#include <math.h>
#include <stdlib.h>
#include <string.h>

#include <SDL3/SDL.h>

#include "pxl8_ase.h"
#include "pxl8_blit.h"
#include "pxl8_font.h"
#include "pxl8_gfx.h"
#include "pxl8_macros.h"
#include "pxl8_math.h"
#include "pxl8_types.h"

static inline void pxl8_color_unpack(u32 color, u8* r, u8* g, u8* b, u8* a) {
    *r = color & 0xFF;
    *g = (color >> 8) & 0xFF;
    *b = (color >> 16) & 0xFF;
    *a = (color >> 24) & 0xFF;
}

static inline u32 pxl8_color_pack(u8 r, u8 g, u8 b, u8 a) {
    return r | (g << 8) | (b << 16) | (a << 24);
}

static inline u8 pxl8_color_lerp_channel(u8 c1, u8 c2, f32 t) {
    return c1 + (i32)((c2 - c1) * t);
}

pxl8_viewport pxl8_gfx_viewport(pxl8_bounds bounds, i32 width, i32 height) {
    pxl8_viewport vp = {0};
    vp.scale = fminf(bounds.w / (f32)width, bounds.h / (f32)height);
    vp.scaled_width = (i32)(width * vp.scale);
    vp.scaled_height = (i32)(height * vp.scale);
    vp.offset_x = (bounds.w - vp.scaled_width) / 2;
    vp.offset_y = (bounds.h - vp.scaled_height) / 2;
    return vp;
}

typedef struct pxl8_atlas_entry {
    char path[256];
    u32 sprite_id;
    i32 x, y, w, h;
} pxl8_atlas_entry;

struct pxl8_gfx {
    SDL_Renderer* renderer;
    SDL_Texture* framebuffer_texture;
    SDL_Texture* sprite_atlas_texture;
    SDL_Window* window;

    u8* atlas;
    bool atlas_dirty;
    pxl8_atlas_entry* atlas_entries;
    u32 atlas_entries_cap;
    u32 atlas_entries_len;

    pxl8_color_mode color_mode;
    u8* framebuffer;
    i32 framebuffer_height;
    i32 framebuffer_width;
    bool initialized;

    u32* palette;
    u32 palette_size;

    u32 sprite_atlas_height;
    u32 sprite_atlas_width;
    u32 sprite_frame_height;
    u32 sprite_frame_width;
    u32 sprite_frames_per_row;

    pxl8_viewport viewport;

    bool backface_culling;
    pxl8_mat4 model;
    pxl8_mat4 projection;
    pxl8_mat4 view;
    bool wireframe;
    f32* zbuffer;
    i32 zbuffer_height;
    i32 zbuffer_width;
};

static u32 pxl8_get_palette_size(pxl8_color_mode mode) {
    switch (mode) {
        case PXL8_COLOR_MODE_HICOLOR:   return 0;
        case PXL8_COLOR_MODE_FAMI:      return 64;
        case PXL8_COLOR_MODE_MEGA:      return 512;
        case PXL8_COLOR_MODE_GBA:       return 32768;
        case PXL8_COLOR_MODE_SUPERFAMI: return 32768;
        default:                        return 256;
    }
}

void pxl8_gfx_get_resolution_dimensions(pxl8_resolution resolution, i32* width, i32* height) {
    switch (resolution) {
        case PXL8_RESOLUTION_240x160:
            *width = 240; *height = 160; break;
        case PXL8_RESOLUTION_320x180:
            *width = 320; *height = 180; break;
        case PXL8_RESOLUTION_320x240:
            *width = 320; *height = 240; break;
        case PXL8_RESOLUTION_640x360:
            *width = 640; *height = 360; break;
        case PXL8_RESOLUTION_640x480:
            *width = 640; *height = 480; break;
        case PXL8_RESOLUTION_800x600:
            *width = 800; *height = 600; break;
        case PXL8_RESOLUTION_960x540:
            *width = 960; *height = 540; break;
        default:
            *width = 640; *height = 360; break;
    }
}

pxl8_bounds pxl8_gfx_get_bounds(pxl8_gfx* gfx) {
    pxl8_bounds bounds = {0};
    if (!gfx || !gfx->window) {
        return bounds;
    }
    SDL_GetWindowPosition(gfx->window, &bounds.x, &bounds.y);
    SDL_GetWindowSize(gfx->window, &bounds.w, &bounds.h);
    return bounds;
}

pxl8_color_mode pxl8_gfx_get_color_mode(pxl8_gfx* gfx) {
    return gfx ? gfx->color_mode : PXL8_COLOR_MODE_FAMI;
}

u8* pxl8_gfx_get_framebuffer(pxl8_gfx* gfx) {
    return gfx ? gfx->framebuffer : NULL;
}

i32 pxl8_gfx_get_height(const pxl8_gfx* gfx) {
    return gfx ? gfx->framebuffer_height : 0;
}

i32 pxl8_gfx_get_width(const pxl8_gfx* gfx) {
    return gfx ? gfx->framebuffer_width : 0;
}

pxl8_gfx* pxl8_gfx_create(pxl8_color_mode mode, pxl8_resolution resolution, const char* title, i32 window_width, i32 window_height) {
    pxl8_gfx* gfx = (pxl8_gfx*)SDL_calloc(1, sizeof(*gfx));
    if (!gfx) {
        pxl8_error("Failed to allocate graphics context");
        return NULL;
    }

    gfx->color_mode = mode;
    pxl8_gfx_get_resolution_dimensions(resolution, &gfx->framebuffer_width, &gfx->framebuffer_height);

    gfx->window = SDL_CreateWindow(
        title,
        window_width, window_height,
        SDL_WINDOW_RESIZABLE
    );

    if (!gfx->window) {
        pxl8_error("Failed to create window: %s", SDL_GetError());
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    gfx->renderer = SDL_CreateRenderer(gfx->window, NULL);
    if (!gfx->renderer) {
        pxl8_error("Failed to create renderer: %s", SDL_GetError());
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    SDL_SetRenderLogicalPresentation(gfx->renderer,
                                     gfx->framebuffer_width,
                                     gfx->framebuffer_height,
                                     SDL_LOGICAL_PRESENTATION_INTEGER_SCALE);

    i32 bytes_per_pixel = (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) ? 4 : 1;
    i32 fb_size = gfx->framebuffer_width * gfx->framebuffer_height * bytes_per_pixel;
    gfx->framebuffer = (u8*)SDL_calloc(1, fb_size);
    if (!gfx->framebuffer) {
        pxl8_error("Failed to allocate framebuffer");
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    gfx->framebuffer_texture = SDL_CreateTexture(
        gfx->renderer,
        SDL_PIXELFORMAT_RGBA32,
        SDL_TEXTUREACCESS_STREAMING,
        gfx->framebuffer_width,
        gfx->framebuffer_height
    );

    SDL_SetTextureScaleMode(gfx->framebuffer_texture, SDL_SCALEMODE_NEAREST);

    if (!gfx->framebuffer_texture) {
        pxl8_error("Failed to create framebuffer texture: %s", SDL_GetError());
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    gfx->palette_size = pxl8_get_palette_size(mode);
    if (gfx->palette_size > 0) {
        gfx->palette = (u32*)SDL_calloc(gfx->palette_size, sizeof(u32));
        if (!gfx->palette) {
            pxl8_error("Failed to allocate palette");
            pxl8_gfx_destroy(gfx);
            return NULL;
        }

        for (u32 i = 0; i < 256 && i < gfx->palette_size; i++) {
            u8 gray = (u8)(i * 255 / 255);
            gfx->palette[i] = 0xFF000000 | (gray << 16) | (gray << 8) | gray;
        }
    }

    gfx->viewport.offset_x = 0;
    gfx->viewport.offset_y = 0;
    gfx->viewport.scaled_width = gfx->framebuffer_width;
    gfx->viewport.scaled_height = gfx->framebuffer_height;
    gfx->viewport.scale = 1.0f;

    gfx->backface_culling = true;
    gfx->model = pxl8_mat4_identity();
    gfx->projection = pxl8_mat4_identity();
    gfx->view = pxl8_mat4_identity();
    gfx->wireframe = false;
    gfx->zbuffer = NULL;
    gfx->zbuffer_height = 0;
    gfx->zbuffer_width = 0;

    gfx->initialized = true;
    return gfx;
}

void pxl8_gfx_destroy(pxl8_gfx* gfx) {
    if (!gfx) return;
    
    if (gfx->framebuffer_texture) {
        SDL_DestroyTexture(gfx->framebuffer_texture);
        gfx->framebuffer_texture = NULL;
    }
    
    if (gfx->sprite_atlas_texture) {
        SDL_DestroyTexture(gfx->sprite_atlas_texture);
        gfx->sprite_atlas_texture = NULL;
    }
    
    if (gfx->renderer) {
        SDL_DestroyRenderer(gfx->renderer);
        gfx->renderer = NULL;
    }
    
    if (gfx->window) {
        SDL_DestroyWindow(gfx->window);
        gfx->window = NULL;
    }
    
    SDL_free(gfx->framebuffer);
    SDL_free(gfx->palette);
    SDL_free(gfx->atlas);
    SDL_free(gfx->atlas_entries);
    SDL_free(gfx->zbuffer);

    SDL_free(gfx);
}

pxl8_result pxl8_gfx_init_atlas(pxl8_gfx* gfx, u32 width, u32 height) {
    if (!gfx || !gfx->initialized) return PXL8_ERROR_INVALID_ARGUMENT;
    
    gfx->sprite_atlas_width = width;
    gfx->sprite_atlas_height = height;
    
    i32 bytes_per_pixel = (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) ? 4 : 1;
    gfx->atlas = (u8*)SDL_calloc(width * height, bytes_per_pixel);
    if (!gfx->atlas) {
        return PXL8_ERROR_OUT_OF_MEMORY;
    }
    
    gfx->sprite_atlas_texture = SDL_CreateTexture(
        gfx->renderer,
        SDL_PIXELFORMAT_RGBA32,
        SDL_TEXTUREACCESS_STREAMING,
        width,
        height
    );
    
    if (!gfx->sprite_atlas_texture) {
        SDL_free(gfx->atlas);
        gfx->atlas = NULL;
        return PXL8_ERROR_SYSTEM_FAILURE;
    }
    
    gfx->atlas_entries_cap = 256;
    gfx->atlas_entries = (pxl8_atlas_entry*)SDL_calloc(gfx->atlas_entries_cap, sizeof(pxl8_atlas_entry));
    if (!gfx->atlas_entries) {
        SDL_DestroyTexture(gfx->sprite_atlas_texture);
        SDL_free(gfx->atlas);
        return PXL8_ERROR_OUT_OF_MEMORY;
    }
    
    gfx->sprite_frames_per_row = width / 128;
    if (gfx->sprite_frames_per_row == 0) gfx->sprite_frames_per_row = 1;
    
    return PXL8_OK;
}

pxl8_result pxl8_gfx_load_sprite(pxl8_gfx* gfx, const char* path) {
    if (!gfx || !gfx->initialized || !path) return PXL8_ERROR_INVALID_ARGUMENT;
    if (!gfx->atlas) {
        pxl8_error("Atlas not initialized");
        return PXL8_ERROR_NOT_INITIALIZED;
    }
    
    for (u32 i = 0; i < gfx->atlas_entries_len; i++) {
        if (strcmp(gfx->atlas_entries[i].path, path) == 0) {
            return PXL8_OK;
        }
    }
    
    pxl8_ase_file ase_file;
    pxl8_result result = pxl8_ase_load(path, &ase_file);
    if (result != PXL8_OK) {
        pxl8_error("Failed to load ASE file: %s", path);
        return result;
    }
    
    if (ase_file.frame_count == 0) {
        pxl8_error("No frames in ASE file");
        pxl8_ase_destroy(&ase_file);
        return PXL8_ERROR_INVALID_FORMAT;
    }
    
    u32 sprite_w = ase_file.header.width;
    u32 sprite_h = ase_file.header.height;
    
    if (gfx->sprite_frame_width == 0 || gfx->sprite_frame_height == 0) {
        gfx->sprite_frame_width = sprite_w;
        gfx->sprite_frame_height = sprite_h;
    }
    
    u32 id = gfx->atlas_entries_len;
    u32 frames_per_row = gfx->sprite_frames_per_row;
    u32 atlas_x = (id % frames_per_row) * gfx->sprite_frame_width;
    u32 atlas_y = (id / frames_per_row) * gfx->sprite_frame_height;
    
    if (atlas_x + sprite_w > gfx->sprite_atlas_width || 
        atlas_y + sprite_h > gfx->sprite_atlas_height) {
        pxl8_error("Sprite doesn't fit in atlas");
        pxl8_ase_destroy(&ase_file);
        return PXL8_ERROR_INVALID_SIZE;
    }
    
    i32 bytes_per_pixel = (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) ? 4 : 1;
    
    for (u32 y = 0; y < sprite_h; y++) {
        for (u32 x = 0; x < sprite_w; x++) {
            u32 frame_idx = y * sprite_w + x;
            u32 atlas_idx = (atlas_y + y) * gfx->sprite_atlas_width + (atlas_x + x);
            
            if (bytes_per_pixel == 4) {
                ((u32*)gfx->atlas)[atlas_idx] = ((u32*)ase_file.frames[0].pixels)[frame_idx];
            } else {
                gfx->atlas[atlas_idx] = ase_file.frames[0].pixels[frame_idx];
            }
        }
    }
    
    if (gfx->atlas_entries_len >= gfx->atlas_entries_cap) {
        gfx->atlas_entries_cap *= 2;
        pxl8_atlas_entry* new_entries = (pxl8_atlas_entry*)SDL_realloc(
            gfx->atlas_entries, 
            gfx->atlas_entries_cap * sizeof(pxl8_atlas_entry)
        );
        if (!new_entries) {
            pxl8_ase_destroy(&ase_file);
            return PXL8_ERROR_OUT_OF_MEMORY;
        }
        gfx->atlas_entries = new_entries;
    }
    
    gfx->atlas_entries[id].x = atlas_x;
    gfx->atlas_entries[id].y = atlas_y;
    gfx->atlas_entries[id].w = sprite_w;
    gfx->atlas_entries[id].h = sprite_h;
    gfx->atlas_entries[id].sprite_id = id;
    strncpy(gfx->atlas_entries[id].path, path, sizeof(gfx->atlas_entries[id].path) - 1);
    gfx->atlas_entries[id].path[sizeof(gfx->atlas_entries[id].path) - 1] = '\0';
    gfx->atlas_entries_len++;
    
    gfx->atlas_dirty = true;
    
    pxl8_ase_destroy(&ase_file);
    pxl8_debug("Loaded sprite %u: %ux%u at (%u,%u)", id, sprite_w, sprite_h, atlas_x, atlas_y);
    
    return id;
}

pxl8_result pxl8_gfx_load_palette(pxl8_gfx* gfx, const char* path) {
    if (!gfx || !gfx->initialized || !path) return PXL8_ERROR_INVALID_ARGUMENT;
    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) return PXL8_OK;
    
    pxl8_debug("Loading palette from: %s", path);
    
    pxl8_ase_file ase_file;
    pxl8_result result = pxl8_ase_load(path, &ase_file);
    if (result != PXL8_OK) {
        pxl8_error("Failed to load ASE file for palette: %s", path);
        return result;
    }
    
    if (ase_file.palette.entry_count == 0) {
        pxl8_error("No palette data in ASE file");
        pxl8_ase_destroy(&ase_file);
        return PXL8_ERROR_INVALID_FORMAT;
    }
    
    u32 copy_size = (ase_file.palette.entry_count < gfx->palette_size) ? ase_file.palette.entry_count : gfx->palette_size;
    memcpy(gfx->palette, ase_file.palette.colors, copy_size * sizeof(u32));
    
    if (gfx->color_mode != PXL8_COLOR_MODE_HICOLOR && gfx->framebuffer_texture) {
        SDL_Color colors[256];
        for (u32 i = 0; i < 256 && i < gfx->palette_size; i++) {
            colors[i].r = (gfx->palette[i] >> 16) & 0xFF;
            colors[i].g = (gfx->palette[i] >> 8) & 0xFF;
            colors[i].b = gfx->palette[i] & 0xFF;
            colors[i].a = (gfx->palette[i] >> 24) & 0xFF;
        }
        SDL_SetPaletteColors(SDL_CreateSurfacePalette(NULL), colors, 0, 256);
    }
    
    pxl8_ase_destroy(&ase_file);
    pxl8_debug("Loaded palette with %u colors", copy_size);
    
    return PXL8_OK;
}

pxl8_result pxl8_gfx_load_font_atlas(pxl8_gfx* gfx) {
    (void)gfx;
    return PXL8_OK;
}

static void pxl8_upload_indexed_texture(SDL_Texture* texture, const u8* indexed, const u32* palette, u32 palette_size, i32 width, i32 height, u32 default_color) {
    static u32* rgba_buffer = NULL;
    static size_t buffer_size = 0;
    size_t needed_size = width * height;

    if (buffer_size < needed_size) {
        rgba_buffer = (u32*)SDL_realloc(rgba_buffer, needed_size * 4);
        buffer_size = needed_size;
    }

    if (!rgba_buffer) return;

    for (i32 i = 0; i < width * height; i++) {
        u8 index = indexed[i];
        rgba_buffer[i] = (index < palette_size) ? palette[index] : default_color;
    }

    SDL_UpdateTexture(texture, NULL, rgba_buffer, width * 4);
}

void pxl8_gfx_upload_framebuffer(pxl8_gfx* gfx) {
    if (!gfx || !gfx->initialized || !gfx->framebuffer_texture) return;

    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        SDL_UpdateTexture(gfx->framebuffer_texture, NULL, gfx->framebuffer,
                         gfx->framebuffer_width * 4);
    } else {
        pxl8_upload_indexed_texture(gfx->framebuffer_texture, gfx->framebuffer,
                                    gfx->palette, gfx->palette_size,
                                    gfx->framebuffer_width, gfx->framebuffer_height,
                                    0xFF000000);
    }
}

void pxl8_gfx_upload_atlas(pxl8_gfx* gfx) {
    if (!gfx || !gfx->initialized || !gfx->sprite_atlas_texture || !gfx->atlas_dirty) return;

    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        SDL_UpdateTexture(gfx->sprite_atlas_texture, NULL, gfx->atlas,
                         gfx->sprite_atlas_width * 4);
    } else {
        pxl8_upload_indexed_texture(gfx->sprite_atlas_texture, gfx->atlas,
                                    gfx->palette, gfx->palette_size,
                                    gfx->sprite_atlas_width, gfx->sprite_atlas_height,
                                    0x00000000);
    }

    gfx->atlas_dirty = false;
    pxl8_debug("Atlas uploaded to GPU");
}

void pxl8_gfx_present(pxl8_gfx* gfx) {
    if (!gfx || !gfx->initialized) return;
    
    SDL_SetRenderDrawColor(gfx->renderer, 0, 0, 0, 255);
    SDL_RenderClear(gfx->renderer);
    
    if (gfx->framebuffer_texture) {
        SDL_RenderTexture(gfx->renderer, gfx->framebuffer_texture, NULL, NULL);
    }
    
    SDL_RenderPresent(gfx->renderer);
}

void pxl8_gfx_set_viewport(pxl8_gfx* gfx, pxl8_viewport vp) {
    if (!gfx) return;
    gfx->viewport = vp;
}

void pxl8_gfx_project(pxl8_gfx* gfx, f32 left, f32 right, f32 top, f32 bottom) {
    (void)gfx; (void)left; (void)right; (void)top; (void)bottom;
}

void pxl8_clr(pxl8_gfx* gfx, u32 color) {
    if (!gfx || !gfx->framebuffer) return;
    
    i32 bytes_per_pixel = (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) ? 4 : 1;
    i32 size = gfx->framebuffer_width * gfx->framebuffer_height;
    
    if (bytes_per_pixel == 4) {
        u32* fb32 = (u32*)gfx->framebuffer;
        for (i32 i = 0; i < size; i++) {
            fb32[i] = color;
        }
    } else {
        memset(gfx->framebuffer, color & 0xFF, size);
    }
}

void pxl8_pixel(pxl8_gfx* gfx, i32 x, i32 y, u32 color) {
    if (!gfx || !gfx->framebuffer) return;
    if (x < 0 || x >= gfx->framebuffer_width || y < 0 || y >= gfx->framebuffer_height) return;
    
    i32 idx = y * gfx->framebuffer_width + x;
    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        ((u32*)gfx->framebuffer)[idx] = color;
    } else {
        gfx->framebuffer[idx] = color & 0xFF;
    }
}

u32 pxl8_get_pixel(pxl8_gfx* gfx, i32 x, i32 y) {
    if (!gfx || !gfx->framebuffer) return 0;
    if (x < 0 || x >= gfx->framebuffer_width || y < 0 || y >= gfx->framebuffer_height) return 0;
    
    i32 idx = y * gfx->framebuffer_width + x;
    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        return ((u32*)gfx->framebuffer)[idx];
    } else {
        return gfx->framebuffer[idx];
    }
}

void pxl8_line(pxl8_gfx* gfx, i32 x0, i32 y0, i32 x1, i32 y1, u32 color) {
    if (!gfx) return;
    
    i32 dx = abs(x1 - x0);
    i32 dy = abs(y1 - y0);
    i32 sx = x0 < x1 ? 1 : -1;
    i32 sy = y0 < y1 ? 1 : -1;
    i32 err = dx - dy;
    
    while (1) {
        pxl8_pixel(gfx, x0, y0, color);
        
        if (x0 == x1 && y0 == y1) break;
        
        i32 e2 = 2 * err;
        if (e2 > -dy) {
            err -= dy;
            x0 += sx;
        }
        if (e2 < dx) {
            err += dx;
            y0 += sy;
        }
    }
}

void pxl8_rect(pxl8_gfx* gfx, i32 x, i32 y, i32 w, i32 h, u32 color) {
    if (!gfx) return;
    
    pxl8_line(gfx, x, y, x + w - 1, y, color);
    pxl8_line(gfx, x + w - 1, y, x + w - 1, y + h - 1, color);
    pxl8_line(gfx, x + w - 1, y + h - 1, x, y + h - 1, color);
    pxl8_line(gfx, x, y + h - 1, x, y, color);
}

static inline void pxl8_pixel_unchecked(pxl8_gfx* gfx, i32 x, i32 y, u32 color) {
    i32 idx = y * gfx->framebuffer_width + x;
    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        ((u32*)gfx->framebuffer)[idx] = color;
    } else {
        gfx->framebuffer[idx] = color & 0xFF;
    }
}

void pxl8_rect_fill(pxl8_gfx* gfx, i32 x, i32 y, i32 w, i32 h, u32 color) {
    if (!gfx || !gfx->framebuffer) return;

    i32 x0 = (x < 0) ? 0 : x;
    i32 y0 = (y < 0) ? 0 : y;
    i32 x1 = (x + w > gfx->framebuffer_width) ? gfx->framebuffer_width : x + w;
    i32 y1 = (y + h > gfx->framebuffer_height) ? gfx->framebuffer_height : y + h;

    for (i32 py = y0; py < y1; py++) {
        for (i32 px = x0; px < x1; px++) {
            pxl8_pixel_unchecked(gfx, px, py, color);
        }
    }
}

void pxl8_circle(pxl8_gfx* gfx, i32 cx, i32 cy, i32 radius, u32 color) {
    if (!gfx) return;
    
    i32 x = radius;
    i32 y = 0;
    i32 err = 0;
    
    while (x >= y) {
        pxl8_pixel(gfx, cx + x, cy + y, color);
        pxl8_pixel(gfx, cx + y, cy + x, color);
        pxl8_pixel(gfx, cx - y, cy + x, color);
        pxl8_pixel(gfx, cx - x, cy + y, color);
        pxl8_pixel(gfx, cx - x, cy - y, color);
        pxl8_pixel(gfx, cx - y, cy - x, color);
        pxl8_pixel(gfx, cx + y, cy - x, color);
        pxl8_pixel(gfx, cx + x, cy - y, color);
        
        if (err <= 0) {
            y += 1;
            err += 2 * y + 1;
        } else {
            x -= 1;
            err -= 2 * x + 1;
        }
    }
}

void pxl8_circle_fill(pxl8_gfx* gfx, i32 cx, i32 cy, i32 radius, u32 color) {
    if (!gfx || !gfx->framebuffer) return;

    i32 x0 = (cx - radius < 0) ? -cx : -radius;
    i32 y0 = (cy - radius < 0) ? -cy : -radius;
    i32 x1 = (cx + radius >= gfx->framebuffer_width) ? gfx->framebuffer_width - cx - 1 : radius;
    i32 y1 = (cy + radius >= gfx->framebuffer_height) ? gfx->framebuffer_height - cy - 1 : radius;

    for (i32 y = y0; y <= y1; y++) {
        for (i32 x = x0; x <= x1; x++) {
            if (x * x + y * y <= radius * radius) {
                pxl8_pixel_unchecked(gfx, cx + x, cy + y, color);
            }
        }
    }
}

void pxl8_text(pxl8_gfx* gfx, const char* text, i32 x, i32 y, u32 color) {
    if (!gfx || !text || !gfx->framebuffer) return;

    const pxl8_font* font = &pxl8_default_font;
    i32 cursor_x = x;
    i32 cursor_y = y;

    for (const char* c = text; *c; c++) {
        const pxl8_glyph* glyph = pxl8_font_find_glyph(font, (u32)*c);
        if (!glyph) continue;

        for (i32 gy = 0; gy < glyph->height; gy++) {
            for (i32 gx = 0; gx < glyph->width; gx++) {
                i32 px = cursor_x + gx;
                i32 py = cursor_y + gy;

                if (px < 0 || px >= gfx->framebuffer_width ||
                    py < 0 || py >= gfx->framebuffer_height) continue;

                u8 pixel_bit = 0;

                if (glyph->format == PXL8_FONT_FORMAT_INDEXED) {
                    u8 pixel_byte = glyph->data.indexed[gy];
                    pixel_bit = (pixel_byte >> gx) & 1;
                } else {
                    i32 glyph_idx = gy * 8 + gx;
                    u32 rgba_pixel = glyph->data.rgba[glyph_idx];
                    pixel_bit = ((rgba_pixel >> 24) & 0xFF) > 128 ? 1 : 0;
                }

                if (pixel_bit) {
                    i32 fb_idx = py * gfx->framebuffer_width + px;
                    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
                        ((u32*)gfx->framebuffer)[fb_idx] = color;
                    } else {
                        gfx->framebuffer[fb_idx] = (u8)color;
                    }
                }
            }
        }

        cursor_x += font->default_width;
    }
}


void pxl8_sprite(pxl8_gfx* gfx, u32 sprite_id, i32 x, i32 y, i32 w, i32 h) {
    if (!gfx || !gfx->atlas || !gfx->framebuffer || sprite_id >= gfx->atlas_entries_len) return;
    
    pxl8_atlas_entry* entry = &gfx->atlas_entries[sprite_id];
    
    i32 clip_left = (x < 0) ? -x : 0;
    i32 clip_top = (y < 0) ? -y : 0;
    i32 clip_right = (x + w > gfx->framebuffer_width) ? x + w - gfx->framebuffer_width : 0;
    i32 clip_bottom = (y + h > gfx->framebuffer_height) ? y + h - gfx->framebuffer_height : 0;
    
    i32 draw_width = w - clip_left - clip_right;
    i32 draw_height = h - clip_top - clip_bottom;
    
    if (draw_width <= 0 || draw_height <= 0) return;
    
    i32 dest_x = x + clip_left;
    i32 dest_y = y + clip_top;
    
    bool is_1to1_scale = (w == entry->w && h == entry->h);
    bool is_unclipped = (clip_left == 0 && clip_top == 0 && clip_right == 0 && clip_bottom == 0);
    
    if (is_1to1_scale && is_unclipped && pxl8_is_simd_aligned(w)) {
        const u8* sprite_data = gfx->atlas + entry->y * gfx->sprite_atlas_width + entry->x;
        
        if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
            pxl8_blit_simd_hicolor((u32*)gfx->framebuffer, gfx->framebuffer_width, 
                                   (const u32*)sprite_data, gfx->sprite_atlas_width, x, y, w, h);
        } else {
            pxl8_blit_simd_indexed(gfx->framebuffer, gfx->framebuffer_width, 
                                   sprite_data, gfx->sprite_atlas_width, x, y, w, h);
        }
    } else {
        for (i32 py = 0; py < draw_height; py++) {
            for (i32 px = 0; px < draw_width; px++) {
                i32 src_x = entry->x + ((px + clip_left) * entry->w) / w;
                i32 src_y = entry->y + ((py + clip_top) * entry->h) / h;
                i32 src_idx = src_y * gfx->sprite_atlas_width + src_x;
                i32 dest_idx = (dest_y + py) * gfx->framebuffer_width + (dest_x + px);
                
                if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
                    u32 pixel = ((u32*)gfx->atlas)[src_idx];
                    if (pixel & 0xFF000000) {
                        ((u32*)gfx->framebuffer)[dest_idx] = pixel;
                    }
                } else {
                    u8 pixel = gfx->atlas[src_idx];
                    if (pixel != 0) {
                        gfx->framebuffer[dest_idx] = pixel;
                    }
                }
            }
        }
    }
}

void pxl8_gfx_cycle_palette(pxl8_gfx* gfx, u8 start, u8 count, i32 step) {
    if (!gfx || !gfx->palette || count == 0) return;
    
    u32 temp[256];
    for (u8 i = 0; i < count; i++) {
        temp[i] = gfx->palette[start + i];
    }
    
    for (u8 i = 0; i < count; i++) {
        u8 src_idx = i;
        u8 dst_idx = (i + step) % count;
        gfx->palette[start + dst_idx] = temp[src_idx];
    }
}

void pxl8_gfx_swap_palette(pxl8_gfx* gfx, u8 start, u8 count, u32* new_colors) {
    if (!gfx || !gfx->palette || !new_colors) return;
    
    for (u8 i = 0; i < count && (start + i) < gfx->palette_size; i++) {
        gfx->palette[start + i] = new_colors[i];
    }
}

void pxl8_gfx_fade_palette(pxl8_gfx* gfx, u8 start, u8 count, f32 amount, u32 target_color) {
    if (!gfx || !gfx->palette || count == 0) return;

    if (amount < 0.0f) amount = 0.0f;
    if (amount > 1.0f) amount = 1.0f;

    u8 target_r, target_g, target_b, target_a;
    pxl8_color_unpack(target_color, &target_r, &target_g, &target_b, &target_a);

    for (u8 i = 0; i < count && (start + i) < gfx->palette_size; i++) {
        u8 cur_r, cur_g, cur_b, cur_a;
        pxl8_color_unpack(gfx->palette[start + i], &cur_r, &cur_g, &cur_b, &cur_a);

        u8 new_r = pxl8_color_lerp_channel(cur_r, target_r, amount);
        u8 new_g = pxl8_color_lerp_channel(cur_g, target_g, amount);
        u8 new_b = pxl8_color_lerp_channel(cur_b, target_b, amount);
        u8 new_a = pxl8_color_lerp_channel(cur_a, target_a, amount);

        gfx->palette[start + i] = pxl8_color_pack(new_r, new_g, new_b, new_a);
    }
}

void pxl8_gfx_interpolate_palettes(pxl8_gfx* gfx, u32* palette1, u32* palette2, u8 start, u8 count, f32 t) {
    if (!gfx || !gfx->palette || !palette1 || !palette2 || count == 0) return;

    if (t < 0.0f) t = 0.0f;
    if (t > 1.0f) t = 1.0f;

    for (u8 i = 0; i < count && (start + i) < gfx->palette_size; i++) {
        u8 r1, g1, b1, a1, r2, g2, b2, a2;
        pxl8_color_unpack(palette1[i], &r1, &g1, &b1, &a1);
        pxl8_color_unpack(palette2[i], &r2, &g2, &b2, &a2);

        u8 r = pxl8_color_lerp_channel(r1, r2, t);
        u8 g = pxl8_color_lerp_channel(g1, g2, t);
        u8 b = pxl8_color_lerp_channel(b1, b2, t);
        u8 a = pxl8_color_lerp_channel(a1, a2, t);

        gfx->palette[start + i] = pxl8_color_pack(r, g, b, a);
    }
}

void pxl8_gfx_color_ramp(pxl8_gfx* gfx, u8 start, u8 count, u32 from_color, u32 to_color) {
    if (!gfx || !gfx->palette || count <= 1) return;

    u8 from_r, from_g, from_b, from_a;
    u8 to_r, to_g, to_b, to_a;
    pxl8_color_unpack(from_color, &from_r, &from_g, &from_b, &from_a);
    pxl8_color_unpack(to_color, &to_r, &to_g, &to_b, &to_a);

    for (u8 i = 0; i < count && (start + i) < gfx->palette_size; i++) {
        f32 t = (f32)i / (f32)(count - 1);

        u8 r = pxl8_color_lerp_channel(from_r, to_r, t);
        u8 g = pxl8_color_lerp_channel(from_g, to_g, t);
        u8 b = pxl8_color_lerp_channel(from_b, to_b, t);
        u8 a = pxl8_color_lerp_channel(from_a, to_a, t);

        gfx->palette[start + i] = pxl8_color_pack(r, g, b, a);
    }
}

void pxl8_gfx_process_effects(pxl8_gfx* gfx, pxl8_effects* effects, f32 dt) {
    if (!gfx || !effects) return;

    effects->time += dt;

    for (i32 i = 0; i < 8; i++) {
        pxl8_palette_cycle* cycle = &effects->palette_cycles[i];
        if (!cycle->active) continue;

        cycle->timer += dt * cycle->speed;
        if (cycle->timer >= 1.0f) {
            cycle->timer -= 1.0f;
            i32 count = cycle->end_index - cycle->start_index + 1;
            pxl8_gfx_cycle_palette(gfx, cycle->start_index, count, 1);
        }
    }
}

static bool pxl8_3d_init_zbuffer(pxl8_gfx* gfx) {
    if (gfx->zbuffer) return true;

    gfx->zbuffer_width = gfx->framebuffer_width;
    gfx->zbuffer_height = gfx->framebuffer_height;

    gfx->zbuffer = (f32*)SDL_calloc(gfx->zbuffer_width * gfx->zbuffer_height, sizeof(f32));
    if (!gfx->zbuffer) {
        return false;
    }

    pxl8_3d_clear_zbuffer(gfx);
    return true;
}

void pxl8_3d_clear_zbuffer(pxl8_gfx* gfx) {
    if (!gfx || !gfx->zbuffer) return;
    for (i32 i = 0; i < gfx->zbuffer_width * gfx->zbuffer_height; i++) {
        gfx->zbuffer[i] = 1e30f;
    }
}

void pxl8_3d_set_backface_culling(pxl8_gfx* gfx, bool culling) {
    if (!gfx) return;
    gfx->backface_culling = culling;
}

void pxl8_3d_set_model(pxl8_gfx* gfx, pxl8_mat4 mat) {
    if (!gfx) return;
    gfx->model = mat;
}

void pxl8_3d_set_projection(pxl8_gfx* gfx, pxl8_mat4 mat) {
    if (!gfx) return;
    gfx->projection = mat;
}

void pxl8_3d_set_view(pxl8_gfx* gfx, pxl8_mat4 mat) {
    if (!gfx) return;
    gfx->view = mat;
}

void pxl8_3d_set_wireframe(pxl8_gfx* gfx, bool wireframe) {
    if (!gfx) return;
    gfx->wireframe = wireframe;
}

static inline pxl8_vec4 pxl8_transform_vertex(pxl8_gfx* gfx, pxl8_vec3 pos) {
    pxl8_vec4 v = {
        .x = pos.x,
        .y = pos.y,
        .z = pos.z,
        .w = 1.0f,
    };

    pxl8_mat4 mvp = pxl8_mat4_multiply(gfx->projection,
                    pxl8_mat4_multiply(gfx->view, gfx->model));

    return pxl8_mat4_multiply_vec4(mvp, v);
}

static inline void pxl8_project_to_screen(pxl8_gfx* gfx, pxl8_vec4 clip, i32* x, i32* y, f32* z) {
    if (fabsf(clip.w) < 1e-6f) {
        *x = *y = 0;
        *z = 1e30f;
        return;
    }

    f32 inv_w = 1.0f / clip.w;
    f32 ndc_x = clip.x * inv_w;
    f32 ndc_y = clip.y * inv_w;

    *x = (i32)((ndc_x + 1.0f) * 0.5f * gfx->framebuffer_width);
    *y = (i32)((1.0f - ndc_y) * 0.5f * gfx->framebuffer_height);
    *z = clip.z * inv_w;
}

void pxl8_3d_draw_line_3d(pxl8_gfx* gfx, pxl8_vec3 p0, pxl8_vec3 p1, u32 color) {
    if (!gfx) return;

    pxl8_vec4 v0 = pxl8_transform_vertex(gfx, p0);
    pxl8_vec4 v1 = pxl8_transform_vertex(gfx, p1);

    i32 x0, y0, x1, y1;
    f32 z0, z1;
    pxl8_project_to_screen(gfx, v0, &x0, &y0, &z0);
    pxl8_project_to_screen(gfx, v1, &x1, &y1, &z1);

    pxl8_line(gfx, x0, y0, x1, y1, color);
}

static inline void pxl8_fill_scanline(pxl8_gfx* gfx, i32 y, i32 xs, i32 xe, f32 z0, f32 z1, u32 color) {
    if (y < 0 || y >= gfx->framebuffer_height) return;
    if (xs > xe) {
        i32 tmp = xs; xs = xe; xe = tmp;
    }

    for (i32 x = xs; x <= xe; x++) {
        if (x >= 0 && x < gfx->framebuffer_width) {
            f32 t = (xe == xs) ? 0.0f : (f32)(x - xs) / (f32)(xe - xs);
            f32 z = z0 + t * (z1 - z0);

            i32 idx = y * gfx->zbuffer_width + x;
            if (z <= gfx->zbuffer[idx]) {
                gfx->zbuffer[idx] = z;
                pxl8_pixel(gfx, x, y, color);
            }
        }
    }
}

static void pxl8_draw_flat_bottom_triangle(
    pxl8_gfx* gfx,
    i32 x0, i32 y0, f32 z0,
    i32 x1, i32 y1, f32 z1,
    i32 x2, i32 y2, f32 z2,
    u32 color
) {
    (void)z2;
    if (y1 == y0) return;

    f32 inv_slope_1 = (f32)(x1 - x0) / (f32)(y1 - y0);
    f32 inv_slope_2 = (f32)(x2 - x0) / (f32)(y2 - y0);

    f32 cur_x1 = (f32)x0;
    f32 cur_x2 = (f32)x0;

    for (i32 y = y0; y <= y1; y++) {
        pxl8_fill_scanline(gfx, y, (i32)cur_x1, (i32)cur_x2, z0, z1, color);
        cur_x1 += inv_slope_1;
        cur_x2 += inv_slope_2;
    }
}

static void pxl8_draw_flat_top_triangle(
    pxl8_gfx* gfx,
    i32 x0, i32 y0, f32 z0,
    i32 x1, i32 y1, f32 z1,
    i32 x2, i32 y2, f32 z2,
    u32 color
) {
    (void)z2;
    if (y2 == y0) return;

    f32 inv_slope_1 = (f32)(x2 - x0) / (f32)(y2 - y0);
    f32 inv_slope_2 = (f32)(x2 - x1) / (f32)(y2 - y1);

    f32 cur_x1 = (f32)x2;
    f32 cur_x2 = (f32)x2;

    for (i32 y = y2; y > y0; y--) {
        pxl8_fill_scanline(gfx, y, (i32)cur_x1, (i32)cur_x2, z0, z1, color);
        cur_x1 -= inv_slope_1;
        cur_x2 -= inv_slope_2;
    }
}

void pxl8_3d_draw_triangle_raw(pxl8_gfx* gfx, pxl8_vec3 v0, pxl8_vec3 v1, pxl8_vec3 v2, u32 color) {
    pxl8_triangle tri;
    tri.v[0].position = v0;
    tri.v[0].color = color;
    tri.v[1].position = v1;
    tri.v[1].color = color;
    tri.v[2].position = v2;
    tri.v[2].color = color;
    pxl8_3d_draw_triangle(gfx, tri);
}

void pxl8_3d_draw_triangle(pxl8_gfx* gfx, pxl8_triangle tri) {
    if (!gfx || !pxl8_3d_init_zbuffer(gfx)) return;

    pxl8_vec4 v0 = pxl8_transform_vertex(gfx, tri.v[0].position);
    pxl8_vec4 v1 = pxl8_transform_vertex(gfx, tri.v[1].position);
    pxl8_vec4 v2 = pxl8_transform_vertex(gfx, tri.v[2].position);

    if (v0.w <= 0 || v1.w <= 0 || v2.w <= 0) return;

    i32 x0, y0, x1, y1, x2, y2;
    f32 z0, z1, z2;
    pxl8_project_to_screen(gfx, v0, &x0, &y0, &z0);
    pxl8_project_to_screen(gfx, v1, &x1, &y1, &z1);
    pxl8_project_to_screen(gfx, v2, &x2, &y2, &z2);

    if (gfx->backface_culling) {
        i32 cross = (x1 - x0) * (y2 - y0) - (y1 - y0) * (x2 - x0);
        if (cross >= 0) return;
    }

    if (gfx->wireframe) {
        u32 color = tri.v[0].color;
        pxl8_line(gfx, x0, y0, x1, y1, color);
        pxl8_line(gfx, x1, y1, x2, y2, color);
        pxl8_line(gfx, x2, y2, x0, y0, color);
        return;
    }

    if (y0 > y1) {
        i32 tmp_i = x0; x0 = x1; x1 = tmp_i;
        tmp_i = y0; y0 = y1; y1 = tmp_i;
        f32 tmp_f = z0; z0 = z1; z1 = tmp_f;
    }
    if (y0 > y2) {
        i32 tmp_i = x0; x0 = x2; x2 = tmp_i;
        tmp_i = y0; y0 = y2; y2 = tmp_i;
        f32 tmp_f = z0; z0 = z2; z2 = tmp_f;
    }
    if (y1 > y2) {
        i32 tmp_i = x1; x1 = x2; x2 = tmp_i;
        tmp_i = y1; y1 = y2; y2 = tmp_i;
        f32 tmp_f = z1; z1 = z2; z2 = tmp_f;
    }

    u32 color = tri.v[0].color;

    if (y1 == y2) {
        pxl8_draw_flat_bottom_triangle(gfx, x0, y0, z0, x1, y1, z1, x2, y2, z2, color);
    } else if (y0 == y1) {
        pxl8_draw_flat_top_triangle(gfx, x0, y0, z0, x1, y1, z1, x2, y2, z2, color);
    } else {
        i32 x3 = x0 + (i32)(((f32)(y1 - y0) / (f32)(y2 - y0)) * (x2 - x0));
        i32 y3 = y1;
        f32 z3 = z0 + ((f32)(y1 - y0) / (f32)(y2 - y0)) * (z2 - z0);

        pxl8_draw_flat_bottom_triangle(gfx, x0, y0, z0, x1, y1, z1, x3, y3, z3, color);
        pxl8_draw_flat_top_triangle(gfx, x1, y1, z1, x3, y3, z3, x2, y2, z2, color);
    }
}