Merge pull request #18 from SamoZ256/metal-air-cache

AIR cache

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

@@ -524,6 +524,7 @@ void LatteSHRC_UpdateVSBaseHash(uint8* vertexShaderPtr, uint32 vertexShaderSize,
 	if (LatteGPUState.contextNew.PA_CL_CLIP_CNTL.get_DX_CLIP_SPACE_DEF())
 		vsHash += 0x1537;
 
+#if ENABLE_METAL
 	if (g_renderer->GetType() == RendererAPI::Metal)
 	{
 	    if (usesGeometryShader || _activeFetchShader->mtlFetchVertexManually)
@@ -542,27 +543,28 @@ void LatteSHRC_UpdateVSBaseHash(uint8* vertexShaderPtr, uint32 vertexShaderSize,
 
 	    if (!usesGeometryShader)
 		{
-  		// Rasterization
+      		// Rasterization
-  		bool rasterizationEnabled = !LatteGPUState.contextNew.PA_CL_CLIP_CNTL.get_DX_RASTERIZATION_KILL();
+      		bool rasterizationEnabled = !LatteGPUState.contextNew.PA_CL_CLIP_CNTL.get_DX_RASTERIZATION_KILL();
 
-  		// HACK
+      		// HACK
-  		if (!LatteGPUState.contextNew.PA_CL_VTE_CNTL.get_VPORT_X_OFFSET_ENA())
+      		if (!LatteGPUState.contextNew.PA_CL_VTE_CNTL.get_VPORT_X_OFFSET_ENA())
- 			rasterizationEnabled = true;
+     			rasterizationEnabled = true;
 
-  		const auto& polygonControlReg = LatteGPUState.contextNew.PA_SU_SC_MODE_CNTL;
+      		const auto& polygonControlReg = LatteGPUState.contextNew.PA_SU_SC_MODE_CNTL;
-  		uint32 cullFront = polygonControlReg.get_CULL_FRONT();
+      		uint32 cullFront = polygonControlReg.get_CULL_FRONT();
-  		uint32 cullBack = polygonControlReg.get_CULL_BACK();
+      		uint32 cullBack = polygonControlReg.get_CULL_BACK();
-  		if (cullFront && cullBack)
+      		if (cullFront && cullBack)
-  		    rasterizationEnabled = false;
+      		    rasterizationEnabled = false;
 
-  		if (rasterizationEnabled)
+      		if (rasterizationEnabled)
-  		    vsHash += 51ULL;
+      		    vsHash += 51ULL;
 
-        // Vertex fetch
+            // Vertex fetch
-        if (_activeFetchShader->mtlFetchVertexManually)
+            if (_activeFetchShader->mtlFetchVertexManually)
-            vsHash += 349ULL;
+                vsHash += 349ULL;
 		}
 	}
+#endif
 
 	_shaderBaseHash_vs = vsHash;
 }

src/Cafe/HW/Latte/Renderer/Metal/LatteTextureMtl.cpp

@@ -2,9 +2,6 @@
 #include "Cafe/HW/Latte/Renderer/Metal/LatteTextureViewMtl.h"
 #include "Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h"
 #include "Cafe/HW/Latte/Renderer/Metal/LatteToMtl.h"
-#include "Common/precompiled.h"
-#include "Metal/MTLResource.hpp"
-#include "Metal/MTLTexture.hpp"
 
 LatteTextureMtl::LatteTextureMtl(class MetalRenderer* mtlRenderer, Latte::E_DIM dim, MPTR physAddress, MPTR physMipAddress, Latte::E_GX2SURFFMT format, uint32 width, uint32 height, uint32 depth, uint32 pitch, uint32 mipLevels, uint32 swizzle,
 	Latte::E_HWTILEMODE tileMode, bool isDepth)
@@ -12,7 +9,7 @@ LatteTextureMtl::LatteTextureMtl(class MetalRenderer* mtlRenderer, Latte::E_DIM
 {
     MTL::TextureDescriptor* desc = MTL::TextureDescriptor::alloc()->init();
     desc->setStorageMode(MTL::StorageModePrivate);
-    desc->setCpuCacheMode(MTL::CPUCacheModeWriteCombined);
+    //desc->setCpuCacheMode(MTL::CPUCacheModeWriteCombined);
 
 	sint32 effectiveBaseWidth = width;
 	sint32 effectiveBaseHeight = height;

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

@@ -101,3 +101,79 @@ inline bool FormatIsRenderable(Latte::E_GX2SURFFMT format)
 {
     return !Latte::IsCompressedFormat(format);
 }
+
+template <typename... T>
+inline bool executeCommand(fmt::format_string<T...> fmt, T&&... args) {
+    std::string command = fmt::format(fmt, std::forward<T>(args)...);
+    int res = system(command.c_str());
+    if (res != 0)
+    {
+        cemuLog_log(LogType::Force, "command \"{}\" failed with exit code {}", command, res);
+        return false;
+    }
+
+    return true;
+}
+
+class MemoryMappedFile
+{
+public:
+    MemoryMappedFile(const std::string& filePath)
+    {
+        // Open the file
+        m_fd = open(filePath.c_str(), O_RDONLY);
+        if (m_fd == -1) {
+            cemuLog_log(LogType::Force, "failed to open file: {}", filePath);
+            return;
+        }
+
+        // Get the file size
+        // Use a loop to handle the case where the file size is 0 (more of a safety net)
+        struct stat fileStat;
+        while (true)
+        {
+            if (fstat(m_fd, &fileStat) == -1)
+            {
+                close(m_fd);
+                cemuLog_log(LogType::Force, "failed to get file size: {}", filePath);
+                return;
+            }
+            m_fileSize = fileStat.st_size;
+
+            if (m_fileSize == 0)
+            {
+                cemuLog_logOnce(LogType::Force, "file size is 0: {}", filePath);
+                std::this_thread::sleep_for(std::chrono::milliseconds(10));
+                continue;
+            }
+
+            break;
+        }
+
+        // Memory map the file
+        m_data = mmap(nullptr, m_fileSize, PROT_READ, MAP_PRIVATE, m_fd, 0);
+        if (m_data == MAP_FAILED)
+        {
+            close(m_fd);
+            cemuLog_log(LogType::Force, "failed to memory map file: {}", filePath);
+            return;
+        }
+    }
+
+    ~MemoryMappedFile()
+    {
+        if (m_data && m_data != MAP_FAILED)
+            munmap(m_data, m_fileSize);
+
+        if (m_fd != -1)
+            close(m_fd);
+    }
+
+    uint8* data() const { return static_cast<uint8*>(m_data); }
+    size_t size() const { return m_fileSize; }
+
+private:
+    int m_fd = -1;
+    void* m_data = nullptr;
+    size_t m_fileSize = 0;
+};

src/Cafe/HW/Latte/Renderer/Metal/RendererShaderMtl.cpp

@@ -2,14 +2,21 @@
 #include "Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h"
 #include "Cafe/HW/Latte/Renderer/Metal/MetalCommon.h"
 
-//#include "Cemu/FileCache/FileCache.h"
+#include "Cemu/FileCache/FileCache.h"
-//#include "config/ActiveSettings.h"
+#include "config/ActiveSettings.h"
 #include "Cemu/Logging/CemuLogging.h"
 #include "Common/precompiled.h"
 #include "GameProfile/GameProfile.h"
 #include "util/helpers/helpers.h"
 
+#define METAL_AIR_CACHE_NAME "Cemu_AIR_cache"
+#define METAL_AIR_CACHE_PATH "/Volumes/" METAL_AIR_CACHE_NAME
+#define METAL_AIR_CACHE_SIZE (16 * 1024 * 1024)
+#define METAL_AIR_CACHE_BLOCK_COUNT (METAL_AIR_CACHE_SIZE / 512)
+
 static bool s_isLoadingShadersMtl{false};
+static bool s_hasRAMFilesystem{false};
+class FileCache* s_airCache{nullptr};
 
 extern std::atomic_int g_compiled_shaders_total;
 extern std::atomic_int g_compiled_shaders_async;
@@ -21,10 +28,14 @@ public:
 	{
 		if (m_threadsActive.exchange(true))
 			return;
-		// create thread pool
+
+		// Create thread pool
 		const uint32 threadCount = 2;
 		for (uint32 i = 0; i < threadCount; ++i)
 			s_threads.emplace_back(&ShaderMtlThreadPool::CompilerThreadFunc, this);
+
+		// Create AIR cache thread
+		s_airCacheThread = new std::thread(&ShaderMtlThreadPool::AIRCacheThreadFunc, this);
 	}
 
 	void StopThreads()
@@ -36,6 +47,9 @@ public:
 		for (auto& it : s_threads)
 			it.join();
 		s_threads.clear();
+
+		s_airCacheThread->join();
+		delete s_airCacheThread;
 	}
 
 	~ShaderMtlThreadPool()
@@ -72,15 +86,48 @@ public:
 		}
 	}
 
+	void AIRCacheThreadFunc()
+    {
+        SetThreadName("mtlAIRCache");
+        while (m_threadsActive.load(std::memory_order::relaxed))
+        {
+            s_airCacheQueueCount.decrementWithWait();
+            s_airCacheQueueMutex.lock();
+            if (s_airCacheQueue.empty())
+            {
+                s_airCacheQueueMutex.unlock();
+                continue;
+            }
+
+            // Create RAM filesystem
+            if (!s_hasRAMFilesystem)
+            {
+                executeCommand("diskutil erasevolume HFS+ {} $(hdiutil attach -nomount ram://{})", METAL_AIR_CACHE_NAME, METAL_AIR_CACHE_BLOCK_COUNT);
+                s_hasRAMFilesystem = true;
+            }
+
+            RendererShaderMtl* job = s_airCacheQueue.front();
+            s_airCacheQueue.pop_front();
+            s_airCacheQueueMutex.unlock();
+            // compile
+            job->CompileToAIR();
+        }
+    }
+
 	bool HasThreadsRunning() const { return m_threadsActive; }
 
 public:
 	std::vector<std::thread> s_threads;
+	std::thread* s_airCacheThread{nullptr};
 
 	std::deque<RendererShaderMtl*> s_compilationQueue;
 	CounterSemaphore s_compilationQueueCount;
 	std::mutex s_compilationQueueMutex;
 
+	std::deque<RendererShaderMtl*> s_airCacheQueue;
+	CounterSemaphore s_airCacheQueueCount;
+	std::mutex s_airCacheQueueMutex;
+
 private:
 	std::atomic<bool> m_threadsActive;
 } shaderMtlThreadPool;
@@ -88,18 +135,45 @@ private:
 // TODO: find out if it would be possible to cache compiled Metal shaders
 void RendererShaderMtl::ShaderCacheLoading_begin(uint64 cacheTitleId)
 {
+    s_isLoadingShadersMtl = true;
+
+    // Open AIR cache
+    if (s_airCache)
+	{
+		delete s_airCache;
+		s_airCache = nullptr;
+	}
+	uint32 airCacheMagic = GeneratePrecompiledCacheId();
+	const std::string cacheFilename = fmt::format("{:016x}_air.bin", cacheTitleId);
+	const fs::path cachePath = ActiveSettings::GetCachePath("shaderCache/precompiled/{}", cacheFilename);
+	s_airCache = FileCache::Open(cachePath, true, airCacheMagic);
+	if (!s_airCache)
+		cemuLog_log(LogType::Force, "Unable to open AIR cache {}", cacheFilename);
+
     // Maximize shader compilation speed
     static_cast<MetalRenderer*>(g_renderer.get())->SetShouldMaximizeConcurrentCompilation(true);
 }
 
 void RendererShaderMtl::ShaderCacheLoading_end()
 {
+    s_isLoadingShadersMtl = false;
+
+    // Reset shader compilation speed
     static_cast<MetalRenderer*>(g_renderer.get())->SetShouldMaximizeConcurrentCompilation(false);
 }
 
 void RendererShaderMtl::ShaderCacheLoading_Close()
 {
-    // Do nothing
+    // Close the AIR cache
+    if (s_airCache)
+    {
+        delete s_airCache;
+        s_airCache = nullptr;
+    }
+
+    // Close RAM filesystem
+    if (s_hasRAMFilesystem)
+        executeCommand("diskutil eject {}", METAL_AIR_CACHE_PATH);
 }
 
 void RendererShaderMtl::Initialize()
@@ -172,10 +246,10 @@ bool RendererShaderMtl::ShouldCountCompilation() const
     return !s_isLoadingShadersMtl && m_isGameShader;
 }
 
-void RendererShaderMtl::CompileInternal()
+MTL::Library* RendererShaderMtl::LibraryFromSource()
 {
+    // Compile from source
     MTL::CompileOptions* options = MTL::CompileOptions::alloc()->init();
-    // TODO: always disable fast math for problematic shaders
     if (g_current_game_profile->GetFastMath())
         options->setFastMathEnabled(true);
     if (g_current_game_profile->GetPositionInvariance())
@@ -186,20 +260,107 @@ void RendererShaderMtl::CompileInternal()
 	options->release();
 	if (error)
     {
-        cemuLog_log(LogType::Force, "failed to create library: {} -> {}", error->localizedDescription()->utf8String(), m_mslCode.c_str());
+        cemuLog_log(LogType::Force, "failed to create library from source: {} -> {}", error->localizedDescription()->utf8String(), m_mslCode.c_str());
-        FinishCompilation();
+        return nullptr;
-        return;
     }
+
+    return library;
+}
+
+MTL::Library* RendererShaderMtl::LibraryFromAIR(std::span<uint8> data)
+{
+    dispatch_data_t dispatchData = dispatch_data_create(data.data(), data.size(), nullptr, DISPATCH_DATA_DESTRUCTOR_DEFAULT);
+
+    NS::Error* error = nullptr;
+	MTL::Library* library = m_mtlr->GetDevice()->newLibrary(dispatchData, &error);
+	if (error)
+    {
+        cemuLog_log(LogType::Force, "failed to create library from AIR: {}", error->localizedDescription()->utf8String());
+        return nullptr;
+    }
+
+    return library;
+}
+
+void RendererShaderMtl::CompileInternal()
+{
+    MTL::Library* library = nullptr;
+
+    // First, try to retrieve the compiled shader from the AIR cache
+    if (s_isLoadingShadersMtl && (m_isGameShader && !m_isGfxPackShader) && s_airCache)
+    {
+        cemu_assert_debug(m_baseHash != 0);
+		uint64 h1, h2;
+		GenerateShaderPrecompiledCacheFilename(m_type, m_baseHash, m_auxHash, h1, h2);
+		std::vector<uint8> cacheFileData;
+		if (s_airCache->GetFile({ h1, h2 }, cacheFileData))
+		{
+			library = LibraryFromAIR(std::span<uint8>(cacheFileData.data(), cacheFileData.size()));
+			FinishCompilation();
+		}
+    }
+
+    // Not in the cache, compile from source
+    if (!library)
+    {
+        // Compile from source
+        library = LibraryFromSource();
+        if (!library)
+            return;
+
+        // Store in the AIR cache
+        shaderMtlThreadPool.s_airCacheQueueMutex.lock();
+        shaderMtlThreadPool.s_airCacheQueue.push_back(this);
+        shaderMtlThreadPool.s_airCacheQueueCount.increment();
+        shaderMtlThreadPool.s_airCacheQueueMutex.unlock();
+    }
+
     m_function = library->newFunction(ToNSString("main0"));
     library->release();
 
-    FinishCompilation();
-
 	// Count shader compilation
 	if (ShouldCountCompilation())
 	    g_compiled_shaders_total++;
 }
 
+void RendererShaderMtl::CompileToAIR()
+{
+    uint64 h1, h2;
+	GenerateShaderPrecompiledCacheFilename(m_type, m_baseHash, m_auxHash, h1, h2);
+
+    // The shader is not in the cache, compile it
+	std::string baseFilename = fmt::format("{}/{}_{}", METAL_AIR_CACHE_PATH, h1, h2);
+
+	// Source
+	std::ofstream mslFile;
+    mslFile.open(fmt::format("{}.metal", baseFilename));
+    mslFile << m_mslCode;
+    mslFile.close();
+
+    // Compile
+	if (!executeCommand("xcrun -sdk macosx metal -o {}.ir -c {}.metal -w", baseFilename, baseFilename))
+	    return;
+	if (!executeCommand("xcrun -sdk macosx metallib -o {}.metallib {}.ir", baseFilename, baseFilename))
+        return;
+
+	// Clean up
+	executeCommand("rm {}.metal", baseFilename);
+	executeCommand("rm {}.ir", baseFilename);
+
+	// Load from the newly generated AIR
+	MemoryMappedFile airFile(fmt::format("{}.metallib", baseFilename));
+	std::span<uint8> airData = std::span<uint8>(airFile.data(), airFile.size());
+	//library = LibraryFromAIR(std::span<uint8>(airData.data(), airData.size()));
+
+	// Store in the cache
+	s_airCache->AddFile({ h1, h2 }, airData.data(), airData.size());
+
+	// Clean up
+	executeCommand("rm {}.metallib", baseFilename);
+
+	FinishCompilation();
+}
+
 void RendererShaderMtl::FinishCompilation()
 {
     m_mslCode.clear();

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

@@ -67,7 +67,13 @@ private:
 
 	bool ShouldCountCompilation() const;
 
+	MTL::Library* LibraryFromSource();
+
+	MTL::Library* LibraryFromAIR(std::span<uint8> data);
+
 	void CompileInternal();
 
+	void CompileToAIR();
+
 	void FinishCompilation();
 };

src/gui/components/wxGameList.cpp

@@ -69,6 +69,7 @@ std::list<fs::path> _getCachesPaths(const TitleId& titleId)
 		ActiveSettings::GetCachePath(L"shaderCache/driver/vk/{:016x}.bin", titleId),
 		ActiveSettings::GetCachePath(L"shaderCache/precompiled/{:016x}_spirv.bin", titleId),
 		ActiveSettings::GetCachePath(L"shaderCache/precompiled/{:016x}_gl.bin", titleId),
+		ActiveSettings::GetCachePath(L"shaderCache/precompiled/{:016x}_air.bin", titleId),
 		ActiveSettings::GetCachePath(L"shaderCache/transferable/{:016x}_shaders.bin", titleId),
 		ActiveSettings::GetCachePath(L"shaderCache/transferable/{:016x}_mtlshaders.bin", titleId),
 		ActiveSettings::GetCachePath(L"shaderCache/transferable/{:016x}_vkpipeline.bin", titleId),