1 // Copyright 2008 Dolphin Emulator Project / 2017 Citra Emulator Project
2 // Licensed under GPLv2+
3 // Refer to the license.txt file included.
4
5 #pragma once
6
7 /**
8 * This is a system to schedule events into the emulated machine's future. Time is measured
9 * in main CPU clock cycles.
10 *
11 * To schedule an event, you first have to register its type. This is where you pass in the
12 * callback. You then schedule events using the type id you get back.
13 *
14 * The int cyclesLate that the callbacks get is how many cycles late it was.
15 * So to schedule a new event on a regular basis:
16 * inside callback:
17 * ScheduleEvent(periodInCycles - cyclesLate, callback, "whatever")
18 */
19
20 #include <chrono>
21 #include <functional>
22 #include <limits>
23 #include <string>
24 #include <unordered_map>
25 #include <vector>
26 #include <boost/serialization/split_member.hpp>
27 #include <boost/serialization/vector.hpp>
28 #include "common/common_types.h"
29 #include "common/logging/log.h"
30 #include "common/threadsafe_queue.h"
31 #include "core/global.h"
32
33 // The timing we get from the assembly is 268,111,855.956 Hz
34 // It is possible that this number isn't just an integer because the compiler could have
35 // optimized the multiplication by a multiply-by-constant division.
36 // Rounding to the nearest integer should be fine
37 constexpr u64 BASE_CLOCK_RATE_ARM11 = 268111856;
38 constexpr u64 MAX_VALUE_TO_MULTIPLY = std::numeric_limits<s64>::max() / BASE_CLOCK_RATE_ARM11;
39
msToCycles(int ms)40 constexpr s64 msToCycles(int ms) {
41 // since ms is int there is no way to overflow
42 return BASE_CLOCK_RATE_ARM11 * static_cast<s64>(ms) / 1000;
43 }
44
msToCycles(float ms)45 constexpr s64 msToCycles(float ms) {
46 return static_cast<s64>(BASE_CLOCK_RATE_ARM11 * (0.001f) * ms);
47 }
48
msToCycles(double ms)49 constexpr s64 msToCycles(double ms) {
50 return static_cast<s64>(BASE_CLOCK_RATE_ARM11 * (0.001) * ms);
51 }
52
usToCycles(float us)53 constexpr s64 usToCycles(float us) {
54 return static_cast<s64>(BASE_CLOCK_RATE_ARM11 * (0.000001f) * us);
55 }
56
usToCycles(int us)57 constexpr s64 usToCycles(int us) {
58 return (BASE_CLOCK_RATE_ARM11 * static_cast<s64>(us) / 1000000);
59 }
60
usToCycles(s64 us)61 inline s64 usToCycles(s64 us) {
62 if (us / 1000000 > static_cast<s64>(MAX_VALUE_TO_MULTIPLY)) {
63 LOG_ERROR(Core_Timing, "Integer overflow, use max value");
64 return std::numeric_limits<s64>::max();
65 }
66 if (us > static_cast<s64>(MAX_VALUE_TO_MULTIPLY)) {
67 LOG_DEBUG(Core_Timing, "Time very big, do rounding");
68 return BASE_CLOCK_RATE_ARM11 * (us / 1000000);
69 }
70 return (BASE_CLOCK_RATE_ARM11 * us) / 1000000;
71 }
72
usToCycles(u64 us)73 inline s64 usToCycles(u64 us) {
74 if (us / 1000000 > MAX_VALUE_TO_MULTIPLY) {
75 LOG_ERROR(Core_Timing, "Integer overflow, use max value");
76 return std::numeric_limits<s64>::max();
77 }
78 if (us > MAX_VALUE_TO_MULTIPLY) {
79 LOG_DEBUG(Core_Timing, "Time very big, do rounding");
80 return BASE_CLOCK_RATE_ARM11 * static_cast<s64>(us / 1000000);
81 }
82 return (BASE_CLOCK_RATE_ARM11 * static_cast<s64>(us)) / 1000000;
83 }
84
nsToCycles(float ns)85 constexpr s64 nsToCycles(float ns) {
86 return static_cast<s64>(BASE_CLOCK_RATE_ARM11 * (0.000000001f) * ns);
87 }
88
nsToCycles(int ns)89 constexpr s64 nsToCycles(int ns) {
90 return BASE_CLOCK_RATE_ARM11 * static_cast<s64>(ns) / 1000000000;
91 }
92
nsToCycles(s64 ns)93 inline s64 nsToCycles(s64 ns) {
94 if (ns / 1000000000 > static_cast<s64>(MAX_VALUE_TO_MULTIPLY)) {
95 LOG_ERROR(Core_Timing, "Integer overflow, use max value");
96 return std::numeric_limits<s64>::max();
97 }
98 if (ns > static_cast<s64>(MAX_VALUE_TO_MULTIPLY)) {
99 LOG_DEBUG(Core_Timing, "Time very big, do rounding");
100 return BASE_CLOCK_RATE_ARM11 * (ns / 1000000000);
101 }
102 return (BASE_CLOCK_RATE_ARM11 * ns) / 1000000000;
103 }
104
nsToCycles(u64 ns)105 inline s64 nsToCycles(u64 ns) {
106 if (ns / 1000000000 > MAX_VALUE_TO_MULTIPLY) {
107 LOG_ERROR(Core_Timing, "Integer overflow, use max value");
108 return std::numeric_limits<s64>::max();
109 }
110 if (ns > MAX_VALUE_TO_MULTIPLY) {
111 LOG_DEBUG(Core_Timing, "Time very big, do rounding");
112 return BASE_CLOCK_RATE_ARM11 * (static_cast<s64>(ns) / 1000000000);
113 }
114 return (BASE_CLOCK_RATE_ARM11 * static_cast<s64>(ns)) / 1000000000;
115 }
116
cyclesToNs(s64 cycles)117 constexpr u64 cyclesToNs(s64 cycles) {
118 return cycles * 1000000000 / BASE_CLOCK_RATE_ARM11;
119 }
120
cyclesToUs(s64 cycles)121 constexpr s64 cyclesToUs(s64 cycles) {
122 return cycles * 1000000 / BASE_CLOCK_RATE_ARM11;
123 }
124
cyclesToMs(s64 cycles)125 constexpr u64 cyclesToMs(s64 cycles) {
126 return cycles * 1000 / BASE_CLOCK_RATE_ARM11;
127 }
128
129 namespace Core {
130
131 using TimedCallback = std::function<void(u64 userdata, int cycles_late)>;
132
133 struct TimingEventType {
134 TimedCallback callback;
135 const std::string* name;
136 };
137
138 class Timing {
139
140 public:
141 struct Event {
142 s64 time;
143 u64 fifo_order;
144 u64 userdata;
145 const TimingEventType* type;
146
147 bool operator>(const Event& right) const;
148 bool operator<(const Event& right) const;
149
150 private:
151 template <class Archive>
saveEvent152 void save(Archive& ar, const unsigned int) const {
153 ar& time;
154 ar& fifo_order;
155 ar& userdata;
156 std::string name = *(type->name);
157 ar << name;
158 }
159
160 template <class Archive>
loadEvent161 void load(Archive& ar, const unsigned int) {
162 ar& time;
163 ar& fifo_order;
164 ar& userdata;
165 std::string name;
166 ar >> name;
167 type = Global<Timing>().RegisterEvent(name, nullptr);
168 }
169 friend class boost::serialization::access;
170
171 BOOST_SERIALIZATION_SPLIT_MEMBER()
172 };
173
174 // currently Service::HID::pad_update_ticks is the smallest interval for an event that gets
175 // always scheduled. Therfore we use this as orientation for the MAX_SLICE_LENGTH
176 // For performance bigger slice length are desired, though this will lead to cores desync
177 // But we never want to schedule events into the current slice, because then cores might to
178 // run small slices to sync up again. This is especially important for events that are always
179 // scheduled and repated.
180 static constexpr int MAX_SLICE_LENGTH = BASE_CLOCK_RATE_ARM11 / 234;
181
182 class Timer {
183 public:
184 Timer();
185 ~Timer();
186
187 s64 GetMaxSliceLength() const;
188
189 void Advance();
190
191 void SetNextSlice(s64 max_slice_length = MAX_SLICE_LENGTH);
192
193 void Idle();
194
195 u64 GetTicks() const;
196 u64 GetIdleTicks() const;
197
198 void AddTicks(u64 ticks);
199
200 s64 GetDowncount() const;
201
202 void ForceExceptionCheck(s64 cycles);
203
204 void MoveEvents();
205
206 private:
207 friend class Timing;
208 // The queue is a min-heap using std::make_heap/push_heap/pop_heap.
209 // We don't use std::priority_queue because we need to be able to serialize, unserialize and
210 // erase arbitrary events (RemoveEvent()) regardless of the queue order. These aren't
211 // accommodated by the standard adaptor class.
212 std::vector<Event> event_queue;
213 u64 event_fifo_id = 0;
214 // the queue for storing the events from other threads threadsafe until they will be added
215 // to the event_queue by the emu thread
216 Common::MPSCQueue<Event> ts_queue;
217 // Are we in a function that has been called from Advance()
218 // If events are sheduled from a function that gets called from Advance(),
219 // don't change slice_length and downcount.
220 // The time between CoreTiming being intialized and the first call to Advance() is
221 // considered the slice boundary between slice -1 and slice 0. Dispatcher loops must call
222 // Advance() before executing the first cycle of each slice to prepare the slice length and
223 // downcount for that slice.
224 bool is_timer_sane = true;
225
226 s64 slice_length = MAX_SLICE_LENGTH;
227 s64 downcount = MAX_SLICE_LENGTH;
228 s64 executed_ticks = 0;
229 u64 idled_cycles = 0;
230 // Stores a scaling for the internal clockspeed. Changing this number results in
231 // under/overclocking the guest cpu
232 double cpu_clock_scale = 1.0;
233
234 template <class Archive>
serialize(Archive & ar,const unsigned int)235 void serialize(Archive& ar, const unsigned int) {
236 MoveEvents();
237 // NOTE: ts_queue should be empty now
238 // TODO(SaveState): Remove the next two lines when we break compatibility
239 s64 x;
240 ar& x; // to keep compatibility with old save states that stored global_timer
241 ar& event_queue;
242 ar& event_fifo_id;
243 ar& slice_length;
244 ar& downcount;
245 ar& executed_ticks;
246 ar& idled_cycles;
247 }
248 friend class boost::serialization::access;
249 };
250
251 explicit Timing(std::size_t num_cores, u32 cpu_clock_percentage);
252
~Timing()253 ~Timing(){};
254
255 /**
256 * Returns the event_type identifier. if name is not unique, it will assert.
257 */
258 TimingEventType* RegisterEvent(const std::string& name, TimedCallback callback);
259
260 void ScheduleEvent(s64 cycles_into_future, const TimingEventType* event_type, u64 userdata = 0,
261 std::size_t core_id = std::numeric_limits<std::size_t>::max());
262
263 void UnscheduleEvent(const TimingEventType* event_type, u64 userdata);
264
265 /// We only permit one event of each type in the queue at a time.
266 void RemoveEvent(const TimingEventType* event_type);
267
268 void SetCurrentTimer(std::size_t core_id);
269
270 s64 GetTicks() const;
271
272 s64 GetGlobalTicks() const;
273
274 /**
275 * Updates the value of the cpu clock scaling to the new percentage.
276 */
277 void UpdateClockSpeed(u32 cpu_clock_percentage);
278
279 std::chrono::microseconds GetGlobalTimeUs() const;
280
281 std::shared_ptr<Timer> GetTimer(std::size_t cpu_id);
282
283 // Used after deserializing to unprotect the event queue.
UnlockEventQueue()284 void UnlockEventQueue() {
285 event_queue_locked = false;
286 }
287
288 private:
289 // unordered_map stores each element separately as a linked list node so pointers to
290 // elements remain stable regardless of rehashes/resizing.
291 std::unordered_map<std::string, TimingEventType> event_types = {};
292
293 std::vector<std::shared_ptr<Timer>> timers;
294 Timer* current_timer = nullptr;
295
296 // Stores a scaling for the internal clockspeed. Changing this number results in
297 // under/overclocking the guest cpu
298 double cpu_clock_scale = 1.0;
299
300 // When true, the event queue can't be modified. Used while deserializing to workaround
301 // destructor side effects.
302 bool event_queue_locked = false;
303
304 template <class Archive>
serialize(Archive & ar,const unsigned int file_version)305 void serialize(Archive& ar, const unsigned int file_version) {
306 // event_types set during initialization of other things
307 ar& timers;
308 if (file_version == 0) {
309 std::shared_ptr<Timer> x;
310 ar& x;
311 current_timer = x.get();
312 } else {
313 ar& current_timer;
314 }
315 if (Archive::is_loading::value) {
316 event_queue_locked = true;
317 }
318 }
319 friend class boost::serialization::access;
320 };
321
322 } // namespace Core
323
324 BOOST_CLASS_VERSION(Core::Timing, 1)
325