From 096366ead51345bcd170e31b6160b14aaf73e996 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 23 Nov 2021 03:29:00 +0100
Subject: [PATCH 01/10] Common: improve native clock.
---
src/common/uint128.h | 5 +++++
src/common/x64/native_clock.cpp | 40 ++++++++++++++++-----------------
src/common/x64/native_clock.h | 13 +++++------
3 files changed, 29 insertions(+), 29 deletions(-)
diff --git a/src/common/uint128.h b/src/common/uint128.h
index f890ffec24..199d0f55e0 100644
--- a/src/common/uint128.h
+++ b/src/common/uint128.h
@@ -30,6 +30,10 @@ namespace Common {
#else
return _udiv128(r[1], r[0], d, &remainder);
#endif
+#else
+#ifdef __SIZEOF_INT128__
+ const auto product = static_cast<unsigned __int128>(a) * static_cast<unsigned __int128>(b);
+ return static_cast<u64>(product / d);
#else
const u64 diva = a / d;
const u64 moda = a % d;
@@ -37,6 +41,7 @@ namespace Common {
const u64 modb = b % d;
return diva * b + moda * divb + moda * modb / d;
#endif
+#endif
}
// This function multiplies 2 u64 values and produces a u128 value;
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index 1b71945037..427a382cdf 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -5,7 +5,6 @@
#include <chrono>
#include <thread>
-#include "common/atomic_ops.h"
#include "common/uint128.h"
#include "common/x64/native_clock.h"
@@ -65,8 +64,10 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
u64 rtsc_frequency_)
: WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
rtsc_frequency_} {
- time_point.inner.last_measure = FencedRDTSC();
- time_point.inner.accumulated_ticks = 0U;
+ TimePoint new_time_point{};
+ new_time_point.last_measure = FencedRDTSC();
+ new_time_point.accumulated_ticks = 0U;
+ time_point.store(new_time_point);
ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency);
us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency);
ms_rtsc_factor = GetFixedPoint64Factor(MS_RATIO, rtsc_frequency);
@@ -76,34 +77,31 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
u64 NativeClock::GetRTSC() {
TimePoint new_time_point{};
- TimePoint current_time_point{};
-
- current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
+ TimePoint current_time_point = time_point.load(std::memory_order_acquire);
do {
const u64 current_measure = FencedRDTSC();
- u64 diff = current_measure - current_time_point.inner.last_measure;
+ u64 diff = current_measure - current_time_point.last_measure;
diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
- new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure
- ? current_measure
- : current_time_point.inner.last_measure;
- new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
- } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
- current_time_point.pack, current_time_point.pack));
+ new_time_point.last_measure = current_measure > current_time_point.last_measure
+ ? current_measure
+ : current_time_point.last_measure;
+ new_time_point.accumulated_ticks = current_time_point.accumulated_ticks + diff;
+ } while (!time_point.compare_exchange_weak(
+ current_time_point, new_time_point, std::memory_order_release, std::memory_order_acquire));
/// The clock cannot be more precise than the guest timer, remove the lower bits
- return new_time_point.inner.accumulated_ticks & inaccuracy_mask;
+ return new_time_point.accumulated_ticks & inaccuracy_mask;
}
void NativeClock::Pause(bool is_paused) {
if (!is_paused) {
- TimePoint current_time_point{};
TimePoint new_time_point{};
-
- current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
+ TimePoint current_time_point = time_point.load(std::memory_order_acquire);
do {
- new_time_point.pack = current_time_point.pack;
- new_time_point.inner.last_measure = FencedRDTSC();
- } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
- current_time_point.pack, current_time_point.pack));
+ new_time_point = current_time_point;
+ new_time_point.last_measure = FencedRDTSC();
+ } while (!time_point.compare_exchange_weak(current_time_point, new_time_point,
+ std::memory_order_release,
+ std::memory_order_acquire));
}
}
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index 30d2ba2e91..e57446cb99 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -3,6 +3,7 @@
#pragma once
+#include <atomic>
#include "common/wall_clock.h"
namespace Common {
@@ -28,13 +29,9 @@ public:
private:
u64 GetRTSC();
- union alignas(16) TimePoint {
- TimePoint() : pack{} {}
- u128 pack{};
- struct Inner {
- u64 last_measure{};
- u64 accumulated_ticks{};
- } inner;
+ struct alignas(16) TimePoint {
+ u64 last_measure{};
+ u64 accumulated_ticks{};
};
/// value used to reduce the native clocks accuracy as some apss rely on
@@ -42,7 +39,7 @@ private:
/// be higher.
static constexpr u64 inaccuracy_mask = ~(UINT64_C(0x400) - 1);
- TimePoint time_point;
+ std::atomic<TimePoint> time_point;
// factors
u64 clock_rtsc_factor{};
u64 cpu_rtsc_factor{};
From 846c994cc9ff3b53d0d3fa3cb3b8fe0418c462c6 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sat, 27 Nov 2021 16:26:48 +0100
Subject: [PATCH 02/10] Core: Reimplement Core Timing.
---
src/core/core_timing.cpp | 130 +++++++++++++++++++++------------
src/core/core_timing.h | 21 +++---
src/tests/core/core_timing.cpp | 1 -
3 files changed, 95 insertions(+), 57 deletions(-)
diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp
index 29e7dba9b1..9185029290 100644
--- a/src/core/core_timing.cpp
+++ b/src/core/core_timing.cpp
@@ -7,6 +7,7 @@
#include <tuple>
#include "common/microprofile.h"
+#include "common/thread.h"
#include "core/core_timing.h"
#include "core/core_timing_util.h"
#include "core/hardware_properties.h"
@@ -59,68 +60,96 @@ void CoreTiming::Initialize(std::function<void()>&& on_thread_init_) {
const auto empty_timed_callback = [](std::uintptr_t, std::chrono::nanoseconds) {};
ev_lost = CreateEvent("_lost_event", empty_timed_callback);
if (is_multicore) {
- timer_thread = std::make_unique<std::thread>(ThreadEntry, std::ref(*this));
+ const auto hardware_concurrency = std::thread::hardware_concurrency();
+ worker_threads.emplace_back(ThreadEntry, std::ref(*this));
+ if (hardware_concurrency > 8) {
+ worker_threads.emplace_back(ThreadEntry, std::ref(*this));
+ }
}
}
void CoreTiming::Shutdown() {
- paused = true;
+ is_paused = true;
shutting_down = true;
- pause_event.Set();
- event.Set();
- if (timer_thread) {
- timer_thread->join();
+ {
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ event_cv.notify_all();
+ wait_pause_cv.notify_all();
}
+ for (auto& thread : worker_threads) {
+ thread.join();
+ }
+ worker_threads.clear();
ClearPendingEvents();
- timer_thread.reset();
has_started = false;
}
-void CoreTiming::Pause(bool is_paused) {
- paused = is_paused;
- pause_event.Set();
-}
-
-void CoreTiming::SyncPause(bool is_paused) {
- if (is_paused == paused && paused_set == paused) {
+void CoreTiming::Pause(bool is_paused_) {
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ if (is_paused_ == paused_state.load(std::memory_order_relaxed)) {
return;
}
- Pause(is_paused);
- if (timer_thread) {
- if (!is_paused) {
- pause_event.Set();
+ if (is_multicore) {
+ is_paused = is_paused_;
+ event_cv.notify_all();
+ if (!is_paused_) {
+ wait_pause_cv.notify_all();
+ }
+ }
+ paused_state.store(is_paused_, std::memory_order_relaxed);
+}
+
+void CoreTiming::SyncPause(bool is_paused_) {
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ if (is_paused_ == paused_state.load(std::memory_order_relaxed)) {
+ return;
+ }
+
+ if (is_multicore) {
+ is_paused = is_paused_;
+ event_cv.notify_all();
+ if (!is_paused_) {
+ wait_pause_cv.notify_all();
+ }
+ }
+ paused_state.store(is_paused_, std::memory_order_relaxed);
+ if (is_multicore) {
+ if (is_paused_) {
+ wait_signal_cv.wait(main_lock, [this] { return pause_count == worker_threads.size(); });
+ } else {
+ wait_signal_cv.wait(main_lock, [this] { return pause_count == 0; });
}
- event.Set();
- while (paused_set != is_paused)
- ;
}
}
bool CoreTiming::IsRunning() const {
- return !paused_set;
+ return !paused_state.load(std::memory_order_acquire);
}
bool CoreTiming::HasPendingEvents() const {
- return !(wait_set && event_queue.empty());
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ return !event_queue.empty();
}
void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
const std::shared_ptr<EventType>& event_type,
std::uintptr_t user_data) {
- {
- std::scoped_lock scope{basic_lock};
- const u64 timeout = static_cast<u64>((GetGlobalTimeNs() + ns_into_future).count());
- event_queue.emplace_back(Event{timeout, event_fifo_id++, user_data, event_type});
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ const u64 timeout = static_cast<u64>((GetGlobalTimeNs() + ns_into_future).count());
- std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+ event_queue.emplace_back(Event{timeout, event_fifo_id++, user_data, event_type});
+
+ std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+
+ if (is_multicore) {
+ event_cv.notify_one();
}
- event.Set();
}
void CoreTiming::UnscheduleEvent(const std::shared_ptr<EventType>& event_type,
std::uintptr_t user_data) {
- std::scoped_lock scope{basic_lock};
+ std::unique_lock<std::mutex> main_lock(event_mutex);
const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
return e.type.lock().get() == event_type.get() && e.user_data == user_data;
});
@@ -168,11 +197,12 @@ u64 CoreTiming::GetClockTicks() const {
}
void CoreTiming::ClearPendingEvents() {
+ std::unique_lock<std::mutex> main_lock(event_mutex);
event_queue.clear();
}
void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
- std::scoped_lock lock{basic_lock};
+ std::unique_lock<std::mutex> main_lock(event_mutex);
const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
return e.type.lock().get() == event_type.get();
@@ -186,21 +216,21 @@ void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
}
std::optional<s64> CoreTiming::Advance() {
- std::scoped_lock lock{advance_lock, basic_lock};
global_timer = GetGlobalTimeNs().count();
+ std::unique_lock<std::mutex> main_lock(event_mutex);
while (!event_queue.empty() && event_queue.front().time <= global_timer) {
Event evt = std::move(event_queue.front());
std::pop_heap(event_queue.begin(), event_queue.end(), std::greater<>());
event_queue.pop_back();
- basic_lock.unlock();
+ event_mutex.unlock();
if (const auto event_type{evt.type.lock()}) {
- event_type->callback(
- evt.user_data, std::chrono::nanoseconds{static_cast<s64>(global_timer - evt.time)});
+ event_type->callback(evt.user_data, std::chrono::nanoseconds{static_cast<s64>(
+ GetGlobalTimeNs().count() - evt.time)});
}
- basic_lock.lock();
+ event_mutex.lock();
global_timer = GetGlobalTimeNs().count();
}
@@ -213,26 +243,34 @@ std::optional<s64> CoreTiming::Advance() {
}
void CoreTiming::ThreadLoop() {
+ const auto predicate = [this] { return !event_queue.empty() || is_paused; };
has_started = true;
while (!shutting_down) {
- while (!paused) {
- paused_set = false;
+ while (!is_paused && !shutting_down) {
const auto next_time = Advance();
if (next_time) {
if (*next_time > 0) {
std::chrono::nanoseconds next_time_ns = std::chrono::nanoseconds(*next_time);
- event.WaitFor(next_time_ns);
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ event_cv.wait_for(main_lock, next_time_ns, predicate);
}
} else {
- wait_set = true;
- event.Wait();
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ event_cv.wait(main_lock, predicate);
}
- wait_set = false;
}
- paused_set = true;
- clock->Pause(true);
- pause_event.Wait();
- clock->Pause(false);
+ std::unique_lock<std::mutex> main_lock(event_mutex);
+ pause_count++;
+ if (pause_count == worker_threads.size()) {
+ clock->Pause(true);
+ wait_signal_cv.notify_all();
+ }
+ wait_pause_cv.wait(main_lock, [this] { return !is_paused || shutting_down; });
+ pause_count--;
+ if (pause_count == 0) {
+ clock->Pause(false);
+ wait_signal_cv.notify_all();
+ }
}
}
diff --git a/src/core/core_timing.h b/src/core/core_timing.h
index d277730096..5c9ee29029 100644
--- a/src/core/core_timing.h
+++ b/src/core/core_timing.h
@@ -14,7 +14,6 @@
#include <vector>
#include "common/common_types.h"
-#include "common/thread.h"
#include "common/wall_clock.h"
namespace Core::Timing {
@@ -146,19 +145,21 @@ private:
u64 event_fifo_id = 0;
std::shared_ptr<EventType> ev_lost;
- Common::Event event{};
- Common::Event pause_event{};
- std::mutex basic_lock;
- std::mutex advance_lock;
- std::unique_ptr<std::thread> timer_thread;
- std::atomic<bool> paused{};
- std::atomic<bool> paused_set{};
- std::atomic<bool> wait_set{};
- std::atomic<bool> shutting_down{};
std::atomic<bool> has_started{};
std::function<void()> on_thread_init{};
+ std::vector<std::thread> worker_threads;
+
+ std::condition_variable event_cv;
+ std::condition_variable wait_pause_cv;
+ std::condition_variable wait_signal_cv;
+ mutable std::mutex event_mutex;
+
+ std::atomic<bool> paused_state{};
+ bool is_paused{};
+ bool shutting_down{};
bool is_multicore{};
+ size_t pause_count{};
/// Cycle timing
u64 ticks{};
diff --git a/src/tests/core/core_timing.cpp b/src/tests/core/core_timing.cpp
index 8358d36b50..62eb437538 100644
--- a/src/tests/core/core_timing.cpp
+++ b/src/tests/core/core_timing.cpp
@@ -27,7 +27,6 @@ void HostCallbackTemplate(std::uintptr_t user_data, std::chrono::nanoseconds ns_
static_assert(IDX < CB_IDS.size(), "IDX out of range");
callbacks_ran_flags.set(IDX);
REQUIRE(CB_IDS[IDX] == user_data);
- REQUIRE(CB_IDS[IDX] == CB_IDS[calls_order[expected_callback]]);
delays[IDX] = ns_late.count();
++expected_callback;
}
From a2d29412cbda3e0dc57c49c5d4c098e8ba73cbb5 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sat, 27 Nov 2021 20:31:46 +0100
Subject: [PATCH 03/10] Core/Common: Corrections to core timing and add
critical priority.
---
src/common/thread.cpp | 13 +++++++++----
src/common/thread.h | 1 +
src/core/core_timing.cpp | 2 +-
3 files changed, 11 insertions(+), 5 deletions(-)
diff --git a/src/common/thread.cpp b/src/common/thread.cpp
index f932a72909..924f0df1b3 100644
--- a/src/common/thread.cpp
+++ b/src/common/thread.cpp
@@ -47,6 +47,9 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
case ThreadPriority::VeryHigh:
windows_priority = THREAD_PRIORITY_HIGHEST;
break;
+ case ThreadPriority::Critical:
+ windows_priority = THREAD_PRIORITY_TIME_CRITICAL;
+ break;
default:
windows_priority = THREAD_PRIORITY_NORMAL;
break;
@@ -59,9 +62,11 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
void SetCurrentThreadPriority(ThreadPriority new_priority) {
pthread_t this_thread = pthread_self();
- s32 max_prio = sched_get_priority_max(SCHED_OTHER);
- s32 min_prio = sched_get_priority_min(SCHED_OTHER);
- u32 level = static_cast<u32>(new_priority) + 1;
+ const auto scheduling_type =
+ new_priority != ThreadPriority::Critical ? SCHED_OTHER : SCHED_FIFO;
+ s32 max_prio = sched_get_priority_max(scheduling_type);
+ s32 min_prio = sched_get_priority_min(scheduling_type);
+ u32 level = std::max(static_cast<u32>(new_priority) + 1, 4U);
struct sched_param params;
if (max_prio > min_prio) {
@@ -70,7 +75,7 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
params.sched_priority = min_prio - ((min_prio - max_prio) * level) / 4;
}
- pthread_setschedparam(this_thread, SCHED_OTHER, ¶ms);
+ pthread_setschedparam(this_thread, scheduling_type, ¶ms);
}
#endif
diff --git a/src/common/thread.h b/src/common/thread.h
index a631225162..1552f58e0f 100644
--- a/src/common/thread.h
+++ b/src/common/thread.h
@@ -92,6 +92,7 @@ enum class ThreadPriority : u32 {
Normal = 1,
High = 2,
VeryHigh = 3,
+ Critical = 4,
};
void SetCurrentThreadPriority(ThreadPriority new_priority);
diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp
index 9185029290..b6c295ada6 100644
--- a/src/core/core_timing.cpp
+++ b/src/core/core_timing.cpp
@@ -46,7 +46,7 @@ void CoreTiming::ThreadEntry(CoreTiming& instance) {
constexpr char name[] = "yuzu:HostTiming";
MicroProfileOnThreadCreate(name);
Common::SetCurrentThreadName(name);
- Common::SetCurrentThreadPriority(Common::ThreadPriority::VeryHigh);
+ Common::SetCurrentThreadPriority(Common::ThreadPriority::Critical);
instance.on_thread_init();
instance.ThreadLoop();
MicroProfileOnThreadExit();
From 00b09de3d9578b29271b33df1b98a37449e7373f Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 28 Nov 2021 11:28:29 +0100
Subject: [PATCH 04/10] Core: add missing include.
---
src/core/core_timing.h | 1 +
1 file changed, 1 insertion(+)
diff --git a/src/core/core_timing.h b/src/core/core_timing.h
index 5c9ee29029..901bf532ed 100644
--- a/src/core/core_timing.h
+++ b/src/core/core_timing.h
@@ -5,6 +5,7 @@
#include <atomic>
#include <chrono>
+#include <condition_variable>
#include <functional>
#include <memory>
#include <mutex>
From 9cafb0d91266210dab2c72e484b493bceae1cb02 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 28 Nov 2021 12:21:45 +0100
Subject: [PATCH 05/10] Core: Fix tests.
---
src/common/thread.cpp | 3 +--
src/common/x64/native_clock.cpp | 1 +
src/tests/core/core_timing.cpp | 3 +++
3 files changed, 5 insertions(+), 2 deletions(-)
diff --git a/src/common/thread.cpp b/src/common/thread.cpp
index 924f0df1b3..919e33af92 100644
--- a/src/common/thread.cpp
+++ b/src/common/thread.cpp
@@ -62,8 +62,7 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
void SetCurrentThreadPriority(ThreadPriority new_priority) {
pthread_t this_thread = pthread_self();
- const auto scheduling_type =
- new_priority != ThreadPriority::Critical ? SCHED_OTHER : SCHED_FIFO;
+ const auto scheduling_type = SCHED_OTHER;
s32 max_prio = sched_get_priority_max(scheduling_type);
s32 min_prio = sched_get_priority_min(scheduling_type);
u32 level = std::max(static_cast<u32>(new_priority) + 1, 4U);
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index 427a382cdf..0b89f9ed2e 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -5,6 +5,7 @@
#include <chrono>
#include <thread>
+#include "common/atomic_ops.h"
#include "common/uint128.h"
#include "common/x64/native_clock.h"
diff --git a/src/tests/core/core_timing.cpp b/src/tests/core/core_timing.cpp
index 62eb437538..e687416a81 100644
--- a/src/tests/core/core_timing.cpp
+++ b/src/tests/core/core_timing.cpp
@@ -8,6 +8,7 @@
#include <chrono>
#include <cstdlib>
#include <memory>
+#include <mutex>
#include <string>
#include "core/core.h"
@@ -21,9 +22,11 @@ std::array<s64, 5> delays{};
std::bitset<CB_IDS.size()> callbacks_ran_flags;
u64 expected_callback = 0;
+std::mutex control_mutex;
template <unsigned int IDX>
void HostCallbackTemplate(std::uintptr_t user_data, std::chrono::nanoseconds ns_late) {
+ std::unique_lock<std::mutex> lk(control_mutex);
static_assert(IDX < CB_IDS.size(), "IDX out of range");
callbacks_ran_flags.set(IDX);
REQUIRE(CB_IDS[IDX] == user_data);
From 38e4a144a1e6f399482eb586c1e0d5646fae9679 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Sun, 28 Nov 2021 13:47:40 +0100
Subject: [PATCH 06/10] Core: Protect each event from race conditions within
it.
---
src/core/core_timing.cpp | 1 +
src/core/core_timing.h | 1 +
2 files changed, 2 insertions(+)
diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp
index b6c295ada6..18dfa07f51 100644
--- a/src/core/core_timing.cpp
+++ b/src/core/core_timing.cpp
@@ -226,6 +226,7 @@ std::optional<s64> CoreTiming::Advance() {
event_mutex.unlock();
if (const auto event_type{evt.type.lock()}) {
+ std::unique_lock<std::mutex> lk(event_type->guard);
event_type->callback(evt.user_data, std::chrono::nanoseconds{static_cast<s64>(
GetGlobalTimeNs().count() - evt.time)});
}
diff --git a/src/core/core_timing.h b/src/core/core_timing.h
index 901bf532ed..4fef6fcce1 100644
--- a/src/core/core_timing.h
+++ b/src/core/core_timing.h
@@ -32,6 +32,7 @@ struct EventType {
TimedCallback callback;
/// A pointer to the name of the event.
const std::string name;
+ mutable std::mutex guard;
};
/**
From 86ccce3721a02338865be74e145255c8a4cb6b4e Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 28 Jun 2022 01:19:30 +0200
Subject: [PATCH 07/10] Address feedback.
---
src/core/core_timing.cpp | 26 +++++++++++++-------------
1 file changed, 13 insertions(+), 13 deletions(-)
diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp
index 18dfa07f51..ac117161c0 100644
--- a/src/core/core_timing.cpp
+++ b/src/core/core_timing.cpp
@@ -72,7 +72,7 @@ void CoreTiming::Shutdown() {
is_paused = true;
shutting_down = true;
{
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
event_cv.notify_all();
wait_pause_cv.notify_all();
}
@@ -85,7 +85,7 @@ void CoreTiming::Shutdown() {
}
void CoreTiming::Pause(bool is_paused_) {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
if (is_paused_ == paused_state.load(std::memory_order_relaxed)) {
return;
}
@@ -100,7 +100,7 @@ void CoreTiming::Pause(bool is_paused_) {
}
void CoreTiming::SyncPause(bool is_paused_) {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
if (is_paused_ == paused_state.load(std::memory_order_relaxed)) {
return;
}
@@ -127,7 +127,7 @@ bool CoreTiming::IsRunning() const {
}
bool CoreTiming::HasPendingEvents() const {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
return !event_queue.empty();
}
@@ -135,7 +135,7 @@ void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
const std::shared_ptr<EventType>& event_type,
std::uintptr_t user_data) {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
const u64 timeout = static_cast<u64>((GetGlobalTimeNs() + ns_into_future).count());
event_queue.emplace_back(Event{timeout, event_fifo_id++, user_data, event_type});
@@ -149,7 +149,7 @@ void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
void CoreTiming::UnscheduleEvent(const std::shared_ptr<EventType>& event_type,
std::uintptr_t user_data) {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
return e.type.lock().get() == event_type.get() && e.user_data == user_data;
});
@@ -197,12 +197,12 @@ u64 CoreTiming::GetClockTicks() const {
}
void CoreTiming::ClearPendingEvents() {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
event_queue.clear();
}
void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
return e.type.lock().get() == event_type.get();
@@ -218,7 +218,7 @@ void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
std::optional<s64> CoreTiming::Advance() {
global_timer = GetGlobalTimeNs().count();
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
while (!event_queue.empty() && event_queue.front().time <= global_timer) {
Event evt = std::move(event_queue.front());
std::pop_heap(event_queue.begin(), event_queue.end(), std::greater<>());
@@ -226,7 +226,7 @@ std::optional<s64> CoreTiming::Advance() {
event_mutex.unlock();
if (const auto event_type{evt.type.lock()}) {
- std::unique_lock<std::mutex> lk(event_type->guard);
+ std::unique_lock lk(event_type->guard);
event_type->callback(evt.user_data, std::chrono::nanoseconds{static_cast<s64>(
GetGlobalTimeNs().count() - evt.time)});
}
@@ -252,15 +252,15 @@ void CoreTiming::ThreadLoop() {
if (next_time) {
if (*next_time > 0) {
std::chrono::nanoseconds next_time_ns = std::chrono::nanoseconds(*next_time);
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
event_cv.wait_for(main_lock, next_time_ns, predicate);
}
} else {
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
event_cv.wait(main_lock, predicate);
}
}
- std::unique_lock<std::mutex> main_lock(event_mutex);
+ std::unique_lock main_lock(event_mutex);
pause_count++;
if (pause_count == worker_threads.size()) {
clock->Pause(true);
From f5c1d7b8c8895b5d6b99685313be9061c8ed8a82 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 28 Jun 2022 01:47:00 +0200
Subject: [PATCH 08/10] Native Clock: remove inaccuracy mask.
---
src/common/x64/native_clock.cpp | 2 +-
src/common/x64/native_clock.h | 5 -----
2 files changed, 1 insertion(+), 6 deletions(-)
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index 0b89f9ed2e..488c8c905c 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -90,7 +90,7 @@ u64 NativeClock::GetRTSC() {
} while (!time_point.compare_exchange_weak(
current_time_point, new_time_point, std::memory_order_release, std::memory_order_acquire));
/// The clock cannot be more precise than the guest timer, remove the lower bits
- return new_time_point.accumulated_ticks & inaccuracy_mask;
+ return new_time_point.accumulated_ticks;
}
void NativeClock::Pause(bool is_paused) {
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index e57446cb99..046cea0952 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -34,11 +34,6 @@ private:
u64 accumulated_ticks{};
};
- /// value used to reduce the native clocks accuracy as some apss rely on
- /// undefined behavior where the level of accuracy in the clock shouldn't
- /// be higher.
- static constexpr u64 inaccuracy_mask = ~(UINT64_C(0x400) - 1);
-
std::atomic<TimePoint> time_point;
// factors
u64 clock_rtsc_factor{};
From 2575a93dc6d15bb4c60c18be1635b48f37355059 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Tue, 28 Jun 2022 22:42:00 +0200
Subject: [PATCH 09/10] Native clock: Use atomic ops as before.
---
src/common/x64/native_clock.cpp | 39 +++++++++++++++++----------------
src/common/x64/native_clock.h | 14 +++++++-----
2 files changed, 29 insertions(+), 24 deletions(-)
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index 488c8c905c..c0d38cf6be 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -65,10 +65,8 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
u64 rtsc_frequency_)
: WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
rtsc_frequency_} {
- TimePoint new_time_point{};
- new_time_point.last_measure = FencedRDTSC();
- new_time_point.accumulated_ticks = 0U;
- time_point.store(new_time_point);
+ time_point.inner.last_measure = FencedRDTSC();
+ time_point.inner.accumulated_ticks = 0U;
ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency);
us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency);
ms_rtsc_factor = GetFixedPoint64Factor(MS_RATIO, rtsc_frequency);
@@ -77,32 +75,35 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
}
u64 NativeClock::GetRTSC() {
+ TimePoint current_time_point{};
TimePoint new_time_point{};
- TimePoint current_time_point = time_point.load(std::memory_order_acquire);
+
+ current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
do {
const u64 current_measure = FencedRDTSC();
- u64 diff = current_measure - current_time_point.last_measure;
+ u64 diff = current_measure - current_time_point.inner.last_measure;
diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
- new_time_point.last_measure = current_measure > current_time_point.last_measure
- ? current_measure
- : current_time_point.last_measure;
- new_time_point.accumulated_ticks = current_time_point.accumulated_ticks + diff;
- } while (!time_point.compare_exchange_weak(
- current_time_point, new_time_point, std::memory_order_release, std::memory_order_acquire));
+ new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure
+ ? current_measure
+ : current_time_point.inner.last_measure;
+ new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
+ } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
+ current_time_point.pack, current_time_point.pack));
/// The clock cannot be more precise than the guest timer, remove the lower bits
- return new_time_point.accumulated_ticks;
+ return new_time_point.inner.accumulated_ticks;
}
void NativeClock::Pause(bool is_paused) {
if (!is_paused) {
+ TimePoint current_time_point{};
TimePoint new_time_point{};
- TimePoint current_time_point = time_point.load(std::memory_order_acquire);
+
+ current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
do {
- new_time_point = current_time_point;
- new_time_point.last_measure = FencedRDTSC();
- } while (!time_point.compare_exchange_weak(current_time_point, new_time_point,
- std::memory_order_release,
- std::memory_order_acquire));
+ new_time_point.pack = current_time_point.pack;
+ new_time_point.inner.last_measure = FencedRDTSC();
+ } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
+ current_time_point.pack, current_time_point.pack));
}
}
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index 046cea0952..38ae7a4625 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -3,7 +3,6 @@
#pragma once
-#include <atomic>
#include "common/wall_clock.h"
namespace Common {
@@ -29,12 +28,17 @@ public:
private:
u64 GetRTSC();
- struct alignas(16) TimePoint {
- u64 last_measure{};
- u64 accumulated_ticks{};
+ union alignas(16) TimePoint {
+ TimePoint() : pack{} {}
+ u128 pack{};
+ struct Inner {
+ u64 last_measure{};
+ u64 accumulated_ticks{};
+ } inner;
};
- std::atomic<TimePoint> time_point;
+ TimePoint time_point;
+
// factors
u64 clock_rtsc_factor{};
u64 cpu_rtsc_factor{};
From 3196d957b02266293b68a60c75c3db9a00faf1f6 Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Wed, 29 Jun 2022 01:29:24 +0200
Subject: [PATCH 10/10] Adress Feedback.
---
src/common/x64/native_clock.cpp | 1 -
src/core/core_timing.cpp | 43 ++++++++++++++++++++-------------
src/core/core_timing.h | 4 ++-
3 files changed, 29 insertions(+), 19 deletions(-)
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index c0d38cf6be..6aaa8cdf99 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -89,7 +89,6 @@ u64 NativeClock::GetRTSC() {
new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
} while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
current_time_point.pack, current_time_point.pack));
- /// The clock cannot be more precise than the guest timer, remove the lower bits
return new_time_point.inner.accumulated_ticks;
}
diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp
index ac117161c0..1405780695 100644
--- a/src/core/core_timing.cpp
+++ b/src/core/core_timing.cpp
@@ -6,6 +6,7 @@
#include <string>
#include <tuple>
+#include "common/logging/log.h"
#include "common/microprofile.h"
#include "common/thread.h"
#include "core/core_timing.h"
@@ -42,10 +43,10 @@ CoreTiming::CoreTiming()
CoreTiming::~CoreTiming() = default;
-void CoreTiming::ThreadEntry(CoreTiming& instance) {
- constexpr char name[] = "yuzu:HostTiming";
- MicroProfileOnThreadCreate(name);
- Common::SetCurrentThreadName(name);
+void CoreTiming::ThreadEntry(CoreTiming& instance, size_t id) {
+ const std::string name = "yuzu:HostTiming_" + std::to_string(id);
+ MicroProfileOnThreadCreate(name.c_str());
+ Common::SetCurrentThreadName(name.c_str());
Common::SetCurrentThreadPriority(Common::ThreadPriority::Critical);
instance.on_thread_init();
instance.ThreadLoop();
@@ -61,9 +62,10 @@ void CoreTiming::Initialize(std::function<void()>&& on_thread_init_) {
ev_lost = CreateEvent("_lost_event", empty_timed_callback);
if (is_multicore) {
const auto hardware_concurrency = std::thread::hardware_concurrency();
- worker_threads.emplace_back(ThreadEntry, std::ref(*this));
+ size_t id = 0;
+ worker_threads.emplace_back(ThreadEntry, std::ref(*this), id++);
if (hardware_concurrency > 8) {
- worker_threads.emplace_back(ThreadEntry, std::ref(*this));
+ worker_threads.emplace_back(ThreadEntry, std::ref(*this), id++);
}
}
}
@@ -71,11 +73,10 @@ void CoreTiming::Initialize(std::function<void()>&& on_thread_init_) {
void CoreTiming::Shutdown() {
is_paused = true;
shutting_down = true;
- {
- std::unique_lock main_lock(event_mutex);
- event_cv.notify_all();
- wait_pause_cv.notify_all();
- }
+ std::atomic_thread_fence(std::memory_order_release);
+
+ event_cv.notify_all();
+ wait_pause_cv.notify_all();
for (auto& thread : worker_threads) {
thread.join();
}
@@ -128,7 +129,7 @@ bool CoreTiming::IsRunning() const {
bool CoreTiming::HasPendingEvents() const {
std::unique_lock main_lock(event_mutex);
- return !event_queue.empty();
+ return !event_queue.empty() || pending_events.load(std::memory_order_relaxed) != 0;
}
void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
@@ -139,6 +140,7 @@ void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
const u64 timeout = static_cast<u64>((GetGlobalTimeNs() + ns_into_future).count());
event_queue.emplace_back(Event{timeout, event_fifo_id++, user_data, event_type});
+ pending_events.fetch_add(1, std::memory_order_relaxed);
std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
@@ -158,6 +160,7 @@ void CoreTiming::UnscheduleEvent(const std::shared_ptr<EventType>& event_type,
if (itr != event_queue.end()) {
event_queue.erase(itr, event_queue.end());
std::make_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+ pending_events.fetch_sub(1, std::memory_order_relaxed);
}
}
@@ -223,15 +226,21 @@ std::optional<s64> CoreTiming::Advance() {
Event evt = std::move(event_queue.front());
std::pop_heap(event_queue.begin(), event_queue.end(), std::greater<>());
event_queue.pop_back();
- event_mutex.unlock();
if (const auto event_type{evt.type.lock()}) {
- std::unique_lock lk(event_type->guard);
- event_type->callback(evt.user_data, std::chrono::nanoseconds{static_cast<s64>(
- GetGlobalTimeNs().count() - evt.time)});
+ sequence_mutex.lock();
+ event_mutex.unlock();
+
+ event_type->guard.lock();
+ sequence_mutex.unlock();
+ const s64 delay = static_cast<s64>(GetGlobalTimeNs().count() - evt.time);
+ event_type->callback(evt.user_data, std::chrono::nanoseconds{delay});
+ event_type->guard.unlock();
+
+ event_mutex.lock();
+ pending_events.fetch_sub(1, std::memory_order_relaxed);
}
- event_mutex.lock();
global_timer = GetGlobalTimeNs().count();
}
diff --git a/src/core/core_timing.h b/src/core/core_timing.h
index 4fef6fcce1..a86553e08b 100644
--- a/src/core/core_timing.h
+++ b/src/core/core_timing.h
@@ -132,7 +132,7 @@ private:
/// Clear all pending events. This should ONLY be done on exit.
void ClearPendingEvents();
- static void ThreadEntry(CoreTiming& instance);
+ static void ThreadEntry(CoreTiming& instance, size_t id);
void ThreadLoop();
std::unique_ptr<Common::WallClock> clock;
@@ -145,6 +145,7 @@ private:
// accomodated by the standard adaptor class.
std::vector<Event> event_queue;
u64 event_fifo_id = 0;
+ std::atomic<size_t> pending_events{};
std::shared_ptr<EventType> ev_lost;
std::atomic<bool> has_started{};
@@ -156,6 +157,7 @@ private:
std::condition_variable wait_pause_cv;
std::condition_variable wait_signal_cv;
mutable std::mutex event_mutex;
+ mutable std::mutex sequence_mutex;
std::atomic<bool> paused_state{};
bool is_paused{};