1/*
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) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad
6
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.
11
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.
16
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/>.
19*/
20
21#ifdef _WIN32
22#if _WIN32_WINNT < 0x0601
23#undef _WIN32_WINNT
24#define _WIN32_WINNT 0x0601 // Force to include needed API prototypes
25#endif
26
27#ifndef NOMINMAX
28#define NOMINMAX
29#endif
30
31#include <windows.h>
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.
36extern "C" {
37typedef bool(*fun1_t)(LOGICAL_PROCESSOR_RELATIONSHIP,
38 PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX, PDWORD);
39typedef bool(*fun2_t)(USHORT, PGROUP_AFFINITY);
40typedef bool(*fun3_t)(HANDLE, CONST GROUP_AFFINITY*, PGROUP_AFFINITY);
41}
42#endif
43
44#include <fstream>
45#include <iomanip>
46#include <iostream>
47#include <sstream>
48#include <vector>
49
50#include "misc.h"
51#include "thread.h"
52
53using namespace std;
54
55namespace {
56
57/// Version number. If Version is left empty, then compile date in the format
58/// DD-MM-YY and show in engine_info.
59const string Version = "";
60
61/// Our fancy logging facility. The trick here is to replace cin.rdbuf() and
62/// cout.rdbuf() with two Tie objects that tie cin and cout to a file stream. We
63/// can toggle the logging of std::cout and std:cin at runtime whilst preserving
64/// usual I/O functionality, all without changing a single line of code!
65/// Idea from http://groups.google.com/group/comp.lang.c++/msg/1d941c0f26ea0d81
66
67struct Tie: public streambuf { // MSVC requires split streambuf for cin and cout
68
69 Tie(streambuf* b, streambuf* l) : buf(b), logBuf(l) {}
70
71 int sync() override { return logBuf->pubsync(), buf->pubsync(); }
72 int overflow(int c) override { return log(buf->sputc((char)c), "<< "); }
73 int underflow() override { return buf->sgetc(); }
74 int uflow() override { return log(buf->sbumpc(), ">> "); }
75
76 streambuf *buf, *logBuf;
77
78 int log(int c, const char* prefix) {
79
80 static int last = '\n'; // Single log file
81
82 if (last == '\n')
83 logBuf->sputn(prefix, 3);
84
85 return last = logBuf->sputc((char)c);
86 }
87};
88
89class Logger {
90
91 Logger() : in(cin.rdbuf(), file.rdbuf()), out(cout.rdbuf(), file.rdbuf()) {}
92 ~Logger() { start(""); }
93
94 ofstream file;
95 Tie in, out;
96
97public:
98 static void start(const std::string& fname) {
99
100 static Logger l;
101
102 if (!fname.empty() && !l.file.is_open())
103 {
104 l.file.open(fname, ifstream::out);
105 cin.rdbuf(&l.in);
106 cout.rdbuf(&l.out);
107 }
108 else if (fname.empty() && l.file.is_open())
109 {
110 cout.rdbuf(l.out.buf);
111 cin.rdbuf(l.in.buf);
112 l.file.close();
113 }
114 }
115};
116
117} // namespace
118
119/// engine_info() returns the full name of the current Stockfish version. This
120/// will be either "Stockfish <Tag> DD-MM-YY" (where DD-MM-YY is the date when
121/// the program was compiled) or "Stockfish <Version>", depending on whether
122/// Version is empty.
123
124const string engine_info(bool to_uci) {
125
126 const string months("Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec");
127 string month, day, year;
128 stringstream ss, date(__DATE__); // From compiler, format is "Sep 21 2008"
129
130 ss << "Stockfish " << Version << setfill('0');
131
132 if (Version.empty())
133 {
134 date >> month >> day >> year;
135 ss << setw(2) << day << setw(2) << (1 + months.find(month) / 4) << year.substr(2);
136 }
137
138 ss << (Is64Bit ? " 64" : "")
139 << (HasPext ? " BMI2" : (HasPopCnt ? " POPCNT" : ""))
140 << (to_uci ? "\nid author ": " by ")
141 << "T. Romstad, M. Costalba, J. Kiiski, G. Linscott";
142
143 return ss.str();
144}
145
146
147/// Debug functions used mainly to collect run-time statistics
148static std::atomic<int64_t> hits[2], means[2];
149
150void dbg_hit_on(bool b) { ++hits[0]; if (b) ++hits[1]; }
151void dbg_hit_on(bool c, bool b) { if (c) dbg_hit_on(b); }
152void dbg_mean_of(int v) { ++means[0]; means[1] += v; }
153
154void dbg_print() {
155
156 if (hits[0])
157 cerr << "Total " << hits[0] << " Hits " << hits[1]
158 << " hit rate (%) " << 100 * hits[1] / hits[0] << endl;
159
160 if (means[0])
161 cerr << "Total " << means[0] << " Mean "
162 << (double)means[1] / means[0] << endl;
163}
164
165
166/// Used to serialize access to std::cout to avoid multiple threads writing at
167/// the same time.
168
169std::ostream& operator<<(std::ostream& os, SyncCout sc) {
170
171 static Mutex m;
172
173 if (sc == IO_LOCK)
174 m.lock();
175
176 if (sc == IO_UNLOCK)
177 m.unlock();
178
179 return os;
180}
181
182
183/// Trampoline helper to avoid moving Logger to misc.h
184void start_logger(const std::string& fname) { Logger::start(fname); }
185
186
187/// prefetch() preloads the given address in L1/L2 cache. This is a non-blocking
188/// function that doesn't stall the CPU waiting for data to be loaded from memory,
189/// which can be quite slow.
190#ifdef NO_PREFETCH
191
192void prefetch(void*) {}
193
194#else
195
196void prefetch(void* addr) {
197
198# if defined(__INTEL_COMPILER)
199 // This hack prevents prefetches from being optimized away by
200 // Intel compiler. Both MSVC and gcc seem not be affected by this.
201 __asm__ ("");
202# endif
203
204# if defined(__INTEL_COMPILER) || defined(_MSC_VER)
205 _mm_prefetch((char*)addr, _MM_HINT_T0);
206# else
207 __builtin_prefetch(addr);
208# endif
209}
210
211#endif
212
213namespace WinProcGroup {
214
215#ifndef _WIN32
216
217void bindThisThread(size_t) {}
218
219#else
220
221/// best_group() retrieves logical processor information using Windows specific
222/// API and returns the best group id for the thread with index idx. Original
223/// code from Texel by Peter Ă–sterlund.
224
225int best_group(size_t idx) {
226
227 int threads = 0;
228 int nodes = 0;
229 int cores = 0;
230 DWORD returnLength = 0;
231 DWORD byteOffset = 0;
232
233 // Early exit if the needed API is not available at runtime
234 HMODULE k32 = GetModuleHandle("Kernel32.dll");
235 auto fun1 = (fun1_t)(void(*)())GetProcAddress(k32, "GetLogicalProcessorInformationEx");
236 if (!fun1)
237 return -1;
238
239 // First call to get returnLength. We expect it to fail due to null buffer
240 if (fun1(RelationAll, nullptr, &returnLength))
241 return -1;
242
243 // Once we know returnLength, allocate the buffer
244 SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *buffer, *ptr;
245 ptr = buffer = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)malloc(returnLength);
246
247 // Second call, now we expect to succeed
248 if (!fun1(RelationAll, buffer, &returnLength))
249 {
250 free(buffer);
251 return -1;
252 }
253
254 while (byteOffset < returnLength)
255 {
256 if (ptr->Relationship == RelationNumaNode)
257 nodes++;
258
259 else if (ptr->Relationship == RelationProcessorCore)
260 {
261 cores++;
262 threads += (ptr->Processor.Flags == LTP_PC_SMT) ? 2 : 1;
263 }
264
265 assert(ptr->Size);
266 byteOffset += ptr->Size;
267 ptr = (SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX*)(((char*)ptr) + ptr->Size);
268 }
269
270 free(buffer);
271
272 std::vector<int> groups;
273
274 // Run as many threads as possible on the same node until core limit is
275 // reached, then move on filling the next node.
276 for (int n = 0; n < nodes; n++)
277 for (int i = 0; i < cores / nodes; i++)
278 groups.push_back(n);
279
280 // In case a core has more than one logical processor (we assume 2) and we
281 // have still threads to allocate, then spread them evenly across available
282 // nodes.
283 for (int t = 0; t < threads - cores; t++)
284 groups.push_back(t % nodes);
285
286 // If we still have more threads than the total number of logical processors
287 // then return -1 and let the OS to decide what to do.
288 return idx < groups.size() ? groups[idx] : -1;
289}
290
291
292/// bindThisThread() set the group affinity of the current thread
293
294void bindThisThread(size_t idx) {
295
296 // Use only local variables to be thread-safe
297 int group = best_group(idx);
298
299 if (group == -1)
300 return;
301
302 // Early exit if the needed API are not available at runtime
303 HMODULE k32 = GetModuleHandle("Kernel32.dll");
304 auto fun2 = (fun2_t)(void(*)())GetProcAddress(k32, "GetNumaNodeProcessorMaskEx");
305 auto fun3 = (fun3_t)(void(*)())GetProcAddress(k32, "SetThreadGroupAffinity");
306
307 if (!fun2 || !fun3)
308 return;
309
310 GROUP_AFFINITY affinity;
311 if (fun2(group, &affinity))
312 fun3(GetCurrentThread(), &affinity, nullptr);
313}
314
315#endif
316
317} // namespace WinProcGroup
318