Improve core timing accuracy (#5257)

* Improve core timing accuracy

* remove wrong global_ticks, use biggest ticks over all cores for GetGlobalTicks

* merge max slice length change
This commit is contained in:
Ben 2020-05-12 22:48:30 +02:00 committed by GitHub
parent d11d600b61
commit 57aa18f52e
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
5 changed files with 62 additions and 28 deletions

View file

@ -148,7 +148,11 @@ System::ResultStatus System::RunLoop(bool tight_loop) {
for (auto& cpu_core : cpu_cores) {
if (cpu_core->GetTimer().GetTicks() < global_ticks) {
s64 delay = global_ticks - cpu_core->GetTimer().GetTicks();
cpu_core->GetTimer().Advance(delay);
kernel->SetRunningCPU(cpu_core.get());
cpu_core->GetTimer().Advance();
cpu_core->PrepareReschedule();
kernel->GetThreadManager(cpu_core->GetID()).Reschedule();
cpu_core->GetTimer().SetNextSlice(delay);
if (max_delay < delay) {
max_delay = delay;
current_core_to_execute = cpu_core.get();
@ -156,10 +160,14 @@ System::ResultStatus System::RunLoop(bool tight_loop) {
}
}
if (max_delay > 0) {
// jit sometimes overshoot by a few ticks which might lead to a minimal desync in the cores.
// This small difference shouldn't make it necessary to sync the cores and would only cost
// performance. Thus we don't sync delays below min_delay
static constexpr s64 min_delay = 100;
if (max_delay > min_delay) {
LOG_TRACE(Core_ARM11, "Core {} running (delayed) for {} ticks",
current_core_to_execute->GetID(),
current_core_to_execute->GetTimer().GetDowncount());
current_core_to_execute->GetID(),
current_core_to_execute->GetTimer().GetDowncount());
if (running_core != current_core_to_execute) {
running_core = current_core_to_execute;
kernel->SetRunningCPU(running_core);
@ -181,12 +189,15 @@ System::ResultStatus System::RunLoop(bool tight_loop) {
// TODO: Make special check for idle since we can easily revert the time of idle cores
s64 max_slice = Timing::MAX_SLICE_LENGTH;
for (const auto& cpu_core : cpu_cores) {
kernel->SetRunningCPU(cpu_core.get());
cpu_core->GetTimer().Advance();
cpu_core->PrepareReschedule();
kernel->GetThreadManager(cpu_core->GetID()).Reschedule();
max_slice = std::min(max_slice, cpu_core->GetTimer().GetMaxSliceLength());
}
for (auto& cpu_core : cpu_cores) {
cpu_core->GetTimer().Advance(max_slice);
}
for (auto& cpu_core : cpu_cores) {
cpu_core->GetTimer().SetNextSlice(max_slice);
auto start_ticks = cpu_core->GetTimer().GetTicks();
LOG_TRACE(Core_ARM11, "Core {} running for {} ticks", cpu_core->GetID(),
cpu_core->GetTimer().GetDowncount());
running_core = cpu_core.get();
@ -204,8 +215,8 @@ System::ResultStatus System::RunLoop(bool tight_loop) {
cpu_core->Step();
}
}
max_slice = cpu_core->GetTimer().GetTicks() - start_ticks;
}
timing->AddToGlobalTicks(max_slice);
}
if (GDBStub::IsServerEnabled()) {

View file

@ -111,11 +111,15 @@ s64 Timing::GetTicks() const {
}
s64 Timing::GetGlobalTicks() const {
return global_timer;
const auto& timer =
std::max_element(timers.cbegin(), timers.cend(), [](const auto& a, const auto& b) {
return a->GetTicks() < b->GetTicks();
});
return (*timer)->GetTicks();
}
std::chrono::microseconds Timing::GetGlobalTimeUs() const {
return std::chrono::microseconds{GetTicks() * 1000000 / BASE_CLOCK_RATE_ARM11};
return std::chrono::microseconds{GetGlobalTicks() * 1000000 / BASE_CLOCK_RATE_ARM11};
}
std::shared_ptr<Timing::Timer> Timing::GetTimer(std::size_t cpu_id) {
@ -161,21 +165,22 @@ void Timing::Timer::MoveEvents() {
}
s64 Timing::Timer::GetMaxSliceLength() const {
auto next_event = std::find_if(event_queue.begin(), event_queue.end(),
[&](const Event& e) { return e.time - executed_ticks > 0; });
const auto& next_event = event_queue.begin();
if (next_event != event_queue.end()) {
ASSERT(next_event->time - executed_ticks > 0);
return next_event->time - executed_ticks;
}
return MAX_SLICE_LENGTH;
}
void Timing::Timer::Advance(s64 max_slice_length) {
void Timing::Timer::Advance() {
MoveEvents();
s64 cycles_executed = slice_length - downcount;
idled_cycles = 0;
executed_ticks += cycles_executed;
slice_length = max_slice_length;
slice_length = 0;
downcount = 0;
is_timer_sane = true;
@ -191,6 +196,10 @@ void Timing::Timer::Advance(s64 max_slice_length) {
}
is_timer_sane = false;
}
void Timing::Timer::SetNextSlice(s64 max_slice_length) {
slice_length = max_slice_length;
// Still events left (scheduled in the future)
if (!event_queue.empty()) {

View file

@ -171,7 +171,13 @@ public:
BOOST_SERIALIZATION_SPLIT_MEMBER()
};
static constexpr int MAX_SLICE_LENGTH = 20000;
// currently Service::HID::pad_update_ticks is the smallest interval for an event that gets
// always scheduled. Therfore we use this as orientation for the MAX_SLICE_LENGTH
// For performance bigger slice length are desired, though this will lead to cores desync
// But we never want to schedule events into the current slice, because then cores might to
// run small slices to sync up again. This is especially important for events that are always
// scheduled and repated.
static constexpr int MAX_SLICE_LENGTH = BASE_CLOCK_RATE_ARM11 / 234;
class Timer {
public:
@ -180,7 +186,9 @@ public:
s64 GetMaxSliceLength() const;
void Advance(s64 max_slice_length = MAX_SLICE_LENGTH);
void Advance();
void SetNextSlice(s64 max_slice_length = MAX_SLICE_LENGTH);
void Idle();
@ -227,6 +235,9 @@ public:
void serialize(Archive& ar, const unsigned int) {
MoveEvents();
// NOTE: ts_queue should be empty now
// TODO(SaveState): Remove the next two lines when we break compatibility
s64 x;
ar& x; // to keep compatibility with old save states that stored global_timer
ar& event_queue;
ar& event_fifo_id;
ar& slice_length;
@ -260,10 +271,6 @@ public:
s64 GetGlobalTicks() const;
void AddToGlobalTicks(s64 ticks) {
global_timer += ticks;
}
/**
* Updates the value of the cpu clock scaling to the new percentage.
*/
@ -274,8 +281,6 @@ public:
std::shared_ptr<Timer> GetTimer(std::size_t cpu_id);
private:
s64 global_timer = 0;
// unordered_map stores each element separately as a linked list node so pointers to
// elements remain stable regardless of rehashes/resizing.
std::unordered_map<std::string, TimingEventType> event_types = {};
@ -290,7 +295,6 @@ private:
template <class Archive>
void serialize(Archive& ar, const unsigned int file_version) {
// event_types set during initialization of other things
ar& global_timer;
ar& timers;
if (file_version == 0) {
std::shared_ptr<Timer> x;

View file

@ -111,7 +111,7 @@ void ThreadManager::SwitchContext(Thread* new_thread) {
// Save context for previous thread
if (previous_thread) {
previous_process = previous_thread->owner_process;
previous_thread->last_running_ticks = timing.GetGlobalTicks();
previous_thread->last_running_ticks = cpu->GetTimer().GetTicks();
cpu->SaveContext(previous_thread->context);
if (previous_thread->status == ThreadStatus::Running) {
@ -344,7 +344,7 @@ ResultVal<std::shared_ptr<Thread>> KernelSystem::CreateThread(
thread->entry_point = entry_point;
thread->stack_top = stack_top;
thread->nominal_priority = thread->current_priority = priority;
thread->last_running_ticks = timing.GetGlobalTicks();
thread->last_running_ticks = timing.GetTimer(processor_id)->GetTicks();
thread->processor_id = processor_id;
thread->wait_objects.clear();
thread->wait_address = 0;

View file

@ -13,7 +13,8 @@
// Numbers are chosen randomly to make sure the correct one is given.
static constexpr std::array<u64, 5> CB_IDS{{42, 144, 93, 1026, UINT64_C(0xFFFF7FFFF7FFFF)}};
static constexpr int MAX_SLICE_LENGTH = 20000; // Copied from CoreTiming internals
static constexpr int MAX_SLICE_LENGTH =
BASE_CLOCK_RATE_ARM11 / 234; // Copied from CoreTiming internals
static std::bitset<CB_IDS.size()> callbacks_ran_flags;
static u64 expected_callback = 0;
@ -36,7 +37,9 @@ static void AdvanceAndCheck(Core::Timing& timing, u32 idx, int downcount, int ex
timing.GetTimer(0)->AddTicks(timing.GetTimer(0)->GetDowncount() -
cpu_downcount); // Pretend we executed X cycles of instructions.
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice();
REQUIRE(decltype(callbacks_ran_flags)().set(idx) == callbacks_ran_flags);
REQUIRE(downcount == timing.GetTimer(0)->GetDowncount());
@ -53,6 +56,7 @@ TEST_CASE("CoreTiming[BasicOrder]", "[core]") {
// Enter slice 0
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice();
// D -> B -> C -> A -> E
timing.ScheduleEvent(1000, cb_a, CB_IDS[0], 0);
@ -106,6 +110,7 @@ TEST_CASE("CoreTiming[SharedSlot]", "[core]") {
// Enter slice 0
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice();
REQUIRE(1000 == timing.GetTimer(0)->GetDowncount());
callbacks_ran_flags = 0;
@ -113,6 +118,7 @@ TEST_CASE("CoreTiming[SharedSlot]", "[core]") {
lateness = 0;
timing.GetTimer(0)->AddTicks(timing.GetTimer(0)->GetDowncount());
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice();
REQUIRE(MAX_SLICE_LENGTH == timing.GetTimer(0)->GetDowncount());
REQUIRE(0x1FULL == callbacks_ran_flags.to_ullong());
}
@ -125,6 +131,7 @@ TEST_CASE("CoreTiming[PredictableLateness]", "[core]") {
// Enter slice 0
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice();
timing.ScheduleEvent(100, cb_a, CB_IDS[0], 0);
timing.ScheduleEvent(200, cb_b, CB_IDS[1], 0);
@ -161,6 +168,7 @@ TEST_CASE("CoreTiming[ChainScheduling]", "[core]") {
// Enter slice 0
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice();
timing.ScheduleEvent(800, cb_a, CB_IDS[0], 0);
timing.ScheduleEvent(1000, cb_b, CB_IDS[1], 0);
@ -174,14 +182,16 @@ TEST_CASE("CoreTiming[ChainScheduling]", "[core]") {
REQUIRE(2 == reschedules);
timing.GetTimer(0)->AddTicks(timing.GetTimer(0)->GetDowncount());
timing.GetTimer(0)->Advance(); // cb_rs
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice(); // cb_rs
REQUIRE(1 == reschedules);
REQUIRE(200 == timing.GetTimer(0)->GetDowncount());
AdvanceAndCheck(timing, 2, 800); // cb_c
timing.GetTimer(0)->AddTicks(timing.GetTimer(0)->GetDowncount());
timing.GetTimer(0)->Advance(); // cb_rs
timing.GetTimer(0)->Advance();
timing.GetTimer(0)->SetNextSlice(); // cb_rs
REQUIRE(0 == reschedules);
REQUIRE(MAX_SLICE_LENGTH == timing.GetTimer(0)->GetDowncount());
}