % nodes);
583 // If we still have more threads than the total number of logical processors
584 // then return -1 and let the OS to decide what to do.
585 return idx < groups.size() ? groups[idx] : -1;
589 /// bindThisThread() set the group affinity of the current thread
591 void bindThisThread(size_t idx) {
593 // Use only local variables to be thread-safe
594 int node = best_node(idx);
599 // Early exit if the needed API are not available at runtime
600 HMODULE k32 = GetModuleHandle("Kernel32.dll");
601 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
602 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
603 auto fun4 = (fun4_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMask2");
604 auto fun5 = (fun5_t)(void(*)())GetProcAddress(k32, "GetMaximumProcessorGroupCount");
611 GROUP_AFFINITY affinity;
612 if (fun2(node, &affinity)) // GetNumaNodeProcessorMaskEx
613 fun3(GetCurrentThread(), &affinity, nullptr); // SetThreadGroupAffinity
617 // If a numa node has more than one processor group, we assume they are
618 // sized equal and we spread threads evenly across the groups.
619 USHORT elements, returnedElements;
620 elements = fun5(); // GetMaximumProcessorGroupCount
621 GROUP_AFFINITY *affinity = (GROUP_AFFINITY*)malloc(elements * sizeof(GROUP_AFFINITY));
622 if (fun4(node, affinity, elements, &returnedElements)) // GetNumaNodeProcessorMask2
623 fun3(GetCurrentThread(), &affinity[idx % returnedElements], nullptr); // SetThreadGroupAffinity
630 } // namespace WinProcGroup
634 #define GETCWD _getcwd
637 #define GETCWD getcwd
640 namespace CommandLine {
642 string argv0; // path+name of the executable binary, as given by argv[0]
643 string binaryDirectory; // path of the executable directory
644 string workingDirectory; // path of the working directory
646 void init([[maybe_unused]] int argc, char* argv[]) {
647 string pathSeparator;
649 // extract the path+name of the executable binary
653 pathSeparator = "\\";
655 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
656 // issues in some windows 10 versions, so check returned values carefully.
657 char* pgmptr = nullptr;
658 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
665 // extract the working directory
666 workingDirectory = "";
668 char* cwd = GETCWD(buff, 40000);
670 workingDirectory = cwd;
672 // extract the binary directory path from argv0
673 binaryDirectory = argv0;
674 size_t pos = binaryDirectory.find_last_of("\\/");
675 if (pos == std::string::npos)
676 binaryDirectory = "." + pathSeparator;
678 binaryDirectory.resize(pos + 1);
680 // pattern replacement: "./" at the start of path is replaced by the working directory
681 if (binaryDirectory.find("." + pathSeparator) == 0)
682 binaryDirectory.replace(0, 1, workingDirectory);
686 } // namespace CommandLine
688 } // namespace Stockfish