X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fmisc.cpp;h=7a23b7a9c930791ce41c254ac3834fbff2941978;hp=0f0fca46624da7c0212b2971c72b0d8a02c2eae8;hb=e0504ab876a997321102f040ab88203cb893db12;hpb=7a8429d9f15a45c5a60f98c4aed798d10d511300 diff --git a/src/misc.cpp b/src/misc.cpp index 0f0fca46..7a23b7a9 100644 --- a/src/misc.cpp +++ b/src/misc.cpp @@ -1,7 +1,8 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2012 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2015-2016 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -17,64 +18,133 @@ along with this program. If not, see . */ +#ifdef _WIN32 +#if _WIN32_WINNT < 0x0601 +#undef _WIN32_WINNT +#define _WIN32_WINNT 0x0601 // Force to include needed API prototypes +#endif +#include +// The needed Windows API for processor groups could be missed from old Windows +// versions, so instead of calling them directly (forcing the linker to resolve +// the calls at compile time), try to load them at runtime. To do this we need +// first to define the corresponding function pointers. +extern "C" { +typedef bool(*fun1_t)(LOGICAL_PROCESSOR_RELATIONSHIP, + PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD); +typedef bool(*fun2_t)(USHORT, PGROUP_AFFINITY); +typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY); +} +#endif + +#include #include #include #include +#include #include "misc.h" #include "thread.h" -#if defined(__hpux) -# include -#endif - using namespace std; -/// Version number. If Version is left empty, then Tag plus current -/// date (in the format YYMMDD) is used as a version number. +namespace { + +/// Version number. If Version is left empty, then compile date in the format +/// DD-MM-YY and show in engine_info. +const string Version = ""; + +/// Our fancy logging facility. The trick here is to replace cin.rdbuf() and +/// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We +/// can toggle the logging of std::cout and std:cin at runtime whilst preserving +/// usual I/O functionality, all without changing a single line of code! +/// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81 + +struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout + + Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {} -static const string Version = ""; -static const string Tag = ""; + int sync() { return logBuf->pubsync(), buf->pubsync(); } + int overflow(int c) { return log(buf->sputc((char)c), "<< "); } + int underflow() { return buf->sgetc(); } + int uflow() { return log(buf->sbumpc(), ">> "); } + streambuf *buf, *logBuf; -/// engine_info() returns the full name of the current Stockfish version. -/// This will be either "Stockfish YYMMDD" (where YYMMDD is the date when -/// the program was compiled) or "Stockfish ", depending -/// on whether Version is empty. + int log(int c, const char* prefix) { + + static int last = '\n'; // Single log file + + if (last == '\n') + logBuf->sputn(prefix, 3); + + return last = logBuf->sputc((char)c); + } +}; + +class Logger { + + Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {} + ~Logger() { start(""); } + + ofstream file; + Tie in, out; + +public: + static void start(const std::string& fname) { + + static Logger l; + + if (!fname.empty() && !l.file.is_open()) + { + l.file.open(fname, ifstream::out); + cin.rdbuf(&l.in); + cout.rdbuf(&l.out); + } + else if (fname.empty() && l.file.is_open()) + { + cout.rdbuf(l.out.buf); + cin.rdbuf(l.in.buf); + l.file.close(); + } + } +}; + +} // namespace + +/// engine_info() returns the full name of the current Stockfish version. This +/// will be either "Stockfish DD-MM-YY" (where DD-MM-YY is the date when +/// the program was compiled) or "Stockfish ", depending on whether +/// Version is empty. const string engine_info(bool to_uci) { const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec"); - const string cpu64(Is64Bit ? " 64bit" : ""); - const string popcnt(HasPopCnt ? " SSE4.2" : ""); - string month, day, year; - stringstream s, date(__DATE__); // From compiler, format is "Sep 21 2008" + stringstream ss, date(__DATE__); // From compiler, format is "Sep 21 2008" - s << "Stockfish " << Version; + ss << "Stockfish " << Version << setfill('0'); if (Version.empty()) { date >> month >> day >> year; - - s << Tag << setfill('0') << " " << year.substr(2) - << setw(2) << (1 + months.find(month) / 4) << setw(2) << day; + ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2); } - s << cpu64 << popcnt << (to_uci ? "\nid author ": " by ") - << "Tord Romstad, Marco Costalba and Joona Kiiski"; + ss << (Is64Bit ? " 64" : "") + << (HasPext ? " BMI2" : (HasPopCnt ? " POPCNT" : "")) + << (to_uci ? "\nid author ": " by ") + << "T. Romstad, M. Costalba, J. Kiiski, G. Linscott"; - return s.str(); + return ss.str(); } /// Debug functions used mainly to collect run-time statistics +static int64_t hits[2], means[2]; -static uint64_t hits[2], means[2]; - -void dbg_hit_on(bool b) { hits[0]++; if (b) hits[1]++; } -void dbg_hit_on_c(bool c, bool b) { if (c) dbg_hit_on(b); } -void dbg_mean_of(int v) { means[0]++; means[1] += v; } +void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; } +void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); } +void dbg_mean_of(int v) { ++means[0]; means[1] += v; } void dbg_print() { @@ -84,137 +154,166 @@ void dbg_print() { if (means[0]) cerr << "Total " << means[0] << " Mean " - << (float)means[1] / means[0] << endl; + << (double)means[1] / means[0] << endl; } -/// Our fancy logging facility. The trick here is to replace cin.rdbuf() and -/// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We -/// can toggle the logging of std::cout and std:cin at runtime while preserving -/// usual i/o functionality and without changing a single line of code! -/// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81 - -class Logger { +/// Used to serialize access to std::cout to avoid multiple threads writing at +/// the same time. - Logger() : in(cin.rdbuf(), file), out(cout.rdbuf(), file) {} - ~Logger() { start(false); } +std::ostream& operator<<(std::ostream& os, SyncCout sc) { - struct Tie: public streambuf { // MSVC requires splitted streambuf for cin and cout + static Mutex m; - Tie(streambuf* b, ofstream& f) : buf(b), file(f) {} + if (sc == IO_LOCK) + m.lock(); - int sync() { return file.rdbuf()->pubsync(), buf->pubsync(); } - int overflow(int c) { return log(buf->sputc((char)c), "<< "); } - int underflow() { return buf->sgetc(); } - int uflow() { return log(buf->sbumpc(), ">> "); } + if (sc == IO_UNLOCK) + m.unlock(); - int log(int c, const char* prefix) { + return os; +} - static int last = '\n'; - if (last == '\n') - file.rdbuf()->sputn(prefix, 3); +/// Trampoline helper to avoid moving Logger to misc.h +void start_logger(const std::string& fname) { Logger::start(fname); } - return last = file.rdbuf()->sputc((char)c); - } - streambuf* buf; - ofstream& file; - }; +/// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking +/// function that doesn't stall the CPU waiting for data to be loaded from memory, +/// which can be quite slow. +#ifdef NO_PREFETCH - ofstream file; - Tie in, out; +void prefetch(void*) {} -public: - static void start(bool b) { +#else - static Logger l; +void prefetch(void* addr) { - if (b && !l.file.is_open()) - { - l.file.open("io_log.txt", ifstream::out | ifstream::app); - cin.rdbuf(&l.in); - cout.rdbuf(&l.out); - } - else if (!b && l.file.is_open()) - { - cout.rdbuf(l.out.buf); - cin.rdbuf(l.in.buf); - l.file.close(); - } - } -}; +# if defined(__INTEL_COMPILER) + // This hack prevents prefetches from being optimized away by + // Intel compiler. Both MSVC and gcc seem not be affected by this. + __asm__ (""); +# endif +# if defined(__INTEL_COMPILER) || defined(_MSC_VER) + _mm_prefetch((char*)addr, _MM_HINT_T0); +# else + __builtin_prefetch(addr); +# endif +} -/// Trampoline helper to avoid moving Logger to misc.h -void start_logger(bool b) { Logger::start(b); } +#endif +namespace WinProcGroup { -/// cpu_count() tries to detect the number of CPU cores +#ifndef _WIN32 -int cpu_count() { +void bindThisThread(size_t) {} -#if defined(_WIN32) || defined(_WIN64) - SYSTEM_INFO s; - GetSystemInfo(&s); - return s.dwNumberOfProcessors; #else -# if defined(_SC_NPROCESSORS_ONLN) - return sysconf(_SC_NPROCESSORS_ONLN); -# elif defined(__hpux) - struct pst_dynamic psd; - if (pstat_getdynamic(&psd, sizeof(psd), (size_t)1, 0) == -1) - return 1; - return psd.psd_proc_cnt; -# else - return 1; -# endif +/// get_group() retrieves logical processor information using Windows specific +/// API and returns the best group id for the thread with index idx. Original +/// code from Texel by Peter Österlund. -#endif -} +int get_group(size_t idx) { + int threads = 0; + int nodes = 0; + int cores = 0; + DWORD returnLength = 0; + DWORD byteOffset = 0; -/// timed_wait() waits for msec milliseconds. It is mainly an helper to wrap -/// conversion from milliseconds to struct timespec, as used by pthreads. + // Early exit if the needed API is not available at runtime + HMODULE k32 = GetModuleHandle("Kernel32.dll"); + auto fun1 = (fun1_t)GetProcAddress(k32, "GetLogicalProcessorInformationEx"); + if (!fun1) + return -1; -void timed_wait(WaitCondition& sleepCond, Lock& sleepLock, int msec) { + // First call to get returnLength. We expect it to fail due to null buffer + if (fun1(RelationAll, nullptr, &returnLength)) + return -1; -#if defined(_WIN32) || defined(_WIN64) - int tm = msec; -#else - timespec ts, *tm = &ts; - uint64_t ms = Time::current_time().msec() + msec; + // Once we know returnLength, allocate the buffer + SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr; + ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength); - ts.tv_sec = ms / 1000; - ts.tv_nsec = (ms % 1000) * 1000000LL; -#endif + // Second call, now we expect to succeed + if (!fun1(RelationAll, buffer, &returnLength)) + { + free(buffer); + return -1; + } + + while (ptr->Size > 0 && byteOffset + ptr->Size <= returnLength) + { + if (ptr->Relationship == RelationNumaNode) + nodes++; - cond_timedwait(sleepCond, sleepLock, tm); + else if (ptr->Relationship == RelationProcessorCore) + { + cores++; + threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1; + } + + byteOffset += ptr->Size; + ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size); + } + + free(buffer); + + std::vector groups; + + // Run as many threads as possible on the same node until core limit is + // reached, then move on filling the next node. + for (int n = 0; n < nodes; n++) + for (int i = 0; i < cores / nodes; i++) + groups.push_back(n); + + // In case a core has more than one logical processor (we assume 2) and we + // have still threads to allocate, then spread them evenly across available + // nodes. + for (int t = 0; t < threads - cores; t++) + groups.push_back(t % nodes); + + // If we still have more threads than the total number of logical processors + // then return -1 and let the OS to decide what to do. + return idx < groups.size() ? groups[idx] : -1; } -/// prefetch() preloads the given address in L1/L2 cache. This is a non -/// blocking function and do not stalls the CPU waiting for data to be -/// loaded from memory, that can be quite slow. -#if defined(NO_PREFETCH) +/// bindThisThread() set the group affinity of the current thread -void prefetch(char*) {} +void bindThisThread(size_t idx) { -#else + // If OS already scheduled us on a different group than 0 then don't overwrite + // the choice, eventually we are one of many one-threaded processes running on + // some Windows NUMA hardware, for instance in fishtest. To make it simple, + // just check if running threads are below a threshold, in this case all this + // NUMA machinery is not needed. + if (Threads.size() < 8) + return; -# include + // Use only local variables to be thread-safe + int group = get_group(idx); -void prefetch(char* addr) { + if (group == -1) + return; -# if defined(__INTEL_COMPILER) || defined(__ICL) - // This hack prevents prefetches to be optimized away by - // Intel compiler. Both MSVC and gcc seems not affected. - __asm__ (""); -# endif + // Early exit if the needed API are not available at runtime + HMODULE k32 = GetModuleHandle("Kernel32.dll"); + auto fun2 = (fun2_t)GetProcAddress(k32, "GetNumaNodeProcessorMaskEx"); + auto fun3 = (fun3_t)GetProcAddress(k32, "SetThreadGroupAffinity"); + + if (!fun2 || !fun3) + return; - _mm_prefetch(addr, _MM_HINT_T2); - _mm_prefetch(addr+64, _MM_HINT_T2); // 64 bytes ahead + GROUP_AFFINITY affinity; + if (fun2(group, &affinity)) + fun3(GetCurrentThread(), &affinity, nullptr); } #endif + +} // namespace WinProcGroup