add an option to use the host memory instead of buffer cache

src/Cafe/CafeSystem.cpp

@@ -258,6 +258,7 @@ void InfoLog_PrintActiveSettings()
 	{
 	    cemuLog_log(LogType::Force, "Async compile: {}", GetConfig().async_compile.GetValue() ? "true" : "false");
 		cemuLog_log(LogType::Force, "Fast math: {}", GetConfig().fast_math.GetValue() ? "true" : "false");
+		cemuLog_log(LogType::Force, "Use host memory for cache: {}", g_current_game_profile->UseHostMemForCache() ? "true" : "false");
 		if (!GetConfig().vk_accurate_barriers.GetValue())
 			cemuLog_log(LogType::Force, "Accurate barriers are disabled!");
 	}

src/Cafe/GameProfile/GameProfile.cpp

@@ -127,7 +127,7 @@ bool gameProfile_loadIntegerOption(IniParser& iniParser, const char* optionName,
 	{
 		cemuLog_log(LogType::Force, "Value '{}' is out of range for option '{}' in game profile", *option_value, optionName);
 		return false;
-	}	
+	}
 }
 
 template<typename T>
@@ -224,8 +224,9 @@ bool GameProfile::Load(uint64_t title_id)
 			gameProfile_loadIntegerOption(&iniParser, "graphics_api", &graphicsApi, -1, 0, 1);
 			if (graphicsApi.value != -1)
 				m_graphics_api = (GraphicAPI)graphicsApi.value;
-			
+
 			gameProfile_loadEnumOption(iniParser, "accurateShaderMul", m_accurateShaderMul);
+			gameProfile_loadBooleanOption2(iniParser, "useHostMemForCache", m_useHostMemForCache);
 
 			// legacy support
 			auto option_precompiledShaders = iniParser.FindOption("precompiledShaders");
@@ -277,7 +278,7 @@ bool GameProfile::Load(uint64_t title_id)
 void GameProfile::Save(uint64_t title_id)
 {
 	auto gameProfileDir = ActiveSettings::GetConfigPath("gameProfiles");
-	if (std::error_code ex_ec; !fs::exists(gameProfileDir, ex_ec)) 
+	if (std::error_code ex_ec; !fs::exists(gameProfileDir, ex_ec))
 		fs::create_directories(gameProfileDir, ex_ec);
 	auto gameProfilePath = gameProfileDir / fmt::format("{:016x}.ini", title_id);
 	FileStream* fs = FileStream::createFile2(gameProfilePath);
@@ -308,6 +309,7 @@ void GameProfile::Save(uint64_t title_id)
 
 	fs->writeLine("[Graphics]");
 	WRITE_ENTRY(accurateShaderMul);
+	WRITE_ENTRY(useHostMemForCache);
 	WRITE_OPTIONAL_ENTRY(precompiledShaders);
 	WRITE_OPTIONAL_ENTRY(graphics_api);
 	fs->writeLine("");
@@ -337,6 +339,7 @@ void GameProfile::ResetOptional()
 
 	// graphic settings
 	m_accurateShaderMul = AccurateShaderMulOption::True;
+	m_useHostMemForCache = false;
 	// cpu settings
 	m_threadQuantum = kThreadQuantumDefault;
 	m_cpuMode.reset(); // CPUModeOption::kSingleCoreRecompiler;
@@ -354,9 +357,10 @@ void GameProfile::Reset()
 	// general settings
 	m_loadSharedLibraries = true;
 	m_startWithPadView = false;
-	
+
 	// graphic settings
 	m_accurateShaderMul = AccurateShaderMulOption::True;
+	m_useHostMemForCache = false;
 	m_precompiledShaders = PrecompiledShaderOption::Auto;
 	// cpu settings
 	m_threadQuantum = kThreadQuantumDefault;
@@ -366,4 +370,4 @@ void GameProfile::Reset()
 	// controller settings
 	for (auto& profile : m_controllerProfile)
 		profile.reset();
-}
+}

src/Cafe/GameProfile/GameProfile.h

@@ -31,6 +31,7 @@ class GameProfile
 
 	[[nodiscard]] const std::optional<GraphicAPI>& GetGraphicsAPI() const { return m_graphics_api; }
 	[[nodiscard]] const AccurateShaderMulOption& GetAccurateShaderMul() const { return m_accurateShaderMul; }
+	[[nodiscard]] bool UseHostMemForCache() const { return m_useHostMemForCache; }
 	[[nodiscard]] const std::optional<PrecompiledShaderOption>& GetPrecompiledShadersState() const { return m_precompiledShaders; }
 
 	[[nodiscard]] uint32 GetThreadQuantum() const { return m_threadQuantum; }
@@ -54,6 +55,7 @@ class GameProfile
 	// graphic settings
 	std::optional<GraphicAPI> m_graphics_api{};
 	AccurateShaderMulOption m_accurateShaderMul = AccurateShaderMulOption::True;
+	bool m_useHostMemForCache = false;
 	std::optional<PrecompiledShaderOption> m_precompiledShaders{};
 	// cpu settings
 	uint32 m_threadQuantum = kThreadQuantumDefault; // values: 20000 45000 60000 80000 100000

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

@@ -1,8 +1,10 @@
 #include "Cafe/HW/Latte/Renderer/Metal/MetalCommon.h"
 #include "Cafe/HW/Latte/Renderer/Metal/MetalMemoryManager.h"
 #include "Cafe/HW/Latte/Renderer/Metal/MetalVoidVertexPipeline.h"
+
 #include "Cemu/Logging/CemuLogging.h"
 #include "Common/precompiled.h"
+#include "GameProfile/GameProfile.h"
 
 /*
 MetalVertexBufferCache::~MetalVertexBufferCache()
@@ -117,8 +119,7 @@ void MetalMemoryManager::InitBufferCache(size_t size)
     cemu_assert_debug(!m_bufferCache);
 
     // First, try to import the host memory as a buffer
-    // TODO: only import if the option is ticked in game profile
-    if (m_mtlr->IsAppleGPU())
+    if (g_current_game_profile->UseHostMemForCache() && m_mtlr->IsAppleGPU())
     {
         m_importedMemBaseAddress = 0x10000000;
     	size_t hostAllocationSize = 0x40000000ull;
@@ -165,7 +166,6 @@ void MetalMemoryManager::UploadToBufferCache(const void* data, size_t offset, si
 
 void MetalMemoryManager::CopyBufferCache(size_t srcOffset, size_t dstOffset, size_t size)
 {
-    cemu_assert_debug(!m_useHostMemoryForCache);
     cemu_assert_debug(m_bufferCache);
 
     m_mtlr->CopyBufferToBuffer(m_bufferCache, srcOffset, m_bufferCache, dstOffset, size, ALL_MTL_RENDER_STAGES, ALL_MTL_RENDER_STAGES);

src/gui/GameProfileWindow.cpp

@@ -127,6 +127,13 @@ GameProfileWindow::GameProfileWindow(wxWindow* parent, uint64_t title_id)
 		m_shader_mul_accuracy->SetToolTip(_("EXPERT OPTION\nControls the accuracy of floating point multiplication in shaders.\n\nRecommended: true"));
 		first_row->Add(m_shader_mul_accuracy, 0, wxALL, 5);
 
+		first_row->Add(new wxStaticText(panel, wxID_ANY, _("Use host memory for cache")), 0, wxALIGN_CENTER_VERTICAL | wxALL, 5);
+
+		wxString mem_values[] = { _("false"), _("true")};
+		m_use_host_mem_for_cache = new wxChoice(panel, wxID_ANY, wxDefaultPosition, wxDefaultSize, (int)std::size(mem_values), mem_values);
+		m_use_host_mem_for_cache->SetToolTip(_("EXPERT OPTION\nAllows the GPU to access data directly without the need for an intermediate cache. May increase performance and reduce memory usage, but can also cause flickering.\n\nMetal only\n\nRecommended: false"));
+		first_row->Add(m_use_host_mem_for_cache, 0, wxALL, 5);
+
 		/*first_row->Add(new wxStaticText(panel, wxID_ANY, _("GPU buffer cache accuracy")), 0, wxALIGN_CENTER_VERTICAL | wxALL, 5);
 		wxString accuarcy_values[] = { _("high"), _("medium"), _("low") };
 		m_cache_accuracy = new wxChoice(panel, wxID_ANY, wxDefaultPosition, wxDefaultSize, (int)std::size(accuarcy_values), accuarcy_values);
@@ -273,6 +280,7 @@ void GameProfileWindow::ApplyProfile()
 	else
 		m_graphic_api->SetSelection(1 + m_game_profile.m_graphics_api.value()); // "", OpenGL, Vulkan, Metal
 	m_shader_mul_accuracy->SetSelection((int)m_game_profile.m_accurateShaderMul);
+	m_use_host_mem_for_cache->SetSelection((int)m_game_profile.m_useHostMemForCache);
 
 	//// audio
 	//m_disable_audio->Set3StateValue(GetCheckboxState(m_game_profile.disableAudio));
@@ -332,6 +340,7 @@ void GameProfileWindow::SaveProfile()
 
 	// gpu
 	m_game_profile.m_accurateShaderMul = (AccurateShaderMulOption)m_shader_mul_accuracy->GetSelection();
+	m_game_profile.m_useHostMemForCache = (bool)m_use_host_mem_for_cache->GetSelection();
 	if (m_game_profile.m_accurateShaderMul != AccurateShaderMulOption::False && m_game_profile.m_accurateShaderMul != AccurateShaderMulOption::True)
 		m_game_profile.m_accurateShaderMul = AccurateShaderMulOption::True; // force a legal value
 

src/gui/GameProfileWindow.h

@@ -40,11 +40,12 @@ class GameProfileWindow : public wxFrame
 	wxChoice* m_graphic_api;
 
 	wxChoice* m_shader_mul_accuracy;
+	wxChoice* m_use_host_mem_for_cache;
 	//wxChoice* m_cache_accuracy;
 
 	// audio
 	//wxCheckBox* m_disable_audio;
 
 	// controller
 	wxComboBox* m_controller_profile[8];
-};
+};