2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2020 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);
49 #if defined(__linux__) && !defined(__ANDROID__)
54 #if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32))
55 #define POSIXALIGNEDALLOC
66 /// Version number. If Version is left empty, then compile date in the format
67 /// DD-MM-YY and show in engine_info.
68 const string Version = "";
70 /// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
71 /// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
72 /// can toggle the logging of std::cout and std:cin at runtime whilst preserving
73 /// usual I/O functionality, all without changing a single line of code!
74 /// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
76 struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout
78 Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {}
80 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
81 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
82 int underflow() override { return buf->sgetc(); }
83 int uflow() override { return log(buf->sbumpc(), ">> "); }
85 streambuf *buf, *logBuf;
87 int log(int c, const char* prefix) {
89 static int last = '\n'; // Single log file
92 logBuf->sputn(prefix, 3);
94 return last = logBuf->sputc((char)c);
100 Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {}
101 ~Logger() { start(""); }
107 static void start(const std::string& fname) {
111 if (!fname.empty() && !l.file.is_open())
113 l.file.open(fname, ifstream::out);
115 if (!l.file.is_open())
117 cerr << "Unable to open debug log file " << fname << endl;
124 else if (fname.empty() && l.file.is_open())
126 cout.rdbuf(l.out.buf);
136 /// engine_info() returns the full name of the current Stockfish version. This
137 /// will be either "Stockfish <Tag> DD-MM-YY" (where DD-MM-YY is the date when
138 /// the program was compiled) or "Stockfish <Version>", depending on whether
139 /// Version is empty.
141 const string engine_info(bool to_uci) {
143 const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
144 string month, day, year;
145 stringstream ss, date(__DATE__); // From compiler, format is "Sep 21 2008"
147 ss << "Stockfish " << Version << setfill('0');
151 date >> month >> day >> year;
152 ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2);
156 ss << (to_uci ? "\nid author ": " by ")
157 << "the Stockfish developers (see AUTHORS file)";
163 /// compiler_info() returns a string trying to describe the compiler we use
165 const std::string compiler_info() {
167 #define stringify2(x) #x
168 #define stringify(x) stringify2(x)
169 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
171 /// Predefined macros hell:
173 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
174 /// __INTEL_COMPILER Compiler is Intel
175 /// _MSC_VER Compiler is MSVC or Intel on Windows
176 /// _WIN32 Building on Windows (any)
177 /// _WIN64 Building on Windows 64 bit
179 std::string compiler = "\nCompiled by ";
182 compiler += "clang++ ";
183 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
184 #elif __INTEL_COMPILER
185 compiler += "Intel compiler ";
186 compiler += "(version ";
187 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
191 compiler += "(version ";
192 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
195 compiler += "g++ (GNUC) ";
196 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
198 compiler += "Unknown compiler ";
199 compiler += "(unknown version)";
202 #if defined(__APPLE__)
203 compiler += " on Apple";
204 #elif defined(__CYGWIN__)
205 compiler += " on Cygwin";
206 #elif defined(__MINGW64__)
207 compiler += " on MinGW64";
208 #elif defined(__MINGW32__)
209 compiler += " on MinGW32";
210 #elif defined(__ANDROID__)
211 compiler += " on Android";
212 #elif defined(__linux__)
213 compiler += " on Linux";
214 #elif defined(_WIN64)
215 compiler += " on Microsoft Windows 64-bit";
216 #elif defined(_WIN32)
217 compiler += " on Microsoft Windows 32-bit";
219 compiler += " on unknown system";
222 compiler += "\nCompilation settings include: ";
223 compiler += (Is64Bit ? " 64bit" : " 32bit");
224 #if defined(USE_VNNI)
227 #if defined(USE_AVX512)
228 compiler += " AVX512";
230 compiler += (HasPext ? " BMI2" : "");
231 #if defined(USE_AVX2)
234 #if defined(USE_SSE41)
235 compiler += " SSE41";
237 #if defined(USE_SSSE3)
238 compiler += " SSSE3";
240 #if defined(USE_SSE2)
243 compiler += (HasPopCnt ? " POPCNT" : "");
247 #if defined(USE_NEON)
252 compiler += " DEBUG";
255 compiler += "\n__VERSION__ macro expands to: ";
257 compiler += __VERSION__;
259 compiler += "(undefined macro)";
267 /// Debug functions used mainly to collect run-time statistics
268 static std::atomic<int64_t> hits[2], means[2];
270 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
271 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
272 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
277 cerr << "Total " << hits[0] << " Hits " << hits[1]
278 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
281 cerr << "Total " << means[0] << " Mean "
282 << (double)means[1] / means[0] << endl;
286 /// Used to serialize access to std::cout to avoid multiple threads writing at
289 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
303 /// Trampoline helper to avoid moving Logger to misc.h
304 void start_logger(const std::string& fname) { Logger::start(fname); }
307 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
308 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
309 /// which can be quite slow.
312 void prefetch(void*) {}
316 void prefetch(void* addr) {
318 # if defined(__INTEL_COMPILER)
319 // This hack prevents prefetches from being optimized away by
320 // Intel compiler. Both MSVC and gcc seem not be affected by this.
324 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
325 _mm_prefetch((char*)addr, _MM_HINT_T0);
327 __builtin_prefetch(addr);
334 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
335 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
336 /// std_aligned_alloc() must be freed with std_aligned_free().
338 void* std_aligned_alloc(size_t alignment, size_t size) {
340 #if defined(POSIXALIGNEDALLOC)
342 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
343 #elif defined(_WIN32)
344 return _mm_malloc(size, alignment);
346 return std::aligned_alloc(alignment, size);
350 void std_aligned_free(void* ptr) {
352 #if defined(POSIXALIGNEDALLOC)
354 #elif defined(_WIN32)
361 /// aligned_ttmem_alloc() will return suitably aligned memory, if possible using large pages.
362 /// The returned pointer is the aligned one, while the mem argument is the one that needs
363 /// to be passed to free. With c++17 some of this functionality could be simplified.
365 #if defined(__linux__) && !defined(__ANDROID__)
367 void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
369 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page sizes
370 size_t size = ((allocSize + alignment - 1) / alignment) * alignment; // multiple of alignment
371 if (posix_memalign(&mem, alignment, size))
373 #if defined(MADV_HUGEPAGE)
374 madvise(mem, allocSize, MADV_HUGEPAGE);
379 #elif defined(_WIN64)
381 static void* aligned_ttmem_alloc_large_pages(size_t allocSize) {
383 HANDLE hProcessToken { };
387 const size_t largePageSize = GetLargePageMinimum();
391 // We need SeLockMemoryPrivilege, so try to enable it for the process
392 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
395 if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid))
397 TOKEN_PRIVILEGES tp { };
398 TOKEN_PRIVILEGES prevTp { };
401 tp.PrivilegeCount = 1;
402 tp.Privileges[0].Luid = luid;
403 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
405 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
406 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
407 if (AdjustTokenPrivileges(
408 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
409 GetLastError() == ERROR_SUCCESS)
411 // Round up size to full pages and allocate
412 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
414 NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
416 // Privilege no longer needed, restore previous state
417 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL);
421 CloseHandle(hProcessToken);
426 void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
428 static bool firstCall = true;
430 // Try to allocate large pages
431 mem = aligned_ttmem_alloc_large_pages(allocSize);
433 // Suppress info strings on the first call. The first call occurs before 'uci'
434 // is received and in that case this output confuses some GUIs.
438 sync_cout << "info string Hash table allocation: Windows large pages used." << sync_endl;
440 sync_cout << "info string Hash table allocation: Windows large pages not used." << sync_endl;
444 // Fall back to regular, page aligned, allocation if necessary
446 mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
453 void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
455 constexpr size_t alignment = 64; // assumed cache line size
456 size_t size = allocSize + alignment - 1; // allocate some extra space
458 void* ret = reinterpret_cast<void*>((uintptr_t(mem) + alignment - 1) & ~uintptr_t(alignment - 1));
465 /// aligned_ttmem_free() will free the previously allocated ttmem
469 void aligned_ttmem_free(void* mem) {
471 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
473 DWORD err = GetLastError();
474 std::cerr << "Failed to free transposition table. Error code: 0x" <<
475 std::hex << err << std::dec << std::endl;
482 void aligned_ttmem_free(void *mem) {
489 namespace WinProcGroup {
493 void bindThisThread(size_t) {}
497 /// best_group() retrieves logical processor information using Windows specific
498 /// API and returns the best group id for the thread with index idx. Original
499 /// code from Texel by Peter Ă–sterlund.
501 int best_group(size_t idx) {
506 DWORD returnLength = 0;
507 DWORD byteOffset = 0;
509 // Early exit if the needed API is not available at runtime
510 HMODULE k32 = GetModuleHandle("Kernel32.dll");
511 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
515 // First call to get returnLength. We expect it to fail due to null buffer
516 if (fun1(RelationAll, nullptr, &returnLength))
519 // Once we know returnLength, allocate the buffer
520 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
521 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
523 // Second call, now we expect to succeed
524 if (!fun1(RelationAll, buffer, &returnLength))
530 while (byteOffset < returnLength)
532 if (ptr->Relationship == RelationNumaNode)
535 else if (ptr->Relationship == RelationProcessorCore)
538 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
542 byteOffset += ptr->Size;
543 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
548 std::vector<int> groups;
550 // Run as many threads as possible on the same node until core limit is
551 // reached, then move on filling the next node.
552 for (int n = 0; n < nodes; n++)
553 for (int i = 0; i < cores / nodes; i++)
556 // In case a core has more than one logical processor (we assume 2) and we
557 // have still threads to allocate, then spread them evenly across available
559 for (int t = 0; t < threads - cores; t++)
560 groups.push_back(t % nodes);
562 // If we still have more threads than the total number of logical processors
563 // then return -1 and let the OS to decide what to do.
564 return idx < groups.size() ? groups[idx] : -1;
568 /// bindThisThread() set the group affinity of the current thread
570 void bindThisThread(size_t idx) {
572 // Use only local variables to be thread-safe
573 int group = best_group(idx);
578 // Early exit if the needed API are not available at runtime
579 HMODULE k32 = GetModuleHandle("Kernel32.dll");
580 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
581 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
586 GROUP_AFFINITY affinity;
587 if (fun2(group, &affinity))
588 fun3(GetCurrentThread(), &affinity, nullptr);
593 } // namespace WinProcGroup
597 #define GETCWD _getcwd
600 #define GETCWD getcwd
603 namespace CommandLine {
605 string argv0; // path+name of the executable binary, as given by argv[0]
606 string binaryDirectory; // path of the executable directory
607 string workingDirectory; // path of the working directory
608 string pathSeparator; // Separator for our current OS
610 void init(int argc, char* argv[]) {
614 // extract the path+name of the executable binary
618 pathSeparator = "\\";
620 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
621 // issues in some windows 10 versions, so check returned values carefully.
622 char* pgmptr = nullptr;
623 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
630 // extract the working directory
631 workingDirectory = "";
633 char* cwd = GETCWD(buff, 40000);
635 workingDirectory = cwd;
637 // extract the binary directory path from argv0
638 binaryDirectory = argv0;
639 size_t pos = binaryDirectory.find_last_of("\\/");
640 if (pos == std::string::npos)
641 binaryDirectory = "." + pathSeparator;
643 binaryDirectory.resize(pos + 1);
645 // pattern replacement: "./" at the start of path is replaced by the working directory
646 if (binaryDirectory.find("." + pathSeparator) == 0)
647 binaryDirectory.replace(0, 1, workingDirectory);
651 } // namespace CommandLine