ladybird/Userland/Libraries/LibSoftGPU/Device.cpp

/*
 * Copyright (c) 2021, Stephan Unverwerth <s.unverwerth@serenityos.org>
 * Copyright (c) 2021, Jesse Buhagiar <jooster669@gmail.com>
 *
 * SPDX-License-Identifier: BSD-2-Clause
 */

#include <AK/Function.h>
#include <AK/SIMDExtras.h>
#include <AK/SIMDMath.h>
#include <LibCore/ElapsedTimer.h>
#include <LibGfx/Painter.h>
#include <LibGfx/Vector2.h>
#include <LibGfx/Vector3.h>
#include <LibSoftGPU/Config.h>
#include <LibSoftGPU/Device.h>
#include <LibSoftGPU/PixelQuad.h>
#include <LibSoftGPU/SIMD.h>

namespace SoftGPU {

static long long g_num_rasterized_triangles;
static long long g_num_pixels;
static long long g_num_pixels_shaded;
static long long g_num_pixels_blended;
static long long g_num_sampler_calls;

using IntVector2 = Gfx::Vector2<int>;
using IntVector3 = Gfx::Vector3<int>;

using AK::SIMD::exp;
using AK::SIMD::expand4;
using AK::SIMD::f32x4;
using AK::SIMD::i32x4;
using AK::SIMD::load4_masked;
using AK::SIMD::maskbits;
using AK::SIMD::maskcount;
using AK::SIMD::none;
using AK::SIMD::store4_masked;
using AK::SIMD::to_f32x4;

constexpr static int edge_function(const IntVector2& a, const IntVector2& b, const IntVector2& c)
{
    return ((c.x() - a.x()) * (b.y() - a.y()) - (c.y() - a.y()) * (b.x() - a.x()));
}

constexpr static i32x4 edge_function4(const IntVector2& a, const IntVector2& b, const Vector2<i32x4>& c)
{
    return ((c.x() - a.x()) * (b.y() - a.y()) - (c.y() - a.y()) * (b.x() - a.x()));
}

template<typename T, typename U>
constexpr static auto interpolate(const T& v0, const T& v1, const T& v2, const Vector3<U>& barycentric_coords)
{
    return v0 * barycentric_coords.x() + v1 * barycentric_coords.y() + v2 * barycentric_coords.z();
}

ALWAYS_INLINE constexpr static Gfx::RGBA32 to_rgba32(const FloatVector4& v)
{
    auto clamped = v.clamped(0, 1);
    u8 r = clamped.x() * 255;
    u8 g = clamped.y() * 255;
    u8 b = clamped.z() * 255;
    u8 a = clamped.w() * 255;
    return a << 24 | r << 16 | g << 8 | b;
}

static FloatVector4 to_vec4(Gfx::RGBA32 rgba)
{
    auto constexpr one_over_255 = 1.0f / 255;
    return {
        ((rgba >> 16) & 0xff) * one_over_255,
        ((rgba >> 8) & 0xff) * one_over_255,
        (rgba & 0xff) * one_over_255,
        ((rgba >> 24) & 0xff) * one_over_255,
    };
}

static Gfx::IntRect scissor_box_to_window_coordinates(Gfx::IntRect const& scissor_box, Gfx::IntRect const& window_rect)
{
    return scissor_box.translated(0, window_rect.height() - 2 * scissor_box.y() - scissor_box.height());
}

static constexpr void setup_blend_factors(BlendFactor mode, FloatVector4& constant, float& src_alpha, float& dst_alpha, float& src_color, float& dst_color)
{
    constant = { 0.0f, 0.0f, 0.0f, 0.0f };
    src_alpha = 0;
    dst_alpha = 0;
    src_color = 0;
    dst_color = 0;

    switch (mode) {
    case BlendFactor::Zero:
        break;
    case BlendFactor::One:
        constant = { 1.0f, 1.0f, 1.0f, 1.0f };
        break;
    case BlendFactor::SrcColor:
        src_color = 1;
        break;
    case BlendFactor::OneMinusSrcColor:
        constant = { 1.0f, 1.0f, 1.0f, 1.0f };
        src_color = -1;
        break;
    case BlendFactor::SrcAlpha:
        src_alpha = 1;
        break;
    case BlendFactor::OneMinusSrcAlpha:
        constant = { 1.0f, 1.0f, 1.0f, 1.0f };
        src_alpha = -1;
        break;
    case BlendFactor::DstAlpha:
        dst_alpha = 1;
        break;
    case BlendFactor::OneMinusDstAlpha:
        constant = { 1.0f, 1.0f, 1.0f, 1.0f };
        dst_alpha = -1;
        break;
    case BlendFactor::DstColor:
        dst_color = 1;
        break;
    case BlendFactor::OneMinusDstColor:
        constant = { 1.0f, 1.0f, 1.0f, 1.0f };
        dst_color = -1;
        break;
    case BlendFactor::SrcAlphaSaturate:
        // FIXME: How do we implement this?
        break;
    default:
        VERIFY_NOT_REACHED();
    }
}

template<typename PS>
static void rasterize_triangle(const RasterizerOptions& options, Gfx::Bitmap& render_target, DepthBuffer& depth_buffer, const Triangle& triangle, PS pixel_shader)
{
    INCREASE_STATISTICS_COUNTER(g_num_rasterized_triangles, 1);

    // Since the algorithm is based on blocks of uniform size, we need
    // to ensure that our render_target size is actually a multiple of the block size
    VERIFY((render_target.width() % 2) == 0);
    VERIFY((render_target.height() % 2) == 0);

    // Return if alpha testing is a no-op
    if (options.enable_alpha_test && options.alpha_test_func == AlphaTestFunction::Never)
        return;

    // Vertices
    Vertex const vertex0 = triangle.vertices[0];
    Vertex const vertex1 = triangle.vertices[1];
    Vertex const vertex2 = triangle.vertices[2];

    constexpr int subpixel_factor = 1 << SUBPIXEL_BITS;

    // Calculate area of the triangle for later tests
    IntVector2 const v0 { static_cast<int>(vertex0.window_coordinates.x() * subpixel_factor), static_cast<int>(vertex0.window_coordinates.y() * subpixel_factor) };
    IntVector2 const v1 { static_cast<int>(vertex1.window_coordinates.x() * subpixel_factor), static_cast<int>(vertex1.window_coordinates.y() * subpixel_factor) };
    IntVector2 const v2 { static_cast<int>(vertex2.window_coordinates.x() * subpixel_factor), static_cast<int>(vertex2.window_coordinates.y() * subpixel_factor) };

    int area = edge_function(v0, v1, v2);
    if (area == 0)
        return;

    auto const one_over_area = 1.0f / area;

    FloatVector4 src_constant {};
    float src_factor_src_alpha = 0;
    float src_factor_dst_alpha = 0;
    float src_factor_src_color = 0;
    float src_factor_dst_color = 0;

    FloatVector4 dst_constant {};
    float dst_factor_src_alpha = 0;
    float dst_factor_dst_alpha = 0;
    float dst_factor_src_color = 0;
    float dst_factor_dst_color = 0;

    if (options.enable_blending) {
        setup_blend_factors(
            options.blend_source_factor,
            src_constant,
            src_factor_src_alpha,
            src_factor_dst_alpha,
            src_factor_src_color,
            src_factor_dst_color);

        setup_blend_factors(
            options.blend_destination_factor,
            dst_constant,
            dst_factor_src_alpha,
            dst_factor_dst_alpha,
            dst_factor_src_color,
            dst_factor_dst_color);
    }

    auto render_bounds = render_target.rect();
    auto window_scissor_rect = scissor_box_to_window_coordinates(options.scissor_box, render_target.rect());
    if (options.scissor_enabled)
        render_bounds.intersect(window_scissor_rect);

    // Obey top-left rule:
    // This sets up "zero" for later pixel coverage tests.
    // Depending on where on the triangle the edge is located
    // it is either tested against 0 or 1, effectively
    // turning "< 0" into "<= 0"
    IntVector3 zero { 1, 1, 1 };
    if (v1.y() > v0.y() || (v1.y() == v0.y() && v1.x() < v0.x()))
        zero.set_z(0);
    if (v2.y() > v1.y() || (v2.y() == v1.y() && v2.x() < v1.x()))
        zero.set_x(0);
    if (v0.y() > v2.y() || (v0.y() == v2.y() && v0.x() < v2.x()))
        zero.set_y(0);

    // This function calculates the 3 edge values for the pixel relative to the triangle.
    auto calculate_edge_values4 = [v0, v1, v2](const Vector2<i32x4>& p) -> Vector3<i32x4> {
        return {
            edge_function4(v1, v2, p),
            edge_function4(v2, v0, p),
            edge_function4(v0, v1, p),
        };
    };

    // This function tests whether a point as identified by its 3 edge values lies within the triangle
    auto test_point4 = [zero](const Vector3<i32x4>& edges) -> i32x4 {
        return edges.x() >= zero.x()
            && edges.y() >= zero.y()
            && edges.z() >= zero.z();
    };

    auto test_scissor4 = [window_scissor_rect](const Vector2<i32x4>& screen_coordinates) -> i32x4 {
        return screen_coordinates.x() >= window_scissor_rect.x()
            && screen_coordinates.x() < window_scissor_rect.x() + window_scissor_rect.width()
            && screen_coordinates.y() >= window_scissor_rect.y()
            && screen_coordinates.y() < window_scissor_rect.y() + window_scissor_rect.height();
    };

    // Calculate block-based bounds
    // clang-format off
    int const bx0 =  max(render_bounds.left(),   min(min(v0.x(), v1.x()), v2.x()) / subpixel_factor) & ~1;
    int const bx1 = (min(render_bounds.right(),  max(max(v0.x(), v1.x()), v2.x()) / subpixel_factor) & ~1) + 2;
    int const by0 =  max(render_bounds.top(),    min(min(v0.y(), v1.y()), v2.y()) / subpixel_factor) & ~1;
    int const by1 = (min(render_bounds.bottom(), max(max(v0.y(), v1.y()), v2.y()) / subpixel_factor) & ~1) + 2;
    // clang-format on

    // Fog depths
    float const vertex0_eye_absz = fabs(vertex0.eye_coordinates.z());
    float const vertex1_eye_absz = fabs(vertex1.eye_coordinates.z());
    float const vertex2_eye_absz = fabs(vertex2.eye_coordinates.z());

    // FIXME: implement stencil testing

    // Iterate over all blocks within the bounds of the triangle
    for (int by = by0; by < by1; by += 2) {
        for (int bx = bx0; bx < bx1; bx += 2) {

            PixelQuad quad;

            quad.screen_coordinates = {
                i32x4 { bx, bx + 1, bx, bx + 1 },
                i32x4 { by, by, by + 1, by + 1 },
            };

            auto edge_values = calculate_edge_values4(quad.screen_coordinates * subpixel_factor);

            // Generate triangle coverage mask
            quad.mask = test_point4(edge_values);
            if (options.scissor_enabled) {
                quad.mask &= test_scissor4(quad.screen_coordinates);
            }

            if (none(quad.mask))
                continue;

            INCREASE_STATISTICS_COUNTER(g_num_pixels, maskcount(quad.mask));

            // Calculate barycentric coordinates from previously calculated edge values
            quad.barycentrics = Vector3<f32x4> {
                to_f32x4(edge_values.x()),
                to_f32x4(edge_values.y()),
                to_f32x4(edge_values.z()),
            } * one_over_area;

            float* depth_ptrs[4] = {
                &depth_buffer.scanline(by)[bx],
                &depth_buffer.scanline(by)[bx + 1],
                &depth_buffer.scanline(by + 1)[bx],
                &depth_buffer.scanline(by + 1)[bx + 1],
            };

            // AND the depth mask onto the coverage mask
            if (options.enable_depth_test) {
                auto depth = load4_masked(depth_ptrs[0], depth_ptrs[1], depth_ptrs[2], depth_ptrs[3], quad.mask);

                quad.depth = interpolate(vertex0.window_coordinates.z(), vertex1.window_coordinates.z(), vertex2.window_coordinates.z(), quad.barycentrics);
                // FIXME: Also apply depth_offset_factor which depends on the depth gradient
                quad.depth += options.depth_offset_constant * NumericLimits<float>::epsilon();

                switch (options.depth_func) {
                case DepthTestFunction::Always:
                    break;
                case DepthTestFunction::Never:
                    quad.mask ^= quad.mask;
                    break;
                case DepthTestFunction::Greater:
                    quad.mask &= quad.depth > depth;
                    break;
                case DepthTestFunction::GreaterOrEqual:
                    quad.mask &= quad.depth >= depth;
                    break;
                case DepthTestFunction::NotEqual:
#ifdef __SSE__
                    quad.mask &= quad.depth != depth;
#else
                    quad.mask[0] = bit_cast<u32>(quad.depth[0]) != bit_cast<u32>(depth[0]) ? -1 : 0;
                    quad.mask[1] = bit_cast<u32>(quad.depth[1]) != bit_cast<u32>(depth[1]) ? -1 : 0;
                    quad.mask[2] = bit_cast<u32>(quad.depth[2]) != bit_cast<u32>(depth[2]) ? -1 : 0;
                    quad.mask[3] = bit_cast<u32>(quad.depth[3]) != bit_cast<u32>(depth[3]) ? -1 : 0;
#endif
                    break;
                case DepthTestFunction::Equal:
#ifdef __SSE__
                    quad.mask &= quad.depth == depth;
#else
                    //
                    // This is an interesting quirk that occurs due to us using the x87 FPU when Serenity is
                    // compiled for the i386 target. When we calculate our depth value to be stored in the buffer,
                    // it is an 80-bit x87 floating point number, however, when stored into the DepthBuffer, this is
                    // truncated to 32 bits. This 38 bit loss of precision means that when x87 `FCOMP` is eventually
                    // used here the comparison fails.
                    // This could be solved by using a `long double` for the depth buffer, however this would take
                    // up significantly more space and is completely overkill for a depth buffer. As such, comparing
                    // the first 32-bits of this depth value is "good enough" that if we get a hit on it being
                    // equal, we can pretty much guarantee that it's actually equal.
                    //
                    quad.mask[0] = bit_cast<u32>(quad.depth[0]) == bit_cast<u32>(depth[0]) ? -1 : 0;
                    quad.mask[1] = bit_cast<u32>(quad.depth[1]) == bit_cast<u32>(depth[1]) ? -1 : 0;
                    quad.mask[2] = bit_cast<u32>(quad.depth[2]) == bit_cast<u32>(depth[2]) ? -1 : 0;
                    quad.mask[3] = bit_cast<u32>(quad.depth[3]) == bit_cast<u32>(depth[3]) ? -1 : 0;
#endif
                    break;
                case DepthTestFunction::LessOrEqual:
                    quad.mask &= quad.depth <= depth;
                    break;
                case DepthTestFunction::Less:
                    quad.mask &= quad.depth < depth;
                    break;
                }

                // Nice, no pixels passed the depth test -> block rejected by early z
                if (none(quad.mask))
                    continue;
            }

            INCREASE_STATISTICS_COUNTER(g_num_pixels_shaded, maskcount(quad.mask));

            // Draw the pixels according to the previously generated mask
            auto const w_coordinates = Vector3<f32x4> {
                expand4(vertex0.window_coordinates.w()),
                expand4(vertex1.window_coordinates.w()),
                expand4(vertex2.window_coordinates.w()),
            };

            auto const interpolated_reciprocal_w = interpolate(w_coordinates.x(), w_coordinates.y(), w_coordinates.z(), quad.barycentrics);
            auto const interpolated_w = 1.0f / interpolated_reciprocal_w;
            quad.barycentrics = quad.barycentrics * w_coordinates * interpolated_w;

            // FIXME: make this more generic. We want to interpolate more than just color and uv
            if (options.shade_smooth) {
                quad.vertex_color = interpolate(expand4(vertex0.color), expand4(vertex1.color), expand4(vertex2.color), quad.barycentrics);
            } else {
                quad.vertex_color = expand4(vertex0.color);
            }

            quad.uv = interpolate(expand4(vertex0.tex_coord), expand4(vertex1.tex_coord), expand4(vertex2.tex_coord), quad.barycentrics);

            // Calculate depth of fragment for fog
            //
            // OpenGL 1.5 spec chapter 3.10: "An implementation may choose to approximate the
            // eye-coordinate distance from the eye to each fragment center by |Ze|."

            quad.fog_depth = interpolate(expand4(vertex0_eye_absz), expand4(vertex1_eye_absz), expand4(vertex2_eye_absz), quad.barycentrics);

            pixel_shader(quad);

            if (options.enable_alpha_test && options.alpha_test_func != AlphaTestFunction::Always) {
                switch (options.alpha_test_func) {
                case AlphaTestFunction::Less:
                    quad.mask &= quad.out_color.w() < options.alpha_test_ref_value;
                    break;
                case AlphaTestFunction::Equal:
                    quad.mask &= quad.out_color.w() == options.alpha_test_ref_value;
                    break;
                case AlphaTestFunction::LessOrEqual:
                    quad.mask &= quad.out_color.w() <= options.alpha_test_ref_value;
                    break;
                case AlphaTestFunction::Greater:
                    quad.mask &= quad.out_color.w() > options.alpha_test_ref_value;
                    break;
                case AlphaTestFunction::NotEqual:
                    quad.mask &= quad.out_color.w() != options.alpha_test_ref_value;
                    break;
                case AlphaTestFunction::GreaterOrEqual:
                    quad.mask &= quad.out_color.w() >= options.alpha_test_ref_value;
                    break;
                case AlphaTestFunction::Never:
                case AlphaTestFunction::Always:
                    VERIFY_NOT_REACHED();
                }
            }

            // Write to depth buffer
            if (options.enable_depth_test && options.enable_depth_write) {
                store4_masked(quad.depth, depth_ptrs[0], depth_ptrs[1], depth_ptrs[2], depth_ptrs[3], quad.mask);
            }

            // We will not update the color buffer at all
            if (!options.color_mask || !options.enable_color_write)
                continue;

            Gfx::RGBA32* color_ptrs[4] = {
                &render_target.scanline(by)[bx],
                &render_target.scanline(by)[bx + 1],
                &render_target.scanline(by + 1)[bx],
                &render_target.scanline(by + 1)[bx + 1],
            };

            int bits = maskbits(quad.mask);

            if (options.enable_blending) {
                INCREASE_STATISTICS_COUNTER(g_num_pixels_blended, maskcount(quad.mask));

                // Blend color values from pixel_staging into render_target
                FloatVector4 dst_aos[4] {
                    bits & 1 ? to_vec4(*color_ptrs[0]) : FloatVector4 { 0, 0, 0, 0 },
                    bits & 2 ? to_vec4(*color_ptrs[1]) : FloatVector4 { 0, 0, 0, 0 },
                    bits & 4 ? to_vec4(*color_ptrs[2]) : FloatVector4 { 0, 0, 0, 0 },
                    bits & 8 ? to_vec4(*color_ptrs[3]) : FloatVector4 { 0, 0, 0, 0 },
                };

                auto dst = Vector4<f32x4> {
                    f32x4 { dst_aos[0].x(), dst_aos[1].x(), dst_aos[2].x(), dst_aos[3].x() },
                    f32x4 { dst_aos[0].y(), dst_aos[1].y(), dst_aos[2].y(), dst_aos[3].y() },
                    f32x4 { dst_aos[0].z(), dst_aos[1].z(), dst_aos[2].z(), dst_aos[3].z() },
                    f32x4 { dst_aos[0].w(), dst_aos[1].w(), dst_aos[2].w(), dst_aos[3].w() },
                };
                Vector4<f32x4> const& src = quad.out_color;

                auto src_factor = expand4(src_constant)
                    + src * src_factor_src_color
                    + Vector4<f32x4> { src.w(), src.w(), src.w(), src.w() } * src_factor_src_alpha
                    + dst * src_factor_dst_color
                    + Vector4<f32x4> { dst.w(), dst.w(), dst.w(), dst.w() } * src_factor_dst_alpha;

                auto dst_factor = expand4(dst_constant)
                    + src * dst_factor_src_color
                    + Vector4<f32x4> { src.w(), src.w(), src.w(), src.w() } * dst_factor_src_alpha
                    + dst * dst_factor_dst_color
                    + Vector4<f32x4> { dst.w(), dst.w(), dst.w(), dst.w() } * dst_factor_dst_alpha;

                quad.out_color = src * src_factor + dst * dst_factor;
            }

            if (bits & 1)
                *color_ptrs[0] = to_rgba32(FloatVector4 { quad.out_color.x()[0], quad.out_color.y()[0], quad.out_color.z()[0], quad.out_color.w()[0] });
            if (bits & 2)
                *color_ptrs[1] = to_rgba32(FloatVector4 { quad.out_color.x()[1], quad.out_color.y()[1], quad.out_color.z()[1], quad.out_color.w()[1] });
            if (bits & 4)
                *color_ptrs[2] = to_rgba32(FloatVector4 { quad.out_color.x()[2], quad.out_color.y()[2], quad.out_color.z()[2], quad.out_color.w()[2] });
            if (bits & 8)
                *color_ptrs[3] = to_rgba32(FloatVector4 { quad.out_color.x()[3], quad.out_color.y()[3], quad.out_color.z()[3], quad.out_color.w()[3] });
        }
    }
}

static Gfx::IntSize closest_multiple(const Gfx::IntSize& min_size, size_t step)
{
    int width = ((min_size.width() + step - 1) / step) * step;
    int height = ((min_size.height() + step - 1) / step) * step;
    return { width, height };
}

Device::Device(const Gfx::IntSize& min_size)
    : m_render_target { Gfx::Bitmap::try_create(Gfx::BitmapFormat::BGRA8888, closest_multiple(min_size, 2)).release_value_but_fixme_should_propagate_errors() }
    , m_depth_buffer { adopt_own(*new DepthBuffer(closest_multiple(min_size, 2))) }
{
    m_options.scissor_box = m_render_target->rect();
}

DeviceInfo Device::info() const
{
    return {
        .vendor_name = "SerenityOS",
        .device_name = "SoftGPU",
        .num_texture_units = NUM_SAMPLERS
    };
}

static void generate_texture_coordinates(Vertex& vertex, RasterizerOptions const& options)
{
    auto generate_coordinate = [&](size_t config_index) -> float {
        auto mode = options.texcoord_generation_config[config_index].mode;

        switch (mode) {
        case TexCoordGenerationMode::ObjectLinear: {
            auto coefficients = options.texcoord_generation_config[config_index].coefficients;
            return coefficients.dot(vertex.position);
        }
        case TexCoordGenerationMode::EyeLinear: {
            auto coefficients = options.texcoord_generation_config[config_index].coefficients;
            return coefficients.dot(vertex.eye_coordinates);
        }
        case TexCoordGenerationMode::SphereMap: {
            auto const eye_unit = vertex.eye_coordinates.normalized();
            FloatVector3 const eye_unit_xyz = { eye_unit.x(), eye_unit.y(), eye_unit.z() };
            auto const normal = vertex.normal;
            auto reflection = eye_unit_xyz - normal * 2 * normal.dot(eye_unit_xyz);
            reflection.set_z(reflection.z() + 1);
            auto const reflection_value = (config_index == 0) ? reflection.x() : reflection.y();
            return reflection_value / (2 * reflection.length()) + 0.5f;
        }
        case TexCoordGenerationMode::ReflectionMap: {
            auto const eye_unit = vertex.eye_coordinates.normalized();
            FloatVector3 const eye_unit_xyz = { eye_unit.x(), eye_unit.y(), eye_unit.z() };
            auto const normal = vertex.normal;
            auto reflection = eye_unit_xyz - normal * 2 * normal.dot(eye_unit_xyz);
            switch (config_index) {
            case 0:
                return reflection.x();
            case 1:
                return reflection.y();
            case 2:
                return reflection.z();
            default:
                VERIFY_NOT_REACHED();
            }
        }
        case TexCoordGenerationMode::NormalMap: {
            auto const normal = vertex.normal;
            switch (config_index) {
            case 0:
                return normal.x();
            case 1:
                return normal.y();
            case 2:
                return normal.z();
            default:
                VERIFY_NOT_REACHED();
            }
        }
        default:
            VERIFY_NOT_REACHED();
        }
    };

    auto const enabled_coords = options.texcoord_generation_enabled_coordinates;
    vertex.tex_coord = {
        ((enabled_coords & TexCoordGenerationCoordinate::S) > 0) ? generate_coordinate(0) : vertex.tex_coord.x(),
        ((enabled_coords & TexCoordGenerationCoordinate::T) > 0) ? generate_coordinate(1) : vertex.tex_coord.y(),
        ((enabled_coords & TexCoordGenerationCoordinate::R) > 0) ? generate_coordinate(2) : vertex.tex_coord.z(),
        ((enabled_coords & TexCoordGenerationCoordinate::Q) > 0) ? generate_coordinate(3) : vertex.tex_coord.w(),
    };
}

void Device::draw_primitives(PrimitiveType primitive_type, FloatMatrix4x4 const& model_view_transform, FloatMatrix3x3 const& normal_transform,
    FloatMatrix4x4 const& projection_transform, FloatMatrix4x4 const& texture_transform, Vector<Vertex> const& vertices,
    Vector<size_t> const& enabled_texture_units)
{
    // At this point, the user has effectively specified that they are done with defining the geometry
    // of what they want to draw. We now need to do a few things (https://www.khronos.org/opengl/wiki/Rendering_Pipeline_Overview):
    //
    // 1.   Transform all of the vertices in the current vertex list into eye space by multiplying the model-view matrix
    // 2.   Transform all of the vertices from eye space into clip space by multiplying by the projection matrix
    // 3.   If culling is enabled, we cull the desired faces (https://learnopengl.com/Advanced-OpenGL/Face-culling)
    // 4.   Each element of the vertex is then divided by w to bring the positions into NDC (Normalized Device Coordinates)
    // 5.   The vertices are sorted (for the rasterizer, how are we doing this? 3Dfx did this top to bottom in terms of vertex y coordinates)
    // 6.   The vertices are then sent off to the rasterizer and drawn to the screen

    float scr_width = m_render_target->width();
    float scr_height = m_render_target->height();

    m_triangle_list.clear_with_capacity();
    m_processed_triangles.clear_with_capacity();

    // Let's construct some triangles
    if (primitive_type == PrimitiveType::Triangles) {
        Triangle triangle;
        for (size_t i = 0; i < vertices.size(); i += 3) {
            triangle.vertices[0] = vertices.at(i);
            triangle.vertices[1] = vertices.at(i + 1);
            triangle.vertices[2] = vertices.at(i + 2);

            m_triangle_list.append(triangle);
        }
    } else if (primitive_type == PrimitiveType::Quads) {
        // We need to construct two triangles to form the quad
        Triangle triangle;
        VERIFY(vertices.size() % 4 == 0);
        for (size_t i = 0; i < vertices.size(); i += 4) {
            // Triangle 1
            triangle.vertices[0] = vertices.at(i);
            triangle.vertices[1] = vertices.at(i + 1);
            triangle.vertices[2] = vertices.at(i + 2);
            m_triangle_list.append(triangle);

            // Triangle 2
            triangle.vertices[0] = vertices.at(i + 2);
            triangle.vertices[1] = vertices.at(i + 3);
            triangle.vertices[2] = vertices.at(i);
            m_triangle_list.append(triangle);
        }
    } else if (primitive_type == PrimitiveType::TriangleFan) {
        Triangle triangle;
        triangle.vertices[0] = vertices.at(0); // Root vertex is always the vertex defined first

        for (size_t i = 1; i < vertices.size() - 1; i++) // This is technically `n-2` triangles. We start at index 1
        {
            triangle.vertices[1] = vertices.at(i);
            triangle.vertices[2] = vertices.at(i + 1);
            m_triangle_list.append(triangle);
        }
    } else if (primitive_type == PrimitiveType::TriangleStrip) {
        Triangle triangle;
        for (size_t i = 0; i < vertices.size() - 2; i++) {
            if (i % 2 == 0) {
                triangle.vertices[0] = vertices.at(i);
                triangle.vertices[1] = vertices.at(i + 1);
                triangle.vertices[2] = vertices.at(i + 2);
            } else {
                triangle.vertices[0] = vertices.at(i + 1);
                triangle.vertices[1] = vertices.at(i);
                triangle.vertices[2] = vertices.at(i + 2);
            }
            m_triangle_list.append(triangle);
        }
    }

    // Now let's transform each triangle and send that to the GPU
    auto const depth_half_range = (m_options.depth_max - m_options.depth_min) / 2;
    auto const depth_halfway = (m_options.depth_min + m_options.depth_max) / 2;
    for (auto& triangle : m_triangle_list) {
        // Transform vertices into eye coordinates using the model-view transform
        triangle.vertices[0].eye_coordinates = model_view_transform * triangle.vertices[0].position;
        triangle.vertices[1].eye_coordinates = model_view_transform * triangle.vertices[1].position;
        triangle.vertices[2].eye_coordinates = model_view_transform * triangle.vertices[2].position;

        // Transform eye coordinates into clip coordinates using the projection transform
        triangle.vertices[0].clip_coordinates = projection_transform * triangle.vertices[0].eye_coordinates;
        triangle.vertices[1].clip_coordinates = projection_transform * triangle.vertices[1].eye_coordinates;
        triangle.vertices[2].clip_coordinates = projection_transform * triangle.vertices[2].eye_coordinates;

        // At this point, we're in clip space
        // Here's where we do the clipping. This is a really crude implementation of the
        // https://learnopengl.com/Getting-started/Coordinate-Systems
        // "Note that if only a part of a primitive e.g. a triangle is outside the clipping volume OpenGL
        // will reconstruct the triangle as one or more triangles to fit inside the clipping range. "
        //
        // ALL VERTICES ARE DEFINED IN A CLOCKWISE ORDER

        // Okay, let's do some face culling first

        m_clipped_vertices.clear_with_capacity();
        m_clipped_vertices.append(triangle.vertices[0]);
        m_clipped_vertices.append(triangle.vertices[1]);
        m_clipped_vertices.append(triangle.vertices[2]);
        m_clipper.clip_triangle_against_frustum(m_clipped_vertices);

        if (m_clipped_vertices.size() < 3)
            continue;

        for (auto& vec : m_clipped_vertices) {
            // To normalized device coordinates (NDC)
            auto const one_over_w = 1 / vec.clip_coordinates.w();
            auto const ndc_coordinates = FloatVector4 {
                vec.clip_coordinates.x() * one_over_w,
                vec.clip_coordinates.y() * one_over_w,
                vec.clip_coordinates.z() * one_over_w,
                one_over_w,
            };

            // To window coordinates
            // FIXME: implement viewport functionality
            vec.window_coordinates = {
                scr_width / 2 + ndc_coordinates.x() * scr_width / 2,
                scr_height / 2 - ndc_coordinates.y() * scr_height / 2,
                depth_half_range * ndc_coordinates.z() + depth_halfway,
                ndc_coordinates.w(),
            };
        }

        Triangle tri;
        tri.vertices[0] = m_clipped_vertices[0];
        for (size_t i = 1; i < m_clipped_vertices.size() - 1; i++) {
            tri.vertices[1] = m_clipped_vertices[i];
            tri.vertices[2] = m_clipped_vertices[i + 1];
            m_processed_triangles.append(tri);
        }
    }

    for (auto& triangle : m_processed_triangles) {
        // Let's calculate the (signed) area of the triangle
        // https://cp-algorithms.com/geometry/oriented-triangle-area.html
        float dxAB = triangle.vertices[0].window_coordinates.x() - triangle.vertices[1].window_coordinates.x(); // A.x - B.x
        float dxBC = triangle.vertices[1].window_coordinates.x() - triangle.vertices[2].window_coordinates.x(); // B.X - C.x
        float dyAB = triangle.vertices[0].window_coordinates.y() - triangle.vertices[1].window_coordinates.y();
        float dyBC = triangle.vertices[1].window_coordinates.y() - triangle.vertices[2].window_coordinates.y();
        float area = (dxAB * dyBC) - (dxBC * dyAB);

        if (area == 0.0f)
            continue;

        if (m_options.enable_culling) {
            bool is_front = (m_options.front_face == WindingOrder::CounterClockwise ? area < 0 : area > 0);

            if (!is_front && m_options.cull_back)
                continue;

            if (is_front && m_options.cull_front)
                continue;
        }

        if (area > 0)
            swap(triangle.vertices[0], triangle.vertices[1]);

        // Transform normals
        triangle.vertices[0].normal = normal_transform * triangle.vertices[0].normal;
        triangle.vertices[1].normal = normal_transform * triangle.vertices[1].normal;
        triangle.vertices[2].normal = normal_transform * triangle.vertices[2].normal;
        if (m_options.normalization_enabled) {
            triangle.vertices[0].normal.normalize();
            triangle.vertices[1].normal.normalize();
            triangle.vertices[2].normal.normalize();
        }

        // Generate texture coordinates if at least one coordinate is enabled
        if (m_options.texcoord_generation_enabled_coordinates != TexCoordGenerationCoordinate::None) {
            generate_texture_coordinates(triangle.vertices[0], m_options);
            generate_texture_coordinates(triangle.vertices[1], m_options);
            generate_texture_coordinates(triangle.vertices[2], m_options);
        }

        // Apply texture transformation
        // FIXME: implement multi-texturing: texcoords should be stored per texture unit
        triangle.vertices[0].tex_coord = texture_transform * triangle.vertices[0].tex_coord;
        triangle.vertices[1].tex_coord = texture_transform * triangle.vertices[1].tex_coord;
        triangle.vertices[2].tex_coord = texture_transform * triangle.vertices[2].tex_coord;

        submit_triangle(triangle, enabled_texture_units);
    }
}

void Device::submit_triangle(const Triangle& triangle, Vector<size_t> const& enabled_texture_units)
{
    rasterize_triangle(m_options, *m_render_target, *m_depth_buffer, triangle, [this, &enabled_texture_units](PixelQuad& quad) {
        quad.out_color = quad.vertex_color;

        for (size_t i : enabled_texture_units) {
            // FIXME: implement GL_TEXTURE_1D, GL_TEXTURE_3D and GL_TEXTURE_CUBE_MAP
            auto const& sampler = m_samplers[i];

            auto texel = sampler.sample_2d({ quad.uv.x(), quad.uv.y() });
            INCREASE_STATISTICS_COUNTER(g_num_sampler_calls, 1);

            // FIXME: Implement more blend modes
            switch (sampler.config().fixed_function_texture_env_mode) {
            case TextureEnvMode::Modulate:
                quad.out_color = quad.out_color * texel;
                break;
            case TextureEnvMode::Replace:
                quad.out_color = texel;
                break;
            case TextureEnvMode::Decal: {
                auto src_alpha = quad.out_color.w();
                quad.out_color.set_x(mix(quad.out_color.x(), texel.x(), src_alpha));
                quad.out_color.set_y(mix(quad.out_color.y(), texel.y(), src_alpha));
                quad.out_color.set_z(mix(quad.out_color.z(), texel.z(), src_alpha));
                break;
            }
            default:
                VERIFY_NOT_REACHED();
            }
        }

        // Calculate fog
        // Math from here: https://opengl-notes.readthedocs.io/en/latest/topics/texturing/aliasing.html

        // FIXME: exponential fog is not vectorized, we should add a SIMD exp function that calculates an approximation.
        if (m_options.fog_enabled) {
            auto factor = expand4(0.0f);
            switch (m_options.fog_mode) {
            case FogMode::Linear:
                factor = (m_options.fog_end - quad.fog_depth) / (m_options.fog_end - m_options.fog_start);
                break;
            case FogMode::Exp: {
                auto argument = -m_options.fog_density * quad.fog_depth;
                factor = exp(argument);
            } break;
            case FogMode::Exp2: {
                auto argument = m_options.fog_density * quad.fog_depth;
                argument *= -argument;
                factor = exp(argument);
            } break;
            default:
                VERIFY_NOT_REACHED();
            }

            // Mix texel's RGB with fog's RBG - leave alpha alone
            auto fog_color = expand4(m_options.fog_color);
            quad.out_color.set_x(mix(fog_color.x(), quad.out_color.x(), factor));
            quad.out_color.set_y(mix(fog_color.y(), quad.out_color.y(), factor));
            quad.out_color.set_z(mix(fog_color.z(), quad.out_color.z(), factor));
        }
    });
}

void Device::resize(const Gfx::IntSize& min_size)
{
    wait_for_all_threads();

    m_render_target = Gfx::Bitmap::try_create(Gfx::BitmapFormat::BGRA8888, closest_multiple(min_size, 2)).release_value_but_fixme_should_propagate_errors();
    m_depth_buffer = adopt_own(*new DepthBuffer(m_render_target->size()));
}

void Device::clear_color(const FloatVector4& color)
{
    wait_for_all_threads();

    uint8_t r = static_cast<uint8_t>(clamp(color.x(), 0.0f, 1.0f) * 255);
    uint8_t g = static_cast<uint8_t>(clamp(color.y(), 0.0f, 1.0f) * 255);
    uint8_t b = static_cast<uint8_t>(clamp(color.z(), 0.0f, 1.0f) * 255);
    uint8_t a = static_cast<uint8_t>(clamp(color.w(), 0.0f, 1.0f) * 255);
    auto const fill_color = Gfx::Color(r, g, b, a);

    if (m_options.scissor_enabled) {
        auto fill_rect = m_render_target->rect();
        fill_rect.intersect(scissor_box_to_window_coordinates(m_options.scissor_box, fill_rect));
        Gfx::Painter painter { *m_render_target };
        painter.fill_rect(fill_rect, fill_color);
        return;
    }

    m_render_target->fill(fill_color);
}

void Device::clear_depth(float depth)
{
    wait_for_all_threads();

    if (m_options.scissor_enabled) {
        m_depth_buffer->clear(scissor_box_to_window_coordinates(m_options.scissor_box, m_render_target->rect()), depth);
        return;
    }

    m_depth_buffer->clear(depth);
}

void Device::blit(Gfx::Bitmap const& source, int x, int y)
{
    wait_for_all_threads();

    INCREASE_STATISTICS_COUNTER(g_num_pixels, source.width() * source.height());
    INCREASE_STATISTICS_COUNTER(g_num_pixels_shaded, source.width() * source.height());

    Gfx::Painter painter { *m_render_target };
    painter.blit({ x, y }, source, source.rect(), 1.0f, true);
}

void Device::blit_to(Gfx::Bitmap& target)
{
    wait_for_all_threads();

    Gfx::Painter painter { target };
    painter.blit({ 0, 0 }, *m_render_target, m_render_target->rect(), 1.0f, false);

    if constexpr (ENABLE_STATISTICS_OVERLAY)
        draw_statistics_overlay(target);
}

void Device::draw_statistics_overlay(Gfx::Bitmap& target)
{
    static Core::ElapsedTimer timer;
    static String debug_string;
    static int frame_counter;

    frame_counter++;
    int milliseconds = 0;
    if (timer.is_valid())
        milliseconds = timer.elapsed();
    else
        timer.start();

    Gfx::Painter painter { target };

    if (milliseconds > 500) {

        if (g_num_pixels == 0)
            g_num_pixels = 1;

        int num_rendertarget_pixels = m_render_target->width() * m_render_target->height();

        StringBuilder builder;
        builder.append(String::formatted("Timings      : {:.1}ms {:.1}FPS\n",
            static_cast<double>(milliseconds) / frame_counter,
            (milliseconds > 0) ? 1000.0 * frame_counter / milliseconds : 9999.0));
        builder.append(String::formatted("Triangles    : {}\n", g_num_rasterized_triangles));
        builder.append(String::formatted("Pixels       : {}, Shaded: {}%, Blended: {}%, Overdraw: {}%\n",
            g_num_pixels,
            g_num_pixels_shaded * 100 / g_num_pixels,
            g_num_pixels_blended * 100 / g_num_pixels_shaded,
            g_num_pixels_shaded * 100 / num_rendertarget_pixels - 100));
        builder.append(String::formatted("Sampler calls: {}\n", g_num_sampler_calls));

        debug_string = builder.to_string();

        frame_counter = 0;
        timer.start();
    }

    g_num_rasterized_triangles = 0;
    g_num_pixels = 0;
    g_num_pixels_shaded = 0;
    g_num_pixels_blended = 0;
    g_num_sampler_calls = 0;

    auto& font = Gfx::FontDatabase::default_fixed_width_font();

    for (int y = -1; y < 2; y++)
        for (int x = -1; x < 2; x++)
            if (x != 0 && y != 0)
                painter.draw_text(target.rect().translated(x + 2, y + 2), debug_string, font, Gfx::TextAlignment::TopLeft, Gfx::Color::Black);

    painter.draw_text(target.rect().translated(2, 2), debug_string, font, Gfx::TextAlignment::TopLeft, Gfx::Color::White);
}

void Device::wait_for_all_threads() const
{
    // FIXME: Wait for all render threads to finish when multithreading is being implemented
}

void Device::set_options(const RasterizerOptions& options)
{
    wait_for_all_threads();

    m_options = options;

    // FIXME: Recreate or reinitialize render threads here when multithreading is being implemented
}

Gfx::RGBA32 Device::get_backbuffer_pixel(int x, int y)
{
    // FIXME: Reading individual pixels is very slow, rewrite this to transfer whole blocks
    if (x < 0 || y < 0 || x >= m_render_target->width() || y >= m_render_target->height())
        return 0;

    return m_render_target->scanline(y)[x];
}

float Device::get_depthbuffer_value(int x, int y)
{
    // FIXME: Reading individual pixels is very slow, rewrite this to transfer whole blocks
    if (x < 0 || y < 0 || x >= m_render_target->width() || y >= m_render_target->height())
        return 1.0f;

    return m_depth_buffer->scanline(y)[x];
}

NonnullRefPtr<Image> Device::create_image(ImageFormat format, unsigned width, unsigned height, unsigned depth, unsigned levels, unsigned layers)
{
    VERIFY(width > 0);
    VERIFY(height > 0);
    VERIFY(depth > 0);
    VERIFY(levels > 0);
    VERIFY(layers > 0);

    return adopt_ref(*new Image(format, width, height, depth, levels, layers));
}

void Device::set_sampler_config(unsigned sampler, SamplerConfig const& config)
{
    m_samplers[sampler].set_config(config);
}

}