6 files changed, 89 insertions, 38 deletions

diff --git a/src/core/arm/dynarmic/arm_dynarmic.cpp b/src/core/arm/dynarmic/arm_dynarmic.cpp
index f1506b372..4d2e99ed0 100644
--- a/src/core/arm/dynarmic/arm_dynarmic.cpp
+++ b/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<s64>(parent.system.CoreTiming().GetDowncount(), 0LL);
     }
     u64 GetCNTPCT() override {
         return Timing::CpuCyclesToClockCycles(parent.system.CoreTiming().GetTicks());
diff --git a/src/core/arm/unicorn/arm_unicorn.cpp b/src/core/arm/unicorn/arm_unicorn.cpp
index 97d5c2a8a..3f91b06d4 100644
--- a/src/core/arm/unicorn/arm_unicorn.cpp
+++ b/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<s64>(system.CoreTiming().GetDowncount(), 0LL));
     }
 }
 
diff --git a/src/core/core_cpu.cpp b/src/core/core_cpu.cpp
index 21c410e34..6bd9639c6 100644
--- a/src/core/core_cpu.cpp
+++ b/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();
diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp
index a58f7b131..6da2dcfb4 100644
--- a/src/core/core_timing.cpp
+++ b/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 +38,14 @@ CoreTiming::CoreTiming() = default;
 CoreTiming::~CoreTiming() = default;
 
 void CoreTiming::Initialize() {
-    downcount = MAX_SLICE_LENGTH;
+    for (std::size_t core = 0; core < num_cpu_cores; core++) {
+        downcounts[core] = MAX_SLICE_LENGTH;
+        time_slice[core] = 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 +114,7 @@ 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 += time_slice[current_context] - downcounts[current_context];
     }
     return ticks;
 }
@@ -120,7 +124,7 @@ u64 CoreTiming::GetIdleTicks() const {
 }
 
 void CoreTiming::AddTicks(u64 ticks) {
-    downcount -= static_cast<int>(ticks);
+    downcounts[current_context] -= static_cast<s64>(ticks);
 }
 
 void CoreTiming::ClearPendingEvents() {
@@ -141,22 +145,36 @@ 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);
+    slice_length -= downcounts[current_context] - 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 {};
+        }
+        next_context = (next_context + 1) % num_cpu_cores;
+    }
+    return {};
 }
 
 void CoreTiming::Advance() {
     std::unique_lock<std::mutex> guard(inner_mutex);
 
-    const int cycles_executed = slice_length - downcount;
+    const int cycles_executed = time_slice[current_context] - downcounts[current_context];
+    time_slice[current_context] = std::max<s64>(0, downcounts[current_context]);
     global_timer += cycles_executed;
-    slice_length = MAX_SLICE_LENGTH;
 
     is_global_timer_sane = true;
 
@@ -173,24 +191,40 @@ 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));
+        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;
+        }
     }
 
-    downcount = slice_length;
+    downcounts[current_context] = time_slice[current_context];
+}
+
+void CoreTiming::ResetRun() {
+    for (std::size_t core = 0; core < num_cpu_cores; core++) {
+        downcounts[core] = MAX_SLICE_LENGTH;
+        time_slice[core] = MAX_SLICE_LENGTH;
+    }
+    current_context = 0;
+    // Still events left (scheduled in the future)
+    if (!event_queue.empty()) {
+        s64 needed_ticks = std::min<s64>(event_queue.front().time - global_timer, MAX_SLICE_LENGTH);
+        downcounts[current_context] = needed_ticks;
+    }
 }
 
 void CoreTiming::Idle() {
-    idled_cycles += downcount;
-    downcount = 0;
+    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
diff --git a/src/core/core_timing.h b/src/core/core_timing.h
index 161c7007d..ec0a6d2c0 100644
--- a/src/core/core_timing.h
+++ b/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 CurrentContextCanRun() const {
+        return time_slice[current_context] > 0;
+    }
+
+    std::optional<u64> NextAvailableCore(const s64 needed_ticks) const;
 
 private:
     struct Event;
@@ -112,10 +125,15 @@ 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;
+    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(),
diff --git a/src/core/cpu_core_manager.cpp b/src/core/cpu_core_manager.cpp
index 8fcb4eeb1..e022e6a60 100644
--- a/src/core/cpu_core_manager.cpp
+++ b/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.CurrentContextCanRun()) {
+                cores[active_core]->RunLoop(tight_loop);
+            }
+            keep_running |= core_timing.CurrentContextCanRun();
         }
-    }
+    } while (keep_running);
 
     if (GDBStub::IsServerEnabled()) {
         GDBStub::SetCpuStepFlag(false);