Vulkan: Experimental barrier skip when the game does not request it

src/Cafe/HW/Latte/Core/LatteCommandProcessor.cpp

@@ -214,6 +214,10 @@ LatteCMDPtr LatteCP_itSurfaceSync(LatteCMDPtr cmd)
 	MPTR addressPhys = LatteReadCMD() << 8;
 	uint32 pollInterval = LatteReadCMD();
 
+	// let the renderer know about colorbuffer invalidation
+	if (static_cast<uint32>(invalidationFlags & (Latte::E_COHER_CNTL::CB_ACTION_ENA | Latte::E_COHER_CNTL::CB_ALL_DEST_BASE_ENA)) != 0)
+		g_renderer->SurfaceSync(invalidationFlags, addressPhys, size);
+
 	if (addressPhys == MPTR_NULL || size == 0xFFFFFFFF)
 		return cmd; // block global invalidations because they are too expensive
 

src/Cafe/HW/Latte/ISA/LatteReg.h

@@ -359,7 +359,7 @@ namespace Latte
 		SO1_DEST_BASE_ENA     = (1<<3),
 		SO2_DEST_BASE_ENA     = (1<<4),
 		SO3_DEST_BASE_ENA     = (1<<5),
-		CB0_DEST_BASE_ENA     = (1<<6),
+		CB0_DEST_BASE_ENA     = (1<<6), // flush color buffer 0
 		CB1_DEST_BASE_ENA     = (1<<7),
 		CB2_DEST_BASE_ENA     = (1<<8),
 		CB3_DEST_BASE_ENA     = (1<<9),
@@ -367,6 +367,8 @@ namespace Latte
 		CB5_DEST_BASE_ENA     = (1<<11),
 		CB6_DEST_BASE_ENA     = (1<<12),
 		CB7_DEST_BASE_ENA     = (1<<13),
+		CB_ALL_DEST_BASE_ENA  = (0xFF<<6), // CB0-CB7 combined
+
 		DB_DEST_BASE_ENA      = (1<<14),
 
 		FULL_CACHE_ENA       = (1<<20),

src/Cafe/HW/Latte/Renderer/Renderer.h

@@ -82,6 +82,7 @@ public:
 	// flush control
 	virtual void Flush(bool waitIdle = false) = 0; // called when explicit flush is required (e.g. by imgui)
 	virtual void NotifyLatteCommandProcessorIdle() = 0; // called when command processor has no more commands available or when stalled
+	virtual void SurfaceSync(Latte::E_COHER_CNTL coher, MPTR address, uint32 size) {} // triggered by game via GX2Invalidate, can be utilized by the render backend as an optimization hint
 
 	// imgui
 	virtual bool ImguiBegin(bool mainWindow);

src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRenderer.h

@@ -240,6 +240,7 @@ public:
 
 	void Flush(bool waitIdle = false) override;
 	void NotifyLatteCommandProcessorIdle() override;
+	void SurfaceSync(Latte::E_COHER_CNTL coher, MPTR address, uint32 size) override;
 
 	uint64 GenUniqueId(); // return unique id (uses incrementing counter)
 
@@ -419,6 +420,7 @@ private:
 
 		// invalidation / flushing
 		uint64 currentFlushIndex{0};
+		bool colorBufferSyncPending{false}; // guest color-buffer sync since the previous draw; survives command-buffer resets
 		bool requestFlush{ false }; // flush after every draw operation. The renderpass dependencies dont handle dependencies across multiple drawcalls inside a single renderpass
 
 		// draw sequence

src/Cafe/HW/Latte/Renderer/Vulkan/VulkanRendererCore.cpp

@@ -995,6 +995,12 @@ VkDescriptorSetInfo* VulkanRenderer::draw_getOrCreateDescriptorSet(PipelineInfo*
 	return dsInfo;
 }
 
+void VulkanRenderer::SurfaceSync(Latte::E_COHER_CNTL coher, MPTR address, uint32 size)
+{
+	if (static_cast<uint32>(coher & Latte::E_COHER_CNTL::CB_ALL_DEST_BASE_ENA) != 0)
+		m_state.colorBufferSyncPending = true;
+}
+
 void VulkanRenderer::sync_inputTexturesChanged(bool withinFeedbackLoopRenderPass)
 {
 	bool writeFlushRequired = withinFeedbackLoopRenderPass; // feedback loop still requires us to emit a barrier
@@ -1029,6 +1035,12 @@ void VulkanRenderer::sync_inputTexturesChanged(bool withinFeedbackLoopRenderPass
 	// barrier here
 	if (writeFlushRequired)
 	{
+		// Continued draws with unchanged descriptors in the same renderpass only introduce feedback hazards.
+		// Relax color feedback without a guest sync, but keep the read indices above updated for later passes.
+		if (withinFeedbackLoopRenderPass && !m_state.descriptorSetsChanged && !m_state.colorBufferSyncPending
+			&& m_state.m_curRenderpassSelfDependencyInfo.GetAspectMask() == VK_IMAGE_ASPECT_COLOR_BIT)
+			return;
+
 		VkMemoryBarrier memoryBarrier{};
 		memoryBarrier.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER;
 		memoryBarrier.srcAccessMask = 0;
@@ -1740,6 +1752,7 @@ void VulkanRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
 		draw_execute_first(baseVertex, baseInstance, instanceCount, count, indexDataMPTR, indexType, drawcallContext);
 	else
 		draw_execute_continued(baseVertex, baseInstance, instanceCount, count, indexDataMPTR, indexType, drawcallContext);
+	m_state.colorBufferSyncPending = false;
 	LatteGPUState.drawCallCounter++;
 }
 

src/Cafe/OS/libs/gx2/GX2_Misc.cpp

@@ -286,11 +286,7 @@ namespace GX2
 
 	    if (HAS_FLAG(invalidationFlags, GX2InvalidationFlag::GPU_COLOR_BUFFER))
 	    {
-	        coherCntl |= Latte::E_COHER_CNTL::CB_ACTION_ENA
-	                   | Latte::E_COHER_CNTL::CB0_DEST_BASE_ENA | Latte::E_COHER_CNTL::CB1_DEST_BASE_ENA
-	                   | Latte::E_COHER_CNTL::CB2_DEST_BASE_ENA | Latte::E_COHER_CNTL::CB3_DEST_BASE_ENA
-	                   | Latte::E_COHER_CNTL::CB4_DEST_BASE_ENA | Latte::E_COHER_CNTL::CB5_DEST_BASE_ENA
-	                   | Latte::E_COHER_CNTL::CB6_DEST_BASE_ENA | Latte::E_COHER_CNTL::CB7_DEST_BASE_ENA;
+	        coherCntl |= Latte::E_COHER_CNTL::CB_ACTION_ENA | Latte::E_COHER_CNTL::CB_ALL_DEST_BASE_ENA;
 	    }
 
 	    if (HAS_FLAG(invalidationFlags, GX2InvalidationFlag::GPU_DEPTH_BUFFER))