// Copyright 2014 Citra Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.

#include <algorithm>
#include <cstddef>
#include <cstdlib>
#include <memory>
#include <glad/glad.h>
#include "common/assert.h"
#include "common/bit_field.h"
#include "common/logging/log.h"
#include "core/core.h"
#include "core/core_timing.h"
#include "core/dumping/backend.h"
#include "core/frontend/emu_window.h"
#include "core/frontend/framebuffer_layout.h"
#include "core/hw/gpu.h"
#include "core/hw/hw.h"
#include "core/hw/lcd.h"
#include "core/memory.h"
#include "core/settings.h"
#include "core/tracer/recorder.h"
#include "video_core/debug_utils/debug_utils.h"
#include "video_core/rasterizer_interface.h"
#include "video_core/renderer_opengl/gl_vars.h"
#include "video_core/renderer_opengl/post_processing_opengl.h"
#include "video_core/renderer_opengl/renderer_opengl.h"
#include "video_core/video_core.h"

namespace Frontend {

struct Frame {
    GLuint index;
    GLsync render_sync;
    GLsync present_sync;
};
} // namespace Frontend

namespace OpenGL {

class OGLTextureMailbox : public Frontend::TextureMailbox {
public:
    Frontend::Frame render_tex = {0, 0, 0}, present_tex = {1, 0, 0}, off_tex = {2, 0, 0};
    bool swapped = false;
    std::mutex swap_mutex{};

    OGLTextureMailbox() = default;

    ~OGLTextureMailbox() = default;

    Frontend::Frame& GetRenderFrame() {
        return render_tex;
    }

    void RenderComplete() {
        std::scoped_lock lock(swap_mutex);
        swapped = true;
        std::swap(render_tex, off_tex);
    }

    Frontend::Frame& GetPresentationFrame() {
        return present_tex;
    }

    void PresentationComplete() {
        std::scoped_lock lock(swap_mutex);
        if (swapped) {
            swapped = false;
            std::swap(present_tex, off_tex);
        }
    }
};

static const char vertex_shader[] = R"(
in vec2 vert_position;
in vec2 vert_tex_coord;
out vec2 frag_tex_coord;

// This is a truncated 3x3 matrix for 2D transformations:
// The upper-left 2x2 submatrix performs scaling/rotation/mirroring.
// The third column performs translation.
// The third row could be used for projection, which we don't need in 2D. It hence is assumed to
// implicitly be [0, 0, 1]
uniform mat3x2 modelview_matrix;

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(modelview_matrix) * vert_position + modelview_matrix[2], 0.0, 1.0);
    frag_tex_coord = vert_tex_coord;
}
)";

static const char fragment_shader[] = R"(
in vec2 frag_tex_coord;
layout(location = 0) out vec4 color;

uniform vec4 i_resolution;
uniform vec4 o_resolution;
uniform int layer;

uniform sampler2D color_texture;

void main() {
    color = texture(color_texture, frag_tex_coord);
}
)";

static const char fragment_shader_anaglyph[] = R"(

// Anaglyph Red-Cyan shader based on Dubois algorithm
// Constants taken from the paper:
// "Conversion of a Stereo Pair to Anaglyph with
// the Least-Squares Projection Method"
// Eric Dubois, March 2009
const mat3 l = mat3( 0.437, 0.449, 0.164,
              -0.062,-0.062,-0.024,
              -0.048,-0.050,-0.017);
const mat3 r = mat3(-0.011,-0.032,-0.007,
               0.377, 0.761, 0.009,
              -0.026,-0.093, 1.234);

in vec2 frag_tex_coord;
out vec4 color;

uniform vec4 resolution;
uniform int layer;

uniform sampler2D color_texture;
uniform sampler2D color_texture_r;

void main() {
    vec4 color_tex_l = texture(color_texture, frag_tex_coord);
    vec4 color_tex_r = texture(color_texture_r, frag_tex_coord);
    color = vec4(color_tex_l.rgb*l+color_tex_r.rgb*r, color_tex_l.a);
}
)";

/**
 * Vertex structure that the drawn screen rectangles are composed of.
 */
struct ScreenRectVertex {
    ScreenRectVertex(GLfloat x, GLfloat y, GLfloat u, GLfloat v) {
        position[0] = x;
        position[1] = y;
        tex_coord[0] = u;
        tex_coord[1] = v;
    }

    GLfloat position[2];
    GLfloat tex_coord[2];
};

/**
 * Defines a 1:1 pixel ortographic projection matrix with (0,0) on the top-left
 * corner and (width, height) on the lower-bottom.
 *
 * The projection part of the matrix is trivial, hence these operations are represented
 * by a 3x2 matrix.
 */
static std::array<GLfloat, 3 * 2> MakeOrthographicMatrix(const float width, const float height) {
    std::array<GLfloat, 3 * 2> matrix; // Laid out in column-major order

    // clang-format off
    matrix[0] = 2.f / width; matrix[2] = 0.f;           matrix[4] = -1.f;
    matrix[1] = 0.f;         matrix[3] = -2.f / height; matrix[5] = 1.f;
    // Last matrix row is implicitly assumed to be [0, 0, 1].
    // clang-format on

    return matrix;
}

RendererOpenGL::RendererOpenGL(Frontend::EmuWindow& window) : RendererBase{window} {
    window.mailbox = std::make_unique<OGLTextureMailbox>();
}

RendererOpenGL::~RendererOpenGL() = default;

/// Swap buffers (render frame)
void RendererOpenGL::SwapBuffers() {
    // Maintain the rasterizer's state as a priority
    OpenGLState prev_state = OpenGLState::GetCurState();
    state.Apply();

    for (int i : {0, 1, 2}) {
        int fb_id = i == 2 ? 1 : 0;
        const auto& framebuffer = GPU::g_regs.framebuffer_config[fb_id];

        // Main LCD (0): 0x1ED02204, Sub LCD (1): 0x1ED02A04
        u32 lcd_color_addr =
            (fb_id == 0) ? LCD_REG_INDEX(color_fill_top) : LCD_REG_INDEX(color_fill_bottom);
        lcd_color_addr = HW::VADDR_LCD + 4 * lcd_color_addr;
        LCD::Regs::ColorFill color_fill = {0};
        LCD::Read(color_fill.raw, lcd_color_addr);

        if (color_fill.is_enabled) {
            LoadColorToActiveGLTexture(color_fill.color_r, color_fill.color_g, color_fill.color_b,
                                       screen_infos[i].texture);

            // Resize the texture in case the framebuffer size has changed
            screen_infos[i].texture.width = 1;
            screen_infos[i].texture.height = 1;
        } else {
            if (screen_infos[i].texture.width != (GLsizei)framebuffer.width ||
                screen_infos[i].texture.height != (GLsizei)framebuffer.height ||
                screen_infos[i].texture.format != framebuffer.color_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_infos[i].texture, framebuffer);
            }
            LoadFBToScreenInfo(framebuffer, screen_infos[i], i == 1);

            // Resize the texture in case the framebuffer size has changed
            screen_infos[i].texture.width = framebuffer.width;
            screen_infos[i].texture.height = framebuffer.height;
        }
    }

    if (VideoCore::g_renderer_screenshot_requested) {
        // Draw this frame to the screenshot framebuffer
        screenshot_framebuffer.Create();
        GLuint old_read_fb = state.draw.read_framebuffer;
        GLuint old_draw_fb = state.draw.draw_framebuffer;
        state.draw.read_framebuffer = state.draw.draw_framebuffer = screenshot_framebuffer.handle;
        state.Apply();

        Layout::FramebufferLayout layout{VideoCore::g_screenshot_framebuffer_layout};

        GLuint renderbuffer;
        glGenRenderbuffers(1, &renderbuffer);
        glBindRenderbuffer(GL_RENDERBUFFER, renderbuffer);
        glRenderbufferStorage(GL_RENDERBUFFER, GL_RGB8, layout.width, layout.height);
        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER,
                                  renderbuffer);

        DrawScreens(layout);

        glReadPixels(0, 0, layout.width, layout.height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV,
                     VideoCore::g_screenshot_bits);

        screenshot_framebuffer.Release();
        state.draw.read_framebuffer = old_read_fb;
        state.draw.draw_framebuffer = old_draw_fb;
        state.Apply();
        glDeleteRenderbuffers(1, &renderbuffer);

        VideoCore::g_screenshot_complete_callback();
        VideoCore::g_renderer_screenshot_requested = false;
    }

    if (cleanup_video_dumping.exchange(false)) {
        ReleaseVideoDumpingGLObjects();
    }

    if (Core::System::GetInstance().VideoDumper().IsDumping()) {
        if (prepare_video_dumping.exchange(false)) {
            InitVideoDumpingGLObjects();
        }

        const auto& layout = Core::System::GetInstance().VideoDumper().GetLayout();
        glBindFramebuffer(GL_READ_FRAMEBUFFER, frame_dumping_framebuffer.handle);
        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, frame_dumping_framebuffer.handle);
        DrawScreens(layout);

        glBindBuffer(GL_PIXEL_PACK_BUFFER, frame_dumping_pbos[current_pbo].handle);
        glReadPixels(0, 0, layout.width, layout.height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, 0);
        glBindBuffer(GL_PIXEL_PACK_BUFFER, frame_dumping_pbos[next_pbo].handle);

        GLubyte* pixels = static_cast<GLubyte*>(glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY));
        VideoDumper::VideoFrame frame_data{layout.width, layout.height, pixels};
        Core::System::GetInstance().VideoDumper().AddVideoFrame(frame_data);

        glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
        glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
        current_pbo = (current_pbo + 1) % 2;
        next_pbo = (current_pbo + 1) % 2;
    }

    const auto& layout = render_window.GetFramebufferLayout();
    auto& frame = render_window.mailbox->GetRenderFrame();
    auto& presentation = presentation_textures[frame.index];

    // Clean up sync objects before drawing

    // INTEL driver workaround. We can't delete the previous render sync object until we are sure
    // that the presentation is done
    if (frame.present_sync) {
        glClientWaitSync(frame.present_sync, 0, GL_TIMEOUT_IGNORED);
    }

    // delete the draw fence if the frame wasn't presented
    if (frame.render_sync) {
        glDeleteSync(frame.render_sync);
        frame.render_sync = 0;
    }

    // wait for the presentation to be done
    if (frame.present_sync) {
        glWaitSync(frame.present_sync, 0, GL_TIMEOUT_IGNORED);
        glDeleteSync(frame.present_sync);
        frame.present_sync = 0;
    }

    // Recreate the presentation texture if the size of the window has changed
    if (layout.width != presentation.width || layout.height != presentation.height) {
        presentation.width = layout.width;
        presentation.height = layout.height;
        presentation.texture.Release();
        presentation.texture.Create();
        state.texture_units[0].texture_2d = presentation.texture.handle;
        state.Apply();
        glActiveTexture(GL_TEXTURE0);
        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, layout.width, layout.height, 0, GL_RGBA,
                     GL_UNSIGNED_BYTE, 0);
        state.texture_units[0].texture_2d = 0;
        state.Apply();
    }

    GLuint render_texture = presentation.texture.handle;
    state.draw.draw_framebuffer = draw_framebuffer.handle;
    state.Apply();
    glFramebufferTexture(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, render_texture, 0);
    GLenum DrawBuffers[1] = {GL_COLOR_ATTACHMENT0};
    glDrawBuffers(1, DrawBuffers); // "1" is the size of DrawBuffers
    DrawScreens(layout);
    // Create a fence for the frontend to wait on and swap this frame to OffTex
    frame.render_sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
    glFlush();
    render_window.mailbox->RenderComplete();
    m_current_frame++;

    Core::System::GetInstance().perf_stats->EndSystemFrame();

    render_window.PollEvents();

    Core::System::GetInstance().frame_limiter.DoFrameLimiting(
        Core::System::GetInstance().CoreTiming().GetGlobalTimeUs());
    Core::System::GetInstance().perf_stats->BeginSystemFrame();

    prev_state.Apply();
    RefreshRasterizerSetting();

    if (Pica::g_debug_context && Pica::g_debug_context->recorder) {
        Pica::g_debug_context->recorder->FrameFinished();
    }
}

/**
 * Loads framebuffer from emulated memory into the active OpenGL texture.
 */
void RendererOpenGL::LoadFBToScreenInfo(const GPU::Regs::FramebufferConfig& framebuffer,
                                        ScreenInfo& screen_info, bool right_eye) {

    if (framebuffer.address_right1 == 0 || framebuffer.address_right2 == 0)
        right_eye = false;

    const PAddr framebuffer_addr =
        framebuffer.active_fb == 0
            ? (!right_eye ? framebuffer.address_left1 : framebuffer.address_right1)
            : (!right_eye ? framebuffer.address_left2 : framebuffer.address_right2);

    LOG_TRACE(Render_OpenGL, "0x{:08x} bytes from 0x{:08x}({}x{}), fmt {:x}",
              framebuffer.stride * framebuffer.height, framebuffer_addr, (int)framebuffer.width,
              (int)framebuffer.height, (int)framebuffer.format);

    int bpp = GPU::Regs::BytesPerPixel(framebuffer.color_format);
    std::size_t pixel_stride = framebuffer.stride / bpp;

    // OpenGL only supports specifying a stride in units of pixels, not bytes, unfortunately
    ASSERT(pixel_stride * bpp == framebuffer.stride);

    // Ensure no bad interactions with GL_UNPACK_ALIGNMENT, which by default
    // only allows rows to have a memory alignement of 4.
    ASSERT(pixel_stride % 4 == 0);

    if (!Rasterizer()->AccelerateDisplay(framebuffer, framebuffer_addr,
                                         static_cast<u32>(pixel_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 = Common::Rectangle<float>(0.f, 0.f, 1.f, 1.f);

        Memory::RasterizerFlushRegion(framebuffer_addr, framebuffer.stride * framebuffer.height);

        const u8* framebuffer_data = VideoCore::g_memory->GetPhysicalPointer(framebuffer_addr);

        state.texture_units[0].texture_2d = screen_info.texture.resource.handle;
        state.Apply();

        glActiveTexture(GL_TEXTURE0);
        glPixelStorei(GL_UNPACK_ROW_LENGTH, (GLint)pixel_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.width, framebuffer.height,
                        screen_info.texture.gl_format, screen_info.texture.gl_type,
                        framebuffer_data);

        glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);

        state.texture_units[0].texture_2d = 0;
        state.Apply();
    }
}

/**
 * Fills active OpenGL texture with the given RGB color. Since the color is solid, the texture can
 * be 1x1 but will stretch across whatever it's rendered on.
 */
void RendererOpenGL::LoadColorToActiveGLTexture(u8 color_r, u8 color_g, u8 color_b,
                                                const TextureInfo& texture) {
    state.texture_units[0].texture_2d = texture.resource.handle;
    state.Apply();

    glActiveTexture(GL_TEXTURE0);
    u8 framebuffer_data[3] = {color_r, color_g, color_b};

    // Update existing texture
    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB, 1, 1, 0, GL_RGB, GL_UNSIGNED_BYTE, framebuffer_data);

    state.texture_units[0].texture_2d = 0;
    state.Apply();
}

/**
 * Initializes the OpenGL state and creates persistent objects.
 */
void RendererOpenGL::InitOpenGLObjects() {
    glClearColor(Settings::values.bg_red, Settings::values.bg_green, Settings::values.bg_blue,
                 0.0f);

    filter_sampler.Create();
    ReloadSampler();

    ReloadShader();

    // Generate VBO handle for drawing
    vertex_buffer.Create();

    // Generate VAO
    vertex_array.Create();

    state.draw.vertex_array = vertex_array.handle;
    state.draw.vertex_buffer = vertex_buffer.handle;
    state.draw.uniform_buffer = 0;
    state.Apply();

    // Attach vertex data to VAO
    glBufferData(GL_ARRAY_BUFFER, sizeof(ScreenRectVertex) * 4, nullptr, GL_STREAM_DRAW);
    glVertexAttribPointer(attrib_position, 2, GL_FLOAT, GL_FALSE, sizeof(ScreenRectVertex),
                          (GLvoid*)offsetof(ScreenRectVertex, position));
    glVertexAttribPointer(attrib_tex_coord, 2, GL_FLOAT, GL_FALSE, sizeof(ScreenRectVertex),
                          (GLvoid*)offsetof(ScreenRectVertex, tex_coord));
    glEnableVertexAttribArray(attrib_position);
    glEnableVertexAttribArray(attrib_tex_coord);

    // Allocate textures for each screen
    for (auto& screen_info : screen_infos) {
        screen_info.texture.resource.Create();

        // Allocation of storage is deferred until the first frame, when we
        // know the framebuffer size.

        state.texture_units[0].texture_2d = screen_info.texture.resource.handle;
        state.Apply();

        glActiveTexture(GL_TEXTURE0);
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 0);
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);

        screen_info.display_texture = screen_info.texture.resource.handle;
    }

    draw_framebuffer.Create();
    presentation_framebuffer.Create();
    presentation_textures[0].texture.Create();
    presentation_textures[1].texture.Create();
    presentation_textures[2].texture.Create();
    state.texture_units[0].texture_2d = 0;
    state.Apply();
}

void RendererOpenGL::ReloadSampler() {
    glSamplerParameteri(filter_sampler.handle, GL_TEXTURE_MIN_FILTER,
                        Settings::values.filter_mode ? GL_LINEAR : GL_NEAREST);
    glSamplerParameteri(filter_sampler.handle, GL_TEXTURE_MAG_FILTER,
                        Settings::values.filter_mode ? GL_LINEAR : GL_NEAREST);
    glSamplerParameteri(filter_sampler.handle, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
    glSamplerParameteri(filter_sampler.handle, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
}

void RendererOpenGL::ReloadShader() {
    // Link shaders and get variable locations
    std::string shader_data;
    if (GLES) {
        shader_data += fragment_shader_precision_OES;
    }
    if (Settings::values.render_3d == Settings::StereoRenderOption::Anaglyph) {
        if (Settings::values.pp_shader_name == "dubois (builtin)") {
            shader_data += fragment_shader_anaglyph;
        } else {
            std::string shader_text =
                OpenGL::GetPostProcessingShaderCode(true, Settings::values.pp_shader_name);
            if (shader_text.empty()) {
                // Should probably provide some information that the shader couldn't load
                shader_data += fragment_shader_anaglyph;
            } else {
                shader_data += shader_text;
            }
        }
    } else {
        if (Settings::values.pp_shader_name == "none (builtin)") {
            shader_data += fragment_shader;
        } else {
            std::string shader_text =
                OpenGL::GetPostProcessingShaderCode(false, Settings::values.pp_shader_name);
            if (shader_text.empty()) {
                // Should probably provide some information that the shader couldn't load
                shader_data += fragment_shader;
            } else {
                shader_data += shader_text;
            }
        }
    }
    shader.Create(vertex_shader, shader_data.c_str());
    state.draw.shader_program = shader.handle;
    state.Apply();
    uniform_modelview_matrix = glGetUniformLocation(shader.handle, "modelview_matrix");
    uniform_color_texture = glGetUniformLocation(shader.handle, "color_texture");
    if (Settings::values.render_3d == Settings::StereoRenderOption::Anaglyph) {
        uniform_color_texture_r = glGetUniformLocation(shader.handle, "color_texture_r");
    }
    uniform_i_resolution = glGetUniformLocation(shader.handle, "i_resolution");
    uniform_o_resolution = glGetUniformLocation(shader.handle, "o_resolution");
    uniform_layer = glGetUniformLocation(shader.handle, "layer");
    attrib_position = glGetAttribLocation(shader.handle, "vert_position");
    attrib_tex_coord = glGetAttribLocation(shader.handle, "vert_tex_coord");
}

void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
                                                 const GPU::Regs::FramebufferConfig& framebuffer) {
    GPU::Regs::PixelFormat format = framebuffer.color_format;
    GLint internal_format;

    texture.format = format;
    texture.width = framebuffer.width;
    texture.height = framebuffer.height;

    switch (format) {
    case GPU::Regs::PixelFormat::RGBA8:
        internal_format = GL_RGBA;
        texture.gl_format = GL_RGBA;
        texture.gl_type = GLES ? GL_UNSIGNED_BYTE : GL_UNSIGNED_INT_8_8_8_8;
        break;

    case GPU::Regs::PixelFormat::RGB8:
        // This pixel format uses BGR since GL_UNSIGNED_BYTE specifies byte-order, unlike every
        // specific OpenGL type used in this function using native-endian (that is, little-endian
        // mostly everywhere) for words or half-words.
        // TODO: check how those behave on big-endian processors.
        internal_format = GL_RGB;

        // GLES Dosen't support BGR , Use RGB instead
        texture.gl_format = GLES ? GL_RGB : GL_BGR;
        texture.gl_type = GL_UNSIGNED_BYTE;
        break;

    case GPU::Regs::PixelFormat::RGB565:
        internal_format = GL_RGB;
        texture.gl_format = GL_RGB;
        texture.gl_type = GL_UNSIGNED_SHORT_5_6_5;
        break;

    case GPU::Regs::PixelFormat::RGB5A1:
        internal_format = GL_RGBA;
        texture.gl_format = GL_RGBA;
        texture.gl_type = GL_UNSIGNED_SHORT_5_5_5_1;
        break;

    case GPU::Regs::PixelFormat::RGBA4:
        internal_format = GL_RGBA;
        texture.gl_format = GL_RGBA;
        texture.gl_type = GL_UNSIGNED_SHORT_4_4_4_4;
        break;

    default:
        UNIMPLEMENTED();
    }

    state.texture_units[0].texture_2d = texture.resource.handle;
    state.Apply();

    glActiveTexture(GL_TEXTURE0);
    glTexImage2D(GL_TEXTURE_2D, 0, internal_format, texture.width, texture.height, 0,
                 texture.gl_format, texture.gl_type, nullptr);

    state.texture_units[0].texture_2d = 0;
    state.Apply();
}

/**
 * Draws a single texture to the emulator window, rotating the texture to correct for the 3DS's LCD
 * rotation.
 */
void RendererOpenGL::DrawSingleScreenRotated(const ScreenInfo& screen_info, float x, float y,
                                             float w, float h) {
    const auto& texcoords = screen_info.display_texcoords;

    const std::array<ScreenRectVertex, 4> vertices = {{
        ScreenRectVertex(x, y, texcoords.bottom, texcoords.left),
        ScreenRectVertex(x + w, y, texcoords.bottom, texcoords.right),
        ScreenRectVertex(x, y + h, texcoords.top, texcoords.left),
        ScreenRectVertex(x + w, y + h, texcoords.top, texcoords.right),
    }};

    // As this is the "DrawSingleScreenRotated" function, the output resolution dimensions have been
    // swapped. If a non-rotated draw-screen function were to be added for book-mode games, those
    // should probably be set to the standard (w, h, 1.0 / w, 1.0 / h) ordering.
    u16 scale_factor = VideoCore::GetResolutionScaleFactor();
    glUniform4f(uniform_i_resolution, screen_info.texture.width * scale_factor,
                screen_info.texture.height * scale_factor,
                1.0 / (screen_info.texture.width * scale_factor),
                1.0 / (screen_info.texture.height * scale_factor));
    glUniform4f(uniform_o_resolution, h, w, 1.0f / h, 1.0f / w);
    state.texture_units[0].texture_2d = screen_info.display_texture;
    state.texture_units[0].sampler = filter_sampler.handle;
    state.Apply();

    glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(vertices), vertices.data());
    glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);

    state.texture_units[0].texture_2d = 0;
    state.texture_units[0].sampler = 0;
    state.Apply();
}

/**
 * Draws a single texture to the emulator window, rotating the texture to correct for the 3DS's LCD
 * rotation.
 */
void RendererOpenGL::DrawSingleScreenAnaglyphRotated(const ScreenInfo& screen_info_l,
                                                     const ScreenInfo& screen_info_r, float x,
                                                     float y, float w, float h) {
    const auto& texcoords = screen_info_l.display_texcoords;

    const std::array<ScreenRectVertex, 4> vertices = {{
        ScreenRectVertex(x, y, texcoords.bottom, texcoords.left),
        ScreenRectVertex(x + w, y, texcoords.bottom, texcoords.right),
        ScreenRectVertex(x, y + h, texcoords.top, texcoords.left),
        ScreenRectVertex(x + w, y + h, texcoords.top, texcoords.right),
    }};

    u16 scale_factor = VideoCore::GetResolutionScaleFactor();
    glUniform4f(uniform_i_resolution, screen_info_l.texture.width * scale_factor,
                screen_info_l.texture.height * scale_factor,
                1.0 / (screen_info_l.texture.width * scale_factor),
                1.0 / (screen_info_l.texture.height * scale_factor));
    glUniform4f(uniform_o_resolution, h, w, 1.0f / h, 1.0f / w);
    state.texture_units[0].texture_2d = screen_info_l.display_texture;
    state.texture_units[1].texture_2d = screen_info_r.display_texture;
    state.texture_units[0].sampler = filter_sampler.handle;
    state.texture_units[1].sampler = filter_sampler.handle;
    state.Apply();

    glBufferSubData(GL_ARRAY_BUFFER, 0, sizeof(vertices), vertices.data());
    glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);

    state.texture_units[0].texture_2d = 0;
    state.texture_units[1].texture_2d = 0;
    state.texture_units[0].sampler = 0;
    state.texture_units[1].sampler = 0;
    state.Apply();
}

/**
 * Draws the emulated screens to the emulator window.
 */
void RendererOpenGL::DrawScreens(const Layout::FramebufferLayout& layout) {
    if (VideoCore::g_renderer_bg_color_update_requested.exchange(false)) {
        // Update background color before drawing
        glClearColor(Settings::values.bg_red, Settings::values.bg_green, Settings::values.bg_blue,
                     0.0f);
    }

    if (VideoCore::g_renderer_sampler_update_requested.exchange(false)) {
        // Set the new filtering mode for the sampler
        ReloadSampler();
    }

    if (VideoCore::g_renderer_shader_update_requested.exchange(false)) {
        // Update fragment shader before drawing
        shader.Release();
        // Link shaders and get variable locations
        ReloadShader();
    }

    const auto& top_screen = layout.top_screen;
    const auto& bottom_screen = layout.bottom_screen;

    glViewport(0, 0, layout.width, layout.height);
    glClear(GL_COLOR_BUFFER_BIT);

    // Set projection matrix
    std::array<GLfloat, 3 * 2> ortho_matrix =
        MakeOrthographicMatrix((float)layout.width, (float)layout.height);
    glUniformMatrix3x2fv(uniform_modelview_matrix, 1, GL_FALSE, ortho_matrix.data());

    // Bind texture in Texture Unit 0
    glUniform1i(uniform_color_texture, 0);

    // Bind a second texture for the right eye if in Anaglyph mode
    if (Settings::values.render_3d == Settings::StereoRenderOption::Anaglyph) {
        glUniform1i(uniform_color_texture_r, 1);
    }

    glUniform1i(uniform_layer, 0);
    if (layout.top_screen_enabled) {
        if (Settings::values.render_3d == Settings::StereoRenderOption::Off) {
            DrawSingleScreenRotated(screen_infos[0], (float)top_screen.left, (float)top_screen.top,
                                    (float)top_screen.GetWidth(), (float)top_screen.GetHeight());
        } else if (Settings::values.render_3d == Settings::StereoRenderOption::SideBySide) {
            DrawSingleScreenRotated(screen_infos[0], (float)top_screen.left / 2,
                                    (float)top_screen.top, (float)top_screen.GetWidth() / 2,
                                    (float)top_screen.GetHeight());
            glUniform1i(uniform_layer, 1);
            DrawSingleScreenRotated(screen_infos[1],
                                    ((float)top_screen.left / 2) + ((float)layout.width / 2),
                                    (float)top_screen.top, (float)top_screen.GetWidth() / 2,
                                    (float)top_screen.GetHeight());
        } else if (Settings::values.render_3d == Settings::StereoRenderOption::Anaglyph) {
            DrawSingleScreenAnaglyphRotated(
                screen_infos[0], screen_infos[1], (float)top_screen.left, (float)top_screen.top,
                (float)top_screen.GetWidth(), (float)top_screen.GetHeight());
        }
    }
    glUniform1i(uniform_layer, 0);
    if (layout.bottom_screen_enabled) {
        if (Settings::values.render_3d == Settings::StereoRenderOption::Off) {
            DrawSingleScreenRotated(screen_infos[2], (float)bottom_screen.left,
                                    (float)bottom_screen.top, (float)bottom_screen.GetWidth(),
                                    (float)bottom_screen.GetHeight());
        } else if (Settings::values.render_3d == Settings::StereoRenderOption::SideBySide) {
            DrawSingleScreenRotated(screen_infos[2], (float)bottom_screen.left / 2,
                                    (float)bottom_screen.top, (float)bottom_screen.GetWidth() / 2,
                                    (float)bottom_screen.GetHeight());
            glUniform1i(uniform_layer, 1);
            DrawSingleScreenRotated(screen_infos[2],
                                    ((float)bottom_screen.left / 2) + ((float)layout.width / 2),
                                    (float)bottom_screen.top, (float)bottom_screen.GetWidth() / 2,
                                    (float)bottom_screen.GetHeight());
        } else if (Settings::values.render_3d == Settings::StereoRenderOption::Anaglyph) {
            DrawSingleScreenAnaglyphRotated(screen_infos[2], screen_infos[2],
                                            (float)bottom_screen.left, (float)bottom_screen.top,
                                            (float)bottom_screen.GetWidth(),
                                            (float)bottom_screen.GetHeight());
        }
    }
}

void RendererOpenGL::Present() {
    const auto& layout = render_window.GetFramebufferLayout();
    auto& frame = render_window.mailbox->GetPresentationFrame();
    const auto& presentation = presentation_textures[frame.index];
    const GLuint texture_handle = presentation.texture.handle;

    glWaitSync(frame.render_sync, 0, GL_TIMEOUT_IGNORED);
    // INTEL workaround.
    // Normally we could just delete the draw fence here, but due to driver bugs, we can just delete
    // it on the emulation thread without too much penalty
    // glDeleteSync(frame.render_sync);
    // frame.render_sync = 0;

    glClearColor(Settings::values.bg_red, Settings::values.bg_green, Settings::values.bg_blue,
                 0.0f);
    glClear(GL_COLOR_BUFFER_BIT);

    glBindFramebuffer(GL_READ_FRAMEBUFFER, presentation_framebuffer.handle);

    glActiveTexture(GL_TEXTURE0);
    glBindTexture(GL_TEXTURE_2D, texture_handle);
    glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture_handle,
                           0);
    glBlitFramebuffer(0, 0, presentation.width, presentation.height, 0, 0, layout.width,
                      layout.height, GL_COLOR_BUFFER_BIT, GL_LINEAR);

    /* insert fence for the main thread to block on */
    frame.present_sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
    glFlush();
    render_window.mailbox->PresentationComplete();
}

/// Updates the framerate
void RendererOpenGL::UpdateFramerate() {}

void RendererOpenGL::PrepareVideoDumping() {
    prepare_video_dumping = true;
}

void RendererOpenGL::CleanupVideoDumping() {
    cleanup_video_dumping = true;
}

void RendererOpenGL::InitVideoDumpingGLObjects() {
    const auto& layout = Core::System::GetInstance().VideoDumper().GetLayout();

    frame_dumping_framebuffer.Create();
    glGenRenderbuffers(1, &frame_dumping_renderbuffer);
    glBindRenderbuffer(GL_RENDERBUFFER, frame_dumping_renderbuffer);
    glRenderbufferStorage(GL_RENDERBUFFER, GL_RGB8, layout.width, layout.height);
    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, frame_dumping_framebuffer.handle);
    glFramebufferRenderbuffer(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER,
                              frame_dumping_renderbuffer);
    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);

    for (auto& buffer : frame_dumping_pbos) {
        buffer.Create();
        glBindBuffer(GL_PIXEL_PACK_BUFFER, buffer.handle);
        glBufferData(GL_PIXEL_PACK_BUFFER, layout.width * layout.height * 4, nullptr,
                     GL_STREAM_READ);
        glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
    }
}

void RendererOpenGL::ReleaseVideoDumpingGLObjects() {
    frame_dumping_framebuffer.Release();
    glDeleteRenderbuffers(1, &frame_dumping_renderbuffer);

    for (auto& buffer : frame_dumping_pbos) {
        buffer.Release();
    }
}

static const char* GetSource(GLenum source) {
#define RET(s)                                                                                     \
    case GL_DEBUG_SOURCE_##s:                                                                      \
        return #s
    switch (source) {
        RET(API);
        RET(WINDOW_SYSTEM);
        RET(SHADER_COMPILER);
        RET(THIRD_PARTY);
        RET(APPLICATION);
        RET(OTHER);
    default:
        UNREACHABLE();
    }
#undef RET
}

static const char* GetType(GLenum type) {
#define RET(t)                                                                                     \
    case GL_DEBUG_TYPE_##t:                                                                        \
        return #t
    switch (type) {
        RET(ERROR);
        RET(DEPRECATED_BEHAVIOR);
        RET(UNDEFINED_BEHAVIOR);
        RET(PORTABILITY);
        RET(PERFORMANCE);
        RET(OTHER);
        RET(MARKER);
    default:
        UNREACHABLE();
    }
#undef RET
}

static void APIENTRY DebugHandler(GLenum source, GLenum type, GLuint id, GLenum severity,
                                  GLsizei length, const GLchar* message, const void* user_param) {
    Log::Level level;
    switch (severity) {
    case GL_DEBUG_SEVERITY_HIGH:
        level = Log::Level::Critical;
        break;
    case GL_DEBUG_SEVERITY_MEDIUM:
        level = Log::Level::Warning;
        break;
    case GL_DEBUG_SEVERITY_NOTIFICATION:
    case GL_DEBUG_SEVERITY_LOW:
        level = Log::Level::Debug;
        break;
    }
    LOG_GENERIC(Log::Class::Render_OpenGL, level, "{} {} {}: {}", GetSource(source), GetType(type),
                id, message);
}

/// Initialize the renderer
Core::System::ResultStatus RendererOpenGL::Init() {
    if (!gladLoadGL()) {
        return Core::System::ResultStatus::ErrorVideoCore_ErrorBelowGL33;
    }

    if (GLAD_GL_KHR_debug) {
        glEnable(GL_DEBUG_OUTPUT);
        glDebugMessageCallback(DebugHandler, nullptr);
    }

    const char* gl_version{reinterpret_cast<char const*>(glGetString(GL_VERSION))};
    const char* gpu_vendor{reinterpret_cast<char const*>(glGetString(GL_VENDOR))};
    const char* gpu_model{reinterpret_cast<char const*>(glGetString(GL_RENDERER))};

    LOG_INFO(Render_OpenGL, "GL_VERSION: {}", gl_version);
    LOG_INFO(Render_OpenGL, "GL_VENDOR: {}", gpu_vendor);
    LOG_INFO(Render_OpenGL, "GL_RENDERER: {}", gpu_model);

    auto& telemetry_session = Core::System::GetInstance().TelemetrySession();
    telemetry_session.AddField(Telemetry::FieldType::UserSystem, "GPU_Vendor", gpu_vendor);
    telemetry_session.AddField(Telemetry::FieldType::UserSystem, "GPU_Model", gpu_model);
    telemetry_session.AddField(Telemetry::FieldType::UserSystem, "GPU_OpenGL_Version", gl_version);

    if (!strcmp(gpu_vendor, "GDI Generic")) {
        return Core::System::ResultStatus::ErrorVideoCore_ErrorGenericDrivers;
    }

    if (!(GLAD_GL_VERSION_3_3 || GLAD_GL_ES_VERSION_3_1)) {
        return Core::System::ResultStatus::ErrorVideoCore_ErrorBelowGL33;
    }

    InitOpenGLObjects();

    RefreshRasterizerSetting();

    return Core::System::ResultStatus::Success;
}

/// Shutdown the renderer
void RendererOpenGL::ShutDown() {}

} // namespace OpenGL