graphics: stall waitUntilSafeForRendering in its command processor instead of the GPU thread

CMakeLists.txt

@@ -676,6 +676,8 @@ if(BUILD_TESTING)
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --stream-buffer-only)
 	add_test(NAME gpu_command_lane
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --gpu-command-lane-only)
+	add_test(NAME wait_until_safe_for_rendering
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --wait-until-safe-for-rendering-only)
 	add_test(NAME pm4_context_state
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --context-state-only)
 	add_test(NAME gpu_tiler

src/graphics/guest_gpu/graphicsRun.cpp

@@ -30,6 +30,7 @@
 #include <mutex>
 #include <semaphore>
 #include <thread>
+#include <utility>
 #include <vector>
 
 namespace Libs::Graphics {
@@ -91,6 +92,12 @@ void GuestGpu::SendCommand(Common::UniqueFunction<void>&& command) {
 	m_work_available.Signal();
 }
 
+void GuestGpu::Wake() {
+	Common::LockGuard lock(m_queue_mutex);
+	m_wake_pending = true;
+	m_work_available.Signal();
+}
+
 void GuestGpu::ProcessCommands() {
 	EXIT_IF(!IsGpuThread());
 	while (m_pending_commands.load(std::memory_order_acquire) != 0) {
@@ -464,7 +471,10 @@ void GuestGpu::ThreadRun(void* data) {
 				}
 				if (selected_queue < 0) {
 					gpu->m_processing = false;
-					gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
+					// A wake that came before this wait must not be lost.
+					if (!std::exchange(gpu->m_wake_pending, false)) {
+						gpu->m_work_available.WaitFor(&gpu->m_queue_mutex, 100);
+					}
 					for (auto& queue: gpu->m_queues) {
 						if (!queue.empty()) {
 							queue.front().blocked = false;
@@ -998,10 +1008,11 @@ void CommandProcessor::DrawIndexAuto(DrawAutoArgs args) {
 }
 
 void CommandProcessor::WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index) {
-	BufferFlush();
-
-	m_renderer.GetVideoOut().WaitFlipDone(static_cast<int>(video_out_handle),
-	                                      static_cast<int>(display_buffer_index));
+	// The SDK stalls only this command processor, not the host command lane or other queues.
+	if (!m_renderer.GetVideoOut().IsSafeForRendering(static_cast<int>(video_out_handle),
+	                                                 static_cast<int>(display_buffer_index))) {
+		SuspendPm4();
+	}
 }
 
 template <typename T>

src/graphics/guest_gpu/graphicsRun.h

@@ -31,6 +31,8 @@ public:
 	[[nodiscard]] bool IsStopping();
 	void               SendCommand(Common::UniqueFunction<void>&& command);
 	void               SendCommandSync(Common::UniqueFunction<void>&& command);
+	// Retries blocked submissions now instead of after the 100 ms stall timeout.
+	void               Wake();
 
 	// Submitted command memory is borrowed and must remain valid until GPU execution completes.
 	void Submit(std::span<const uint32_t> draw_commands,
@@ -87,6 +89,7 @@ private:
 	bool                                           m_processing        = false;
 	bool                                           m_accepting         = true;
 	bool                                           m_stopping          = false;
+	bool                                           m_wake_pending      = false;
 	bool                                           m_shutdown_complete = false;
 	// Completion callbacks can outlive GuestGpu during renderer shutdown.
 	std::shared_ptr<std::binary_semaphore> m_suspend_point_ready =

src/graphics/presentation/videoOut.cpp

@@ -221,7 +221,7 @@ public:
 	void WaitForSubmitSlot(VideoOutConfig& cfg);
 	bool Flip(uint32_t micros);
 	void GetFlipStatus(VideoOutConfig& cfg, VideoOutFlipStatus& out);
-	void Wait(VideoOutConfig& cfg, int index);
+	bool IsSafeForRendering(VideoOutConfig& cfg, int index);
 
 private:
 	enum class RequestState { Reserved, Recording, Ready, Presenting };
@@ -240,6 +240,7 @@ private:
 		Graphics::Presenter::Frame* frame;
 		bool                        premultiplied_alpha;
 	};
+	void SignalDoneLocked();
 
 	Graphics::Presenter& m_presenter;
 	Common::Mutex        m_mutex;
@@ -250,6 +251,7 @@ private:
 	std::list<Request>   m_cpu_requests;
 	std::list<Request>   m_cancelled_requests;
 	bool                 m_processing      = false;
+	bool                 m_gpu_waiting     = false;
 	uint64_t             m_next_request_id = 1;
 };
 
@@ -953,7 +955,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
 				}
 			}
 		}
-		m_done_cond_var.SignalAll();
+		SignalDoneLocked();
 	}
 	if (cfg.flip_status.flipPendingNum >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
 		Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRejected);
@@ -1033,7 +1035,7 @@ void FlipQueue::Cancel(VideoOutConfig& cfg) {
 			it = queue->erase(it);
 		}
 	}
-	m_done_cond_var.SignalAll();
+	SignalDoneLocked();
 	m_submit_slot_cond_var.SignalAll();
 	m_submit_cond_var.SignalAll();
 	m_mutex.Unlock();
@@ -1205,19 +1207,26 @@ void FlipQueue::WaitForSubmitSlot(VideoOutConfig& cfg) {
 	}
 }
 
-void FlipQueue::Wait(VideoOutConfig& cfg, int index) {
-	Common::LockGuard lock(m_mutex);
-
-	auto has_request = [this, &cfg, index] {
-		auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
-		return std::any_of(m_requests.begin(), m_requests.end(), matches) ||
-		       std::any_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
-	};
-	while (has_request()) {
-		m_done_cond_var.Wait(&m_mutex);
+void FlipQueue::SignalDoneLocked() {
+	m_done_cond_var.SignalAll();
+	// A stalled GPU cannot drain and shut down without re-checking under m_mutex, so it is alive here.
+	if (m_gpu_waiting) {
+		m_gpu_waiting = false;
+		m_presenter.Renderer().GetGpu().Wake();
 	}
 }
 
+// A display buffer is safe for rendering once it is neither displayed nor waiting to be.
+bool FlipQueue::IsSafeForRendering(VideoOutConfig& cfg, int index) {
+	Common::LockGuard lock(m_mutex);
+	auto matches = [&cfg, index](const auto& r) { return r.cfg == &cfg && r.index == index; };
+	const bool safe = cfg.flip_status.currentBuffer != index &&
+	                  std::none_of(m_requests.begin(), m_requests.end(), matches) &&
+	                  std::none_of(m_cpu_requests.begin(), m_cpu_requests.end(), matches);
+	m_gpu_waiting = !safe;
+	return safe;
+}
+
 bool FlipQueue::Flip(uint32_t micros) {
 	KYTY_PROFILER_BLOCK("FlipQueue::Flip");
 
@@ -1311,7 +1320,7 @@ bool FlipQueue::Flip(uint32_t micros) {
 		it = m_requests.erase(it);
 	}
 	m_processing = false;
-	m_done_cond_var.SignalAll();
+	SignalDoneLocked();
 	m_submit_slot_cond_var.SignalAll();
 	m_mutex.Unlock();
 	for (size_t i = count; i != 0; i--) {
@@ -1694,12 +1703,12 @@ void VideoOutDriver::WaitForSubmitSlot(int handle) {
 	m_impl->GetFlipQueue().WaitForSubmitSlot(*cfg);
 }
 
-void VideoOutDriver::WaitFlipDone(int handle, int index) {
+bool VideoOutDriver::IsSafeForRendering(int handle, int index) {
 	auto* ctx = m_impl->Get(handle);
 	EXIT_IF(ctx == nullptr);
 
 	EXIT_NOT_IMPLEMENTED(!IsValidBufferIndex(index));
-	m_impl->GetFlipQueue().Wait(*ctx, index);
+	return m_impl->GetFlipQueue().IsSafeForRendering(*ctx, index);
 }
 
 KYTY_SYSV_ABI int VideoOutGetFlipStatus(int handle, VideoOutFlipStatus* status) {

src/graphics/presentation/videoOut.h

@@ -38,7 +38,7 @@ public:
 	void CompleteFlip(uint64_t request_id);
 	void SubmitFlipPreparation(uint64_t request_id);
 	void WaitForSubmitSlot(int handle);
-	void WaitFlipDone(int handle, int index);
+	[[nodiscard]] bool IsSafeForRendering(int handle, int index);
 
 	[[nodiscard]] Impl& State() noexcept;
 

tests/ShaderRecompilerComputeTests.cpp

@@ -39,6 +39,7 @@
 #include "graphics/host_gpu/renderer/sync.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 #include <vulkan/vulkan_format_traits.hpp>
+#include "graphics/presentation/videoOut.h"
 #include "graphics/presentation/window/windowInternal.h"
 #include "graphics/shader/recompiler/BufferFormat.h"
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
@@ -49108,6 +49109,94 @@ void CheckPm4IndirectControlFlow(RenderContext &renderer) {
   std::printf("[host]    %-32s ok\n", name);
 }
 
+// waitUntilSafeForRendering stalls only its own command processor (SDK), and a display buffer
+// stays busy while its flip is pending or while it is on screen.
+void CheckWaitUntilSafeForRendering() {
+  constexpr const char *name = "WaitUntilSafeForRendering";
+  Require(name, "video", SDL_InitSubSystem(SDL_INIT_VIDEO), "SDL video init failed");
+  WindowContext window;
+  window.graphic_ctx.screen_width = 64;
+  window.graphic_ctx.screen_height = 64;
+  window.window = SDL_CreateWindow(name, 64, 64, SDL_WINDOW_VULKAN | SDL_WINDOW_HIDDEN);
+  Require(name, "window", window.window != nullptr, "hidden Vulkan window failed");
+  window.CreateVulkan();
+  auto &video_out = VideoOut::VideoOutInit(64, 64, *window.presenter);
+  auto &renderer = *window.render_context;
+  renderer.InitializeGpu(&video_out);
+  auto &gpu = renderer.GetGpu();
+
+  constexpr int user_system = 255;
+  constexpr int bus_main = 0;
+  constexpr int flip_mode_vsync = 1;
+  // Special buffer indices flip without a registered surface.
+  constexpr int index_blank = -1;
+  constexpr int index_black = -2;
+  const int handle = VideoOut::VideoOutOpen(user_system, bus_main, 0, nullptr);
+  Require(name, "open", handle > 0, "video-out open failed");
+  const auto wait_until = [](const auto &done) {
+    const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(2);
+    while (!done() && std::chrono::steady_clock::now() < deadline) {
+      std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    }
+    return done();
+  };
+  const auto address = [](const void *value) { return reinterpret_cast<uint64_t>(value); };
+
+  // Hold the graphics queue so the CPU flip of BLACK, prepared behind it, stays pending.
+  std::atomic_uint32_t gate{0};
+  std::array<uint32_t, 7> hold{};
+  hold[0] = KYTY_PM4(7, Pm4::IT_WAIT_REG_MEM, 0);
+  hold[1] = 0x10u | 3u;
+  hold[2] = static_cast<uint32_t>(address(&gate));
+  hold[3] = static_cast<uint32_t>(address(&gate) >> 32u);
+  hold[4] = 1;
+  hold[5] = UINT32_MAX;
+  gpu.Submit(hold, {});
+  Require(name, "flip black",
+          VideoOut::VideoOutSubmitFlip(handle, index_black, flip_mode_vsync, 0) == 0,
+          "CPU flip failed");
+
+  std::atomic_uint32_t rendered{0};
+  std::array<uint32_t, 12> render{};
+  render[0] = KYTY_PM4(7, Pm4::IT_NOP, Pm4::R_WAIT_FLIP_DONE);
+  render[1] = static_cast<uint32_t>(handle);
+  render[2] = static_cast<uint32_t>(index_black);
+  render[7] = KYTY_PM4(5, Pm4::IT_WRITE_DATA, 0);
+  render[9] = static_cast<uint32_t>(address(&rendered));
+  render[10] = static_cast<uint32_t>(address(&rendered) >> 32u);
+  render[11] = 1;
+  gpu.SubmitCompute(0x20, render);
+
+  std::binary_semaphore lane{0};
+  std::jthread command_lane([&] {
+    gpu.SendCommandSync([] {});
+    lane.release();
+  });
+  Require(name, "command lane", lane.try_acquire_for(std::chrono::seconds(2)),
+          "a busy display buffer blocked the GPU thread");
+  Require(name, "pending", rendered.load() == 0,
+          "rendering started while its display buffer had a pending flip");
+
+  gate.store(1);
+  Require(name, "flip shown", wait_until([handle] { return VideoOut::VideoOutIsFlipPending(handle) == 0; }),
+          "the BLACK flip was not presented");
+  std::this_thread::sleep_for(std::chrono::milliseconds(100));
+  Require(name, "displayed", rendered.load() == 0,
+          "rendering started while its display buffer was on screen");
+
+  Require(name, "flip blank",
+          VideoOut::VideoOutSubmitFlip(handle, index_blank, flip_mode_vsync, 0) == 0,
+          "CPU flip failed");
+  Require(name, "resume", wait_until([&] { return rendered.load() == 1; }),
+          "rendering did not resume once its display buffer left the screen");
+
+  gpu.WaitForIdle();
+  Require(name, "close", VideoOut::VideoOutClose(handle) == 0, "video-out close failed");
+  renderer.ShutdownGpu();
+  VideoOut::VideoOutShutdown();
+  std::printf("[host]    %-32s ok\n", name);
+}
+
 void CheckPm4WaitResume(RenderContext &renderer) {
   GraphicsInitJmpTables();
   CommandProcessor processor(renderer, 0);
@@ -49987,6 +50076,10 @@ int main(int argc, char **argv) {
     vulkan.CheckScalarLeafCall();
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--wait-until-safe-for-rendering-only") == 0) {
+    CheckWaitUntilSafeForRendering();
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--gpu-command-lane-only") == 0) {
     VulkanHarness vulkan;
     vulkan.CheckGpuCommandLane();
@@ -50540,6 +50633,8 @@ int main(int argc, char **argv) {
   vulkan.CheckGpuSuspendPoint();
   vulkan.CheckGpuCommandLane();
   vulkan.CheckScalarLeafCall();
+  // Last: its window installs global GPU resources.
+  CheckWaitUntilSafeForRendering();
   if (skipped_device_checks) {
     std::printf(
         "ShaderRecompilerComputeTests: device rasterization checks skipped, this "