Merge pull request #2965 from FernandoS27/fair-core-timing

Core Timing: Rework Core Timing to run all cores evenly.

src/core/arm/dynarmic/arm_dynarmic.cpp

@@ -116,7 +116,7 @@ public:
         num_interpreted_instructions = 0;
     }
     u64 GetTicksRemaining() override {
-        return std::max(parent.system.CoreTiming().GetDowncount(), 0);
+        return std::max(parent.system.CoreTiming().GetDowncount(), s64{0});
     }
     u64 GetCNTPCT() override {
         return Timing::CpuCyclesToClockCycles(parent.system.CoreTiming().GetTicks());

src/core/arm/unicorn/arm_unicorn.cpp

@@ -156,7 +156,7 @@ void ARM_Unicorn::Run() {
     if (GDBStub::IsServerEnabled()) {
         ExecuteInstructions(std::max(4000000, 0));
     } else {
-        ExecuteInstructions(std::max(system.CoreTiming().GetDowncount(), 0));
+        ExecuteInstructions(std::max(system.CoreTiming().GetDowncount(), s64{0}));
     }
 }
 

src/core/core_cpu.cpp

@@ -85,24 +85,16 @@ void Cpu::RunLoop(bool tight_loop) {
     // instead advance to the next event and try to yield to the next thread
     if (Kernel::GetCurrentThread() == nullptr) {
         LOG_TRACE(Core, "Core-{} idling", core_index);
-
-        if (IsMainCore()) {
-            // TODO(Subv): Only let CoreTiming idle if all 4 cores are idling.
-            core_timing.Idle();
-            core_timing.Advance();
-        }
-
+        core_timing.Idle();
+        core_timing.Advance();
         PrepareReschedule();
     } else {
-        if (IsMainCore()) {
-            core_timing.Advance();
-        }
-
         if (tight_loop) {
             arm_interface->Run();
         } else {
             arm_interface->Step();
         }
+        core_timing.Advance();
     }
 
     Reschedule();

src/core/core_timing.cpp

@@ -15,7 +15,7 @@
 
 namespace Core::Timing {
 
-constexpr int MAX_SLICE_LENGTH = 20000;
+constexpr int MAX_SLICE_LENGTH = 10000;
 
 struct CoreTiming::Event {
     s64 time;
@@ -38,10 +38,12 @@ CoreTiming::CoreTiming() = default;
 CoreTiming::~CoreTiming() = default;
 
 void CoreTiming::Initialize() {
-    downcount = MAX_SLICE_LENGTH;
+    downcounts.fill(MAX_SLICE_LENGTH);
+    time_slice.fill(MAX_SLICE_LENGTH);
     slice_length = MAX_SLICE_LENGTH;
     global_timer = 0;
     idled_cycles = 0;
+    current_context = 0;
 
     // The time between CoreTiming being initialized and the first call to Advance() is considered
     // the slice boundary between slice -1 and slice 0. Dispatcher loops must call Advance() before
@@ -110,7 +112,7 @@ void CoreTiming::UnscheduleEvent(const EventType* event_type, u64 userdata) {
 u64 CoreTiming::GetTicks() const {
     u64 ticks = static_cast<u64>(global_timer);
     if (!is_global_timer_sane) {
-        ticks += slice_length - downcount;
+        ticks += accumulated_ticks;
     }
     return ticks;
 }
@@ -120,7 +122,8 @@ u64 CoreTiming::GetIdleTicks() const {
 }
 
 void CoreTiming::AddTicks(u64 ticks) {
-    downcount -= static_cast<int>(ticks);
+    accumulated_ticks += ticks;
+    downcounts[current_context] -= static_cast<s64>(ticks);
 }
 
 void CoreTiming::ClearPendingEvents() {
@@ -141,22 +144,35 @@ void CoreTiming::RemoveEvent(const EventType* event_type) {
 
 void CoreTiming::ForceExceptionCheck(s64 cycles) {
     cycles = std::max<s64>(0, cycles);
-    if (downcount <= cycles) {
+    if (downcounts[current_context] <= cycles) {
         return;
     }
 
     // downcount is always (much) smaller than MAX_INT so we can safely cast cycles to an int
     // here. Account for cycles already executed by adjusting the g.slice_length
-    slice_length -= downcount - static_cast<int>(cycles);
-    downcount = static_cast<int>(cycles);
+    downcounts[current_context] = static_cast<int>(cycles);
+}
+
+std::optional<u64> CoreTiming::NextAvailableCore(const s64 needed_ticks) const {
+    const u64 original_context = current_context;
+    u64 next_context = (original_context + 1) % num_cpu_cores;
+    while (next_context != original_context) {
+        if (time_slice[next_context] >= needed_ticks) {
+            return {next_context};
+        } else if (time_slice[next_context] >= 0) {
+            return std::nullopt;
+        }
+        next_context = (next_context + 1) % num_cpu_cores;
+    }
+    return std::nullopt;
 }
 
 void CoreTiming::Advance() {
     std::unique_lock<std::mutex> guard(inner_mutex);
 
-    const int cycles_executed = slice_length - downcount;
+    const u64 cycles_executed = accumulated_ticks;
+    time_slice[current_context] = std::max<s64>(0, time_slice[current_context] - accumulated_ticks);
     global_timer += cycles_executed;
-    slice_length = MAX_SLICE_LENGTH;
 
     is_global_timer_sane = true;
 
@@ -173,24 +189,46 @@ void CoreTiming::Advance() {
 
     // Still events left (scheduled in the future)
     if (!event_queue.empty()) {
-        slice_length = static_cast<int>(
-            std::min<s64>(event_queue.front().time - global_timer, MAX_SLICE_LENGTH));
+        const s64 needed_ticks =
+            std::min<s64>(event_queue.front().time - global_timer, MAX_SLICE_LENGTH);
+        const auto next_core = NextAvailableCore(needed_ticks);
+        if (next_core) {
+            downcounts[*next_core] = needed_ticks;
+        }
+    }
+
+    accumulated_ticks = 0;
+
+    downcounts[current_context] = time_slice[current_context];
+}
+
+void CoreTiming::ResetRun() {
+    downcounts.fill(MAX_SLICE_LENGTH);
+    time_slice.fill(MAX_SLICE_LENGTH);
+    current_context = 0;
+    // Still events left (scheduled in the future)
+    if (!event_queue.empty()) {
+        const s64 needed_ticks =
+            std::min<s64>(event_queue.front().time - global_timer, MAX_SLICE_LENGTH);
+        downcounts[current_context] = needed_ticks;
     }
 
-    downcount = slice_length;
+    is_global_timer_sane = false;
+    accumulated_ticks = 0;
 }
 
 void CoreTiming::Idle() {
-    idled_cycles += downcount;
-    downcount = 0;
+    accumulated_ticks += downcounts[current_context];
+    idled_cycles += downcounts[current_context];
+    downcounts[current_context] = 0;
 }
 
 std::chrono::microseconds CoreTiming::GetGlobalTimeUs() const {
     return std::chrono::microseconds{GetTicks() * 1000000 / BASE_CLOCK_RATE};
 }
 
-int CoreTiming::GetDowncount() const {
-    return downcount;
+s64 CoreTiming::GetDowncount() const {
+    return downcounts[current_context];
 }
 
 } // namespace Core::Timing

src/core/core_timing.h

@@ -7,6 +7,7 @@
 #include <chrono>
 #include <functional>
 #include <mutex>
+#include <optional>
 #include <string>
 #include <unordered_map>
 #include <vector>
@@ -104,7 +105,19 @@ public:
 
     std::chrono::microseconds GetGlobalTimeUs() const;
 
-    int GetDowncount() const;
+    void ResetRun();
+
+    s64 GetDowncount() const;
+
+    void SwitchContext(u64 new_context) {
+        current_context = new_context;
+    }
+
+    bool CanCurrentContextRun() const {
+        return time_slice[current_context] > 0;
+    }
+
+    std::optional<u64> NextAvailableCore(const s64 needed_ticks) const;
 
 private:
     struct Event;
@@ -112,10 +125,16 @@ private:
     /// Clear all pending events. This should ONLY be done on exit.
     void ClearPendingEvents();
 
+    static constexpr u64 num_cpu_cores = 4;
+
     s64 global_timer = 0;
     s64 idled_cycles = 0;
-    int slice_length = 0;
-    int downcount = 0;
+    s64 slice_length = 0;
+    u64 accumulated_ticks = 0;
+    std::array<s64, num_cpu_cores> downcounts{};
+    // Slice of time assigned to each core per run.
+    std::array<s64, num_cpu_cores> time_slice{};
+    u64 current_context = 0;
 
     // Are we in a function that has been called from Advance()
     // If events are scheduled from a function that gets called from Advance(),

src/core/cpu_core_manager.cpp

@@ -6,6 +6,7 @@
 #include "core/arm/exclusive_monitor.h"
 #include "core/core.h"
 #include "core/core_cpu.h"
+#include "core/core_timing.h"
 #include "core/cpu_core_manager.h"
 #include "core/gdbstub/gdbstub.h"
 #include "core/settings.h"
@@ -122,13 +123,19 @@ void CpuCoreManager::RunLoop(bool tight_loop) {
         }
     }
 
-    for (active_core = 0; active_core < NUM_CPU_CORES; ++active_core) {
-        cores[active_core]->RunLoop(tight_loop);
-        if (Settings::values.use_multi_core) {
-            // Cores 1-3 are run on other threads in this mode
-            break;
+    auto& core_timing = system.CoreTiming();
+    core_timing.ResetRun();
+    bool keep_running{};
+    do {
+        keep_running = false;
+        for (active_core = 0; active_core < NUM_CPU_CORES; ++active_core) {
+            core_timing.SwitchContext(active_core);
+            if (core_timing.CanCurrentContextRun()) {
+                cores[active_core]->RunLoop(tight_loop);
+            }
+            keep_running |= core_timing.CanCurrentContextRun();
         }
-    }
+    } while (keep_running);
 
     if (GDBStub::IsServerEnabled()) {
         GDBStub::SetCpuStepFlag(false);

src/tests/core/core_timing.cpp

@@ -6,6 +6,7 @@
 
 #include <array>
 #include <bitset>
+#include <cstdlib>
 #include <string>
 #include "common/file_util.h"
 #include "core/core.h"
@@ -13,7 +14,7 @@
 
 // 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 = 10000; // Copied from CoreTiming internals
 
 static std::bitset<CB_IDS.size()> callbacks_ran_flags;
 static u64 expected_callback = 0;
@@ -28,6 +29,12 @@ void CallbackTemplate(u64 userdata, s64 cycles_late) {
     REQUIRE(lateness == cycles_late);
 }
 
+static u64 callbacks_done = 0;
+
+void EmptyCallback(u64 userdata, s64 cycles_late) {
+    ++callbacks_done;
+}
+
 struct ScopeInit final {
     ScopeInit() {
         core_timing.Initialize();
@@ -39,18 +46,19 @@ struct ScopeInit final {
     Core::Timing::CoreTiming core_timing;
 };
 
-static void AdvanceAndCheck(Core::Timing::CoreTiming& core_timing, u32 idx, int downcount,
+static void AdvanceAndCheck(Core::Timing::CoreTiming& core_timing, u32 idx, u32 context = 0,
                             int expected_lateness = 0, int cpu_downcount = 0) {
     callbacks_ran_flags = 0;
     expected_callback = CB_IDS[idx];
     lateness = expected_lateness;
 
     // Pretend we executed X cycles of instructions.
+    core_timing.SwitchContext(context);
     core_timing.AddTicks(core_timing.GetDowncount() - cpu_downcount);
     core_timing.Advance();
+    core_timing.SwitchContext((context + 1) % 4);
 
     REQUIRE(decltype(callbacks_ran_flags)().set(idx) == callbacks_ran_flags);
-    REQUIRE(downcount == core_timing.GetDowncount());
 }
 
 TEST_CASE("CoreTiming[BasicOrder]", "[core]") {
@@ -64,112 +72,61 @@ TEST_CASE("CoreTiming[BasicOrder]", "[core]") {
     Core::Timing::EventType* cb_e = core_timing.RegisterEvent("callbackE", CallbackTemplate<4>);
 
     // Enter slice 0
-    core_timing.Advance();
-
-    // D -> B -> C -> A -> E
-    core_timing.ScheduleEvent(1000, cb_a, CB_IDS[0]);
-    REQUIRE(1000 == core_timing.GetDowncount());
-    core_timing.ScheduleEvent(500, cb_b, CB_IDS[1]);
-    REQUIRE(500 == core_timing.GetDowncount());
-    core_timing.ScheduleEvent(800, cb_c, CB_IDS[2]);
-    REQUIRE(500 == core_timing.GetDowncount());
-    core_timing.ScheduleEvent(100, cb_d, CB_IDS[3]);
-    REQUIRE(100 == core_timing.GetDowncount());
-    core_timing.ScheduleEvent(1200, cb_e, CB_IDS[4]);
-    REQUIRE(100 == core_timing.GetDowncount());
-
-    AdvanceAndCheck(core_timing, 3, 400);
-    AdvanceAndCheck(core_timing, 1, 300);
-    AdvanceAndCheck(core_timing, 2, 200);
-    AdvanceAndCheck(core_timing, 0, 200);
-    AdvanceAndCheck(core_timing, 4, MAX_SLICE_LENGTH);
-}
-
-TEST_CASE("CoreTiming[Threadsave]", "[core]") {
-    ScopeInit guard;
-    auto& core_timing = guard.core_timing;
-
-    Core::Timing::EventType* cb_a = core_timing.RegisterEvent("callbackA", CallbackTemplate<0>);
-    Core::Timing::EventType* cb_b = core_timing.RegisterEvent("callbackB", CallbackTemplate<1>);
-    Core::Timing::EventType* cb_c = core_timing.RegisterEvent("callbackC", CallbackTemplate<2>);
-    Core::Timing::EventType* cb_d = core_timing.RegisterEvent("callbackD", CallbackTemplate<3>);
-    Core::Timing::EventType* cb_e = core_timing.RegisterEvent("callbackE", CallbackTemplate<4>);
-
-    // Enter slice 0
-    core_timing.Advance();
+    core_timing.ResetRun();
 
     // D -> B -> C -> A -> E
+    core_timing.SwitchContext(0);
     core_timing.ScheduleEvent(1000, cb_a, CB_IDS[0]);
-    // Manually force since ScheduleEvent doesn't call it
-    core_timing.ForceExceptionCheck(1000);
     REQUIRE(1000 == core_timing.GetDowncount());
     core_timing.ScheduleEvent(500, cb_b, CB_IDS[1]);
-    // Manually force since ScheduleEvent doesn't call it
-    core_timing.ForceExceptionCheck(500);
     REQUIRE(500 == core_timing.GetDowncount());
     core_timing.ScheduleEvent(800, cb_c, CB_IDS[2]);
-    // Manually force since ScheduleEvent doesn't call it
-    core_timing.ForceExceptionCheck(800);
     REQUIRE(500 == core_timing.GetDowncount());
     core_timing.ScheduleEvent(100, cb_d, CB_IDS[3]);
-    // Manually force since ScheduleEvent doesn't call it
-    core_timing.ForceExceptionCheck(100);
     REQUIRE(100 == core_timing.GetDowncount());
     core_timing.ScheduleEvent(1200, cb_e, CB_IDS[4]);
-    // Manually force since ScheduleEvent doesn't call it
-    core_timing.ForceExceptionCheck(1200);
     REQUIRE(100 == core_timing.GetDowncount());
 
-    AdvanceAndCheck(core_timing, 3, 400);
-    AdvanceAndCheck(core_timing, 1, 300);
-    AdvanceAndCheck(core_timing, 2, 200);
-    AdvanceAndCheck(core_timing, 0, 200);
-    AdvanceAndCheck(core_timing, 4, MAX_SLICE_LENGTH);
-}
-
-namespace SharedSlotTest {
-static unsigned int counter = 0;
-
-template <unsigned int ID>
-void FifoCallback(u64 userdata, s64 cycles_late) {
-    static_assert(ID < CB_IDS.size(), "ID out of range");
-    callbacks_ran_flags.set(ID);
-    REQUIRE(CB_IDS[ID] == userdata);
-    REQUIRE(ID == counter);
-    REQUIRE(lateness == cycles_late);
-    ++counter;
+    AdvanceAndCheck(core_timing, 3, 0);
+    AdvanceAndCheck(core_timing, 1, 1);
+    AdvanceAndCheck(core_timing, 2, 2);
+    AdvanceAndCheck(core_timing, 0, 3);
+    AdvanceAndCheck(core_timing, 4, 0);
 }
-} // namespace SharedSlotTest
 
-TEST_CASE("CoreTiming[SharedSlot]", "[core]") {
-    using namespace SharedSlotTest;
+TEST_CASE("CoreTiming[FairSharing]", "[core]") {
 
     ScopeInit guard;
     auto& core_timing = guard.core_timing;
 
-    Core::Timing::EventType* cb_a = core_timing.RegisterEvent("callbackA", FifoCallback<0>);
-    Core::Timing::EventType* cb_b = core_timing.RegisterEvent("callbackB", FifoCallback<1>);
-    Core::Timing::EventType* cb_c = core_timing.RegisterEvent("callbackC", FifoCallback<2>);
-    Core::Timing::EventType* cb_d = core_timing.RegisterEvent("callbackD", FifoCallback<3>);
-    Core::Timing::EventType* cb_e = core_timing.RegisterEvent("callbackE", FifoCallback<4>);
+    Core::Timing::EventType* empty_callback =
+        core_timing.RegisterEvent("empty_callback", EmptyCallback);
 
-    core_timing.ScheduleEvent(1000, cb_a, CB_IDS[0]);
-    core_timing.ScheduleEvent(1000, cb_b, CB_IDS[1]);
-    core_timing.ScheduleEvent(1000, cb_c, CB_IDS[2]);
-    core_timing.ScheduleEvent(1000, cb_d, CB_IDS[3]);
-    core_timing.ScheduleEvent(1000, cb_e, CB_IDS[4]);
-
-    // Enter slice 0
-    core_timing.Advance();
-    REQUIRE(1000 == core_timing.GetDowncount());
+    callbacks_done = 0;
+    u64 MAX_CALLBACKS = 10;
+    for (std::size_t i = 0; i < 10; i++) {
+        core_timing.ScheduleEvent(i * 3333U, empty_callback, 0);
+    }
 
-    callbacks_ran_flags = 0;
-    counter = 0;
-    lateness = 0;
-    core_timing.AddTicks(core_timing.GetDowncount());
-    core_timing.Advance();
-    REQUIRE(MAX_SLICE_LENGTH == core_timing.GetDowncount());
-    REQUIRE(0x1FULL == callbacks_ran_flags.to_ullong());
+    const s64 advances = MAX_SLICE_LENGTH / 10;
+    core_timing.ResetRun();
+    u64 current_time = core_timing.GetTicks();
+    bool keep_running{};
+    do {
+        keep_running = false;
+        for (u32 active_core = 0; active_core < 4; ++active_core) {
+            core_timing.SwitchContext(active_core);
+            if (core_timing.CanCurrentContextRun()) {
+                core_timing.AddTicks(std::min<s64>(advances, core_timing.GetDowncount()));
+                core_timing.Advance();
+            }
+            keep_running |= core_timing.CanCurrentContextRun();
+        }
+    } while (keep_running);
+    u64 current_time_2 = core_timing.GetTicks();
+
+    REQUIRE(MAX_CALLBACKS == callbacks_done);
+    REQUIRE(current_time_2 == current_time + MAX_SLICE_LENGTH * 4);
 }
 
 TEST_CASE("Core::Timing[PredictableLateness]", "[core]") {
@@ -180,13 +137,13 @@ TEST_CASE("Core::Timing[PredictableLateness]", "[core]") {
     Core::Timing::EventType* cb_b = core_timing.RegisterEvent("callbackB", CallbackTemplate<1>);
 
     // Enter slice 0
-    core_timing.Advance();
+    core_timing.ResetRun();
 
     core_timing.ScheduleEvent(100, cb_a, CB_IDS[0]);
     core_timing.ScheduleEvent(200, cb_b, CB_IDS[1]);
 
-    AdvanceAndCheck(core_timing, 0, 90, 10, -10); // (100 - 10)
-    AdvanceAndCheck(core_timing, 1, MAX_SLICE_LENGTH, 50, -50);
+    AdvanceAndCheck(core_timing, 0, 0, 10, -10); // (100 - 10)
+    AdvanceAndCheck(core_timing, 1, 1, 50, -50);
 }
 
 namespace ChainSchedulingTest {
@@ -220,7 +177,7 @@ TEST_CASE("CoreTiming[ChainScheduling]", "[core]") {
         });
 
     // Enter slice 0
-    core_timing.Advance();
+    core_timing.ResetRun();
 
     core_timing.ScheduleEvent(800, cb_a, CB_IDS[0]);
     core_timing.ScheduleEvent(1000, cb_b, CB_IDS[1]);
@@ -229,19 +186,19 @@ TEST_CASE("CoreTiming[ChainScheduling]", "[core]") {
     REQUIRE(800 == core_timing.GetDowncount());
 
     reschedules = 3;
-    AdvanceAndCheck(core_timing, 0, 200);  // cb_a
-    AdvanceAndCheck(core_timing, 1, 1000); // cb_b, cb_rs
+    AdvanceAndCheck(core_timing, 0, 0); // cb_a
+    AdvanceAndCheck(core_timing, 1, 1); // cb_b, cb_rs
     REQUIRE(2 == reschedules);
 
     core_timing.AddTicks(core_timing.GetDowncount());
     core_timing.Advance(); // cb_rs
+    core_timing.SwitchContext(3);
     REQUIRE(1 == reschedules);
     REQUIRE(200 == core_timing.GetDowncount());
 
-    AdvanceAndCheck(core_timing, 2, 800); // cb_c
+    AdvanceAndCheck(core_timing, 2, 3); // cb_c
 
     core_timing.AddTicks(core_timing.GetDowncount());
     core_timing.Advance(); // cb_rs
     REQUIRE(0 == reschedules);
-    REQUIRE(MAX_SLICE_LENGTH == core_timing.GetDowncount());
 }