citra/src/video_core/swrasterizer/rasterizer.cpp

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

#include <algorithm>
#include <array>
#include <cmath>
#include <tuple>
#include "common/assert.h"
#include "common/bit_field.h"
#include "common/color.h"
#include "common/common_types.h"
#include "common/logging/log.h"
#include "common/math_util.h"
#include "common/microprofile.h"
#include "common/quaternion.h"
#include "common/vector_math.h"
#include "core/hw/gpu.h"
#include "core/memory.h"
#include "video_core/debug_utils/debug_utils.h"
#include "video_core/pica_state.h"
#include "video_core/pica_types.h"
#include "video_core/regs_framebuffer.h"
#include "video_core/regs_rasterizer.h"
#include "video_core/regs_texturing.h"
#include "video_core/shader/shader.h"
#include "video_core/swrasterizer/framebuffer.h"
#include "video_core/swrasterizer/proctex.h"
#include "video_core/swrasterizer/rasterizer.h"
#include "video_core/swrasterizer/texturing.h"
#include "video_core/texture/texture_decode.h"
#include "video_core/utils.h"

namespace Pica {
namespace Rasterizer {

// NOTE: Assuming that rasterizer coordinates are 12.4 fixed-point values
struct Fix12P4 {
    Fix12P4() {}
    Fix12P4(u16 val) : val(val) {}

    static u16 FracMask() {
        return 0xF;
    }
    static u16 IntMask() {
        return (u16)~0xF;
    }

    operator u16() const {
        return val;
    }

    bool operator<(const Fix12P4& oth) const {
        return (u16) * this < (u16)oth;
    }

private:
    u16 val;
};

/**
 * Calculate signed area of the triangle spanned by the three argument vertices.
 * The sign denotes an orientation.
 *
 * @todo define orientation concretely.
 */
static int SignedArea(const Math::Vec2<Fix12P4>& vtx1, const Math::Vec2<Fix12P4>& vtx2,
                      const Math::Vec2<Fix12P4>& vtx3) {
    const auto vec1 = Math::MakeVec(vtx2 - vtx1, 0);
    const auto vec2 = Math::MakeVec(vtx3 - vtx1, 0);
    // TODO: There is a very small chance this will overflow for sizeof(int) == 4
    return Math::Cross(vec1, vec2).z;
};

/// Convert a 3D vector for cube map coordinates to 2D texture coordinates along with the face name
static std::tuple<float24, float24, PAddr> ConvertCubeCoord(float24 u, float24 v, float24 w,
                                                            const TexturingRegs& regs) {
    const float abs_u = std::abs(u.ToFloat32());
    const float abs_v = std::abs(v.ToFloat32());
    const float abs_w = std::abs(w.ToFloat32());
    float24 x, y, z;
    PAddr addr;
    if (abs_u > abs_v && abs_u > abs_w) {
        if (u > float24::FromFloat32(0)) {
            addr = regs.GetCubePhysicalAddress(TexturingRegs::CubeFace::PositiveX);
            y = -v;
        } else {
            addr = regs.GetCubePhysicalAddress(TexturingRegs::CubeFace::NegativeX);
            y = v;
        }
        x = -w;
        z = u;
    } else if (abs_v > abs_w) {
        if (v > float24::FromFloat32(0)) {
            addr = regs.GetCubePhysicalAddress(TexturingRegs::CubeFace::PositiveY);
            x = u;
        } else {
            addr = regs.GetCubePhysicalAddress(TexturingRegs::CubeFace::NegativeY);
            x = -u;
        }
        y = w;
        z = v;
    } else {
        if (w > float24::FromFloat32(0)) {
            addr = regs.GetCubePhysicalAddress(TexturingRegs::CubeFace::PositiveZ);
            y = -v;
        } else {
            addr = regs.GetCubePhysicalAddress(TexturingRegs::CubeFace::NegativeZ);
            y = v;
        }
        x = u;
        z = w;
    }
    const float24 half = float24::FromFloat32(0.5f);
    return std::make_tuple(x / z * half + half, y / z * half + half, addr);
}


float LookupLightingLut(size_t lut_index, float index) {
    unsigned index_i = static_cast<unsigned>(MathUtil::Clamp(floor(index * 256), 0.0f, 1.0f));

    float index_f = index - index_i;

    ASSERT_MSG(lut_index < g_state.lighting.luts.size(), "Out of range lut");

    float lut_value = g_state.lighting.luts[lut_index][index_i].ToFloat();
    float lut_diff = g_state.lighting.luts[lut_index][index_i].DiffToFloat();

    return lut_value + lut_diff * index_f;
}

std::tuple<Math::Vec4<u8>, Math::Vec4<u8>> ComputeFragmentsColors(const Math::Quaternion<float>& normquat, const Math::Vec3<float>& view) {
    const auto& lighting = g_state.regs.lighting;

    if (lighting.disable)
        return {{}, {}};

    // TODO(Subv): Bump mapping
    Math::Vec3<float> surface_normal = {0.0f, 0.0f, 1.0f};

    if (lighting.config0.bump_mode != LightingRegs::LightingBumpMode::None) {
        LOG_CRITICAL(HW_GPU, "unimplemented bump mapping");
        UNIMPLEMENTED();
    }

    // TODO(Subv): Do we need to normalize the quaternion here?
    auto normal = Math::QuaternionRotate(normquat, surface_normal);

    Math::Vec3<float> light_vector = {};
    Math::Vec4<float> diffuse_sum = {0.f, 0.f, 0.f, 1.f};
    Math::Vec4<float> specular_sum = {0.f, 0.f, 0.f, 1.f};
    Math::Vec3<float> refl_value = {};

    for (unsigned light_index = 0; light_index <= lighting.max_light_index; ++light_index) {
        unsigned num = lighting.light_enable.GetNum(light_index);
        const auto& light_config = g_state.regs.lighting.light[num];

        Math::Vec3<float> position = {float16::FromRaw(light_config.x).ToFloat32(), float16::FromRaw(light_config.y).ToFloat32(), float16::FromRaw(light_config.z).ToFloat32()};

        if (light_config.config.directional)
            light_vector = position;
        else
            light_vector = position + view;

        light_vector.Normalize();

        auto LV_N = Math::Dot(light_vector, normal);
        auto dot_product = LV_N;

        if (light_config.config.two_sided_diffuse)
            dot_product = std::abs(dot_product);
        else
            dot_product = std::max(dot_product, 0.0f);

        float dist_atten = 1.0f;
        if (!lighting.IsDistAttenDisabled(num)) {
            auto distance = (-view - position).Length();
            float scale = Pica::float20::FromRaw(light_config.dist_atten_scale).ToFloat32();
            float bias = Pica::float20::FromRaw(light_config.dist_atten_scale).ToFloat32();
            size_t lut = static_cast<size_t>(LightingRegs::LightingSampler::DistanceAttenuation) + num;

            float sample_loc = scale * distance + bias;
            dist_atten = LookupLightingLut(lut, sample_loc);
        }

        float clamp_highlights = 1.0f;

        if (lighting.config0.clamp_highlights) {
            if (LV_N <= 0.f)
                clamp_highlights = 0.f;
            else
                clamp_highlights = 1.f;
        }

        auto GetLutIndex = [&](unsigned num, LightingRegs::LightingLutInput input,
                              bool abs) -> float {

            Math::Vec3<float> norm_view = view.Normalized();
            Math::Vec3<float> half_angle = (norm_view + light_vector).Normalized();
            float result = 0.0f;

            switch (input) {
            case LightingRegs::LightingLutInput::NH:
                result = Math::Dot(normal, half_angle);
                break;

            case LightingRegs::LightingLutInput::VH:
                result = Math::Dot(norm_view, half_angle);
                break;

            case LightingRegs::LightingLutInput::NV:
                result = Math::Dot(normal, norm_view);
                break;

            case LightingRegs::LightingLutInput::LN:
                result = Math::Dot(light_vector, normal);
                break;

            default:
                LOG_CRITICAL(HW_GPU, "Unknown lighting LUT input %d\n", (int)input);
                UNIMPLEMENTED();
                result = 0.f;
            }

            if (abs) {
                if (light_config.config.two_sided_diffuse)
                    result = std::abs(result);
                else
                    result = std::max(result, 0.0f);
            } else {
                if (result < 0.f)
                    result += 2.f;

                result /= 2.f;
            }

            return MathUtil::Clamp(result, 0.0f, 1.0f);
        };

        // Specular 0 component
        float d0_lut_value = 1.0f;
        if (lighting.config1.disable_lut_d0 == 0 &&
            LightingRegs::IsLightingSamplerSupported(
                lighting.config0.config, LightingRegs::LightingSampler::Distribution0)) {

            // Lookup specular "distribution 0" LUT value
            float index = GetLutIndex(num, lighting.lut_input.d0.Value(), lighting.abs_lut_input.disable_d0 == 0);

            float scale = lighting.lut_scale.GetScale(lighting.lut_scale.d0);

            d0_lut_value = scale * LookupLightingLut(static_cast<size_t>(LightingRegs::LightingSampler::Distribution0), index);
        }

        Math::Vec3<float> specular_0 = d0_lut_value * light_config.specular_0.ToVec3f();

        // If enabled, lookup ReflectRed value, otherwise, 1.0 is used
        if (lighting.config1.disable_lut_rr == 0 &&
            LightingRegs::IsLightingSamplerSupported(lighting.config0.config,
                                                     LightingRegs::LightingSampler::ReflectRed)) {

            float index = GetLutIndex(num, lighting.lut_input.rr, lighting.abs_lut_input.disable_rr == 0);

            float scale = lighting.lut_scale.GetScale(lighting.lut_scale.rr);

            refl_value.x = scale * LookupLightingLut(static_cast<size_t>(LightingRegs::LightingSampler::ReflectRed), index);
        } else {
            refl_value.x = 1.0f;
        }

        // If enabled, lookup ReflectGreen value, otherwise, ReflectRed value is used
        if (lighting.config1.disable_lut_rg == 0 &&
            LightingRegs::IsLightingSamplerSupported(lighting.config0.config,
                                                     LightingRegs::LightingSampler::ReflectGreen)) {

            float index = GetLutIndex(num, lighting.lut_input.rg, lighting.abs_lut_input.disable_rg == 0);

            float scale = lighting.lut_scale.GetScale(lighting.lut_scale.rg);

            refl_value.y = scale * LookupLightingLut(static_cast<size_t>(LightingRegs::LightingSampler::ReflectGreen), index);
        } else {
            refl_value.y = refl_value.x;
        }

        // If enabled, lookup ReflectBlue value, otherwise, ReflectRed value is used
        if (lighting.config1.disable_lut_rb == 0 &&
            LightingRegs::IsLightingSamplerSupported(lighting.config0.config,
                                                     LightingRegs::LightingSampler::ReflectBlue)) {

            float index = GetLutIndex(num, lighting.lut_input.rb, lighting.abs_lut_input.disable_rb == 0);

            float scale = lighting.lut_scale.GetScale(lighting.lut_scale.rb);

            refl_value.z = scale * LookupLightingLut(static_cast<size_t>(LightingRegs::LightingSampler::ReflectBlue), index);
        } else {
            refl_value.z = refl_value.x;
        }

        float d1_lut_value = 1.0f;
        if (lighting.config1.disable_lut_d1 == 0 &&
            LightingRegs::IsLightingSamplerSupported(
                lighting.config0.config, LightingRegs::LightingSampler::Distribution1)) {

            // Lookup specular "distribution 1" LUT value
            float index = GetLutIndex(num, lighting.lut_input.d1.Value(), lighting.abs_lut_input.disable_d1 == 0);

            float scale = lighting.lut_scale.GetScale(lighting.lut_scale.d1);

            d1_lut_value = scale * LookupLightingLut(static_cast<size_t>(LightingRegs::LightingSampler::Distribution1), index);
        }

        Math::Vec3<float> specular_1 = d1_lut_value * refl_value * light_config.specular_1.ToVec3f();

        if (lighting.config1.disable_lut_fr == 0 &&
            LightingRegs::IsLightingSamplerSupported(
                lighting.config0.config, LightingRegs::LightingSampler::Fresnel)) {

            // Lookup fresnel LUT value
            float index = GetLutIndex(num, lighting.lut_input.fr.Value(), lighting.abs_lut_input.disable_fr == 0);

            float scale = lighting.lut_scale.GetScale(lighting.lut_scale.fr);

            float lut_value = scale * LookupLightingLut(static_cast<size_t>(LightingRegs::LightingSampler::Fresnel), index);

            // Enabled for difffuse lighting alpha component
            if (lighting.config0.fresnel_selector == LightingRegs::LightingFresnelSelector::PrimaryAlpha ||
                lighting.config0.fresnel_selector == LightingRegs::LightingFresnelSelector::Both) {
                diffuse_sum.a() *= lut_value;
            }

            // Enabled for the specular lighting alpha component
            if (lighting.config0.fresnel_selector ==
                LightingRegs::LightingFresnelSelector::SecondaryAlpha ||
                lighting.config0.fresnel_selector == LightingRegs::LightingFresnelSelector::Both) {
                specular_sum.a() *= lut_value;
            }
        }


        auto diffuse = light_config.diffuse.ToVec3f() * dot_product + light_config.ambient.ToVec3f();
        diffuse_sum += Math::MakeVec(diffuse * dist_atten, 0.0f);

        specular_sum += Math::MakeVec((specular_0 + specular_1) * clamp_highlights * dist_atten, 0.f);
    }

    diffuse_sum += Math::MakeVec(lighting.global_ambient.ToVec3f(), 0.0f);
    return {
        Math::MakeVec<float>(MathUtil::Clamp(diffuse_sum.x, 0.0f, 1.0f) * 255, MathUtil::Clamp(diffuse_sum.y, 0.0f, 1.0f) * 255, MathUtil::Clamp(diffuse_sum.z, 0.0f, 1.0f) * 255, MathUtil::Clamp(diffuse_sum.w, 0.0f, 1.0f) * 255).Cast<u8>(),
        Math::MakeVec<float>(MathUtil::Clamp(specular_sum.x, 0.0f, 1.0f) * 255, MathUtil::Clamp(specular_sum.y, 0.0f, 1.0f) * 255, MathUtil::Clamp(specular_sum.z, 0.0f, 1.0f) * 255, MathUtil::Clamp(specular_sum.w, 0.0f, 1.0f) * 255).Cast<u8>()
    };
}

static bool AreQuaternionsOpposite(Math::Vec4<Pica::float24> qa, Math::Vec4<Pica::float24> qb) {
    Math::Vec4f a{ qa.x.ToFloat32(), qa.y.ToFloat32(), qa.z.ToFloat32(), qa.w.ToFloat32() };
    Math::Vec4f b{ qb.x.ToFloat32(), qb.y.ToFloat32(), qb.z.ToFloat32(), qb.w.ToFloat32() };

    return (Math::Dot(a, b) < 0.f);
}

MICROPROFILE_DEFINE(GPU_Rasterization, "GPU", "Rasterization", MP_RGB(50, 50, 240));

/**
 * Helper function for ProcessTriangle with the "reversed" flag to allow for implementing
 * culling via recursion.
 */
static void ProcessTriangleInternal(const Vertex& v0, const Vertex& v1, const Vertex& v2,
                                    bool reversed = false) {
    const auto& regs = g_state.regs;
    MICROPROFILE_SCOPE(GPU_Rasterization);

    // vertex positions in rasterizer coordinates
    static auto FloatToFix = [](float24 flt) {
        // TODO: Rounding here is necessary to prevent garbage pixels at
        //       triangle borders. Is it that the correct solution, though?
        return Fix12P4(static_cast<unsigned short>(round(flt.ToFloat32() * 16.0f)));
    };
    static auto ScreenToRasterizerCoordinates = [](const Math::Vec3<float24>& vec) {
        return Math::Vec3<Fix12P4>{FloatToFix(vec.x), FloatToFix(vec.y), FloatToFix(vec.z)};
    };

    Math::Vec3<Fix12P4> vtxpos[3]{ScreenToRasterizerCoordinates(v0.screenpos),
                                  ScreenToRasterizerCoordinates(v1.screenpos),
                                  ScreenToRasterizerCoordinates(v2.screenpos)};

    if (regs.rasterizer.cull_mode == RasterizerRegs::CullMode::KeepAll) {
        // Make sure we always end up with a triangle wound counter-clockwise
        if (!reversed && SignedArea(vtxpos[0].xy(), vtxpos[1].xy(), vtxpos[2].xy()) <= 0) {
            ProcessTriangleInternal(v0, v2, v1, true);
            return;
        }
    } else {
        if (!reversed && regs.rasterizer.cull_mode == RasterizerRegs::CullMode::KeepClockWise) {
            // Reverse vertex order and use the CCW code path.
            ProcessTriangleInternal(v0, v2, v1, true);
            return;
        }

        // Cull away triangles which are wound clockwise.
        if (SignedArea(vtxpos[0].xy(), vtxpos[1].xy(), vtxpos[2].xy()) <= 0)
            return;
    }

    u16 min_x = std::min({vtxpos[0].x, vtxpos[1].x, vtxpos[2].x});
    u16 min_y = std::min({vtxpos[0].y, vtxpos[1].y, vtxpos[2].y});
    u16 max_x = std::max({vtxpos[0].x, vtxpos[1].x, vtxpos[2].x});
    u16 max_y = std::max({vtxpos[0].y, vtxpos[1].y, vtxpos[2].y});

    // Convert the scissor box coordinates to 12.4 fixed point
    u16 scissor_x1 = (u16)(regs.rasterizer.scissor_test.x1 << 4);
    u16 scissor_y1 = (u16)(regs.rasterizer.scissor_test.y1 << 4);
    // x2,y2 have +1 added to cover the entire sub-pixel area
    u16 scissor_x2 = (u16)((regs.rasterizer.scissor_test.x2 + 1) << 4);
    u16 scissor_y2 = (u16)((regs.rasterizer.scissor_test.y2 + 1) << 4);

    if (regs.rasterizer.scissor_test.mode == RasterizerRegs::ScissorMode::Include) {
        // Calculate the new bounds
        min_x = std::max(min_x, scissor_x1);
        min_y = std::max(min_y, scissor_y1);
        max_x = std::min(max_x, scissor_x2);
        max_y = std::min(max_y, scissor_y2);
    }

    min_x &= Fix12P4::IntMask();
    min_y &= Fix12P4::IntMask();
    max_x = ((max_x + Fix12P4::FracMask()) & Fix12P4::IntMask());
    max_y = ((max_y + Fix12P4::FracMask()) & Fix12P4::IntMask());

    // Triangle filling rules: Pixels on the right-sided edge or on flat bottom edges are not
    // drawn. Pixels on any other triangle border are drawn. This is implemented with three bias
    // values which are added to the barycentric coordinates w0, w1 and w2, respectively.
    // NOTE: These are the PSP filling rules. Not sure if the 3DS uses the same ones...
    auto IsRightSideOrFlatBottomEdge = [](const Math::Vec2<Fix12P4>& vtx,
                                          const Math::Vec2<Fix12P4>& line1,
                                          const Math::Vec2<Fix12P4>& line2) {
        if (line1.y == line2.y) {
            // just check if vertex is above us => bottom line parallel to x-axis
            return vtx.y < line1.y;
        } else {
            // check if vertex is on our left => right side
            // TODO: Not sure how likely this is to overflow
            return (int)vtx.x < (int)line1.x +
                                    ((int)line2.x - (int)line1.x) * ((int)vtx.y - (int)line1.y) /
                                        ((int)line2.y - (int)line1.y);
        }
    };
    int bias0 =
        IsRightSideOrFlatBottomEdge(vtxpos[0].xy(), vtxpos[1].xy(), vtxpos[2].xy()) ? -1 : 0;
    int bias1 =
        IsRightSideOrFlatBottomEdge(vtxpos[1].xy(), vtxpos[2].xy(), vtxpos[0].xy()) ? -1 : 0;
    int bias2 =
        IsRightSideOrFlatBottomEdge(vtxpos[2].xy(), vtxpos[0].xy(), vtxpos[1].xy()) ? -1 : 0;

    // Flip the quaternions if they are opposite to prevent interpolating them over the wrong direction.
    auto v1_quat = v1.quat;
    auto v2_quat = v2.quat;

    if (AreQuaternionsOpposite(v0.quat, v1.quat))
        v1_quat = v1_quat * float24::FromFloat32(-1.0f);
    if (AreQuaternionsOpposite(v0.quat, v2.quat))
        v2_quat = v2_quat * float24::FromFloat32(-1.0f);

    auto w_inverse = Math::MakeVec(v0.pos.w, v1.pos.w, v2.pos.w);

    auto textures = regs.texturing.GetTextures();
    auto tev_stages = regs.texturing.GetTevStages();

    bool stencil_action_enable =
        g_state.regs.framebuffer.output_merger.stencil_test.enable &&
        g_state.regs.framebuffer.framebuffer.depth_format == FramebufferRegs::DepthFormat::D24S8;
    const auto stencil_test = g_state.regs.framebuffer.output_merger.stencil_test;

    // Enter rasterization loop, starting at the center of the topleft bounding box corner.
    // TODO: Not sure if looping through x first might be faster
    for (u16 y = min_y + 8; y < max_y; y += 0x10) {
        for (u16 x = min_x + 8; x < max_x; x += 0x10) {

            // Do not process the pixel if it's inside the scissor box and the scissor mode is set
            // to Exclude
            if (regs.rasterizer.scissor_test.mode == RasterizerRegs::ScissorMode::Exclude) {
                if (x >= scissor_x1 && x < scissor_x2 && y >= scissor_y1 && y < scissor_y2)
                    continue;
            }

            // Calculate the barycentric coordinates w0, w1 and w2
            int w0 = bias0 + SignedArea(vtxpos[1].xy(), vtxpos[2].xy(), {x, y});
            int w1 = bias1 + SignedArea(vtxpos[2].xy(), vtxpos[0].xy(), {x, y});
            int w2 = bias2 + SignedArea(vtxpos[0].xy(), vtxpos[1].xy(), {x, y});
            int wsum = w0 + w1 + w2;

            // If current pixel is not covered by the current primitive
            if (w0 < 0 || w1 < 0 || w2 < 0)
                continue;

            auto baricentric_coordinates =
                Math::MakeVec(float24::FromFloat32(static_cast<float>(w0)),
                              float24::FromFloat32(static_cast<float>(w1)),
                              float24::FromFloat32(static_cast<float>(w2)));
            float24 interpolated_w_inverse =
                float24::FromFloat32(1.0f) / Math::Dot(w_inverse, baricentric_coordinates);

            // interpolated_z = z / w
            float interpolated_z_over_w =
                (v0.screenpos[2].ToFloat32() * w0 + v1.screenpos[2].ToFloat32() * w1 +
                 v2.screenpos[2].ToFloat32() * w2) /
                wsum;

            // Not fully accurate. About 3 bits in precision are missing.
            // Z-Buffer (z / w * scale + offset)
            float depth_scale = float24::FromRaw(regs.rasterizer.viewport_depth_range).ToFloat32();
            float depth_offset =
                float24::FromRaw(regs.rasterizer.viewport_depth_near_plane).ToFloat32();
            float depth = interpolated_z_over_w * depth_scale + depth_offset;

            // Potentially switch to W-Buffer
            if (regs.rasterizer.depthmap_enable ==
                Pica::RasterizerRegs::DepthBuffering::WBuffering) {
                // W-Buffer (z * scale + w * offset = (z / w * scale + offset) * w)
                depth *= interpolated_w_inverse.ToFloat32() * wsum;
            }

            // Clamp the result
            depth = MathUtil::Clamp(depth, 0.0f, 1.0f);

            // Perspective correct attribute interpolation:
            // Attribute values cannot be calculated by simple linear interpolation since
            // they are not linear in screen space. For example, when interpolating a
            // texture coordinate across two vertices, something simple like
            //     u = (u0*w0 + u1*w1)/(w0+w1)
            // will not work. However, the attribute value divided by the
            // clipspace w-coordinate (u/w) and and the inverse w-coordinate (1/w) are linear
            // in screenspace. Hence, we can linearly interpolate these two independently and
            // calculate the interpolated attribute by dividing the results.
            // I.e.
            //     u_over_w   = ((u0/v0.pos.w)*w0 + (u1/v1.pos.w)*w1)/(w0+w1)
            //     one_over_w = (( 1/v0.pos.w)*w0 + ( 1/v1.pos.w)*w1)/(w0+w1)
            //     u = u_over_w / one_over_w
            //
            // The generalization to three vertices is straightforward in baricentric coordinates.
            auto GetInterpolatedAttribute = [&](float24 attr0, float24 attr1, float24 attr2) {
                auto attr_over_w = Math::MakeVec(attr0, attr1, attr2);
                float24 interpolated_attr_over_w = Math::Dot(attr_over_w, baricentric_coordinates);
                return interpolated_attr_over_w * interpolated_w_inverse;
            };

            Math::Vec4<u8> primary_color{
                (u8)(
                    GetInterpolatedAttribute(v0.color.r(), v1.color.r(), v2.color.r()).ToFloat32() *
                    255),
                (u8)(
                    GetInterpolatedAttribute(v0.color.g(), v1.color.g(), v2.color.g()).ToFloat32() *
                    255),
                (u8)(
                    GetInterpolatedAttribute(v0.color.b(), v1.color.b(), v2.color.b()).ToFloat32() *
                    255),
                (u8)(
                    GetInterpolatedAttribute(v0.color.a(), v1.color.a(), v2.color.a()).ToFloat32() *
                    255),
            };

            Math::Quaternion<float> normquat{
                {
                    GetInterpolatedAttribute(v0.quat.x, v1_quat.x, v2_quat.x).ToFloat32(),
                    GetInterpolatedAttribute(v0.quat.y, v1_quat.y, v2_quat.y).ToFloat32(),
                    GetInterpolatedAttribute(v0.quat.z, v1_quat.z, v2_quat.z).ToFloat32()
                },
                GetInterpolatedAttribute(v0.quat.w, v1_quat.w, v2_quat.w).ToFloat32(),
            };

            Math::Vec3<float> fragment_position{
                GetInterpolatedAttribute(v0.view.x, v1.view.x, v2.view.x).ToFloat32(),
                GetInterpolatedAttribute(v0.view.y, v1.view.y, v2.view.y).ToFloat32(),
                GetInterpolatedAttribute(v0.view.z, v1.view.z, v2.view.z).ToFloat32()
            };

            Math::Vec2<float24> uv[3];
            uv[0].u() = GetInterpolatedAttribute(v0.tc0.u(), v1.tc0.u(), v2.tc0.u());
            uv[0].v() = GetInterpolatedAttribute(v0.tc0.v(), v1.tc0.v(), v2.tc0.v());
            uv[1].u() = GetInterpolatedAttribute(v0.tc1.u(), v1.tc1.u(), v2.tc1.u());
            uv[1].v() = GetInterpolatedAttribute(v0.tc1.v(), v1.tc1.v(), v2.tc1.v());
            uv[2].u() = GetInterpolatedAttribute(v0.tc2.u(), v1.tc2.u(), v2.tc2.u());
            uv[2].v() = GetInterpolatedAttribute(v0.tc2.v(), v1.tc2.v(), v2.tc2.v());

            Math::Vec4<u8> texture_color[4]{};
            for (int i = 0; i < 3; ++i) {
                const auto& texture = textures[i];
                if (!texture.enabled)
                    continue;

                DEBUG_ASSERT(0 != texture.config.address);

                int coordinate_i =
                    (i == 2 && regs.texturing.main_config.texture2_use_coord1) ? 1 : i;
                float24 u = uv[coordinate_i].u();
                float24 v = uv[coordinate_i].v();

                // Only unit 0 respects the texturing type (according to 3DBrew)
                // TODO: Refactor so cubemaps and shadowmaps can be handled
                PAddr texture_address = texture.config.GetPhysicalAddress();
                if (i == 0) {
                    switch (texture.config.type) {
                    case TexturingRegs::TextureConfig::Texture2D:
                        break;
                    case TexturingRegs::TextureConfig::TextureCube: {
                        auto w = GetInterpolatedAttribute(v0.tc0_w, v1.tc0_w, v2.tc0_w);
                        std::tie(u, v, texture_address) = ConvertCubeCoord(u, v, w, regs.texturing);
                        break;
                    }
                    case TexturingRegs::TextureConfig::Projection2D: {
                        auto tc0_w = GetInterpolatedAttribute(v0.tc0_w, v1.tc0_w, v2.tc0_w);
                        u /= tc0_w;
                        v /= tc0_w;
                        break;
                    }
                    default:
                        // TODO: Change to LOG_ERROR when more types are handled.
                        LOG_DEBUG(HW_GPU, "Unhandled texture type %x", (int)texture.config.type);
                        UNIMPLEMENTED();
                        break;
                    }
                }

                int s = (int)(u * float24::FromFloat32(static_cast<float>(texture.config.width)))
                            .ToFloat32();
                int t = (int)(v * float24::FromFloat32(static_cast<float>(texture.config.height)))
                            .ToFloat32();

                bool use_border_s = false;
                bool use_border_t = false;

                if (texture.config.wrap_s == TexturingRegs::TextureConfig::ClampToBorder) {
                    use_border_s = s < 0 || s >= static_cast<int>(texture.config.width);
                } else if (texture.config.wrap_s == TexturingRegs::TextureConfig::ClampToBorder2) {
                    use_border_s = s >= static_cast<int>(texture.config.width);
                }

                if (texture.config.wrap_t == TexturingRegs::TextureConfig::ClampToBorder) {
                    use_border_t = t < 0 || t >= static_cast<int>(texture.config.height);
                } else if (texture.config.wrap_t == TexturingRegs::TextureConfig::ClampToBorder2) {
                    use_border_t = t >= static_cast<int>(texture.config.height);
                }

                if (use_border_s || use_border_t) {
                    auto border_color = texture.config.border_color;
                    texture_color[i] = {border_color.r, border_color.g, border_color.b,
                                        border_color.a};
                } else {
                    // Textures are laid out from bottom to top, hence we invert the t coordinate.
                    // NOTE: This may not be the right place for the inversion.
                    // TODO: Check if this applies to ETC textures, too.
                    s = GetWrappedTexCoord(texture.config.wrap_s, s, texture.config.width);
                    t = texture.config.height - 1 -
                        GetWrappedTexCoord(texture.config.wrap_t, t, texture.config.height);

                    const u8* texture_data = Memory::GetPhysicalPointer(texture_address);
                    auto info =
                        Texture::TextureInfo::FromPicaRegister(texture.config, texture.format);

                    // TODO: Apply the min and mag filters to the texture
                    texture_color[i] = Texture::LookupTexture(texture_data, s, t, info);
#if PICA_DUMP_TEXTURES
                    DebugUtils::DumpTexture(texture.config, texture_data);
#endif
                }
            }

            // sample procedural texture
            if (regs.texturing.main_config.texture3_enable) {
                const auto& proctex_uv = uv[regs.texturing.main_config.texture3_coordinates];
                texture_color[3] = ProcTex(proctex_uv.u().ToFloat32(), proctex_uv.v().ToFloat32(),
                                           g_state.regs.texturing, g_state.proctex);
            }

            // Texture environment - consists of 6 stages of color and alpha combining.
            //
            // Color combiners take three input color values from some source (e.g. interpolated
            // vertex color, texture color, previous stage, etc), perform some very simple
            // operations on each of them (e.g. inversion) and then calculate the output color
            // with some basic arithmetic. Alpha combiners can be configured separately but work
            // analogously.
            Math::Vec4<u8> combiner_output;
            Math::Vec4<u8> combiner_buffer = {0, 0, 0, 0};
            Math::Vec4<u8> next_combiner_buffer = {
                regs.texturing.tev_combiner_buffer_color.r,
                regs.texturing.tev_combiner_buffer_color.g,
                regs.texturing.tev_combiner_buffer_color.b,
                regs.texturing.tev_combiner_buffer_color.a,
            };

            Math::Vec4<u8> primary_fragment_color;
            Math::Vec4<u8> secondary_fragment_color;

            std::tie(primary_fragment_color, secondary_fragment_color) = ComputeFragmentsColors(normquat, fragment_position);

            for (unsigned tev_stage_index = 0; tev_stage_index < tev_stages.size();
                 ++tev_stage_index) {
                const auto& tev_stage = tev_stages[tev_stage_index];
                using Source = TexturingRegs::TevStageConfig::Source;

                auto GetSource = [&](Source source) -> Math::Vec4<u8> {
                    switch (source) {
                    case Source::PrimaryColor:
                        return primary_color;

                    case Source::PrimaryFragmentColor:
                        return primary_fragment_color;

                    case Source::SecondaryFragmentColor:
                        return secondary_fragment_color;

                    case Source::Texture0:
                        return texture_color[0];

                    case Source::Texture1:
                        return texture_color[1];

                    case Source::Texture2:
                        return texture_color[2];

                    case Source::Texture3:
                        return texture_color[3];

                    case Source::PreviousBuffer:
                        return combiner_buffer;

                    case Source::Constant:
                        return {tev_stage.const_r, tev_stage.const_g, tev_stage.const_b,
                                tev_stage.const_a};

                    case Source::Previous:
                        return combiner_output;

                    default:
                        LOG_ERROR(HW_GPU, "Unknown color combiner source %d", (int)source);
                        UNIMPLEMENTED();
                        return {0, 0, 0, 0};
                    }
                };

                // color combiner
                // NOTE: Not sure if the alpha combiner might use the color output of the previous
                //       stage as input. Hence, we currently don't directly write the result to
                //       combiner_output.rgb(), but instead store it in a temporary variable until
                //       alpha combining has been done.
                Math::Vec3<u8> color_result[3] = {
                    GetColorModifier(tev_stage.color_modifier1, GetSource(tev_stage.color_source1)),
                    GetColorModifier(tev_stage.color_modifier2, GetSource(tev_stage.color_source2)),
                    GetColorModifier(tev_stage.color_modifier3, GetSource(tev_stage.color_source3)),
                };
                auto color_output = ColorCombine(tev_stage.color_op, color_result);

                u8 alpha_output;
                if (tev_stage.color_op == TexturingRegs::TevStageConfig::Operation::Dot3_RGBA) {
                    // result of Dot3_RGBA operation is also placed to the alpha component
                    alpha_output = color_output.x;
                } else {
                    // alpha combiner
                    std::array<u8, 3> alpha_result = {{
                        GetAlphaModifier(tev_stage.alpha_modifier1,
                                         GetSource(tev_stage.alpha_source1)),
                        GetAlphaModifier(tev_stage.alpha_modifier2,
                                         GetSource(tev_stage.alpha_source2)),
                        GetAlphaModifier(tev_stage.alpha_modifier3,
                                         GetSource(tev_stage.alpha_source3)),
                    }};
                    alpha_output = AlphaCombine(tev_stage.alpha_op, alpha_result);
                }

                combiner_output[0] =
                    std::min((unsigned)255, color_output.r() * tev_stage.GetColorMultiplier());
                combiner_output[1] =
                    std::min((unsigned)255, color_output.g() * tev_stage.GetColorMultiplier());
                combiner_output[2] =
                    std::min((unsigned)255, color_output.b() * tev_stage.GetColorMultiplier());
                combiner_output[3] =
                    std::min((unsigned)255, alpha_output * tev_stage.GetAlphaMultiplier());

                combiner_buffer = next_combiner_buffer;

                if (regs.texturing.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferColor(
                        tev_stage_index)) {
                    next_combiner_buffer.r() = combiner_output.r();
                    next_combiner_buffer.g() = combiner_output.g();
                    next_combiner_buffer.b() = combiner_output.b();
                }

                if (regs.texturing.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferAlpha(
                        tev_stage_index)) {
                    next_combiner_buffer.a() = combiner_output.a();
                }
            }

            const auto& output_merger = regs.framebuffer.output_merger;
            // TODO: Does alpha testing happen before or after stencil?
            if (output_merger.alpha_test.enable) {
                bool pass = false;

                switch (output_merger.alpha_test.func) {
                case FramebufferRegs::CompareFunc::Never:
                    pass = false;
                    break;

                case FramebufferRegs::CompareFunc::Always:
                    pass = true;
                    break;

                case FramebufferRegs::CompareFunc::Equal:
                    pass = combiner_output.a() == output_merger.alpha_test.ref;
                    break;

                case FramebufferRegs::CompareFunc::NotEqual:
                    pass = combiner_output.a() != output_merger.alpha_test.ref;
                    break;

                case FramebufferRegs::CompareFunc::LessThan:
                    pass = combiner_output.a() < output_merger.alpha_test.ref;
                    break;

                case FramebufferRegs::CompareFunc::LessThanOrEqual:
                    pass = combiner_output.a() <= output_merger.alpha_test.ref;
                    break;

                case FramebufferRegs::CompareFunc::GreaterThan:
                    pass = combiner_output.a() > output_merger.alpha_test.ref;
                    break;

                case FramebufferRegs::CompareFunc::GreaterThanOrEqual:
                    pass = combiner_output.a() >= output_merger.alpha_test.ref;
                    break;
                }

                if (!pass)
                    continue;
            }

            // Apply fog combiner
            // Not fully accurate. We'd have to know what data type is used to
            // store the depth etc. Using float for now until we know more
            // about Pica datatypes
            if (regs.texturing.fog_mode == TexturingRegs::FogMode::Fog) {
                const Math::Vec3<u8> fog_color = {
                    static_cast<u8>(regs.texturing.fog_color.r.Value()),
                    static_cast<u8>(regs.texturing.fog_color.g.Value()),
                    static_cast<u8>(regs.texturing.fog_color.b.Value()),
                };

                // Get index into fog LUT
                float fog_index;
                if (g_state.regs.texturing.fog_flip) {
                    fog_index = (1.0f - depth) * 128.0f;
                } else {
                    fog_index = depth * 128.0f;
                }

                // Generate clamped fog factor from LUT for given fog index
                float fog_i = MathUtil::Clamp(floorf(fog_index), 0.0f, 127.0f);
                float fog_f = fog_index - fog_i;
                const auto& fog_lut_entry = g_state.fog.lut[static_cast<unsigned int>(fog_i)];
                float fog_factor = fog_lut_entry.ToFloat() + fog_lut_entry.DiffToFloat() * fog_f;
                fog_factor = MathUtil::Clamp(fog_factor, 0.0f, 1.0f);

                // Blend the fog
                for (unsigned i = 0; i < 3; i++) {
                    combiner_output[i] = static_cast<u8>(fog_factor * combiner_output[i] +
                                                         (1.0f - fog_factor) * fog_color[i]);
                }
            }

            u8 old_stencil = 0;

            auto UpdateStencil = [stencil_test, x, y,
                                  &old_stencil](Pica::FramebufferRegs::StencilAction action) {
                u8 new_stencil =
                    PerformStencilAction(action, old_stencil, stencil_test.reference_value);
                if (g_state.regs.framebuffer.framebuffer.allow_depth_stencil_write != 0)
                    SetStencil(x >> 4, y >> 4, (new_stencil & stencil_test.write_mask) |
                                                   (old_stencil & ~stencil_test.write_mask));
            };

            if (stencil_action_enable) {
                old_stencil = GetStencil(x >> 4, y >> 4);
                u8 dest = old_stencil & stencil_test.input_mask;
                u8 ref = stencil_test.reference_value & stencil_test.input_mask;

                bool pass = false;
                switch (stencil_test.func) {
                case FramebufferRegs::CompareFunc::Never:
                    pass = false;
                    break;

                case FramebufferRegs::CompareFunc::Always:
                    pass = true;
                    break;

                case FramebufferRegs::CompareFunc::Equal:
                    pass = (ref == dest);
                    break;

                case FramebufferRegs::CompareFunc::NotEqual:
                    pass = (ref != dest);
                    break;

                case FramebufferRegs::CompareFunc::LessThan:
                    pass = (ref < dest);
                    break;

                case FramebufferRegs::CompareFunc::LessThanOrEqual:
                    pass = (ref <= dest);
                    break;

                case FramebufferRegs::CompareFunc::GreaterThan:
                    pass = (ref > dest);
                    break;

                case FramebufferRegs::CompareFunc::GreaterThanOrEqual:
                    pass = (ref >= dest);
                    break;
                }

                if (!pass) {
                    UpdateStencil(stencil_test.action_stencil_fail);
                    continue;
                }
            }

            // Convert float to integer
            unsigned num_bits =
                FramebufferRegs::DepthBitsPerPixel(regs.framebuffer.framebuffer.depth_format);
            u32 z = (u32)(depth * ((1 << num_bits) - 1));

            if (output_merger.depth_test_enable) {
                u32 ref_z = GetDepth(x >> 4, y >> 4);

                bool pass = false;

                switch (output_merger.depth_test_func) {
                case FramebufferRegs::CompareFunc::Never:
                    pass = false;
                    break;

                case FramebufferRegs::CompareFunc::Always:
                    pass = true;
                    break;

                case FramebufferRegs::CompareFunc::Equal:
                    pass = z == ref_z;
                    break;

                case FramebufferRegs::CompareFunc::NotEqual:
                    pass = z != ref_z;
                    break;

                case FramebufferRegs::CompareFunc::LessThan:
                    pass = z < ref_z;
                    break;

                case FramebufferRegs::CompareFunc::LessThanOrEqual:
                    pass = z <= ref_z;
                    break;

                case FramebufferRegs::CompareFunc::GreaterThan:
                    pass = z > ref_z;
                    break;

                case FramebufferRegs::CompareFunc::GreaterThanOrEqual:
                    pass = z >= ref_z;
                    break;
                }

                if (!pass) {
                    if (stencil_action_enable)
                        UpdateStencil(stencil_test.action_depth_fail);
                    continue;
                }
            }

            if (regs.framebuffer.framebuffer.allow_depth_stencil_write != 0 &&
                output_merger.depth_write_enable) {

                SetDepth(x >> 4, y >> 4, z);
            }

            // The stencil depth_pass action is executed even if depth testing is disabled
            if (stencil_action_enable)
                UpdateStencil(stencil_test.action_depth_pass);

            auto dest = GetPixel(x >> 4, y >> 4);
            Math::Vec4<u8> blend_output = combiner_output;

            if (output_merger.alphablend_enable) {
                auto params = output_merger.alpha_blending;

                auto LookupFactor = [&](unsigned channel,
                                        FramebufferRegs::BlendFactor factor) -> u8 {
                    DEBUG_ASSERT(channel < 4);

                    const Math::Vec4<u8> blend_const = {
                        static_cast<u8>(output_merger.blend_const.r),
                        static_cast<u8>(output_merger.blend_const.g),
                        static_cast<u8>(output_merger.blend_const.b),
                        static_cast<u8>(output_merger.blend_const.a),
                    };

                    switch (factor) {
                    case FramebufferRegs::BlendFactor::Zero:
                        return 0;

                    case FramebufferRegs::BlendFactor::One:
                        return 255;

                    case FramebufferRegs::BlendFactor::SourceColor:
                        return combiner_output[channel];

                    case FramebufferRegs::BlendFactor::OneMinusSourceColor:
                        return 255 - combiner_output[channel];

                    case FramebufferRegs::BlendFactor::DestColor:
                        return dest[channel];

                    case FramebufferRegs::BlendFactor::OneMinusDestColor:
                        return 255 - dest[channel];

                    case FramebufferRegs::BlendFactor::SourceAlpha:
                        return combiner_output.a();

                    case FramebufferRegs::BlendFactor::OneMinusSourceAlpha:
                        return 255 - combiner_output.a();

                    case FramebufferRegs::BlendFactor::DestAlpha:
                        return dest.a();

                    case FramebufferRegs::BlendFactor::OneMinusDestAlpha:
                        return 255 - dest.a();

                    case FramebufferRegs::BlendFactor::ConstantColor:
                        return blend_const[channel];

                    case FramebufferRegs::BlendFactor::OneMinusConstantColor:
                        return 255 - blend_const[channel];

                    case FramebufferRegs::BlendFactor::ConstantAlpha:
                        return blend_const.a();

                    case FramebufferRegs::BlendFactor::OneMinusConstantAlpha:
                        return 255 - blend_const.a();

                    case FramebufferRegs::BlendFactor::SourceAlphaSaturate:
                        // Returns 1.0 for the alpha channel
                        if (channel == 3)
                            return 255;
                        return std::min(combiner_output.a(), static_cast<u8>(255 - dest.a()));

                    default:
                        LOG_CRITICAL(HW_GPU, "Unknown blend factor %x", factor);
                        UNIMPLEMENTED();
                        break;
                    }

                    return combiner_output[channel];
                };

                auto srcfactor = Math::MakeVec(LookupFactor(0, params.factor_source_rgb),
                                               LookupFactor(1, params.factor_source_rgb),
                                               LookupFactor(2, params.factor_source_rgb),
                                               LookupFactor(3, params.factor_source_a));

                auto dstfactor = Math::MakeVec(LookupFactor(0, params.factor_dest_rgb),
                                               LookupFactor(1, params.factor_dest_rgb),
                                               LookupFactor(2, params.factor_dest_rgb),
                                               LookupFactor(3, params.factor_dest_a));

                blend_output = EvaluateBlendEquation(combiner_output, srcfactor, dest, dstfactor,
                                                     params.blend_equation_rgb);
                blend_output.a() = EvaluateBlendEquation(combiner_output, srcfactor, dest,
                                                         dstfactor, params.blend_equation_a)
                                       .a();
            } else {
                blend_output =
                    Math::MakeVec(LogicOp(combiner_output.r(), dest.r(), output_merger.logic_op),
                                  LogicOp(combiner_output.g(), dest.g(), output_merger.logic_op),
                                  LogicOp(combiner_output.b(), dest.b(), output_merger.logic_op),
                                  LogicOp(combiner_output.a(), dest.a(), output_merger.logic_op));
            }

            const Math::Vec4<u8> result = {
                output_merger.red_enable ? blend_output.r() : dest.r(),
                output_merger.green_enable ? blend_output.g() : dest.g(),
                output_merger.blue_enable ? blend_output.b() : dest.b(),
                output_merger.alpha_enable ? blend_output.a() : dest.a(),
            };

            if (regs.framebuffer.framebuffer.allow_color_write != 0)
                DrawPixel(x >> 4, y >> 4, result);
        }
    }
}

void ProcessTriangle(const Vertex& v0, const Vertex& v1, const Vertex& v2) {
    ProcessTriangleInternal(v0, v1, v2);
}

} // namespace Rasterizer

} // namespace Pica