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

#include "pxl8_ase.h"
#include "pxl8_atlas.h"
#include "pxl8_blit.h"
#include "pxl8_font.h"
#include "pxl8_gfx.h"
#include "pxl8_hal.h"
#include "pxl8_macros.h"
#include "pxl8_math.h"
#include "pxl8_types.h"

typedef struct pxl8_sprite_cache_entry {
    char path[256];
    u32 sprite_id;
    bool active;
} pxl8_sprite_cache_entry;

struct pxl8_gfx {
    const pxl8_hal* hal;
    void* platform_data;

    pxl8_atlas* atlas;
    pxl8_sprite_cache_entry* sprite_cache;
    u32 sprite_cache_capacity;
    u32 sprite_cache_count;

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

    u32* palette;
    u32 palette_size;

    pxl8_viewport viewport;

    bool backface_culling;
    pxl8_mat4 model;
    pxl8_mat4 projection;
    pxl8_mat4 view;
    pxl8_mat4 mvp;
    bool mvp_dirty;
    pxl8_frustum frustum;
    bool wireframe;
    f32* zbuffer;
    i32 zbuffer_height;
    i32 zbuffer_width;

    bool affine_textures;

    u32 frame_triangle_count;
    u32 frame_pixel_count;
};

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);
}

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_SNES:      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) {
        return bounds;
    }
    bounds.w = gfx->framebuffer_width;
    bounds.h = gfx->framebuffer_height;
    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;
}

u32 pxl8_gfx_get_palette_size(const pxl8_gfx* gfx) {
    return gfx ? gfx->palette_size : 0;
}

pxl8_gfx* pxl8_gfx_create(
    const pxl8_hal* hal,
    pxl8_color_mode mode,
    pxl8_resolution resolution,
    const char* title,
    i32 window_width,
    i32 window_height
) {
    pxl8_gfx* gfx = (pxl8_gfx*)calloc(1, sizeof(*gfx));
    if (!gfx) {
        pxl8_error("Failed to allocate graphics context");
        return NULL;
    }

    gfx->hal = hal;

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

    gfx->platform_data = gfx->hal->create(mode, resolution, title, window_width, window_height);
    if (!gfx->platform_data) {
        pxl8_error("Failed to create platform context");
        free(gfx);
        return NULL;
    }

    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*)calloc(1, fb_size);
    if (!gfx->framebuffer) {
        pxl8_error("Failed to allocate framebuffer");
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    gfx->palette_size = pxl8_get_palette_size(mode);
    if (gfx->palette_size > 0) {
        gfx->palette = (u32*)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->mvp_dirty = true;
    gfx->wireframe = false;
    gfx->zbuffer = NULL;
    gfx->zbuffer_height = 0;
    gfx->zbuffer_width = 0;
    gfx->affine_textures = false;

    gfx->initialized = true;
    return gfx;
}

void pxl8_gfx_destroy(pxl8_gfx* gfx) {
    if (!gfx) return;

    pxl8_atlas_destroy(gfx->atlas);
    free(gfx->sprite_cache);

    if (gfx->hal && gfx->platform_data) {
        gfx->hal->destroy(gfx->platform_data);
    }

    free(gfx->framebuffer);
    free(gfx->palette);
    free(gfx->zbuffer);

    free(gfx);
}


pxl8_result pxl8_gfx_create_texture(pxl8_gfx* gfx, const u8* pixels, u32 width, u32 height) {
    if (!gfx || !gfx->initialized || !pixels) return PXL8_ERROR_INVALID_ARGUMENT;

    if (!gfx->atlas) {
        gfx->atlas = pxl8_atlas_create(1024, 1024, gfx->color_mode);
        if (!gfx->atlas) return PXL8_ERROR_OUT_OF_MEMORY;
    }

    u32 texture_id = pxl8_atlas_add_texture(gfx->atlas, pixels, width, height, gfx->color_mode);
    if (texture_id == UINT32_MAX) {
        pxl8_error("Texture doesn't fit in atlas");
        return PXL8_ERROR_INVALID_SIZE;
    }

    return texture_id;
}

pxl8_result pxl8_gfx_load_sprite(pxl8_gfx* gfx, const char* path) {
    if (!gfx || !gfx->initialized || !path) return PXL8_ERROR_INVALID_ARGUMENT;

    if (!gfx->sprite_cache) {
        gfx->sprite_cache_capacity = 64;
        gfx->sprite_cache = (pxl8_sprite_cache_entry*)calloc(
            gfx->sprite_cache_capacity, sizeof(pxl8_sprite_cache_entry)
        );
        if (!gfx->sprite_cache) return PXL8_ERROR_OUT_OF_MEMORY;
    }

    for (u32 i = 0; i < gfx->sprite_cache_count; i++) {
        if (gfx->sprite_cache[i].active && strcmp(gfx->sprite_cache[i].path, path) == 0) {
            return gfx->sprite_cache[i].sprite_id;
        }
    }

    if (!gfx->atlas) {
        gfx->atlas = pxl8_atlas_create(1024, 1024, gfx->color_mode);
        if (!gfx->atlas) return PXL8_ERROR_OUT_OF_MEMORY;
    }

    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_id = pxl8_atlas_add_texture(
        gfx->atlas,
        ase_file.frames[0].pixels,
        ase_file.header.width,
        ase_file.header.height,
        gfx->color_mode
    );

    pxl8_ase_destroy(&ase_file);

    if (sprite_id == UINT32_MAX) {
        pxl8_error("Sprite doesn't fit in atlas");
        return PXL8_ERROR_INVALID_SIZE;
    }

    if (gfx->sprite_cache_count >= gfx->sprite_cache_capacity) {
        gfx->sprite_cache_capacity *= 2;
        gfx->sprite_cache = (pxl8_sprite_cache_entry*)realloc(
            gfx->sprite_cache,
            gfx->sprite_cache_capacity * sizeof(pxl8_sprite_cache_entry)
        );
    }

    pxl8_sprite_cache_entry* entry = &gfx->sprite_cache[gfx->sprite_cache_count++];
    entry->active = true;
    entry->sprite_id = sprite_id;
    strncpy(entry->path, path, sizeof(entry->path) - 1);
    entry->path[sizeof(entry->path) - 1] = '\0';

    return sprite_id;
}

pxl8_atlas* pxl8_gfx_get_atlas(pxl8_gfx* gfx) {
    if (!gfx || !gfx->initialized) return NULL;

    if (!gfx->atlas) {
        gfx->atlas = pxl8_atlas_create(1024, 1024, gfx->color_mode);
    }

    return gfx->atlas;
}

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 || !ase_file.palette.colors) {
        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));

    u32 last_color = (copy_size > 0) ? gfx->palette[copy_size - 1] : 0xFF000000;
    for (u32 i = copy_size; i < gfx->palette_size; i++) {
        gfx->palette[i] = last_color;
    }

    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;
}

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

    if (gfx->hal && gfx->hal->upload_atlas) {
        gfx->hal->upload_atlas(gfx->platform_data, gfx->atlas, gfx->palette, gfx->color_mode);
        pxl8_atlas_mark_clean(gfx->atlas);
    }
}

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

    gfx->hal->upload_framebuffer(
        gfx->platform_data,
        gfx->framebuffer,
        gfx->framebuffer_width,
        gfx->framebuffer_height,
        gfx->palette,
        gfx->color_mode
    );
}

void pxl8_gfx_present(pxl8_gfx* gfx) {
    if (!gfx || !gfx->initialized || !gfx->hal) return;

    gfx->hal->present(gfx->platform_data);
}

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;
}

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;

    static u32 frame_count = 0;
    if (gfx->frame_triangle_count > 0) {
        if (frame_count % 60 == 0) {
            i32 fb_pixels = gfx->framebuffer_width * gfx->framebuffer_height;
            f32 overdraw = (f32)gfx->frame_pixel_count / (f32)fb_pixels;
            pxl8_trace("Frame triangles: %u, pixels: %u, overdraw: %.2fx",
                      gfx->frame_triangle_count, gfx->frame_pixel_count, overdraw);
        }
        frame_count++;
    }
    gfx->frame_triangle_count = 0;
    gfx->frame_pixel_count = 0;

    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) return;

    const pxl8_atlas_entry* entry = pxl8_atlas_get_entry(gfx->atlas, sprite_id);
    if (!entry || !entry->active) return;
    
    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);

    u32 atlas_width = pxl8_atlas_get_width(gfx->atlas);
    const u8* atlas_pixels = pxl8_atlas_get_pixels(gfx->atlas);

    if (is_1to1_scale && is_unclipped && pxl8_is_simd_aligned(w)) {
        const u8* sprite_data = atlas_pixels + entry->y * 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,
                atlas_width,
                x, y, w, h
            );
        } else {
            pxl8_blit_simd_indexed(
                gfx->framebuffer,
                gfx->framebuffer_width,
                sprite_data,
                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 * 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 = ((const u32*)atlas_pixels)[src_idx];
                    if (pixel & 0xFF000000) {
                        ((u32*)gfx->framebuffer)[dest_idx] = pixel;
                    }
                } else {
                    u8 pixel = atlas_pixels[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*)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;

    i32 count = gfx->zbuffer_width * gfx->zbuffer_height;
    const f32 far_z = 1e30f;

#if !defined(PXL8_SIMD_SCALAR)
    pxl8_simd_vec_f32 far_vec = pxl8_simd_set1_f32(far_z);
    i32 i = 0;
    for (; i + PXL8_SIMD_WIDTH_F32 <= count; i += PXL8_SIMD_WIDTH_F32) {
        pxl8_simd_store_f32(&gfx->zbuffer[i], far_vec);
    }
    for (; i < count; i++) {
        gfx->zbuffer[i] = far_z;
    }
#else
    for (i32 i = 0; i < count; i++) {
        gfx->zbuffer[i] = far_z;
    }
#endif
}

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;
    gfx->mvp_dirty = true;
}

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

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

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

void pxl8_3d_set_affine_textures(pxl8_gfx* gfx, bool affine) {
    if (!gfx) return;
    gfx->affine_textures = affine;
}

static inline void pxl8_update_mvp(pxl8_gfx* gfx) {
    if (!gfx->mvp_dirty) return;

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

    pxl8_mat4 vp = pxl8_mat4_multiply(gfx->projection, gfx->view);
    gfx->frustum = pxl8_frustum_from_matrix(vp);

    gfx->mvp_dirty = false;
}

const pxl8_frustum* pxl8_3d_get_frustum(pxl8_gfx* gfx) {
    if (!gfx) return NULL;
    pxl8_update_mvp(gfx);
    return &gfx->frustum;
}

static inline pxl8_vec4 pxl8_transform_vertex(pxl8_gfx* gfx, pxl8_vec3 pos) {
    pxl8_update_mvp(gfx);

    pxl8_vec4 v = {
        .x = pos.x,
        .y = pos.y,
        .z = pos.z,
        .w = 1.0f,
    };

    return pxl8_mat4_multiply_vec4(gfx->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_hicolor(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;
        f32 tmpz = z0; z0 = z1; z1 = tmpz;
    }

    if (xs < 0) xs = 0;
    if (xe >= gfx->framebuffer_width) xe = gfx->framebuffer_width - 1;
    if (xs > xe) return;

    i32 width = xe - xs;
    if (width == 0) {
        i32 idx = y * gfx->zbuffer_width + xs;
        if (z0 <= gfx->zbuffer[idx]) {
            gfx->zbuffer[idx] = z0;
            i32 fb_idx = y * gfx->framebuffer_width + xs;
            ((u32*)gfx->framebuffer)[fb_idx] = color;
        }
        return;
    }

    f32 dz = (z1 - z0) / (f32)width;
    f32 z = z0;

    i32 zbuf_offset = y * gfx->zbuffer_width;
    i32 fb_offset = y * gfx->framebuffer_width;
    u32* fb = (u32*)gfx->framebuffer;

    for (i32 x = xs; x <= xe; x++) {
        i32 idx = zbuf_offset + x;
        if (z <= gfx->zbuffer[idx]) {
            gfx->zbuffer[idx] = z;
            fb[fb_offset + x] = color;
        }
        z += dz;
    }
}

static inline void pxl8_fill_scanline_indexed(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;
        f32 tmpz = z0; z0 = z1; z1 = tmpz;
    }

    if (xs < 0) xs = 0;
    if (xe >= gfx->framebuffer_width) xe = gfx->framebuffer_width - 1;
    if (xs > xe) return;

    i32 width = xe - xs;
    u8 idx_color = color & 0xFF;

    if (width == 0) {
        i32 idx = y * gfx->zbuffer_width + xs;
        if (z0 <= gfx->zbuffer[idx]) {
            gfx->zbuffer[idx] = z0;
            i32 fb_idx = y * gfx->framebuffer_width + xs;
            gfx->framebuffer[fb_idx] = idx_color;
        }
        return;
    }

    f32 dz = (z1 - z0) / (f32)width;
    f32 z = z0;

    i32 zbuf_offset = y * gfx->zbuffer_width;
    i32 fb_offset = y * gfx->framebuffer_width;

    for (i32 x = xs; x <= xe; x++) {
        i32 idx = zbuf_offset + x;
        if (z <= gfx->zbuffer[idx]) {
            gfx->zbuffer[idx] = z;
            gfx->framebuffer[fb_offset + x] = idx_color;
        }
        z += dz;
    }
}

static inline u32 pxl8_sample_texture(pxl8_gfx* gfx, u32 texture_id, f32 u, f32 v) {
    if (!gfx->atlas) return 0;

    const pxl8_atlas_entry* entry = pxl8_atlas_get_entry(gfx->atlas, texture_id);
    if (!entry || !entry->active) return 0;

    u = u - floorf(u);
    v = v - floorf(v);

    i32 tx = (i32)(u * entry->w) % entry->w;
    i32 ty = (i32)(v * entry->h) % entry->h;

    i32 atlas_x = entry->x + tx;
    i32 atlas_y = entry->y + ty;
    u32 atlas_width = pxl8_atlas_get_width(gfx->atlas);
    const u8* atlas_pixels = pxl8_atlas_get_pixels(gfx->atlas);
    i32 idx = atlas_y * atlas_width + atlas_x;

    if (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        return ((const u32*)atlas_pixels)[idx];
    } else {
        return atlas_pixels[idx];
    }
}

static inline void pxl8_fill_scanline_textured_hicolor(
    pxl8_gfx* gfx, i32 y, i32 xs, i32 xe,
    f32 z0, f32 z1,
    f32 u0, f32 v0, f32 w0,
    f32 u1, f32 v1, f32 w1,
    u32 texture_id
) {
    if (y < 0 || y >= gfx->framebuffer_height) return;
    if (xs > xe) {
        i32 tmp = xs; xs = xe; xe = tmp;
        f32 tmpf;
        tmpf = z0; z0 = z1; z1 = tmpf;
        tmpf = u0; u0 = u1; u1 = tmpf;
        tmpf = v0; v0 = v1; v1 = tmpf;
        tmpf = w0; w0 = w1; w1 = tmpf;
    }

    i32 span = xe - xs;
    if (span <= 0) {
        if (xs >= 0 && xs < gfx->framebuffer_width) {
            i32 idx = y * gfx->zbuffer_width + xs;
            if (z0 <= gfx->zbuffer[idx]) {
                gfx->frame_pixel_count++;
                f32 u = u0, v = v0;
                if (!gfx->affine_textures && fabsf(w0) > 1e-6f) {
                    f32 perspective_w = 1.0f / w0;
                    u *= perspective_w;
                    v *= perspective_w;
                }
                u32 color = pxl8_sample_texture(gfx, texture_id, u, v);
                if (color & 0xFF000000) {
                    gfx->zbuffer[idx] = z0;
                    pxl8_pixel(gfx, xs, y, color);
                }
            }
        }
        return;
    }

    f32 inv_span = 1.0f / (f32)span;
    f32 dz = (z1 - z0) * inv_span;
    f32 du = (u1 - u0) * inv_span;
    f32 dv = (v1 - v0) * inv_span;
    f32 dw = (w1 - w0) * inv_span;

    f32 z = z0;
    f32 u = u0;
    f32 v = v0;
    f32 w = w0;

    for (i32 x = xs; x <= xe; x++) {
        if (x >= 0 && x < gfx->framebuffer_width) {
            i32 idx = y * gfx->zbuffer_width + x;
            if (z <= gfx->zbuffer[idx]) {
                gfx->frame_pixel_count++;
                f32 tex_u = u, tex_v = v;
                if (!gfx->affine_textures && fabsf(w) > 1e-6f) {
                    f32 perspective_w = 1.0f / w;
                    tex_u *= perspective_w;
                    tex_v *= perspective_w;
                }
                u32 color = pxl8_sample_texture(gfx, texture_id, tex_u, tex_v);
                if (color & 0xFF000000) {
                    gfx->zbuffer[idx] = z;
                    pxl8_pixel(gfx, x, y, color);
                }
            }
        }
        z += dz;
        u += du;
        v += dv;
        w += dw;
    }
}

static inline void pxl8_fill_scanline_textured_indexed(
    pxl8_gfx* gfx, i32 y, i32 xs, i32 xe,
    f32 z0, f32 z1,
    f32 u0, f32 v0, f32 w0,
    f32 u1, f32 v1, f32 w1,
    u32 texture_id
) {
    if (y < 0 || y >= gfx->framebuffer_height) return;
    if (xs > xe) {
        i32 tmp = xs; xs = xe; xe = tmp;
        f32 tmpf;
        tmpf = z0; z0 = z1; z1 = tmpf;
        tmpf = u0; u0 = u1; u1 = tmpf;
        tmpf = v0; v0 = v1; v1 = tmpf;
        tmpf = w0; w0 = w1; w1 = tmpf;
    }

    i32 span = xe - xs;
    if (span <= 0) {
        if (xs >= 0 && xs < gfx->framebuffer_width) {
            i32 idx = y * gfx->zbuffer_width + xs;
            if (z0 <= gfx->zbuffer[idx]) {
                gfx->frame_pixel_count++;
                f32 u = u0, v = v0;
                if (!gfx->affine_textures && fabsf(w0) > 1e-6f) {
                    f32 perspective_w = 1.0f / w0;
                    u *= perspective_w;
                    v *= perspective_w;
                }
                u32 color = pxl8_sample_texture(gfx, texture_id, u, v);
                if (color != 0) {
                    gfx->zbuffer[idx] = z0;
                    pxl8_pixel(gfx, xs, y, color);
                }
            }
        }
        return;
    }

    f32 inv_span = 1.0f / (f32)span;
    f32 dz = (z1 - z0) * inv_span;
    f32 du = (u1 - u0) * inv_span;
    f32 dv = (v1 - v0) * inv_span;
    f32 dw = (w1 - w0) * inv_span;

    f32 z = z0;
    f32 u = u0;
    f32 v = v0;
    f32 w = w0;

    for (i32 x = xs; x <= xe; x++) {
        if (x >= 0 && x < gfx->framebuffer_width) {
            i32 idx = y * gfx->zbuffer_width + x;
            if (z <= gfx->zbuffer[idx]) {
                gfx->frame_pixel_count++;
                f32 tex_u = u, tex_v = v;
                if (!gfx->affine_textures && fabsf(w) > 1e-6f) {
                    f32 perspective_w = 1.0f / w;
                    tex_u *= perspective_w;
                    tex_v *= perspective_w;
                }
                u32 color = pxl8_sample_texture(gfx, texture_id, tex_u, tex_v);
                if (color != 0) {
                    gfx->zbuffer[idx] = z;
                    pxl8_pixel(gfx, x, y, color);
                }
            }
        }
        z += dz;
        u += du;
        v += dv;
        w += dw;
    }
}

typedef void (*pxl8_scanline_func)(pxl8_gfx*, i32, i32, i32, f32, f32, u32);
typedef void (*pxl8_scanline_textured_func)(pxl8_gfx*, i32, i32, i32, f32, f32, f32, f32, f32, f32, f32, f32, u32);

static void pxl8_scan_triangle(
    pxl8_gfx* gfx,
    i32 x0, i32 y0, f32 z0,
    i32 x1, i32 y1, f32 z1,
    i32 x2, i32 y2, f32 z2,
    u32 color,
    pxl8_scanline_func fill_scanline,
    bool is_bottom
) {
    (void)z2;

    i32 y_start, y_end, y_step;
    f32 x_start1, x_start2, inv_slope_1, inv_slope_2;

    if (is_bottom) {
        if (y1 == y0 || y1 < y0 || y1 - y0 > gfx->framebuffer_height) return;
        y_start = y0;
        y_end = y1;
        y_step = 1;
        x_start1 = x_start2 = (f32)x0;
        inv_slope_1 = (f32)(x1 - x0) / (f32)(y1 - y0);
        inv_slope_2 = (f32)(x2 - x0) / (f32)(y2 - y0);
    } else {
        if (y2 == y0 || y2 < y0 || y2 - y0 > gfx->framebuffer_height) return;
        y_start = y2;
        y_end = y0;
        y_step = -1;
        x_start1 = x_start2 = (f32)x2;
        inv_slope_1 = (f32)(x2 - x0) / (f32)(y2 - y0);
        inv_slope_2 = (f32)(x2 - x1) / (f32)(y2 - y1);
    }

    f32 cur_x1 = x_start1;
    f32 cur_x2 = x_start2;
    f32 slope_step1 = inv_slope_1 * y_step;
    f32 slope_step2 = inv_slope_2 * y_step;

    for (i32 y = y_start; y != y_end + y_step; y += y_step) {
        fill_scanline(gfx, y, (i32)cur_x1, (i32)cur_x2, z0, z1, color);
        cur_x1 += slope_step1;
        cur_x2 += slope_step2;
    }
}

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;

    gfx->frame_triangle_count++;

    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;

    pxl8_scanline_func fill_scanline = (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR)
        ? pxl8_fill_scanline_hicolor
        : pxl8_fill_scanline_indexed;

    if (y1 == y2) {
        pxl8_scan_triangle(gfx, x0, y0, z0, x1, y1, z1, x2, y2, z2, color, fill_scanline, true);
    } else if (y0 == y1) {
        pxl8_scan_triangle(gfx, x0, y0, z0, x1, y1, z1, x2, y2, z2, color, fill_scanline, false);
    } 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_scan_triangle(gfx, x0, y0, z0, x1, y1, z1, x3, y3, z3, color, fill_scanline, true);
        pxl8_scan_triangle(gfx, x1, y1, z1, x3, y3, z3, x2, y2, z2, color, fill_scanline, false);
    }
}

typedef struct pxl8_textured_vertex {
    i32 x, y;
    f32 z, u, v, w;
} pxl8_textured_vertex;

static void pxl8_scan_triangle_textured(
    pxl8_gfx* gfx,
    pxl8_textured_vertex v0,
    pxl8_textured_vertex v1,
    pxl8_textured_vertex v2,
    u32 texture_id,
    pxl8_scanline_textured_func fill_scanline,
    bool is_bottom
) {
    i32 y_start, y_end, y_step;
    f32 x_start1, x_start2;
    f32 z_start1, z_start2, u_start1, u_start2, v_start1, v_start2, w_start1, w_start2;
    f32 inv_slope_1, inv_slope_2, inv_slope_z1, inv_slope_z2;
    f32 inv_slope_u1, inv_slope_u2, inv_slope_v1, inv_slope_v2, inv_slope_w1, inv_slope_w2;

    if (is_bottom) {
        if (v1.y == v0.y || v1.y < v0.y || v1.y - v0.y > gfx->framebuffer_height) return;
        y_start = v0.y;
        y_end = v1.y;
        y_step = 1;
        x_start1 = x_start2 = (f32)v0.x;
        z_start1 = z_start2 = v0.z;
        u_start1 = u_start2 = v0.u;
        v_start1 = v_start2 = v0.v;
        w_start1 = w_start2 = v0.w;
        inv_slope_1 = (f32)(v1.x - v0.x) / (f32)(v1.y - v0.y);
        inv_slope_2 = (f32)(v2.x - v0.x) / (f32)(v2.y - v0.y);
        inv_slope_z1 = (v1.z - v0.z) / (f32)(v1.y - v0.y);
        inv_slope_z2 = (v2.z - v0.z) / (f32)(v2.y - v0.y);
        inv_slope_u1 = (v1.u - v0.u) / (f32)(v1.y - v0.y);
        inv_slope_u2 = (v2.u - v0.u) / (f32)(v2.y - v0.y);
        inv_slope_v1 = (v1.v - v0.v) / (f32)(v1.y - v0.y);
        inv_slope_v2 = (v2.v - v0.v) / (f32)(v2.y - v0.y);
        inv_slope_w1 = (v1.w - v0.w) / (f32)(v1.y - v0.y);
        inv_slope_w2 = (v2.w - v0.w) / (f32)(v2.y - v0.y);
    } else {
        if (v2.y == v0.y || v2.y < v0.y || v2.y - v0.y > gfx->framebuffer_height) return;
        y_start = v2.y;
        y_end = v0.y;
        y_step = -1;
        x_start1 = x_start2 = (f32)v2.x;
        z_start1 = z_start2 = v2.z;
        u_start1 = u_start2 = v2.u;
        v_start1 = v_start2 = v2.v;
        w_start1 = w_start2 = v2.w;
        inv_slope_1 = (f32)(v2.x - v0.x) / (f32)(v2.y - v0.y);
        inv_slope_2 = (f32)(v2.x - v1.x) / (f32)(v2.y - v1.y);
        inv_slope_z1 = (v2.z - v0.z) / (f32)(v2.y - v0.y);
        inv_slope_z2 = (v2.z - v1.z) / (f32)(v2.y - v1.y);
        inv_slope_u1 = (v2.u - v0.u) / (f32)(v2.y - v0.y);
        inv_slope_u2 = (v2.u - v1.u) / (f32)(v2.y - v1.y);
        inv_slope_v1 = (v2.v - v0.v) / (f32)(v2.y - v0.y);
        inv_slope_v2 = (v2.v - v1.v) / (f32)(v2.y - v1.y);
        inv_slope_w1 = (v2.w - v0.w) / (f32)(v2.y - v0.y);
        inv_slope_w2 = (v2.w - v1.w) / (f32)(v2.y - v1.y);
    }

    f32 cur_x1 = x_start1, cur_x2 = x_start2;
    f32 cur_z1 = z_start1, cur_z2 = z_start2;
    f32 cur_u1 = u_start1, cur_u2 = u_start2;
    f32 cur_v1 = v_start1, cur_v2 = v_start2;
    f32 cur_w1 = w_start1, cur_w2 = w_start2;

    f32 slope_step_x1 = inv_slope_1 * y_step;
    f32 slope_step_x2 = inv_slope_2 * y_step;
    f32 slope_step_z1 = inv_slope_z1 * y_step;
    f32 slope_step_z2 = inv_slope_z2 * y_step;
    f32 slope_step_u1 = inv_slope_u1 * y_step;
    f32 slope_step_u2 = inv_slope_u2 * y_step;
    f32 slope_step_v1 = inv_slope_v1 * y_step;
    f32 slope_step_v2 = inv_slope_v2 * y_step;
    f32 slope_step_w1 = inv_slope_w1 * y_step;
    f32 slope_step_w2 = inv_slope_w2 * y_step;

    for (i32 y = y_start; y != y_end + y_step; y += y_step) {
        fill_scanline(gfx, y, (i32)cur_x1, (i32)cur_x2,
            cur_z1, cur_z2, cur_u1, cur_v1, cur_w1, cur_u2, cur_v2, cur_w2, texture_id);
        cur_x1 += slope_step_x1; cur_x2 += slope_step_x2;
        cur_z1 += slope_step_z1; cur_z2 += slope_step_z2;
        cur_u1 += slope_step_u1; cur_u2 += slope_step_u2;
        cur_v1 += slope_step_v1; cur_v2 += slope_step_v2;
        cur_w1 += slope_step_w1; cur_w2 += slope_step_w2;
    }
}

void pxl8_3d_draw_triangle_textured(
    pxl8_gfx* gfx,
    pxl8_vec3 v0,
    pxl8_vec3 v1,
    pxl8_vec3 v2,
    f32 u0,
    f32 v0f,
    f32 u1,
    f32 v1f,
    f32 u2,
    f32 v2f,
    u32 texture_id
) {
    if (!gfx || !pxl8_3d_init_zbuffer(gfx)) return;

    pxl8_vec4 cv0 = pxl8_transform_vertex(gfx, v0);
    pxl8_vec4 cv1 = pxl8_transform_vertex(gfx, v1);
    pxl8_vec4 cv2 = pxl8_transform_vertex(gfx, v2);

    if (cv0.w <= 0 || cv1.w <= 0 || cv2.w <= 0) return;

    pxl8_textured_vertex tv[3];
    f32 z_tmp;
    pxl8_project_to_screen(gfx, cv0, &tv[0].x, &tv[0].y, &z_tmp);
    tv[0].z = z_tmp;
    tv[0].u = gfx->affine_textures ? u0 : u0 / cv0.w;
    tv[0].v = gfx->affine_textures ? v0f : v0f / cv0.w;
    tv[0].w = gfx->affine_textures ? 1.0f : 1.0f / cv0.w;

    pxl8_project_to_screen(gfx, cv1, &tv[1].x, &tv[1].y, &z_tmp);
    tv[1].z = z_tmp;
    tv[1].u = gfx->affine_textures ? u1 : u1 / cv1.w;
    tv[1].v = gfx->affine_textures ? v1f : v1f / cv1.w;
    tv[1].w = gfx->affine_textures ? 1.0f : 1.0f / cv1.w;

    pxl8_project_to_screen(gfx, cv2, &tv[2].x, &tv[2].y, &z_tmp);
    tv[2].z = z_tmp;
    tv[2].u = gfx->affine_textures ? u2 : u2 / cv2.w;
    tv[2].v = gfx->affine_textures ? v2f : v2f / cv2.w;
    tv[2].w = gfx->affine_textures ? 1.0f : 1.0f / cv2.w;

    if (gfx->backface_culling) {
        i32 cross =
            (tv[1].x - tv[0].x) * (tv[2].y - tv[0].y) - (tv[1].y - tv[0].y) * (tv[2].x - tv[0].x);
        if (cross >= 0) return;
    }

    if (tv[0].y > tv[1].y) {
        pxl8_textured_vertex tmp = tv[0]; tv[0] = tv[1]; tv[1] = tmp;
    }
    if (tv[0].y > tv[2].y) {
        pxl8_textured_vertex tmp = tv[0]; tv[0] = tv[2]; tv[2] = tmp;
    }
    if (tv[1].y > tv[2].y) {
        pxl8_textured_vertex tmp = tv[1]; tv[1] = tv[2]; tv[2] = tmp;
    }

    pxl8_scanline_textured_func fill_scanline = (gfx->color_mode == PXL8_COLOR_MODE_HICOLOR)
        ? pxl8_fill_scanline_textured_hicolor
        : pxl8_fill_scanline_textured_indexed;

    if (tv[1].y == tv[2].y) {
        pxl8_scan_triangle_textured(gfx, tv[0], tv[1], tv[2], texture_id, fill_scanline, true);
    } else if (tv[0].y == tv[1].y) {
        pxl8_scan_triangle_textured(gfx, tv[0], tv[1], tv[2], texture_id, fill_scanline, false);
    } else {
        f32 t = (f32)(tv[1].y - tv[0].y) / (f32)(tv[2].y - tv[0].y);
        pxl8_textured_vertex v3;
        v3.x = tv[0].x + (i32)(t * (tv[2].x - tv[0].x));
        v3.y = tv[1].y;
        v3.z = tv[0].z + t * (tv[2].z - tv[0].z);
        v3.u = tv[0].u + t * (tv[2].u - tv[0].u);
        v3.v = tv[0].v + t * (tv[2].v - tv[0].v);
        v3.w = tv[0].w + t * (tv[2].w - tv[0].w);

        pxl8_scan_triangle_textured(gfx, tv[0], tv[1], v3, texture_id, fill_scanline, true);
        pxl8_scan_triangle_textured(gfx, tv[1], v3, tv[2], texture_id, fill_scanline, false);
    }
}