Merge pull request #265 from bunnei/tegra-progress-2

Tegra progress 2

src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp

@@ -26,14 +26,13 @@ void nvdisp_disp0::flip(u32 buffer_handle, u32 offset, u32 format, u32 width, u3
                 "Drawing from address %lx offset %08X Width %u Height %u Stride %u Format %u", addr,
                 offset, width, height, stride, format);
 
-    using PixelFormat = RendererBase::FramebufferInfo::PixelFormat;
-    using Flags = NVFlinger::BufferQueue::BufferTransformFlags;
-    const bool flip_vertical = static_cast<u32>(transform) & static_cast<u32>(Flags::FlipV);
-    const RendererBase::FramebufferInfo framebuffer_info{
-        addr, offset, width, height, stride, static_cast<PixelFormat>(format), flip_vertical};
+    using PixelFormat = Tegra::FramebufferConfig::PixelFormat;
+    const Tegra::FramebufferConfig framebuffer{
+        addr, offset, width, height, stride, static_cast<PixelFormat>(format), transform};
 
     Core::System::GetInstance().perf_stats.EndGameFrame();
-    VideoCore::g_renderer->SwapBuffers(framebuffer_info);
+
+    VideoCore::g_renderer->SwapBuffers(framebuffer);
 }
 
 } // namespace Devices

src/core/hle/service/nvflinger/buffer_queue.h

@@ -47,6 +47,8 @@ public:
     ~BufferQueue() = default;
 
     enum class BufferTransformFlags : u32 {
+        /// No transform flags are set
+        Unset = 0x00,
         /// Flip source image horizontally (around the vertical axis)
         FlipH = 0x01,
         /// Flip source image vertically (around the horizontal axis)

src/core/memory.cpp

@@ -41,6 +41,9 @@ static void MapPages(PageTable& page_table, VAddr base, u64 size, u8* memory, Pa
     LOG_DEBUG(HW_Memory, "Mapping %p onto %016" PRIX64 "-%016" PRIX64, memory, base * PAGE_SIZE,
               (base + size) * PAGE_SIZE);
 
+    RasterizerFlushVirtualRegion(base << PAGE_BITS, size * PAGE_SIZE,
+                                 FlushMode::FlushAndInvalidate);
+
     VAddr end = base + size;
     while (base != end) {
         ASSERT_MSG(base < PAGE_TABLE_NUM_ENTRIES, "out of range mapping at %016" PRIX64, base);
@@ -288,6 +291,43 @@ u8* GetPhysicalPointer(PAddr address) {
     return target_pointer;
 }
 
+void RasterizerFlushVirtualRegion(VAddr start, u64 size, FlushMode mode) {
+    // Since pages are unmapped on shutdown after video core is shutdown, the renderer may be
+    // null here
+    if (VideoCore::g_renderer == nullptr) {
+        return;
+    }
+
+    VAddr end = start + size;
+
+    auto CheckRegion = [&](VAddr region_start, VAddr region_end) {
+        if (start >= region_end || end <= region_start) {
+            // No overlap with region
+            return;
+        }
+
+        VAddr overlap_start = std::max(start, region_start);
+        VAddr overlap_end = std::min(end, region_end);
+        u64 overlap_size = overlap_end - overlap_start;
+
+        auto* rasterizer = VideoCore::g_renderer->Rasterizer();
+        switch (mode) {
+        case FlushMode::Flush:
+            rasterizer->FlushRegion(overlap_start, overlap_size);
+            break;
+        case FlushMode::Invalidate:
+            rasterizer->InvalidateRegion(overlap_start, overlap_size);
+            break;
+        case FlushMode::FlushAndInvalidate:
+            rasterizer->FlushAndInvalidateRegion(overlap_start, overlap_size);
+            break;
+        }
+    };
+
+    CheckRegion(PROCESS_IMAGE_VADDR, PROCESS_IMAGE_VADDR_END);
+    CheckRegion(HEAP_VADDR, HEAP_VADDR_END);
+}
+
 u8 Read8(const VAddr addr) {
     return Read<u8>(addr);
 }

src/core/memory.h

@@ -36,7 +36,10 @@ enum class PageType : u8 {
     Unmapped,
     /// Page is mapped to regular memory. This is the only type you can get pointers to.
     Memory,
-    /// Page is mapped to a memory hook, which intercepts read and write requests.
+    /// Page is mapped to regular memory, but also needs to check for rasterizer cache flushing and
+    /// invalidation
+    RasterizerCachedMemory,
+    /// Page is mapped to a I/O region. Writing and reading to this page is handled by functions.
     Special,
 };
 
@@ -242,4 +245,19 @@ boost::optional<VAddr> PhysicalToVirtualAddress(PAddr addr);
  */
 u8* GetPhysicalPointer(PAddr address);
 
+enum class FlushMode {
+    /// Write back modified surfaces to RAM
+    Flush,
+    /// Remove region from the cache
+    Invalidate,
+    /// Write back modified surfaces to RAM, and also remove them from the cache
+    FlushAndInvalidate,
+};
+
+/**
+ * Flushes and invalidates any externally cached rasterizer resources touching the given virtual
+ * address region.
+ */
+void RasterizerFlushVirtualRegion(VAddr start, u64 size, FlushMode mode);
+
 } // namespace Memory

src/video_core/engines/maxwell_3d.h

@@ -7,6 +7,7 @@
 #include <array>
 #include <unordered_map>
 #include <vector>
+#include "common/assert.h"
 #include "common/bit_field.h"
 #include "common/common_funcs.h"
 #include "common/common_types.h"
@@ -62,6 +63,107 @@ public:
             Fragment = 4,
         };
 
+        enum class VertexSize : u32 {
+            Size_32_32_32_32 = 0x01,
+            Size_32_32_32 = 0x02,
+            Size_16_16_16_16 = 0x03,
+            Size_32_32 = 0x04,
+            Size_16_16_16 = 0x05,
+            Size_8_8_8_8 = 0x0a,
+            Size_16_16 = 0x0f,
+            Size_32 = 0x12,
+            Size_8_8_8 = 0x13,
+            Size_8_8 = 0x18,
+            Size_16 = 0x1b,
+            Size_8 = 0x1d,
+            Size_10_10_10_2 = 0x30,
+            Size_11_11_10 = 0x31,
+        };
+
+        static std::string VertexSizeToString(VertexSize vertex_size) {
+            switch (vertex_size) {
+            case VertexSize::Size_32_32_32_32:
+                return "32_32_32_32";
+            case VertexSize::Size_32_32_32:
+                return "32_32_32";
+            case VertexSize::Size_16_16_16_16:
+                return "16_16_16_16";
+            case VertexSize::Size_32_32:
+                return "32_32";
+            case VertexSize::Size_16_16_16:
+                return "16_16_16";
+            case VertexSize::Size_8_8_8_8:
+                return "8_8_8_8";
+            case VertexSize::Size_16_16:
+                return "16_16";
+            case VertexSize::Size_32:
+                return "32";
+            case VertexSize::Size_8_8_8:
+                return "8_8_8";
+            case VertexSize::Size_8_8:
+                return "8_8";
+            case VertexSize::Size_16:
+                return "16";
+            case VertexSize::Size_8:
+                return "8";
+            case VertexSize::Size_10_10_10_2:
+                return "10_10_10_2";
+            case VertexSize::Size_11_11_10:
+                return "11_11_10";
+            }
+            UNIMPLEMENTED();
+            return {};
+        }
+
+        enum class VertexType : u32 {
+            SignedNorm = 1,
+            UnsignedNorm = 2,
+            SignedInt = 3,
+            UnsignedInt = 4,
+            UnsignedScaled = 5,
+            SignedScaled = 6,
+            Float = 7,
+        };
+
+        static std::string VertexTypeToString(VertexType vertex_type) {
+            switch (vertex_type) {
+            case VertexType::SignedNorm:
+                return "SignedNorm";
+            case VertexType::UnsignedNorm:
+                return "UnsignedNorm";
+            case VertexType::SignedInt:
+                return "SignedInt";
+            case VertexType::UnsignedInt:
+                return "UnsignedInt";
+            case VertexType::UnsignedScaled:
+                return "UnsignedScaled";
+            case VertexType::SignedScaled:
+                return "SignedScaled";
+            case VertexType::Float:
+                return "Float";
+            }
+            UNIMPLEMENTED();
+            return {};
+        }
+
+        enum class PrimitiveTopology : u32 {
+            Points = 0x0,
+            Lines = 0x1,
+            LineLoop = 0x2,
+            LineStrip = 0x3,
+            Triangles = 0x4,
+            TriangleStrip = 0x5,
+            TriangleFan = 0x6,
+            Quads = 0x7,
+            QuadStrip = 0x8,
+            Polygon = 0x9,
+            LinesAdjacency = 0xa,
+            LineStripAdjacency = 0xb,
+            TrianglesAdjacency = 0xc,
+            TriangleStripAdjacency = 0xd,
+            Patches = 0xe,
+        };
+
         union {
             struct {
                 INSERT_PADDING_WORDS(0x200);
@@ -112,8 +214,8 @@ public:
                     BitField<0, 5, u32> buffer;
                     BitField<6, 1, u32> constant;
                     BitField<7, 14, u32> offset;
-                    BitField<21, 6, u32> size;
-                    BitField<27, 3, u32> type;
+                    BitField<21, 6, VertexSize> size;
+                    BitField<27, 3, VertexType> type;
                     BitField<31, 1, u32> bgra;
                 } vertex_attrib_format[NumVertexAttributes];
 
@@ -163,13 +265,15 @@ public:
                     }
                 } code_address;
                 INSERT_PADDING_WORDS(1);
+
                 struct {
                     u32 vertex_end_gl;
                     union {
                         u32 vertex_begin_gl;
-                        BitField<0, 16, u32> topology;
+                        BitField<0, 16, PrimitiveTopology> topology;
                     };
                 } draw;
+
                 INSERT_PADDING_WORDS(0x139);
                 struct {
                     u32 query_address_high;

src/video_core/gpu.h

@@ -8,10 +8,42 @@
 #include <unordered_map>
 #include <vector>
 #include "common/common_types.h"
+#include "core/hle/service/nvflinger/buffer_queue.h"
 #include "video_core/memory_manager.h"
 
 namespace Tegra {
 
+/**
+ * Struct describing framebuffer configuration
+ */
+struct FramebufferConfig {
+    enum class PixelFormat : u32 {
+        ABGR8 = 1,
+    };
+
+    /**
+     * Returns the number of bytes per pixel.
+     */
+    static u32 BytesPerPixel(PixelFormat format) {
+        switch (format) {
+        case PixelFormat::ABGR8:
+            return 4;
+        }
+
+        UNREACHABLE();
+    }
+
+    VAddr address;
+    u32 offset;
+    u32 width;
+    u32 height;
+    u32 stride;
+    PixelFormat pixel_format;
+
+    using TransformFlags = Service::NVFlinger::BufferQueue::BufferTransformFlags;
+    TransformFlags transform_flags;
+};
+
 namespace Engines {
 class Fermi2D;
 class Maxwell3D;
@@ -36,6 +68,10 @@ public:
 
     std::unique_ptr<MemoryManager> memory_manager;
 
+    Engines::Maxwell3D& Maxwell3D() {
+        return *maxwell_3d;
+    }
+
 private:
     static constexpr u32 InvalidGraphMacroEntry = 0xFFFFFFFF;
 

src/video_core/rasterizer_interface.h

@@ -5,6 +5,7 @@
 #pragma once
 
 #include "common/common_types.h"
+#include "video_core/gpu.h"
 
 struct ScreenInfo;
 
@@ -24,14 +25,14 @@ public:
     virtual void FlushAll() = 0;
 
     /// Notify rasterizer that any caches of the specified region should be flushed to 3DS memory
-    virtual void FlushRegion(PAddr addr, u32 size) = 0;
+    virtual void FlushRegion(VAddr addr, u64 size) = 0;
 
     /// Notify rasterizer that any caches of the specified region should be invalidated
-    virtual void InvalidateRegion(PAddr addr, u32 size) = 0;
+    virtual void InvalidateRegion(VAddr addr, u64 size) = 0;
 
     /// Notify rasterizer that any caches of the specified region should be flushed to 3DS memory
     /// and invalidated
-    virtual void FlushAndInvalidateRegion(PAddr addr, u32 size) = 0;
+    virtual void FlushAndInvalidateRegion(VAddr addr, u64 size) = 0;
 
     /// Attempt to use a faster method to perform a display transfer with is_texture_copy = 0
     virtual bool AccelerateDisplayTransfer(const void* config) {
@@ -49,7 +50,8 @@ public:
     }
 
     /// Attempt to use a faster method to display the framebuffer to screen
-    virtual bool AccelerateDisplay(const void* config, PAddr framebuffer_addr, u32 pixel_stride,
+    virtual bool AccelerateDisplay(const Tegra::FramebufferConfig& framebuffer,
+                                   VAddr framebuffer_addr, u32 pixel_stride,
                                    ScreenInfo& screen_info) {
         return false;
     }

src/video_core/renderer_base.cpp

@@ -5,6 +5,11 @@
 #include <atomic>
 #include <memory>
 #include "video_core/renderer_base.h"
+#include "video_core/renderer_opengl/gl_rasterizer.h"
 #include "video_core/video_core.h"
 
-void RendererBase::RefreshRasterizerSetting() {}
+void RendererBase::RefreshRasterizerSetting() {
+    if (rasterizer == nullptr) {
+        rasterizer = std::make_unique<RasterizerOpenGL>();
+    }
+}

src/video_core/renderer_base.h

@@ -8,6 +8,8 @@
 #include <boost/optional.hpp>
 #include "common/assert.h"
 #include "common/common_types.h"
+#include "video_core/gpu.h"
+#include "video_core/rasterizer_interface.h"
 
 class EmuWindow;
 
@@ -16,40 +18,10 @@ public:
     /// Used to reference a framebuffer
     enum kFramebuffer { kFramebuffer_VirtualXFB = 0, kFramebuffer_EFB, kFramebuffer_Texture };
 
-    /**
-     * Struct describing framebuffer metadata
-     * TODO(bunnei): This struct belongs in the GPU code, but we don't have a good place for it yet.
-     */
-    struct FramebufferInfo {
-        enum class PixelFormat : u32 {
-            ABGR8 = 1,
-        };
-
-        /**
-         * Returns the number of bytes per pixel.
-         */
-        static u32 BytesPerPixel(PixelFormat format) {
-            switch (format) {
-            case PixelFormat::ABGR8:
-                return 4;
-            }
-
-            UNREACHABLE();
-        }
-
-        VAddr address;
-        u32 offset;
-        u32 width;
-        u32 height;
-        u32 stride;
-        PixelFormat pixel_format;
-        bool flip_vertical;
-    };
-
     virtual ~RendererBase() {}
 
     /// Swap buffers (render frame)
-    virtual void SwapBuffers(boost::optional<const FramebufferInfo&> framebuffer_info) = 0;
+    virtual void SwapBuffers(boost::optional<const Tegra::FramebufferConfig&> framebuffer) = 0;
 
     /**
      * Set the emulator window to use for renderer
@@ -74,12 +46,16 @@ public:
         return m_current_frame;
     }
 
+    VideoCore::RasterizerInterface* Rasterizer() const {
+        return rasterizer.get();
+    }
+
     void RefreshRasterizerSetting();
 
 protected:
+    std::unique_ptr<VideoCore::RasterizerInterface> rasterizer;
     f32 m_current_fps = 0.0f; ///< Current framerate, should be set by the renderer
     int m_current_frame = 0;  ///< Current frame, should be set by the renderer
 
 private:
-    bool opengl_rasterizer_active = false;
 };

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -54,6 +54,8 @@ static void SetShaderUniformBlockBindings(GLuint shader) {
 }
 
 RasterizerOpenGL::RasterizerOpenGL() {
+    shader_dirty = true;
+
     has_ARB_buffer_storage = false;
     has_ARB_direct_state_access = false;
     has_ARB_separate_shader_objects = false;
@@ -106,8 +108,6 @@ RasterizerOpenGL::RasterizerOpenGL() {
         state.draw.vertex_buffer = stream_buffer->GetHandle();
 
         pipeline.Create();
-        vs_input_index_min = 0;
-        vs_input_index_max = 0;
         state.draw.program_pipeline = pipeline.handle;
         state.draw.shader_program = 0;
         state.draw.vertex_array = hw_vao.handle;
@@ -120,20 +120,14 @@ RasterizerOpenGL::RasterizerOpenGL() {
         glBufferData(GL_UNIFORM_BUFFER, sizeof(VSUniformData), nullptr, GL_STREAM_COPY);
         glBindBufferBase(GL_UNIFORM_BUFFER, 1, vs_uniform_buffer.handle);
     } else {
-        UNIMPLEMENTED();
+        ASSERT_MSG(false, "Unimplemented");
     }
 
     accelerate_draw = AccelDraw::Disabled;
 
     glEnable(GL_BLEND);
 
-    // Sync fixed function OpenGL state
-    SyncClipEnabled();
-    SyncClipCoef();
-    SyncCullMode();
-    SyncBlendEnabled();
-    SyncBlendFuncs();
-    SyncBlendColor();
+    LOG_WARNING(HW_GPU, "Sync fixed function OpenGL state here when ready");
 }
 
 RasterizerOpenGL::~RasterizerOpenGL() {
@@ -167,12 +161,12 @@ void RasterizerOpenGL::SetupVertexShader(VSUniformData* ub_ptr, GLintptr buffer_
 
 void RasterizerOpenGL::SetupFragmentShader(FSUniformData* ub_ptr, GLintptr buffer_offset) {
     MICROPROFILE_SCOPE(OpenGL_FS);
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 bool RasterizerOpenGL::AccelerateDrawBatch(bool is_indexed) {
     if (!has_ARB_separate_shader_objects) {
-        UNIMPLEMENTED();
+        ASSERT_MSG(false, "Unimplemented");
         return false;
     }
 
@@ -194,17 +188,17 @@ void RasterizerOpenGL::FlushAll() {
     res_cache.FlushAll();
 }
 
-void RasterizerOpenGL::FlushRegion(PAddr addr, u32 size) {
+void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
     res_cache.FlushRegion(addr, size);
 }
 
-void RasterizerOpenGL::InvalidateRegion(PAddr addr, u32 size) {
+void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
     res_cache.InvalidateRegion(addr, size, nullptr);
 }
 
-void RasterizerOpenGL::FlushAndInvalidateRegion(PAddr addr, u32 size) {
+void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) {
     MICROPROFILE_SCOPE(OpenGL_CacheManagement);
     res_cache.FlushRegion(addr, size);
     res_cache.InvalidateRegion(addr, size, nullptr);
@@ -212,58 +206,144 @@ void RasterizerOpenGL::FlushAndInvalidateRegion(PAddr addr, u32 size) {
 
 bool RasterizerOpenGL::AccelerateDisplayTransfer(const void* config) {
     MICROPROFILE_SCOPE(OpenGL_Blits);
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
     return true;
 }
 
 bool RasterizerOpenGL::AccelerateTextureCopy(const void* config) {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
     return true;
 }
 
 bool RasterizerOpenGL::AccelerateFill(const void* config) {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
     return true;
 }
 
-bool RasterizerOpenGL::AccelerateDisplay(const void* config, PAddr framebuffer_addr,
-                                         u32 pixel_stride, ScreenInfo& screen_info) {
-    UNIMPLEMENTED();
+bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& framebuffer,
+                                         VAddr framebuffer_addr, u32 pixel_stride,
+                                         ScreenInfo& screen_info) {
+    if (framebuffer_addr == 0) {
+        return false;
+    }
+    MICROPROFILE_SCOPE(OpenGL_CacheManagement);
+
+    SurfaceParams src_params;
+    src_params.addr = framebuffer_addr;
+    src_params.width = std::min(framebuffer.width, pixel_stride);
+    src_params.height = framebuffer.height;
+    src_params.stride = pixel_stride;
+    src_params.is_tiled = false;
+    src_params.pixel_format =
+        SurfaceParams::PixelFormatFromGPUPixelFormat(framebuffer.pixel_format);
+    src_params.UpdateParams();
+
+    MathUtil::Rectangle<u32> src_rect;
+    Surface src_surface;
+    std::tie(src_surface, src_rect) =
+        res_cache.GetSurfaceSubRect(src_params, ScaleMatch::Ignore, true);
+
+    if (src_surface == nullptr) {
+        return false;
+    }
+
+    u32 scaled_width = src_surface->GetScaledWidth();
+    u32 scaled_height = src_surface->GetScaledHeight();
+
+    screen_info.display_texcoords = MathUtil::Rectangle<float>(
+        (float)src_rect.bottom / (float)scaled_height, (float)src_rect.left / (float)scaled_width,
+        (float)src_rect.top / (float)scaled_height, (float)src_rect.right / (float)scaled_width);
+
+    screen_info.display_texture = src_surface->texture.handle;
+
     return true;
 }
 
 void RasterizerOpenGL::SetShader() {
-    UNIMPLEMENTED();
+    // TODO(bunnei): The below sets up a static test shader for passing untransformed vertices to
+    // OpenGL for rendering. This should be removed/replaced when we start emulating Maxwell
+    // shaders.
+
+    static constexpr char vertex_shader[] = R"(
+#version 150 core
+
+in vec2 vert_position;
+in vec2 vert_tex_coord;
+out vec2 frag_tex_coord;
+
+void main() {
+    // Multiply input position by the rotscale part of the matrix and then manually translate by
+    // the last column. This is equivalent to using a full 3x3 matrix and expanding the vector
+    // to `vec3(vert_position.xy, 1.0)`
+    gl_Position = vec4(mat2(mat3x2(0.0015625f, 0.0, 0.0, -0.0027778, -1.0, 1.0)) * vert_position + mat3x2(0.0015625f, 0.0, 0.0, -0.0027778, -1.0, 1.0)[2], 0.0, 1.0);
+    frag_tex_coord = vert_tex_coord;
+}
+)";
+
+    static constexpr char fragment_shader[] = R"(
+#version 150 core
+
+in vec2 frag_tex_coord;
+out vec4 color;
+
+uniform sampler2D color_texture;
+
+void main() {
+    color = vec4(1.0, 0.0, 1.0, 0.0);
+}
+)";
+
+    if (current_shader) {
+        return;
+    }
+
+    LOG_ERROR(HW_GPU, "Emulated shaders are not supported! Using a passthrough shader.");
+
+    current_shader = &test_shader;
+    if (has_ARB_separate_shader_objects) {
+        test_shader.shader.Create(vertex_shader, nullptr, fragment_shader, {}, true);
+        glActiveShaderProgram(pipeline.handle, test_shader.shader.handle);
+    } else {
+        ASSERT_MSG(false, "Unimplemented");
+    }
+
+    state.draw.shader_program = test_shader.shader.handle;
+    state.Apply();
+
+    if (has_ARB_separate_shader_objects) {
+        state.draw.shader_program = 0;
+        state.Apply();
+    }
 }
 
 void RasterizerOpenGL::SyncClipEnabled() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 void RasterizerOpenGL::SyncClipCoef() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 void RasterizerOpenGL::SyncCullMode() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 void RasterizerOpenGL::SyncDepthScale() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 void RasterizerOpenGL::SyncDepthOffset() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 void RasterizerOpenGL::SyncBlendEnabled() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 void RasterizerOpenGL::SyncBlendFuncs() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }
 
 void RasterizerOpenGL::SyncBlendColor() {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
 }

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -32,16 +32,22 @@ public:
     void DrawTriangles() override;
     void NotifyMaxwellRegisterChanged(u32 id) override;
     void FlushAll() override;
-    void FlushRegion(PAddr addr, u32 size) override;
-    void InvalidateRegion(PAddr addr, u32 size) override;
-    void FlushAndInvalidateRegion(PAddr addr, u32 size) override;
+    void FlushRegion(VAddr addr, u64 size) override;
+    void InvalidateRegion(VAddr addr, u64 size) override;
+    void FlushAndInvalidateRegion(VAddr addr, u64 size) override;
     bool AccelerateDisplayTransfer(const void* config) override;
     bool AccelerateTextureCopy(const void* config) override;
     bool AccelerateFill(const void* config) override;
-    bool AccelerateDisplay(const void* config, PAddr framebuffer_addr, u32 pixel_stride,
-                           ScreenInfo& screen_info) override;
+    bool AccelerateDisplay(const Tegra::FramebufferConfig& framebuffer, VAddr framebuffer_addr,
+                           u32 pixel_stride, ScreenInfo& screen_info) override;
     bool AccelerateDrawBatch(bool is_indexed) override;
 
+    /// OpenGL shader generated for a given Maxwell register state
+    struct MaxwellShader {
+        /// OpenGL shader resource
+        OGLShader shader;
+    };
+
     struct VertexShader {
         OGLShader shader;
     };
@@ -117,6 +123,12 @@ private:
 
     RasterizerCacheOpenGL res_cache;
 
+    /// Shader used for test renderering - to be removed once we have emulated shaders
+    MaxwellShader test_shader{};
+
+    const MaxwellShader* current_shader{};
+    bool shader_dirty{};
+
     struct {
         UniformData data;
         bool dirty;
@@ -136,8 +148,6 @@ private:
     static constexpr size_t STREAM_BUFFER_SIZE = 4 * 1024 * 1024;
     std::unique_ptr<OGLStreamBuffer> stream_buffer;
 
-    GLint vs_input_index_min;
-    GLint vs_input_index_max;
     GLsizeiptr vs_input_size;
 
     void AnalyzeVertexArray(bool is_indexed);

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -22,6 +22,7 @@
 #include "common/scope_exit.h"
 #include "common/vector_math.h"
 #include "core/frontend/emu_window.h"
+#include "core/hle/kernel/vm_manager.h"
 #include "core/memory.h"
 #include "core/settings.h"
 #include "video_core/renderer_opengl/gl_rasterizer_cache.h"
@@ -107,7 +108,7 @@ static void MortonCopyTile(u32 stride, u8* tile_buffer, u8* gl_buffer) {
 }
 
 template <bool morton_to_gl, PixelFormat format>
-static void MortonCopy(u32 stride, u32 height, u8* gl_buffer, PAddr base, PAddr start, PAddr end) {
+static void MortonCopy(u32 stride, u32 height, u8* gl_buffer, VAddr base, VAddr start, VAddr end) {
     constexpr u32 bytes_per_pixel = SurfaceParams::GetFormatBpp(format) / 8;
     constexpr u32 tile_size = bytes_per_pixel * 64;
 
@@ -115,9 +116,9 @@ static void MortonCopy(u32 stride, u32 height, u8* gl_buffer, PAddr base, PAddr
     static_assert(gl_bytes_per_pixel >= bytes_per_pixel, "");
     gl_buffer += gl_bytes_per_pixel - bytes_per_pixel;
 
-    const PAddr aligned_down_start = base + Common::AlignDown(start - base, tile_size);
-    const PAddr aligned_start = base + Common::AlignUp(start - base, tile_size);
-    const PAddr aligned_end = base + Common::AlignDown(end - base, tile_size);
+    const VAddr aligned_down_start = base + Common::AlignDown(start - base, tile_size);
+    const VAddr aligned_start = base + Common::AlignUp(start - base, tile_size);
+    const VAddr aligned_end = base + Common::AlignDown(end - base, tile_size);
 
     ASSERT(!morton_to_gl || (aligned_start == start && aligned_end == end));
 
@@ -136,7 +137,7 @@ static void MortonCopy(u32 stride, u32 height, u8* gl_buffer, PAddr base, PAddr
         }
     };
 
-    u8* tile_buffer = Memory::GetPhysicalPointer(start);
+    u8* tile_buffer = Memory::GetPointer(start);
 
     if (start < aligned_start && !morton_to_gl) {
         std::array<u8, tile_size> tmp_buf;
@@ -162,7 +163,7 @@ static void MortonCopy(u32 stride, u32 height, u8* gl_buffer, PAddr base, PAddr
     }
 }
 
-static constexpr std::array<void (*)(u32, u32, u8*, PAddr, PAddr, PAddr), 18> morton_to_gl_fns = {
+static constexpr std::array<void (*)(u32, u32, u8*, VAddr, VAddr, VAddr), 18> morton_to_gl_fns = {
     MortonCopy<true, PixelFormat::RGBA8>,  // 0
     MortonCopy<true, PixelFormat::RGB8>,   // 1
     MortonCopy<true, PixelFormat::RGB5A1>, // 2
@@ -183,7 +184,7 @@ static constexpr std::array<void (*)(u32, u32, u8*, PAddr, PAddr, PAddr), 18> mo
     MortonCopy<true, PixelFormat::D24S8> // 17
 };
 
-static constexpr std::array<void (*)(u32, u32, u8*, PAddr, PAddr, PAddr), 18> gl_to_morton_fns = {
+static constexpr std::array<void (*)(u32, u32, u8*, VAddr, VAddr, VAddr), 18> gl_to_morton_fns = {
     MortonCopy<false, PixelFormat::RGBA8>,  // 0
     MortonCopy<false, PixelFormat::RGB8>,   // 1
     MortonCopy<false, PixelFormat::RGB5A1>, // 2
@@ -290,7 +291,7 @@ static bool BlitTextures(GLuint src_tex, const MathUtil::Rectangle<u32>& src_rec
 
 static bool FillSurface(const Surface& surface, const u8* fill_data,
                         const MathUtil::Rectangle<u32>& fill_rect, GLuint draw_fb_handle) {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
     return true;
 }
 
@@ -298,9 +299,9 @@ SurfaceParams SurfaceParams::FromInterval(SurfaceInterval interval) const {
     SurfaceParams params = *this;
     const u32 tiled_size = is_tiled ? 8 : 1;
     const u64 stride_tiled_bytes = BytesInPixels(stride * tiled_size);
-    PAddr aligned_start =
+    VAddr aligned_start =
         addr + Common::AlignDown(boost::icl::first(interval) - addr, stride_tiled_bytes);
-    PAddr aligned_end =
+    VAddr aligned_end =
         addr + Common::AlignUp(boost::icl::last_next(interval) - addr, stride_tiled_bytes);
 
     if (aligned_end - aligned_start > stride_tiled_bytes) {
@@ -527,10 +528,10 @@ void RasterizerCacheOpenGL::CopySurface(const Surface& src_surface, const Surfac
 }
 
 MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 64, 192));
-void CachedSurface::LoadGLBuffer(PAddr load_start, PAddr load_end) {
+void CachedSurface::LoadGLBuffer(VAddr load_start, VAddr load_end) {
     ASSERT(type != SurfaceType::Fill);
 
-    const u8* const texture_src_data = Memory::GetPhysicalPointer(addr);
+    u8* texture_src_data = Memory::GetPointer(addr);
     if (texture_src_data == nullptr)
         return;
 
@@ -539,35 +540,25 @@ void CachedSurface::LoadGLBuffer(PAddr load_start, PAddr load_end) {
         gl_buffer.reset(new u8[gl_buffer_size]);
     }
 
-    // TODO: Should probably be done in ::Memory:: and check for other regions too
-    if (load_start < Memory::VRAM_VADDR_END && load_end > Memory::VRAM_VADDR_END)
-        load_end = Memory::VRAM_VADDR_END;
-
-    if (load_start < Memory::VRAM_VADDR && load_end > Memory::VRAM_VADDR)
-        load_start = Memory::VRAM_VADDR;
-
     MICROPROFILE_SCOPE(OpenGL_SurfaceLoad);
 
     ASSERT(load_start >= addr && load_end <= end);
-    const u32 start_offset = load_start - addr;
+    const u64 start_offset = load_start - addr;
 
     if (!is_tiled) {
         ASSERT(type == SurfaceType::Color);
-        std::memcpy(&gl_buffer[start_offset], texture_src_data + start_offset,
-                    load_end - load_start);
+        const u32 bytes_per_pixel{GetFormatBpp() >> 3};
+        VideoCore::MortonCopyPixels128(width, height, bytes_per_pixel, 4,
+                                       texture_src_data + start_offset, &gl_buffer[start_offset],
+                                       true);
     } else {
-        if (type == SurfaceType::Texture) {
-            UNIMPLEMENTED();
-        } else {
-            morton_to_gl_fns[static_cast<size_t>(pixel_format)](stride, height, &gl_buffer[0], addr,
-                                                                load_start, load_end);
-        }
+        ASSERT_MSG(false, "Unimplemented");
     }
 }
 
 MICROPROFILE_DEFINE(OpenGL_SurfaceFlush, "OpenGL", "Surface Flush", MP_RGB(128, 192, 64));
-void CachedSurface::FlushGLBuffer(PAddr flush_start, PAddr flush_end) {
-    u8* const dst_buffer = Memory::GetPhysicalPointer(addr);
+void CachedSurface::FlushGLBuffer(VAddr flush_start, VAddr flush_end) {
+    u8* const dst_buffer = Memory::GetPointer(addr);
     if (dst_buffer == nullptr)
         return;
 
@@ -1102,7 +1093,7 @@ SurfaceRect_Tuple RasterizerCacheOpenGL::GetSurfaceSubRect(const SurfaceParams&
 }
 
 Surface RasterizerCacheOpenGL::GetTextureSurface(const void* config) {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
     return {};
 }
 
@@ -1113,7 +1104,7 @@ SurfaceSurfaceRect_Tuple RasterizerCacheOpenGL::GetFramebufferSurfaces(
 }
 
 Surface RasterizerCacheOpenGL::GetFillSurface(const void* config) {
-    UNIMPLEMENTED();
+    ASSERT_MSG(false, "Unimplemented");
     return {};
 }
 
@@ -1167,7 +1158,7 @@ void RasterizerCacheOpenGL::DuplicateSurface(const Surface& src_surface,
     }
 }
 
-void RasterizerCacheOpenGL::ValidateSurface(const Surface& surface, PAddr addr, u64 size) {
+void RasterizerCacheOpenGL::ValidateSurface(const Surface& surface, VAddr addr, u64 size) {
     if (size == 0)
         return;
 
@@ -1227,7 +1218,7 @@ void RasterizerCacheOpenGL::ValidateSurface(const Surface& surface, PAddr addr,
     }
 }
 
-void RasterizerCacheOpenGL::FlushRegion(PAddr addr, u64 size, Surface flush_surface) {
+void RasterizerCacheOpenGL::FlushRegion(VAddr addr, u64 size, Surface flush_surface) {
     if (size == 0)
         return;
 
@@ -1260,10 +1251,10 @@ void RasterizerCacheOpenGL::FlushRegion(PAddr addr, u64 size, Surface flush_surf
 }
 
 void RasterizerCacheOpenGL::FlushAll() {
-    FlushRegion(0, 0xFFFFFFFF);
+    FlushRegion(0, Kernel::VMManager::MAX_ADDRESS);
 }
 
-void RasterizerCacheOpenGL::InvalidateRegion(PAddr addr, u64 size, const Surface& region_owner) {
+void RasterizerCacheOpenGL::InvalidateRegion(VAddr addr, u64 size, const Surface& region_owner) {
     if (size == 0)
         return;
 
@@ -1356,6 +1347,6 @@ void RasterizerCacheOpenGL::UnregisterSurface(const Surface& surface) {
     surface_cache.subtract({surface->GetInterval(), SurfaceSet{surface}});
 }
 
-void RasterizerCacheOpenGL::UpdatePagesCachedCount(PAddr addr, u64 size, int delta) {
-    UNIMPLEMENTED();
+void RasterizerCacheOpenGL::UpdatePagesCachedCount(VAddr addr, u64 size, int delta) {
+    // ASSERT_MSG(false, "Unimplemented");
 }

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -22,15 +22,16 @@
 #include "common/common_funcs.h"
 #include "common/common_types.h"
 #include "common/math_util.h"
+#include "video_core/gpu.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 
 struct CachedSurface;
 using Surface = std::shared_ptr<CachedSurface>;
 using SurfaceSet = std::set<Surface>;
 
-using SurfaceRegions = boost::icl::interval_set<PAddr>;
-using SurfaceMap = boost::icl::interval_map<PAddr, Surface>;
-using SurfaceCache = boost::icl::interval_map<PAddr, SurfaceSet>;
+using SurfaceRegions = boost::icl::interval_set<VAddr>;
+using SurfaceMap = boost::icl::interval_map<VAddr, Surface>;
+using SurfaceCache = boost::icl::interval_map<VAddr, SurfaceSet>;
 
 using SurfaceInterval = SurfaceCache::interval_type;
 static_assert(std::is_same<SurfaceRegions::interval_type, SurfaceCache::interval_type>() &&
@@ -115,6 +116,15 @@ struct SurfaceParams {
         return GetFormatBpp(pixel_format);
     }
 
+    static PixelFormat PixelFormatFromGPUPixelFormat(Tegra::FramebufferConfig::PixelFormat format) {
+        switch (format) {
+        case Tegra::FramebufferConfig::PixelFormat::ABGR8:
+            return PixelFormat::RGBA8;
+        default:
+            UNREACHABLE();
+        }
+    }
+
     static bool CheckFormatsBlittable(PixelFormat pixel_format_a, PixelFormat pixel_format_b) {
         SurfaceType a_type = GetFormatType(pixel_format_a);
         SurfaceType b_type = GetFormatType(pixel_format_b);
@@ -211,8 +221,8 @@ struct SurfaceParams {
     MathUtil::Rectangle<u32> GetSubRect(const SurfaceParams& sub_surface) const;
     MathUtil::Rectangle<u32> GetScaledSubRect(const SurfaceParams& sub_surface) const;
 
-    PAddr addr = 0;
-    PAddr end = 0;
+    VAddr addr = 0;
+    VAddr end = 0;
     u64 size = 0;
 
     u32 width = 0;
@@ -257,9 +267,9 @@ struct CachedSurface : SurfaceParams {
     std::unique_ptr<u8[]> gl_buffer;
     size_t gl_buffer_size = 0;
 
-    // Read/Write data in 3DS memory to/from gl_buffer
-    void LoadGLBuffer(PAddr load_start, PAddr load_end);
-    void FlushGLBuffer(PAddr flush_start, PAddr flush_end);
+    // Read/Write data in Switch memory to/from gl_buffer
+    void LoadGLBuffer(VAddr load_start, VAddr load_end);
+    void FlushGLBuffer(VAddr flush_start, VAddr flush_end);
 
     // Upload/Download data in gl_buffer in/to this surface's texture
     void UploadGLTexture(const MathUtil::Rectangle<u32>& rect, GLuint read_fb_handle,
@@ -307,10 +317,10 @@ public:
     SurfaceRect_Tuple GetTexCopySurface(const SurfaceParams& params);
 
     /// Write any cached resources overlapping the region back to memory (if dirty)
-    void FlushRegion(PAddr addr, u64 size, Surface flush_surface = nullptr);
+    void FlushRegion(VAddr addr, u64 size, Surface flush_surface = nullptr);
 
     /// Mark region as being invalidated by region_owner (nullptr if 3DS memory)
-    void InvalidateRegion(PAddr addr, u64 size, const Surface& region_owner);
+    void InvalidateRegion(VAddr addr, u64 size, const Surface& region_owner);
 
     /// Flush all cached resources tracked by this cache manager
     void FlushAll();
@@ -319,7 +329,7 @@ private:
     void DuplicateSurface(const Surface& src_surface, const Surface& dest_surface);
 
     /// Update surface's texture for given region when necessary
-    void ValidateSurface(const Surface& surface, PAddr addr, u64 size);
+    void ValidateSurface(const Surface& surface, VAddr addr, u64 size);
 
     /// Create a new surface
     Surface CreateSurface(const SurfaceParams& params);
@@ -331,7 +341,7 @@ private:
     void UnregisterSurface(const Surface& surface);
 
     /// Increase/decrease the number of surface in pages touching the specified region
-    void UpdatePagesCachedCount(PAddr addr, u64 size, int delta);
+    void UpdatePagesCachedCount(VAddr addr, u64 size, int delta);
 
     SurfaceCache surface_cache;
     PageMap cached_pages;

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -20,6 +20,7 @@
 #include "core/settings.h"
 #include "core/tracer/recorder.h"
 #include "video_core/renderer_opengl/renderer_opengl.h"
+#include "video_core/utils.h"
 #include "video_core/video_core.h"
 
 static const char vertex_shader[] = R"(
@@ -98,22 +99,22 @@ RendererOpenGL::RendererOpenGL() = default;
 RendererOpenGL::~RendererOpenGL() = default;
 
 /// Swap buffers (render frame)
-void RendererOpenGL::SwapBuffers(boost::optional<const FramebufferInfo&> framebuffer_info) {
+void RendererOpenGL::SwapBuffers(boost::optional<const Tegra::FramebufferConfig&> framebuffer) {
     // Maintain the rasterizer's state as a priority
     OpenGLState prev_state = OpenGLState::GetCurState();
     state.Apply();
 
-    if (framebuffer_info != boost::none) {
-        // If framebuffer_info is provided, reload it from memory to a texture
-        if (screen_info.texture.width != (GLsizei)framebuffer_info->width ||
-            screen_info.texture.height != (GLsizei)framebuffer_info->height ||
-            screen_info.texture.pixel_format != framebuffer_info->pixel_format) {
+    if (framebuffer != boost::none) {
+        // If framebuffer is provided, reload it from memory to a texture
+        if (screen_info.texture.width != (GLsizei)framebuffer->width ||
+            screen_info.texture.height != (GLsizei)framebuffer->height ||
+            screen_info.texture.pixel_format != framebuffer->pixel_format) {
             // Reallocate texture if the framebuffer size has changed.
             // This is expected to not happen very often and hence should not be a
             // performance problem.
-            ConfigureFramebufferTexture(screen_info.texture, *framebuffer_info);
+            ConfigureFramebufferTexture(screen_info.texture, *framebuffer);
         }
-        LoadFBToScreenInfo(*framebuffer_info, screen_info);
+        LoadFBToScreenInfo(*framebuffer, screen_info);
     }
 
     DrawScreens();
@@ -131,164 +132,59 @@ void RendererOpenGL::SwapBuffers(boost::optional<const FramebufferInfo&> framebu
     RefreshRasterizerSetting();
 }
 
-static inline u32 MortonInterleave128(u32 x, u32 y) {
-    // 128x128 Z-Order coordinate from 2D coordinates
-    static constexpr u32 xlut[] = {
-        0x0000, 0x0001, 0x0002, 0x0003, 0x0008, 0x0009, 0x000a, 0x000b, 0x0040, 0x0041, 0x0042,
-        0x0043, 0x0048, 0x0049, 0x004a, 0x004b, 0x0800, 0x0801, 0x0802, 0x0803, 0x0808, 0x0809,
-        0x080a, 0x080b, 0x0840, 0x0841, 0x0842, 0x0843, 0x0848, 0x0849, 0x084a, 0x084b, 0x1000,
-        0x1001, 0x1002, 0x1003, 0x1008, 0x1009, 0x100a, 0x100b, 0x1040, 0x1041, 0x1042, 0x1043,
-        0x1048, 0x1049, 0x104a, 0x104b, 0x1800, 0x1801, 0x1802, 0x1803, 0x1808, 0x1809, 0x180a,
-        0x180b, 0x1840, 0x1841, 0x1842, 0x1843, 0x1848, 0x1849, 0x184a, 0x184b, 0x2000, 0x2001,
-        0x2002, 0x2003, 0x2008, 0x2009, 0x200a, 0x200b, 0x2040, 0x2041, 0x2042, 0x2043, 0x2048,
-        0x2049, 0x204a, 0x204b, 0x2800, 0x2801, 0x2802, 0x2803, 0x2808, 0x2809, 0x280a, 0x280b,
-        0x2840, 0x2841, 0x2842, 0x2843, 0x2848, 0x2849, 0x284a, 0x284b, 0x3000, 0x3001, 0x3002,
-        0x3003, 0x3008, 0x3009, 0x300a, 0x300b, 0x3040, 0x3041, 0x3042, 0x3043, 0x3048, 0x3049,
-        0x304a, 0x304b, 0x3800, 0x3801, 0x3802, 0x3803, 0x3808, 0x3809, 0x380a, 0x380b, 0x3840,
-        0x3841, 0x3842, 0x3843, 0x3848, 0x3849, 0x384a, 0x384b, 0x0000, 0x0001, 0x0002, 0x0003,
-        0x0008, 0x0009, 0x000a, 0x000b, 0x0040, 0x0041, 0x0042, 0x0043, 0x0048, 0x0049, 0x004a,
-        0x004b, 0x0800, 0x0801, 0x0802, 0x0803, 0x0808, 0x0809, 0x080a, 0x080b, 0x0840, 0x0841,
-        0x0842, 0x0843, 0x0848, 0x0849, 0x084a, 0x084b, 0x1000, 0x1001, 0x1002, 0x1003, 0x1008,
-        0x1009, 0x100a, 0x100b, 0x1040, 0x1041, 0x1042, 0x1043, 0x1048, 0x1049, 0x104a, 0x104b,
-        0x1800, 0x1801, 0x1802, 0x1803, 0x1808, 0x1809, 0x180a, 0x180b, 0x1840, 0x1841, 0x1842,
-        0x1843, 0x1848, 0x1849, 0x184a, 0x184b, 0x2000, 0x2001, 0x2002, 0x2003, 0x2008, 0x2009,
-        0x200a, 0x200b, 0x2040, 0x2041, 0x2042, 0x2043, 0x2048, 0x2049, 0x204a, 0x204b, 0x2800,
-        0x2801, 0x2802, 0x2803, 0x2808, 0x2809, 0x280a, 0x280b, 0x2840, 0x2841, 0x2842, 0x2843,
-        0x2848, 0x2849, 0x284a, 0x284b, 0x3000, 0x3001, 0x3002, 0x3003, 0x3008, 0x3009, 0x300a,
-        0x300b, 0x3040, 0x3041, 0x3042, 0x3043, 0x3048, 0x3049, 0x304a, 0x304b, 0x3800, 0x3801,
-        0x3802, 0x3803, 0x3808, 0x3809, 0x380a, 0x380b, 0x3840, 0x3841, 0x3842, 0x3843, 0x3848,
-        0x3849, 0x384a, 0x384b, 0x0000, 0x0001, 0x0002, 0x0003, 0x0008, 0x0009, 0x000a, 0x000b,
-        0x0040, 0x0041, 0x0042, 0x0043, 0x0048, 0x0049, 0x004a, 0x004b, 0x0800, 0x0801, 0x0802,
-        0x0803, 0x0808, 0x0809, 0x080a, 0x080b, 0x0840, 0x0841, 0x0842, 0x0843, 0x0848, 0x0849,
-        0x084a, 0x084b, 0x1000, 0x1001, 0x1002, 0x1003, 0x1008, 0x1009, 0x100a, 0x100b, 0x1040,
-        0x1041, 0x1042, 0x1043, 0x1048, 0x1049, 0x104a, 0x104b, 0x1800, 0x1801, 0x1802, 0x1803,
-        0x1808, 0x1809, 0x180a, 0x180b, 0x1840, 0x1841, 0x1842, 0x1843, 0x1848, 0x1849, 0x184a,
-        0x184b, 0x2000, 0x2001, 0x2002, 0x2003, 0x2008, 0x2009, 0x200a, 0x200b, 0x2040, 0x2041,
-        0x2042, 0x2043, 0x2048, 0x2049, 0x204a, 0x204b, 0x2800, 0x2801, 0x2802, 0x2803, 0x2808,
-        0x2809, 0x280a, 0x280b, 0x2840, 0x2841, 0x2842, 0x2843, 0x2848, 0x2849, 0x284a, 0x284b,
-        0x3000, 0x3001, 0x3002, 0x3003, 0x3008, 0x3009, 0x300a, 0x300b, 0x3040, 0x3041, 0x3042,
-        0x3043, 0x3048, 0x3049, 0x304a, 0x304b, 0x3800, 0x3801, 0x3802, 0x3803, 0x3808, 0x3809,
-        0x380a, 0x380b, 0x3840, 0x3841, 0x3842, 0x3843, 0x3848, 0x3849, 0x384a, 0x384b,
-    };
-    static constexpr u32 ylut[] = {
-        0x0000, 0x0004, 0x0010, 0x0014, 0x0020, 0x0024, 0x0030, 0x0034, 0x0080, 0x0084, 0x0090,
-        0x0094, 0x00a0, 0x00a4, 0x00b0, 0x00b4, 0x0100, 0x0104, 0x0110, 0x0114, 0x0120, 0x0124,
-        0x0130, 0x0134, 0x0180, 0x0184, 0x0190, 0x0194, 0x01a0, 0x01a4, 0x01b0, 0x01b4, 0x0200,
-        0x0204, 0x0210, 0x0214, 0x0220, 0x0224, 0x0230, 0x0234, 0x0280, 0x0284, 0x0290, 0x0294,
-        0x02a0, 0x02a4, 0x02b0, 0x02b4, 0x0300, 0x0304, 0x0310, 0x0314, 0x0320, 0x0324, 0x0330,
-        0x0334, 0x0380, 0x0384, 0x0390, 0x0394, 0x03a0, 0x03a4, 0x03b0, 0x03b4, 0x0400, 0x0404,
-        0x0410, 0x0414, 0x0420, 0x0424, 0x0430, 0x0434, 0x0480, 0x0484, 0x0490, 0x0494, 0x04a0,
-        0x04a4, 0x04b0, 0x04b4, 0x0500, 0x0504, 0x0510, 0x0514, 0x0520, 0x0524, 0x0530, 0x0534,
-        0x0580, 0x0584, 0x0590, 0x0594, 0x05a0, 0x05a4, 0x05b0, 0x05b4, 0x0600, 0x0604, 0x0610,
-        0x0614, 0x0620, 0x0624, 0x0630, 0x0634, 0x0680, 0x0684, 0x0690, 0x0694, 0x06a0, 0x06a4,
-        0x06b0, 0x06b4, 0x0700, 0x0704, 0x0710, 0x0714, 0x0720, 0x0724, 0x0730, 0x0734, 0x0780,
-        0x0784, 0x0790, 0x0794, 0x07a0, 0x07a4, 0x07b0, 0x07b4, 0x0000, 0x0004, 0x0010, 0x0014,
-        0x0020, 0x0024, 0x0030, 0x0034, 0x0080, 0x0084, 0x0090, 0x0094, 0x00a0, 0x00a4, 0x00b0,
-        0x00b4, 0x0100, 0x0104, 0x0110, 0x0114, 0x0120, 0x0124, 0x0130, 0x0134, 0x0180, 0x0184,
-        0x0190, 0x0194, 0x01a0, 0x01a4, 0x01b0, 0x01b4, 0x0200, 0x0204, 0x0210, 0x0214, 0x0220,
-        0x0224, 0x0230, 0x0234, 0x0280, 0x0284, 0x0290, 0x0294, 0x02a0, 0x02a4, 0x02b0, 0x02b4,
-        0x0300, 0x0304, 0x0310, 0x0314, 0x0320, 0x0324, 0x0330, 0x0334, 0x0380, 0x0384, 0x0390,
-        0x0394, 0x03a0, 0x03a4, 0x03b0, 0x03b4, 0x0400, 0x0404, 0x0410, 0x0414, 0x0420, 0x0424,
-        0x0430, 0x0434, 0x0480, 0x0484, 0x0490, 0x0494, 0x04a0, 0x04a4, 0x04b0, 0x04b4, 0x0500,
-        0x0504, 0x0510, 0x0514, 0x0520, 0x0524, 0x0530, 0x0534, 0x0580, 0x0584, 0x0590, 0x0594,
-        0x05a0, 0x05a4, 0x05b0, 0x05b4, 0x0600, 0x0604, 0x0610, 0x0614, 0x0620, 0x0624, 0x0630,
-        0x0634, 0x0680, 0x0684, 0x0690, 0x0694, 0x06a0, 0x06a4, 0x06b0, 0x06b4, 0x0700, 0x0704,
-        0x0710, 0x0714, 0x0720, 0x0724, 0x0730, 0x0734, 0x0780, 0x0784, 0x0790, 0x0794, 0x07a0,
-        0x07a4, 0x07b0, 0x07b4, 0x0000, 0x0004, 0x0010, 0x0014, 0x0020, 0x0024, 0x0030, 0x0034,
-        0x0080, 0x0084, 0x0090, 0x0094, 0x00a0, 0x00a4, 0x00b0, 0x00b4, 0x0100, 0x0104, 0x0110,
-        0x0114, 0x0120, 0x0124, 0x0130, 0x0134, 0x0180, 0x0184, 0x0190, 0x0194, 0x01a0, 0x01a4,
-        0x01b0, 0x01b4, 0x0200, 0x0204, 0x0210, 0x0214, 0x0220, 0x0224, 0x0230, 0x0234, 0x0280,
-        0x0284, 0x0290, 0x0294, 0x02a0, 0x02a4, 0x02b0, 0x02b4, 0x0300, 0x0304, 0x0310, 0x0314,
-        0x0320, 0x0324, 0x0330, 0x0334, 0x0380, 0x0384, 0x0390, 0x0394, 0x03a0, 0x03a4, 0x03b0,
-        0x03b4, 0x0400, 0x0404, 0x0410, 0x0414, 0x0420, 0x0424, 0x0430, 0x0434, 0x0480, 0x0484,
-        0x0490, 0x0494, 0x04a0, 0x04a4, 0x04b0, 0x04b4, 0x0500, 0x0504, 0x0510, 0x0514, 0x0520,
-        0x0524, 0x0530, 0x0534, 0x0580, 0x0584, 0x0590, 0x0594, 0x05a0, 0x05a4, 0x05b0, 0x05b4,
-        0x0600, 0x0604, 0x0610, 0x0614, 0x0620, 0x0624, 0x0630, 0x0634, 0x0680, 0x0684, 0x0690,
-        0x0694, 0x06a0, 0x06a4, 0x06b0, 0x06b4, 0x0700, 0x0704, 0x0710, 0x0714, 0x0720, 0x0724,
-        0x0730, 0x0734, 0x0780, 0x0784, 0x0790, 0x0794, 0x07a0, 0x07a4, 0x07b0, 0x07b4,
-    };
-    return xlut[x % 128] + ylut[y % 128];
-}
-
-static inline u32 GetMortonOffset128(u32 x, u32 y, u32 bytes_per_pixel) {
-    // Calculates the offset of the position of the pixel in Morton order
-    // Framebuffer images are split into 128x128 tiles.
-
-    const unsigned int block_height = 128;
-    const unsigned int coarse_x = x & ~127;
-
-    u32 i = MortonInterleave128(x, y);
-
-    const unsigned int offset = coarse_x * block_height;
-
-    return (i + offset) * bytes_per_pixel;
-}
-
-static void MortonCopyPixels128(u32 width, u32 height, u32 bytes_per_pixel, u32 gl_bytes_per_pixel,
-                                u8* morton_data, u8* gl_data, bool morton_to_gl) {
-    u8* data_ptrs[2];
-    for (unsigned y = 0; y < height; ++y) {
-        for (unsigned x = 0; x < width; ++x) {
-            const u32 coarse_y = y & ~127;
-            u32 morton_offset =
-                GetMortonOffset128(x, y, bytes_per_pixel) + coarse_y * width * bytes_per_pixel;
-            u32 gl_pixel_index = (x + (height - 1 - y) * width) * gl_bytes_per_pixel;
-
-            data_ptrs[morton_to_gl] = morton_data + morton_offset;
-            data_ptrs[!morton_to_gl] = &gl_data[gl_pixel_index];
-
-            memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel);
-        }
-    }
-}
-
 /**
  * Loads framebuffer from emulated memory into the active OpenGL texture.
  */
-void RendererOpenGL::LoadFBToScreenInfo(const FramebufferInfo& framebuffer_info,
+void RendererOpenGL::LoadFBToScreenInfo(const Tegra::FramebufferConfig& framebuffer,
                                         ScreenInfo& screen_info) {
-    const u32 bpp{FramebufferInfo::BytesPerPixel(framebuffer_info.pixel_format)};
-    const u32 size_in_bytes{framebuffer_info.stride * framebuffer_info.height * bpp};
+    const u32 bytes_per_pixel{Tegra::FramebufferConfig::BytesPerPixel(framebuffer.pixel_format)};
+    const u64 size_in_bytes{framebuffer.stride * framebuffer.height * bytes_per_pixel};
+    const VAddr framebuffer_addr{framebuffer.address + framebuffer.offset};
 
-    MortonCopyPixels128(framebuffer_info.width, framebuffer_info.height, bpp, 4,
-                        Memory::GetPointer(framebuffer_info.address), gl_framebuffer_data.data(),
-                        true);
+    // TODO(bunnei): The framebuffer region should only be invalidated if it is written to, not
+    // every frame. When we find the right place for this, the below line can be removed.
+    Memory::RasterizerFlushVirtualRegion(framebuffer_addr, size_in_bytes,
+                                         Memory::FlushMode::Invalidate);
 
-    LOG_TRACE(Render_OpenGL, "0x%08x bytes from 0x%llx(%dx%d), fmt %x", size_in_bytes,
-              framebuffer_info.address, framebuffer_info.width, framebuffer_info.height,
-              (int)framebuffer_info.pixel_format);
+    // Framebuffer orientation handling
+    framebuffer_transform_flags = framebuffer.transform_flags;
 
     // Ensure no bad interactions with GL_UNPACK_ALIGNMENT, which by default
     // only allows rows to have a memory alignement of 4.
-    ASSERT(framebuffer_info.stride % 4 == 0);
+    ASSERT(framebuffer.stride % 4 == 0);
 
-    framebuffer_flip_vertical = framebuffer_info.flip_vertical;
+    if (!Rasterizer()->AccelerateDisplay(framebuffer, framebuffer_addr, framebuffer.stride,
+                                         screen_info)) {
+        // Reset the screen info's display texture to its own permanent texture
+        screen_info.display_texture = screen_info.texture.resource.handle;
+        screen_info.display_texcoords = MathUtil::Rectangle<float>(0.f, 0.f, 1.f, 1.f);
 
-    // Reset the screen info's display texture to its own permanent texture
-    screen_info.display_texture = screen_info.texture.resource.handle;
-    screen_info.display_texcoords = MathUtil::Rectangle<float>(0.f, 0.f, 1.f, 1.f);
+        Rasterizer()->FlushRegion(framebuffer_addr, size_in_bytes);
 
-    // Memory::RasterizerFlushRegion(framebuffer_info.address, size_in_bytes);
+        VideoCore::MortonCopyPixels128(framebuffer.width, framebuffer.height, bytes_per_pixel, 4,
+                                       Memory::GetPointer(framebuffer_addr),
+                                       gl_framebuffer_data.data(), true);
 
-    state.texture_units[0].texture_2d = screen_info.texture.resource.handle;
-    state.Apply();
+        state.texture_units[0].texture_2d = screen_info.texture.resource.handle;
+        state.Apply();
 
-    glActiveTexture(GL_TEXTURE0);
-    glPixelStorei(GL_UNPACK_ROW_LENGTH, (GLint)framebuffer_info.stride);
+        glActiveTexture(GL_TEXTURE0);
+        glPixelStorei(GL_UNPACK_ROW_LENGTH, static_cast<GLint>(framebuffer.stride));
 
-    // Update existing texture
-    // TODO: Test what happens on hardware when you change the framebuffer dimensions so that
-    //       they differ from the LCD resolution.
-    // TODO: Applications could theoretically crash Citra here by specifying too large
-    //       framebuffer sizes. We should make sure that this cannot happen.
-    glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, framebuffer_info.width, framebuffer_info.height,
-                    screen_info.texture.gl_format, screen_info.texture.gl_type,
-                    gl_framebuffer_data.data());
+        // Update existing texture
+        // TODO: Test what happens on hardware when you change the framebuffer dimensions so that
+        //       they differ from the LCD resolution.
+        // TODO: Applications could theoretically crash yuzu here by specifying too large
+        //       framebuffer sizes. We should make sure that this cannot happen.
+        glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, framebuffer.width, framebuffer.height,
+                        screen_info.texture.gl_format, screen_info.texture.gl_type,
+                        gl_framebuffer_data.data());
 
-    glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
+        glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 
-    state.texture_units[0].texture_2d = 0;
-    state.Apply();
+        state.texture_units[0].texture_2d = 0;
+        state.Apply();
+    }
 }
 
 /**
@@ -372,14 +268,14 @@ void RendererOpenGL::InitOpenGLObjects() {
 }
 
 void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
-                                                 const FramebufferInfo& framebuffer_info) {
+                                                 const Tegra::FramebufferConfig& framebuffer) {
 
-    texture.width = framebuffer_info.width;
-    texture.height = framebuffer_info.height;
+    texture.width = framebuffer.width;
+    texture.height = framebuffer.height;
 
     GLint internal_format;
-    switch (framebuffer_info.pixel_format) {
-    case FramebufferInfo::PixelFormat::ABGR8:
+    switch (framebuffer.pixel_format) {
+    case Tegra::FramebufferConfig::PixelFormat::ABGR8:
         // Use RGBA8 and swap in the fragment shader
         internal_format = GL_RGBA;
         texture.gl_format = GL_RGBA;
@@ -404,8 +300,19 @@ void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
 void RendererOpenGL::DrawSingleScreen(const ScreenInfo& screen_info, float x, float y, float w,
                                       float h) {
     const auto& texcoords = screen_info.display_texcoords;
-    const auto& left = framebuffer_flip_vertical ? texcoords.right : texcoords.left;
-    const auto& right = framebuffer_flip_vertical ? texcoords.left : texcoords.right;
+    auto left = texcoords.left;
+    auto right = texcoords.right;
+    if (framebuffer_transform_flags != Tegra::FramebufferConfig::TransformFlags::Unset)
+        if (framebuffer_transform_flags == Tegra::FramebufferConfig::TransformFlags::FlipV) {
+            // Flip the framebuffer vertically
+            left = texcoords.right;
+            right = texcoords.left;
+        } else {
+            // Other transformations are unsupported
+            LOG_CRITICAL(HW_GPU, "unsupported framebuffer_transform_flags=%d",
+                         framebuffer_transform_flags);
+            UNIMPLEMENTED();
+        }
 
     std::array<ScreenRectVertex, 4> vertices = {{
         ScreenRectVertex(x, y, texcoords.top, right),

src/video_core/renderer_opengl/renderer_opengl.h

@@ -21,7 +21,7 @@ struct TextureInfo {
     GLsizei height;
     GLenum gl_format;
     GLenum gl_type;
-    RendererBase::FramebufferInfo::PixelFormat pixel_format;
+    Tegra::FramebufferConfig::PixelFormat pixel_format;
 };
 
 /// Structure used for storing information about the display target for each 3DS screen
@@ -37,7 +37,7 @@ public:
     ~RendererOpenGL() override;
 
     /// Swap buffers (render frame)
-    void SwapBuffers(boost::optional<const FramebufferInfo&> framebuffer_info) override;
+    void SwapBuffers(boost::optional<const Tegra::FramebufferConfig&> framebuffer) override;
 
     /**
      * Set the emulator window to use for renderer
@@ -53,13 +53,14 @@ public:
 
 private:
     void InitOpenGLObjects();
-    void ConfigureFramebufferTexture(TextureInfo& texture, const FramebufferInfo& framebuffer_info);
+    void ConfigureFramebufferTexture(TextureInfo& texture,
+                                     const Tegra::FramebufferConfig& framebuffer);
     void DrawScreens();
     void DrawSingleScreen(const ScreenInfo& screen_info, float x, float y, float w, float h);
     void UpdateFramerate();
 
     // Loads framebuffer from emulated memory into the display information structure
-    void LoadFBToScreenInfo(const FramebufferInfo& framebuffer_info, ScreenInfo& screen_info);
+    void LoadFBToScreenInfo(const Tegra::FramebufferConfig& framebuffer, ScreenInfo& screen_info);
     // Fills active OpenGL texture with the given RGBA color.
     void LoadColorToActiveGLTexture(u8 color_r, u8 color_g, u8 color_b, u8 color_a,
                                     const TextureInfo& texture);
@@ -87,6 +88,6 @@ private:
     GLuint attrib_position;
     GLuint attrib_tex_coord;
 
-    /// Flips the framebuffer vertically when true
-    bool framebuffer_flip_vertical;
+    /// Used for transforming the framebuffer orientation
+    Tegra::FramebufferConfig::TransformFlags framebuffer_transform_flags;
 };

src/video_core/utils.h

@@ -49,4 +49,116 @@ static inline u32 GetMortonOffset(u32 x, u32 y, u32 bytes_per_pixel) {
     return (i + offset) * bytes_per_pixel;
 }
 
+static inline u32 MortonInterleave128(u32 x, u32 y) {
+    // 128x128 Z-Order coordinate from 2D coordinates
+    static constexpr u32 xlut[] = {
+        0x0000, 0x0001, 0x0002, 0x0003, 0x0008, 0x0009, 0x000a, 0x000b, 0x0040, 0x0041, 0x0042,
+        0x0043, 0x0048, 0x0049, 0x004a, 0x004b, 0x0800, 0x0801, 0x0802, 0x0803, 0x0808, 0x0809,
+        0x080a, 0x080b, 0x0840, 0x0841, 0x0842, 0x0843, 0x0848, 0x0849, 0x084a, 0x084b, 0x1000,
+        0x1001, 0x1002, 0x1003, 0x1008, 0x1009, 0x100a, 0x100b, 0x1040, 0x1041, 0x1042, 0x1043,
+        0x1048, 0x1049, 0x104a, 0x104b, 0x1800, 0x1801, 0x1802, 0x1803, 0x1808, 0x1809, 0x180a,
+        0x180b, 0x1840, 0x1841, 0x1842, 0x1843, 0x1848, 0x1849, 0x184a, 0x184b, 0x2000, 0x2001,
+        0x2002, 0x2003, 0x2008, 0x2009, 0x200a, 0x200b, 0x2040, 0x2041, 0x2042, 0x2043, 0x2048,
+        0x2049, 0x204a, 0x204b, 0x2800, 0x2801, 0x2802, 0x2803, 0x2808, 0x2809, 0x280a, 0x280b,
+        0x2840, 0x2841, 0x2842, 0x2843, 0x2848, 0x2849, 0x284a, 0x284b, 0x3000, 0x3001, 0x3002,
+        0x3003, 0x3008, 0x3009, 0x300a, 0x300b, 0x3040, 0x3041, 0x3042, 0x3043, 0x3048, 0x3049,
+        0x304a, 0x304b, 0x3800, 0x3801, 0x3802, 0x3803, 0x3808, 0x3809, 0x380a, 0x380b, 0x3840,
+        0x3841, 0x3842, 0x3843, 0x3848, 0x3849, 0x384a, 0x384b, 0x0000, 0x0001, 0x0002, 0x0003,
+        0x0008, 0x0009, 0x000a, 0x000b, 0x0040, 0x0041, 0x0042, 0x0043, 0x0048, 0x0049, 0x004a,
+        0x004b, 0x0800, 0x0801, 0x0802, 0x0803, 0x0808, 0x0809, 0x080a, 0x080b, 0x0840, 0x0841,
+        0x0842, 0x0843, 0x0848, 0x0849, 0x084a, 0x084b, 0x1000, 0x1001, 0x1002, 0x1003, 0x1008,
+        0x1009, 0x100a, 0x100b, 0x1040, 0x1041, 0x1042, 0x1043, 0x1048, 0x1049, 0x104a, 0x104b,
+        0x1800, 0x1801, 0x1802, 0x1803, 0x1808, 0x1809, 0x180a, 0x180b, 0x1840, 0x1841, 0x1842,
+        0x1843, 0x1848, 0x1849, 0x184a, 0x184b, 0x2000, 0x2001, 0x2002, 0x2003, 0x2008, 0x2009,
+        0x200a, 0x200b, 0x2040, 0x2041, 0x2042, 0x2043, 0x2048, 0x2049, 0x204a, 0x204b, 0x2800,
+        0x2801, 0x2802, 0x2803, 0x2808, 0x2809, 0x280a, 0x280b, 0x2840, 0x2841, 0x2842, 0x2843,
+        0x2848, 0x2849, 0x284a, 0x284b, 0x3000, 0x3001, 0x3002, 0x3003, 0x3008, 0x3009, 0x300a,
+        0x300b, 0x3040, 0x3041, 0x3042, 0x3043, 0x3048, 0x3049, 0x304a, 0x304b, 0x3800, 0x3801,
+        0x3802, 0x3803, 0x3808, 0x3809, 0x380a, 0x380b, 0x3840, 0x3841, 0x3842, 0x3843, 0x3848,
+        0x3849, 0x384a, 0x384b, 0x0000, 0x0001, 0x0002, 0x0003, 0x0008, 0x0009, 0x000a, 0x000b,
+        0x0040, 0x0041, 0x0042, 0x0043, 0x0048, 0x0049, 0x004a, 0x004b, 0x0800, 0x0801, 0x0802,
+        0x0803, 0x0808, 0x0809, 0x080a, 0x080b, 0x0840, 0x0841, 0x0842, 0x0843, 0x0848, 0x0849,
+        0x084a, 0x084b, 0x1000, 0x1001, 0x1002, 0x1003, 0x1008, 0x1009, 0x100a, 0x100b, 0x1040,
+        0x1041, 0x1042, 0x1043, 0x1048, 0x1049, 0x104a, 0x104b, 0x1800, 0x1801, 0x1802, 0x1803,
+        0x1808, 0x1809, 0x180a, 0x180b, 0x1840, 0x1841, 0x1842, 0x1843, 0x1848, 0x1849, 0x184a,
+        0x184b, 0x2000, 0x2001, 0x2002, 0x2003, 0x2008, 0x2009, 0x200a, 0x200b, 0x2040, 0x2041,
+        0x2042, 0x2043, 0x2048, 0x2049, 0x204a, 0x204b, 0x2800, 0x2801, 0x2802, 0x2803, 0x2808,
+        0x2809, 0x280a, 0x280b, 0x2840, 0x2841, 0x2842, 0x2843, 0x2848, 0x2849, 0x284a, 0x284b,
+        0x3000, 0x3001, 0x3002, 0x3003, 0x3008, 0x3009, 0x300a, 0x300b, 0x3040, 0x3041, 0x3042,
+        0x3043, 0x3048, 0x3049, 0x304a, 0x304b, 0x3800, 0x3801, 0x3802, 0x3803, 0x3808, 0x3809,
+        0x380a, 0x380b, 0x3840, 0x3841, 0x3842, 0x3843, 0x3848, 0x3849, 0x384a, 0x384b,
+    };
+    static constexpr u32 ylut[] = {
+        0x0000, 0x0004, 0x0010, 0x0014, 0x0020, 0x0024, 0x0030, 0x0034, 0x0080, 0x0084, 0x0090,
+        0x0094, 0x00a0, 0x00a4, 0x00b0, 0x00b4, 0x0100, 0x0104, 0x0110, 0x0114, 0x0120, 0x0124,
+        0x0130, 0x0134, 0x0180, 0x0184, 0x0190, 0x0194, 0x01a0, 0x01a4, 0x01b0, 0x01b4, 0x0200,
+        0x0204, 0x0210, 0x0214, 0x0220, 0x0224, 0x0230, 0x0234, 0x0280, 0x0284, 0x0290, 0x0294,
+        0x02a0, 0x02a4, 0x02b0, 0x02b4, 0x0300, 0x0304, 0x0310, 0x0314, 0x0320, 0x0324, 0x0330,
+        0x0334, 0x0380, 0x0384, 0x0390, 0x0394, 0x03a0, 0x03a4, 0x03b0, 0x03b4, 0x0400, 0x0404,
+        0x0410, 0x0414, 0x0420, 0x0424, 0x0430, 0x0434, 0x0480, 0x0484, 0x0490, 0x0494, 0x04a0,
+        0x04a4, 0x04b0, 0x04b4, 0x0500, 0x0504, 0x0510, 0x0514, 0x0520, 0x0524, 0x0530, 0x0534,
+        0x0580, 0x0584, 0x0590, 0x0594, 0x05a0, 0x05a4, 0x05b0, 0x05b4, 0x0600, 0x0604, 0x0610,
+        0x0614, 0x0620, 0x0624, 0x0630, 0x0634, 0x0680, 0x0684, 0x0690, 0x0694, 0x06a0, 0x06a4,
+        0x06b0, 0x06b4, 0x0700, 0x0704, 0x0710, 0x0714, 0x0720, 0x0724, 0x0730, 0x0734, 0x0780,
+        0x0784, 0x0790, 0x0794, 0x07a0, 0x07a4, 0x07b0, 0x07b4, 0x0000, 0x0004, 0x0010, 0x0014,
+        0x0020, 0x0024, 0x0030, 0x0034, 0x0080, 0x0084, 0x0090, 0x0094, 0x00a0, 0x00a4, 0x00b0,
+        0x00b4, 0x0100, 0x0104, 0x0110, 0x0114, 0x0120, 0x0124, 0x0130, 0x0134, 0x0180, 0x0184,
+        0x0190, 0x0194, 0x01a0, 0x01a4, 0x01b0, 0x01b4, 0x0200, 0x0204, 0x0210, 0x0214, 0x0220,
+        0x0224, 0x0230, 0x0234, 0x0280, 0x0284, 0x0290, 0x0294, 0x02a0, 0x02a4, 0x02b0, 0x02b4,
+        0x0300, 0x0304, 0x0310, 0x0314, 0x0320, 0x0324, 0x0330, 0x0334, 0x0380, 0x0384, 0x0390,
+        0x0394, 0x03a0, 0x03a4, 0x03b0, 0x03b4, 0x0400, 0x0404, 0x0410, 0x0414, 0x0420, 0x0424,
+        0x0430, 0x0434, 0x0480, 0x0484, 0x0490, 0x0494, 0x04a0, 0x04a4, 0x04b0, 0x04b4, 0x0500,
+        0x0504, 0x0510, 0x0514, 0x0520, 0x0524, 0x0530, 0x0534, 0x0580, 0x0584, 0x0590, 0x0594,
+        0x05a0, 0x05a4, 0x05b0, 0x05b4, 0x0600, 0x0604, 0x0610, 0x0614, 0x0620, 0x0624, 0x0630,
+        0x0634, 0x0680, 0x0684, 0x0690, 0x0694, 0x06a0, 0x06a4, 0x06b0, 0x06b4, 0x0700, 0x0704,
+        0x0710, 0x0714, 0x0720, 0x0724, 0x0730, 0x0734, 0x0780, 0x0784, 0x0790, 0x0794, 0x07a0,
+        0x07a4, 0x07b0, 0x07b4, 0x0000, 0x0004, 0x0010, 0x0014, 0x0020, 0x0024, 0x0030, 0x0034,
+        0x0080, 0x0084, 0x0090, 0x0094, 0x00a0, 0x00a4, 0x00b0, 0x00b4, 0x0100, 0x0104, 0x0110,
+        0x0114, 0x0120, 0x0124, 0x0130, 0x0134, 0x0180, 0x0184, 0x0190, 0x0194, 0x01a0, 0x01a4,
+        0x01b0, 0x01b4, 0x0200, 0x0204, 0x0210, 0x0214, 0x0220, 0x0224, 0x0230, 0x0234, 0x0280,
+        0x0284, 0x0290, 0x0294, 0x02a0, 0x02a4, 0x02b0, 0x02b4, 0x0300, 0x0304, 0x0310, 0x0314,
+        0x0320, 0x0324, 0x0330, 0x0334, 0x0380, 0x0384, 0x0390, 0x0394, 0x03a0, 0x03a4, 0x03b0,
+        0x03b4, 0x0400, 0x0404, 0x0410, 0x0414, 0x0420, 0x0424, 0x0430, 0x0434, 0x0480, 0x0484,
+        0x0490, 0x0494, 0x04a0, 0x04a4, 0x04b0, 0x04b4, 0x0500, 0x0504, 0x0510, 0x0514, 0x0520,
+        0x0524, 0x0530, 0x0534, 0x0580, 0x0584, 0x0590, 0x0594, 0x05a0, 0x05a4, 0x05b0, 0x05b4,
+        0x0600, 0x0604, 0x0610, 0x0614, 0x0620, 0x0624, 0x0630, 0x0634, 0x0680, 0x0684, 0x0690,
+        0x0694, 0x06a0, 0x06a4, 0x06b0, 0x06b4, 0x0700, 0x0704, 0x0710, 0x0714, 0x0720, 0x0724,
+        0x0730, 0x0734, 0x0780, 0x0784, 0x0790, 0x0794, 0x07a0, 0x07a4, 0x07b0, 0x07b4,
+    };
+    return xlut[x % 128] + ylut[y % 128];
+}
+
+static inline u32 GetMortonOffset128(u32 x, u32 y, u32 bytes_per_pixel) {
+    // Calculates the offset of the position of the pixel in Morton order
+    // Framebuffer images are split into 128x128 tiles.
+
+    const unsigned int block_height = 128;
+    const unsigned int coarse_x = x & ~127;
+
+    u32 i = MortonInterleave128(x, y);
+
+    const unsigned int offset = coarse_x * block_height;
+
+    return (i + offset) * bytes_per_pixel;
+}
+
+static inline void MortonCopyPixels128(u32 width, u32 height, u32 bytes_per_pixel,
+                                       u32 gl_bytes_per_pixel, u8* morton_data, u8* gl_data,
+                                       bool morton_to_gl) {
+    u8* data_ptrs[2];
+    for (unsigned y = 0; y < height; ++y) {
+        for (unsigned x = 0; x < width; ++x) {
+            const u32 coarse_y = y & ~127;
+            u32 morton_offset =
+                GetMortonOffset128(x, y, bytes_per_pixel) + coarse_y * width * bytes_per_pixel;
+            u32 gl_pixel_index = (x + (height - 1 - y) * width) * gl_bytes_per_pixel;
+
+            data_ptrs[morton_to_gl] = morton_data + morton_offset;
+            data_ptrs[!morton_to_gl] = &gl_data[gl_pixel_index];
+
+            memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel);
+        }
+    }
+}
+
 } // namespace VideoCore

src/video_core/video_core.h

@@ -15,6 +15,8 @@ class RendererBase;
 
 namespace VideoCore {
 
+enum class Renderer { Software, OpenGL };
+
 extern std::unique_ptr<RendererBase> g_renderer; ///< Renderer plugin
 extern EmuWindow* g_emu_window;                  ///< Emu window