Vulkan: Rework inter-renderpass barrier code

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

@@ -124,12 +124,8 @@ void LatteTextureReadbackInfoVk::StartTransfer()
 
 	renderer->barrier_image<VulkanRenderer::ANY_TRANSFER | VulkanRenderer::IMAGE_WRITE, VulkanRenderer::TRANSFER_READ>(baseTexture, region.imageSubresource, VK_IMAGE_LAYOUT_GENERAL);
 
-	renderer->barrier_sequentializeTransfer();
-
 	vkCmdCopyImageToBuffer(renderer->getCurrentCommandBuffer(), baseTexture->GetImageObj()->m_image, VK_IMAGE_LAYOUT_GENERAL, m_buffer, 1, &region);
 
-	renderer->barrier_sequentializeTransfer();
-
 	renderer->barrier_image<VulkanRenderer::TRANSFER_READ, VulkanRenderer::ANY_TRANSFER | VulkanRenderer::IMAGE_WRITE>(baseTexture, region.imageSubresource, VK_IMAGE_LAYOUT_GENERAL); // make sure transfer is finished before image is modified
 	renderer->barrier_bufferRange<VulkanRenderer::TRANSFER_WRITE, VulkanRenderer::HOST_READ>(m_buffer, m_buffer_offset, m_image_size); // make sure transfer is finished before result is read
 

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

@@ -3222,6 +3222,14 @@ void VulkanRenderer::ProcessDestructionQueue()
 	m_spinlockDestructionQueue.unlock();
 }
 
+void VkDescriptorSetInfo::ForEachView(const std::function<void(LatteTextureViewVk*)>& fun)
+{
+	for (auto& view : list_referencedViews)
+	{
+		fun(view);
+	}
+}
+
 VkDescriptorSetInfo::~VkDescriptorSetInfo()
 {
 	for (auto& it : list_referencedViews)
@@ -3701,9 +3709,9 @@ void VulkanRenderer::bufferCache_upload(uint8* buffer, sint32 size, uint32 buffe
 
 	vkMemAllocator.FlushReservation(uploadResv);
 
-	barrier_bufferRange<ANY_TRANSFER | HOST_WRITE, ANY_TRANSFER,
-		BUFFER_SHADER_READ, TRANSFER_WRITE>(
-			uploadResv.vkBuffer, uploadResv.bufferOffset, uploadResv.size, // make sure any in-flight transfers are completed
+	barrier_bufferRange<TRANSFER_WRITE | HOST_WRITE, TRANSFER_READ,
+						TRANSFER_READ | BUFFER_SHADER_READ, TRANSFER_WRITE>(
+			uploadResv.vkBuffer, uploadResv.bufferOffset, uploadResv.size, // make sure source data is visible
 			m_bufferCache, bufferOffset, size); // make sure all reads are completed before we overwrite the data
 
 	VkBufferCopy region;
@@ -3712,15 +3720,18 @@ void VulkanRenderer::bufferCache_upload(uint8* buffer, sint32 size, uint32 buffe
 	region.size = size;
 	vkCmdCopyBuffer(m_state.currentCommandBuffer, uploadResv.vkBuffer, m_bufferCache, 1, &region);
 
-	barrier_sequentializeTransfer();
+	barrier_bufferRange<TRANSFER_WRITE, BUFFER_SHADER_READ>(m_bufferCache, bufferOffset, size);
 }
 
 void VulkanRenderer::bufferCache_copy(uint32 srcOffset, uint32 dstOffset, uint32 size)
 {
 	cemu_assert_debug(!m_useHostMemoryForCache);
 	draw_endRenderPass();
 
-	barrier_sequentializeTransfer();
+	barrier_bufferRange<BUFFER_SHADER_WRITE | TRANSFER_WRITE, TRANSFER_READ,
+		TRANSFER_READ | BUFFER_SHADER_READ, TRANSFER_WRITE>(
+			m_bufferCache, srcOffset, size, // make sure source data is visible
+			m_bufferCache, dstOffset, size); // make sure all reads are completed before we overwrite the data
 
 	bool isOverlapping = (srcOffset + size) > dstOffset && (srcOffset) < (dstOffset + size);
 	cemu_assert_debug(!isOverlapping);
@@ -3731,7 +3742,7 @@ void VulkanRenderer::bufferCache_copy(uint32 srcOffset, uint32 dstOffset, uint32
 	bufferCopy.size = size;
 	vkCmdCopyBuffer(m_state.currentCommandBuffer, m_bufferCache, m_bufferCache, 1, &bufferCopy);
 
-	barrier_sequentializeTransfer();
+	barrier_bufferRange<TRANSFER_WRITE, BUFFER_SHADER_READ>(m_bufferCache, dstOffset, size);
 }
 
 void VulkanRenderer::bufferCache_copyStreamoutToMainBuffer(uint32 srcOffset, uint32 dstOffset, uint32 size)
@@ -3748,20 +3759,18 @@ void VulkanRenderer::bufferCache_copyStreamoutToMainBuffer(uint32 srcOffset, uin
 	else
 		dstBuffer = m_bufferCache;
 
-	barrier_bufferRange<BUFFER_SHADER_WRITE, TRANSFER_READ,
-		ANY_TRANSFER | BUFFER_SHADER_READ, TRANSFER_WRITE>(
-			m_xfbRingBuffer, srcOffset, size, // wait for all writes to finish
-			dstBuffer, dstOffset, size); // wait for all reads to finish
-
-	barrier_sequentializeTransfer();
+	barrier_bufferRange<BUFFER_SHADER_WRITE | TRANSFER_WRITE, TRANSFER_READ,
+							TRANSFER_READ | BUFFER_SHADER_READ, TRANSFER_WRITE>(
+			m_xfbRingBuffer, srcOffset, size, // make sure source data is visible
+			dstBuffer, dstOffset, size); // make sure all reads are completed before we overwrite the data
 
 	VkBufferCopy bufferCopy{};
 	bufferCopy.srcOffset = srcOffset;
 	bufferCopy.dstOffset = dstOffset;
 	bufferCopy.size = size;
 	vkCmdCopyBuffer(m_state.currentCommandBuffer, m_xfbRingBuffer, dstBuffer, 1, &bufferCopy);
 
-	barrier_sequentializeTransfer();
+	barrier_bufferRange<TRANSFER_WRITE, BUFFER_SHADER_READ>(dstBuffer, dstOffset, size); // make sure writes are visible to host
 }
 
 void VulkanRenderer::AppendOverlayDebugInfo()

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

@@ -8,6 +8,7 @@
 #include "Cafe/HW/Latte/Renderer/Vulkan/CachedFBOVk.h"
 #include "Cafe/HW/Latte/Renderer/Vulkan/VKRMemoryManager.h"
 #include "Cafe/HW/Latte/Renderer/Vulkan/SwapchainInfoVk.h"
+#include "Cafe/HW/Latte/Core/LattePerformanceMonitor.h"
 #include "util/math/vector2.h"
 #include "util/helpers/Semaphore.h"
 #include "util/containers/flat_hash_map.hpp"
@@ -26,6 +27,8 @@ struct VkDescriptorSetInfo
 {
 	VKRObjectDescriptorSet* m_vkObjDescriptorSet{};
 
+	void ForEachView(const std::function<void(LatteTextureViewVk*)>& fun);
+
 	~VkDescriptorSetInfo();
 
 	std::vector<LatteTextureViewVk*> list_referencedViews;
@@ -409,7 +412,7 @@ class VulkanRenderer : public Renderer
 		}
 
 		// invalidation / flushing
-		uint64 currentFlushIndex{0};
+		uint64 currentFlushIndex{1};
 		bool requestFlush{ false }; // flush after every draw operation. The renderpass dependencies dont handle dependencies across multiple drawcalls inside a single renderpass
 
 		// draw sequence
@@ -546,7 +549,8 @@ class VulkanRenderer : public Renderer
 	void draw_handleSpecialState5();
 
 	// draw synchronization helper
-	void sync_inputTexturesChanged();
+	void sync_performFlushBarrier(CachedFBOVk* fboVk);
+	bool sync_isInputTexturesSyncRequired();
 	void sync_RenderPassLoadTextures(CachedFBOVk* fboVk);
 	void sync_RenderPassStoreTextures(CachedFBOVk* fboVk);
 
@@ -825,6 +829,7 @@ class VulkanRenderer : public Renderer
 		bufMemBarrier.offset = offset;
 		bufMemBarrier.size = size;
 		vkCmdPipelineBarrier(m_state.currentCommandBuffer, srcStages, dstStages, 0, 0, nullptr, 1, &bufMemBarrier, 0, nullptr);
+		performanceMonitor.vk.numDrawBarriersPerFrame.increment();
 	}
 
 	template<uint32 TSrcSyncOpA, uint32 TDstSyncOpA, uint32 TSrcSyncOpB, uint32 TDstSyncOpB>
@@ -863,32 +868,7 @@ class VulkanRenderer : public Renderer
 		bufMemBarrier[1].size = sizeB;
 
 		vkCmdPipelineBarrier(m_state.currentCommandBuffer, srcStagesA|srcStagesB, dstStagesA|dstStagesB, 0, 0, nullptr, 2, bufMemBarrier, 0, nullptr);
-	}
-
-	void barrier_sequentializeTransfer()
-	{
-		VkMemoryBarrier memBarrier{};
-		memBarrier.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER;
-		memBarrier.pNext = nullptr;
-
-		VkPipelineStageFlags srcStages = VK_PIPELINE_STAGE_TRANSFER_BIT;
-		VkPipelineStageFlags dstStages = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
-
-		memBarrier.srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT;
-		memBarrier.dstAccessMask = 0;
-
-		memBarrier.srcAccessMask |= (VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT);
-		memBarrier.dstAccessMask |= (VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT);
-
-		vkCmdPipelineBarrier(m_state.currentCommandBuffer, srcStages, dstStages, 0, 1, &memBarrier, 0, nullptr, 0, nullptr);
-	}
-
-	void barrier_sequentializeCommand()
-	{
-		VkPipelineStageFlags srcStages = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;
-		VkPipelineStageFlags dstStages = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
-
-		vkCmdPipelineBarrier(m_state.currentCommandBuffer, srcStages, dstStages, 0, 0, nullptr, 0, nullptr, 0, nullptr);
+		performanceMonitor.vk.numDrawBarriersPerFrame.increment();
 	}
 
 	template<uint32 TSrcSyncOp, uint32 TDstSyncOp>
@@ -916,6 +896,7 @@ class VulkanRenderer : public Renderer
 							 0, NULL,
 							 0, NULL,
 							 1, &imageMemBarrier);
+		performanceMonitor.vk.numDrawBarriersPerFrame.increment();
 	}
 
 	template<uint32 TSrcSyncOp, uint32 TDstSyncOp>