#include "imgui_impl_sdlsurface3.h"
#include "imgui.h"
#ifndef IMGUI_DISABLE
#include <SDL3/SDL.h>
#include <stdint.h>
#include <algorithm>
#include <cmath>
#include <cstring>

struct ImGui_ImplSDLSurface3_FastFormat32
{
    Uint32 Rmask;
    Uint32 Gmask;
    Uint32 Bmask;
    Uint32 Amask;
    Uint8 Rshift;
    Uint8 Gshift;
    Uint8 Bshift;
    Uint8 Ashift;
    bool HasAlpha;
    bool Valid;

    ImGui_ImplSDLSurface3_FastFormat32() { memset((void*)this, 0, sizeof(*this)); }
};

struct ImGui_ImplSDLSurface3_SurfaceInfo
{
    SDL_Surface*                              Surface;
    Uint8*                                    Pixels;
    int                                       Pitch;
    int                                       Width;
    int                                       Height;
    const SDL_PixelFormatDetails*             Format;
    SDL_Palette*                              Palette;
    ImGui_ImplSDLSurface3_FastFormat32        FastFormat;
    bool                                      IsFast;

    ImGui_ImplSDLSurface3_SurfaceInfo() { memset((void*)this, 0, sizeof(*this)); }
};

struct ImGui_ImplSDLSurface3_Data
{
    SDL_Surface*                              TargetSurface;
    SDL_Surface*                              FontSurface;
    ImGui_ImplSDLSurface3_SurfaceInfo         TargetInfo;

    ImGui_ImplSDLSurface3_Data() { memset((void*)this, 0, sizeof(*this)); }
};

struct ImGui_ImplSDLSurface3_RasterVertex
{
    float X;
    float Y;
    float U;
    float V;
    float R;
    float G;
    float B;
    float A;
};

static ImGui_ImplSDLSurface3_Data* ImGui_ImplSDLSurface3_GetBackendData()
{
    return ImGui::GetCurrentContext() ? (ImGui_ImplSDLSurface3_Data*)ImGui::GetIO().BackendRendererUserData : nullptr;
}

static inline float ImGui_ImplSDLSurface3_Edge(float ax, float ay, float bx, float by, float px, float py)
{
    return (bx - ax) * (py - ay) - (by - ay) * (px - ax);
}

static inline int ImGui_ImplSDLSurface3_ClampInt(int v, int lo, int hi)
{
    return (v < lo) ? lo : (v > hi ? hi : v);
}

static inline Uint8 ImGui_ImplSDLSurface3_ClampByte(float v)
{
    if (v <= 0.0f)
        return 0;
    if (v >= 255.0f)
        return 255;
    return (Uint8)(v + 0.5f);
}

static inline Uint8 ImGui_ImplSDLSurface3_Mul255(Uint32 a, Uint32 b)
{
    Uint32 t = a * b + 128;
    return (Uint8)((t + (t >> 8)) >> 8);
}

static ImGui_ImplSDLSurface3_FastFormat32 ImGui_ImplSDLSurface3_BuildFastFormat(const SDL_PixelFormatDetails* format)
{
    ImGui_ImplSDLSurface3_FastFormat32 out;
    if (format == nullptr)
        return out;
    if (format->bits_per_pixel != 32 || format->bytes_per_pixel != 4)
        return out;
    if (format->Rbits != 8 || format->Gbits != 8 || format->Bbits != 8)
        return out;
    if (format->Amask != 0 && format->Abits != 8)
        return out;
    out.Rmask = format->Rmask;
    out.Gmask = format->Gmask;
    out.Bmask = format->Bmask;
    out.Amask = format->Amask;
    out.Rshift = format->Rshift;
    out.Gshift = format->Gshift;
    out.Bshift = format->Bshift;
    out.Ashift = format->Ashift;
    out.HasAlpha = format->Amask != 0;
    out.Valid = true;
    return out;
}

static bool ImGui_ImplSDLSurface3_CanUseFastPath(SDL_Surface* surface, const ImGui_ImplSDLSurface3_FastFormat32& format)
{
#ifdef IMGUI_IMPL_SDLSURFACE_DISABLE_FAST_PATH
    IM_UNUSED(surface);
    IM_UNUSED(format);
    return false;
#else
    if (surface == nullptr || !format.Valid)
        return false;
    if ((surface->pitch & 3) != 0)
        return false;
    if ((((uintptr_t)surface->pixels) & 3) != 0)
        return false;
    return true;
#endif
}

static void ImGui_ImplSDLSurface3_InitSurfaceInfo(ImGui_ImplSDLSurface3_SurfaceInfo* out, SDL_Surface* surface)
{
    out->Surface = surface;
    out->Pixels = surface ? (Uint8*)surface->pixels : nullptr;
    out->Pitch = surface ? surface->pitch : 0;
    out->Width = surface ? surface->w : 0;
    out->Height = surface ? surface->h : 0;
    out->Format = surface ? SDL_GetPixelFormatDetails(surface->format) : nullptr;
    out->Palette = surface ? SDL_GetSurfacePalette(surface) : nullptr;
    out->FastFormat = ImGui_ImplSDLSurface3_BuildFastFormat(out->Format);
    out->IsFast = ImGui_ImplSDLSurface3_CanUseFastPath(surface, out->FastFormat);
}

static inline Uint32 ImGui_ImplSDLSurface3_ReadPixelRaw(const ImGui_ImplSDLSurface3_SurfaceInfo& surface, int x, int y)
{
    const Uint8* p = surface.Pixels + y * surface.Pitch + x * surface.Format->bytes_per_pixel;
    switch (surface.Format->bytes_per_pixel)
    {
    case 1:
        return *p;
    case 2:
        return *(const Uint16*)p;
    case 3:
        if (SDL_BYTEORDER == SDL_BIG_ENDIAN)
            return (Uint32)((p[0] << 16) | (p[1] << 8) | p[2]);
        return (Uint32)(p[0] | (p[1] << 8) | (p[2] << 16));
    case 4:
        return *(const Uint32*)p;
    default:
        return 0;
    }
}

static inline void ImGui_ImplSDLSurface3_WritePixelRaw(const ImGui_ImplSDLSurface3_SurfaceInfo& surface, int x, int y, Uint32 pixel)
{
    Uint8* p = surface.Pixels + y * surface.Pitch + x * surface.Format->bytes_per_pixel;
    switch (surface.Format->bytes_per_pixel)
    {
    case 1:
        *p = (Uint8)pixel;
        return;
    case 2:
        *(Uint16*)p = (Uint16)pixel;
        return;
    case 3:
        if (SDL_BYTEORDER == SDL_BIG_ENDIAN)
        {
            p[0] = (Uint8)((pixel >> 16) & 0xFF);
            p[1] = (Uint8)((pixel >> 8) & 0xFF);
            p[2] = (Uint8)(pixel & 0xFF);
        }
        else
        {
            p[0] = (Uint8)(pixel & 0xFF);
            p[1] = (Uint8)((pixel >> 8) & 0xFF);
            p[2] = (Uint8)((pixel >> 16) & 0xFF);
        }
        return;
    case 4:
        *(Uint32*)p = pixel;
        return;
    default:
        return;
    }
}

static inline void ImGui_ImplSDLSurface3_UnpackFastRGBA(const ImGui_ImplSDLSurface3_FastFormat32& format, Uint32 pixel, Uint8* r, Uint8* g, Uint8* b, Uint8* a)
{
    *r = (Uint8)((pixel & format.Rmask) >> format.Rshift);
    *g = (Uint8)((pixel & format.Gmask) >> format.Gshift);
    *b = (Uint8)((pixel & format.Bmask) >> format.Bshift);
    *a = format.HasAlpha ? (Uint8)((pixel & format.Amask) >> format.Ashift) : 255;
}

static inline Uint32 ImGui_ImplSDLSurface3_PackFastRGBA(const ImGui_ImplSDLSurface3_FastFormat32& format, Uint8 r, Uint8 g, Uint8 b, Uint8 a)
{
    Uint32 pixel = (((Uint32)r << format.Rshift) & format.Rmask)
        | (((Uint32)g << format.Gshift) & format.Gmask)
        | (((Uint32)b << format.Bshift) & format.Bmask);
    if (format.HasAlpha)
        pixel |= (((Uint32)a << format.Ashift) & format.Amask);
    return pixel;
}

static inline void ImGui_ImplSDLSurface3_ReadPixelRGBA(const ImGui_ImplSDLSurface3_SurfaceInfo& surface, int x, int y, Uint8* r, Uint8* g, Uint8* b, Uint8* a)
{
    if (surface.IsFast)
    {
        const Uint32* src = (const Uint32*)(surface.Pixels + y * surface.Pitch) + x;
        ImGui_ImplSDLSurface3_UnpackFastRGBA(surface.FastFormat, *src, r, g, b, a);
        return;
    }

    Uint32 pixel = ImGui_ImplSDLSurface3_ReadPixelRaw(surface, x, y);
    SDL_GetRGBA(pixel, surface.Format, surface.Palette, r, g, b, a);
}

static inline void ImGui_ImplSDLSurface3_BlendPixelFast(const ImGui_ImplSDLSurface3_SurfaceInfo& target, Uint32* dst, Uint8 sr, Uint8 sg, Uint8 sb, Uint8 sa)
{
    if (sa == 0)
        return;
    if (sa == 255)
    {
        *dst = ImGui_ImplSDLSurface3_PackFastRGBA(target.FastFormat, sr, sg, sb, 255);
        return;
    }

    Uint8 dr, dg, db, da;
    ImGui_ImplSDLSurface3_UnpackFastRGBA(target.FastFormat, *dst, &dr, &dg, &db, &da);
    Uint8 inv_a = (Uint8)(255 - sa);
    Uint8 out_r = (Uint8)(ImGui_ImplSDLSurface3_Mul255(sr, sa) + ImGui_ImplSDLSurface3_Mul255(dr, inv_a));
    Uint8 out_g = (Uint8)(ImGui_ImplSDLSurface3_Mul255(sg, sa) + ImGui_ImplSDLSurface3_Mul255(dg, inv_a));
    Uint8 out_b = (Uint8)(ImGui_ImplSDLSurface3_Mul255(sb, sa) + ImGui_ImplSDLSurface3_Mul255(db, inv_a));
    Uint8 out_a = (target.Format != nullptr && target.Format->Amask != 0) ? (Uint8)(sa + ImGui_ImplSDLSurface3_Mul255(da, inv_a)) : 255;
    *dst = ImGui_ImplSDLSurface3_PackFastRGBA(target.FastFormat, out_r, out_g, out_b, out_a);
}

static inline void ImGui_ImplSDLSurface3_BlendPixelGeneric(const ImGui_ImplSDLSurface3_SurfaceInfo& target, int x, int y, Uint8 sr, Uint8 sg, Uint8 sb, Uint8 sa)
{
    if (sa == 0)
        return;
    if (sa == 255)
    {
        Uint32 pixel = SDL_MapRGBA(target.Format, target.Palette, sr, sg, sb, 255);
        ImGui_ImplSDLSurface3_WritePixelRaw(target, x, y, pixel);
        return;
    }

    Uint8 dr, dg, db, da;
    Uint32 dst_pixel = ImGui_ImplSDLSurface3_ReadPixelRaw(target, x, y);
    SDL_GetRGBA(dst_pixel, target.Format, target.Palette, &dr, &dg, &db, &da);
    Uint8 inv_a = (Uint8)(255 - sa);
    Uint8 out_r = (Uint8)(ImGui_ImplSDLSurface3_Mul255(sr, sa) + ImGui_ImplSDLSurface3_Mul255(dr, inv_a));
    Uint8 out_g = (Uint8)(ImGui_ImplSDLSurface3_Mul255(sg, sa) + ImGui_ImplSDLSurface3_Mul255(dg, inv_a));
    Uint8 out_b = (Uint8)(ImGui_ImplSDLSurface3_Mul255(sb, sa) + ImGui_ImplSDLSurface3_Mul255(db, inv_a));
    Uint8 out_a = target.FastFormat.HasAlpha ? (Uint8)(sa + ImGui_ImplSDLSurface3_Mul255(da, inv_a)) : 255;
    Uint32 pixel = SDL_MapRGBA(target.Format, target.Palette, out_r, out_g, out_b, out_a);
    ImGui_ImplSDLSurface3_WritePixelRaw(target, x, y, pixel);
}

static inline void ImGui_ImplSDLSurface3_SetupRasterVertex(ImGui_ImplSDLSurface3_RasterVertex* out, const ImDrawVert& v, const ImVec2& display_pos, const ImVec2& framebuffer_scale)
{
    out->X = (v.pos.x - display_pos.x) * framebuffer_scale.x;
    out->Y = (v.pos.y - display_pos.y) * framebuffer_scale.y;
    out->U = v.uv.x;
    out->V = v.uv.y;
    out->R = (float)((v.col >> IM_COL32_R_SHIFT) & 0xFF);
    out->G = (float)((v.col >> IM_COL32_G_SHIFT) & 0xFF);
    out->B = (float)((v.col >> IM_COL32_B_SHIFT) & 0xFF);
    out->A = (float)((v.col >> IM_COL32_A_SHIFT) & 0xFF);
}

static void ImGui_ImplSDLSurface3_RenderTriangleFast(const ImGui_ImplSDLSurface3_SurfaceInfo& target, const ImGui_ImplSDLSurface3_SurfaceInfo* texture, const ImGui_ImplSDLSurface3_RasterVertex& v0, const ImGui_ImplSDLSurface3_RasterVertex& v1, const ImGui_ImplSDLSurface3_RasterVertex& v2, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y)
{
    int min_x = std::max((int)std::floor(std::min(v0.X, std::min(v1.X, v2.X))), clip_min_x);
    int min_y = std::max((int)std::floor(std::min(v0.Y, std::min(v1.Y, v2.Y))), clip_min_y);
    int max_x = std::min((int)std::ceil(std::max(v0.X, std::max(v1.X, v2.X))) - 1, clip_max_x - 1);
    int max_y = std::min((int)std::ceil(std::max(v0.Y, std::max(v1.Y, v2.Y))) - 1, clip_max_y - 1);
    if (min_x > max_x || min_y > max_y)
        return;

    float area = ImGui_ImplSDLSurface3_Edge(v0.X, v0.Y, v1.X, v1.Y, v2.X, v2.Y);
    if (area == 0.0f)
        return;

    float sign = (area < 0.0f) ? -1.0f : 1.0f;
    area *= sign;

    float e0_dx = (v1.Y - v2.Y) * sign;
    float e0_dy = (v2.X - v1.X) * sign;
    float e1_dx = (v2.Y - v0.Y) * sign;
    float e1_dy = (v0.X - v2.X) * sign;
    float e2_dx = (v0.Y - v1.Y) * sign;
    float e2_dy = (v1.X - v0.X) * sign;

    float inv_area = 1.0f / area;
    float dw0_dx = e0_dx * inv_area;
    float dw0_dy = e0_dy * inv_area;
    float dw1_dx = e1_dx * inv_area;
    float dw1_dy = e1_dy * inv_area;
    float dw2_dx = e2_dx * inv_area;
    float dw2_dy = e2_dy * inv_area;

    float dr_dx = dw0_dx * v0.R + dw1_dx * v1.R + dw2_dx * v2.R;
    float dg_dx = dw0_dx * v0.G + dw1_dx * v1.G + dw2_dx * v2.G;
    float db_dx = dw0_dx * v0.B + dw1_dx * v1.B + dw2_dx * v2.B;
    float da_dx = dw0_dx * v0.A + dw1_dx * v1.A + dw2_dx * v2.A;
    float du_dx = dw0_dx * v0.U + dw1_dx * v1.U + dw2_dx * v2.U;
    float dv_dx = dw0_dx * v0.V + dw1_dx * v1.V + dw2_dx * v2.V;

    float dr_dy = dw0_dy * v0.R + dw1_dy * v1.R + dw2_dy * v2.R;
    float dg_dy = dw0_dy * v0.G + dw1_dy * v1.G + dw2_dy * v2.G;
    float db_dy = dw0_dy * v0.B + dw1_dy * v1.B + dw2_dy * v2.B;
    float da_dy = dw0_dy * v0.A + dw1_dy * v1.A + dw2_dy * v2.A;
    float du_dy = dw0_dy * v0.U + dw1_dy * v1.U + dw2_dy * v2.U;
    float dv_dy = dw0_dy * v0.V + dw1_dy * v1.V + dw2_dy * v2.V;

    float start_x = (float)min_x + 0.5f;
    float start_y = (float)min_y + 0.5f;

    float e0_row = ImGui_ImplSDLSurface3_Edge(v1.X, v1.Y, v2.X, v2.Y, start_x, start_y) * sign;
    float e1_row = ImGui_ImplSDLSurface3_Edge(v2.X, v2.Y, v0.X, v0.Y, start_x, start_y) * sign;
    float e2_row = ImGui_ImplSDLSurface3_Edge(v0.X, v0.Y, v1.X, v1.Y, start_x, start_y) * sign;

    float w0_row = e0_row * inv_area;
    float w1_row = e1_row * inv_area;
    float w2_row = e2_row * inv_area;

    float r_row = w0_row * v0.R + w1_row * v1.R + w2_row * v2.R;
    float g_row = w0_row * v0.G + w1_row * v1.G + w2_row * v2.G;
    float b_row = w0_row * v0.B + w1_row * v1.B + w2_row * v2.B;
    float a_row = w0_row * v0.A + w1_row * v1.A + w2_row * v2.A;
    float u_row = w0_row * v0.U + w1_row * v1.U + w2_row * v2.U;
    float v_row = w0_row * v0.V + w1_row * v1.V + w2_row * v2.V;

    int texture_max_x = texture ? (texture->Width - 1) : 0;
    int texture_max_y = texture ? (texture->Height - 1) : 0;

    for (int y = min_y; y <= max_y; y++)
    {
        float e0 = e0_row;
        float e1 = e1_row;
        float e2 = e2_row;
        float r = r_row;
        float g = g_row;
        float b = b_row;
        float a = a_row;
        float u = u_row;
        float v = v_row;
        Uint32* dst = (Uint32*)(target.Pixels + y * target.Pitch) + min_x;

        for (int x = min_x; x <= max_x; x++)
        {
            if (e0 >= 0.0f && e1 >= 0.0f && e2 >= 0.0f)
            {
                Uint8 out_r = ImGui_ImplSDLSurface3_ClampByte(r);
                Uint8 out_g = ImGui_ImplSDLSurface3_ClampByte(g);
                Uint8 out_b = ImGui_ImplSDLSurface3_ClampByte(b);
                Uint8 out_a = ImGui_ImplSDLSurface3_ClampByte(a);

                if (texture != nullptr)
                {
                    int tx = ImGui_ImplSDLSurface3_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                    int ty = ImGui_ImplSDLSurface3_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                    Uint8 tr, tg, tb, ta;
                    ImGui_ImplSDLSurface3_ReadPixelRGBA(*texture, tx, ty, &tr, &tg, &tb, &ta);
                    out_r = ImGui_ImplSDLSurface3_Mul255(out_r, tr);
                    out_g = ImGui_ImplSDLSurface3_Mul255(out_g, tg);
                    out_b = ImGui_ImplSDLSurface3_Mul255(out_b, tb);
                    out_a = ImGui_ImplSDLSurface3_Mul255(out_a, ta);
                }

                ImGui_ImplSDLSurface3_BlendPixelFast(target, dst, out_r, out_g, out_b, out_a);
            }

            dst++;
            e0 += e0_dx;
            e1 += e1_dx;
            e2 += e2_dx;
            r += dr_dx;
            g += dg_dx;
            b += db_dx;
            a += da_dx;
            u += du_dx;
            v += dv_dx;
        }

        e0_row += e0_dy;
        e1_row += e1_dy;
        e2_row += e2_dy;
        r_row += dr_dy;
        g_row += dg_dy;
        b_row += db_dy;
        a_row += da_dy;
        u_row += du_dy;
        v_row += dv_dy;
    }
}

static void ImGui_ImplSDLSurface3_RenderTriangleGeneric(const ImGui_ImplSDLSurface3_SurfaceInfo& target, const ImGui_ImplSDLSurface3_SurfaceInfo* texture, const ImGui_ImplSDLSurface3_RasterVertex& v0, const ImGui_ImplSDLSurface3_RasterVertex& v1, const ImGui_ImplSDLSurface3_RasterVertex& v2, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y)
{
    int min_x = std::max((int)std::floor(std::min(v0.X, std::min(v1.X, v2.X))), clip_min_x);
    int min_y = std::max((int)std::floor(std::min(v0.Y, std::min(v1.Y, v2.Y))), clip_min_y);
    int max_x = std::min((int)std::ceil(std::max(v0.X, std::max(v1.X, v2.X))) - 1, clip_max_x - 1);
    int max_y = std::min((int)std::ceil(std::max(v0.Y, std::max(v1.Y, v2.Y))) - 1, clip_max_y - 1);
    if (min_x > max_x || min_y > max_y)
        return;

    float area = ImGui_ImplSDLSurface3_Edge(v0.X, v0.Y, v1.X, v1.Y, v2.X, v2.Y);
    if (area == 0.0f)
        return;

    float sign = (area < 0.0f) ? -1.0f : 1.0f;
    area *= sign;

    float e0_dx = (v1.Y - v2.Y) * sign;
    float e0_dy = (v2.X - v1.X) * sign;
    float e1_dx = (v2.Y - v0.Y) * sign;
    float e1_dy = (v0.X - v2.X) * sign;
    float e2_dx = (v0.Y - v1.Y) * sign;
    float e2_dy = (v1.X - v0.X) * sign;

    float inv_area = 1.0f / area;
    float dw0_dx = e0_dx * inv_area;
    float dw0_dy = e0_dy * inv_area;
    float dw1_dx = e1_dx * inv_area;
    float dw1_dy = e1_dy * inv_area;
    float dw2_dx = e2_dx * inv_area;
    float dw2_dy = e2_dy * inv_area;

    float dr_dx = dw0_dx * v0.R + dw1_dx * v1.R + dw2_dx * v2.R;
    float dg_dx = dw0_dx * v0.G + dw1_dx * v1.G + dw2_dx * v2.G;
    float db_dx = dw0_dx * v0.B + dw1_dx * v1.B + dw2_dx * v2.B;
    float da_dx = dw0_dx * v0.A + dw1_dx * v1.A + dw2_dx * v2.A;
    float du_dx = dw0_dx * v0.U + dw1_dx * v1.U + dw2_dx * v2.U;
    float dv_dx = dw0_dx * v0.V + dw1_dx * v1.V + dw2_dx * v2.V;

    float dr_dy = dw0_dy * v0.R + dw1_dy * v1.R + dw2_dy * v2.R;
    float dg_dy = dw0_dy * v0.G + dw1_dy * v1.G + dw2_dy * v2.G;
    float db_dy = dw0_dy * v0.B + dw1_dy * v1.B + dw2_dy * v2.B;
    float da_dy = dw0_dy * v0.A + dw1_dy * v1.A + dw2_dy * v2.A;
    float du_dy = dw0_dy * v0.U + dw1_dy * v1.U + dw2_dy * v2.U;
    float dv_dy = dw0_dy * v0.V + dw1_dy * v1.V + dw2_dy * v2.V;

    float start_x = (float)min_x + 0.5f;
    float start_y = (float)min_y + 0.5f;

    float e0_row = ImGui_ImplSDLSurface3_Edge(v1.X, v1.Y, v2.X, v2.Y, start_x, start_y) * sign;
    float e1_row = ImGui_ImplSDLSurface3_Edge(v2.X, v2.Y, v0.X, v0.Y, start_x, start_y) * sign;
    float e2_row = ImGui_ImplSDLSurface3_Edge(v0.X, v0.Y, v1.X, v1.Y, start_x, start_y) * sign;

    float w0_row = e0_row * inv_area;
    float w1_row = e1_row * inv_area;
    float w2_row = e2_row * inv_area;

    float r_row = w0_row * v0.R + w1_row * v1.R + w2_row * v2.R;
    float g_row = w0_row * v0.G + w1_row * v1.G + w2_row * v2.G;
    float b_row = w0_row * v0.B + w1_row * v1.B + w2_row * v2.B;
    float a_row = w0_row * v0.A + w1_row * v1.A + w2_row * v2.A;
    float u_row = w0_row * v0.U + w1_row * v1.U + w2_row * v2.U;
    float v_row = w0_row * v0.V + w1_row * v1.V + w2_row * v2.V;

    int texture_max_x = texture ? (texture->Width - 1) : 0;
    int texture_max_y = texture ? (texture->Height - 1) : 0;

    for (int y = min_y; y <= max_y; y++)
    {
        float e0 = e0_row;
        float e1 = e1_row;
        float e2 = e2_row;
        float r = r_row;
        float g = g_row;
        float b = b_row;
        float a = a_row;
        float u = u_row;
        float v = v_row;

        for (int x = min_x; x <= max_x; x++)
        {
            if (e0 >= 0.0f && e1 >= 0.0f && e2 >= 0.0f)
            {
                Uint8 out_r = ImGui_ImplSDLSurface3_ClampByte(r);
                Uint8 out_g = ImGui_ImplSDLSurface3_ClampByte(g);
                Uint8 out_b = ImGui_ImplSDLSurface3_ClampByte(b);
                Uint8 out_a = ImGui_ImplSDLSurface3_ClampByte(a);

                if (texture != nullptr)
                {
                    int tx = ImGui_ImplSDLSurface3_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                    int ty = ImGui_ImplSDLSurface3_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                    Uint8 tr, tg, tb, ta;
                    ImGui_ImplSDLSurface3_ReadPixelRGBA(*texture, tx, ty, &tr, &tg, &tb, &ta);
                    out_r = ImGui_ImplSDLSurface3_Mul255(out_r, tr);
                    out_g = ImGui_ImplSDLSurface3_Mul255(out_g, tg);
                    out_b = ImGui_ImplSDLSurface3_Mul255(out_b, tb);
                    out_a = ImGui_ImplSDLSurface3_Mul255(out_a, ta);
                }

                ImGui_ImplSDLSurface3_BlendPixelGeneric(target, x, y, out_r, out_g, out_b, out_a);
            }

            e0 += e0_dx;
            e1 += e1_dx;
            e2 += e2_dx;
            r += dr_dx;
            g += dg_dx;
            b += db_dx;
            a += da_dx;
            u += du_dx;
            v += dv_dx;
        }

        e0_row += e0_dy;
        e1_row += e1_dy;
        e2_row += e2_dy;
        r_row += dr_dy;
        g_row += dg_dy;
        b_row += db_dy;
        a_row += da_dy;
        u_row += du_dy;
        v_row += dv_dy;
    }
}

SDL_Surface* ImGui_ImplSDLSurface3_CreateFontAtlasSurface()
{
    ImGuiIO& io = ImGui::GetIO();
    unsigned char* pixels = nullptr;
    int width = 0;
    int height = 0;
    io.Fonts->GetTexDataAsRGBA32(&pixels, &width, &height);
    if (pixels == nullptr || width <= 0 || height <= 0)
        return nullptr;

    SDL_Surface* surface = SDL_CreateSurface(width, height, SDL_PIXELFORMAT_RGBA32);
    if (surface == nullptr)
        return nullptr;
    if (SDL_MUSTLOCK(surface) && !SDL_LockSurface(surface))
    {
        SDL_DestroySurface(surface);
        return nullptr;
    }

    for (int y = 0; y < height; y++)
        memcpy((Uint8*)surface->pixels + y * surface->pitch, pixels + (size_t)y * (size_t)width * 4, (size_t)width * 4);

    if (SDL_MUSTLOCK(surface))
        SDL_UnlockSurface(surface);

    SDL_SetSurfaceBlendMode(surface, SDL_BLENDMODE_NONE);
    return surface;
}

bool ImGui_ImplSDLSurface3_Init(SDL_Surface* surface)
{
    if (surface == nullptr)
        return false;

    ImGuiIO& io = ImGui::GetIO();
    IMGUI_CHECKVERSION();
    IM_ASSERT(io.BackendRendererUserData == nullptr && "Already initialized a renderer backend!");

    ImGui_ImplSDLSurface3_Data* bd = IM_NEW(ImGui_ImplSDLSurface3_Data)();
    io.BackendRendererUserData = (void*)bd;
    io.BackendRendererName = "imgui_impl_sdlsurface3";

    bd->TargetSurface = surface;
    ImGui_ImplSDLSurface3_InitSurfaceInfo(&bd->TargetInfo, surface);
    bd->FontSurface = ImGui_ImplSDLSurface3_CreateFontAtlasSurface();
    if (bd->FontSurface != nullptr)
        io.Fonts->TexID = (ImTextureID)bd->FontSurface;

    return true;
}

void ImGui_ImplSDLSurface3_Shutdown()
{
    ImGui_ImplSDLSurface3_Data* bd = ImGui_ImplSDLSurface3_GetBackendData();
    IM_ASSERT(bd != nullptr && "No renderer backend to shutdown, or already shutdown?");
    ImGuiIO& io = ImGui::GetIO();
    ImGuiPlatformIO& platform_io = ImGui::GetPlatformIO();

    io.Fonts->TexID = nullptr;
    io.BackendRendererName = nullptr;
    io.BackendRendererUserData = nullptr;
    platform_io.ClearRendererHandlers();

    if (bd->FontSurface != nullptr)
        SDL_DestroySurface(bd->FontSurface);

    IM_DELETE(bd);
}

void ImGui_ImplSDLSurface3_NewFrame()
{
    ImGui_ImplSDLSurface3_Data* bd = ImGui_ImplSDLSurface3_GetBackendData();
    IM_ASSERT(bd != nullptr && "Context or backend not initialized! Did you call ImGui_ImplSDLSurface3_Init()?");
    IM_UNUSED(bd);
}

void ImGui_ImplSDLSurface3_RenderDrawData(ImDrawData* draw_data)
{
    ImGui_ImplSDLSurface3_Data* bd = ImGui_ImplSDLSurface3_GetBackendData();
    if (bd == nullptr || draw_data == nullptr || bd->TargetSurface == nullptr)
        return;

    if (SDL_MUSTLOCK(bd->TargetSurface) && !SDL_LockSurface(bd->TargetSurface))
        return;

    ImGui_ImplSDLSurface3_InitSurfaceInfo(&bd->TargetInfo, bd->TargetSurface);
    if (bd->TargetInfo.Width <= 0 || bd->TargetInfo.Height <= 0)
    {
        if (SDL_MUSTLOCK(bd->TargetSurface))
            SDL_UnlockSurface(bd->TargetSurface);
        return;
    }

    const ImVec2 clip_off = draw_data->DisplayPos;
    const ImVec2 clip_scale = draw_data->FramebufferScale;

    for (int n = 0; n < draw_data->CmdListsCount; n++)
    {
        const ImDrawList* draw_list = draw_data->CmdLists[n];
        const ImDrawVert* vtx_buffer = draw_list->VtxBuffer.Data;
        const ImDrawIdx* idx_buffer = draw_list->IdxBuffer.Data;

        for (int cmd_i = 0; cmd_i < draw_list->CmdBuffer.Size; cmd_i++)
        {
            const ImDrawCmd* pcmd = &draw_list->CmdBuffer[cmd_i];
            if (pcmd->UserCallback != nullptr)
            {
                if (pcmd->UserCallback != ImDrawCallback_ResetRenderState)
                    pcmd->UserCallback(draw_list, pcmd);
                continue;
            }
            if (pcmd->ElemCount == 0)
                continue;

            ImVec2 clip_min((pcmd->ClipRect.x - clip_off.x) * clip_scale.x, (pcmd->ClipRect.y - clip_off.y) * clip_scale.y);
            ImVec2 clip_max((pcmd->ClipRect.z - clip_off.x) * clip_scale.x, (pcmd->ClipRect.w - clip_off.y) * clip_scale.y);
            if (clip_min.x < 0.0f) clip_min.x = 0.0f;
            if (clip_min.y < 0.0f) clip_min.y = 0.0f;
            if (clip_max.x > (float)bd->TargetInfo.Width) clip_max.x = (float)bd->TargetInfo.Width;
            if (clip_max.y > (float)bd->TargetInfo.Height) clip_max.y = (float)bd->TargetInfo.Height;
            if (clip_max.x <= clip_min.x || clip_max.y <= clip_min.y)
                continue;

            int clip_min_x = (int)clip_min.x;
            int clip_min_y = (int)clip_min.y;
            int clip_max_x = (int)clip_max.x;
            int clip_max_y = (int)clip_max.y;

            SDL_Surface* texture_surface = (SDL_Surface*)pcmd->GetTexID();
            ImGui_ImplSDLSurface3_SurfaceInfo texture_info;
            ImGui_ImplSDLSurface3_SurfaceInfo* texture_ptr = nullptr;
            bool texture_locked = false;

            if (texture_surface != nullptr)
            {
                if (SDL_MUSTLOCK(texture_surface))
                {
                    if (!SDL_LockSurface(texture_surface))
                        texture_surface = nullptr;
                    else
                        texture_locked = true;
                }
                if (texture_surface != nullptr)
                {
                    ImGui_ImplSDLSurface3_InitSurfaceInfo(&texture_info, texture_surface);
                    if (texture_info.Width > 0 && texture_info.Height > 0)
                        texture_ptr = &texture_info;
                }
            }

            for (unsigned int idx = 0; idx + 2 < (unsigned int)pcmd->ElemCount; idx += 3)
            {
                ImDrawIdx idx0 = idx_buffer[pcmd->IdxOffset + idx + 0];
                ImDrawIdx idx1 = idx_buffer[pcmd->IdxOffset + idx + 1];
                ImDrawIdx idx2 = idx_buffer[pcmd->IdxOffset + idx + 2];

                ImGui_ImplSDLSurface3_RasterVertex v0;
                ImGui_ImplSDLSurface3_RasterVertex v1;
                ImGui_ImplSDLSurface3_RasterVertex v2;
                ImGui_ImplSDLSurface3_SetupRasterVertex(&v0, vtx_buffer[pcmd->VtxOffset + idx0], clip_off, clip_scale);
                ImGui_ImplSDLSurface3_SetupRasterVertex(&v1, vtx_buffer[pcmd->VtxOffset + idx1], clip_off, clip_scale);
                ImGui_ImplSDLSurface3_SetupRasterVertex(&v2, vtx_buffer[pcmd->VtxOffset + idx2], clip_off, clip_scale);

                if (bd->TargetInfo.IsFast)
                    ImGui_ImplSDLSurface3_RenderTriangleFast(bd->TargetInfo, texture_ptr, v0, v1, v2, clip_min_x, clip_min_y, clip_max_x, clip_max_y);
                else
                    ImGui_ImplSDLSurface3_RenderTriangleGeneric(bd->TargetInfo, texture_ptr, v0, v1, v2, clip_min_x, clip_min_y, clip_max_x, clip_max_y);
            }

            if (texture_locked)
                SDL_UnlockSurface(texture_surface);
        }
    }

    if (SDL_MUSTLOCK(bd->TargetSurface))
        SDL_UnlockSurface(bd->TargetSurface);
}

#endif