1 files changed, 62 insertions, 93 deletions

diff --git a/src/video_core/gpu_thread.cpp b/src/video_core/gpu_thread.cpp
index c5bdd2a17..cc56cf467 100644
--- a/src/video_core/gpu_thread.cpp
+++ b/src/video_core/gpu_thread.cpp
@@ -4,8 +4,10 @@
 
 #include "common/assert.h"
 #include "common/microprofile.h"
+#include "core/core.h"
+#include "core/core_timing.h"
+#include "core/core_timing_util.h"
 #include "core/frontend/scope_acquire_window_context.h"
-#include "core/settings.h"
 #include "video_core/dma_pusher.h"
 #include "video_core/gpu.h"
 #include "video_core/gpu_thread.h"
@@ -13,38 +15,13 @@
 
 namespace VideoCommon::GPUThread {
 
-/// Executes a single GPU thread command
-static void ExecuteCommand(CommandData* command, VideoCore::RendererBase& renderer,
-                           Tegra::DmaPusher& dma_pusher) {
-    if (const auto submit_list = std::get_if<SubmitListCommand>(command)) {
-        dma_pusher.Push(std::move(submit_list->entries));
-        dma_pusher.DispatchCalls();
-    } else if (const auto data = std::get_if<SwapBuffersCommand>(command)) {
-        renderer.SwapBuffers(data->framebuffer);
-    } else if (const auto data = std::get_if<FlushRegionCommand>(command)) {
-        renderer.Rasterizer().FlushRegion(data->addr, data->size);
-    } else if (const auto data = std::get_if<InvalidateRegionCommand>(command)) {
-        renderer.Rasterizer().InvalidateRegion(data->addr, data->size);
-    } else if (const auto data = std::get_if<FlushAndInvalidateRegionCommand>(command)) {
-        renderer.Rasterizer().FlushAndInvalidateRegion(data->addr, data->size);
-    } else {
-        UNREACHABLE();
-    }
-}
-
 /// Runs the GPU thread
 static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher,
                       SynchState& state) {
-
     MicroProfileOnThreadCreate("GpuThread");
 
-    auto WaitForWakeup = [&]() {
-        std::unique_lock<std::mutex> lock{state.signal_mutex};
-        state.signal_condition.wait(lock, [&] { return !state.is_idle || !state.is_running; });
-    };
-
     // Wait for first GPU command before acquiring the window context
-    WaitForWakeup();
+    state.WaitForCommands();
 
     // If emulation was stopped during disk shader loading, abort before trying to acquire context
     if (!state.is_running) {
@@ -53,99 +30,91 @@ static void RunThread(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_p
 
     Core::Frontend::ScopeAcquireWindowContext acquire_context{renderer.GetRenderWindow()};
 
+    CommandDataContainer next;
     while (state.is_running) {
-        if (!state.is_running) {
-            return;
-        }
-
-        {
-            // Thread has been woken up, so make the previous write queue the next read queue
-            std::lock_guard<std::mutex> lock{state.signal_mutex};
-            std::swap(state.push_queue, state.pop_queue);
+        state.WaitForCommands();
+        while (!state.queue.Empty()) {
+            state.queue.Pop(next);
+            if (const auto submit_list = std::get_if<SubmitListCommand>(&next.data)) {
+                dma_pusher.Push(std::move(submit_list->entries));
+                dma_pusher.DispatchCalls();
+            } else if (const auto data = std::get_if<SwapBuffersCommand>(&next.data)) {
+                renderer.SwapBuffers(std::move(data->framebuffer));
+            } else if (const auto data = std::get_if<FlushRegionCommand>(&next.data)) {
+                renderer.Rasterizer().FlushRegion(data->addr, data->size);
+            } else if (const auto data = std::get_if<InvalidateRegionCommand>(&next.data)) {
+                renderer.Rasterizer().InvalidateRegion(data->addr, data->size);
+            } else if (const auto data = std::get_if<EndProcessingCommand>(&next.data)) {
+                return;
+            } else {
+                UNREACHABLE();
+            }
+            state.signaled_fence = next.fence;
+            state.TrySynchronize();
         }
-
-        // Execute all of the GPU commands
-        while (!state.pop_queue->empty()) {
-            ExecuteCommand(&state.pop_queue->front(), renderer, dma_pusher);
-            state.pop_queue->pop();
-        }
-
-        state.UpdateIdleState();
-
-        // Signal that the GPU thread has finished processing commands
-        if (state.is_idle) {
-            state.idle_condition.notify_one();
-        }
-
-        // Wait for CPU thread to send more GPU commands
-        WaitForWakeup();
     }
 }
 
-ThreadManager::ThreadManager(VideoCore::RendererBase& renderer, Tegra::DmaPusher& dma_pusher)
-    : renderer{renderer}, dma_pusher{dma_pusher}, thread{RunThread, std::ref(renderer),
-                                                         std::ref(dma_pusher), std::ref(state)},
-      thread_id{thread.get_id()} {}
+ThreadManager::ThreadManager(Core::System& system, VideoCore::RendererBase& renderer,
+                             Tegra::DmaPusher& dma_pusher)
+    : system{system}, thread{RunThread, std::ref(renderer), std::ref(dma_pusher), std::ref(state)} {
+    synchronization_event = system.CoreTiming().RegisterEvent(
+        "GPUThreadSynch", [this](u64 fence, s64) { state.WaitForSynchronization(fence); });
+}
 
 ThreadManager::~ThreadManager() {
-    {
-        // Notify GPU thread that a shutdown is pending
-        std::lock_guard<std::mutex> lock{state.signal_mutex};
-        state.is_running = false;
-    }
-
-    state.signal_condition.notify_one();
+    // Notify GPU thread that a shutdown is pending
+    PushCommand(EndProcessingCommand());
     thread.join();
 }
 
 void ThreadManager::SubmitList(Tegra::CommandList&& entries) {
-    if (entries.empty()) {
-        return;
-    }
-
-    PushCommand(SubmitListCommand(std::move(entries)), false, false);
+    const u64 fence{PushCommand(SubmitListCommand(std::move(entries)))};
+    const s64 synchronization_ticks{Core::Timing::usToCycles(9000)};
+    system.CoreTiming().ScheduleEvent(synchronization_ticks, synchronization_event, fence);
 }
 
 void ThreadManager::SwapBuffers(
     std::optional<std::reference_wrapper<const Tegra::FramebufferConfig>> framebuffer) {
-    PushCommand(SwapBuffersCommand(std::move(framebuffer)), true, false);
+    PushCommand(SwapBuffersCommand(std::move(framebuffer)));
 }
 
-void ThreadManager::FlushRegion(VAddr addr, u64 size) {
-    // Block the CPU when using accurate emulation
-    PushCommand(FlushRegionCommand(addr, size), Settings::values.use_accurate_gpu_emulation, false);
+void ThreadManager::FlushRegion(CacheAddr addr, u64 size) {
+    PushCommand(FlushRegionCommand(addr, size));
 }
 
-void ThreadManager::InvalidateRegion(VAddr addr, u64 size) {
-    PushCommand(InvalidateRegionCommand(addr, size), true, true);
+void ThreadManager::InvalidateRegion(CacheAddr addr, u64 size) {
+    if (state.queue.Empty()) {
+        // It's quicker to invalidate a single region on the CPU if the queue is already empty
+        system.Renderer().Rasterizer().InvalidateRegion(addr, size);
+    } else {
+        PushCommand(InvalidateRegionCommand(addr, size));
+    }
 }
 
-void ThreadManager::FlushAndInvalidateRegion(VAddr addr, u64 size) {
+void ThreadManager::FlushAndInvalidateRegion(CacheAddr addr, u64 size) {
+    // Skip flush on asynch mode, as FlushAndInvalidateRegion is not used for anything too important
     InvalidateRegion(addr, size);
 }
 
-void ThreadManager::PushCommand(CommandData&& command_data, bool wait_for_idle, bool allow_on_cpu) {
-    {
-        std::lock_guard<std::mutex> lock{state.signal_mutex};
-
-        if ((allow_on_cpu && state.is_idle) || IsGpuThread()) {
-            // Execute the command synchronously on the current thread
-            ExecuteCommand(&command_data, renderer, dma_pusher);
-            return;
-        }
+u64 ThreadManager::PushCommand(CommandData&& command_data) {
+    const u64 fence{++state.last_fence};
+    state.queue.Push(CommandDataContainer(std::move(command_data), fence));
+    state.SignalCommands();
+    return fence;
+}
 
-        // Push the command to the GPU thread
-        state.UpdateIdleState();
-        state.push_queue->emplace(command_data);
+MICROPROFILE_DEFINE(GPU_wait, "GPU", "Wait for the GPU", MP_RGB(128, 128, 192));
+void SynchState::WaitForSynchronization(u64 fence) {
+    if (signaled_fence >= fence) {
+        return;
     }
 
-    // Signal the GPU thread that commands are pending
-    state.signal_condition.notify_one();
-
-    if (wait_for_idle) {
-        // Wait for the GPU to be idle (all commands to be executed)
-        std::unique_lock<std::mutex> lock{state.idle_mutex};
-        state.idle_condition.wait(lock, [this] { return static_cast<bool>(state.is_idle); });
+    // Wait for the GPU to be idle (all commands to be executed)
+    {
+        MICROPROFILE_SCOPE(GPU_wait);
+        std::unique_lock<std::mutex> lock{synchronization_mutex};
+        synchronization_condition.wait(lock, [this, fence] { return signaled_fence >= fence; });
     }
 }