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)();
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 or dev.
71 const string version = "dev";
73 /// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
74 /// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
75 /// can toggle the logging of std::cout and std:cin at runtime whilst preserving
76 /// usual I/O functionality, all without changing a single line of code!
77 /// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
79 struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout
81 Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {}
83 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
84 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
85 int underflow() override { return buf->sgetc(); }
86 int uflow() override { return log(buf->sbumpc(), ">> "); }
88 streambuf *buf, *logBuf;
90 int log(int c, const char* prefix) {
92 static int last = '\n'; // Single log file
95 logBuf->sputn(prefix, 3);
97 return last = logBuf->sputc((char)c);
103 Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {}
104 ~Logger() { start(""); }
110 static void start(const std::string& fname) {
114 if (l.file.is_open())
116 cout.rdbuf(l.out.buf);
123 l.file.open(fname, ifstream::out);
125 if (!l.file.is_open())
127 cerr << "Unable to open debug log file " << fname << endl;
140 /// engine_info() returns the full name of the current Stockfish version.
141 /// For local dev compiles we try to append the commit sha and commit date
142 /// from git if that fails only the local compilation date is set and "nogit" is specified:
143 /// Stockfish dev-YYYYMMDD-SHA
145 /// Stockfish dev-YYYYMMDD-nogit
147 /// For releases (non dev builds) we only include the version number:
148 /// Stockfish version
150 string engine_info(bool to_uci) {
152 ss << "Stockfish " << version << setfill('0');
154 if (version == "dev")
160 const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
161 string month, day, year;
162 stringstream date(__DATE__); // From compiler, format is "Sep 21 2008"
164 date >> month >> day >> year;
165 ss << year << setw(2) << setfill('0') << (1 + months.find(month) / 4) << setw(2) << setfill('0') << day;
177 ss << (to_uci ? "\nid author ": " by ")
178 << "the Stockfish developers (see AUTHORS file)";
184 /// compiler_info() returns a string trying to describe the compiler we use
186 std::string compiler_info() {
188 #define stringify2(x) #x
189 #define stringify(x) stringify2(x)
190 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
192 /// Predefined macros hell:
194 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
195 /// __INTEL_COMPILER Compiler is Intel
196 /// _MSC_VER Compiler is MSVC or Intel on Windows
197 /// _WIN32 Building on Windows (any)
198 /// _WIN64 Building on Windows 64 bit
200 std::string compiler = "\nCompiled by ";
203 compiler += "clang++ ";
204 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
205 #elif __INTEL_COMPILER
206 compiler += "Intel compiler ";
207 compiler += "(version ";
208 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
212 compiler += "(version ";
213 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
215 #elif defined(__e2k__) && defined(__LCC__)
216 #define dot_ver2(n) \
217 compiler += (char)'.'; \
218 compiler += (char)('0' + (n) / 10); \
219 compiler += (char)('0' + (n) % 10);
221 compiler += "MCST LCC ";
222 compiler += "(version ";
223 compiler += std::to_string(__LCC__ / 100);
224 dot_ver2(__LCC__ % 100)
225 dot_ver2(__LCC_MINOR__)
228 compiler += "g++ (GNUC) ";
229 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
231 compiler += "Unknown compiler ";
232 compiler += "(unknown version)";
235 #if defined(__APPLE__)
236 compiler += " on Apple";
237 #elif defined(__CYGWIN__)
238 compiler += " on Cygwin";
239 #elif defined(__MINGW64__)
240 compiler += " on MinGW64";
241 #elif defined(__MINGW32__)
242 compiler += " on MinGW32";
243 #elif defined(__ANDROID__)
244 compiler += " on Android";
245 #elif defined(__linux__)
246 compiler += " on Linux";
247 #elif defined(_WIN64)
248 compiler += " on Microsoft Windows 64-bit";
249 #elif defined(_WIN32)
250 compiler += " on Microsoft Windows 32-bit";
252 compiler += " on unknown system";
255 compiler += "\nCompilation settings include: ";
256 compiler += (Is64Bit ? " 64bit" : " 32bit");
257 #if defined(USE_VNNI)
260 #if defined(USE_AVX512)
261 compiler += " AVX512";
263 compiler += (HasPext ? " BMI2" : "");
264 #if defined(USE_AVX2)
267 #if defined(USE_SSE41)
268 compiler += " SSE41";
270 #if defined(USE_SSSE3)
271 compiler += " SSSE3";
273 #if defined(USE_SSE2)
276 compiler += (HasPopCnt ? " POPCNT" : "");
280 #if defined(USE_NEON)
285 compiler += " DEBUG";
288 compiler += "\n__VERSION__ macro expands to: ";
290 compiler += __VERSION__;
292 compiler += "(undefined macro)";
300 /// Debug functions used mainly to collect run-time statistics
301 static std::atomic<int64_t> hits[2], means[2];
303 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
304 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
305 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
310 cerr << "Total " << hits[0] << " Hits " << hits[1]
311 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
314 cerr << "Total " << means[0] << " Mean "
315 << (double)means[1] / means[0] << endl;
319 /// Used to serialize access to std::cout to avoid multiple threads writing at
322 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
336 /// Trampoline helper to avoid moving Logger to misc.h
337 void start_logger(const std::string& fname) { Logger::start(fname); }
340 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
341 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
342 /// which can be quite slow.
345 void prefetch(void*) {}
349 void prefetch(void* addr) {
351 # if defined(__INTEL_COMPILER)
352 // This hack prevents prefetches from being optimized away by
353 // Intel compiler. Both MSVC and gcc seem not be affected by this.
357 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
358 _mm_prefetch((char*)addr, _MM_HINT_T0);
360 __builtin_prefetch(addr);
367 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
368 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
369 /// std_aligned_alloc() must be freed with std_aligned_free().
371 void* std_aligned_alloc(size_t alignment, size_t size) {
373 #if defined(POSIXALIGNEDALLOC)
375 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
376 #elif defined(_WIN32)
377 return _mm_malloc(size, alignment);
379 return std::aligned_alloc(alignment, size);
383 void std_aligned_free(void* ptr) {
385 #if defined(POSIXALIGNEDALLOC)
387 #elif defined(_WIN32)
394 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
398 static void* aligned_large_pages_alloc_windows([[maybe_unused]] size_t allocSize) {
404 HANDLE hProcessToken { };
408 const size_t largePageSize = GetLargePageMinimum();
412 // We need SeLockMemoryPrivilege, so try to enable it for the process
413 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
416 if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid))
418 TOKEN_PRIVILEGES tp { };
419 TOKEN_PRIVILEGES prevTp { };
422 tp.PrivilegeCount = 1;
423 tp.Privileges[0].Luid = luid;
424 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
426 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
427 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
428 if (AdjustTokenPrivileges(
429 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
430 GetLastError() == ERROR_SUCCESS)
432 // Round up size to full pages and allocate
433 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
435 NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
437 // Privilege no longer needed, restore previous state
438 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL);
442 CloseHandle(hProcessToken);
449 void* aligned_large_pages_alloc(size_t allocSize) {
451 // Try to allocate large pages
452 void* mem = aligned_large_pages_alloc_windows(allocSize);
454 // Fall back to regular, page aligned, allocation if necessary
456 mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
463 void* aligned_large_pages_alloc(size_t allocSize) {
465 #if defined(__linux__)
466 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
468 constexpr size_t alignment = 4096; // assumed small page size
471 // round up to multiples of alignment
472 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
473 void *mem = std_aligned_alloc(alignment, size);
474 #if defined(MADV_HUGEPAGE)
475 madvise(mem, size, MADV_HUGEPAGE);
483 /// aligned_large_pages_free() will free the previously allocated ttmem
487 void aligned_large_pages_free(void* mem) {
489 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
491 DWORD err = GetLastError();
492 std::cerr << "Failed to free large page memory. Error code: 0x"
494 << std::dec << std::endl;
501 void aligned_large_pages_free(void *mem) {
502 std_aligned_free(mem);
508 namespace WinProcGroup {
512 void bindThisThread(size_t) {}
516 /// best_node() retrieves logical processor information using Windows specific
517 /// API and returns the best node id for the thread with index idx. Original
518 /// code from Texel by Peter Ă–sterlund.
520 int best_node(size_t idx) {
525 DWORD returnLength = 0;
526 DWORD byteOffset = 0;
528 // Early exit if the needed API is not available at runtime
529 HMODULE k32 = GetModuleHandle("Kernel32.dll");
530 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
534 // First call to GetLogicalProcessorInformationEx() to get returnLength.
535 // We expect the call to fail due to null buffer.
536 if (fun1(RelationAll, nullptr, &returnLength))
539 // Once we know returnLength, allocate the buffer
540 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
541 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
543 // Second call to GetLogicalProcessorInformationEx(), now we expect to succeed
544 if (!fun1(RelationAll, buffer, &returnLength))
550 while (byteOffset < returnLength)
552 if (ptr->Relationship == RelationNumaNode)
555 else if (ptr->Relationship == RelationProcessorCore)
558 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
562 byteOffset += ptr->Size;
563 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
568 std::vector<int> groups;
570 // Run as many threads as possible on the same node until core limit is
571 // reached, then move on filling the next node.
572 for (int n = 0; n < nodes; n++)
573 for (int i = 0; i < cores / nodes; i++)
576 // In case a core has more than one logical processor (we assume 2) and we
577 // have still threads to allocate, then spread them evenly across available
579 for (int t = 0; t < threads - cores; t++)
580 groups.push_back(t % nodes);
582 // If we still have more threads than the total number of logical processors
583 // then return -1 and let the OS to decide what to do.
584 return idx < groups.size() ? groups[idx] : -1;
588 /// bindThisThread() set the group affinity of the current thread
590 void bindThisThread(size_t idx) {
592 // Use only local variables to be thread-safe
593 int node = best_node(idx);
598 // Early exit if the needed API are not available at runtime
599 HMODULE k32 = GetModuleHandle("Kernel32.dll");
600 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
601 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
602 auto fun4 = (fun4_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMask2");
603 auto fun5 = (fun5_t)(void(*)())GetProcAddress(k32, "GetMaximumProcessorGroupCount");
610 GROUP_AFFINITY affinity;
611 if (fun2(node, &affinity)) // GetNumaNodeProcessorMaskEx
612 fun3(GetCurrentThread(), &affinity, nullptr); // SetThreadGroupAffinity
616 // If a numa node has more than one processor group, we assume they are
617 // sized equal and we spread threads evenly across the groups.
618 USHORT elements, returnedElements;
619 elements = fun5(); // GetMaximumProcessorGroupCount
620 GROUP_AFFINITY *affinity = (GROUP_AFFINITY*)malloc(elements * sizeof(GROUP_AFFINITY));
621 if (fun4(node, affinity, elements, &returnedElements)) // GetNumaNodeProcessorMask2
622 fun3(GetCurrentThread(), &affinity[idx % returnedElements], nullptr); // SetThreadGroupAffinity
629 } // namespace WinProcGroup
633 #define GETCWD _getcwd
636 #define GETCWD getcwd
639 namespace CommandLine {
641 string argv0; // path+name of the executable binary, as given by argv[0]
642 string binaryDirectory; // path of the executable directory
643 string workingDirectory; // path of the working directory
645 void init([[maybe_unused]] int argc, char* argv[]) {
646 string pathSeparator;
648 // extract the path+name of the executable binary
652 pathSeparator = "\\";
654 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
655 // issues in some windows 10 versions, so check returned values carefully.
656 char* pgmptr = nullptr;
657 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
664 // extract the working directory
665 workingDirectory = "";
667 char* cwd = GETCWD(buff, 40000);
669 workingDirectory = cwd;
671 // extract the binary directory path from argv0
672 binaryDirectory = argv0;
673 size_t pos = binaryDirectory.find_last_of("\\/");
674 if (pos == std::string::npos)
675 binaryDirectory = "." + pathSeparator;
677 binaryDirectory.resize(pos + 1);
679 // pattern replacement: "./" at the start of path is replaced by the working directory
680 if (binaryDirectory.find("." + pathSeparator) == 0)
681 binaryDirectory.replace(0, 1, workingDirectory);
685 } // namespace CommandLine
687 } // namespace Stockfish