2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2022 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)();
51 #if defined(__linux__) && !defined(__ANDROID__)
56 #if defined(__APPLE__) || defined(__ANDROID__) || defined(__OpenBSD__) || (defined(__GLIBCXX__) && !defined(_GLIBCXX_HAVE_ALIGNED_ALLOC) && !defined(_WIN32)) || defined(__e2k__)
57 #define POSIXALIGNEDALLOC
70 /// Version number. If Version is left empty, then compile date in the format
71 /// DD-MM-YY and show in engine_info.
72 const string Version = "";
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. This
142 /// will be either "Stockfish <Tag> DD-MM-YY" (where DD-MM-YY is the date when
143 /// the program was compiled) or "Stockfish <Version>", depending on whether
144 /// Version is empty.
146 string engine_info(bool to_uci) {
148 const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
149 string month, day, year;
150 stringstream ss, date(__DATE__); // From compiler, format is "Sep 21 2008"
152 ss << "Stockfish " << Version << setfill('0');
156 date >> month >> day >> year;
157 ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2);
160 ss << (to_uci ? "\nid author ": " by ")
161 << "the Stockfish developers (see AUTHORS file)";
167 /// compiler_info() returns a string trying to describe the compiler we use
169 std::string compiler_info() {
171 #define stringify2(x) #x
172 #define stringify(x) stringify2(x)
173 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
175 /// Predefined macros hell:
177 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
178 /// __INTEL_COMPILER Compiler is Intel
179 /// _MSC_VER Compiler is MSVC or Intel on Windows
180 /// _WIN32 Building on Windows (any)
181 /// _WIN64 Building on Windows 64 bit
183 std::string compiler = "\nCompiled by ";
186 compiler += "clang++ ";
187 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
188 #elif __INTEL_COMPILER
189 compiler += "Intel compiler ";
190 compiler += "(version ";
191 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
195 compiler += "(version ";
196 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
198 #elif defined(__e2k__) && defined(__LCC__)
199 #define dot_ver2(n) \
200 compiler += (char)'.'; \
201 compiler += (char)('0' + (n) / 10); \
202 compiler += (char)('0' + (n) % 10);
204 compiler += "MCST LCC ";
205 compiler += "(version ";
206 compiler += std::to_string(__LCC__ / 100);
207 dot_ver2(__LCC__ % 100)
208 dot_ver2(__LCC_MINOR__)
211 compiler += "g++ (GNUC) ";
212 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
214 compiler += "Unknown compiler ";
215 compiler += "(unknown version)";
218 #if defined(__APPLE__)
219 compiler += " on Apple";
220 #elif defined(__CYGWIN__)
221 compiler += " on Cygwin";
222 #elif defined(__MINGW64__)
223 compiler += " on MinGW64";
224 #elif defined(__MINGW32__)
225 compiler += " on MinGW32";
226 #elif defined(__ANDROID__)
227 compiler += " on Android";
228 #elif defined(__linux__)
229 compiler += " on Linux";
230 #elif defined(_WIN64)
231 compiler += " on Microsoft Windows 64-bit";
232 #elif defined(_WIN32)
233 compiler += " on Microsoft Windows 32-bit";
235 compiler += " on unknown system";
238 compiler += "\nCompilation settings include: ";
239 compiler += (Is64Bit ? " 64bit" : " 32bit");
240 #if defined(USE_VNNI)
243 #if defined(USE_AVX512)
244 compiler += " AVX512";
246 compiler += (HasPext ? " BMI2" : "");
247 #if defined(USE_AVX2)
250 #if defined(USE_SSE41)
251 compiler += " SSE41";
253 #if defined(USE_SSSE3)
254 compiler += " SSSE3";
256 #if defined(USE_SSE2)
259 compiler += (HasPopCnt ? " POPCNT" : "");
263 #if defined(USE_NEON)
268 compiler += " DEBUG";
271 compiler += "\n__VERSION__ macro expands to: ";
273 compiler += __VERSION__;
275 compiler += "(undefined macro)";
283 /// Debug functions used mainly to collect run-time statistics
284 static std::atomic<int64_t> hits[2], means[2];
286 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
287 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
288 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
293 cerr << "Total " << hits[0] << " Hits " << hits[1]
294 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
297 cerr << "Total " << means[0] << " Mean "
298 << (double)means[1] / means[0] << endl;
302 /// Used to serialize access to std::cout to avoid multiple threads writing at
305 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
319 /// Trampoline helper to avoid moving Logger to misc.h
320 void start_logger(const std::string& fname) { Logger::start(fname); }
323 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
324 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
325 /// which can be quite slow.
328 void prefetch(void*) {}
332 void prefetch(void* addr) {
334 # if defined(__INTEL_COMPILER)
335 // This hack prevents prefetches from being optimized away by
336 // Intel compiler. Both MSVC and gcc seem not be affected by this.
340 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
341 _mm_prefetch((char*)addr, _MM_HINT_T0);
343 __builtin_prefetch(addr);
350 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
351 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
352 /// std_aligned_alloc() must be freed with std_aligned_free().
354 void* std_aligned_alloc(size_t alignment, size_t size) {
356 #if defined(POSIXALIGNEDALLOC)
358 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
359 #elif defined(_WIN32)
360 return _mm_malloc(size, alignment);
362 return std::aligned_alloc(alignment, size);
366 void std_aligned_free(void* ptr) {
368 #if defined(POSIXALIGNEDALLOC)
370 #elif defined(_WIN32)
377 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
381 static void* aligned_large_pages_alloc_windows(size_t allocSize) {
384 (void)allocSize; // suppress unused-parameter compiler warning
388 HANDLE hProcessToken { };
392 const size_t largePageSize = GetLargePageMinimum();
396 // We need SeLockMemoryPrivilege, so try to enable it for the process
397 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
400 if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid))
402 TOKEN_PRIVILEGES tp { };
403 TOKEN_PRIVILEGES prevTp { };
406 tp.PrivilegeCount = 1;
407 tp.Privileges[0].Luid = luid;
408 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
410 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
411 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
412 if (AdjustTokenPrivileges(
413 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
414 GetLastError() == ERROR_SUCCESS)
416 // Round up size to full pages and allocate
417 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
419 NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
421 // Privilege no longer needed, restore previous state
422 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL);
426 CloseHandle(hProcessToken);
433 void* aligned_large_pages_alloc(size_t allocSize) {
435 // Try to allocate large pages
436 void* mem = aligned_large_pages_alloc_windows(allocSize);
438 // Fall back to regular, page aligned, allocation if necessary
440 mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
447 void* aligned_large_pages_alloc(size_t allocSize) {
449 #if defined(__linux__)
450 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
452 constexpr size_t alignment = 4096; // assumed small page size
455 // round up to multiples of alignment
456 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
457 void *mem = std_aligned_alloc(alignment, size);
458 #if defined(MADV_HUGEPAGE)
459 madvise(mem, size, MADV_HUGEPAGE);
467 /// aligned_large_pages_free() will free the previously allocated ttmem
471 void aligned_large_pages_free(void* mem) {
473 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
475 DWORD err = GetLastError();
476 std::cerr << "Failed to free large page memory. Error code: 0x"
478 << std::dec << std::endl;
485 void aligned_large_pages_free(void *mem) {
486 std_aligned_free(mem);
492 namespace WinProcGroup {
496 void bindThisThread(size_t) {}
500 /// best_node() retrieves logical processor information using Windows specific
501 /// API and returns the best node id for the thread with index idx. Original
502 /// code from Texel by Peter Ă–sterlund.
504 int best_node(size_t idx) {
509 DWORD returnLength = 0;
510 DWORD byteOffset = 0;
512 // Early exit if the needed API is not available at runtime
513 HMODULE k32 = GetModuleHandle("Kernel32.dll");
514 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
518 // First call to GetLogicalProcessorInformationEx() to get returnLength.
519 // We expect the call to fail due to null buffer.
520 if (fun1(RelationAll, nullptr, &returnLength))
523 // Once we know returnLength, allocate the buffer
524 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
525 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
527 // Second call to GetLogicalProcessorInformationEx(), now we expect to succeed
528 if (!fun1(RelationAll, buffer, &returnLength))
534 while (byteOffset < returnLength)
536 if (ptr->Relationship == RelationNumaNode)
539 else if (ptr->Relationship == RelationProcessorCore)
542 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
546 byteOffset += ptr->Size;
547 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
552 std::vector<int> groups;
554 // Run as many threads as possible on the same node until core limit is
555 // reached, then move on filling the next node.
556 for (int n = 0; n < nodes; n++)
557 for (int i = 0; i < cores / nodes; i++)
560 // In case a core has more than one logical processor (we assume 2) and we
561 // have still threads to allocate, then spread them evenly across available
563 for (int t = 0; t < threads - cores; t++)
564 groups.push_back(t % nodes);
566 // If we still have more threads than the total number of logical processors
567 // then return -1 and let the OS to decide what to do.
568 return idx < groups.size() ? groups[idx] : -1;
572 /// bindThisThread() set the group affinity of the current thread
574 void bindThisThread(size_t idx) {
576 // Use only local variables to be thread-safe
577 int node = best_node(idx);
582 // Early exit if the needed API are not available at runtime
583 HMODULE k32 = GetModuleHandle("Kernel32.dll");
584 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
585 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
586 auto fun4 = (fun4_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMask2");
587 auto fun5 = (fun5_t)(void(*)())GetProcAddress(k32, "GetMaximumProcessorGroupCount");
594 GROUP_AFFINITY affinity;
595 if (fun2(node, &affinity)) // GetNumaNodeProcessorMaskEx
596 fun3(GetCurrentThread(), &affinity, nullptr); // SetThreadGroupAffinity
600 // If a numa node has more than one processor group, we assume they are
601 // sized equal and we spread threads evenly across the groups.
602 USHORT elements, returnedElements;
603 elements = fun5(); // GetMaximumProcessorGroupCount
604 GROUP_AFFINITY *affinity = (GROUP_AFFINITY*)malloc(elements * sizeof(GROUP_AFFINITY));
605 if (fun4(node, affinity, elements, &returnedElements)) // GetNumaNodeProcessorMask2
606 fun3(GetCurrentThread(), &affinity[idx % returnedElements], nullptr); // SetThreadGroupAffinity
613 } // namespace WinProcGroup
617 #define GETCWD _getcwd
620 #define GETCWD getcwd
623 namespace CommandLine {
625 string argv0; // path+name of the executable binary, as given by argv[0]
626 string binaryDirectory; // path of the executable directory
627 string workingDirectory; // path of the working directory
629 void init(int argc, char* argv[]) {
631 string pathSeparator;
633 // extract the path+name of the executable binary
637 pathSeparator = "\\";
639 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
640 // issues in some windows 10 versions, so check returned values carefully.
641 char* pgmptr = nullptr;
642 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
649 // extract the working directory
650 workingDirectory = "";
652 char* cwd = GETCWD(buff, 40000);
654 workingDirectory = cwd;
656 // extract the binary directory path from argv0
657 binaryDirectory = argv0;
658 size_t pos = binaryDirectory.find_last_of("\\/");
659 if (pos == std::string::npos)
660 binaryDirectory = "." + pathSeparator;
662 binaryDirectory.resize(pos + 1);
664 // pattern replacement: "./" at the start of path is replaced by the working directory
665 if (binaryDirectory.find("." + pathSeparator) == 0)
666 binaryDirectory.replace(0, 1, workingDirectory);
670 } // namespace CommandLine
672 } // namespace Stockfish