Merge pull request #17868 from hrydgard/present-wait-thread-fixes

Turn off the present-wait thread

Common/GPU/MiscTypes.h

@@ -21,6 +21,9 @@ typedef std::function<void(InvalidationCallbackFlags)> InvalidationCallback;
 // Also, this might be joined with more non-GPU timing information later.
 struct FrameTimeData {
 	uint64_t frameId;
+
+	int waitCount;
+
 	double frameBegin;
 	double afterFenceWait;
 	double firstSubmit;

Common/GPU/Vulkan/VulkanFrameData.cpp

@@ -113,6 +113,8 @@ VkResult FrameData::QueuePresent(VulkanContext *vulkan, FrameDataShared &shared)
 	// Can't move these into the if.
 	VkPresentIdKHR presentID{ VK_STRUCTURE_TYPE_PRESENT_ID_KHR };
 	VkPresentTimesInfoGOOGLE presentGOOGLE{ VK_STRUCTURE_TYPE_PRESENT_TIMES_INFO_GOOGLE };
+
+	uint64_t frameId = this->frameId;
 	VkPresentTimeGOOGLE presentTimeGOOGLE{ (uint32_t)frameId, 0 };  // it's ok to truncate this. it'll wrap around and work (if we ever reach 4 billion frames..)
 	if (vulkan->Extensions().KHR_present_id && vulkan->GetDeviceFeatures().enabled.presentId.presentId) {
 		presentID.pPresentIds = &frameId;

Common/GPU/Vulkan/VulkanRenderManager.cpp

@@ -25,6 +25,9 @@
 #define UINT64_MAX 0xFFFFFFFFFFFFFFFFULL
 #endif
 
+
+#define USE_PRESENT_WAIT 0
+
 using namespace PPSSPP_VK;
 
 // renderPass is an example of the "compatibility class" or RenderPassType type.
@@ -304,7 +307,7 @@ bool VulkanRenderManager::CreateBackbuffers() {
 		INFO_LOG(G3D, "Starting Vulkan compiler thread");
 		compileThread_ = std::thread(&VulkanRenderManager::CompileThreadFunc, this);
 
-		if (vulkan_->GetDeviceFeatures().enabled.presentWait.presentWait) {
+		if (USE_PRESENT_WAIT && vulkan_->Extensions().KHR_present_wait && vulkan_->GetPresentMode() == VK_PRESENT_MODE_FIFO_KHR) {
 			INFO_LOG(G3D, "Starting Vulkan present wait thread");
 			presentWaitThread_ = std::thread(&VulkanRenderManager::PresentWaitThreadFunc, this);
 		}
@@ -550,7 +553,12 @@ void VulkanRenderManager::PresentWaitThreadFunc() {
 		const uint64_t timeout = 1000000000ULL;  // 1 sec
 		if (VK_SUCCESS == vkWaitForPresentKHR(vulkan_->GetDevice(), vulkan_->GetSwapchain(), waitedId, timeout)) {
 			frameTimeData_[waitedId].actualPresent = time_now_d();
+			frameTimeData_[waitedId].waitCount++;
 			waitedId++;
+		} else {
+			// We caught up somehow, which is a bad sign (we should have blocked, right?). Maybe we should break out of the loop?
+			sleep_ms(1);
+			frameTimeData_[waitedId].waitCount++;
 		}
 		_dbg_assert_(waitedId <= frameIdGen_);
 	}
@@ -583,7 +591,6 @@ void VulkanRenderManager::PollPresentTiming() {
 	}
 }
 
-
 void VulkanRenderManager::BeginFrame(bool enableProfiling, bool enableLogProfiler) {
 	double frameBeginTime = time_now_d()
 	VLOG("BeginFrame");

UI/DebugOverlay.cpp

@@ -114,9 +114,11 @@ static void DrawFrameTiming(UIContext *ctx, const Bounds &bounds) {
 
 	for (int i = 0; i < 8; i++) {
 		FrameTimeData data = ctx->GetDrawContext()->GetFrameTimeData(6 + i);
+		FrameTimeData prevData = ctx->GetDrawContext()->GetFrameTimeData(7 + i);
 		if (data.frameBegin == 0.0) {
 			snprintf(statBuf, sizeof(statBuf), "(No frame time data)");
 		} else {
+			double stride = data.frameBegin - prevData.frameBegin;
 			double fenceLatency_s = data.afterFenceWait - data.frameBegin;
 			double submitLatency_s = data.firstSubmit - data.frameBegin;
 			double queuePresentLatency_s = data.queuePresent - data.frameBegin;
@@ -135,11 +137,14 @@ static void DrawFrameTiming(UIContext *ctx, const Bounds &bounds) {
 					computedMargin * 1000.0);
 			}
 			snprintf(statBuf, sizeof(statBuf),
-				"%llu: From start of frame to event:\n"
+				"* Stride: %0.1f (waits: %d)\n"
+				"%llu: From start:\n"
 				"* Past fence: %0.1f ms\n"
 				"* Submit #1: %0.1f ms\n"
 				"* Queue-p: %0.1f ms\n"
 				"%s",
+				stride * 1000.0,
+				data.waitCount,
 				(long long)data.frameId,
 				fenceLatency_s * 1000.0,
 				submitLatency_s * 1000.0,