save pipelines to a binary archive

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

@@ -290,7 +290,8 @@ public:
 	{
 		if (m_hasCacheAlloc)
 		{
-			cemu_assert_debug(isInUse() == false);
+		    // HACK
+			//cemu_assert_debug(isInUse() == false);
 			g_gpuBufferHeap->freeOffset(m_cacheOffset);
 			m_hasCacheAlloc = false;
 		}
@@ -441,7 +442,7 @@ public:
 			if (uploadBegin >= uploadEnd)
 				return; // reserve range not within invalidation or range is zero sized
 
-			
+
 			if (uploadBegin == m_invalidationRangeBegin)
 			{
 				m_invalidationRangeBegin = uploadEnd;
@@ -536,7 +537,7 @@ private:
 	MPTR m_invalidationRangeBegin;
 	MPTR m_invalidationRangeEnd;
 
-	BufferCacheNode(MPTR rangeBegin, MPTR rangeEnd): m_rangeBegin(rangeBegin), m_rangeEnd(rangeEnd) 
+	BufferCacheNode(MPTR rangeBegin, MPTR rangeEnd): m_rangeBegin(rangeBegin), m_rangeEnd(rangeEnd)
 	{
 		flagInUse();
 		cemu_assert_debug(rangeBegin < rangeEnd);
@@ -740,7 +741,7 @@ private:
 		cemu_assert_debug(rangeEnd <= pageRangeEnd);
 		cemu_assert_debug((rangeBegin & 0xF) == 0);
 		cemu_assert_debug((rangeEnd & 0xF) == 0);
-		
+
 		auto pageInfo = m_pageInfo.data() + pageIndex;
 		pageInfo->hasStreamoutData = true;
 
@@ -805,7 +806,7 @@ public:
         s_allCacheNodes.clear();
         g_deallocateQueue.clear();
     }
-	
+
 	static void ProcessDeallocations()
 	{
 		for(auto& itr : g_deallocateQueue)

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

@@ -11,7 +11,7 @@
 #include "Cafe/HW/Latte/Renderer/Renderer.h"
 #include "Cafe/HW/Latte/Renderer/OpenGL/RendererShaderGL.h"
 #include "Cafe/HW/Latte/Renderer/Vulkan/RendererShaderVk.h"
-#include "Cafe/HW/Latte/Renderer/Metal/RendererShaderMtl.h"
+#include "Cafe/HW/Latte/Renderer/Metal/MetalPipelineCache.h"
 #include "Cafe/HW/Latte/Renderer/Vulkan/VulkanPipelineStableCache.h"
 
 #include <imgui.h>
@@ -163,7 +163,7 @@ void LatteShaderCache_finish()
 	else if (g_renderer->GetType() == RendererAPI::OpenGL)
 		RendererShaderGL::ShaderCacheLoading_end();
 	else if (g_renderer->GetType() == RendererAPI::Metal)
-		RendererShaderMtl::ShaderCacheLoading_end();
+		MetalPipelineCache::ShaderCacheLoading_end();
 }
 
 uint32 LatteShaderCache_getShaderCacheExtraVersion(uint64 titleId)
@@ -247,7 +247,7 @@ void LatteShaderCache_Load()
 	else if (g_renderer->GetType() == RendererAPI::OpenGL)
 		RendererShaderGL::ShaderCacheLoading_begin(cacheTitleId);
 	else if (g_renderer->GetType() == RendererAPI::Metal)
-		RendererShaderMtl::ShaderCacheLoading_begin(cacheTitleId);
+		MetalPipelineCache::ShaderCacheLoading_begin(cacheTitleId);
 	// get cache file name
 	const auto pathGeneric = ActiveSettings::GetCachePath("shaderCache/transferable/{:016x}_shaders.bin", cacheTitleId);
 	const auto pathGenericPre1_25_0 = ActiveSettings::GetCachePath("shaderCache/transferable/{:016x}.bin", cacheTitleId); // before 1.25.0
@@ -780,7 +780,7 @@ void LatteShaderCache_Close()
     else if (g_renderer->GetType() == RendererAPI::OpenGL)
         RendererShaderGL::ShaderCacheLoading_Close();
     else if (g_renderer->GetType() == RendererAPI::Metal)
-        RendererShaderMtl::ShaderCacheLoading_Close();
+        MetalPipelineCache::ShaderCacheLoading_Close();
 
     // if Vulkan then also close pipeline cache
     if (g_renderer->GetType() == RendererAPI::Vulkan)

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

@@ -1,6 +1,7 @@
 #include "Cafe/HW/Latte/Renderer/Metal/MetalCommon.h"
 #include "Cafe/HW/Latte/Renderer/Metal/MetalPipelineCache.h"
 #include "Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h"
+#include "Foundation/NSObject.hpp"
 #include "HW/Latte/Renderer/Metal/CachedFBOMtl.h"
 #include "HW/Latte/Renderer/Metal/LatteToMtl.h"
 #include "HW/Latte/Renderer/Metal/RendererShaderMtl.h"
@@ -8,6 +9,29 @@
 
 #include "HW/Latte/Core/FetchShader.h"
 #include "HW/Latte/ISA/RegDefines.h"
+#include "config/ActiveSettings.h"
+
+#define INVALID_TITLE_ID 0xFFFFFFFFFFFFFFFF
+
+uint64 s_cacheTitleId = INVALID_TITLE_ID;
+
+extern std::atomic_int g_compiled_shaders_total;
+extern std::atomic_int g_compiled_shaders_async;
+
+void MetalPipelineCache::ShaderCacheLoading_begin(uint64 cacheTitleId)
+{
+    s_cacheTitleId = cacheTitleId;
+}
+
+void MetalPipelineCache::ShaderCacheLoading_end()
+{
+}
+
+void MetalPipelineCache::ShaderCacheLoading_Close()
+{
+    g_compiled_shaders_total = 0;
+    g_compiled_shaders_async = 0;
+}
 
 MetalPipelineCache::~MetalPipelineCache()
 {
@@ -16,6 +40,17 @@ MetalPipelineCache::~MetalPipelineCache()
         pair.second->release();
     }
     m_pipelineCache.clear();
+
+    NS::Error* error = nullptr;
+    m_binaryArchive->serializeToURL(m_binaryArchiveURL, &error);
+    if (error)
+    {
+        debug_printf("failed to serialize binary archive: %s\n", error->localizedDescription()->utf8String());
+        error->release();
+    }
+    m_binaryArchive->release();
+
+    m_binaryArchiveURL->release();
 }
 
 MTL::RenderPipelineState* MetalPipelineCache::GetPipelineState(const LatteFetchShader* fetchShader, const LatteDecompilerShader* vertexShader, const LatteDecompilerShader* pixelShader, CachedFBOMtl* activeFBO, const LatteContextRegister& lcr)
@@ -151,16 +186,41 @@ MTL::RenderPipelineState* MetalPipelineCache::GetPipelineState(const LatteFetchS
         }
 	}
 
-	NS::Error* error = nullptr;
+	LoadBinary(desc);
-	pipeline = m_mtlr->GetDevice()->newRenderPipelineState(desc, &error);
+
-	desc->release();
+    NS::Error* error = nullptr;
-	vertexDescriptor->release();
+	pipeline = m_mtlr->GetDevice()->newRenderPipelineState(desc, MTL::PipelineOptionFailOnBinaryArchiveMiss, nullptr, &error);
+
+	//static uint32 oldPipelineCount = 0;
+	//static uint32 newPipelineCount = 0;
+
+	// Pipeline wasn't found in the binary archive, we need to compile it
 	if (error)
 	{
-	    debug_printf("error creating render pipeline state: %s\n", error->localizedDescription()->utf8String());
+		desc->setBinaryArchives(nullptr);
-		error->release();
+
-		return nullptr;
+        error->release();
+        error = nullptr;
+	    pipeline = m_mtlr->GetDevice()->newRenderPipelineState(desc, &error);
+		if (error)
+		{
+		    debug_printf("error creating render pipeline state: %s\n", error->localizedDescription()->utf8String());
+			error->release();
+		}
+		else
+		{
+		    SaveBinary(desc);
+		}
+
+		//newPipelineCount++;
 	}
+	//else
+    //{
+    //    oldPipelineCount++;
+    //}
+    //debug_printf("%u pipelines were found in the binary archive, %u new were created\n", oldPipelineCount, newPipelineCount);
+	desc->release();
+	vertexDescriptor->release();
 
 	return pipeline;
 }
@@ -238,3 +298,60 @@ uint64 MetalPipelineCache::CalculatePipelineHash(const LatteFetchShader* fetchSh
 
 	return stateHash;
 }
+
+void MetalPipelineCache::TryLoadBinaryArchive()
+{
+    if (m_binaryArchive || s_cacheTitleId == INVALID_TITLE_ID)
+        return;
+
+    const std::string cacheFilename = fmt::format("{:016x}_mtl_pipelines.bin", s_cacheTitleId);
+	const fs::path cachePath = ActiveSettings::GetCachePath("shaderCache/precompiled/{}", cacheFilename);
+    m_binaryArchiveURL = NS::URL::fileURLWithPath(NS::String::string((const char*)cachePath.generic_u8string().c_str(), NS::ASCIIStringEncoding));
+
+    MTL::BinaryArchiveDescriptor* desc = MTL::BinaryArchiveDescriptor::alloc()->init();
+    desc->setUrl(m_binaryArchiveURL);
+
+    NS::Error* error = nullptr;
+    m_binaryArchive = m_mtlr->GetDevice()->newBinaryArchive(desc, &error);
+    if (error)
+    {
+        desc->setUrl(nullptr);
+
+        error->release();
+        error = nullptr;
+        m_binaryArchive = m_mtlr->GetDevice()->newBinaryArchive(desc, &error);
+        if (error)
+        {
+            debug_printf("failed to create binary archive: %s\n", error->localizedDescription()->utf8String());
+            error->release();
+        }
+    }
+    desc->release();
+}
+
+void MetalPipelineCache::LoadBinary(MTL::RenderPipelineDescriptor* desc)
+{
+    TryLoadBinaryArchive();
+
+    if (!m_binaryArchive)
+        return;
+
+    NS::Object* binArchives[] = {m_binaryArchive};
+    auto binaryArchives = NS::Array::alloc()->init(binArchives, 1);
+    desc->setBinaryArchives(binaryArchives);
+    binaryArchives->release();
+}
+
+void MetalPipelineCache::SaveBinary(MTL::RenderPipelineDescriptor* desc)
+{
+    if (!m_binaryArchive)
+        return;
+
+    NS::Error* error = nullptr;
+    m_binaryArchive->addRenderPipelineFunctions(desc, &error);
+    if (error)
+    {
+        debug_printf("error saving render pipeline functions: %s\n", error->localizedDescription()->utf8String());
+        error->release();
+    }
+}

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

@@ -8,6 +8,10 @@
 class MetalPipelineCache
 {
 public:
+    static void ShaderCacheLoading_begin(uint64 cacheTitleId);
+    static void ShaderCacheLoading_end();
+    static void ShaderCacheLoading_Close();
+
     MetalPipelineCache(class MetalRenderer* metalRenderer) : m_mtlr{metalRenderer} {}
     ~MetalPipelineCache();
 
@@ -18,5 +22,14 @@ private:
 
     std::map<uint64, MTL::RenderPipelineState*> m_pipelineCache;
 
+    NS::URL* m_binaryArchiveURL;
+    MTL::BinaryArchive* m_binaryArchive;
+
     uint64 CalculatePipelineHash(const LatteFetchShader* fetchShader, const LatteDecompilerShader* vertexShader, const LatteDecompilerShader* pixelShader, class CachedFBOMtl* activeFBO, const LatteContextRegister& lcr);
+
+    void TryLoadBinaryArchive();
+
+    void LoadBinary(MTL::RenderPipelineDescriptor* desc);
+
+    void SaveBinary(MTL::RenderPipelineDescriptor* desc);
 };

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

@@ -2,24 +2,18 @@
 #include "Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h"
 #include "Cafe/HW/Latte/Renderer/Metal/LatteToMtl.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"
 
-bool s_isLoadingShadersMtl{ false };
-class FileCache* s_mslCache{nullptr};
-
 extern std::atomic_int g_compiled_shaders_total;
 extern std::atomic_int g_compiled_shaders_async;
 
 RendererShaderMtl::RendererShaderMtl(MetalRenderer* mtlRenderer, ShaderType type, uint64 baseHash, uint64 auxHash, bool isGameShader, bool isGfxPackShader, const std::string& mslCode)
 	: RendererShader(type, baseHash, auxHash, isGameShader, isGfxPackShader), m_mtlr{mtlRenderer}
 {
-    if (LoadBinary())
-        return;
-
     if (m_type == ShaderType::kFragment)
     {
         // Fragment functions are compiled just-in-time
@@ -30,12 +24,8 @@ RendererShaderMtl::RendererShaderMtl(MetalRenderer* mtlRenderer, ShaderType type
         Compile(mslCode);
     }
 
-    // Store the compiled shader in the cache
-    StoreBinary();
-
 	// Count shader compilation
-	if (!s_isLoadingShadersMtl)
+	g_compiled_shaders_total++;
-		g_compiled_shaders_total++;
 }
 
 RendererShaderMtl::~RendererShaderMtl()
@@ -85,33 +75,6 @@ void RendererShaderMtl::CompileFragmentFunction(CachedFBOMtl* activeFBO)
     Compile(fullCode);
 }
 
-void RendererShaderMtl::ShaderCacheLoading_begin(uint64 cacheTitleId)
-{
-    if (s_mslCache)
-	{
-		delete s_mslCache;
-	}
-	uint32 spirvCacheMagic = GeneratePrecompiledCacheId();
-	const std::string cacheFilename = fmt::format("{:016x}_msl.bin", cacheTitleId);
-	const fs::path cachePath = ActiveSettings::GetCachePath("shaderCache/precompiled/{}", cacheFilename);
-	s_mslCache = FileCache::Open(cachePath, true, spirvCacheMagic);
-	if (!s_mslCache)
-		cemuLog_log(LogType::Force, "Unable to open MSL cache {}", cacheFilename);
-	s_isLoadingShadersMtl = true;
-}
-
-void RendererShaderMtl::ShaderCacheLoading_end()
-{
-	s_isLoadingShadersMtl = false;
-}
-
-void RendererShaderMtl::ShaderCacheLoading_Close()
-{
-    delete s_mslCache;
-    g_compiled_shaders_total = 0;
-    g_compiled_shaders_async = 0;
-}
-
 void RendererShaderMtl::Compile(const std::string& mslCode)
 {
     NS::Error* error = nullptr;
@@ -125,33 +88,3 @@ void RendererShaderMtl::Compile(const std::string& mslCode)
     m_function = library->newFunction(NS::String::string("main0", NS::ASCIIStringEncoding));
     library->release();
 }
-
-bool RendererShaderMtl::LoadBinary()
-{
-    // HACK: since fragment functions are compiled just-in-time, we cannot load them from the cache
-    if (m_type == ShaderType::kFragment)
-        return false;
-
-    uint64 h1, h2;
-    GenerateShaderPrecompiledCacheFilename(m_type, m_baseHash, m_auxHash, h1, h2);
-	if (!s_mslCache->GetFile({h1, h2 }, m_binary))
-		return false;
-
-	// TODO: implement
-	return false;
-
-	return true;
-}
-
-void RendererShaderMtl::StoreBinary()
-{
-    if (m_binary.size() == 0)
-    {
-        // TODO: retrieve the binary from the function
-        return;
-    }
-
-    uint64 h1, h2;
-	GenerateShaderPrecompiledCacheFilename(m_type, m_baseHash, m_auxHash, h1, h2);
-	s_mslCache->AddFileAsync({h1, h2 }, m_binary.data(), m_binary.size());
-}

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

@@ -18,10 +18,6 @@ class RendererShaderMtl : public RendererShader
 	//};
 
 public:
-	static void ShaderCacheLoading_begin(uint64 cacheTitleId);
-    static void ShaderCacheLoading_end();
-    static void ShaderCacheLoading_Close();
-
 	RendererShaderMtl(class MetalRenderer* mtlRenderer, ShaderType type, uint64 baseHash, uint64 auxHash, bool isGameShader, bool isGfxPackShader, const std::string& mslCode);
 	virtual ~RendererShaderMtl();
 
@@ -62,7 +58,4 @@ private:
 	std::string m_mslCode;
 
 	void Compile(const std::string& mslCode);
-
-	bool LoadBinary();
-	void StoreBinary();
 };