Merge branch 'main' into metal

.github/workflows/generate_pot.yml

@@ -35,7 +35,7 @@ jobs:
           -o cemu.pot
 
       - name: Upload artifact
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: POT file
           path: ./cemu.pot

src/Cafe/GameProfile/GameProfile.cpp

@@ -140,7 +140,7 @@ bool gameProfile_loadEnumOption(IniParser& iniParser, const char* optionName, T&
 	for(const T& v : T())
 	{
 		// test integer option
-		if (boost::iequals(fmt::format("{}", static_cast<typename std::underlying_type<T>::type>(v)), *option_value))
+		if (boost::iequals(fmt::format("{}", fmt::underlying(v)), *option_value))
 		{
 			option = v;
 			return true;

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

@@ -502,6 +502,9 @@ void LatteShaderCache_Load()
 		g_renderer->DeleteTexture(g_shaderCacheLoaderState.textureDRCId);
 
 	g_bootSndPlayer.FadeOutSound();
+
+	if(Latte_GetStopSignal())
+		LatteThread_Exit();
 }
 
 void LatteShaderCache_ShowProgress(const std::function <bool(void)>& loadUpdateFunc, bool isPipelines)
@@ -651,8 +654,6 @@ void LatteShaderCache_LoadPipelineCache(uint64 cacheTitleId)
 	else if (g_renderer->GetType() == RendererAPI::Metal)
 		MetalPipelineCache::GetInstance().EndLoading();
 #endif
-    if(Latte_GetStopSignal())
-        LatteThread_Exit();
 }
 
 bool LatteShaderCache_updatePipelineLoadingProgress()

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

@@ -257,6 +257,7 @@ void LatteThread_Exit()
     LatteSHRC_UnloadAll();
     // close disk cache
     LatteShaderCache_Close();
+	RendererOutputShader::ShutdownStatic();
     // destroy renderer but make sure that g_renderer remains valid until the destructor has finished
 	if (g_renderer)
 	{

src/Cafe/HW/Latte/Renderer/RendererOuputShader.cpp

@@ -251,8 +251,8 @@ RendererOutputShader::RendererOutputShader(const std::string& vertex_source, con
     else
         finalFragmentSrc = PrependFragmentPreamble(fragment_source);
 
-	m_vertex_shader = g_renderer->shader_create(RendererShader::ShaderType::kVertex, 0, 0, vertex_source, false, false);
-	m_fragment_shader = g_renderer->shader_create(RendererShader::ShaderType::kFragment, 0, 0, finalFragmentSrc, false, false);
+	m_vertex_shader.reset(g_renderer->shader_create(RendererShader::ShaderType::kVertex, 0, 0, vertex_source, false, false));
+	m_fragment_shader.reset(g_renderer->shader_create(RendererShader::ShaderType::kFragment, 0, 0, finalFragmentSrc, false, false));
 
 	m_vertex_shader->PreponeCompilation(true);
 	m_fragment_shader->PreponeCompilation(true);
@@ -302,8 +302,8 @@ void RendererOutputShader::SetUniformParameters(const LatteTextureView& texture_
 		  shader->SetUniform2fv(locations.m_loc_outputResolution, res, 1);
 	  }
 	};
-	setUniforms(m_vertex_shader, m_uniformLocations[0]);
-	setUniforms(m_fragment_shader, m_uniformLocations[1]);
+	setUniforms(m_vertex_shader.get(), m_uniformLocations[0]);
+	setUniforms(m_fragment_shader.get(), m_uniformLocations[1]);
 }
 
 RendererOutputShader* RendererOutputShader::s_copy_shader;
@@ -513,3 +513,15 @@ void RendererOutputShader::InitializeStatic()
     	s_hermit_shader_ud = new RendererOutputShader(vertex_source_ud, s_hermite_shader_source);
     }
 }
+
+void RendererOutputShader::ShutdownStatic()
+{
+	delete s_copy_shader;
+	delete s_copy_shader_ud;
+
+	delete s_bicubic_shader;
+	delete s_bicubic_shader_ud;
+
+	delete s_hermit_shader;
+	delete s_hermit_shader_ud;
+}

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

@@ -21,15 +21,16 @@ class RendererOutputShader
 
 	RendererShader* GetVertexShader() const
 	{
-		return m_vertex_shader;
+		return m_vertex_shader.get();
 	}
 
 	RendererShader* GetFragmentShader() const
 	{
-		return m_fragment_shader;
+		return m_fragment_shader.get();
 	}
 
 	static void InitializeStatic();
+	static void ShutdownStatic();
 
 	static RendererOutputShader* s_copy_shader;
 	static RendererOutputShader* s_copy_shader_ud;
@@ -47,8 +48,8 @@ class RendererOutputShader
 	static std::string PrependFragmentPreamble(const std::string& shaderSrc);
 
 protected:
-	RendererShader* m_vertex_shader;
-	RendererShader* m_fragment_shader;
+	std::unique_ptr<RendererShader> m_vertex_shader;
+	std::unique_ptr<RendererShader> m_fragment_shader;
 
 	struct UniformLocations
 	{

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

@@ -211,6 +211,9 @@ RendererShaderVk::~RendererShaderVk()
 {
 	while (!list_pipelineInfo.empty())
 		delete list_pipelineInfo[0];
+
+	VkDevice vkDev = VulkanRenderer::GetInstance()->GetLogicalDevice();
+	vkDestroyShaderModule(vkDev, m_shader_module, nullptr);
 }
 
 void RendererShaderVk::Init()

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

@@ -60,7 +60,7 @@ void SwapchainInfoVk::Create()
 	VkAttachmentDescription colorAttachment = {};
 	colorAttachment.format = m_surfaceFormat.format;
 	colorAttachment.samples = VK_SAMPLE_COUNT_1_BIT;
-	colorAttachment.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
+	colorAttachment.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 	colorAttachment.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
 	colorAttachment.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 	colorAttachment.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;

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

@@ -70,6 +70,7 @@ struct SwapchainInfoVk
 	VkSurfaceFormatKHR m_surfaceFormat{};
 	VkSwapchainKHR m_swapchain{};
 	Vector2i m_desiredExtent{};
+	VkExtent2D m_actualExtent{};
 	uint32 swapchainImageIndex = (uint32)-1;
 	uint64 m_presentId = 1;
 	uint64 m_queueDepth = 0; // number of frames with pending presentation requests
@@ -92,5 +93,4 @@ struct SwapchainInfoVk
 	VkSemaphore m_currentSemaphore = VK_NULL_HANDLE;
 
 	std::array<uint32, 2> m_swapchainQueueFamilyIndices;
-	VkExtent2D m_actualExtent{};
 };

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

@@ -4,6 +4,14 @@
 
 /* VKRSynchronizedMemoryBuffer */
 
+VKRSynchronizedRingAllocator::~VKRSynchronizedRingAllocator()
+{
+	for(auto& buf : m_buffers)
+	{
+		m_vkrMemMgr->DeleteBuffer(buf.vk_buffer, buf.vk_mem);
+	}
+}
+
 void VKRSynchronizedRingAllocator::addUploadBufferSyncPoint(AllocatorBuffer_t& buffer, uint32 offset)
 {
 	auto cmdBufferId = m_vkr->GetCurrentCommandBufferId();
@@ -233,6 +241,15 @@ void VKRSynchronizedHeapAllocator::GetStats(uint32& numBuffers, size_t& totalBuf
 
 /* VkTextureChunkedHeap */
 
+VkTextureChunkedHeap::~VkTextureChunkedHeap()
+{
+	VkDevice device = VulkanRenderer::GetInstance()->GetLogicalDevice();
+	for (auto& i : m_list_chunkInfo)
+	{
+		vkFreeMemory(device, i.mem, nullptr);
+	}
+}
+
 uint32 VkTextureChunkedHeap::allocateNewChunk(uint32 chunkIndex, uint32 minimumAllocationSize)
 {
 	cemu_assert_debug(m_list_chunkInfo.size() == chunkIndex);
@@ -310,11 +327,11 @@ VKRBuffer* VKRBuffer::Create(VKR_BUFFER_TYPE bufferType, size_t bufferSize, VkMe
 	VkDeviceMemory bufferMemory;
 	bool allocSuccess;
 	if (bufferType == VKR_BUFFER_TYPE::STAGING)
-		allocSuccess = memMgr->CreateBuffer2(bufferSize, VK_BUFFER_USAGE_TRANSFER_SRC_BIT, properties, buffer, bufferMemory);
+		allocSuccess = memMgr->CreateBuffer(bufferSize, VK_BUFFER_USAGE_TRANSFER_SRC_BIT, properties, buffer, bufferMemory);
 	else if (bufferType == VKR_BUFFER_TYPE::INDEX)
-		allocSuccess = memMgr->CreateBuffer2(bufferSize, VK_BUFFER_USAGE_INDEX_BUFFER_BIT, properties, buffer, bufferMemory);
+		allocSuccess = memMgr->CreateBuffer(bufferSize, VK_BUFFER_USAGE_INDEX_BUFFER_BIT, properties, buffer, bufferMemory);
 	else if (bufferType == VKR_BUFFER_TYPE::STRIDE)
-		allocSuccess = memMgr->CreateBuffer2(bufferSize, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, properties, buffer, bufferMemory);
+		allocSuccess = memMgr->CreateBuffer(bufferSize, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT, properties, buffer, bufferMemory);
 	else
 		cemu_assert_debug(false);
 	if (!allocSuccess)
@@ -363,28 +380,14 @@ uint32 VkBufferChunkedHeap::allocateNewChunk(uint32 chunkIndex, uint32 minimumAl
 	return allocationSize;
 }
 
-uint32_t VKRMemoryManager::FindMemoryType(uint32_t typeFilter, VkMemoryPropertyFlags properties) const
-{
-	VkPhysicalDeviceMemoryProperties memProperties;
-	vkGetPhysicalDeviceMemoryProperties(m_vkr->GetPhysicalDevice(), &memProperties);
-
-	for (uint32 i = 0; i < memProperties.memoryTypeCount; i++)
-	{
-		if ((typeFilter & (1 << i)) != 0 && (memProperties.memoryTypes[i].propertyFlags & properties) == properties)
-			return i;
-	}
-	m_vkr->UnrecoverableError(fmt::format("failed to find suitable memory type ({0:#08x} {1:#08x})", typeFilter, properties).c_str());
-	return 0;
-}
-
-bool VKRMemoryManager::FindMemoryType2(uint32 typeFilter, VkMemoryPropertyFlags properties, uint32& memoryIndex) const
+bool VKRMemoryManager::FindMemoryType(uint32 typeFilter, VkMemoryPropertyFlags properties, uint32& memoryIndex) const
 {
 	VkPhysicalDeviceMemoryProperties memProperties;
 	vkGetPhysicalDeviceMemoryProperties(m_vkr->GetPhysicalDevice(), &memProperties);
 
 	for (uint32_t i = 0; i < memProperties.memoryTypeCount; i++)
 	{
-		if (typeFilter & (1 << i) && memProperties.memoryTypes[i].propertyFlags == properties)
+		if (typeFilter & (1 << i) && (memProperties.memoryTypes[i].propertyFlags & properties) == properties)
 		{
 			memoryIndex = i;
 			return true;
@@ -455,31 +458,7 @@ size_t VKRMemoryManager::GetTotalMemoryForBufferType(VkBufferUsageFlags usage, V
 	return total;
 }
 
-void VKRMemoryManager::CreateBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, VkBuffer& buffer, VkDeviceMemory& bufferMemory) const
-{
-	VkBufferCreateInfo bufferInfo{};
-	bufferInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
-	bufferInfo.usage = usage;
-	bufferInfo.size = size;
-	bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
-	if (vkCreateBuffer(m_vkr->GetLogicalDevice(), &bufferInfo, nullptr, &buffer) != VK_SUCCESS)
-		m_vkr->UnrecoverableError("Failed to create buffer");
-
-	VkMemoryRequirements memRequirements;
-	vkGetBufferMemoryRequirements(m_vkr->GetLogicalDevice(), buffer, &memRequirements);
-
-	VkMemoryAllocateInfo allocInfo{};
-	allocInfo.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
-	allocInfo.allocationSize = memRequirements.size;
-	allocInfo.memoryTypeIndex = FindMemoryType(memRequirements.memoryTypeBits, properties);
-
-	if (vkAllocateMemory(m_vkr->GetLogicalDevice(), &allocInfo, nullptr, &bufferMemory) != VK_SUCCESS)
-		m_vkr->UnrecoverableError("Failed to allocate buffer memory");
-	if (vkBindBufferMemory(m_vkr->GetLogicalDevice(), buffer, bufferMemory, 0) != VK_SUCCESS)
-		m_vkr->UnrecoverableError("Failed to bind buffer memory");
-}
-
-bool VKRMemoryManager::CreateBuffer2(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, VkBuffer& buffer, VkDeviceMemory& bufferMemory) const
+bool VKRMemoryManager::CreateBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, VkBuffer& buffer, VkDeviceMemory& bufferMemory) const
 {
 	VkBufferCreateInfo bufferInfo{};
 	bufferInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
@@ -488,7 +467,7 @@ bool VKRMemoryManager::CreateBuffer2(VkDeviceSize size, VkBufferUsageFlags usage
 	bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
 	if (vkCreateBuffer(m_vkr->GetLogicalDevice(), &bufferInfo, nullptr, &buffer) != VK_SUCCESS)
 	{
-		cemuLog_log(LogType::Force, "Failed to create buffer (CreateBuffer2)");
+		cemuLog_log(LogType::Force, "Failed to create buffer (CreateBuffer)");
 		return false;
 	}
 
@@ -498,7 +477,7 @@ bool VKRMemoryManager::CreateBuffer2(VkDeviceSize size, VkBufferUsageFlags usage
 	VkMemoryAllocateInfo allocInfo{};
 	allocInfo.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
 	allocInfo.allocationSize = memRequirements.size;
-	if (!FindMemoryType2(memRequirements.memoryTypeBits, properties, allocInfo.memoryTypeIndex))
+	if (!FindMemoryType(memRequirements.memoryTypeBits, properties, allocInfo.memoryTypeIndex))
 	{
 		vkDestroyBuffer(m_vkr->GetLogicalDevice(), buffer, nullptr);
 		return false;
@@ -511,7 +490,7 @@ bool VKRMemoryManager::CreateBuffer2(VkDeviceSize size, VkBufferUsageFlags usage
 	if (vkBindBufferMemory(m_vkr->GetLogicalDevice(), buffer, bufferMemory, 0) != VK_SUCCESS)
 	{
 		vkDestroyBuffer(m_vkr->GetLogicalDevice(), buffer, nullptr);
-		cemuLog_log(LogType::Force, "Failed to bind buffer (CreateBuffer2)");
+		cemuLog_log(LogType::Force, "Failed to bind buffer (CreateBuffer)");
 		return false;
 	}
 	return true;
@@ -533,7 +512,7 @@ bool VKRMemoryManager::CreateBufferFromHostMemory(void* hostPointer, VkDeviceSiz
 
 	if (vkCreateBuffer(m_vkr->GetLogicalDevice(), &bufferInfo, nullptr, &buffer) != VK_SUCCESS)
 	{
-		cemuLog_log(LogType::Force, "Failed to create buffer (CreateBuffer2)");
+		cemuLog_log(LogType::Force, "Failed to create buffer (CreateBuffer)");
 		return false;
 	}
 
@@ -554,7 +533,7 @@ bool VKRMemoryManager::CreateBufferFromHostMemory(void* hostPointer, VkDeviceSiz
 
 	allocInfo.pNext = &importHostMem;
 
-	if (!FindMemoryType2(memRequirements.memoryTypeBits, properties, allocInfo.memoryTypeIndex))
+	if (!FindMemoryType(memRequirements.memoryTypeBits, properties, allocInfo.memoryTypeIndex))
 	{
 		vkDestroyBuffer(m_vkr->GetLogicalDevice(), buffer, nullptr);
 		return false;
@@ -598,7 +577,7 @@ VkImageMemAllocation* VKRMemoryManager::imageMemoryAllocate(VkImage image)
 		map_textureHeap.emplace(typeFilter, texHeap);
 	}
 	else
-		texHeap = it->second;
+		texHeap = it->second.get();
 
 	// alloc mem from heap
 	uint32 allocationSize = (uint32)memRequirements.size;

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

@@ -48,6 +48,7 @@ class VkTextureChunkedHeap : private ChunkedHeap<>
 {
 public:
 	VkTextureChunkedHeap(class VKRMemoryManager* memoryManager, uint32 typeFilter) : m_vkrMemoryManager(memoryManager), m_typeFilter(typeFilter) { };
+	~VkTextureChunkedHeap();
 
 	struct ChunkInfo
 	{
@@ -148,6 +149,7 @@ class VKRSynchronizedRingAllocator
 public:
 	VKRSynchronizedRingAllocator(class VulkanRenderer* vkRenderer, class VKRMemoryManager* vkMemoryManager, VKR_BUFFER_TYPE bufferType, uint32 minimumBufferAllocSize) : m_vkr(vkRenderer), m_vkrMemMgr(vkMemoryManager), m_bufferType(bufferType), m_minimumBufferAllocSize(minimumBufferAllocSize) {};
 	VKRSynchronizedRingAllocator(const VKRSynchronizedRingAllocator&) = delete; // disallow copy
+	~VKRSynchronizedRingAllocator();
 
 	struct BufferSyncPoint_t
 	{
@@ -256,7 +258,7 @@ class VKRMemoryManager
 	}
 
 	// texture memory management
-	std::unordered_map<uint32, VkTextureChunkedHeap*> map_textureHeap; // one heap per memory type
+	std::unordered_map<uint32, std::unique_ptr<VkTextureChunkedHeap>> map_textureHeap; // one heap per memory type
 	std::vector<uint8> m_textureUploadBuffer;
 
 	// texture upload buffer
@@ -286,9 +288,7 @@ class VKRMemoryManager
 		m_vertexStrideMetalBuffer.CleanupBuffer(latestFinishedCommandBufferId);
 	}
 
-	// memory helpers
-	uint32_t FindMemoryType(uint32_t typeFilter, VkMemoryPropertyFlags properties) const;
-	bool FindMemoryType2(uint32 typeFilter, VkMemoryPropertyFlags properties, uint32& memoryIndex) const; // searches for exact properties. Can gracefully fail without throwing exception (returns false)
+	bool FindMemoryType(uint32 typeFilter, VkMemoryPropertyFlags properties, uint32& memoryIndex) const; // searches for exact properties. Can gracefully fail without throwing exception (returns false)
 	std::vector<uint32> FindMemoryTypes(uint32_t typeFilter, VkMemoryPropertyFlags properties) const;
 
 	// image memory allocation
@@ -298,8 +298,7 @@ class VKRMemoryManager
 	// buffer management
 	size_t GetTotalMemoryForBufferType(VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, size_t minimumBufferSize = 16 * 1024 * 1024);
 
-	void CreateBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, VkBuffer& buffer, VkDeviceMemory& bufferMemory) const;
-	bool CreateBuffer2(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, VkBuffer& buffer, VkDeviceMemory& bufferMemory) const; // same as CreateBuffer but doesn't throw exception on failure
+	bool CreateBuffer(VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, VkBuffer& buffer, VkDeviceMemory& bufferMemory) const; // same as CreateBuffer but doesn't throw exception on failure
 	bool CreateBufferFromHostMemory(void* hostPointer, VkDeviceSize size, VkBufferUsageFlags usage, VkMemoryPropertyFlags properties, VkBuffer& buffer, VkDeviceMemory& bufferMemory) const;
 
 	void DeleteBuffer(VkBuffer& buffer, VkDeviceMemory& deviceMem) const;

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

@@ -165,6 +165,7 @@ VKFUNC_DEVICE(vkCmdDraw);
 VKFUNC_DEVICE(vkCmdCopyBufferToImage);
 VKFUNC_DEVICE(vkCmdCopyImageToBuffer);
 VKFUNC_DEVICE(vkCmdClearColorImage);
+VKFUNC_DEVICE(vkCmdClearAttachments);
 VKFUNC_DEVICE(vkCmdBindIndexBuffer);
 VKFUNC_DEVICE(vkCmdBindVertexBuffers);
 VKFUNC_DEVICE(vkCmdDrawIndexed);
@@ -198,6 +199,7 @@ VKFUNC_DEVICE(vkCmdEndTransformFeedbackEXT);
 
 // query
 VKFUNC_DEVICE(vkCreateQueryPool);
+VKFUNC_DEVICE(vkDestroyQueryPool);
 VKFUNC_DEVICE(vkCmdResetQueryPool);
 VKFUNC_DEVICE(vkCmdBeginQuery);
 VKFUNC_DEVICE(vkCmdEndQuery);
@@ -236,6 +238,7 @@ VKFUNC_DEVICE(vkAllocateDescriptorSets);
 VKFUNC_DEVICE(vkFreeDescriptorSets);
 VKFUNC_DEVICE(vkUpdateDescriptorSets);
 VKFUNC_DEVICE(vkCreateDescriptorPool);
+VKFUNC_DEVICE(vkDestroyDescriptorPool);
 VKFUNC_DEVICE(vkDestroyDescriptorSetLayout);
 
 #undef VKFUNC_INIT