Vulkan: Additional smaller optimizations

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

@@ -311,6 +311,7 @@ void LatteCP_itSetRegistersGeneric_handleSpecialRanges(uint32 registerStartIndex
 template<uint32 TRegisterBase>
 LatteCMDPtr LatteCP_itSetRegistersGeneric(LatteCMDPtr cmd, uint32 nWords)
 {
+	nWords--; // subtract the register offset field
 	uint32 registerOffset = LatteReadCMD();
 	uint32 registerIndex = TRegisterBase + registerOffset;
 	uint32 registerStartIndex = registerIndex;
@@ -324,7 +325,7 @@ LatteCMDPtr LatteCP_itSetRegistersGeneric(LatteCMDPtr cmd, uint32 nWords)
 		// state shadowing enabled
 		uint32* __restrict shadowAddrs = LatteGPUState.contextRegisterShadowAddr + registerIndex;
 		sint32 indexCounter = 0;
-		while (--nWords)
+		while (nWords--)
 		{
 			uint32 dataWord = LatteReadCMD();
 			MPTR regShadowAddr = shadowAddrs[indexCounter];
@@ -337,10 +338,19 @@ LatteCMDPtr LatteCP_itSetRegistersGeneric(LatteCMDPtr cmd, uint32 nWords)
 	else
 	{
 		// state shadowing disabled
-		while (--nWords)
+		if (nWords == 1) // common case
 		{
 			*outputReg = LatteReadCMD();
-			outputReg++;
+		}
+		else
+		{
+			sint32 i = 0;
+			while (i < nWords)
+			{
+				outputReg[i] = cmd[i];
+				i++;
+			}
+			cmd += nWords;
 		}
 	}
 	// some register writes trigger special behavior
@@ -352,6 +362,7 @@ LatteCMDPtr LatteCP_itSetRegistersGeneric(LatteCMDPtr cmd, uint32 nWords)
 template<uint32 TRegisterBase, typename TRegRangeCallback>
 bool LatteCP_itSetRegistersGeneric2(LatteCMDPtr cmd, uint32 nWords, TRegRangeCallback cbRegRange)
 {
+	nWords--;
 	const uint32 registerOffset = LatteReadCMD();
 	const uint32 registerIndex = TRegisterBase + registerOffset;
 	const uint32 registerStartIndex = registerIndex;
@@ -365,7 +376,7 @@ bool LatteCP_itSetRegistersGeneric2(LatteCMDPtr cmd, uint32 nWords, TRegRangeCal
 		// state shadowing enabled
 		uint32* shadowAddrs = LatteGPUState.contextRegisterShadowAddr + registerIndex;
 		sint32 indexCounter = 0;
-		while (--nWords)
+		while (nWords--)
 		{
 			uint32 dataWord = LatteReadCMD();
 			MPTR regShadowAddr = shadowAddrs[indexCounter];
@@ -379,13 +390,23 @@ bool LatteCP_itSetRegistersGeneric2(LatteCMDPtr cmd, uint32 nWords, TRegRangeCal
 	else
 	{
 		// state shadowing disabled
-		sint32 indexCounter = 0;
+		if (nWords == 1) // common case
-		while (--nWords)
 		{
-			uint32 dataWord = LatteReadCMD();
+			uint32 v = LatteReadCMD();
-			hasRegChange |= (*outputReg != dataWord);
+			hasRegChange |= (*outputReg != v);
-			*outputReg = dataWord;
+			*outputReg = v;
-			outputReg++;
+		}
+		else
+		{
+			sint32 i = 0;
+			while (i < nWords)
+			{
+				uint32 v = cmd[i];
+				hasRegChange |= (outputReg[i] != v);
+				outputReg[i] = v;
+				i++;
+			}
+			cmd += nWords;
 		}
 	}
 	// some register writes trigger special behavior

src/Cafe/HW/Latte/Core/LatteConst.h

@@ -109,16 +109,13 @@ namespace LatteConst
 {
 	enum class ShaderType : uint32
 	{
-		Reserved = 0,
 		// shaders for drawing
-		FirstRender = 1,
+		Vertex = 0,
-		Vertex = 1,
+		Pixel = 1,
-		Pixel = 2,
+		Geometry = 2,
-		Geometry = 3,
-		LastRender = 3,
 		// compute shader
-		Compute = 4,
+		Compute = 3,
-		TotalCount = 5
+		TotalCount = 4
 	};
 
 	enum class VertexFetchNFA

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

@@ -58,20 +58,13 @@ PipelineInfo::~PipelineInfo()
 	}
 
 	// delete descriptor sets
-	while (!pixel_ds_cache.empty())
+	for (auto& it : ds_cache)
 	{
-		VkDescriptorSetInfo* dsInfo = pixel_ds_cache.begin()->second;
+		while (!it.empty())
-		delete dsInfo;
+		{
-	}
+			VkDescriptorSetInfo* dsInfo = it.begin()->second;
-	while (!geometry_ds_cache.empty())
+			delete dsInfo;
-	{
+		}
-		VkDescriptorSetInfo* dsInfo = geometry_ds_cache.begin()->second;
-		delete dsInfo;
-	}
-	while (!vertex_ds_cache.empty())
-	{
-		VkDescriptorSetInfo* dsInfo = vertex_ds_cache.begin()->second;
-		delete dsInfo;
 	}
 
 	// disassociate from shaders

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

@@ -3374,29 +3374,8 @@ VkDescriptorSetInfo::~VkDescriptorSetInfo()
 	for (auto& it : list_referencedViews)
 		it->RemoveDescriptorSetReference(this);
 	// unregister
-	switch (shaderType)
+	auto r = pipeline_info->GetDescriptorSetCache(shaderType).erase(stateHash);
-	{
+	cemu_assert_debug(r == 1);
-	case LatteConst::ShaderType::Vertex:
-	{
-		auto r = pipeline_info->vertex_ds_cache.erase(stateHash);
-		cemu_assert_debug(r == 1);
-		break;
-	}
-	case LatteConst::ShaderType::Pixel:
-	{
-		auto r = pipeline_info->pixel_ds_cache.erase(stateHash);
-		cemu_assert_debug(r == 1);
-		break;
-	}
-	case LatteConst::ShaderType::Geometry:
-	{
-		auto r = pipeline_info->geometry_ds_cache.erase(stateHash);
-		cemu_assert_debug(r == 1);
-		break;
-	}
-	default:
-		UNREACHABLE;
-	}
 	// update global stats
 	performanceMonitor.vk.numDescriptorSamplerTextures.decrement(statsNumSamplerTextures);
 	performanceMonitor.vk.numDescriptorDynUniformBuffers.decrement(statsNumDynUniformBuffers);

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

@@ -55,12 +55,17 @@ private:
 
 namespace VulkanRendererConst
 {
-	static const inline int SHADER_STAGE_INDEX_VERTEX = 0;
+	static const inline int SHADER_STAGE_INDEX_VERTEX = static_cast<int>(LatteConst::ShaderType::Vertex);
-	static const inline int SHADER_STAGE_INDEX_FRAGMENT = 1;
+	static const inline int SHADER_STAGE_INDEX_FRAGMENT = static_cast<int>(LatteConst::ShaderType::Pixel);
-	static const inline int SHADER_STAGE_INDEX_GEOMETRY = 2;
+	static const inline int SHADER_STAGE_INDEX_GEOMETRY = static_cast<int>(LatteConst::ShaderType::Geometry);
-	static const inline int SHADER_STAGE_INDEX_COUNT = 3;
+	static const inline int SHADER_STAGE_INDEX_COUNT = 4;
 };
 
+// the order doesnt really matter but the types should cover range 0-2 since we use them as an array index
+static_assert(static_cast<int>(LatteConst::ShaderType::Vertex) == 0);
+static_assert(static_cast<int>(LatteConst::ShaderType::Pixel) == 1);
+static_assert(static_cast<int>(LatteConst::ShaderType::Geometry) == 2);
+
 class PipelineInfo
 {
 public:
@@ -82,11 +87,18 @@ public:
 			return k;
 		}
 	};
+	using DescriptorSetCache = ska::flat_hash_map<uint64, VkDescriptorSetInfo*, direct_hash<uint64>>;
+
+	FORCEINLINE DescriptorSetCache& GetDescriptorSetCache(LatteConst::ShaderType shaderType)
+	{
+		cemu_assert_debug(shaderType == LatteConst::ShaderType::Vertex || shaderType == LatteConst::ShaderType::Pixel || shaderType == LatteConst::ShaderType::Geometry);
+		return ds_cache[static_cast<size_t>(shaderType)];
+	}
 
 	// std::unordered_map<uint64, VkDescriptorSetInfo*> 3.16% (total CPU time)
-	// robin_hood::unordered_flat_map<uint64, VkDescriptorSetInfo*> vertex_ds_cache, pixel_ds_cache, geometry_ds_cache; ~1.80%
+	// robin_hood::unordered_flat_map<uint64, VkDescriptorSetInfo*> descriptor set cache; ~1.80%
-	// ska::bytell_hash_map<uint64, VkDescriptorSetInfo*, direct_hash<uint64>> vertex_ds_cache, pixel_ds_cache, geometry_ds_cache; -> 1.91%
+	// ska::bytell_hash_map<uint64, VkDescriptorSetInfo*, direct_hash<uint64>> descriptor set cache; -> 1.91%
-	ska::flat_hash_map<uint64, VkDescriptorSetInfo*, direct_hash<uint64>> vertex_ds_cache, pixel_ds_cache, geometry_ds_cache; // 1.71%
+	DescriptorSetCache ds_cache[VulkanRendererConst::SHADER_STAGE_INDEX_COUNT]; // 1.71%
 
 	VKRObjectPipeline* m_vkrObjPipeline;
 

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

@@ -590,31 +590,27 @@ uint64 VulkanRenderer::GetDescriptorSetStateHash(LatteDecompilerShader* shader)
 VkDescriptorSetInfo* VulkanRenderer::draw_getOrCreateDescriptorSet(PipelineInfo* pipeline_info, LatteDecompilerShader* shader)
 {
 	const uint64 stateHash = GetDescriptorSetStateHash(shader);
+	cemu_assert_debug(shader->shaderType == LatteConst::ShaderType::Vertex || shader->shaderType == LatteConst::ShaderType::Pixel || shader->shaderType == LatteConst::ShaderType::Geometry);
+	auto& ds_cache = pipeline_info->GetDescriptorSetCache(shader->shaderType);
+	const auto it = ds_cache.find(stateHash);
+	if (it != ds_cache.cend())
+		return it->second;
 
 	VkDescriptorSetLayout descriptor_set_layout;
 	switch (shader->shaderType)
 	{
 	case LatteConst::ShaderType::Vertex:
 	{
-		const auto it = pipeline_info->vertex_ds_cache.find(stateHash);
-		if (it != pipeline_info->vertex_ds_cache.cend())
-			return it->second;
 		descriptor_set_layout = pipeline_info->m_vkrObjPipeline->m_vertexDSL;
 		break;
 	}
 	case LatteConst::ShaderType::Pixel:
 	{
-		const auto it = pipeline_info->pixel_ds_cache.find(stateHash);
-		if (it != pipeline_info->pixel_ds_cache.cend())
-			return it->second;
 		descriptor_set_layout = pipeline_info->m_vkrObjPipeline->m_pixelDSL;
 		break;
 	}
 	case LatteConst::ShaderType::Geometry:
 	{
-		const auto it = pipeline_info->geometry_ds_cache.find(stateHash);
-		if (it != pipeline_info->geometry_ds_cache.cend())
-			return it->second;
 		descriptor_set_layout = pipeline_info->m_vkrObjPipeline->m_geometryDSL;
 		break;
 	}
@@ -1006,26 +1002,7 @@ VkDescriptorSetInfo* VulkanRenderer::draw_getOrCreateDescriptorSet(PipelineInfo*
 	if (!descriptorWrites.empty())
 		vkUpdateDescriptorSets(m_logicalDevice, (uint32)descriptorWrites.size(), descriptorWrites.data(), 0, nullptr);
 
-	switch (shader->shaderType)
+	ds_cache[stateHash] = dsInfo;
-	{
-	case LatteConst::ShaderType::Vertex:
-	{
-		pipeline_info->vertex_ds_cache[stateHash] = dsInfo;
-		break;
-	}
-	case LatteConst::ShaderType::Pixel:
-	{
-		pipeline_info->pixel_ds_cache[stateHash] = dsInfo;
-		break;
-	}
-	case LatteConst::ShaderType::Geometry:
-	{
-		pipeline_info->geometry_ds_cache[stateHash] = dsInfo;
-		break;
-	}
-	default:
-		UNREACHABLE;
-	}
 
 	return dsInfo;
 }