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 (l.file.is_open())
115 cout.rdbuf(l.out.buf);
122 l.file.open(fname, ifstream::out);
124 if (!l.file.is_open())
126 cerr << "Unable to open debug log file " << fname << endl;
139 /// engine_info() returns the full name of the current Stockfish version. This
140 /// will be either "Stockfish <Tag> DD-MM-YY" (where DD-MM-YY is the date when
141 /// the program was compiled) or "Stockfish <Version>", depending on whether
142 /// Version is empty.
144 string engine_info(bool to_uci) {
146 const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
147 string month, day, year;
148 stringstream ss, date(__DATE__); // From compiler, format is "Sep 21 2008"
150 ss << "Stockfish " << Version << setfill('0');
154 date >> month >> day >> year;
155 ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2);
158 ss << (to_uci ? "\nid author ": " by ")
159 << "the Stockfish developers (see AUTHORS file)";
165 /// compiler_info() returns a string trying to describe the compiler we use
167 std::string compiler_info() {
169 #define stringify2(x) #x
170 #define stringify(x) stringify2(x)
171 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
173 /// Predefined macros hell:
175 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
176 /// __INTEL_COMPILER Compiler is Intel
177 /// _MSC_VER Compiler is MSVC or Intel on Windows
178 /// _WIN32 Building on Windows (any)
179 /// _WIN64 Building on Windows 64 bit
181 std::string compiler = "\nCompiled by ";
184 compiler += "clang++ ";
185 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
186 #elif __INTEL_COMPILER
187 compiler += "Intel compiler ";
188 compiler += "(version ";
189 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
193 compiler += "(version ";
194 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
196 #elif defined(__e2k__) && defined(__LCC__)
197 #define dot_ver2(n) \
198 compiler += (char)'.'; \
199 compiler += (char)('0' + (n) / 10); \
200 compiler += (char)('0' + (n) % 10);
202 compiler += "MCST LCC ";
203 compiler += "(version ";
204 compiler += std::to_string(__LCC__ / 100);
205 dot_ver2(__LCC__ % 100)
206 dot_ver2(__LCC_MINOR__)
209 compiler += "g++ (GNUC) ";
210 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
212 compiler += "Unknown compiler ";
213 compiler += "(unknown version)";
216 #if defined(__APPLE__)
217 compiler += " on Apple";
218 #elif defined(__CYGWIN__)
219 compiler += " on Cygwin";
220 #elif defined(__MINGW64__)
221 compiler += " on MinGW64";
222 #elif defined(__MINGW32__)
223 compiler += " on MinGW32";
224 #elif defined(__ANDROID__)
225 compiler += " on Android";
226 #elif defined(__linux__)
227 compiler += " on Linux";
228 #elif defined(_WIN64)
229 compiler += " on Microsoft Windows 64-bit";
230 #elif defined(_WIN32)
231 compiler += " on Microsoft Windows 32-bit";
233 compiler += " on unknown system";
236 compiler += "\nCompilation settings include: ";
237 compiler += (Is64Bit ? " 64bit" : " 32bit");
238 #if defined(USE_VNNI)
241 #if defined(USE_AVX512)
242 compiler += " AVX512";
244 compiler += (HasPext ? " BMI2" : "");
245 #if defined(USE_AVX2)
248 #if defined(USE_SSE41)
249 compiler += " SSE41";
251 #if defined(USE_SSSE3)
252 compiler += " SSSE3";
254 #if defined(USE_SSE2)
257 compiler += (HasPopCnt ? " POPCNT" : "");
261 #if defined(USE_NEON)
266 compiler += " DEBUG";
269 compiler += "\n__VERSION__ macro expands to: ";
271 compiler += __VERSION__;
273 compiler += "(undefined macro)";
281 /// Debug functions used mainly to collect run-time statistics
282 static std::atomic<int64_t> hits[2], means[2];
284 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
285 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
286 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
291 cerr << "Total " << hits[0] << " Hits " << hits[1]
292 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
295 cerr << "Total " << means[0] << " Mean "
296 << (double)means[1] / means[0] << endl;
300 /// Used to serialize access to std::cout to avoid multiple threads writing at
303 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
317 /// Trampoline helper to avoid moving Logger to misc.h
318 void start_logger(const std::string& fname) { Logger::start(fname); }
321 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
322 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
323 /// which can be quite slow.
326 void prefetch(void*) {}
330 void prefetch(void* addr) {
332 # if defined(__INTEL_COMPILER)
333 // This hack prevents prefetches from being optimized away by
334 // Intel compiler. Both MSVC and gcc seem not be affected by this.
338 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
339 _mm_prefetch((char*)addr, _MM_HINT_T0);
341 __builtin_prefetch(addr);
348 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
349 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
350 /// std_aligned_alloc() must be freed with std_aligned_free().
352 void* std_aligned_alloc(size_t alignment, size_t size) {
354 #if defined(POSIXALIGNEDALLOC)
356 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
357 #elif defined(_WIN32)
358 return _mm_malloc(size, alignment);
360 return std::aligned_alloc(alignment, size);
364 void std_aligned_free(void* ptr) {
366 #if defined(POSIXALIGNEDALLOC)
368 #elif defined(_WIN32)
375 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
379 static void* aligned_large_pages_alloc_windows(size_t allocSize) {
382 (void)allocSize; // suppress unused-parameter compiler warning
386 HANDLE hProcessToken { };
390 const size_t largePageSize = GetLargePageMinimum();
394 // We need SeLockMemoryPrivilege, so try to enable it for the process
395 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
398 if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid))
400 TOKEN_PRIVILEGES tp { };
401 TOKEN_PRIVILEGES prevTp { };
404 tp.PrivilegeCount = 1;
405 tp.Privileges[0].Luid = luid;
406 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
408 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
409 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
410 if (AdjustTokenPrivileges(
411 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
412 GetLastError() == ERROR_SUCCESS)
414 // Round up size to full pages and allocate
415 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
417 NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
419 // Privilege no longer needed, restore previous state
420 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL);
424 CloseHandle(hProcessToken);
431 void* aligned_large_pages_alloc(size_t allocSize) {
433 // Try to allocate large pages
434 void* mem = aligned_large_pages_alloc_windows(allocSize);
436 // Fall back to regular, page aligned, allocation if necessary
438 mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
445 void* aligned_large_pages_alloc(size_t allocSize) {
447 #if defined(__linux__)
448 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
450 constexpr size_t alignment = 4096; // assumed small page size
453 // round up to multiples of alignment
454 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
455 void *mem = std_aligned_alloc(alignment, size);
456 #if defined(MADV_HUGEPAGE)
457 madvise(mem, size, MADV_HUGEPAGE);
465 /// aligned_large_pages_free() will free the previously allocated ttmem
469 void aligned_large_pages_free(void* mem) {
471 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
473 DWORD err = GetLastError();
474 std::cerr << "Failed to free large page memory. Error code: 0x"
476 << std::dec << std::endl;
483 void aligned_large_pages_free(void *mem) {
484 std_aligned_free(mem);
490 namespace WinProcGroup {
494 void bindThisThread(size_t) {}
498 /// best_group() retrieves logical processor information using Windows specific
499 /// API and returns the best group id for the thread with index idx. Original
500 /// code from Texel by Peter Ă–sterlund.
502 int best_group(size_t idx) {
507 DWORD returnLength = 0;
508 DWORD byteOffset = 0;
510 // Early exit if the needed API is not available at runtime
511 HMODULE k32 = GetModuleHandle("Kernel32.dll");
512 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
516 // First call to get returnLength. We expect it to fail due to null buffer
517 if (fun1(RelationAll, nullptr, &returnLength))
520 // Once we know returnLength, allocate the buffer
521 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
522 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
524 // Second call, now we expect to succeed
525 if (!fun1(RelationAll, buffer, &returnLength))
531 while (byteOffset < returnLength)
533 if (ptr->Relationship == RelationNumaNode)
536 else if (ptr->Relationship == RelationProcessorCore)
539 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
543 byteOffset += ptr->Size;
544 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
549 std::vector<int> groups;
551 // Run as many threads as possible on the same node until core limit is
552 // reached, then move on filling the next node.
553 for (int n = 0; n < nodes; n++)
554 for (int i = 0; i < cores / nodes; i++)
557 // In case a core has more than one logical processor (we assume 2) and we
558 // have still threads to allocate, then spread them evenly across available
560 for (int t = 0; t < threads - cores; t++)
561 groups.push_back(t % nodes);
563 // If we still have more threads than the total number of logical processors
564 // then return -1 and let the OS to decide what to do.
565 return idx < groups.size() ? groups[idx] : -1;
569 /// bindThisThread() set the group affinity of the current thread
571 void bindThisThread(size_t idx) {
573 // Use only local variables to be thread-safe
574 int group = best_group(idx);
579 // Early exit if the needed API are not available at runtime
580 HMODULE k32 = GetModuleHandle("Kernel32.dll");
581 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
582 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
587 GROUP_AFFINITY affinity;
588 if (fun2(group, &affinity))
589 fun3(GetCurrentThread(), &affinity, nullptr);
594 } // namespace WinProcGroup
598 #define GETCWD _getcwd
601 #define GETCWD getcwd
604 namespace CommandLine {
606 string argv0; // path+name of the executable binary, as given by argv[0]
607 string binaryDirectory; // path of the executable directory
608 string workingDirectory; // path of the working directory
610 void init(int argc, char* argv[]) {
612 string pathSeparator;
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
653 } // namespace Stockfish