2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2008 Tord Romstad (Glaurung author)
4 Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad
5 Copyright (C) 2015-2020 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad
7 Stockfish is free software: you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation, either version 3 of the License, or
10 (at your option) any later version.
12 Stockfish is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with this program. If not, see <http://www.gnu.org/licenses/>.
22 #if _WIN32_WINNT < 0x0601
24 #define _WIN32_WINNT 0x0601 // Force to include needed API prototypes
32 // The needed Windows API for processor groups could be missed from old Windows
33 // versions, so instead of calling them directly (forcing the linker to resolve
34 // the calls at compile time), try to load them at runtime. To do this we need
35 // first to define the corresponding function pointers.
37 typedef bool(*fun1_t)(LOGICAL_PROCESSOR_RELATIONSHIP,
38 PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);
39 typedef bool(*fun2_t)(USHORT, PGROUP_AFFINITY);
40 typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
50 #if defined(__linux__) && !defined(__ANDROID__)
62 /// Version number. If Version is left empty, then compile date in the format
63 /// DD-MM-YY and show in engine_info.
64 const string Version = "";
66 /// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
67 /// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
68 /// can toggle the logging of std::cout and std:cin at runtime whilst preserving
69 /// usual I/O functionality, all without changing a single line of code!
70 /// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
72 struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout
74 Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {}
76 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
77 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
78 int underflow() override { return buf->sgetc(); }
79 int uflow() override { return log(buf->sbumpc(), ">> "); }
81 streambuf *buf, *logBuf;
83 int log(int c, const char* prefix) {
85 static int last = '\n'; // Single log file
88 logBuf->sputn(prefix, 3);
90 return last = logBuf->sputc((char)c);
96 Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {}
97 ~Logger() { start(""); }
103 static void start(const std::string& fname) {
107 if (!fname.empty() && !l.file.is_open())
109 l.file.open(fname, ifstream::out);
111 if (!l.file.is_open())
113 cerr << "Unable to open debug log file " << fname << endl;
120 else if (fname.empty() && l.file.is_open())
122 cout.rdbuf(l.out.buf);
131 /// engine_info() returns the full name of the current Stockfish version. This
132 /// will be either "Stockfish <Tag> DD-MM-YY" (where DD-MM-YY is the date when
133 /// the program was compiled) or "Stockfish <Version>", depending on whether
134 /// Version is empty.
136 const string engine_info(bool to_uci) {
138 const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
139 string month, day, year;
140 stringstream ss, date(__DATE__); // From compiler, format is "Sep 21 2008"
142 ss << "Stockfish " << Version << setfill('0');
146 date >> month >> day >> year;
147 ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2);
150 ss << (Is64Bit ? " 64" : "")
151 << (HasPext ? " BMI2" : (HasPopCnt ? " POPCNT" : ""))
152 << (to_uci ? "\nid author ": " by ")
153 << "T. Romstad, M. Costalba, J. Kiiski, G. Linscott";
159 /// compiler_info() returns a string trying to describe the compiler we use
161 const std::string compiler_info() {
163 #define stringify2(x) #x
164 #define stringify(x) stringify2(x)
165 #define make_version_string(major, minor, patch) stringify(major) "." stringify(minor) "." stringify(patch)
167 /// Predefined macros hell:
169 /// __GNUC__ Compiler is gcc, Clang or Intel on Linux
170 /// __INTEL_COMPILER Compiler is Intel
171 /// _MSC_VER Compiler is MSVC or Intel on Windows
172 /// _WIN32 Building on Windows (any)
173 /// _WIN64 Building on Windows 64 bit
175 std::string compiler = "\nCompiled by ";
178 compiler += "clang++ ";
179 compiler += make_version_string(__clang_major__, __clang_minor__, __clang_patchlevel__);
180 #elif __INTEL_COMPILER
181 compiler += "Intel compiler ";
182 compiler += "(version ";
183 compiler += stringify(__INTEL_COMPILER) " update " stringify(__INTEL_COMPILER_UPDATE);
187 compiler += "(version ";
188 compiler += stringify(_MSC_FULL_VER) "." stringify(_MSC_BUILD);
191 compiler += "g++ (GNUC) ";
192 compiler += make_version_string(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__);
194 compiler += "Unknown compiler ";
195 compiler += "(unknown version)";
198 #if defined(__APPLE__)
199 compiler += " on Apple";
200 #elif defined(__CYGWIN__)
201 compiler += " on Cygwin";
202 #elif defined(__MINGW64__)
203 compiler += " on MinGW64";
204 #elif defined(__MINGW32__)
205 compiler += " on MinGW32";
206 #elif defined(__ANDROID__)
207 compiler += " on Android";
208 #elif defined(__linux__)
209 compiler += " on Linux";
210 #elif defined(_WIN64)
211 compiler += " on Microsoft Windows 64-bit";
212 #elif defined(_WIN32)
213 compiler += " on Microsoft Windows 32-bit";
215 compiler += " on unknown system";
218 compiler += "\n __VERSION__ macro expands to: ";
220 compiler += __VERSION__;
222 compiler += "(undefined macro)";
230 /// Debug functions used mainly to collect run-time statistics
231 static std::atomic<int64_t> hits[2], means[2];
233 void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
234 void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
235 void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
240 cerr << "Total " << hits[0] << " Hits " << hits[1]
241 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
244 cerr << "Total " << means[0] << " Mean "
245 << (double)means[1] / means[0] << endl;
249 /// Used to serialize access to std::cout to avoid multiple threads writing at
252 std::ostream& operator<<(std::ostream& os, SyncCout sc) {
266 /// Trampoline helper to avoid moving Logger to misc.h
267 void start_logger(const std::string& fname) { Logger::start(fname); }
270 /// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
271 /// function that doesn't stall the CPU waiting for data to be loaded from memory,
272 /// which can be quite slow.
275 void prefetch(void*) {}
279 void prefetch(void* addr) {
281 # if defined(__INTEL_COMPILER)
282 // This hack prevents prefetches from being optimized away by
283 // Intel compiler. Both MSVC and gcc seem not be affected by this.
287 # if defined(__INTEL_COMPILER) || defined(_MSC_VER)
288 _mm_prefetch((char*)addr, _MM_HINT_T0);
290 __builtin_prefetch(addr);
297 /// aligned_ttmem_alloc will return suitably aligned memory, and if possible use large pages.
298 /// The returned pointer is the aligned one, while the mem argument is the one that needs to be passed to free.
299 /// With c++17 some of this functionality can be simplified.
300 #if defined(__linux__) && !defined(__ANDROID__)
302 void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
304 constexpr size_t alignment = 2 * 1024 * 1024; // assumed 2MB page sizes
305 size_t size = ((allocSize + alignment - 1) / alignment) * alignment; // multiple of alignment
306 if (posix_memalign(&mem, alignment, size))
308 madvise(mem, allocSize, MADV_HUGEPAGE);
312 #elif defined(_WIN64)
314 static void* aligned_ttmem_alloc_large_pages(size_t allocSize) {
316 HANDLE hProcessToken { };
320 const size_t largePageSize = GetLargePageMinimum();
324 // We need SeLockMemoryPrivilege, so try to enable it for the process
325 if (!OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &hProcessToken))
328 if (LookupPrivilegeValue(NULL, SE_LOCK_MEMORY_NAME, &luid))
330 TOKEN_PRIVILEGES tp { };
331 TOKEN_PRIVILEGES prevTp { };
334 tp.PrivilegeCount = 1;
335 tp.Privileges[0].Luid = luid;
336 tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
338 // Try to enable SeLockMemoryPrivilege. Note that even if AdjustTokenPrivileges() succeeds,
339 // we still need to query GetLastError() to ensure that the privileges were actually obtained...
340 if (AdjustTokenPrivileges(
341 hProcessToken, FALSE, &tp, sizeof(TOKEN_PRIVILEGES), &prevTp, &prevTpLen) &&
342 GetLastError() == ERROR_SUCCESS)
344 // round up size to full pages and allocate
345 allocSize = (allocSize + largePageSize - 1) & ~size_t(largePageSize - 1);
347 NULL, allocSize, MEM_RESERVE | MEM_COMMIT | MEM_LARGE_PAGES, PAGE_READWRITE);
349 // privilege no longer needed, restore previous state
350 AdjustTokenPrivileges(hProcessToken, FALSE, &prevTp, 0, NULL, NULL);
354 CloseHandle(hProcessToken);
359 void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
361 static bool firstCall = true;
363 // try to allocate large pages
364 mem = aligned_ttmem_alloc_large_pages(allocSize);
366 // Suppress info strings on the first call. The first call occurs before 'uci'
367 // is received and in that case this output confuses some GUIs.
371 sync_cout << "info string Hash table allocation: Windows large pages used." << sync_endl;
373 sync_cout << "info string Hash table allocation: Windows large pages not used." << sync_endl;
377 // fall back to regular, page aligned, allocation if necessary
379 mem = VirtualAlloc(NULL, allocSize, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
386 void* aligned_ttmem_alloc(size_t allocSize, void*& mem) {
388 constexpr size_t alignment = 64; // assumed cache line size
389 size_t size = allocSize + alignment - 1; // allocate some extra space
391 void* ret = reinterpret_cast<void*>((uintptr_t(mem) + alignment - 1) & ~uintptr_t(alignment - 1));
397 /// aligned_ttmem_free will free the previously allocated ttmem
400 void aligned_ttmem_free(void* mem) {
402 if (mem && !VirtualFree(mem, 0, MEM_RELEASE))
404 DWORD err = GetLastError();
405 std::cerr << "Failed to free transposition table. Error code: 0x" <<
406 std::hex << err << std::dec << std::endl;
413 void aligned_ttmem_free(void *mem) {
420 namespace WinProcGroup {
424 void bindThisThread(size_t) {}
428 /// best_group() retrieves logical processor information using Windows specific
429 /// API and returns the best group id for the thread with index idx. Original
430 /// code from Texel by Peter Ă–sterlund.
432 int best_group(size_t idx) {
437 DWORD returnLength = 0;
438 DWORD byteOffset = 0;
440 // Early exit if the needed API is not available at runtime
441 HMODULE k32 = GetModuleHandle("Kernel32.dll");
442 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
446 // First call to get returnLength. We expect it to fail due to null buffer
447 if (fun1(RelationAll, nullptr, &returnLength))
450 // Once we know returnLength, allocate the buffer
451 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
452 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
454 // Second call, now we expect to succeed
455 if (!fun1(RelationAll, buffer, &returnLength))
461 while (byteOffset < returnLength)
463 if (ptr->Relationship == RelationNumaNode)
466 else if (ptr->Relationship == RelationProcessorCore)
469 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
473 byteOffset += ptr->Size;
474 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
479 std::vector<int> groups;
481 // Run as many threads as possible on the same node until core limit is
482 // reached, then move on filling the next node.
483 for (int n = 0; n < nodes; n++)
484 for (int i = 0; i < cores / nodes; i++)
487 // In case a core has more than one logical processor (we assume 2) and we
488 // have still threads to allocate, then spread them evenly across available
490 for (int t = 0; t < threads - cores; t++)
491 groups.push_back(t % nodes);
493 // If we still have more threads than the total number of logical processors
494 // then return -1 and let the OS to decide what to do.
495 return idx < groups.size() ? groups[idx] : -1;
499 /// bindThisThread() set the group affinity of the current thread
501 void bindThisThread(size_t idx) {
503 // Use only local variables to be thread-safe
504 int group = best_group(idx);
509 // Early exit if the needed API are not available at runtime
510 HMODULE k32 = GetModuleHandle("Kernel32.dll");
511 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
512 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
517 GROUP_AFFINITY affinity;
518 if (fun2(group, &affinity))
519 fun3(GetCurrentThread(), &affinity, nullptr);
524 } // namespace WinProcGroup