From 863326b5cdad7dd0a7775cc6944f4d5d2babe5b2 Mon Sep 17 00:00:00 2001
From: Asuka <6732524+Inori@users.noreply.github.com>
Date: Tue, 5 Apr 2022 07:54:28 +0800
Subject: [PATCH] fix presenter bug

caused by forget to update frame index
add some comments on submission and presentation loop
disable GPCS4_NO_GRAPHICS again
---
 GPCS4/GPCS4Common.h                  |  2 +-
 GPCS4/GPCS4Config.h                  |  2 +-
 GPCS4/Graphics/Gnm/GnmDataFormat.cpp |  2 +-
 GPCS4/Graphics/Sce/SceGnmDriver.cpp  |  8 ++++++++
 GPCS4/Graphics/Sce/ScePresenter.cpp  |  8 +++++++-
 GPCS4/Graphics/Violet/VltQueue.cpp   | 15 ++++++++++++++-
 GPCS4/Graphics/VirtualGPU.h          |  4 ++--
 7 files changed, 34 insertions(+), 7 deletions(-)

diff --git a/GPCS4/GPCS4Common.h b/GPCS4/GPCS4Common.h
index 9356188f..de553830 100644
--- a/GPCS4/GPCS4Common.h
+++ b/GPCS4/GPCS4Common.h
@@ -4,6 +4,6 @@
 #include "GPCS4Types.h"
 #include "GPCS4Log.h"
 #include "GPCS4Decoration.h"
-#include "IntelliSenseClang.h"
+#include "IntellisenseClang.h"
 #include "UtilMath.h"
 
diff --git a/GPCS4/GPCS4Config.h b/GPCS4/GPCS4Config.h
index 99cb3f65..8751395c 100644
--- a/GPCS4/GPCS4Config.h
+++ b/GPCS4/GPCS4Config.h
@@ -24,5 +24,5 @@
 // Graphics switch
 // Define this will turn off graphics output,
 // useful when developing non-graphics parts of GPCS4
-#define GPCS4_NO_GRAPHICS
+// #define GPCS4_NO_GRAPHICS
 
diff --git a/GPCS4/Graphics/Gnm/GnmDataFormat.cpp b/GPCS4/Graphics/Gnm/GnmDataFormat.cpp
index 9946352b..ef851de4 100644
--- a/GPCS4/Graphics/Gnm/GnmDataFormat.cpp
+++ b/GPCS4/Graphics/Gnm/GnmDataFormat.cpp
@@ -625,7 +625,7 @@ bool DataFormat::getRenderTargetChannelOrder(RenderTargetChannelOrder* outOrder)
 		}
 		goto LABEL_42;
 	}
-	v11 = 288231505728184512LL;
+	v11 = 0x4000107000120C0LL;
 	v12 = v5 == 4 && v6 == 5;
 	v13 = v12 && v8 == 6;
 	if (_bittest64(&v11, v3))
diff --git a/GPCS4/Graphics/Sce/SceGnmDriver.cpp b/GPCS4/Graphics/Sce/SceGnmDriver.cpp
index 9e292f76..17c147f2 100644
--- a/GPCS4/Graphics/Sce/SceGnmDriver.cpp
+++ b/GPCS4/Graphics/Sce/SceGnmDriver.cpp
@@ -173,6 +173,9 @@ namespace sce
 
 		LOG_ASSERT(count == 1, "Currently only support 1 cmdbuff at one call.");
 
+		// track currently display buffer
+		// so that we can find it during command buffer recording
+		// and use it as render target.
 		trackSwapImage(displayBufferIndex);
 
 		SceGpuCommand cmd = {};
@@ -182,6 +185,7 @@ namespace sce
 
 		submitPresent(cmdList, displayBufferIndex);
 
+		// clear resource tracker every frame
 		resetResourceTracker();
 
 		return SCE_OK;
@@ -222,6 +226,10 @@ namespace sce
 		// Currently I didn't find a very good place, so I place it here.
 		
 		glfwPollEvents();
+
+		// TODO:
+		// Execute the Gnm::DrawCommandBuffer::InitializeDefaultHardwareState command.
+
 		return SCE_OK;
 	}
 
diff --git a/GPCS4/Graphics/Sce/ScePresenter.cpp b/GPCS4/Graphics/Sce/ScePresenter.cpp
index e02cf10e..59ea96e5 100644
--- a/GPCS4/Graphics/Sce/ScePresenter.cpp
+++ b/GPCS4/Graphics/Sce/ScePresenter.cpp
@@ -64,7 +64,13 @@ namespace sce
 		info.pSwapchains        = &m_swapchain;
 		info.pImageIndices      = &m_imageIndex;
 		info.pResults           = nullptr;
-		return vkQueuePresentKHR(m_device.queue, &info);
+
+		VkResult status = vkQueuePresentKHR(m_device.queue, &info);
+
+		m_frameIndex += 1;
+		m_frameIndex %= m_semaphores.size();
+
+		return status;
 	}
 
 	VkResult ScePresenter::recreateSwapChain(const PresenterDesc& desc)
diff --git a/GPCS4/Graphics/Violet/VltQueue.cpp b/GPCS4/Graphics/Violet/VltQueue.cpp
index fafa3f77..d59164ae 100644
--- a/GPCS4/Graphics/Violet/VltQueue.cpp
+++ b/GPCS4/Graphics/Violet/VltQueue.cpp
@@ -18,6 +18,8 @@ namespace sce::vlt
 	{
 		auto& cmdList = submission.cmdList;
 		cmdList->submit(submission.waitSync, submission.wakeSync);
+
+		// simulate async
 		m_submitQueue.push(std::move(cmdList));
 	}
 
@@ -28,10 +30,21 @@ namespace sce::vlt
 		auto& presenter = presentInfo.presenter;
 		presenter->presentImage();
 
-
+		// simulate async
 		auto cmdList = m_submitQueue.front();
 		m_submitQueue.pop();
 
+		// TODO:
+		// Calling synchronize will block the CPU waiting for the submission done on GPU,
+		// thus will block the submit thread, slow down the submit speed and drop FPS.
+		// DXVK's solution is to use another thread dedicated for cmdlist submission 
+		// thus overcome the performance problem,
+		// see https://github.com/doitsujin/dxvk/blob/master/src/dxvk/dxvk_queue.cpp
+		// 
+		// This can be easily implemented, but for now, our bottleneck is not here obviously.
+		// So I keep it single threaded to make debugging easier.
+
+
 		// Wait for command buffer submit finish.
 		cmdList->synchronize();
 
diff --git a/GPCS4/Graphics/VirtualGPU.h b/GPCS4/Graphics/VirtualGPU.h
index 92823e1d..26288d52 100644
--- a/GPCS4/Graphics/VirtualGPU.h
+++ b/GPCS4/Graphics/VirtualGPU.h
@@ -47,9 +47,9 @@ namespace sce
 		/**
 		 * \brief Global GPU mode.
 		 * 
-		 * Any calls from HLE or emulator to
+		 * Equalrant/implementation of 
 		 * Gnm::GpuMode getGpuMode(void);
-		 * should be replaced by this mode.
+		 * 
 		 */
 		Gnm::GpuMode mode();