2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2021 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)) || defined(__e2k__)
55 #define POSIXALIGNEDALLOC
68 /// Version number. If Version is left empty, then compile date in the format
69 /// DD-MM-YY and show in engine_info.
70 const string Version = "";
72 /// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
73 /// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
74 /// can toggle the logging of std::cout and std:cin at runtime whilst preserving
75 /// usual I/O functionality, all without changing a single line of code!
76 /// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
78 struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout
80 Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {}
82 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
83 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
84 int underflow() override { return buf->sgetc(); }
85 int uflow() override { return log(buf->sbumpc(), ">> "); }
87 streambuf *buf, *logBuf;
89 int log(int c, const char* prefix) {
91 static int last = '\n'; // Single log file
94 logBuf->sputn(prefix, 3);
96 return last = logBuf->sputc((char)c);
102 Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {}
103 ~Logger() { start(""); }
109 static void start(const std::string& fname) {
113 if (!fname.empty() && !l.file.is_open())
115 l.file.open(fname, ifstream::out);
117 if (!l.file.is_open())
119 cerr << "Unable to open debug log file " << fname << endl;
126 else if (fname.empty() && l.file.is_open())
128 cout.rdbuf(l.out.buf);
138 /// engine_info() returns the full name of the current Stockfish version. This
139 /// will be either "Stockfish <Tag> DD-MM-YY" (where DD-MM-YY is the date when
140 /// the program was compiled) or "Stockfish <Version>", depending on whether
141 /// Version is empty.
143 string engine_info(bool to_uci) {
145 const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
146 string month, day, year;
147 stringstream ss, date(__DATE__); // From compiler, format is "Sep 21 2008"
149 ss << "Stockfish " << Version << setfill('0');
153 date >> month >> day >> year;
154 ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2);
157 ss << (to_uci ? "\nid author ": " by ")
158 << "the Stockfish developers (see AUTHORS file)";
164 /// compiler_info() returns a string trying to describe the compiler we use
166 std::string compiler_info() {
168 #define stringify2(x) #x
169 #define stringify(x) stringify2(x)
170 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
172 /// Predefined macros hell:
174 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
175 /// __INTEL_COMPILER Compiler is Intel
176 /// _MSC_VER Compiler is MSVC or Intel on Windows
177 /// _WIN32 Building on Windows (any)
178 /// _WIN64 Building on Windows 64 bit
180 std::string compiler = "\nCompiled by ";
183 compiler += "clang++ ";
184 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
185 #elif __INTEL_COMPILER
186 compiler += "Intel compiler ";
187 compiler += "(version ";
188 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
192 compiler += "(version ";
193 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
195 #elif defined(__e2k__) && defined(__LCC__)
196 #define dot_ver2(n) \
197 compiler += (char)'.'; \
198 compiler += (char)('0' + (n) / 10); \
199 compiler += (char)('0' + (n) % 10);
201 compiler += "MCST LCC ";
202 compiler += "(version ";
203 compiler += std::to_string(__LCC__ / 100);
204 dot_ver2(__LCC__ % 100)
205 dot_ver2(__LCC_MINOR__)
208 compiler += "g++ (GNUC) ";
209 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
211 compiler += "Unknown compiler ";
212 compiler += "(unknown version)";
215 #if defined(__APPLE__)
216 compiler += " on Apple";
217 #elif defined(__CYGWIN__)
218 compiler += " on Cygwin";
219 #elif defined(__MINGW64__)
220 compiler += " on MinGW64";
221 #elif defined(__MINGW32__)
222 compiler += " on MinGW32";
223 #elif defined(__ANDROID__)
224 compiler += " on Android";
225 #elif defined(__linux__)
226 compiler += " on Linux";
227 #elif defined(_WIN64)
228 compiler += " on Microsoft Windows 64-bit";
229 #elif defined(_WIN32)
230 compiler += " on Microsoft Windows 32-bit";
232 compiler += " on unknown system";
235 compiler += "\nCompilation settings include: ";
236 compiler += (Is64Bit ? " 64bit" : " 32bit");
237 #if defined(USE_VNNI)
240 #if defined(USE_AVX512)
241 compiler += " AVX512";
243 compiler += (HasPext ? " BMI2" : "");
244 #if defined(USE_AVX2)
247 #if defined(USE_SSE41)
248 compiler += " SSE41";
250 #if defined(USE_SSSE3)
251 compiler += " SSSE3";
253 #if defined(USE_SSE2)
256 compiler += (HasPopCnt ? " POPCNT" : "");
260 #if defined(USE_NEON)
265 compiler += " DEBUG";
268 compiler += "\n__VERSION__ macro expands to: ";
270 compiler += __VERSION__;
272 compiler += "(undefined macro)";
280 /// Debug functions used mainly to collect run-time statistics
281 static std::atomic<int64_t> hits[2], means[2];
283 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
284 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
285 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
290 cerr << "Total " << hits[0] << " Hits " << hits[1]
291 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
294 cerr << "Total " << means[0] << " Mean "
295 << (double)means[1] / means[0] << endl;
299 /// Used to serialize access to std::cout to avoid multiple threads writing at
302 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
316 /// Trampoline helper to avoid moving Logger to misc.h
317 void start_logger(const std::string& fname) { Logger::start(fname); }
320 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
321 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
322 /// which can be quite slow.
325 void prefetch(void*) {}
329 void prefetch(void* addr) {
331 # if defined(__INTEL_COMPILER)
332 // This hack prevents prefetches from being optimized away by
333 // Intel compiler. Both MSVC and gcc seem not be affected by this.
337 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
338 _mm_prefetch((char*)addr, _MM_HINT_T0);
340 __builtin_prefetch(addr);
347 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
348 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
349 /// std_aligned_alloc() must be freed with std_aligned_free().
351 void* std_aligned_alloc(size_t alignment, size_t size) {
353 #if defined(POSIXALIGNEDALLOC)
355 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
356 #elif defined(_WIN32)
357 return _mm_malloc(size, alignment);
359 return std::aligned_alloc(alignment, size);
363 void std_aligned_free(void* ptr) {
365 #if defined(POSIXALIGNEDALLOC)
367 #elif defined(_WIN32)
374 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
378 static void* aligned_large_pages_alloc_windows(size_t allocSize) {
381 (void)allocSize; // suppress unused-parameter compiler warning
385 HANDLE hProcessToken { };
389 const size_t largePageSize = GetLargePageMinimum();
393 // We need SeLockMemoryPrivilege, so try to enable it for the process
394 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
397 if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid))
399 TOKEN_PRIVILEGES tp { };
400 TOKEN_PRIVILEGES prevTp { };
403 tp.PrivilegeCount = 1;
404 tp.Privileges[0].Luid = luid;
405 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
407 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
408 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
409 if (AdjustTokenPrivileges(
410 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
411 GetLastError() == ERROR_SUCCESS)
413 // Round up size to full pages and allocate
414 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
416 NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
418 // Privilege no longer needed, restore previous state
419 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL);
423 CloseHandle(hProcessToken);
430 void* aligned_large_pages_alloc(size_t allocSize) {
432 // Try to allocate large pages
433 void* mem = aligned_large_pages_alloc_windows(allocSize);
435 // Fall back to regular, page aligned, allocation if necessary
437 mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
444 void* aligned_large_pages_alloc(size_t allocSize) {
446 #if defined(__linux__)
447 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
449 constexpr size_t alignment = 4096; // assumed small page size
452 // round up to multiples of alignment
453 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
454 void *mem = std_aligned_alloc(alignment, size);
455 #if defined(MADV_HUGEPAGE)
456 madvise(mem, size, MADV_HUGEPAGE);
464 /// aligned_large_pages_free() will free the previously allocated ttmem
468 void aligned_large_pages_free(void* mem) {
470 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
472 DWORD err = GetLastError();
473 std::cerr << "Failed to free large page memory. Error code: 0x"
475 << std::dec << std::endl;
482 void aligned_large_pages_free(void *mem) {
483 std_aligned_free(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
609 void init(int argc, char* argv[]) {
611 string pathSeparator;
613 // extract the path+name of the executable binary
617 pathSeparator = "\\";
619 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
620 // issues in some windows 10 versions, so check returned values carefully.
621 char* pgmptr = nullptr;
622 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
629 // extract the working directory
630 workingDirectory = "";
632 char* cwd = GETCWD(buff, 40000);
634 workingDirectory = cwd;
636 // extract the binary directory path from argv0
637 binaryDirectory = argv0;
638 size_t pos = binaryDirectory.find_last_of("\\/");
639 if (pos == std::string::npos)
640 binaryDirectory = "." + pathSeparator;
642 binaryDirectory.resize(pos + 1);
644 // pattern replacement: "./" at the start of path is replaced by the working directory
645 if (binaryDirectory.find("." + pathSeparator) == 0)
646 binaryDirectory.replace(0, 1, workingDirectory);
650 } // namespace CommandLine
652 } // namespace Stockfish