video_out: Move presentation to separate thread

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -2131,6 +2131,7 @@ int PS4_SYSV_ABI sceGnmSubmitDone() {
     if (!liverpool->IsGpuIdle()) {
         submission_lock = true;
     }
+    liverpool->SubmitDone();
     send_init_packet = true;
     ++frames_submitted;
     return ORBIS_OK;

src/core/libraries/videoout/driver.cpp

@@ -3,6 +3,7 @@
 
 #include <pthread.h>
 #include "common/assert.h"
+#include "common/debug.h"
 #include "core/libraries/error_codes.h"
 #include "core/libraries/kernel/time_management.h"
 #include "core/libraries/videoout/driver.h"
@@ -41,6 +42,7 @@ VideoOutDriver::VideoOutDriver(u32 width, u32 height) {
     main_port.resolution.fullHeight = height;
     main_port.resolution.paneWidth = width;
     main_port.resolution.paneHeight = height;
+    present_thread = std::jthread([&](std::stop_token token) { PresentThread(token); });
 }
 
 VideoOutDriver::~VideoOutDriver() = default;
@@ -158,27 +160,7 @@ int VideoOutDriver::UnregisterBuffers(VideoOutPort* port, s32 attributeIndex) {
     return ORBIS_OK;
 }
 
-void VideoOutDriver::Flip(std::chrono::microseconds timeout) {
-    Request req;
-    {
-        std::unique_lock lock{mutex};
-        submit_cond.wait_for(lock, timeout, [&] { return !requests.empty(); });
-        if (requests.empty()) {
-            renderer->ShowSplash();
-            return;
-        }
-
-        // Retrieve the request.
-        req = requests.front();
-        requests.pop();
-    }
-
-    // Whatever the game is rendering show splash if it is active
-    if (!renderer->ShowSplash(req.frame)) {
-        // Present the frame.
-        renderer->Present(req.frame);
-    }
-
+void VideoOutDriver::Flip(const Request& req) {
     std::scoped_lock lock{mutex};
 
     // Update flip status.
@@ -256,4 +238,33 @@ void VideoOutDriver::Vblank() {
     }
 }
 
+void VideoOutDriver::PresentThread(std::stop_token token) {
+    static constexpr std::chrono::milliseconds FlipPeriod{16};
+    while (!token.stop_requested()) {
+        Request req;
+        {
+            std::unique_lock lock{mutex};
+            submit_cond.wait_for(lock, FlipPeriod, [&] { return !requests.empty(); });
+            if (requests.empty()) {
+                renderer->ShowSplash();
+                continue;
+            }
+
+            // Retrieve the request.
+            req = requests.front();
+            requests.pop();
+        }
+
+        // Whatever the game is rendering show splash if it is active
+        if (!renderer->ShowSplash(req.frame)) {
+            // Present the frame.
+            renderer->Present(req.frame);
+        }
+
+        Flip(req);
+        Vblank();
+        FRAME_END;
+    }
+}
+
 } // namespace Libraries::VideoOut

src/core/libraries/videoout/driver.h

@@ -6,6 +6,7 @@
 #include <condition_variable>
 #include <mutex>
 #include <queue>
+#include "common/polyfill_thread.h"
 #include "core/libraries/videoout/video_out.h"
 
 namespace Vulkan {
@@ -63,7 +64,6 @@ public:
                         const BufferAttribute* attribute);
     int UnregisterBuffers(VideoOutPort* port, s32 attributeIndex);
 
-    void Flip(std::chrono::microseconds timeout);
     bool SubmitFlip(VideoOutPort* port, s32 index, s64 flip_arg, bool is_eop = false);
 
     void Vblank();
@@ -78,10 +78,14 @@ private:
         bool eop;
     };
 
+    void Flip(const Request& req);
+    void PresentThread(std::stop_token token);
+
     std::mutex mutex;
     VideoOutPort main_port{};
     std::condition_variable_any submit_cond;
     std::condition_variable done_cond;
+    std::jthread present_thread;
     std::queue<Request> requests;
     bool is_neo{};
 };

src/core/libraries/videoout/video_out.cpp

@@ -141,10 +141,12 @@ s32 PS4_SYSV_ABI sceVideoOutSubmitFlip(s32 handle, s32 bufferIndex, s32 flipMode
     LOG_INFO(Lib_VideoOut, "bufferIndex = {}, flipMode = {}, flipArg = {}", bufferIndex, flipMode,
              flipArg);
 
-    if (!driver->SubmitFlip(port, bufferIndex, flipArg)) {
-        LOG_ERROR(Lib_VideoOut, "Flip queue is full");
-        return ORBIS_VIDEO_OUT_ERROR_FLIP_QUEUE_FULL;
-    }
+    // Next time the gpu enters idle state, submit the flip
+    Platform::IrqC::Instance()->RegisterOnce(
+        Platform::InterruptId::GpuIdle, [=](Platform::InterruptId irq) {
+            const auto result = driver->SubmitFlip(port, bufferIndex, flipArg);
+            ASSERT_MSG(result, "Flip submission failed");
+        });
 
     return ORBIS_OK;
 }
@@ -229,14 +231,6 @@ s32 PS4_SYSV_ABI sceVideoOutUnregisterBuffers(s32 handle, s32 attributeIndex) {
     return driver->UnregisterBuffers(port, attributeIndex);
 }
 
-void Flip(std::chrono::microseconds micros) {
-    return driver->Flip(micros);
-}
-
-void Vblank() {
-    return driver->Vblank();
-}
-
 void sceVideoOutGetBufferLabelAddress(s32 handle, uintptr_t* label_addr) {
     auto* port = driver->GetPort(handle);
     ASSERT(port);

src/core/libraries/videoout/video_out.h

@@ -104,9 +104,6 @@ s32 PS4_SYSV_ABI sceVideoOutOpen(SceUserServiceUserId userId, s32 busType, s32 i
                                  const void* param);
 s32 PS4_SYSV_ABI sceVideoOutClose(s32 handle);
 
-void Flip(std::chrono::microseconds micros);
-void Vblank();
-
 // Internal system functions
 void sceVideoOutGetBufferLabelAddress(s32 handle, uintptr_t* label_addr);
 s32 sceVideoOutSubmitEopFlip(s32 handle, u32 buf_id, u32 mode, u32 arg, void** unk);

src/emulator.cpp

@@ -136,14 +136,8 @@ void Emulator::Run(const std::filesystem::path& file) {
     std::jthread mainthread =
         std::jthread([this](std::stop_token stop_token) { linker->Execute(); });
 
-    // Begin main window loop until the application exits
-    static constexpr std::chrono::milliseconds FlipPeriod{16};
-
     while (window->isOpen()) {
         window->waitEvent();
-        Libraries::VideoOut::Flip(FlipPeriod);
-        Libraries::VideoOut::Vblank();
-        FRAME_END;
     }
 
     std::exit(0);

src/sdl_window.cpp

@@ -71,7 +71,7 @@ void WindowSDL::waitEvent() {
     // Called on main thread
     SDL_Event event;
 
-    if (!SDL_PollEvent(&event)) {
+    if (!SDL_WaitEvent(&event)) {
         return;
     }
 

src/video_core/amdgpu/liverpool.cpp

@@ -32,7 +32,7 @@ void Liverpool::Process(std::stop_token stoken) {
     while (!stoken.stop_requested()) {
         {
             std::unique_lock lk{submit_mutex};
-            Common::CondvarWait(submit_cv, lk, stoken, [this] { return num_submits != 0; });
+            Common::CondvarWait(submit_cv, lk, stoken, [this] { return num_submits != 0 || submit_done; });
         }
         if (stoken.stop_requested()) {
             break;
@@ -67,6 +67,13 @@ void Liverpool::Process(std::stop_token stoken) {
             }
         }
 
+        if (submit_done) {
+            if (rasterizer) {
+                rasterizer->Flush();
+            }
+            submit_done = false;
+        }
+
         Platform::IrqC::Instance()->Signal(Platform::InterruptId::GpuIdle);
     }
 }

src/video_core/amdgpu/liverpool.h

@@ -991,6 +991,12 @@ public:
     void SubmitGfx(std::span<const u32> dcb, std::span<const u32> ccb);
     void SubmitAsc(u32 vqid, std::span<const u32> acb);
 
+    void SubmitDone() noexcept {
+        std::scoped_lock lk{submit_mutex};
+        submit_done = true;
+        submit_cv.notify_one();
+    }
+
     bool IsGpuIdle() const {
         return num_submits == 0;
     }
@@ -1061,6 +1067,7 @@ private:
     Vulkan::Rasterizer* rasterizer{};
     std::jthread process_thread{};
     std::atomic<u32> num_submits{};
+    std::atomic<bool> submit_done{};
     std::mutex submit_mutex;
     std::condition_variable_any submit_cv;
 };

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -96,6 +96,10 @@ void Rasterizer::DispatchDirect() {
     cmdbuf.dispatch(cs_program.dim_x, cs_program.dim_y, cs_program.dim_z);
 }
 
+void Rasterizer::Flush() {
+    scheduler.Flush();
+}
+
 void Rasterizer::BeginRendering() {
     const auto& regs = liverpool->regs;
     RenderState state;

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -35,6 +35,8 @@ public:
 
     void ScopeMarkerBegin(const std::string& str);
     void ScopeMarkerEnd();
+    
+    void Flush();
 
 private:
     u32 SetupIndexBuffer(bool& is_indexed, u32 index_offset);