Fix cubemaps to use a less hacky method of loading

- This will be used as a catch-all for slow-but-accurate GPU emulation paths.

src/core/settings.h

@@ -136,7 +136,7 @@ struct Values {
     float resolution_factor;
     bool use_frame_limit;
     u16 frame_limit;
-    bool use_accurate_framebuffers;
+    bool use_accurate_gpu_emulation;
 
     float bg_red;
     float bg_green;

src/core/telemetry_session.cpp

@@ -163,8 +163,8 @@ TelemetrySession::TelemetrySession() {
     AddField(Telemetry::FieldType::UserConfig, "Renderer_UseFrameLimit",
              Settings::values.use_frame_limit);
     AddField(Telemetry::FieldType::UserConfig, "Renderer_FrameLimit", Settings::values.frame_limit);
-    AddField(Telemetry::FieldType::UserConfig, "Renderer_UseAccurateFramebuffers",
-             Settings::values.use_accurate_framebuffers);
+    AddField(Telemetry::FieldType::UserConfig, "Renderer_UseAccurateGpuEmulation",
+             Settings::values.use_accurate_gpu_emulation);
     AddField(Telemetry::FieldType::UserConfig, "System_UseDockedMode",
              Settings::values.use_docked_mode);
 }

src/video_core/engines/maxwell_3d.h

@@ -448,7 +448,10 @@ public:
                 BitField<8, 3, u32> block_depth;
                 BitField<12, 1, InvMemoryLayout> type;
             } memory_layout;
-            u32 array_mode;
+            union {
+                BitField<0, 16, u32> array_mode;
+                BitField<16, 1, u32> volume;
+            };
             u32 layer_stride;
             u32 base_layer;
             INSERT_PADDING_WORDS(7);

src/video_core/engines/shader_bytecode.h

@@ -1095,13 +1095,11 @@ public:
         KIL,
         SSY,
         SYNC,
-        BRK,
         DEPBAR,
         BFE_C,
         BFE_R,
         BFE_IMM,
         BRA,
-        PBK,
         LD_A,
         LD_C,
         ST_A,
@@ -1241,7 +1239,7 @@ public:
     /// conditionally executed).
     static bool IsPredicatedInstruction(Id opcode) {
         // TODO(Subv): Add the rest of unpredicated instructions.
-        return opcode != Id::SSY && opcode != Id::PBK;
+        return opcode != Id::SSY;
     }
 
     class Matcher {
@@ -1337,11 +1335,9 @@ private:
 #define INST(bitstring, op, type, name) Detail::GetMatcher(bitstring, op, type, name)
             INST("111000110011----", Id::KIL, Type::Flow, "KIL"),
             INST("111000101001----", Id::SSY, Type::Flow, "SSY"),
-            INST("111000101010----", Id::PBK, Type::Flow, "PBK"),
             INST("111000100100----", Id::BRA, Type::Flow, "BRA"),
-            INST("1111000011111---", Id::SYNC, Type::Flow, "SYNC"),
-            INST("111000110100---", Id::BRK, Type::Flow, "BRK"),
             INST("1111000011110---", Id::DEPBAR, Type::Synch, "DEPBAR"),
+            INST("1111000011111---", Id::SYNC, Type::Synch, "SYNC"),
             INST("1110111111011---", Id::LD_A, Type::Memory, "LD_A"),
             INST("1110111110010---", Id::LD_C, Type::Memory, "LD_C"),
             INST("1110111111110---", Id::ST_A, Type::Memory, "ST_A"),
@@ -1467,4 +1463,4 @@ private:
     }
 };
 
-} // namespace Tegra::Shader
+} // namespace Tegra::Shader

src/video_core/memory_manager.cpp

@@ -87,6 +87,16 @@ GPUVAddr MemoryManager::UnmapBuffer(GPUVAddr gpu_addr, u64 size) {
     return gpu_addr;
 }
 
+GPUVAddr MemoryManager::GetRegionEnd(GPUVAddr region_start) const {
+    for (const auto& region : mapped_regions) {
+        const GPUVAddr region_end{region.gpu_addr + region.size};
+        if (region_start >= region.gpu_addr && region_start < region_end) {
+            return region_end;
+        }
+    }
+    return {};
+}
+
 boost::optional<GPUVAddr> MemoryManager::FindFreeBlock(u64 size, u64 align) {
     GPUVAddr gpu_addr = 0;
     u64 free_space = 0;

src/video_core/memory_manager.h

@@ -26,6 +26,7 @@ public:
     GPUVAddr MapBufferEx(VAddr cpu_addr, u64 size);
     GPUVAddr MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size);
     GPUVAddr UnmapBuffer(GPUVAddr gpu_addr, u64 size);
+    GPUVAddr GetRegionEnd(GPUVAddr region_start) const;
     boost::optional<VAddr> GpuToCpuAddress(GPUVAddr gpu_addr);
     std::vector<GPUVAddr> CpuToGpuAddress(VAddr cpu_addr) const;
 

src/video_core/rasterizer_cache.h

@@ -11,32 +11,77 @@
 
 #include "common/common_types.h"
 #include "core/core.h"
+#include "core/settings.h"
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_base.h"
 
+class RasterizerCacheObject {
+public:
+    /// Gets the address of the shader in guest memory, required for cache management
+    virtual VAddr GetAddr() const = 0;
+
+    /// Gets the size of the shader in guest memory, required for cache management
+    virtual std::size_t GetSizeInBytes() const = 0;
+
+    /// Wriets any cached resources back to memory
+    virtual void Flush() = 0;
+
+    /// Sets whether the cached object should be considered registered
+    void SetIsRegistered(bool registered) {
+        is_registered = registered;
+    }
+
+    /// Returns true if the cached object is registered
+    bool IsRegistered() const {
+        return is_registered;
+    }
+
+    /// Returns true if the cached object is dirty
+    bool IsDirty() const {
+        return is_dirty;
+    }
+
+    /// Returns ticks from when this cached object was last modified
+    u64 GetLastModifiedTicks() const {
+        return last_modified_ticks;
+    }
+
+    /// Marks an object as recently modified, used to specify whether it is clean or dirty
+    template <class T>
+    void MarkAsModified(bool dirty, T& cache) {
+        is_dirty = dirty;
+        last_modified_ticks = cache.GetModifiedTicks();
+    }
+
+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)
+    u64 last_modified_ticks{}; ///< When the object was last modified, used for in-order flushing
+};
+
 template <class T>
 class RasterizerCache : NonCopyable {
+    friend class RasterizerCacheObject;
+
 public:
+    /// Write any cached resources overlapping the specified region back to memory
+    void FlushRegion(Tegra::GPUVAddr addr, size_t size) {
+        const auto& objects{GetSortedObjectsFromRegion(addr, size)};
+        for (auto& object : objects) {
+            FlushObject(object);
+        }
+    }
+
     /// Mark the specified region as being invalidated
     void InvalidateRegion(VAddr addr, u64 size) {
-        if (size == 0)
-            return;
-
-        const ObjectInterval interval{addr, addr + size};
-        for (auto& pair : boost::make_iterator_range(object_cache.equal_range(interval))) {
-            for (auto& cached_object : pair.second) {
-                if (!cached_object)
-                    continue;
-
-                remove_objects.emplace(cached_object);
+        const auto& objects{GetSortedObjectsFromRegion(addr, size)};
+        for (auto& object : objects) {
+            if (!object->IsRegistered()) {
+                // Skip duplicates
+                continue;
             }
+            Unregister(object);
         }
-
-        for (auto& remove_object : remove_objects) {
-            Unregister(remove_object);
-        }
-
-        remove_objects.clear();
     }
 
     /// Invalidates everything in the cache
@@ -62,6 +107,7 @@ protected:
 
     /// Register an object into the cache
     void Register(const T& object) {
+        object->SetIsRegistered(true);
         object_cache.add({GetInterval(object), ObjectSet{object}});
         auto& rasterizer = Core::System::GetInstance().Renderer().Rasterizer();
         rasterizer.UpdatePagesCachedCount(object->GetAddr(), object->GetSizeInBytes(), 1);
@@ -69,12 +115,57 @@ protected:
 
     /// Unregisters an object from the cache
     void Unregister(const T& object) {
+        object->SetIsRegistered(false);
         auto& rasterizer = Core::System::GetInstance().Renderer().Rasterizer();
         rasterizer.UpdatePagesCachedCount(object->GetAddr(), object->GetSizeInBytes(), -1);
+
+        // Only flush if use_accurate_gpu_emulation is enabled, as it incurs a performance hit
+        if (Settings::values.use_accurate_gpu_emulation) {
+            FlushObject(object);
+        }
+
         object_cache.subtract({GetInterval(object), ObjectSet{object}});
     }
 
+    /// Returns a ticks counter used for tracking when cached objects were last modified
+    u64 GetModifiedTicks() {
+        return ++modified_ticks;
+    }
+
 private:
+    /// Returns a list of cached objects from the specified memory region, ordered by access time
+    std::vector<T> GetSortedObjectsFromRegion(VAddr addr, u64 size) {
+        if (size == 0) {
+            return {};
+        }
+
+        std::vector<T> objects;
+        const ObjectInterval interval{addr, addr + size};
+        for (auto& pair : boost::make_iterator_range(object_cache.equal_range(interval))) {
+            for (auto& cached_object : pair.second) {
+                if (!cached_object) {
+                    continue;
+                }
+                objects.push_back(cached_object);
+            }
+        }
+
+        std::sort(objects.begin(), objects.end(), [](const T& a, const T& b) -> bool {
+            return a->GetLastModifiedTicks() < b->GetLastModifiedTicks();
+        });
+
+        return objects;
+    }
+
+    /// Flushes the specified object, updating appropriate cache state as needed
+    void FlushObject(const T& object) {
+        if (!object->IsDirty()) {
+            return;
+        }
+        object->Flush();
+        object->MarkAsModified(false, *this);
+    }
+
     using ObjectSet = std::set<T>;
     using ObjectCache = boost::icl::interval_map<VAddr, ObjectSet>;
     using ObjectInterval = typename ObjectCache::interval_type;
@@ -84,6 +175,6 @@ private:
                                           object->GetAddr() + object->GetSizeInBytes());
     }
 
-    ObjectCache object_cache;
-    ObjectSet remove_objects;
+    ObjectCache object_cache; ///< Cache of objects
+    u64 modified_ticks{};     ///< Counter of cache state ticks, used for in-order flushing
 };

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -15,15 +15,18 @@
 
 namespace OpenGL {
 
-struct CachedBufferEntry final {
-    VAddr GetAddr() const {
+struct CachedBufferEntry final : public RasterizerCacheObject {
+    VAddr GetAddr() const override {
         return addr;
     }
 
-    std::size_t GetSizeInBytes() const {
+    std::size_t GetSizeInBytes() const override {
         return size;
     }
 
+    // We do not have to flush this cache as things in it are never modified by us.
+    void Flush() override {}
+
     VAddr addr;
     std::size_t size;
     GLintptr offset;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -424,6 +424,13 @@ void RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb, bool using_dep
             // Used when just a single color attachment is enabled, e.g. for clearing a color buffer
             Surface color_surface =
                 res_cache.GetColorBufferSurface(*single_color_target, preserve_contents);
+
+            if (color_surface) {
+                // Assume that a surface will be written to if it is used as a framebuffer, even if
+                // the shader doesn't actually write to it.
+                color_surface->MarkAsModified(true, res_cache);
+            }
+
             glFramebufferTexture2D(
                 GL_DRAW_FRAMEBUFFER,
                 GL_COLOR_ATTACHMENT0 + static_cast<GLenum>(*single_color_target), GL_TEXTURE_2D,
@@ -434,6 +441,13 @@ void RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb, bool using_dep
             std::array<GLenum, Maxwell::NumRenderTargets> buffers;
             for (std::size_t index = 0; index < Maxwell::NumRenderTargets; ++index) {
                 Surface color_surface = res_cache.GetColorBufferSurface(index, preserve_contents);
+
+                if (color_surface) {
+                    // Assume that a surface will be written to if it is used as a framebuffer, even
+                    // if the shader doesn't actually write to it.
+                    color_surface->MarkAsModified(true, res_cache);
+                }
+
                 buffers[index] = GL_COLOR_ATTACHMENT0 + regs.rt_control.GetMap(index);
                 glFramebufferTexture2D(
                     GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + static_cast<GLenum>(index),
@@ -453,6 +467,10 @@ void RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb, bool using_dep
     }
 
     if (depth_surface) {
+        // Assume that a surface will be written to if it is used as a framebuffer, even if
+        // the shader doesn't actually write to it.
+        depth_surface->MarkAsModified(true, res_cache);
+
         if (regs.stencil_enable) {
             // Attach both depth and stencil
             glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D,
@@ -617,7 +635,14 @@ void RasterizerOpenGL::DrawArrays() {
 
 void RasterizerOpenGL::FlushAll() {}
 
-void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {}
+void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
+    MICROPROFILE_SCOPE(OpenGL_CacheManagement);
+
+    if (Settings::values.use_accurate_gpu_emulation) {
+        // Only flush if use_accurate_gpu_emulation is enabled, as it incurs a performance hit
+        res_cache.FlushRegion(addr, size);
+    }
+}
 
 void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
@@ -627,6 +652,7 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
 }
 
 void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
+    FlushRegion(addr, size);
     InvalidateRegion(addr, size);
 }
 

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -34,16 +34,53 @@ struct FormatTuple {
     bool compressed;
 };
 
-static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) {
-    auto& gpu{Core::System::GetInstance().GPU()};
-    const auto cpu_addr{gpu.MemoryManager().GpuToCpuAddress(gpu_addr)};
-    return cpu_addr ? *cpu_addr : 0;
+static bool IsPixelFormatASTC(PixelFormat format) {
+    switch (format) {
+    case PixelFormat::ASTC_2D_4X4:
+    case PixelFormat::ASTC_2D_5X4:
+    case PixelFormat::ASTC_2D_8X8:
+    case PixelFormat::ASTC_2D_8X5:
+        return true;
+    default:
+        return false;
+    }
+}
+
+static std::pair<u32, u32> GetASTCBlockSize(PixelFormat format) {
+    switch (format) {
+    case PixelFormat::ASTC_2D_4X4:
+        return {4, 4};
+    case PixelFormat::ASTC_2D_5X4:
+        return {5, 4};
+    case PixelFormat::ASTC_2D_8X8:
+        return {8, 8};
+    case PixelFormat::ASTC_2D_8X5:
+        return {8, 5};
+    default:
+        LOG_CRITICAL(HW_GPU, "Unhandled format: {}", static_cast<u32>(format));
+        UNREACHABLE();
+    }
+}
+
+void SurfaceParams::InitCacheParameters(Tegra::GPUVAddr gpu_addr_) {
+    auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()};
+    const auto cpu_addr{memory_manager.GpuToCpuAddress(gpu_addr_)};
+
+    addr = cpu_addr ? *cpu_addr : 0;
+    gpu_addr = gpu_addr_;
+    size_in_bytes = SizeInBytesRaw();
+
+    if (IsPixelFormatASTC(pixel_format)) {
+        // ASTC is uncompressed in software, in emulated as RGBA8
+        size_in_bytes_gl = width * height * depth * 4;
+    } else {
+        size_in_bytes_gl = SizeInBytesGL();
+    }
 }
 
 /*static*/ SurfaceParams SurfaceParams::CreateForTexture(
     const Tegra::Texture::FullTextureInfo& config, const GLShader::SamplerEntry& entry) {
     SurfaceParams params{};
-    params.addr = TryGetCpuAddr(config.tic.Address());
     params.is_tiled = config.tic.IsTiled();
     params.block_width = params.is_tiled ? config.tic.BlockWidth() : 0,
     params.block_height = params.is_tiled ? config.tic.BlockHeight() : 0,
@@ -87,18 +124,18 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) {
         break;
     }
 
-    params.size_in_bytes_total = params.SizeInBytesTotal();
-    params.size_in_bytes_2d = params.SizeInBytes2D();
     params.max_mip_level = config.tic.max_mip_level + 1;
     params.rt = {};
 
+    params.InitCacheParameters(config.tic.Address());
+
     return params;
 }
 
 /*static*/ SurfaceParams SurfaceParams::CreateForFramebuffer(std::size_t index) {
     const auto& config{Core::System::GetInstance().GPU().Maxwell3D().regs.rt[index]};
     SurfaceParams params{};
-    params.addr = TryGetCpuAddr(config.Address());
+
     params.is_tiled =
         config.memory_layout.type == Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout::BlockLinear;
     params.block_width = 1 << config.memory_layout.block_width;
@@ -112,16 +149,17 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) {
     params.unaligned_height = config.height;
     params.target = SurfaceTarget::Texture2D;
     params.depth = 1;
-    params.size_in_bytes_total = params.SizeInBytesTotal();
-    params.size_in_bytes_2d = params.SizeInBytes2D();
     params.max_mip_level = 0;
 
     // Render target specific parameters, not used for caching
     params.rt.index = static_cast<u32>(index);
     params.rt.array_mode = config.array_mode;
     params.rt.layer_stride = config.layer_stride;
+    params.rt.volume = config.volume;
     params.rt.base_layer = config.base_layer;
 
+    params.InitCacheParameters(config.Address());
+
     return params;
 }
 
@@ -130,7 +168,7 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) {
     u32 block_width, u32 block_height, u32 block_depth,
     Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type) {
     SurfaceParams params{};
-    params.addr = TryGetCpuAddr(zeta_address);
+
     params.is_tiled = type == Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout::BlockLinear;
     params.block_width = 1 << std::min(block_width, 5U);
     params.block_height = 1 << std::min(block_height, 5U);
@@ -143,18 +181,18 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) {
     params.unaligned_height = zeta_height;
     params.target = SurfaceTarget::Texture2D;
     params.depth = 1;
-    params.size_in_bytes_total = params.SizeInBytesTotal();
-    params.size_in_bytes_2d = params.SizeInBytes2D();
     params.max_mip_level = 0;
     params.rt = {};
 
+    params.InitCacheParameters(zeta_address);
+
     return params;
 }
 
 /*static*/ SurfaceParams SurfaceParams::CreateForFermiCopySurface(
     const Tegra::Engines::Fermi2D::Regs::Surface& config) {
     SurfaceParams params{};
-    params.addr = TryGetCpuAddr(config.Address());
+
     params.is_tiled = !config.linear;
     params.block_width = params.is_tiled ? std::min(config.BlockWidth(), 32U) : 0,
     params.block_height = params.is_tiled ? std::min(config.BlockHeight(), 32U) : 0,
@@ -167,11 +205,11 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) {
     params.unaligned_height = config.height;
     params.target = SurfaceTarget::Texture2D;
     params.depth = 1;
-    params.size_in_bytes_total = params.SizeInBytesTotal();
-    params.size_in_bytes_2d = params.SizeInBytes2D();
     params.max_mip_level = 0;
     params.rt = {};
 
+    params.InitCacheParameters(config.Address());
+
     return params;
 }
 
@@ -276,34 +314,6 @@ static const FormatTuple& GetFormatTuple(PixelFormat pixel_format, ComponentType
     return format;
 }
 
-static bool IsPixelFormatASTC(PixelFormat format) {
-    switch (format) {
-    case PixelFormat::ASTC_2D_4X4:
-    case PixelFormat::ASTC_2D_5X4:
-    case PixelFormat::ASTC_2D_8X8:
-    case PixelFormat::ASTC_2D_8X5:
-        return true;
-    default:
-        return false;
-    }
-}
-
-static std::pair<u32, u32> GetASTCBlockSize(PixelFormat format) {
-    switch (format) {
-    case PixelFormat::ASTC_2D_4X4:
-        return {4, 4};
-    case PixelFormat::ASTC_2D_5X4:
-        return {5, 4};
-    case PixelFormat::ASTC_2D_8X8:
-        return {8, 8};
-    case PixelFormat::ASTC_2D_8X5:
-        return {8, 5};
-    default:
-        LOG_CRITICAL(HW_GPU, "Unhandled format: {}", static_cast<u32>(format));
-        UNREACHABLE();
-    }
-}
-
 MathUtil::Rectangle<u32> SurfaceParams::GetRect() const {
     u32 actual_height{unaligned_height};
     if (IsPixelFormatASTC(pixel_format)) {
@@ -333,23 +343,24 @@ static bool IsFormatBCn(PixelFormat format) {
 template <bool morton_to_gl, PixelFormat format>
 void MortonCopy(u32 stride, u32 block_height, u32 height, u32 block_depth, u32 depth, u8* gl_buffer,
                 std::size_t gl_buffer_size, VAddr addr) {
-    constexpr u32 bytes_per_pixel = SurfaceParams::GetFormatBpp(format) / CHAR_BIT;
-    constexpr u32 gl_bytes_per_pixel = CachedSurface::GetGLBytesPerPixel(format);
+    constexpr u32 bytes_per_pixel = SurfaceParams::GetBytesPerPixel(format);
+
+    // With the BCn formats (DXT and DXN), each 4x4 tile is swizzled instead of just individual
+    // pixel values.
+    const u32 tile_size{IsFormatBCn(format) ? 4U : 1U};
 
     if (morton_to_gl) {
-        // With the BCn formats (DXT and DXN), each 4x4 tile is swizzled instead of just individual
-        // pixel values.
-        const u32 tile_size{IsFormatBCn(format) ? 4U : 1U};
         const std::vector<u8> data = Tegra::Texture::UnswizzleTexture(
             addr, tile_size, bytes_per_pixel, stride, height, depth, block_height, block_depth);
         const std::size_t size_to_copy{std::min(gl_buffer_size, data.size())};
         memcpy(gl_buffer, data.data(), size_to_copy);
     } else {
-        // TODO(bunnei): Assumes the default rendering GOB size of 16 (128 lines). We should
-        // check the configuration for this and perform more generic un/swizzle
-        LOG_WARNING(Render_OpenGL, "need to use correct swizzle/GOB parameters!");
-        VideoCore::MortonCopyPixels128(stride, height, bytes_per_pixel, gl_bytes_per_pixel,
-                                       Memory::GetPointer(addr), gl_buffer, morton_to_gl);
+        std::vector<u8> data(gl_buffer_size);
+        Tegra::Texture::CopySwizzledData(stride / tile_size, height / tile_size, depth,
+                                         bytes_per_pixel, bytes_per_pixel, data.data(), gl_buffer,
+                                         false, block_height, block_depth);
+        const std::size_t size_to_copy{std::min(gl_buffer_size, data.size())};
+        memcpy(Memory::GetPointer(addr), data.data(), size_to_copy);
     }
 }
 
@@ -430,17 +441,16 @@ static constexpr std::array<void (*)(u32, u32, u32, u32, u32, u8*, std::size_t,
         MortonCopy<false, PixelFormat::RGBA16UI>,
         MortonCopy<false, PixelFormat::R11FG11FB10F>,
         MortonCopy<false, PixelFormat::RGBA32UI>,
-        // TODO(Subv): Swizzling DXT1/DXT23/DXT45/DXN1/DXN2/BC7U/BC6H_UF16/BC6H_SF16/ASTC_2D_4X4
-        // formats are not supported
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
+        MortonCopy<false, PixelFormat::DXT1>,
+        MortonCopy<false, PixelFormat::DXT23>,
+        MortonCopy<false, PixelFormat::DXT45>,
+        MortonCopy<false, PixelFormat::DXN1>,
+        MortonCopy<false, PixelFormat::DXN2UNORM>,
+        MortonCopy<false, PixelFormat::DXN2SNORM>,
+        MortonCopy<false, PixelFormat::BC7U>,
+        MortonCopy<false, PixelFormat::BC6H_UF16>,
+        MortonCopy<false, PixelFormat::BC6H_SF16>,
+        // TODO(Subv): Swizzling ASTC formats are not supported
         nullptr,
         MortonCopy<false, PixelFormat::G8R8U>,
         MortonCopy<false, PixelFormat::G8R8S>,
@@ -626,22 +636,21 @@ static void CopySurface(const Surface& src_surface, const Surface& dst_surface,
     auto source_format = GetFormatTuple(src_params.pixel_format, src_params.component_type);
     auto dest_format = GetFormatTuple(dst_params.pixel_format, dst_params.component_type);
 
-    std::size_t buffer_size =
-        std::max(src_params.size_in_bytes_total, dst_params.size_in_bytes_total);
+    std::size_t buffer_size = std::max(src_params.size_in_bytes, dst_params.size_in_bytes);
 
     glBindBuffer(GL_PIXEL_PACK_BUFFER, copy_pbo_handle);
     glBufferData(GL_PIXEL_PACK_BUFFER, buffer_size, nullptr, GL_STREAM_DRAW_ARB);
     if (source_format.compressed) {
         glGetCompressedTextureImage(src_surface->Texture().handle, src_attachment,
-                                    static_cast<GLsizei>(src_params.size_in_bytes_total), nullptr);
+                                    static_cast<GLsizei>(src_params.size_in_bytes), nullptr);
     } else {
         glGetTextureImage(src_surface->Texture().handle, src_attachment, source_format.format,
-                          source_format.type, static_cast<GLsizei>(src_params.size_in_bytes_total),
+                          source_format.type, static_cast<GLsizei>(src_params.size_in_bytes),
                           nullptr);
     }
     // If the new texture is bigger than the previous one, we need to fill in the rest with data
     // from the CPU.
-    if (src_params.size_in_bytes_total < dst_params.size_in_bytes_total) {
+    if (src_params.size_in_bytes < dst_params.size_in_bytes) {
         // Upload the rest of the memory.
         if (dst_params.is_tiled) {
             // TODO(Subv): We might have to de-tile the subtexture and re-tile it with the rest
@@ -651,12 +660,12 @@ static void CopySurface(const Surface& src_surface, const Surface& dst_surface,
             LOG_DEBUG(HW_GPU, "Trying to upload extra texture data from the CPU during "
                               "reinterpretation but the texture is tiled.");
         }
-        std::size_t remaining_size =
-            dst_params.size_in_bytes_total - src_params.size_in_bytes_total;
+        std::size_t remaining_size = dst_params.size_in_bytes - src_params.size_in_bytes;
         std::vector<u8> data(remaining_size);
-        Memory::ReadBlock(dst_params.addr + src_params.size_in_bytes_total, data.data(),
-                          data.size());
-        glBufferSubData(GL_PIXEL_PACK_BUFFER, src_params.size_in_bytes_total, remaining_size,
+        std::memcpy(data.data(), Memory::GetPointer(dst_params.addr + src_params.size_in_bytes),
+                    data.size());
+
+        glBufferSubData(GL_PIXEL_PACK_BUFFER, src_params.size_in_bytes, remaining_size,
                         data.data());
     }
 
@@ -702,7 +711,8 @@ static void CopySurface(const Surface& src_surface, const Surface& dst_surface,
 }
 
 CachedSurface::CachedSurface(const SurfaceParams& params)
-    : params(params), gl_target(SurfaceTargetToGL(params.target)) {
+    : params(params), gl_target(SurfaceTargetToGL(params.target)),
+      cached_size_in_bytes(params.size_in_bytes) {
     texture.Create();
     const auto& rect{params.GetRect()};
 
@@ -752,9 +762,21 @@ CachedSurface::CachedSurface(const SurfaceParams& params)
 
     VideoCore::LabelGLObject(GL_TEXTURE, texture.handle, params.addr,
                              SurfaceParams::SurfaceTargetName(params.target));
+
+    // Clamp size to mapped GPU memory region
+    // TODO(bunnei): Super Mario Odyssey maps a 0x40000 byte region and then uses it for a 0x80000
+    // R32F render buffer. We do not yet know if this is a game bug or something else, but this
+    // check is necessary to prevent flushing from overwriting unmapped memory.
+
+    auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()};
+    const u64 max_size{memory_manager.GetRegionEnd(params.gpu_addr) - params.gpu_addr};
+    if (cached_size_in_bytes > max_size) {
+        LOG_ERROR(HW_GPU, "Surface size {} exceeds region size {}", params.size_in_bytes, max_size);
+        cached_size_in_bytes = max_size;
+    }
 }
 
-static void ConvertS8Z24ToZ24S8(std::vector<u8>& data, u32 width, u32 height) {
+static void ConvertS8Z24ToZ24S8(std::vector<u8>& data, u32 width, u32 height, bool reverse) {
     union S8Z24 {
         BitField<0, 24, u32> z24;
         BitField<24, 8, u32> s8;
@@ -767,22 +789,29 @@ static void ConvertS8Z24ToZ24S8(std::vector<u8>& data, u32 width, u32 height) {
     };
     static_assert(sizeof(Z24S8) == 4, "Z24S8 is incorrect size");
 
-    S8Z24 input_pixel{};
-    Z24S8 output_pixel{};
-    constexpr auto bpp{CachedSurface::GetGLBytesPerPixel(PixelFormat::S8Z24)};
+    S8Z24 s8z24_pixel{};
+    Z24S8 z24s8_pixel{};
+    constexpr auto bpp{SurfaceParams::GetBytesPerPixel(PixelFormat::S8Z24)};
     for (std::size_t y = 0; y < height; ++y) {
         for (std::size_t x = 0; x < width; ++x) {
             const std::size_t offset{bpp * (y * width + x)};
-            std::memcpy(&input_pixel, &data[offset], sizeof(S8Z24));
-            output_pixel.s8.Assign(input_pixel.s8);
-            output_pixel.z24.Assign(input_pixel.z24);
-            std::memcpy(&data[offset], &output_pixel, sizeof(Z24S8));
+            if (reverse) {
+                std::memcpy(&z24s8_pixel, &data[offset], sizeof(Z24S8));
+                s8z24_pixel.s8.Assign(z24s8_pixel.s8);
+                s8z24_pixel.z24.Assign(z24s8_pixel.z24);
+                std::memcpy(&data[offset], &s8z24_pixel, sizeof(S8Z24));
+            } else {
+                std::memcpy(&s8z24_pixel, &data[offset], sizeof(S8Z24));
+                z24s8_pixel.s8.Assign(s8z24_pixel.s8);
+                z24s8_pixel.z24.Assign(s8z24_pixel.z24);
+                std::memcpy(&data[offset], &z24s8_pixel, sizeof(Z24S8));
+            }
         }
     }
 }
 
 static void ConvertG8R8ToR8G8(std::vector<u8>& data, u32 width, u32 height) {
-    constexpr auto bpp{CachedSurface::GetGLBytesPerPixel(PixelFormat::G8R8U)};
+    constexpr auto bpp{SurfaceParams::GetBytesPerPixel(PixelFormat::G8R8U)};
     for (std::size_t y = 0; y < height; ++y) {
         for (std::size_t x = 0; x < width; ++x) {
             const std::size_t offset{bpp * (y * width + x)};
@@ -814,7 +843,7 @@ static void ConvertFormatAsNeeded_LoadGLBuffer(std::vector<u8>& data, PixelForma
     }
     case PixelFormat::S8Z24:
         // Convert the S8Z24 depth format to Z24S8, as OpenGL does not support S8Z24.
-        ConvertS8Z24ToZ24S8(data, width, height);
+        ConvertS8Z24ToZ24S8(data, width, height, false);
         break;
 
     case PixelFormat::G8R8U:
@@ -825,22 +854,36 @@ static void ConvertFormatAsNeeded_LoadGLBuffer(std::vector<u8>& data, PixelForma
     }
 }
 
+/**
+ * Helper function to perform software conversion (as needed) when flushing a buffer from OpenGL to
+ * Switch memory. This is for Maxwell pixel formats that cannot be represented as-is in OpenGL or
+ * with typical desktop GPUs.
+ */
+static void ConvertFormatAsNeeded_FlushGLBuffer(std::vector<u8>& data, PixelFormat pixel_format,
+                                                u32 width, u32 height) {
+    switch (pixel_format) {
+    case PixelFormat::G8R8U:
+    case PixelFormat::G8R8S:
+    case PixelFormat::ASTC_2D_4X4:
+    case PixelFormat::ASTC_2D_8X8: {
+        LOG_CRITICAL(HW_GPU, "Conversion of format {} after texture flushing is not implemented",
+                     static_cast<u32>(pixel_format));
+        UNREACHABLE();
+        break;
+    }
+    case PixelFormat::S8Z24:
+        // Convert the Z24S8 depth format to S8Z24, as OpenGL does not support S8Z24.
+        ConvertS8Z24ToZ24S8(data, width, height, true);
+        break;
+    }
+}
+
 MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 64, 192));
 void CachedSurface::LoadGLBuffer() {
-    ASSERT(params.type != SurfaceType::Fill);
-
-    const u8* const texture_src_data = Memory::GetPointer(params.addr);
-
-    ASSERT(texture_src_data);
-
-    const u32 bytes_per_pixel = GetGLBytesPerPixel(params.pixel_format);
-    const u32 copy_size = params.width * params.height * bytes_per_pixel;
-    const std::size_t total_size = copy_size * params.depth;
-
     MICROPROFILE_SCOPE(OpenGL_SurfaceLoad);
 
+    gl_buffer.resize(params.size_in_bytes_gl);
     if (params.is_tiled) {
-        gl_buffer.resize(total_size);
         u32 depth = params.depth;
         u32 block_depth = params.block_depth;
 
@@ -853,13 +896,29 @@ void CachedSurface::LoadGLBuffer() {
             block_depth = 1U;
         }
 
-        const std::size_t size = copy_size * depth;
-
-        morton_to_gl_fns[static_cast<std::size_t>(params.pixel_format)](
-            params.width, params.block_height, params.height, block_depth, depth, gl_buffer.data(),
-            size, params.addr);
+        if (params.target == SurfaceParams::SurfaceTarget::TextureCubemap) {
+            // TODO(Blinkhawk): Figure where this number comes from and if it's constant or depends
+            // on the objects size and/or address.
+            u64 magic_number = 0x6000;
+            u64 offset = 0;
+            u64 offset_gl = 0;
+            u64 size = params.SizeInBytesCubeFace();
+            u64 gl_size = params.SizeInBytesCubeFaceGL();
+            for (u32 i = 0; i < depth; i++) {
+                morton_to_gl_fns[static_cast<std::size_t>(params.pixel_format)](
+                    params.width, params.block_height, params.height, block_depth, 1,
+                    gl_buffer.data() + offset_gl, gl_size, params.addr + offset);
+                offset += size + magic_number;
+                offset_gl += gl_size;
+            }
+        } else {
+            morton_to_gl_fns[static_cast<std::size_t>(params.pixel_format)](
+                params.width, params.block_height, params.height, block_depth, depth,
+                gl_buffer.data(), gl_buffer.size(), params.addr);
+        }
     } else {
-        const u8* const texture_src_data_end{texture_src_data + total_size};
+        const auto texture_src_data{Memory::GetPointer(params.addr)};
+        const auto texture_src_data_end{texture_src_data + params.size_in_bytes_gl};
         gl_buffer.assign(texture_src_data, texture_src_data_end);
     }
 
@@ -868,7 +927,45 @@ void CachedSurface::LoadGLBuffer() {
 
 MICROPROFILE_DEFINE(OpenGL_SurfaceFlush, "OpenGL", "Surface Flush", MP_RGB(128, 192, 64));
 void CachedSurface::FlushGLBuffer() {
-    ASSERT_MSG(false, "Unimplemented");
+    MICROPROFILE_SCOPE(OpenGL_SurfaceFlush);
+
+    ASSERT_MSG(!IsPixelFormatASTC(params.pixel_format), "Unimplemented");
+
+    // OpenGL temporary buffer needs to be big enough to store raw texture size
+    gl_buffer.resize(GetSizeInBytes());
+
+    const FormatTuple& tuple = GetFormatTuple(params.pixel_format, params.component_type);
+    // Ensure no bad interactions with GL_UNPACK_ALIGNMENT
+    ASSERT(params.width * SurfaceParams::GetBytesPerPixel(params.pixel_format) % 4 == 0);
+    glPixelStorei(GL_PACK_ROW_LENGTH, static_cast<GLint>(params.width));
+    ASSERT(!tuple.compressed);
+    glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+    glGetTextureImage(texture.handle, 0, tuple.format, tuple.type, gl_buffer.size(),
+                      gl_buffer.data());
+    glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+    ConvertFormatAsNeeded_FlushGLBuffer(gl_buffer, params.pixel_format, params.width,
+                                        params.height);
+    ASSERT(params.type != SurfaceType::Fill);
+    const u8* const texture_src_data = Memory::GetPointer(params.addr);
+    ASSERT(texture_src_data);
+    if (params.is_tiled) {
+        u32 depth = params.depth;
+        u32 block_depth = params.block_depth;
+
+        ASSERT_MSG(params.block_width == 1, "Block width is defined as {} on texture type {}",
+                   params.block_width, static_cast<u32>(params.target));
+
+        if (params.target == SurfaceParams::SurfaceTarget::Texture2D) {
+            // TODO(Blinkhawk): Eliminate this condition once all texture types are implemented.
+            depth = 1U;
+            block_depth = 1U;
+        }
+        gl_to_morton_fns[static_cast<size_t>(params.pixel_format)](
+            params.width, params.block_height, params.height, block_depth, depth, gl_buffer.data(),
+            gl_buffer.size(), GetAddr());
+    } else {
+        std::memcpy(Memory::GetPointer(GetAddr()), gl_buffer.data(), GetSizeInBytes());
+    }
 }
 
 MICROPROFILE_DEFINE(OpenGL_TextureUL, "OpenGL", "Texture Upload", MP_RGB(128, 64, 192));
@@ -878,9 +975,6 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle
 
     MICROPROFILE_SCOPE(OpenGL_TextureUL);
 
-    ASSERT(gl_buffer.size() == static_cast<std::size_t>(params.width) * params.height *
-                                   GetGLBytesPerPixel(params.pixel_format) * params.depth);
-
     const auto& rect{params.GetRect()};
 
     // Load data from memory to the surface
@@ -889,7 +983,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle
     std::size_t buffer_offset =
         static_cast<std::size_t>(static_cast<std::size_t>(y0) * params.width +
                                  static_cast<std::size_t>(x0)) *
-        GetGLBytesPerPixel(params.pixel_format);
+        SurfaceParams::GetBytesPerPixel(params.pixel_format);
 
     const FormatTuple& tuple = GetFormatTuple(params.pixel_format, params.component_type);
     const GLuint target_tex = texture.handle;
@@ -905,7 +999,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle
     cur_state.Apply();
 
     // Ensure no bad interactions with GL_UNPACK_ALIGNMENT
-    ASSERT(params.width * GetGLBytesPerPixel(params.pixel_format) % 4 == 0);
+    ASSERT(params.width * SurfaceParams::GetBytesPerPixel(params.pixel_format) % 4 == 0);
     glPixelStorei(GL_UNPACK_ROW_LENGTH, static_cast<GLint>(params.width));
 
     glActiveTexture(GL_TEXTURE0);
@@ -915,7 +1009,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle
             glCompressedTexImage2D(
                 SurfaceTargetToGL(params.target), 0, tuple.internal_format,
                 static_cast<GLsizei>(params.width), static_cast<GLsizei>(params.height), 0,
-                static_cast<GLsizei>(params.size_in_bytes_2d), &gl_buffer[buffer_offset]);
+                static_cast<GLsizei>(params.size_in_bytes_gl), &gl_buffer[buffer_offset]);
             break;
         case SurfaceParams::SurfaceTarget::Texture3D:
         case SurfaceParams::SurfaceTarget::Texture2DArray:
@@ -923,16 +1017,16 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle
                 SurfaceTargetToGL(params.target), 0, tuple.internal_format,
                 static_cast<GLsizei>(params.width), static_cast<GLsizei>(params.height),
                 static_cast<GLsizei>(params.depth), 0,
-                static_cast<GLsizei>(params.size_in_bytes_total), &gl_buffer[buffer_offset]);
+                static_cast<GLsizei>(params.size_in_bytes_gl), &gl_buffer[buffer_offset]);
             break;
         case SurfaceParams::SurfaceTarget::TextureCubemap:
             for (std::size_t face = 0; face < params.depth; ++face) {
                 glCompressedTexImage2D(static_cast<GLenum>(GL_TEXTURE_CUBE_MAP_POSITIVE_X + face),
                                        0, tuple.internal_format, static_cast<GLsizei>(params.width),
                                        static_cast<GLsizei>(params.height), 0,
-                                       static_cast<GLsizei>(params.size_in_bytes_2d),
+                                       static_cast<GLsizei>(params.SizeInBytesCubeFaceGL()),
                                        &gl_buffer[buffer_offset]);
-                buffer_offset += params.size_in_bytes_2d;
+                buffer_offset += params.SizeInBytesCubeFace();
             }
             break;
         default:
@@ -942,7 +1036,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle
             glCompressedTexImage2D(
                 GL_TEXTURE_2D, 0, tuple.internal_format, static_cast<GLsizei>(params.width),
                 static_cast<GLsizei>(params.height), 0,
-                static_cast<GLsizei>(params.size_in_bytes_2d), &gl_buffer[buffer_offset]);
+                static_cast<GLsizei>(params.size_in_bytes_gl), &gl_buffer[buffer_offset]);
         }
     } else {
 
@@ -971,7 +1065,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle
                                 y0, static_cast<GLsizei>(rect.GetWidth()),
                                 static_cast<GLsizei>(rect.GetHeight()), tuple.format, tuple.type,
                                 &gl_buffer[buffer_offset]);
-                buffer_offset += params.size_in_bytes_2d;
+                buffer_offset += params.SizeInBytesCubeFace();
             }
             break;
         default:
@@ -1033,10 +1127,7 @@ Surface RasterizerCacheOpenGL::GetColorBufferSurface(std::size_t index, bool pre
 void RasterizerCacheOpenGL::LoadSurface(const Surface& surface) {
     surface->LoadGLBuffer();
     surface->UploadGLTexture(read_framebuffer.handle, draw_framebuffer.handle);
-}
-
-void RasterizerCacheOpenGL::FlushSurface(const Surface& surface) {
-    surface->FlushGLBuffer();
+    surface->MarkAsModified(false, *this);
 }
 
 Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool preserve_contents) {
@@ -1053,8 +1144,8 @@ Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool pres
         } else if (preserve_contents) {
             // If surface parameters changed and we care about keeping the previous data, recreate
             // the surface from the old one
-            Unregister(surface);
             Surface new_surface{RecreateSurface(surface, params)};
+            Unregister(surface);
             Register(new_surface);
             return new_surface;
         } else {
@@ -1105,6 +1196,14 @@ void RasterizerCacheOpenGL::FermiCopySurface(
     FastCopySurface(GetSurface(src_params, true), GetSurface(dst_params, false));
 }
 
+void RasterizerCacheOpenGL::AccurateCopySurface(const Surface& src_surface,
+                                                const Surface& dst_surface) {
+    const auto& src_params{src_surface->GetSurfaceParams()};
+    const auto& dst_params{dst_surface->GetSurfaceParams()};
+    FlushRegion(src_params.addr, dst_params.size_in_bytes);
+    LoadSurface(dst_surface);
+}
+
 Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface,
                                                const SurfaceParams& new_params) {
     // Verify surface is compatible for blitting
@@ -1113,6 +1212,12 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface,
     // Get a new surface with the new parameters, and blit the previous surface to it
     Surface new_surface{GetUncachedSurface(new_params)};
 
+    // With use_accurate_gpu_emulation enabled, do an accurate surface copy
+    if (Settings::values.use_accurate_gpu_emulation) {
+        AccurateCopySurface(old_surface, new_surface);
+        return new_surface;
+    }
+
     // For compatible surfaces, we can just do fast glCopyImageSubData based copy
     if (old_params.target == new_params.target && old_params.type == new_params.type &&
         old_params.depth == new_params.depth && old_params.depth == 1 &&
@@ -1124,11 +1229,10 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface,
 
     // If the format is the same, just do a framebuffer blit. This is significantly faster than
     // using PBOs. The is also likely less accurate, as textures will be converted rather than
-    // reinterpreted. When use_accurate_framebuffers setting is enabled, perform a more accurate
+    // reinterpreted. When use_accurate_gpu_emulation setting is enabled, perform a more accurate
     // surface copy, where pixels are reinterpreted as a new format (without conversion). This
     // code path uses OpenGL PBOs and is quite slow.
-    const bool is_blit{old_params.pixel_format == new_params.pixel_format ||
-                       !Settings::values.use_accurate_framebuffers};
+    const bool is_blit{old_params.pixel_format == new_params.pixel_format};
 
     switch (new_params.target) {
     case SurfaceParams::SurfaceTarget::Texture2D:
@@ -1138,41 +1242,10 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface,
             CopySurface(old_surface, new_surface, copy_pbo.handle);
         }
         break;
-    case SurfaceParams::SurfaceTarget::TextureCubemap: {
-        if (old_params.rt.array_mode != 1) {
-            // TODO(bunnei): This is used by Breath of the Wild, I'm not sure how to implement this
-            // yet (array rendering used as a cubemap texture).
-            LOG_CRITICAL(HW_GPU, "Unhandled rendertarget array_mode {}", old_params.rt.array_mode);
-            UNREACHABLE();
-            return new_surface;
-        }
-
-        // This seems to be used for render-to-cubemap texture
-        ASSERT_MSG(old_params.target == SurfaceParams::SurfaceTarget::Texture2D, "Unexpected");
-        ASSERT_MSG(old_params.pixel_format == new_params.pixel_format, "Unexpected");
-        ASSERT_MSG(old_params.rt.base_layer == 0, "Unimplemented");
-
-        // TODO(bunnei): Verify the below - this stride seems to be in 32-bit words, not pixels.
-        // Tested with Splatoon 2, Super Mario Odyssey, and Breath of the Wild.
-        const std::size_t byte_stride{old_params.rt.layer_stride * sizeof(u32)};
-
-        for (std::size_t index = 0; index < new_params.depth; ++index) {
-            Surface face_surface{TryGetReservedSurface(old_params)};
-            ASSERT_MSG(face_surface, "Unexpected");
-
-            if (is_blit) {
-                BlitSurface(face_surface, new_surface, read_framebuffer.handle,
-                            draw_framebuffer.handle, face_surface->GetSurfaceParams().rt.index,
-                            new_params.rt.index, index);
-            } else {
-                CopySurface(face_surface, new_surface, copy_pbo.handle,
-                            face_surface->GetSurfaceParams().rt.index, new_params.rt.index, index);
-            }
-
-            old_params.addr += byte_stride;
-        }
+    case SurfaceParams::SurfaceTarget::TextureCubemap:
+    case SurfaceParams::SurfaceTarget::Texture3D:
+        AccurateCopySurface(old_surface, new_surface);
         break;
-    }
     default:
         LOG_CRITICAL(Render_OpenGL, "Unimplemented surface target={}",
                      static_cast<u32>(new_params.target));

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -18,6 +18,7 @@
 #include "video_core/rasterizer_cache.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/gl_shader_gen.h"
+#include "video_core/textures/decoders.h"
 #include "video_core/textures/texture.h"
 
 namespace OpenGL {
@@ -131,6 +132,8 @@ struct SurfaceParams {
         case Tegra::Texture::TextureType::Texture2D:
         case Tegra::Texture::TextureType::Texture2DNoMipmap:
             return SurfaceTarget::Texture2D;
+        case Tegra::Texture::TextureType::Texture3D:
+            return SurfaceTarget::Texture3D;
         case Tegra::Texture::TextureType::TextureCubemap:
             return SurfaceTarget::TextureCubemap;
         case Tegra::Texture::TextureType::Texture1DArray:
@@ -701,21 +704,42 @@ struct SurfaceParams {
         return SurfaceType::Invalid;
     }
 
+    /// Returns the sizer in bytes of the specified pixel format
+    static constexpr u32 GetBytesPerPixel(PixelFormat pixel_format) {
+        if (pixel_format == SurfaceParams::PixelFormat::Invalid) {
+            return 0;
+        }
+        return GetFormatBpp(pixel_format) / CHAR_BIT;
+    }
+
     /// Returns the rectangle corresponding to this surface
     MathUtil::Rectangle<u32> GetRect() const;
 
-    /// Returns the size of this surface as a 2D texture in bytes, adjusted for compression
-    std::size_t SizeInBytes2D() const {
+    /// Returns the total size of this surface in bytes, adjusted for compression
+    std::size_t SizeInBytesRaw(bool ignore_tiled = false) const {
         const u32 compression_factor{GetCompressionFactor(pixel_format)};
-        ASSERT(width % compression_factor == 0);
-        ASSERT(height % compression_factor == 0);
-        return (width / compression_factor) * (height / compression_factor) *
-               GetFormatBpp(pixel_format) / CHAR_BIT;
+        const u32 bytes_per_pixel{GetBytesPerPixel(pixel_format)};
+        const size_t uncompressed_size{
+            Tegra::Texture::CalculateSize((ignore_tiled ? false : is_tiled), bytes_per_pixel, width,
+                                          height, depth, block_height, block_depth)};
+
+        // Divide by compression_factor^2, as height and width are factored by this
+        return uncompressed_size / (compression_factor * compression_factor);
     }
 
-    /// Returns the total size of this surface in bytes, adjusted for compression
-    std::size_t SizeInBytesTotal() const {
-        return SizeInBytes2D() * depth;
+    /// Returns the size of this surface as an OpenGL texture in bytes
+    std::size_t SizeInBytesGL() const {
+        return SizeInBytesRaw(true);
+    }
+
+    /// Returns the size of this surface as a cube face in bytes
+    std::size_t SizeInBytesCubeFace() const {
+        return size_in_bytes / 6;
+    }
+
+    /// Returns the size of this surface as an OpenGL cube face in bytes
+    std::size_t SizeInBytesCubeFaceGL() const {
+        return size_in_bytes_gl / 6;
     }
 
     /// Creates SurfaceParams from a texture configuration
@@ -742,7 +766,9 @@ struct SurfaceParams {
                         other.depth);
     }
 
-    VAddr addr;
+    /// Initializes parameters for caching, should be called after everything has been initialized
+    void InitCacheParameters(Tegra::GPUVAddr gpu_addr);
+
     bool is_tiled;
     u32 block_width;
     u32 block_height;
@@ -754,15 +780,20 @@ struct SurfaceParams {
     u32 height;
     u32 depth;
     u32 unaligned_height;
-    std::size_t size_in_bytes_total;
-    std::size_t size_in_bytes_2d;
     SurfaceTarget target;
     u32 max_mip_level;
 
+    // Parameters used for caching
+    VAddr addr;
+    Tegra::GPUVAddr gpu_addr;
+    std::size_t size_in_bytes;
+    std::size_t size_in_bytes_gl;
+
     // Render target specific parameters, not used in caching
     struct {
         u32 index;
         u32 array_mode;
+        u32 volume;
         u32 layer_stride;
         u32 base_layer;
     } rt;
@@ -775,7 +806,8 @@ struct SurfaceReserveKey : Common::HashableStruct<OpenGL::SurfaceParams> {
     static SurfaceReserveKey Create(const OpenGL::SurfaceParams& params) {
         SurfaceReserveKey res;
         res.state = params;
-        res.state.rt = {}; // Ignore rt config in caching
+        res.state.gpu_addr = {}; // Ignore GPU vaddr in caching
+        res.state.rt = {};       // Ignore rt config in caching
         return res;
     }
 };
@@ -790,16 +822,20 @@ struct hash<SurfaceReserveKey> {
 
 namespace OpenGL {
 
-class CachedSurface final {
+class CachedSurface final : public RasterizerCacheObject {
 public:
     CachedSurface(const SurfaceParams& params);
 
-    VAddr GetAddr() const {
+    VAddr GetAddr() const override {
         return params.addr;
     }
 
-    std::size_t GetSizeInBytes() const {
-        return params.size_in_bytes_total;
+    std::size_t GetSizeInBytes() const override {
+        return cached_size_in_bytes;
+    }
+
+    void Flush() override {
+        FlushGLBuffer();
     }
 
     const OGLTexture& Texture() const {
@@ -810,13 +846,6 @@ public:
         return gl_target;
     }
 
-    static constexpr unsigned int GetGLBytesPerPixel(SurfaceParams::PixelFormat format) {
-        if (format == SurfaceParams::PixelFormat::Invalid)
-            return 0;
-
-        return SurfaceParams::GetFormatBpp(format) / CHAR_BIT;
-    }
-
     const SurfaceParams& GetSurfaceParams() const {
         return params;
     }
@@ -833,6 +862,7 @@ private:
     std::vector<u8> gl_buffer;
     SurfaceParams params;
     GLenum gl_target;
+    std::size_t cached_size_in_bytes;
 };
 
 class RasterizerCacheOpenGL final : public RasterizerCache<Surface> {
@@ -849,9 +879,6 @@ public:
     /// Get the color surface based on the framebuffer configuration and the specified render target
     Surface GetColorBufferSurface(std::size_t index, bool preserve_contents);
 
-    /// Flushes the surface to Switch memory
-    void FlushSurface(const Surface& surface);
-
     /// Tries to find a framebuffer using on the provided CPU address
     Surface TryFindFramebufferSurface(VAddr addr) const;
 
@@ -875,6 +902,9 @@ private:
     /// Tries to get a reserved surface for the specified parameters
     Surface TryGetReservedSurface(const SurfaceParams& params);
 
+    /// Performs a slow but accurate surface copy, flushing to RAM and reinterpreting the data
+    void AccurateCopySurface(const Surface& src_surface, const Surface& dst_surface);
+
     /// The surface reserve is a "backup" cache, this is where we put unique surfaces that have
     /// previously been used. This is to prevent surfaces from being constantly created and
     /// destroyed when used with different surface parameters.

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -19,20 +19,21 @@ class CachedShader;
 using Shader = std::shared_ptr<CachedShader>;
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
-class CachedShader final {
+class CachedShader final : public RasterizerCacheObject {
 public:
     CachedShader(VAddr addr, Maxwell::ShaderProgram program_type);
 
-    /// Gets the address of the shader in guest memory, required for cache management
-    VAddr GetAddr() const {
+    VAddr GetAddr() const override {
         return addr;
     }
 
-    /// Gets the size of the shader in guest memory, required for cache management
-    std::size_t GetSizeInBytes() const {
+    std::size_t GetSizeInBytes() const override {
         return GLShader::MAX_PROGRAM_CODE_LENGTH * sizeof(u64);
     }
 
+    // We do not have to flush this cache as things in it are never modified by us.
+    void Flush() override {}
+
     /// Gets the shader entries for the shader
     const GLShader::ShaderEntries& GetShaderEntries() const {
         return entries;

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -165,11 +165,10 @@ private:
                     const ExitMethod jmp = Scan(target, end, labels);
                     return exit_method = ParallelExit(no_jmp, jmp);
                 }
-                case OpCode::Id::SSY:
-                case OpCode::Id::PBK: {
-                    // The SSY and PBK use a similar encoding as the BRA instruction.
+                case OpCode::Id::SSY: {
+                    // The SSY instruction uses a similar encoding as the BRA instruction.
                     ASSERT_MSG(instr.bra.constant_buffer == 0,
-                               "Constant buffer branching is not supported");
+                               "Constant buffer SSY is not supported");
                     const u32 target = offset + instr.bra.GetBranchTarget();
                     labels.insert(target);
                     // Continue scanning for an exit method.
@@ -1143,6 +1142,7 @@ private:
         case Tegra::Shader::TextureType::Texture2D: {
             return 2;
         }
+        case Tegra::Shader::TextureType::Texture3D:
         case Tegra::Shader::TextureType::TextureCube: {
             return 3;
         }
@@ -1154,27 +1154,27 @@ private:
     }
 
     /*
-     * Emits code to push the input target address to the flow address stack, incrementing the stack
+     * Emits code to push the input target address to the SSY address stack, incrementing the stack
      * top.
      */
-    void EmitPushToFlowStack(u32 target) {
+    void EmitPushToSSYStack(u32 target) {
         shader.AddLine('{');
         ++shader.scope;
-        shader.AddLine("flow_stack[flow_stack_top] = " + std::to_string(target) + "u;");
-        shader.AddLine("flow_stack_top++;");
+        shader.AddLine("ssy_stack[ssy_stack_top] = " + std::to_string(target) + "u;");
+        shader.AddLine("ssy_stack_top++;");
         --shader.scope;
         shader.AddLine('}');
     }
 
     /*
-     * Emits code to pop an address from the flow address stack, setting the jump address to the
+     * Emits code to pop an address from the SSY address stack, setting the jump address to the
      * popped address and decrementing the stack top.
      */
-    void EmitPopFromFlowStack() {
+    void EmitPopFromSSYStack() {
         shader.AddLine('{');
         ++shader.scope;
-        shader.AddLine("flow_stack_top--;");
-        shader.AddLine("jmp_to = flow_stack[flow_stack_top];");
+        shader.AddLine("ssy_stack_top--;");
+        shader.AddLine("jmp_to = ssy_stack[ssy_stack_top];");
         shader.AddLine("break;");
         --shader.scope;
         shader.AddLine('}');
@@ -2934,32 +2934,16 @@ private:
                 // The SSY opcode tells the GPU where to re-converge divergent execution paths, it
                 // sets the target of the jump that the SYNC instruction will make. The SSY opcode
                 // has a similar structure to the BRA opcode.
-                ASSERT_MSG(instr.bra.constant_buffer == 0, "Constant buffer flow is not supported");
+                ASSERT_MSG(instr.bra.constant_buffer == 0, "Constant buffer SSY is not supported");
 
                 const u32 target = offset + instr.bra.GetBranchTarget();
-                EmitPushToFlowStack(target);
-                break;
-            }
-            case OpCode::Id::PBK: {
-                // PBK pushes to a stack the address where BRK will jump to. This shares stack with
-                // SSY but using SYNC on a PBK address will kill the shader execution. We don't
-                // emulate this because it's very unlikely a driver will emit such invalid shader.
-                ASSERT_MSG(instr.bra.constant_buffer == 0, "Constant buffer PBK is not supported");
-
-                const u32 target = offset + instr.bra.GetBranchTarget();
-                EmitPushToFlowStack(target);
+                EmitPushToSSYStack(target);
                 break;
             }
             case OpCode::Id::SYNC: {
                 // The SYNC opcode jumps to the address previously set by the SSY opcode
                 ASSERT(instr.flow.cond == Tegra::Shader::FlowCondition::Always);
-                EmitPopFromFlowStack();
-                break;
-            }
-            case OpCode::Id::BRK: {
-                // The BRK opcode jumps to the address previously set by the PBK opcode
-                ASSERT(instr.flow.cond == Tegra::Shader::FlowCondition::Always);
-                EmitPopFromFlowStack();
+                EmitPopFromSSYStack();
                 break;
             }
             case OpCode::Id::DEPBAR: {
@@ -3113,11 +3097,11 @@ private:
                 labels.insert(subroutine.begin);
                 shader.AddLine("uint jmp_to = " + std::to_string(subroutine.begin) + "u;");
 
-                // TODO(Subv): Figure out the actual depth of the flow stack, for now it seems
-                // unlikely that shaders will use 20 nested SSYs and PBKs.
-                constexpr u32 FLOW_STACK_SIZE = 20;
-                shader.AddLine("uint flow_stack[" + std::to_string(FLOW_STACK_SIZE) + "];");
-                shader.AddLine("uint flow_stack_top = 0u;");
+                // TODO(Subv): Figure out the actual depth of the SSY stack, for now it seems
+                // unlikely that shaders will use 20 nested SSYs.
+                constexpr u32 SSY_STACK_SIZE = 20;
+                shader.AddLine("uint ssy_stack[" + std::to_string(SSY_STACK_SIZE) + "];");
+                shader.AddLine("uint ssy_stack_top = 0u;");
 
                 shader.AddLine("while (true) {");
                 ++shader.scope;

src/yuzu/configuration/config.cpp

@@ -85,8 +85,8 @@ void Config::ReadValues() {
     Settings::values.resolution_factor = qt_config->value("resolution_factor", 1.0).toFloat();
     Settings::values.use_frame_limit = qt_config->value("use_frame_limit", true).toBool();
     Settings::values.frame_limit = qt_config->value("frame_limit", 100).toInt();
-    Settings::values.use_accurate_framebuffers =
-        qt_config->value("use_accurate_framebuffers", false).toBool();
+    Settings::values.use_accurate_gpu_emulation =
+        qt_config->value("use_accurate_gpu_emulation", false).toBool();
 
     Settings::values.bg_red = qt_config->value("bg_red", 0.0).toFloat();
     Settings::values.bg_green = qt_config->value("bg_green", 0.0).toFloat();
@@ -233,7 +233,7 @@ void Config::SaveValues() {
     qt_config->setValue("resolution_factor", (double)Settings::values.resolution_factor);
     qt_config->setValue("use_frame_limit", Settings::values.use_frame_limit);
     qt_config->setValue("frame_limit", Settings::values.frame_limit);
-    qt_config->setValue("use_accurate_framebuffers", Settings::values.use_accurate_framebuffers);
+    qt_config->setValue("use_accurate_gpu_emulation", Settings::values.use_accurate_gpu_emulation);
 
     // Cast to double because Qt's written float values are not human-readable
     qt_config->setValue("bg_red", (double)Settings::values.bg_red);

src/yuzu/configuration/configure_graphics.cpp

@@ -75,7 +75,7 @@ void ConfigureGraphics::setConfiguration() {
         static_cast<int>(FromResolutionFactor(Settings::values.resolution_factor)));
     ui->toggle_frame_limit->setChecked(Settings::values.use_frame_limit);
     ui->frame_limit->setValue(Settings::values.frame_limit);
-    ui->use_accurate_framebuffers->setChecked(Settings::values.use_accurate_framebuffers);
+    ui->use_accurate_gpu_emulation->setChecked(Settings::values.use_accurate_gpu_emulation);
     bg_color = QColor::fromRgbF(Settings::values.bg_red, Settings::values.bg_green,
                                 Settings::values.bg_blue);
     ui->bg_button->setStyleSheet(
@@ -87,7 +87,7 @@ void ConfigureGraphics::applyConfiguration() {
         ToResolutionFactor(static_cast<Resolution>(ui->resolution_factor_combobox->currentIndex()));
     Settings::values.use_frame_limit = ui->toggle_frame_limit->isChecked();
     Settings::values.frame_limit = ui->frame_limit->value();
-    Settings::values.use_accurate_framebuffers = ui->use_accurate_framebuffers->isChecked();
+    Settings::values.use_accurate_gpu_emulation = ui->use_accurate_gpu_emulation->isChecked();
     Settings::values.bg_red = static_cast<float>(bg_color.redF());
     Settings::values.bg_green = static_cast<float>(bg_color.greenF());
     Settings::values.bg_blue = static_cast<float>(bg_color.blueF());

src/yuzu/configuration/configure_graphics.ui

@@ -50,9 +50,9 @@
           </layout>
         </item>
         <item>
-         <widget class="QCheckBox" name="use_accurate_framebuffers">
+         <widget class="QCheckBox" name="use_accurate_gpu_emulation">
           <property name="text">
-           <string>Use accurate framebuffers (slow)</string>
+           <string>Use accurate GPU emulation (slow)</string>
           </property>
          </widget>
         </item>

src/yuzu_cmd/config.cpp

@@ -99,8 +99,8 @@ void Config::ReadValues() {
     Settings::values.use_frame_limit = sdl2_config->GetBoolean("Renderer", "use_frame_limit", true);
     Settings::values.frame_limit =
         static_cast<u16>(sdl2_config->GetInteger("Renderer", "frame_limit", 100));
-    Settings::values.use_accurate_framebuffers =
-        sdl2_config->GetBoolean("Renderer", "use_accurate_framebuffers", false);
+    Settings::values.use_accurate_gpu_emulation =
+        sdl2_config->GetBoolean("Renderer", "use_accurate_gpu_emulation", false);
 
     Settings::values.bg_red = (float)sdl2_config->GetReal("Renderer", "bg_red", 0.0);
     Settings::values.bg_green = (float)sdl2_config->GetReal("Renderer", "bg_green", 0.0);

src/yuzu_cmd/default_ini.h

@@ -110,9 +110,9 @@ use_frame_limit =
 # 1 - 9999: Speed limit as a percentage of target game speed. 100 (default)
 frame_limit =
 
-# Whether to use accurate framebuffers
+# Whether to use accurate GPU emulation
 # 0 (default): Off (fast), 1 : On (slow)
-use_accurate_framebuffers =
+use_accurate_gpu_emulation =
 
 # The clear color for the renderer. What shows up on the sides of the bottom screen.
 # Must be in range of 0.0-1.0. Defaults to 1.0 for all.