"Merge Tagged PR 1340"

texture_cache: Implement depth stencil texture swizzles

src/common/CMakeLists.txt

@@ -123,6 +123,8 @@ add_library(common STATIC
     lz4_compression.cpp
     lz4_compression.h
     math_util.h
+    memory_detect.cpp
+    memory_detect.h
     memory_hook.cpp
     memory_hook.h
     microprofile.cpp

src/common/memory_detect.cpp

@@ -0,0 +1,60 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#ifdef _WIN32
+// clang-format off
+#include <windows.h>
+#include <sysinfoapi.h>
+// clang-format on
+#else
+#include <sys/types.h>
+#ifdef __APPLE__
+#include <sys/sysctl.h>
+#else
+#include <sys/sysinfo.h>
+#endif
+#endif
+
+#include "common/memory_detect.h"
+
+namespace Common {
+
+// Detects the RAM and Swapfile sizes
+static MemoryInfo Detect() {
+    MemoryInfo mem_info{};
+
+#ifdef _WIN32
+    MEMORYSTATUSEX memorystatus;
+    memorystatus.dwLength = sizeof(memorystatus);
+    GlobalMemoryStatusEx(&memorystatus);
+    mem_info.TotalPhysicalMemory = memorystatus.ullTotalPhys;
+    mem_info.TotalSwapMemory = memorystatus.ullTotalPageFile - mem_info.TotalPhysicalMemory;
+#elif defined(__APPLE__)
+    u64 ramsize;
+    struct xsw_usage vmusage;
+    std::size_t sizeof_ramsize = sizeof(ramsize);
+    std::size_t sizeof_vmusage = sizeof(vmusage);
+    // hw and vm are defined in sysctl.h
+    // https://github.com/apple/darwin-xnu/blob/master/bsd/sys/sysctl.h#L471
+    // sysctlbyname(const char *, void *, size_t *, void *, size_t);
+    sysctlbyname("hw.memsize", &ramsize, &sizeof_ramsize, NULL, 0);
+    sysctlbyname("vm.swapusage", &vmusage, &sizeof_vmusage, NULL, 0);
+    mem_info.TotalPhysicalMemory = ramsize;
+    mem_info.TotalSwapMemory = vmusage.xsu_total;
+#else
+    struct sysinfo meminfo;
+    sysinfo(&meminfo);
+    mem_info.TotalPhysicalMemory = meminfo.totalram;
+    mem_info.TotalSwapMemory = meminfo.totalswap;
+#endif
+
+    return mem_info;
+}
+
+const MemoryInfo& GetMemInfo() {
+    static MemoryInfo mem_info = Detect();
+    return mem_info;
+}
+
+} // namespace Common

src/common/memory_detect.h

@@ -0,0 +1,22 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "common/common_types.h"
+
+namespace Common {
+
+struct MemoryInfo {
+    u64 TotalPhysicalMemory{};
+    u64 TotalSwapMemory{};
+};
+
+/**
+ * Gets the memory info of the host system
+ * @return Reference to a MemoryInfo struct with the physical and swap memory sizes in bytes
+ */
+const MemoryInfo& GetMemInfo();
+
+} // namespace Common

src/core/file_sys/patch_manager.cpp

@@ -10,6 +10,7 @@
 #include "common/file_util.h"
 #include "common/hex_util.h"
 #include "common/logging/log.h"
+#include "common/string_util.h"
 #include "core/core.h"
 #include "core/file_sys/content_archive.h"
 #include "core/file_sys/control_metadata.h"
@@ -48,6 +49,23 @@ std::string FormatTitleVersion(u32 version, TitleVersionFormat format) {
     return fmt::format("v{}.{}.{}", bytes[3], bytes[2], bytes[1]);
 }
 
+std::shared_ptr<VfsDirectory> FindSubdirectoryCaseless(const std::shared_ptr<VfsDirectory> dir,
+                                                       std::string_view name) {
+#ifdef _WIN32
+    return dir->GetSubdirectory(name);
+#else
+    const auto subdirs = dir->GetSubdirectories();
+    for (const auto& subdir : subdirs) {
+        std::string dir_name = Common::ToLower(subdir->GetName());
+        if (dir_name == name) {
+            return subdir;
+        }
+    }
+
+    return nullptr;
+#endif
+}
+
 PatchManager::PatchManager(u64 title_id) : title_id(title_id) {}
 
 PatchManager::~PatchManager() = default;
@@ -104,7 +122,7 @@ VirtualDir PatchManager::PatchExeFS(VirtualDir exefs) const {
             if (std::find(disabled.begin(), disabled.end(), subdir->GetName()) != disabled.end())
                 continue;
 
-            auto exefs_dir = subdir->GetSubdirectory("exefs");
+            auto exefs_dir = FindSubdirectoryCaseless(subdir, "exefs");
             if (exefs_dir != nullptr)
                 layers.push_back(std::move(exefs_dir));
         }
@@ -130,7 +148,7 @@ std::vector<VirtualFile> PatchManager::CollectPatches(const std::vector<VirtualD
         if (std::find(disabled.cbegin(), disabled.cend(), subdir->GetName()) != disabled.cend())
             continue;
 
-        auto exefs_dir = subdir->GetSubdirectory("exefs");
+        auto exefs_dir = FindSubdirectoryCaseless(subdir, "exefs");
         if (exefs_dir != nullptr) {
             for (const auto& file : exefs_dir->GetFiles()) {
                 if (file->GetExtension() == "ips") {
@@ -295,7 +313,7 @@ std::vector<Core::Memory::CheatEntry> PatchManager::CreateCheatList(
             continue;
         }
 
-        auto cheats_dir = subdir->GetSubdirectory("cheats");
+        auto cheats_dir = FindSubdirectoryCaseless(subdir, "cheats");
         if (cheats_dir != nullptr) {
             auto res = ReadCheatFileFromFolder(system, title_id, build_id_, cheats_dir, true);
             if (res.has_value()) {
@@ -340,11 +358,11 @@ static void ApplyLayeredFS(VirtualFile& romfs, u64 title_id, ContentRecordType t
             continue;
         }
 
-        auto romfs_dir = subdir->GetSubdirectory("romfs");
+        auto romfs_dir = FindSubdirectoryCaseless(subdir, "romfs");
         if (romfs_dir != nullptr)
             layers.push_back(std::move(romfs_dir));
 
-        auto ext_dir = subdir->GetSubdirectory("romfs_ext");
+        auto ext_dir = FindSubdirectoryCaseless(subdir, "romfs_ext");
         if (ext_dir != nullptr)
             layers_ext.push_back(std::move(ext_dir));
     }
@@ -470,7 +488,7 @@ std::map<std::string, std::string, std::less<>> PatchManager::GetPatchVersionNam
         for (const auto& mod : mod_dir->GetSubdirectories()) {
             std::string types;
 
-            const auto exefs_dir = mod->GetSubdirectory("exefs");
+            const auto exefs_dir = FindSubdirectoryCaseless(mod, "exefs");
             if (IsDirValidAndNonEmpty(exefs_dir)) {
                 bool ips = false;
                 bool ipswitch = false;
@@ -494,9 +512,9 @@ std::map<std::string, std::string, std::less<>> PatchManager::GetPatchVersionNam
                 if (layeredfs)
                     AppendCommaIfNotEmpty(types, "LayeredExeFS");
             }
-            if (IsDirValidAndNonEmpty(mod->GetSubdirectory("romfs")))
+            if (IsDirValidAndNonEmpty(FindSubdirectoryCaseless(mod, "romfs")))
                 AppendCommaIfNotEmpty(types, "LayeredFS");
-            if (IsDirValidAndNonEmpty(mod->GetSubdirectory("cheats")))
+            if (IsDirValidAndNonEmpty(FindSubdirectoryCaseless(mod, "cheats")))
                 AppendCommaIfNotEmpty(types, "Cheats");
 
             if (types.empty())

src/core/file_sys/patch_manager.h

@@ -29,6 +29,11 @@ enum class TitleVersionFormat : u8 {
 std::string FormatTitleVersion(u32 version,
                                TitleVersionFormat format = TitleVersionFormat::ThreeElements);
 
+// Returns a directory with name matching name case-insensitive. Returns nullptr if directory
+// doesn't have a directory with name.
+std::shared_ptr<VfsDirectory> FindSubdirectoryCaseless(const std::shared_ptr<VfsDirectory> dir,
+                                                       std::string_view name);
+
 // A centralized class to manage patches to games.
 class PatchManager {
 public:

src/video_core/rasterizer_cache.h

@@ -56,9 +56,27 @@ public:
         last_modified_ticks = cache.GetModifiedTicks();
     }
 
+    void SetMemoryMarked(bool is_memory_marked_) {
+        is_memory_marked = is_memory_marked_;
+    }
+
+    bool IsMemoryMarked() const {
+        return is_memory_marked;
+    }
+
+    void SetSyncPending(bool is_sync_pending_) {
+        is_sync_pending = is_sync_pending_;
+    }
+
+    bool IsSyncPending() const {
+        return is_sync_pending;
+    }
+
 private:
     bool is_registered{};      ///< Whether the object is currently registered with the cache
     bool is_dirty{};           ///< Whether the object is dirty (out of sync with guest memory)
+    bool is_memory_marked{};   ///< Whether the object is marking rasterizer memory.
+    bool is_sync_pending{};    ///< Whether the object is pending deletion.
     u64 last_modified_ticks{}; ///< When the object was last modified, used for in-order flushing
     VAddr cpu_addr{};          ///< Cpu address memory, unique from emulated virtual address space
 };
@@ -94,6 +112,30 @@ public:
         }
     }
 
+    void OnCPUWrite(VAddr addr, std::size_t size) {
+        std::lock_guard lock{mutex};
+
+        for (const auto& object : GetSortedObjectsFromRegion(addr, size)) {
+            if (object->IsRegistered()) {
+                UnmarkMemory(object);
+                object->SetSyncPending(true);
+                marked_for_unregister.emplace_back(object);
+            }
+        }
+    }
+
+    void SyncGuestHost() {
+        std::lock_guard lock{mutex};
+
+        for (const auto& object : marked_for_unregister) {
+            if (object->IsRegistered()) {
+                object->SetSyncPending(false);
+                Unregister(object);
+            }
+        }
+        marked_for_unregister.clear();
+    }
+
     /// Invalidates everything in the cache
     void InvalidateAll() {
         std::lock_guard lock{mutex};
@@ -120,19 +162,32 @@ protected:
         interval_cache.add({GetInterval(object), ObjectSet{object}});
         map_cache.insert({object->GetCpuAddr(), object});
         rasterizer.UpdatePagesCachedCount(object->GetCpuAddr(), object->GetSizeInBytes(), 1);
+        object->SetMemoryMarked(true);
     }
 
     /// Unregisters an object from the cache
     virtual void Unregister(const T& object) {
         std::lock_guard lock{mutex};
 
+        UnmarkMemory(object);
         object->SetIsRegistered(false);
-        rasterizer.UpdatePagesCachedCount(object->GetCpuAddr(), object->GetSizeInBytes(), -1);
+        if (object->IsSyncPending()) {
+            marked_for_unregister.remove(object);
+            object->SetSyncPending(false);
+        }
         const VAddr addr = object->GetCpuAddr();
         interval_cache.subtract({GetInterval(object), ObjectSet{object}});
         map_cache.erase(addr);
     }
 
+    void UnmarkMemory(const T& object) {
+        if (!object->IsMemoryMarked()) {
+            return;
+        }
+        rasterizer.UpdatePagesCachedCount(object->GetCpuAddr(), object->GetSizeInBytes(), -1);
+        object->SetMemoryMarked(false);
+    }
+
     /// Returns a ticks counter used for tracking when cached objects were last modified
     u64 GetModifiedTicks() {
         std::lock_guard lock{mutex};
@@ -194,4 +249,5 @@ private:
     IntervalCache interval_cache; ///< Cache of objects
     u64 modified_ticks{};         ///< Counter of cache state ticks, used for in-order flushing
     VideoCore::RasterizerInterface& rasterizer;
+    std::list<T> marked_for_unregister;
 };

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -727,15 +727,15 @@ void RasterizerOpenGL::OnCPUWrite(VAddr addr, u64 size) {
         return;
     }
     texture_cache.OnCPUWrite(addr, size);
-    shader_cache.InvalidateRegion(addr, size);
+    shader_cache.OnCPUWrite(addr, size);
     buffer_cache.OnCPUWrite(addr, size);
-    query_cache.InvalidateRegion(addr, size);
 }
 
 void RasterizerOpenGL::SyncGuestHost() {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
     texture_cache.SyncGuestHost();
     buffer_cache.SyncGuestHost();
+    shader_cache.SyncGuestHost();
 }
 
 void RasterizerOpenGL::SignalSemaphore(GPUVAddr addr, u32 value) {
@@ -977,16 +977,12 @@ void RasterizerOpenGL::SetupTexture(u32 binding, const Tegra::Texture::FullTextu
         glBindTextureUnit(binding, 0);
         return;
     }
-    glBindTextureUnit(binding, view->GetTexture());
-
-    if (view->GetSurfaceParams().IsBuffer()) {
-        return;
+    const GLuint handle = view->GetTexture(texture.tic.x_source, texture.tic.y_source,
+                                           texture.tic.z_source, texture.tic.w_source);
+    glBindTextureUnit(binding, handle);
+    if (!view->GetSurfaceParams().IsBuffer()) {
+        glBindSampler(binding, sampler_cache.GetSampler(texture.tsc));
     }
-    // Apply swizzle to textures that are not buffers.
-    view->ApplySwizzle(texture.tic.x_source, texture.tic.y_source, texture.tic.z_source,
-                       texture.tic.w_source);
-
-    glBindSampler(binding, sampler_cache.GetSampler(texture.tsc));
 }
 
 void RasterizerOpenGL::SetupDrawImages(std::size_t stage_index, const Shader& shader) {
@@ -1015,14 +1011,11 @@ void RasterizerOpenGL::SetupImage(u32 binding, const Tegra::Texture::TICEntry& t
         glBindImageTexture(binding, 0, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R8);
         return;
     }
-    if (!tic.IsBuffer()) {
-        view->ApplySwizzle(tic.x_source, tic.y_source, tic.z_source, tic.w_source);
-    }
     if (entry.is_written) {
         view->MarkAsModified(texture_cache.Tick());
     }
-    glBindImageTexture(binding, view->GetTexture(), 0, GL_TRUE, 0, GL_READ_WRITE,
-                       view->GetFormat());
+    const GLuint handle = view->GetTexture(tic.x_source, tic.y_source, tic.z_source, tic.w_source);
+    glBindImageTexture(binding, handle, 0, GL_TRUE, 0, GL_READ_WRITE, view->GetFormat());
 }
 
 void RasterizerOpenGL::SyncViewport() {

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -1538,7 +1538,9 @@ private:
         Expression target;
         if (const auto gpr = std::get_if<GprNode>(&*dest)) {
             if (gpr->GetIndex() == Register::ZeroIndex) {
-                // Writing to Register::ZeroIndex is a no op
+                // Writing to Register::ZeroIndex is a no op but we still have to visit the source
+                // as it might have side effects.
+                code.AddLine("{};", Visit(src).GetCode());
                 return {};
             }
             target = {GetRegister(gpr->GetIndex()), Type::Float};
@@ -2333,6 +2335,15 @@ private:
         return {fmt::format("readInvocationARB({}, {})", value, index), Type::Float};
     }
 
+    Expression Barrier(Operation) {
+        if (!ir.IsDecompiled()) {
+            LOG_ERROR(Render_OpenGL, "barrier() used but shader is not decompiled");
+            return {};
+        }
+        code.AddLine("barrier();");
+        return {};
+    }
+
     Expression MemoryBarrierGL(Operation) {
         code.AddLine("memoryBarrier();");
         return {};
@@ -2579,6 +2590,7 @@ private:
         &GLSLDecompiler::ThreadMask<Func::Lt>,
         &GLSLDecompiler::ShuffleIndexed,
 
+        &GLSLDecompiler::Barrier,
         &GLSLDecompiler::MemoryBarrierGL,
     };
     static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));

src/video_core/renderer_opengl/gl_shader_manager.cpp

@@ -47,6 +47,10 @@ void ProgramManager::BindHostPipeline(GLuint pipeline) {
             old_state.geometry = 0;
             glDisable(GL_GEOMETRY_PROGRAM_NV);
         }
+    } else {
+        if (!is_graphics_bound) {
+            glUseProgram(0);
+        }
     }
     glBindProgramPipeline(pipeline);
 }

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -35,7 +35,7 @@ MICROPROFILE_DEFINE(OpenGL_Texture_Buffer_Copy, "OpenGL", "Texture Buffer Copy",
 namespace {
 
 struct FormatTuple {
-    GLint internal_format;
+    GLenum internal_format;
     GLenum format = GL_NONE;
     GLenum type = GL_NONE;
 };
@@ -238,6 +238,12 @@ OGLTexture CreateTexture(const SurfaceParams& params, GLenum target, GLenum inte
     return texture;
 }
 
+constexpr u32 EncodeSwizzle(SwizzleSource x_source, SwizzleSource y_source, SwizzleSource z_source,
+                            SwizzleSource w_source) {
+    return (static_cast<u32>(x_source) << 24) | (static_cast<u32>(y_source) << 16) |
+           (static_cast<u32>(z_source) << 8) | static_cast<u32>(w_source);
+}
+
 } // Anonymous namespace
 
 CachedSurface::CachedSurface(const GPUVAddr gpu_addr, const SurfaceParams& params,
@@ -381,7 +387,7 @@ void CachedSurface::DecorateSurfaceName() {
 }
 
 void CachedSurfaceView::DecorateViewName(GPUVAddr gpu_addr, std::string prefix) {
-    LabelGLObject(GL_TEXTURE, texture_view.handle, gpu_addr, prefix);
+    LabelGLObject(GL_TEXTURE, main_view.handle, gpu_addr, prefix);
 }
 
 View CachedSurface::CreateView(const ViewParams& view_key) {
@@ -397,14 +403,13 @@ View CachedSurface::CreateViewInner(const ViewParams& view_key, const bool is_pr
 }
 
 CachedSurfaceView::CachedSurfaceView(CachedSurface& surface, const ViewParams& params,
-                                     const bool is_proxy)
-    : VideoCommon::ViewBase(params), surface{surface}, is_proxy{is_proxy} {
-    target = GetTextureTarget(params.target);
-    format = GetFormatTuple(surface.GetSurfaceParams().pixel_format).internal_format;
+                                     bool is_proxy)
+    : VideoCommon::ViewBase(params), surface{surface},
+      format{GetFormatTuple(surface.GetSurfaceParams().pixel_format).internal_format},
+      target{GetTextureTarget(params.target)}, is_proxy{is_proxy} {
     if (!is_proxy) {
-        texture_view = CreateTextureView();
+        main_view = CreateTextureView();
     }
-    swizzle = EncodeSwizzle(SwizzleSource::R, SwizzleSource::G, SwizzleSource::B, SwizzleSource::A);
 }
 
 CachedSurfaceView::~CachedSurfaceView() = default;
@@ -447,27 +452,49 @@ void CachedSurfaceView::Attach(GLenum attachment, GLenum target) const {
     }
 }
 
-void CachedSurfaceView::ApplySwizzle(SwizzleSource x_source, SwizzleSource y_source,
+GLuint CachedSurfaceView::GetTexture(SwizzleSource x_source, SwizzleSource y_source,
                                      SwizzleSource z_source, SwizzleSource w_source) {
-    u32 new_swizzle = EncodeSwizzle(x_source, y_source, z_source, w_source);
-    if (new_swizzle == swizzle)
-        return;
-    swizzle = new_swizzle;
-    const std::array gl_swizzle = {GetSwizzleSource(x_source), GetSwizzleSource(y_source),
-                                   GetSwizzleSource(z_source), GetSwizzleSource(w_source)};
-    const GLuint handle = GetTexture();
-    const PixelFormat format = surface.GetSurfaceParams().pixel_format;
-    switch (format) {
+    if (GetSurfaceParams().IsBuffer()) {
+        return GetTexture();
+    }
+    const u32 new_swizzle = EncodeSwizzle(x_source, y_source, z_source, w_source);
+    if (current_swizzle == new_swizzle) {
+        return current_view;
+    }
+    current_swizzle = new_swizzle;
+
+    const auto [entry, is_cache_miss] = view_cache.try_emplace(new_swizzle);
+    OGLTextureView& view = entry->second;
+    if (!is_cache_miss) {
+        current_view = view.handle;
+        return view.handle;
+    }
+    view = CreateTextureView();
+    current_view = view.handle;
+
+    std::array swizzle{x_source, y_source, z_source, w_source};
+
+    switch (const PixelFormat format = GetSurfaceParams().pixel_format) {
     case PixelFormat::Z24S8:
     case PixelFormat::Z32FS8:
     case PixelFormat::S8Z24:
-        glTextureParameteri(handle, GL_DEPTH_STENCIL_TEXTURE_MODE,
+        UNIMPLEMENTED_IF(x_source != SwizzleSource::R && x_source != SwizzleSource::G);
+        glTextureParameteri(view.handle, GL_DEPTH_STENCIL_TEXTURE_MODE,
                             GetComponent(format, x_source == SwizzleSource::R));
-        break;
-    default:
-        glTextureParameteriv(handle, GL_TEXTURE_SWIZZLE_RGBA, gl_swizzle.data());
+
+        // Make sure we sample the first component
+        std::transform(swizzle.begin(), swizzle.end(), swizzle.begin(), [](SwizzleSource value) {
+            return value == SwizzleSource::G ? SwizzleSource::R : value;
+        });
+        [[fallthrough]];
+    default: {
+        const std::array gl_swizzle = {GetSwizzleSource(swizzle[0]), GetSwizzleSource(swizzle[1]),
+                                       GetSwizzleSource(swizzle[2]), GetSwizzleSource(swizzle[3])};
+        glTextureParameteriv(view.handle, GL_TEXTURE_SWIZZLE_RGBA, gl_swizzle.data());
         break;
     }
+    }
+    return view.handle;
 }
 
 OGLTextureView CachedSurfaceView::CreateTextureView() const {

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -83,7 +83,7 @@ public:
     /// Attaches this texture view to the current bound GL_DRAW_FRAMEBUFFER
     void Attach(GLenum attachment, GLenum target) const;
 
-    void ApplySwizzle(Tegra::Texture::SwizzleSource x_source,
+    GLuint GetTexture(Tegra::Texture::SwizzleSource x_source,
                       Tegra::Texture::SwizzleSource y_source,
                       Tegra::Texture::SwizzleSource z_source,
                       Tegra::Texture::SwizzleSource w_source);
@@ -98,7 +98,7 @@ public:
         if (is_proxy) {
             return surface.GetTexture();
         }
-        return texture_view.handle;
+        return main_view.handle;
     }
 
     GLenum GetFormat() const {
@@ -110,23 +110,19 @@ public:
     }
 
 private:
-    u32 EncodeSwizzle(Tegra::Texture::SwizzleSource x_source,
-                      Tegra::Texture::SwizzleSource y_source,
-                      Tegra::Texture::SwizzleSource z_source,
-                      Tegra::Texture::SwizzleSource w_source) const {
-        return (static_cast<u32>(x_source) << 24) | (static_cast<u32>(y_source) << 16) |
-               (static_cast<u32>(z_source) << 8) | static_cast<u32>(w_source);
-    }
-
     OGLTextureView CreateTextureView() const;
 
     CachedSurface& surface;
-    GLenum target{};
-    GLenum format{};
+    const GLenum format;
+    const GLenum target;
+    const bool is_proxy;
 
-    OGLTextureView texture_view;
-    u32 swizzle{};
-    bool is_proxy{};
+    std::unordered_map<u32, OGLTextureView> view_cache;
+    OGLTextureView main_view;
+
+    // Use an invalid default so it always fails the comparison test
+    u32 current_swizzle = 0xffffffff;
+    GLuint current_view = 0;
 };
 
 class TextureCacheOpenGL final : public TextureCacheBase {

src/video_core/renderer_vulkan/maxwell_to_vk.cpp

@@ -142,7 +142,7 @@ struct FormatTuple {
     {VK_FORMAT_BC6H_UFLOAT_BLOCK},                              // BC6H_UF16
     {VK_FORMAT_BC6H_SFLOAT_BLOCK},                              // BC6H_SF16
     {VK_FORMAT_ASTC_4x4_UNORM_BLOCK},                           // ASTC_2D_4X4
-    {VK_FORMAT_B8G8R8A8_UNORM},                                 // BGRA8
+    {VK_FORMAT_B8G8R8A8_UNORM, Attachable},                     // BGRA8
     {VK_FORMAT_R32G32B32A32_SFLOAT, Attachable | Storage},      // RGBA32F
     {VK_FORMAT_R32G32_SFLOAT, Attachable | Storage},            // RG32F
     {VK_FORMAT_R32_SFLOAT, Attachable | Storage},               // R32F
@@ -168,7 +168,7 @@ struct FormatTuple {
     {VK_FORMAT_ASTC_8x8_UNORM_BLOCK},                           // ASTC_2D_8X8
     {VK_FORMAT_UNDEFINED},                                      // ASTC_2D_8X5
     {VK_FORMAT_UNDEFINED},                                      // ASTC_2D_5X4
-    {VK_FORMAT_UNDEFINED},                                      // BGRA8_SRGB
+    {VK_FORMAT_B8G8R8A8_SRGB, Attachable},                      // BGRA8_SRGB
     {VK_FORMAT_BC1_RGBA_SRGB_BLOCK},                            // DXT1_SRGB
     {VK_FORMAT_BC2_SRGB_BLOCK},                                 // DXT23_SRGB
     {VK_FORMAT_BC3_SRGB_BLOCK},                                 // DXT45_SRGB

src/video_core/renderer_vulkan/vk_device.cpp

@@ -104,6 +104,7 @@ std::unordered_map<VkFormat, VkFormatProperties> GetFormatProperties(
                                         VK_FORMAT_R16_SFLOAT,
                                         VK_FORMAT_R16G16B16A16_SFLOAT,
                                         VK_FORMAT_B8G8R8A8_UNORM,
+                                        VK_FORMAT_B8G8R8A8_SRGB,
                                         VK_FORMAT_R4G4B4A4_UNORM_PACK16,
                                         VK_FORMAT_D32_SFLOAT,
                                         VK_FORMAT_D16_UNORM,

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -329,8 +329,7 @@ VKPipelineCache::DecompileShaders(const GraphicsPipelineCacheKey& key) {
 
         const GPUVAddr gpu_addr = GetShaderAddress(system, program_enum);
         const auto cpu_addr = memory_manager.GpuToCpuAddress(gpu_addr);
-        ASSERT(cpu_addr);
-        const auto shader = TryGet(*cpu_addr);
+        const auto shader = cpu_addr ? TryGet(*cpu_addr) : null_shader;
         ASSERT(shader);
 
         const std::size_t stage = index == 0 ? 0 : index - 1; // Stage indices are 0 - 5

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -532,14 +532,14 @@ void RasterizerVulkan::OnCPUWrite(VAddr addr, u64 size) {
         return;
     }
     texture_cache.OnCPUWrite(addr, size);
-    pipeline_cache.InvalidateRegion(addr, size);
+    pipeline_cache.OnCPUWrite(addr, size);
     buffer_cache.OnCPUWrite(addr, size);
-    query_cache.InvalidateRegion(addr, size);
 }
 
 void RasterizerVulkan::SyncGuestHost() {
     texture_cache.SyncGuestHost();
     buffer_cache.SyncGuestHost();
+    pipeline_cache.SyncGuestHost();
 }
 
 void RasterizerVulkan::SignalSemaphore(GPUVAddr addr, u32 value) {

src/video_core/renderer_vulkan/vk_shader_decompiler.cpp

@@ -1081,8 +1081,7 @@ private:
 
     void VisitBasicBlock(const NodeBlock& bb) {
         for (const auto& node : bb) {
-            [[maybe_unused]] const Type type = Visit(node).type;
-            ASSERT(type == Type::Void);
+            Visit(node);
         }
     }
 
@@ -1372,7 +1371,9 @@ private:
         Expression target{};
         if (const auto gpr = std::get_if<GprNode>(&*dest)) {
             if (gpr->GetIndex() == Register::ZeroIndex) {
-                // Writing to Register::ZeroIndex is a no op
+                // Writing to Register::ZeroIndex is a no op but we still have to visit its source
+                // because it might have side effects.
+                Visit(src);
                 return {};
             }
             target = {registers.at(gpr->GetIndex()), Type::Float};
@@ -2198,6 +2199,22 @@ private:
         return {OpSubgroupReadInvocationKHR(t_float, value, index), Type::Float};
     }
 
+    Expression Barrier(Operation) {
+        if (!ir.IsDecompiled()) {
+            LOG_ERROR(Render_Vulkan, "OpBarrier used by shader is not decompiled");
+            return {};
+        }
+
+        const auto scope = spv::Scope::Workgroup;
+        const auto memory = spv::Scope::Workgroup;
+        const auto semantics =
+            spv::MemorySemanticsMask::WorkgroupMemory | spv::MemorySemanticsMask::AcquireRelease;
+        OpControlBarrier(Constant(t_uint, static_cast<u32>(scope)),
+                         Constant(t_uint, static_cast<u32>(memory)),
+                         Constant(t_uint, static_cast<u32>(semantics)));
+        return {};
+    }
+
     Expression MemoryBarrierGL(Operation) {
         const auto scope = spv::Scope::Device;
         const auto semantics =
@@ -2663,6 +2680,7 @@ private:
         &SPIRVDecompiler::ThreadMask<4>, // Lt
         &SPIRVDecompiler::ShuffleIndexed,
 
+        &SPIRVDecompiler::Barrier,
         &SPIRVDecompiler::MemoryBarrierGL,
     };
     static_assert(operation_decompilers.size() == static_cast<std::size_t>(OperationCode::Amount));

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -354,26 +354,23 @@ CachedSurfaceView::~CachedSurfaceView() = default;
 
 VkImageView CachedSurfaceView::GetHandle(SwizzleSource x_source, SwizzleSource y_source,
                                          SwizzleSource z_source, SwizzleSource w_source) {
-    const u32 swizzle = EncodeSwizzle(x_source, y_source, z_source, w_source);
-    if (last_image_view && last_swizzle == swizzle) {
+    const u32 new_swizzle = EncodeSwizzle(x_source, y_source, z_source, w_source);
+    if (last_image_view && last_swizzle == new_swizzle) {
         return last_image_view;
     }
-    last_swizzle = swizzle;
+    last_swizzle = new_swizzle;
 
-    const auto [entry, is_cache_miss] = view_cache.try_emplace(swizzle);
+    const auto [entry, is_cache_miss] = view_cache.try_emplace(new_swizzle);
     auto& image_view = entry->second;
     if (!is_cache_miss) {
         return last_image_view = *image_view;
     }
 
-    auto swizzle_x = MaxwellToVK::SwizzleSource(x_source);
-    auto swizzle_y = MaxwellToVK::SwizzleSource(y_source);
-    auto swizzle_z = MaxwellToVK::SwizzleSource(z_source);
-    auto swizzle_w = MaxwellToVK::SwizzleSource(w_source);
-
+    std::array swizzle{MaxwellToVK::SwizzleSource(x_source), MaxwellToVK::SwizzleSource(y_source),
+                       MaxwellToVK::SwizzleSource(z_source), MaxwellToVK::SwizzleSource(w_source)};
     if (params.pixel_format == VideoCore::Surface::PixelFormat::A1B5G5R5U) {
         // A1B5G5R5 is implemented as A1R5G5B5, we have to change the swizzle here.
-        std::swap(swizzle_x, swizzle_z);
+        std::swap(swizzle[0], swizzle[2]);
     }
 
     // Games can sample depth or stencil values on textures. This is decided by the swizzle value on
@@ -395,11 +392,11 @@ VkImageView CachedSurfaceView::GetHandle(SwizzleSource x_source, SwizzleSource y
             UNIMPLEMENTED();
         }
 
-        // Vulkan doesn't seem to understand swizzling of a depth stencil image, use identity
-        swizzle_x = VK_COMPONENT_SWIZZLE_R;
-        swizzle_y = VK_COMPONENT_SWIZZLE_G;
-        swizzle_z = VK_COMPONENT_SWIZZLE_B;
-        swizzle_w = VK_COMPONENT_SWIZZLE_A;
+        // Make sure we sample the first component
+        std::transform(
+            swizzle.begin(), swizzle.end(), swizzle.begin(), [](VkComponentSwizzle component) {
+                return component == VK_COMPONENT_SWIZZLE_G ? VK_COMPONENT_SWIZZLE_R : component;
+            });
     }
 
     VkImageViewCreateInfo ci;
@@ -409,7 +406,7 @@ VkImageView CachedSurfaceView::GetHandle(SwizzleSource x_source, SwizzleSource y
     ci.image = surface.GetImageHandle();
     ci.viewType = image_view_type;
     ci.format = surface.GetImage().GetFormat();
-    ci.components = {swizzle_x, swizzle_y, swizzle_z, swizzle_w};
+    ci.components = {swizzle[0], swizzle[1], swizzle[2], swizzle[3]};
     ci.subresourceRange.aspectMask = aspect;
     ci.subresourceRange.baseMipLevel = base_level;
     ci.subresourceRange.levelCount = num_levels;

src/video_core/shader/decode/memory.cpp

@@ -387,7 +387,6 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
     }
     case OpCode::Id::RED: {
         UNIMPLEMENTED_IF_MSG(instr.red.type != GlobalAtomicType::U32);
-        UNIMPLEMENTED_IF_MSG(instr.red.operation != AtomicOp::Add);
         const auto [real_address, base_address, descriptor] =
             TrackGlobalMemory(bb, instr, true, true);
         if (!real_address || !base_address) {
@@ -396,7 +395,7 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
         }
         Node gmem = MakeNode<GmemNode>(real_address, base_address, descriptor);
         Node value = GetRegister(instr.gpr0);
-        bb.push_back(Operation(OperationCode::ReduceIAdd, move(gmem), move(value)));
+        bb.push_back(Operation(GetAtomOperation(instr.red.operation), move(gmem), move(value)));
         break;
     }
     case OpCode::Id::ATOM: {

src/video_core/shader/decode/other.cpp

@@ -293,6 +293,11 @@ u32 ShaderIR::DecodeOther(NodeBlock& bb, u32 pc) {
         SetRegister(bb, instr.gpr0, GetRegister(instr.gpr8));
         break;
     }
+    case OpCode::Id::BAR: {
+        UNIMPLEMENTED_IF_MSG(instr.value != 0xF0A81B8000070000ULL, "BAR is not BAR.SYNC 0x0");
+        bb.push_back(Operation(OperationCode::Barrier));
+        break;
+    }
     case OpCode::Id::MEMBAR: {
         UNIMPLEMENTED_IF(instr.membar.type != Tegra::Shader::MembarType::GL);
         UNIMPLEMENTED_IF(instr.membar.unknown != Tegra::Shader::MembarUnknown::Default);

src/video_core/shader/node.h

@@ -233,6 +233,7 @@ enum class OperationCode {
     ThreadLtMask,   /// () -> uint
     ShuffleIndexed, /// (uint value, uint index) -> uint
 
+    Barrier,         /// () -> void
     MemoryBarrierGL, /// () -> void
 
     Amount,

src/video_core/texture_cache/texture_cache.h

@@ -991,7 +991,9 @@ private:
         params.target = target;
         params.is_tiled = false;
         params.srgb_conversion = false;
-        params.is_layered = false;
+        params.is_layered =
+            target == SurfaceTarget::Texture1DArray || target == SurfaceTarget::Texture2DArray ||
+            target == SurfaceTarget::TextureCubemap || target == SurfaceTarget::TextureCubeArray;
         params.block_width = 0;
         params.block_height = 0;
         params.block_depth = 0;

src/yuzu/main.cpp

@@ -65,6 +65,7 @@ static FileSys::VirtualFile VfsDirectoryCreateFileWrapper(const FileSys::Virtual
 #include "common/logging/backend.h"
 #include "common/logging/filter.h"
 #include "common/logging/log.h"
+#include "common/memory_detect.h"
 #include "common/microprofile.h"
 #include "common/scm_rev.h"
 #include "common/scope_exit.h"
@@ -219,6 +220,10 @@ GMainWindow::GMainWindow()
     LOG_INFO(Frontend, "Host CPU: {}", Common::GetCPUCaps().cpu_string);
 #endif
     LOG_INFO(Frontend, "Host OS: {}", QSysInfo::prettyProductName().toStdString());
+    LOG_INFO(Frontend, "Host RAM: {:.2f} GB",
+             Common::GetMemInfo().TotalPhysicalMemory / 1024.0f / 1024 / 1024);
+    LOG_INFO(Frontend, "Host Swap: {:.2f} GB",
+             Common::GetMemInfo().TotalSwapMemory / 1024.0f / 1024 / 1024);
     UpdateWindowTitle();
 
     show();