// dear imgui: CPU framebuffer renderer backend
// (supports caller-owned framebuffers/textures in generic packed pixel formats)
//
// Implemented features:
//  [X] Renderer: User texture binding. Use 'ImGui_ImplCPU_Texture*' as texture identifier.
//  [X] Renderer: Large meshes support (64k+ vertices) even with 16-bit indices (ImGuiBackendFlags_RendererHasVtxOffset).
//  [X] Renderer: Texture updates support for dynamic font atlas (ImGuiBackendFlags_RendererHasTextures).
//
// CHANGELOG
//  2026-03-16: Added generic packed-pixel format support, builtin formats, Alpha8 textures, and parity-oriented generic paths.
//  2026-03-15: Initial version, extracted from the SDL surface software renderer.

#include "imgui.h"
#ifndef IMGUI_DISABLE
#include "imgui_impl_cpu.h"

#include <algorithm>
#include <cmath>
#include <cstring>
#include <stdint.h>

#if defined(__clang__)
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wfloat-equal"
#pragma clang diagnostic ignored "-Wold-style-cast"
#elif defined(__GNUC__)
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif

#if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
#define IMGUI_IMPL_CPU_BIG_ENDIAN 1
#else
#define IMGUI_IMPL_CPU_BIG_ENDIAN 0
#endif

#if IMGUI_IMPL_CPU_BIG_ENDIAN
#define IMGUI_IMPL_CPU_RGBA8888_RMASK 0xFF000000u
#define IMGUI_IMPL_CPU_RGBA8888_GMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_RGBA8888_BMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_RGBA8888_AMASK 0x000000FFu
#define IMGUI_IMPL_CPU_BGRA8888_RMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_BGRA8888_GMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_BGRA8888_BMASK 0xFF000000u
#define IMGUI_IMPL_CPU_BGRA8888_AMASK 0x000000FFu
#define IMGUI_IMPL_CPU_ARGB8888_RMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_ARGB8888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_ARGB8888_BMASK 0x000000FFu
#define IMGUI_IMPL_CPU_ARGB8888_AMASK 0xFF000000u
#define IMGUI_IMPL_CPU_ABGR8888_RMASK 0x000000FFu
#define IMGUI_IMPL_CPU_ABGR8888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_ABGR8888_BMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_ABGR8888_AMASK 0xFF000000u
#define IMGUI_IMPL_CPU_RGB888_RMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_RGB888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_RGB888_BMASK 0x000000FFu
#define IMGUI_IMPL_CPU_BGR888_RMASK 0x000000FFu
#define IMGUI_IMPL_CPU_BGR888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_BGR888_BMASK 0x00FF0000u
#else
#define IMGUI_IMPL_CPU_RGBA8888_RMASK 0x000000FFu
#define IMGUI_IMPL_CPU_RGBA8888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_RGBA8888_BMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_RGBA8888_AMASK 0xFF000000u
#define IMGUI_IMPL_CPU_BGRA8888_RMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_BGRA8888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_BGRA8888_BMASK 0x000000FFu
#define IMGUI_IMPL_CPU_BGRA8888_AMASK 0xFF000000u
#define IMGUI_IMPL_CPU_ARGB8888_RMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_ARGB8888_GMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_ARGB8888_BMASK 0xFF000000u
#define IMGUI_IMPL_CPU_ARGB8888_AMASK 0x000000FFu
#define IMGUI_IMPL_CPU_ABGR8888_RMASK 0xFF000000u
#define IMGUI_IMPL_CPU_ABGR8888_GMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_ABGR8888_BMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_ABGR8888_AMASK 0x000000FFu
#define IMGUI_IMPL_CPU_RGB888_RMASK 0x000000FFu
#define IMGUI_IMPL_CPU_RGB888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_RGB888_BMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_BGR888_RMASK 0x00FF0000u
#define IMGUI_IMPL_CPU_BGR888_GMASK 0x0000FF00u
#define IMGUI_IMPL_CPU_BGR888_BMASK 0x000000FFu
#endif

static const ImGui_ImplCPU_PixelFormat GImGui_ImplCPU_BuiltinFormats[ImGui_ImplCPU_BuiltinFormat_COUNT] =
{
    { 32, 4, IMGUI_IMPL_CPU_RGBA8888_RMASK, IMGUI_IMPL_CPU_RGBA8888_GMASK, IMGUI_IMPL_CPU_RGBA8888_BMASK, IMGUI_IMPL_CPU_RGBA8888_AMASK, 0, 8, 16, 24, 0, 0, 0, 0, nullptr, 0, nullptr },
    { 32, 4, IMGUI_IMPL_CPU_BGRA8888_RMASK, IMGUI_IMPL_CPU_BGRA8888_GMASK, IMGUI_IMPL_CPU_BGRA8888_BMASK, IMGUI_IMPL_CPU_BGRA8888_AMASK, 16, 8, 0, 24, 0, 0, 0, 0, nullptr, 0, nullptr },
    { 32, 4, IMGUI_IMPL_CPU_ARGB8888_RMASK, IMGUI_IMPL_CPU_ARGB8888_GMASK, IMGUI_IMPL_CPU_ARGB8888_BMASK, IMGUI_IMPL_CPU_ARGB8888_AMASK, 8, 16, 24, 0, 0, 0, 0, 0, nullptr, 0, nullptr },
    { 32, 4, IMGUI_IMPL_CPU_ABGR8888_RMASK, IMGUI_IMPL_CPU_ABGR8888_GMASK, IMGUI_IMPL_CPU_ABGR8888_BMASK, IMGUI_IMPL_CPU_ABGR8888_AMASK, 24, 16, 8, 0, 0, 0, 0, 0, nullptr, 0, nullptr },
    { 16, 2, 0xF800u, 0x07E0u, 0x001Fu, 0x0000u, 11, 5, 0, 0, 3, 2, 3, 8, nullptr, 0, nullptr },
    { 16, 2, 0x7C00u, 0x03E0u, 0x001Fu, 0x8000u, 10, 5, 0, 15, 3, 3, 3, 7, nullptr, 0, nullptr },
    { 16, 2, 0xF000u, 0x0F00u, 0x00F0u, 0x000Fu, 12, 8, 4, 0, 4, 4, 4, 4, nullptr, 0, nullptr },
    { 24, 3, IMGUI_IMPL_CPU_RGB888_RMASK, IMGUI_IMPL_CPU_RGB888_GMASK, IMGUI_IMPL_CPU_RGB888_BMASK, 0x00000000u, 0, 8, 16, 0, 0, 0, 0, 8, nullptr, 0, nullptr },
    { 24, 3, IMGUI_IMPL_CPU_BGR888_RMASK, IMGUI_IMPL_CPU_BGR888_GMASK, IMGUI_IMPL_CPU_BGR888_BMASK, 0x00000000u, 16, 8, 0, 0, 0, 0, 0, 8, nullptr, 0, nullptr },
    { 8, 1, 0x00000000u, 0x00000000u, 0x00000000u, 0x000000FFu, 0, 0, 0, 0, 8, 8, 8, 0, nullptr, 0, nullptr }
};

const ImGui_ImplCPU_PixelFormat* ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat format)
{
    if (format < 0 || format >= ImGui_ImplCPU_BuiltinFormat_COUNT)
        return nullptr;
    return &GImGui_ImplCPU_BuiltinFormats[format];
}

enum ImGui_ImplCPU_FormatClass
{
    ImGui_ImplCPU_FormatClass_FastRGBA32,
    ImGui_ImplCPU_FormatClass_FastBGRA32,
    ImGui_ImplCPU_FormatClass_FastARGB32,
    ImGui_ImplCPU_FormatClass_FastABGR32,
    ImGui_ImplCPU_FormatClass_RGB565,
    ImGui_ImplCPU_FormatClass_ARGB1555,
    ImGui_ImplCPU_FormatClass_RGBA4444,
    ImGui_ImplCPU_FormatClass_RGB888,
    ImGui_ImplCPU_FormatClass_BGR888,
    ImGui_ImplCPU_FormatClass_Alpha8,
    ImGui_ImplCPU_FormatClass_Indexed,
    ImGui_ImplCPU_FormatClass_Generic
};

struct ImGui_ImplCPU_FastFormat32
{
    ImU32 Rmask;
    ImU32 Gmask;
    ImU32 Bmask;
    ImU32 Amask;
    ImU8 Rshift;
    ImU8 Gshift;
    ImU8 Bshift;
    ImU8 Ashift;
    bool HasAlpha;
    bool Valid;

    ImGui_ImplCPU_FastFormat32() { memset((void*)this, 0, sizeof(*this)); }
};

struct ImGui_ImplCPU_SurfaceInfo;
typedef void (*ImGui_ImplCPU_ReadRGBAFn)(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a);
typedef unsigned char (*ImGui_ImplCPU_ReadAlphaFn)(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y);
typedef ImU32 (*ImGui_ImplCPU_PackRGBAFn)(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a);

struct ImGui_ImplCPU_SurfaceInfo
{
    unsigned char*               Pixels;
    int                          Pitch;
    int                          Width;
    int                          Height;
    const ImGui_ImplCPU_PixelFormat* Format;
    ImGui_ImplCPU_FormatClass    Class;
    ImGui_ImplCPU_FastFormat32   FastFormat;
    ImGui_ImplCPU_ReadRGBAFn     ReadRGBAFn;
    ImGui_ImplCPU_ReadAlphaFn    ReadAlphaFn;
    ImGui_ImplCPU_PackRGBAFn     PackRGBAFn;
    bool                         IsFast;
    bool                         HasAlpha;
    bool                         IsAlphaOnly;
    bool                         IsIndexed;

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

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

struct ImGui_ImplCPU_Quad
{
    const ImGui_ImplCPU_RasterVertex* TL;
    const ImGui_ImplCPU_RasterVertex* TR;
    const ImGui_ImplCPU_RasterVertex* BR;
    const ImGui_ImplCPU_RasterVertex* BL;
};

struct ImGui_ImplCPU_Data
{
    ImVector<ImGui_ImplCPU_RasterVertex> VertexCache;
};

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

static inline const ImGui_ImplCPU_PixelFormat* ImGui_ImplCPU_GetDefaultPixelFormat()
{
    return ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_RGBA8888);
}

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

static inline bool ImGui_ImplCPU_NearEqual(float a, float b)
{
    return fabsf(a - b) <= 0.0001f;
}

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

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

static inline unsigned char ImGui_ImplCPU_Mul255(unsigned int a, unsigned int b)
{
    unsigned int t = a * b + 128;
    return (unsigned char)((t + (t >> 8)) >> 8);
}

static inline unsigned char ImGui_ImplCPU_ExpandBits(ImU32 value, int bits)
{
    if (bits <= 0)
        return 0;
    if (bits >= 8)
        return (unsigned char)value;
    ImU32 max_value = (1u << bits) - 1u;
    return (unsigned char)((value * 255u + (max_value >> 1)) / max_value);
}

static inline ImU32 ImGui_ImplCPU_CompressBits(unsigned char value, int bits)
{
    if (bits <= 0)
        return 0;
    if (bits >= 8)
        return value;
    ImU32 max_value = (1u << bits) - 1u;
    return (ImU32)((value * max_value + 127u) / 255u);
}

static inline void ImGui_ImplCPU_UnpackCanonicalRGBA(ImU32 pixel, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    *r = (unsigned char)((pixel >> IM_COL32_R_SHIFT) & 0xFF);
    *g = (unsigned char)((pixel >> IM_COL32_G_SHIFT) & 0xFF);
    *b = (unsigned char)((pixel >> IM_COL32_B_SHIFT) & 0xFF);
    *a = (unsigned char)((pixel >> IM_COL32_A_SHIFT) & 0xFF);
}

static inline bool ImGui_ImplCPU_CanUseFastPath(const void* pixels, int pitch)
{
#ifdef IMGUI_IMPL_CPU_DISABLE_FAST_PATH
    IM_UNUSED(pixels);
    IM_UNUSED(pitch);
    return false;
#else
    if (pixels == nullptr || pitch <= 0)
        return false;
    if ((pitch & 3) != 0)
        return false;
    if ((((uintptr_t)pixels) & 3) != 0)
        return false;
    return true;
#endif
}

static inline bool ImGui_ImplCPU_MatchesFormat(const ImGui_ImplCPU_PixelFormat* format, const ImGui_ImplCPU_PixelFormat* builtin)
{
    return format->BitsPerPixel == builtin->BitsPerPixel
        && format->BytesPerPixel == builtin->BytesPerPixel
        && format->Rmask == builtin->Rmask
        && format->Gmask == builtin->Gmask
        && format->Bmask == builtin->Bmask
        && format->Amask == builtin->Amask
        && format->Rshift == builtin->Rshift
        && format->Gshift == builtin->Gshift
        && format->Bshift == builtin->Bshift
        && format->Ashift == builtin->Ashift
        && format->Rloss == builtin->Rloss
        && format->Gloss == builtin->Gloss
        && format->Bloss == builtin->Bloss
        && format->Aloss == builtin->Aloss;
}

static inline ImGui_ImplCPU_FormatClass ImGui_ImplCPU_ClassifyFormat(const ImGui_ImplCPU_PixelFormat* format)
{
    if (format == nullptr)
        format = ImGui_ImplCPU_GetDefaultPixelFormat();

    if (format->Palette != nullptr && format->PaletteSize > 0)
        return ImGui_ImplCPU_FormatClass_Indexed;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_RGBA8888)))
        return ImGui_ImplCPU_FormatClass_FastRGBA32;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_BGRA8888)))
        return ImGui_ImplCPU_FormatClass_FastBGRA32;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_ARGB8888)))
        return ImGui_ImplCPU_FormatClass_FastARGB32;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_ABGR8888)))
        return ImGui_ImplCPU_FormatClass_FastABGR32;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_RGB565)))
        return ImGui_ImplCPU_FormatClass_RGB565;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_ARGB1555)))
        return ImGui_ImplCPU_FormatClass_ARGB1555;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_RGBA4444)))
        return ImGui_ImplCPU_FormatClass_RGBA4444;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_RGB888)))
        return ImGui_ImplCPU_FormatClass_RGB888;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_BGR888)))
        return ImGui_ImplCPU_FormatClass_BGR888;
    if (ImGui_ImplCPU_MatchesFormat(format, ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_Alpha8)))
        return ImGui_ImplCPU_FormatClass_Alpha8;
    return ImGui_ImplCPU_FormatClass_Generic;
}

static inline ImGui_ImplCPU_FastFormat32 ImGui_ImplCPU_BuildFastFormat(const ImGui_ImplCPU_PixelFormat* format)
{
    ImGui_ImplCPU_FastFormat32 out;
    if (format == nullptr)
        return out;
    if (format->BitsPerPixel != 32 || format->BytesPerPixel != 4)
        return out;
    if (format->Rloss != 0 || format->Gloss != 0 || format->Bloss != 0)
        return out;
    if (format->Amask != 0 && format->Aloss != 0)
        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 inline ImU32 ImGui_ImplCPU_ReadPixelRaw(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    const unsigned char* p = surface.Pixels + (size_t)y * (size_t)surface.Pitch + (size_t)x * (size_t)surface.Format->BytesPerPixel;
    switch (surface.Format->BytesPerPixel)
    {
    case 1:
        return *p;
    case 2:
    {
        ImU16 pixel = 0;
        memcpy(&pixel, p, sizeof(pixel));
        return pixel;
    }
    case 3:
#if IMGUI_IMPL_CPU_BIG_ENDIAN
        return (ImU32)((p[0] << 16) | (p[1] << 8) | p[2]);
#else
        return (ImU32)(p[0] | (p[1] << 8) | (p[2] << 16));
#endif
    case 4:
    {
        ImU32 pixel = 0;
        memcpy(&pixel, p, sizeof(pixel));
        return pixel;
    }
    default:
        return 0;
    }
}

static inline void ImGui_ImplCPU_WritePixelRaw(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, ImU32 pixel)
{
    unsigned char* p = surface.Pixels + (size_t)y * (size_t)surface.Pitch + (size_t)x * (size_t)surface.Format->BytesPerPixel;
    switch (surface.Format->BytesPerPixel)
    {
    case 1:
        *p = (unsigned char)pixel;
        return;
    case 2:
    {
        ImU16 v = (ImU16)pixel;
        memcpy(p, &v, sizeof(v));
        return;
    }
    case 3:
#if IMGUI_IMPL_CPU_BIG_ENDIAN
        p[0] = (unsigned char)((pixel >> 16) & 0xFF);
        p[1] = (unsigned char)((pixel >> 8) & 0xFF);
        p[2] = (unsigned char)(pixel & 0xFF);
#else
        p[0] = (unsigned char)(pixel & 0xFF);
        p[1] = (unsigned char)((pixel >> 8) & 0xFF);
        p[2] = (unsigned char)((pixel >> 16) & 0xFF);
#endif
        return;
    case 4:
        memcpy(p, &pixel, sizeof(pixel));
        return;
    default:
        return;
    }
}

static inline void ImGui_ImplCPU_UnpackFastRGBA(const ImGui_ImplCPU_FastFormat32& format, ImU32 pixel, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    *r = (unsigned char)((pixel & format.Rmask) >> format.Rshift);
    *g = (unsigned char)((pixel & format.Gmask) >> format.Gshift);
    *b = (unsigned char)((pixel & format.Bmask) >> format.Bshift);
    *a = format.HasAlpha ? (unsigned char)((pixel & format.Amask) >> format.Ashift) : 255;
}

static inline ImU32 ImGui_ImplCPU_PackFastRGBA(const ImGui_ImplCPU_FastFormat32& format, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    ImU32 pixel = (((ImU32)r << format.Rshift) & format.Rmask)
        | (((ImU32)g << format.Gshift) & format.Gmask)
        | (((ImU32)b << format.Bshift) & format.Bmask);
    if (format.HasAlpha)
        pixel |= (((ImU32)a << format.Ashift) & format.Amask);
    return pixel;
}

static inline void ImGui_ImplCPU_UnpackGenericRGBA(const ImGui_ImplCPU_PixelFormat* format, ImU32 pixel, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    *r = (format->Rmask != 0) ? ImGui_ImplCPU_ExpandBits((pixel & format->Rmask) >> format->Rshift, 8 - format->Rloss) : 0;
    *g = (format->Gmask != 0) ? ImGui_ImplCPU_ExpandBits((pixel & format->Gmask) >> format->Gshift, 8 - format->Gloss) : 0;
    *b = (format->Bmask != 0) ? ImGui_ImplCPU_ExpandBits((pixel & format->Bmask) >> format->Bshift, 8 - format->Bloss) : 0;
    *a = (format->Amask != 0) ? ImGui_ImplCPU_ExpandBits((pixel & format->Amask) >> format->Ashift, 8 - format->Aloss) : 255;
}

static inline ImU32 ImGui_ImplCPU_PackGenericRGBA(const ImGui_ImplCPU_PixelFormat* format, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    ImU32 pixel = 0;
    if (format->Rmask != 0)
        pixel |= (ImGui_ImplCPU_CompressBits(r, 8 - format->Rloss) << format->Rshift) & format->Rmask;
    if (format->Gmask != 0)
        pixel |= (ImGui_ImplCPU_CompressBits(g, 8 - format->Gloss) << format->Gshift) & format->Gmask;
    if (format->Bmask != 0)
        pixel |= (ImGui_ImplCPU_CompressBits(b, 8 - format->Bloss) << format->Bshift) & format->Bmask;
    if (format->Amask != 0)
        pixel |= (ImGui_ImplCPU_CompressBits(a, 8 - format->Aloss) << format->Ashift) & format->Amask;
    return pixel;
}

static inline ImU32 ImGui_ImplCPU_MapNearestPaletteColor(const ImGui_ImplCPU_PixelFormat* format, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    if (format->Palette == nullptr || format->PaletteSize <= 0)
        return 0;

    unsigned int best_distance = 0xFFFFFFFFu;
    int best_index = 0;
    for (int i = 0; i < format->PaletteSize; i++)
    {
        unsigned char pr, pg, pb, pa;
        ImGui_ImplCPU_UnpackCanonicalRGBA(format->Palette[i], &pr, &pg, &pb, &pa);
        int dr = (int)pr - (int)r;
        int dg = (int)pg - (int)g;
        int db = (int)pb - (int)b;
        int da = (int)pa - (int)a;
        unsigned int distance = (unsigned int)(dr * dr + dg * dg + db * db + da * da);
        if (distance < best_distance)
        {
            best_distance = distance;
            best_index = i;
            if (distance == 0)
                break;
        }
    }
    return (ImU32)best_index;
}

static inline void ImGui_ImplCPU_ReadRGBAFast32(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    ImGui_ImplCPU_UnpackFastRGBA(surface.FastFormat, ImGui_ImplCPU_ReadPixelRaw(surface, x, y), r, g, b, a);
}

static inline unsigned char ImGui_ImplCPU_ReadAlphaFast32(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    if (!surface.FastFormat.HasAlpha)
        return 255;
    ImU32 pixel = ImGui_ImplCPU_ReadPixelRaw(surface, x, y);
    return (unsigned char)((pixel & surface.FastFormat.Amask) >> surface.FastFormat.Ashift);
}

static inline ImU32 ImGui_ImplCPU_PackFast32(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    return ImGui_ImplCPU_PackFastRGBA(surface.FastFormat, r, g, b, a);
}

static inline void ImGui_ImplCPU_ReadRGBAGenericSurface(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    ImGui_ImplCPU_UnpackGenericRGBA(surface.Format, ImGui_ImplCPU_ReadPixelRaw(surface, x, y), r, g, b, a);
}

static inline unsigned char ImGui_ImplCPU_ReadAlphaGenericSurface(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    if (surface.Format->Amask == 0)
        return 255;
    ImU32 pixel = ImGui_ImplCPU_ReadPixelRaw(surface, x, y);
    return ImGui_ImplCPU_ExpandBits((pixel & surface.Format->Amask) >> surface.Format->Ashift, 8 - surface.Format->Aloss);
}

static inline ImU32 ImGui_ImplCPU_PackGenericSurface(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    return ImGui_ImplCPU_PackGenericRGBA(surface.Format, r, g, b, a);
}

static inline void ImGui_ImplCPU_ReadRGBAAlpha8Surface(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    IM_UNUSED(surface);
    IM_UNUSED(x);
    IM_UNUSED(y);
    *r = 255;
    *g = 255;
    *b = 255;
    *a = surface.Pixels[(size_t)y * (size_t)surface.Pitch + (size_t)x];
}

static inline unsigned char ImGui_ImplCPU_ReadAlphaAlpha8Surface(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    return surface.Pixels[(size_t)y * (size_t)surface.Pitch + (size_t)x];
}

static inline ImU32 ImGui_ImplCPU_PackAlpha8Surface(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    IM_UNUSED(surface);
    IM_UNUSED(r);
    IM_UNUSED(g);
    IM_UNUSED(b);
    return a;
}

static inline void ImGui_ImplCPU_ReadRGBAIndexedSurface(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    ImU32 index = ImGui_ImplCPU_ReadPixelRaw(surface, x, y);
    if (surface.Format->Palette == nullptr || index >= (ImU32)surface.Format->PaletteSize)
    {
        *r = *g = *b = *a = 0;
        return;
    }
    ImGui_ImplCPU_UnpackCanonicalRGBA(surface.Format->Palette[index], r, g, b, a);
}

static inline unsigned char ImGui_ImplCPU_ReadAlphaIndexedSurface(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    ImU32 index = ImGui_ImplCPU_ReadPixelRaw(surface, x, y);
    if (surface.Format->Palette == nullptr || index >= (ImU32)surface.Format->PaletteSize)
        return 0;
    return (unsigned char)((surface.Format->Palette[index] >> IM_COL32_A_SHIFT) & 0xFF);
}

static inline ImU32 ImGui_ImplCPU_PackIndexedSurface(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    if (surface.Format->MapColorFn != nullptr)
        return surface.Format->MapColorFn(surface.Format, r, g, b, a);
    return ImGui_ImplCPU_MapNearestPaletteColor(surface.Format, r, g, b, a);
}

static inline void ImGui_ImplCPU_ReadRGBARGB565(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    ImU32 pixel = ImGui_ImplCPU_ReadPixelRaw(surface, x, y);
    *r = ImGui_ImplCPU_ExpandBits((pixel >> 11) & 0x1Fu, 5);
    *g = ImGui_ImplCPU_ExpandBits((pixel >> 5) & 0x3Fu, 6);
    *b = ImGui_ImplCPU_ExpandBits(pixel & 0x1Fu, 5);
    *a = 255;
}

static inline unsigned char ImGui_ImplCPU_ReadAlphaOpaqueSurface(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    IM_UNUSED(surface);
    IM_UNUSED(x);
    IM_UNUSED(y);
    return 255;
}

static inline ImU32 ImGui_ImplCPU_PackRGB565(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    IM_UNUSED(surface);
    IM_UNUSED(a);
    return (ImU32)((ImGui_ImplCPU_CompressBits(r, 5) << 11) | (ImGui_ImplCPU_CompressBits(g, 6) << 5) | ImGui_ImplCPU_CompressBits(b, 5));
}

static inline void ImGui_ImplCPU_ReadRGBAARGB1555(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    ImU32 pixel = ImGui_ImplCPU_ReadPixelRaw(surface, x, y);
    *r = ImGui_ImplCPU_ExpandBits((pixel >> 10) & 0x1Fu, 5);
    *g = ImGui_ImplCPU_ExpandBits((pixel >> 5) & 0x1Fu, 5);
    *b = ImGui_ImplCPU_ExpandBits(pixel & 0x1Fu, 5);
    *a = (pixel & 0x8000u) ? 255 : 0;
}

static inline unsigned char ImGui_ImplCPU_ReadAlphaARGB1555(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    return (ImGui_ImplCPU_ReadPixelRaw(surface, x, y) & 0x8000u) ? 255 : 0;
}

static inline ImU32 ImGui_ImplCPU_PackARGB1555(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    IM_UNUSED(surface);
    return (ImU32)(((a >= 128 ? 1u : 0u) << 15) | (ImGui_ImplCPU_CompressBits(r, 5) << 10) | (ImGui_ImplCPU_CompressBits(g, 5) << 5) | ImGui_ImplCPU_CompressBits(b, 5));
}

static inline void ImGui_ImplCPU_ReadRGBARGBA4444(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    ImU32 pixel = ImGui_ImplCPU_ReadPixelRaw(surface, x, y);
    *r = ImGui_ImplCPU_ExpandBits((pixel >> 12) & 0x0Fu, 4);
    *g = ImGui_ImplCPU_ExpandBits((pixel >> 8) & 0x0Fu, 4);
    *b = ImGui_ImplCPU_ExpandBits((pixel >> 4) & 0x0Fu, 4);
    *a = ImGui_ImplCPU_ExpandBits(pixel & 0x0Fu, 4);
}

static inline unsigned char ImGui_ImplCPU_ReadAlphaRGBA4444(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    return ImGui_ImplCPU_ExpandBits(ImGui_ImplCPU_ReadPixelRaw(surface, x, y) & 0x0Fu, 4);
}

static inline ImU32 ImGui_ImplCPU_PackRGBA4444(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    IM_UNUSED(surface);
    return (ImU32)((ImGui_ImplCPU_CompressBits(r, 4) << 12) | (ImGui_ImplCPU_CompressBits(g, 4) << 8) | (ImGui_ImplCPU_CompressBits(b, 4) << 4) | ImGui_ImplCPU_CompressBits(a, 4));
}

static inline void ImGui_ImplCPU_ReadRGBARGB888(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    const unsigned char* p = surface.Pixels + (size_t)y * (size_t)surface.Pitch + (size_t)x * 3u;
#if IMGUI_IMPL_CPU_BIG_ENDIAN
    *r = p[0];
    *g = p[1];
    *b = p[2];
#else
    *r = p[0];
    *g = p[1];
    *b = p[2];
#endif
    *a = 255;
}

static inline ImU32 ImGui_ImplCPU_PackRGB888(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    IM_UNUSED(surface);
    IM_UNUSED(a);
#if IMGUI_IMPL_CPU_BIG_ENDIAN
    return (ImU32)((r << 16) | (g << 8) | b);
#else
    return (ImU32)(r | (g << 8) | (b << 16));
#endif
}

static inline void ImGui_ImplCPU_ReadRGBABGR888(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    const unsigned char* p = surface.Pixels + (size_t)y * (size_t)surface.Pitch + (size_t)x * 3u;
#if IMGUI_IMPL_CPU_BIG_ENDIAN
    *b = p[0];
    *g = p[1];
    *r = p[2];
#else
    *b = p[0];
    *g = p[1];
    *r = p[2];
#endif
    *a = 255;
}

static inline ImU32 ImGui_ImplCPU_PackBGR888(const ImGui_ImplCPU_SurfaceInfo& surface, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    IM_UNUSED(surface);
    IM_UNUSED(a);
#if IMGUI_IMPL_CPU_BIG_ENDIAN
    return (ImU32)((b << 16) | (g << 8) | r);
#else
    return (ImU32)(b | (g << 8) | (r << 16));
#endif
}

static void ImGui_ImplCPU_InitSurfaceInfo(ImGui_ImplCPU_SurfaceInfo* out, const unsigned char* pixels, int pitch, int width, int height, const ImGui_ImplCPU_PixelFormat* format)
{
    out->Pixels = (unsigned char*)pixels;
    out->Pitch = pitch;
    out->Width = width;
    out->Height = height;
    out->Format = format ? format : ImGui_ImplCPU_GetDefaultPixelFormat();
    out->Class = ImGui_ImplCPU_ClassifyFormat(out->Format);
    out->FastFormat = ImGui_ImplCPU_BuildFastFormat(out->Format);
    out->HasAlpha = out->Format->Amask != 0 || out->Class == ImGui_ImplCPU_FormatClass_Alpha8;
    out->IsAlphaOnly = out->Class == ImGui_ImplCPU_FormatClass_Alpha8;
    out->IsIndexed = out->Class == ImGui_ImplCPU_FormatClass_Indexed;
    out->IsFast = (out->Class == ImGui_ImplCPU_FormatClass_FastRGBA32
        || out->Class == ImGui_ImplCPU_FormatClass_FastBGRA32
        || out->Class == ImGui_ImplCPU_FormatClass_FastARGB32
        || out->Class == ImGui_ImplCPU_FormatClass_FastABGR32)
        && out->FastFormat.Valid
        && ImGui_ImplCPU_CanUseFastPath(out->Pixels, out->Pitch);

    switch (out->Class)
    {
    case ImGui_ImplCPU_FormatClass_FastRGBA32:
    case ImGui_ImplCPU_FormatClass_FastBGRA32:
    case ImGui_ImplCPU_FormatClass_FastARGB32:
    case ImGui_ImplCPU_FormatClass_FastABGR32:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBAFast32;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaFast32;
        out->PackRGBAFn = ImGui_ImplCPU_PackFast32;
        break;
    case ImGui_ImplCPU_FormatClass_RGB565:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBARGB565;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaOpaqueSurface;
        out->PackRGBAFn = ImGui_ImplCPU_PackRGB565;
        break;
    case ImGui_ImplCPU_FormatClass_ARGB1555:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBAARGB1555;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaARGB1555;
        out->PackRGBAFn = ImGui_ImplCPU_PackARGB1555;
        break;
    case ImGui_ImplCPU_FormatClass_RGBA4444:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBARGBA4444;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaRGBA4444;
        out->PackRGBAFn = ImGui_ImplCPU_PackRGBA4444;
        break;
    case ImGui_ImplCPU_FormatClass_RGB888:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBARGB888;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaOpaqueSurface;
        out->PackRGBAFn = ImGui_ImplCPU_PackRGB888;
        break;
    case ImGui_ImplCPU_FormatClass_BGR888:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBABGR888;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaOpaqueSurface;
        out->PackRGBAFn = ImGui_ImplCPU_PackBGR888;
        break;
    case ImGui_ImplCPU_FormatClass_Alpha8:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBAAlpha8Surface;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaAlpha8Surface;
        out->PackRGBAFn = ImGui_ImplCPU_PackAlpha8Surface;
        break;
    case ImGui_ImplCPU_FormatClass_Indexed:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBAIndexedSurface;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaIndexedSurface;
        out->PackRGBAFn = ImGui_ImplCPU_PackIndexedSurface;
        break;
    case ImGui_ImplCPU_FormatClass_Generic:
    default:
        out->ReadRGBAFn = ImGui_ImplCPU_ReadRGBAGenericSurface;
        out->ReadAlphaFn = ImGui_ImplCPU_ReadAlphaGenericSurface;
        out->PackRGBAFn = ImGui_ImplCPU_PackGenericSurface;
        break;
    }
}

static inline void ImGui_ImplCPU_InitTextureInfo(ImGui_ImplCPU_SurfaceInfo* out, const ImGui_ImplCPU_Texture* texture)
{
    ImGui_ImplCPU_InitSurfaceInfo(out,
        texture ? texture->Pixels : nullptr,
        texture ? texture->Pitch : 0,
        texture ? texture->Width : 0,
        texture ? texture->Height : 0,
        texture ? texture->Format : nullptr);
}

static inline void ImGui_ImplCPU_InitBufferInfo(ImGui_ImplCPU_SurfaceInfo* out, ImGui_ImplCPU_Framebuffer* framebuffer)
{
    ImGui_ImplCPU_InitSurfaceInfo(out,
        framebuffer ? framebuffer->Pixels : nullptr,
        framebuffer ? framebuffer->Pitch : 0,
        framebuffer ? framebuffer->Width : 0,
        framebuffer ? framebuffer->Height : 0,
        framebuffer ? framebuffer->Format : nullptr);
}

static inline void ImGui_ImplCPU_ReadPixelRGBA(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y, unsigned char* r, unsigned char* g, unsigned char* b, unsigned char* a)
{
    surface.ReadRGBAFn(surface, x, y, r, g, b, a);
}

static inline unsigned char ImGui_ImplCPU_ReadPixelAlpha(const ImGui_ImplCPU_SurfaceInfo& surface, int x, int y)
{
    return surface.ReadAlphaFn(surface, x, y);
}

static inline void ImGui_ImplCPU_BlendPixelFast(const ImGui_ImplCPU_SurfaceInfo& target, ImU32* dst, unsigned char sr, unsigned char sg, unsigned char sb, unsigned char sa)
{
    if (sa == 0)
        return;
    if (sa == 255)
    {
        *dst = ImGui_ImplCPU_PackFastRGBA(target.FastFormat, sr, sg, sb, target.FastFormat.HasAlpha ? 255 : 255);
        return;
    }

    unsigned char dr, dg, db, da;
    ImGui_ImplCPU_UnpackFastRGBA(target.FastFormat, *dst, &dr, &dg, &db, &da);
    unsigned char inv_a = (unsigned char)(255 - sa);
    unsigned char out_r = (unsigned char)(ImGui_ImplCPU_Mul255(sr, sa) + ImGui_ImplCPU_Mul255(dr, inv_a));
    unsigned char out_g = (unsigned char)(ImGui_ImplCPU_Mul255(sg, sa) + ImGui_ImplCPU_Mul255(dg, inv_a));
    unsigned char out_b = (unsigned char)(ImGui_ImplCPU_Mul255(sb, sa) + ImGui_ImplCPU_Mul255(db, inv_a));
    unsigned char out_a = target.FastFormat.HasAlpha ? (unsigned char)(sa + ImGui_ImplCPU_Mul255(da, inv_a)) : 255;
    *dst = ImGui_ImplCPU_PackFastRGBA(target.FastFormat, out_r, out_g, out_b, out_a);
}

static inline void ImGui_ImplCPU_BlendPixelGeneric(const ImGui_ImplCPU_SurfaceInfo& target, int x, int y, unsigned char sr, unsigned char sg, unsigned char sb, unsigned char sa)
{
    if (sa == 0)
        return;
    if (sa == 255)
    {
        ImU32 pixel = target.PackRGBAFn(target, sr, sg, sb, target.HasAlpha ? 255 : 255);
        ImGui_ImplCPU_WritePixelRaw(target, x, y, pixel);
        return;
    }

    unsigned char dr, dg, db, da;
    ImGui_ImplCPU_ReadPixelRGBA(target, x, y, &dr, &dg, &db, &da);
    unsigned char inv_a = (unsigned char)(255 - sa);
    unsigned char out_r = (unsigned char)(ImGui_ImplCPU_Mul255(sr, sa) + ImGui_ImplCPU_Mul255(dr, inv_a));
    unsigned char out_g = (unsigned char)(ImGui_ImplCPU_Mul255(sg, sa) + ImGui_ImplCPU_Mul255(dg, inv_a));
    unsigned char out_b = (unsigned char)(ImGui_ImplCPU_Mul255(sb, sa) + ImGui_ImplCPU_Mul255(db, inv_a));
    unsigned char out_a = target.HasAlpha ? (unsigned char)(sa + ImGui_ImplCPU_Mul255(da, inv_a)) : 255;
    ImU32 pixel = target.PackRGBAFn(target, out_r, out_g, out_b, out_a);
    ImGui_ImplCPU_WritePixelRaw(target, x, y, pixel);
}

static inline bool ImGui_ImplCPU_SameColor(const ImGui_ImplCPU_RasterVertex& a, const ImGui_ImplCPU_RasterVertex& b)
{
    return a.R == b.R && a.G == b.G && a.B == b.B && a.A == b.A;
}

static inline void ImGui_ImplCPU_SetupRasterVertex(ImGui_ImplCPU_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_ImplCPU_BuildVertexCache(ImVector<ImGui_ImplCPU_RasterVertex>* out, const ImDrawList* draw_list, const ImVec2& display_pos, const ImVec2& framebuffer_scale)
{
    out->resize(draw_list->VtxBuffer.Size);
    for (int vtx_i = 0; vtx_i < draw_list->VtxBuffer.Size; vtx_i++)
        ImGui_ImplCPU_SetupRasterVertex(&out->Data[vtx_i], draw_list->VtxBuffer[vtx_i], display_pos, framebuffer_scale);
}

static void ImGui_ImplCPU_RenderTriangleFast(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_SurfaceInfo* texture, bool texture_is_alpha, const ImGui_ImplCPU_RasterVertex& v0, const ImGui_ImplCPU_RasterVertex& v1, const ImGui_ImplCPU_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_ImplCPU_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_ImplCPU_Edge(v1.X, v1.Y, v2.X, v2.Y, start_x, start_y) * sign;
    float e1_row = ImGui_ImplCPU_Edge(v2.X, v2.Y, v0.X, v0.Y, start_x, start_y) * sign;
    float e2_row = ImGui_ImplCPU_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;

    if (texture == nullptr)
    {
        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;
            ImU32* dst = (ImU32*)(void*)(target.Pixels + (size_t)y * (size_t)target.Pitch) + min_x;

            for (int x = min_x; x <= max_x; x++)
            {
                if (e0 >= 0.0f && e1 >= 0.0f && e2 >= 0.0f)
                    ImGui_ImplCPU_BlendPixelFast(target, dst, ImGui_ImplCPU_ClampByte(r), ImGui_ImplCPU_ClampByte(g), ImGui_ImplCPU_ClampByte(b), ImGui_ImplCPU_ClampByte(a));

                dst++;
                e0 += e0_dx;
                e1 += e1_dx;
                e2 += e2_dx;
                r += dr_dx;
                g += dg_dx;
                b += db_dx;
                a += da_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;
        }
        return;
    }

    if (texture_is_alpha)
    {
        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;
            ImU32* dst = (ImU32*)(void*)(target.Pixels + (size_t)y * (size_t)target.Pitch) + min_x;

            for (int x = min_x; x <= max_x; x++)
            {
                if (e0 >= 0.0f && e1 >= 0.0f && e2 >= 0.0f)
                {
                    int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                    int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                    ImGui_ImplCPU_BlendPixelFast(target, dst,
                        ImGui_ImplCPU_ClampByte(r),
                        ImGui_ImplCPU_ClampByte(g),
                        ImGui_ImplCPU_ClampByte(b),
                        ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(a), ImGui_ImplCPU_ReadPixelAlpha(*texture, tx, ty)));
                }

                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;
        }
        return;
    }

    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;
        ImU32* dst = (ImU32*)(void*)(target.Pixels + (size_t)y * (size_t)target.Pitch) + min_x;

        for (int x = min_x; x <= max_x; x++)
        {
            if (e0 >= 0.0f && e1 >= 0.0f && e2 >= 0.0f)
            {
                int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                unsigned char tr, tg, tb, ta;
                ImGui_ImplCPU_ReadPixelRGBA(*texture, tx, ty, &tr, &tg, &tb, &ta);
                ImGui_ImplCPU_BlendPixelFast(target, dst,
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(r), tr),
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(g), tg),
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(b), tb),
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(a), ta));
            }

            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_ImplCPU_RenderTriangleGeneric(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_SurfaceInfo* texture, bool texture_is_alpha, const ImGui_ImplCPU_RasterVertex& v0, const ImGui_ImplCPU_RasterVertex& v1, const ImGui_ImplCPU_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_ImplCPU_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_ImplCPU_Edge(v1.X, v1.Y, v2.X, v2.Y, start_x, start_y) * sign;
    float e1_row = ImGui_ImplCPU_Edge(v2.X, v2.Y, v0.X, v0.Y, start_x, start_y) * sign;
    float e2_row = ImGui_ImplCPU_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;

    if (texture == nullptr)
    {
        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;

            for (int x = min_x; x <= max_x; x++)
            {
                if (e0 >= 0.0f && e1 >= 0.0f && e2 >= 0.0f)
                    ImGui_ImplCPU_BlendPixelGeneric(target, x, y, ImGui_ImplCPU_ClampByte(r), ImGui_ImplCPU_ClampByte(g), ImGui_ImplCPU_ClampByte(b), ImGui_ImplCPU_ClampByte(a));

                e0 += e0_dx;
                e1 += e1_dx;
                e2 += e2_dx;
                r += dr_dx;
                g += dg_dx;
                b += db_dx;
                a += da_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;
        }
        return;
    }

    if (texture_is_alpha)
    {
        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)
                {
                    int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                    int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                    ImGui_ImplCPU_BlendPixelGeneric(target, x, y,
                        ImGui_ImplCPU_ClampByte(r),
                        ImGui_ImplCPU_ClampByte(g),
                        ImGui_ImplCPU_ClampByte(b),
                        ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(a), ImGui_ImplCPU_ReadPixelAlpha(*texture, tx, ty)));
                }

                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;
        }
        return;
    }

    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)
            {
                int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                unsigned char tr, tg, tb, ta;
                ImGui_ImplCPU_ReadPixelRGBA(*texture, tx, ty, &tr, &tg, &tb, &ta);
                ImGui_ImplCPU_BlendPixelGeneric(target, x, y,
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(r), tr),
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(g), tg),
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(b), tb),
                    ImGui_ImplCPU_Mul255(ImGui_ImplCPU_ClampByte(a), ta));
            }

            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 bool ImGui_ImplCPU_DetectQuad(const ImGui_ImplCPU_RasterVertex* vertices, const ImDrawIdx* idx_buffer, int vtx_offset, unsigned int elem_offset, ImGui_ImplCPU_Quad* out)
{
    int unique_indices[4];
    int unique_count = 0;
    for (int i = 0; i < 6; i++)
    {
        int idx = vtx_offset + idx_buffer[elem_offset + i];
        bool found = false;
        for (int j = 0; j < unique_count; j++)
            if (unique_indices[j] == idx)
                found = true;
        if (!found)
        {
            if (unique_count >= 4)
                return false;
            unique_indices[unique_count++] = idx;
        }
    }
    if (unique_count != 4)
        return false;

    const ImGui_ImplCPU_RasterVertex* candidates[4];
    for (int i = 0; i < 4; i++)
        candidates[i] = &vertices[unique_indices[i]];

    float min_x = candidates[0]->X;
    float max_x = candidates[0]->X;
    float min_y = candidates[0]->Y;
    float max_y = candidates[0]->Y;
    for (int i = 1; i < 4; i++)
    {
        min_x = std::min(min_x, candidates[i]->X);
        max_x = std::max(max_x, candidates[i]->X);
        min_y = std::min(min_y, candidates[i]->Y);
        max_y = std::max(max_y, candidates[i]->Y);
    }
    if (!(max_x > min_x) || !(max_y > min_y))
        return false;

    out->TL = nullptr;
    out->TR = nullptr;
    out->BR = nullptr;
    out->BL = nullptr;
    for (int i = 0; i < 4; i++)
    {
        const ImGui_ImplCPU_RasterVertex* v = candidates[i];
        if (ImGui_ImplCPU_NearEqual(v->X, min_x) && ImGui_ImplCPU_NearEqual(v->Y, min_y))
            out->TL = v;
        else if (ImGui_ImplCPU_NearEqual(v->X, max_x) && ImGui_ImplCPU_NearEqual(v->Y, min_y))
            out->TR = v;
        else if (ImGui_ImplCPU_NearEqual(v->X, max_x) && ImGui_ImplCPU_NearEqual(v->Y, max_y))
            out->BR = v;
        else if (ImGui_ImplCPU_NearEqual(v->X, min_x) && ImGui_ImplCPU_NearEqual(v->Y, max_y))
            out->BL = v;
        else
            return false;
    }
    return out->TL != nullptr && out->TR != nullptr && out->BR != nullptr && out->BL != nullptr;
}

static bool ImGui_ImplCPU_QuadHasConstantColor(const ImGui_ImplCPU_Quad& quad)
{
    return ImGui_ImplCPU_SameColor(*quad.TL, *quad.TR) && ImGui_ImplCPU_SameColor(*quad.TL, *quad.BR) && ImGui_ImplCPU_SameColor(*quad.TL, *quad.BL);
}

static bool ImGui_ImplCPU_QuadHasLinearUV(const ImGui_ImplCPU_Quad& quad)
{
    if (!ImGui_ImplCPU_NearEqual(quad.TL->Y, quad.TR->Y) || !ImGui_ImplCPU_NearEqual(quad.BL->Y, quad.BR->Y) || !ImGui_ImplCPU_NearEqual(quad.TL->X, quad.BL->X) || !ImGui_ImplCPU_NearEqual(quad.TR->X, quad.BR->X))
        return false;
    float u = quad.TL->U + (quad.TR->U - quad.TL->U) + (quad.BL->U - quad.TL->U);
    float v = quad.TL->V + (quad.TR->V - quad.TL->V) + (quad.BL->V - quad.TL->V);
    return ImGui_ImplCPU_NearEqual(u, quad.BR->U) && ImGui_ImplCPU_NearEqual(v, quad.BR->V);
}

static bool ImGui_ImplCPU_QuadGetConstantTexel(const ImGui_ImplCPU_SurfaceInfo& texture, const ImGui_ImplCPU_Quad& quad, int* out_x, int* out_y)
{
    int max_x = texture.Width - 1;
    int max_y = texture.Height - 1;
    int tx = ImGui_ImplCPU_ClampInt((int)(quad.TL->U * (float)max_x + 0.5f), 0, max_x);
    int ty = ImGui_ImplCPU_ClampInt((int)(quad.TL->V * (float)max_y + 0.5f), 0, max_y);
    int corner_tx[3] =
    {
        ImGui_ImplCPU_ClampInt((int)(quad.TR->U * (float)max_x + 0.5f), 0, max_x),
        ImGui_ImplCPU_ClampInt((int)(quad.BR->U * (float)max_x + 0.5f), 0, max_x),
        ImGui_ImplCPU_ClampInt((int)(quad.BL->U * (float)max_x + 0.5f), 0, max_x)
    };
    int corner_ty[3] =
    {
        ImGui_ImplCPU_ClampInt((int)(quad.TR->V * (float)max_y + 0.5f), 0, max_y),
        ImGui_ImplCPU_ClampInt((int)(quad.BR->V * (float)max_y + 0.5f), 0, max_y),
        ImGui_ImplCPU_ClampInt((int)(quad.BL->V * (float)max_y + 0.5f), 0, max_y)
    };
    for (int i = 0; i < 3; i++)
        if (corner_tx[i] != tx || corner_ty[i] != ty)
            return false;
    *out_x = tx;
    *out_y = ty;
    return true;
}

static void ImGui_ImplCPU_RenderSolidRectFast(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_Quad& quad, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    int min_x = std::max((int)ceilf(quad.TL->X - 0.5f), clip_min_x);
    int max_x = std::min((int)floorf(quad.TR->X - 0.5f), clip_max_x - 1);
    int min_y = std::max((int)ceilf(quad.TL->Y - 0.5f), clip_min_y);
    int max_y = std::min((int)floorf(quad.BL->Y - 0.5f), clip_max_y - 1);
    if (min_x > max_x || min_y > max_y)
        return;

    int count = max_x - min_x + 1;
    if (a == 255)
    {
        ImU32 pixel = ImGui_ImplCPU_PackFastRGBA(target.FastFormat, r, g, b, target.FastFormat.HasAlpha ? 255 : 255);
        for (int y = min_y; y <= max_y; y++)
        {
            ImU32* dst = (ImU32*)(void*)(target.Pixels + (size_t)y * (size_t)target.Pitch) + min_x;
            std::fill_n(dst, count, pixel);
        }
        return;
    }

    for (int y = min_y; y <= max_y; y++)
    {
        ImU32* dst = (ImU32*)(void*)(target.Pixels + (size_t)y * (size_t)target.Pitch) + min_x;
        for (int x = 0; x < count; x++)
            ImGui_ImplCPU_BlendPixelFast(target, dst + x, r, g, b, a);
    }
}

static void ImGui_ImplCPU_RenderSolidRectGeneric(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_Quad& quad, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y, unsigned char r, unsigned char g, unsigned char b, unsigned char a)
{
    int min_x = std::max((int)ceilf(quad.TL->X - 0.5f), clip_min_x);
    int max_x = std::min((int)floorf(quad.TR->X - 0.5f), clip_max_x - 1);
    int min_y = std::max((int)ceilf(quad.TL->Y - 0.5f), clip_min_y);
    int max_y = std::min((int)floorf(quad.BL->Y - 0.5f), clip_max_y - 1);
    if (min_x > max_x || min_y > max_y)
        return;

    ImU32 opaque_pixel = 0;
    if (a == 255)
        opaque_pixel = target.PackRGBAFn(target, r, g, b, target.HasAlpha ? 255 : 255);

    for (int y = min_y; y <= max_y; y++)
        for (int x = min_x; x <= max_x; x++)
            if (a == 255)
                ImGui_ImplCPU_WritePixelRaw(target, x, y, opaque_pixel);
            else
                ImGui_ImplCPU_BlendPixelGeneric(target, x, y, r, g, b, a);
}

static void ImGui_ImplCPU_RenderTexturedRectFast(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_SurfaceInfo& texture, bool texture_is_alpha, const ImGui_ImplCPU_Quad& quad, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y)
{
    int min_x = std::max((int)ceilf(quad.TL->X - 0.5f), clip_min_x);
    int max_x = std::min((int)floorf(quad.TR->X - 0.5f), clip_max_x - 1);
    int min_y = std::max((int)ceilf(quad.TL->Y - 0.5f), clip_min_y);
    int max_y = std::min((int)floorf(quad.BL->Y - 0.5f), clip_max_y - 1);
    if (min_x > max_x || min_y > max_y)
        return;

    unsigned char color_r = (unsigned char)quad.TL->R;
    unsigned char color_g = (unsigned char)quad.TL->G;
    unsigned char color_b = (unsigned char)quad.TL->B;
    unsigned char color_a = (unsigned char)quad.TL->A;
    float du_dx = (quad.TR->U - quad.TL->U) / (quad.TR->X - quad.TL->X);
    float dv_dx = (quad.TR->V - quad.TL->V) / (quad.TR->X - quad.TL->X);
    float du_dy = (quad.BL->U - quad.TL->U) / (quad.BL->Y - quad.TL->Y);
    float dv_dy = (quad.BL->V - quad.TL->V) / (quad.BL->Y - quad.TL->Y);
    float u_const = quad.TL->U - quad.TL->X * du_dx - quad.TL->Y * du_dy;
    float v_const = quad.TL->V - quad.TL->X * dv_dx - quad.TL->Y * dv_dy;
    int texture_max_x = texture.Width - 1;
    int texture_max_y = texture.Height - 1;

    for (int y = min_y; y <= max_y; y++)
    {
        float py = (float)y + 0.5f;
        float px = (float)min_x + 0.5f;
        float u = u_const + px * du_dx + py * du_dy;
        float v = v_const + px * dv_dx + py * dv_dy;
        ImU32* dst = (ImU32*)(void*)(target.Pixels + (size_t)y * (size_t)target.Pitch) + min_x;

        if (texture_is_alpha)
        {
            for (int x = min_x; x <= max_x; x++)
            {
                int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                ImGui_ImplCPU_BlendPixelFast(target, dst, color_r, color_g, color_b, ImGui_ImplCPU_Mul255(color_a, ImGui_ImplCPU_ReadPixelAlpha(texture, tx, ty)));
                dst++;
                u += du_dx;
                v += dv_dx;
            }
        }
        else
        {
            for (int x = min_x; x <= max_x; x++)
            {
                int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                unsigned char tr, tg, tb, ta;
                ImGui_ImplCPU_ReadPixelRGBA(texture, tx, ty, &tr, &tg, &tb, &ta);
                ImGui_ImplCPU_BlendPixelFast(target, dst,
                    ImGui_ImplCPU_Mul255(color_r, tr),
                    ImGui_ImplCPU_Mul255(color_g, tg),
                    ImGui_ImplCPU_Mul255(color_b, tb),
                    ImGui_ImplCPU_Mul255(color_a, ta));
                dst++;
                u += du_dx;
                v += dv_dx;
            }
        }
    }
}

static void ImGui_ImplCPU_RenderTexturedRectGeneric(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_SurfaceInfo& texture, bool texture_is_alpha, const ImGui_ImplCPU_Quad& quad, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y)
{
    int min_x = std::max((int)ceilf(quad.TL->X - 0.5f), clip_min_x);
    int max_x = std::min((int)floorf(quad.TR->X - 0.5f), clip_max_x - 1);
    int min_y = std::max((int)ceilf(quad.TL->Y - 0.5f), clip_min_y);
    int max_y = std::min((int)floorf(quad.BL->Y - 0.5f), clip_max_y - 1);
    if (min_x > max_x || min_y > max_y)
        return;

    unsigned char color_r = (unsigned char)quad.TL->R;
    unsigned char color_g = (unsigned char)quad.TL->G;
    unsigned char color_b = (unsigned char)quad.TL->B;
    unsigned char color_a = (unsigned char)quad.TL->A;
    float du_dx = (quad.TR->U - quad.TL->U) / (quad.TR->X - quad.TL->X);
    float dv_dx = (quad.TR->V - quad.TL->V) / (quad.TR->X - quad.TL->X);
    float du_dy = (quad.BL->U - quad.TL->U) / (quad.BL->Y - quad.TL->Y);
    float dv_dy = (quad.BL->V - quad.TL->V) / (quad.BL->Y - quad.TL->Y);
    float u_const = quad.TL->U - quad.TL->X * du_dx - quad.TL->Y * du_dy;
    float v_const = quad.TL->V - quad.TL->X * dv_dx - quad.TL->Y * dv_dy;
    int texture_max_x = texture.Width - 1;
    int texture_max_y = texture.Height - 1;

    for (int y = min_y; y <= max_y; y++)
    {
        float py = (float)y + 0.5f;
        float px = (float)min_x + 0.5f;
        float u = u_const + px * du_dx + py * du_dy;
        float v = v_const + px * dv_dx + py * dv_dy;

        if (texture_is_alpha)
        {
            for (int x = min_x; x <= max_x; x++)
            {
                int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                ImGui_ImplCPU_BlendPixelGeneric(target, x, y, color_r, color_g, color_b, ImGui_ImplCPU_Mul255(color_a, ImGui_ImplCPU_ReadPixelAlpha(texture, tx, ty)));
                u += du_dx;
                v += dv_dx;
            }
        }
        else
        {
            for (int x = min_x; x <= max_x; x++)
            {
                int tx = ImGui_ImplCPU_ClampInt((int)(u * (float)texture_max_x + 0.5f), 0, texture_max_x);
                int ty = ImGui_ImplCPU_ClampInt((int)(v * (float)texture_max_y + 0.5f), 0, texture_max_y);
                unsigned char tr, tg, tb, ta;
                ImGui_ImplCPU_ReadPixelRGBA(texture, tx, ty, &tr, &tg, &tb, &ta);
                ImGui_ImplCPU_BlendPixelGeneric(target, x, y,
                    ImGui_ImplCPU_Mul255(color_r, tr),
                    ImGui_ImplCPU_Mul255(color_g, tg),
                    ImGui_ImplCPU_Mul255(color_b, tb),
                    ImGui_ImplCPU_Mul255(color_a, ta));
                u += du_dx;
                v += dv_dx;
            }
        }
    }
}

static bool ImGui_ImplCPU_TryRenderQuadFast(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_SurfaceInfo* texture, bool texture_is_alpha, const ImGui_ImplCPU_Quad& quad, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y)
{
    if (!ImGui_ImplCPU_QuadHasConstantColor(quad))
        return false;

    unsigned char r = (unsigned char)quad.TL->R;
    unsigned char g = (unsigned char)quad.TL->G;
    unsigned char b = (unsigned char)quad.TL->B;
    unsigned char a = (unsigned char)quad.TL->A;
    if (texture == nullptr)
    {
        ImGui_ImplCPU_RenderSolidRectFast(target, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y, r, g, b, a);
        return true;
    }

    int tx = 0;
    int ty = 0;
    if (ImGui_ImplCPU_QuadGetConstantTexel(*texture, quad, &tx, &ty))
    {
        if (texture_is_alpha)
            a = ImGui_ImplCPU_Mul255(a, ImGui_ImplCPU_ReadPixelAlpha(*texture, tx, ty));
        else
        {
            unsigned char tr, tg, tb, ta;
            ImGui_ImplCPU_ReadPixelRGBA(*texture, tx, ty, &tr, &tg, &tb, &ta);
            r = ImGui_ImplCPU_Mul255(r, tr);
            g = ImGui_ImplCPU_Mul255(g, tg);
            b = ImGui_ImplCPU_Mul255(b, tb);
            a = ImGui_ImplCPU_Mul255(a, ta);
        }
        ImGui_ImplCPU_RenderSolidRectFast(target, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y, r, g, b, a);
        return true;
    }

    if (!ImGui_ImplCPU_QuadHasLinearUV(quad))
        return false;
    ImGui_ImplCPU_RenderTexturedRectFast(target, *texture, texture_is_alpha, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y);
    return true;
}

static bool ImGui_ImplCPU_TryRenderQuadGeneric(const ImGui_ImplCPU_SurfaceInfo& target, const ImGui_ImplCPU_SurfaceInfo* texture, bool texture_is_alpha, const ImGui_ImplCPU_Quad& quad, int clip_min_x, int clip_min_y, int clip_max_x, int clip_max_y)
{
    if (!ImGui_ImplCPU_QuadHasConstantColor(quad))
        return false;

    unsigned char r = (unsigned char)quad.TL->R;
    unsigned char g = (unsigned char)quad.TL->G;
    unsigned char b = (unsigned char)quad.TL->B;
    unsigned char a = (unsigned char)quad.TL->A;
    if (texture == nullptr)
    {
        ImGui_ImplCPU_RenderSolidRectGeneric(target, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y, r, g, b, a);
        return true;
    }

    int tx = 0;
    int ty = 0;
    if (ImGui_ImplCPU_QuadGetConstantTexel(*texture, quad, &tx, &ty))
    {
        if (texture_is_alpha)
            a = ImGui_ImplCPU_Mul255(a, ImGui_ImplCPU_ReadPixelAlpha(*texture, tx, ty));
        else
        {
            unsigned char tr, tg, tb, ta;
            ImGui_ImplCPU_ReadPixelRGBA(*texture, tx, ty, &tr, &tg, &tb, &ta);
            r = ImGui_ImplCPU_Mul255(r, tr);
            g = ImGui_ImplCPU_Mul255(g, tg);
            b = ImGui_ImplCPU_Mul255(b, tb);
            a = ImGui_ImplCPU_Mul255(a, ta);
        }
        ImGui_ImplCPU_RenderSolidRectGeneric(target, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y, r, g, b, a);
        return true;
    }

    if (!ImGui_ImplCPU_QuadHasLinearUV(quad))
        return false;
    ImGui_ImplCPU_RenderTexturedRectGeneric(target, *texture, texture_is_alpha, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y);
    return true;
}

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

    ImGui_ImplCPU_Data* bd = IM_NEW(ImGui_ImplCPU_Data)();
    io.BackendRendererUserData = (void*)bd;
    io.BackendRendererName = "imgui_impl_cpu";
    io.BackendFlags |= ImGuiBackendFlags_RendererHasVtxOffset;
    io.BackendFlags |= ImGuiBackendFlags_RendererHasTextures;
    return true;
}

void ImGui_ImplCPU_UpdateTexture(ImTextureData* tex)
{
    const ImGui_ImplCPU_PixelFormat* format = nullptr;
    if (tex->Format == ImTextureFormat_RGBA32)
        format = ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_RGBA8888);
    else if (tex->Format == ImTextureFormat_Alpha8)
        format = ImGui_ImplCPU_GetBuiltinPixelFormat(ImGui_ImplCPU_BuiltinFormat_Alpha8);
    else
        IM_ASSERT(false && "Unsupported ImTextureData format for imgui_impl_cpu");

    if (tex->Status == ImTextureStatus_WantCreate)
    {
        IM_ASSERT(tex->TexID == ImTextureID_Invalid && tex->BackendUserData == nullptr);

        ImGui_ImplCPU_Texture* cpu_texture = IM_NEW(ImGui_ImplCPU_Texture)();
        cpu_texture->Pixels = (const unsigned char*)tex->GetPixels();
        cpu_texture->Width = tex->Width;
        cpu_texture->Height = tex->Height;
        cpu_texture->Pitch = tex->GetPitch();
        cpu_texture->Format = format;
        tex->BackendUserData = cpu_texture;
        tex->SetTexID((ImTextureID)(intptr_t)cpu_texture);
        tex->SetStatus(ImTextureStatus_OK);
    }
    else if (tex->Status == ImTextureStatus_WantUpdates)
    {
        ImGui_ImplCPU_Texture* cpu_texture = (ImGui_ImplCPU_Texture*)tex->BackendUserData;
        IM_ASSERT(cpu_texture != nullptr);
        cpu_texture->Pixels = (const unsigned char*)tex->GetPixels();
        cpu_texture->Width = tex->Width;
        cpu_texture->Height = tex->Height;
        cpu_texture->Pitch = tex->GetPitch();
        cpu_texture->Format = format;
        tex->SetStatus(ImTextureStatus_OK);
    }
    else if (tex->Status == ImTextureStatus_WantDestroy)
    {
        ImGui_ImplCPU_Texture* cpu_texture = (ImGui_ImplCPU_Texture*)tex->BackendUserData;
        if (cpu_texture != nullptr)
            IM_DELETE(cpu_texture);
        tex->BackendUserData = nullptr;
        tex->SetTexID(ImTextureID_Invalid);
        tex->SetStatus(ImTextureStatus_Destroyed);
    }
}

static void ImGui_ImplCPU_DestroyDeviceObjects()
{
    for (ImTextureData* tex : ImGui::GetPlatformIO().Textures)
        if (tex->RefCount == 1)
        {
            tex->SetStatus(ImTextureStatus_WantDestroy);
            ImGui_ImplCPU_UpdateTexture(tex);
        }
}

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

    ImGui_ImplCPU_DestroyDeviceObjects();

    io.BackendRendererName = nullptr;
    io.BackendRendererUserData = nullptr;
    io.BackendFlags &= ~(ImGuiBackendFlags_RendererHasVtxOffset | ImGuiBackendFlags_RendererHasTextures);
    platform_io.ClearRendererHandlers();
    IM_DELETE(bd);
}

void ImGui_ImplCPU_NewFrame()
{
    ImGui_ImplCPU_Data* bd = ImGui_ImplCPU_GetBackendData();
    IM_ASSERT(bd != nullptr && "Context or backend not initialized! Did you call ImGui_ImplCPU_Init()?");
    IM_UNUSED(bd);
}

void ImGui_ImplCPU_RenderDrawData(ImDrawData* draw_data, ImGui_ImplCPU_Framebuffer* framebuffer)
{
    ImGui_ImplCPU_Data* bd = ImGui_ImplCPU_GetBackendData();
    if (bd == nullptr || draw_data == nullptr)
        return;

    if (draw_data->Textures != nullptr)
        for (ImTextureData* tex : *draw_data->Textures)
            if (tex->Status != ImTextureStatus_OK)
                ImGui_ImplCPU_UpdateTexture(tex);

    if (framebuffer == nullptr || framebuffer->Pixels == nullptr || framebuffer->Width <= 0 || framebuffer->Height <= 0 || framebuffer->Pitch <= 0)
        return;

    ImGui_ImplCPU_SurfaceInfo target;
    ImGui_ImplCPU_InitBufferInfo(&target, framebuffer);
    if (target.Format == nullptr || target.Format->BytesPerPixel <= 0)
        return;

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

    const ImGui_ImplCPU_Texture* active_texture = nullptr;
    ImGui_ImplCPU_SurfaceInfo active_texture_info;
    bool active_texture_valid = false;
    bool active_texture_is_alpha = false;

    for (int n = 0; n < draw_data->CmdListsCount; n++)
    {
        const ImDrawList* draw_list = draw_data->CmdLists[n];
        const ImDrawIdx* idx_buffer = draw_list->IdxBuffer.Data;
        ImGui_ImplCPU_BuildVertexCache(&bd->VertexCache, draw_list, clip_off, clip_scale);
        const ImGui_ImplCPU_RasterVertex* vertex_cache = bd->VertexCache.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)target.Width) clip_max.x = (float)target.Width;
            if (clip_max.y > (float)target.Height) clip_max.y = (float)target.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;

            const ImGui_ImplCPU_Texture* texture = (const ImGui_ImplCPU_Texture*)(intptr_t)pcmd->GetTexID();
            if (texture != active_texture)
            {
                active_texture = texture;
                active_texture_valid = false;
                active_texture_is_alpha = false;
                if (texture != nullptr)
                {
                    ImGui_ImplCPU_InitTextureInfo(&active_texture_info, texture);
                    active_texture_valid = active_texture_info.Pixels != nullptr && active_texture_info.Width > 0 && active_texture_info.Height > 0 && active_texture_info.Pitch > 0 && active_texture_info.Format != nullptr;
                    active_texture_is_alpha = active_texture_valid && active_texture_info.IsAlphaOnly;
                }
            }

            const ImGui_ImplCPU_SurfaceInfo* texture_ptr = active_texture_valid ? &active_texture_info : nullptr;
            for (unsigned int idx = 0; idx + 2 < (unsigned int)pcmd->ElemCount;)
            {
                if (idx + 5 < (unsigned int)pcmd->ElemCount)
                {
                    ImGui_ImplCPU_Quad quad;
                    if (ImGui_ImplCPU_DetectQuad(vertex_cache, idx_buffer, pcmd->VtxOffset, pcmd->IdxOffset + idx, &quad))
                    {
                        bool rendered = target.IsFast
                            ? ImGui_ImplCPU_TryRenderQuadFast(target, texture_ptr, active_texture_is_alpha, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y)
                            : ImGui_ImplCPU_TryRenderQuadGeneric(target, texture_ptr, active_texture_is_alpha, quad, clip_min_x, clip_min_y, clip_max_x, clip_max_y);
                        if (rendered)
                        {
                            idx += 6;
                            continue;
                        }
                    }
                }

                ImDrawIdx idx0 = idx_buffer[pcmd->IdxOffset + idx + 0];
                ImDrawIdx idx1 = idx_buffer[pcmd->IdxOffset + idx + 1];
                ImDrawIdx idx2 = idx_buffer[pcmd->IdxOffset + idx + 2];
                const ImGui_ImplCPU_RasterVertex& v0 = vertex_cache[pcmd->VtxOffset + idx0];
                const ImGui_ImplCPU_RasterVertex& v1 = vertex_cache[pcmd->VtxOffset + idx1];
                const ImGui_ImplCPU_RasterVertex& v2 = vertex_cache[pcmd->VtxOffset + idx2];
                if (target.IsFast)
                    ImGui_ImplCPU_RenderTriangleFast(target, texture_ptr, active_texture_is_alpha, v0, v1, v2, clip_min_x, clip_min_y, clip_max_x, clip_max_y);
                else
                    ImGui_ImplCPU_RenderTriangleGeneric(target, texture_ptr, active_texture_is_alpha, v0, v1, v2, clip_min_x, clip_min_y, clip_max_x, clip_max_y);
                idx += 3;
            }
        }
    }
}

#if defined(__clang__)
#pragma clang diagnostic pop
#endif

#endif // #ifndef IMGUI_DISABLE