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);
155 ss << (to_uci ? "\nid author ": " by ")
156 << "the Stockfish developers (see AUTHORS file)";
162 /// compiler_info() returns a string trying to describe the compiler we use
164 const std::string compiler_info() {
166 #define stringify2(x) #x
167 #define stringify(x) stringify2(x)
168 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
170 /// Predefined macros hell:
172 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
173 /// __INTEL_COMPILER Compiler is Intel
174 /// _MSC_VER Compiler is MSVC or Intel on Windows
175 /// _WIN32 Building on Windows (any)
176 /// _WIN64 Building on Windows 64 bit
178 std::string compiler = "\nCompiled by ";
181 compiler += "clang++ ";
182 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
183 #elif __INTEL_COMPILER
184 compiler += "Intel compiler ";
185 compiler += "(version ";
186 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
190 compiler += "(version ";
191 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
194 compiler += "g++ (GNUC) ";
195 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
197 compiler += "Unknown compiler ";
198 compiler += "(unknown version)";
201 #if defined(__APPLE__)
202 compiler += " on Apple";
203 #elif defined(__CYGWIN__)
204 compiler += " on Cygwin";
205 #elif defined(__MINGW64__)
206 compiler += " on MinGW64";
207 #elif defined(__MINGW32__)
208 compiler += " on MinGW32";
209 #elif defined(__ANDROID__)
210 compiler += " on Android";
211 #elif defined(__linux__)
212 compiler += " on Linux";
213 #elif defined(_WIN64)
214 compiler += " on Microsoft Windows 64-bit";
215 #elif defined(_WIN32)
216 compiler += " on Microsoft Windows 32-bit";
218 compiler += " on unknown system";
221 compiler += "\nCompilation settings include: ";
222 compiler += (Is64Bit ? " 64bit" : " 32bit");
223 #if defined(USE_VNNI)
226 #if defined(USE_AVX512)
227 compiler += " AVX512";
229 compiler += (HasPext ? " BMI2" : "");
230 #if defined(USE_AVX2)
233 #if defined(USE_SSE41)
234 compiler += " SSE41";
236 #if defined(USE_SSSE3)
237 compiler += " SSSE3";
239 #if defined(USE_SSE2)
242 compiler += (HasPopCnt ? " POPCNT" : "");
246 #if defined(USE_NEON)
251 compiler += " DEBUG";
254 compiler += "\n__VERSION__ macro expands to: ";
256 compiler += __VERSION__;
258 compiler += "(undefined macro)";
266 /// Debug functions used mainly to collect run-time statistics
267 static std::atomic<int64_t> hits[2], means[2];
269 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
270 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
271 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
276 cerr << "Total " << hits[0] << " Hits " << hits[1]
277 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
280 cerr << "Total " << means[0] << " Mean "
281 << (double)means[1] / means[0] << endl;
285 /// Used to serialize access to std::cout to avoid multiple threads writing at
288 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
302 /// Trampoline helper to avoid moving Logger to misc.h
303 void start_logger(const std::string& fname) { Logger::start(fname); }
306 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
307 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
308 /// which can be quite slow.
311 void prefetch(void*) {}
315 void prefetch(void* addr) {
317 # if defined(__INTEL_COMPILER)
318 // This hack prevents prefetches from being optimized away by
319 // Intel compiler. Both MSVC and gcc seem not be affected by this.
323 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
324 _mm_prefetch((char*)addr, _MM_HINT_T0);
326 __builtin_prefetch(addr);
333 /// std_aligned_alloc() is our wrapper for systems where the c++17 implementation
334 /// does not guarantee the availability of aligned_alloc(). Memory allocated with
335 /// std_aligned_alloc() must be freed with std_aligned_free().
337 void* std_aligned_alloc(size_t alignment, size_t size) {
339 #if defined(POSIXALIGNEDALLOC)
341 return posix_memalign(&mem, alignment, size) ? nullptr : mem;
342 #elif defined(_WIN32)
343 return _mm_malloc(size, alignment);
345 return std::aligned_alloc(alignment, size);
349 void std_aligned_free(void* ptr) {
351 #if defined(POSIXALIGNEDALLOC)
353 #elif defined(_WIN32)
360 /// aligned_large_pages_alloc() will return suitably aligned memory, if possible using large pages.
364 static void* aligned_large_pages_alloc_win(size_t allocSize) {
366 HANDLE hProcessToken { };
370 const size_t largePageSize = GetLargePageMinimum();
374 // We need SeLockMemoryPrivilege, so try to enable it for the process
375 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
378 if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid))
380 TOKEN_PRIVILEGES tp { };
381 TOKEN_PRIVILEGES prevTp { };
384 tp.PrivilegeCount = 1;
385 tp.Privileges[0].Luid = luid;
386 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
388 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
389 // we still need to query GetLastError() to ensure that the privileges were actually obtained.
390 if (AdjustTokenPrivileges(
391 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
392 GetLastError() == ERROR_SUCCESS)
394 // Round up size to full pages and allocate
395 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
397 NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
399 // Privilege no longer needed, restore previous state
400 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL);
404 CloseHandle(hProcessToken);
409 void* aligned_large_pages_alloc(size_t allocSize) {
411 // Try to allocate large pages
412 void* mem = aligned_large_pages_alloc_win(allocSize);
414 // Fall back to regular, page aligned, allocation if necessary
416 mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
423 void* aligned_large_pages_alloc(size_t allocSize) {
425 #if defined(__linux__)
426 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page size
428 constexpr size_t alignment = 4096; // assumed small page size
431 // round up to multiples of alignment
432 size_t size = ((allocSize + alignment - 1) / alignment) * alignment;
433 void *mem = std_aligned_alloc(alignment, size);
434 #if defined(MADV_HUGEPAGE)
435 madvise(mem, size, MADV_HUGEPAGE);
443 /// aligned_large_pages_free() will free the previously allocated ttmem
447 void aligned_large_pages_free(void* mem) {
449 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
451 DWORD err = GetLastError();
452 std::cerr << "Failed to free transposition table. Error code: 0x" <<
453 std::hex << err << std::dec << std::endl;
460 void aligned_large_pages_free(void *mem) {
461 std_aligned_free(mem);
467 namespace WinProcGroup {
471 void bindThisThread(size_t) {}
475 /// best_group() retrieves logical processor information using Windows specific
476 /// API and returns the best group id for the thread with index idx. Original
477 /// code from Texel by Peter Ă–sterlund.
479 int best_group(size_t idx) {
484 DWORD returnLength = 0;
485 DWORD byteOffset = 0;
487 // Early exit if the needed API is not available at runtime
488 HMODULE k32 = GetModuleHandle("Kernel32.dll");
489 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
493 // First call to get returnLength. We expect it to fail due to null buffer
494 if (fun1(RelationAll, nullptr, &returnLength))
497 // Once we know returnLength, allocate the buffer
498 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
499 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
501 // Second call, now we expect to succeed
502 if (!fun1(RelationAll, buffer, &returnLength))
508 while (byteOffset < returnLength)
510 if (ptr->Relationship == RelationNumaNode)
513 else if (ptr->Relationship == RelationProcessorCore)
516 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
520 byteOffset += ptr->Size;
521 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
526 std::vector<int> groups;
528 // Run as many threads as possible on the same node until core limit is
529 // reached, then move on filling the next node.
530 for (int n = 0; n < nodes; n++)
531 for (int i = 0; i < cores / nodes; i++)
534 // In case a core has more than one logical processor (we assume 2) and we
535 // have still threads to allocate, then spread them evenly across available
537 for (int t = 0; t < threads - cores; t++)
538 groups.push_back(t % nodes);
540 // If we still have more threads than the total number of logical processors
541 // then return -1 and let the OS to decide what to do.
542 return idx < groups.size() ? groups[idx] : -1;
546 /// bindThisThread() set the group affinity of the current thread
548 void bindThisThread(size_t idx) {
550 // Use only local variables to be thread-safe
551 int group = best_group(idx);
556 // Early exit if the needed API are not available at runtime
557 HMODULE k32 = GetModuleHandle("Kernel32.dll");
558 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
559 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
564 GROUP_AFFINITY affinity;
565 if (fun2(group, &affinity))
566 fun3(GetCurrentThread(), &affinity, nullptr);
571 } // namespace WinProcGroup
575 #define GETCWD _getcwd
578 #define GETCWD getcwd
581 namespace CommandLine {
583 string argv0; // path+name of the executable binary, as given by argv[0]
584 string binaryDirectory; // path of the executable directory
585 string workingDirectory; // path of the working directory
586 string pathSeparator; // Separator for our current OS
588 void init(int argc, char* argv[]) {
592 // extract the path+name of the executable binary
596 pathSeparator = "\\";
598 // Under windows argv[0] may not have the extension. Also _get_pgmptr() had
599 // issues in some windows 10 versions, so check returned values carefully.
600 char* pgmptr = nullptr;
601 if (!_get_pgmptr(&pgmptr) && pgmptr != nullptr && *pgmptr)
608 // extract the working directory
609 workingDirectory = "";
611 char* cwd = GETCWD(buff, 40000);
613 workingDirectory = cwd;
615 // extract the binary directory path from argv0
616 binaryDirectory = argv0;
617 size_t pos = binaryDirectory.find_last_of("\\/");
618 if (pos == std::string::npos)
619 binaryDirectory = "." + pathSeparator;
621 binaryDirectory.resize(pos + 1);
623 // pattern replacement: "./" at the start of path is replaced by the working directory
624 if (binaryDirectory.find("." + pathSeparator) == 0)
625 binaryDirectory.replace(0, 1, workingDirectory);
629 } // namespace CommandLine