#include "pxl8_gfx.h"

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

#include "pxl8_ase.h"
#include "pxl8_atlas.h"
#include "pxl8_blit.h"
#include "pxl8_color.h"
#include "pxl8_colormap.h"
#include "pxl8_font.h"
#include "pxl8_glows.h"
#include "pxl8_hal.h"
#include "pxl8_log.h"
#include "pxl8_macros.h"
#include "pxl8_math.h"
#include "pxl8_mem.h"
#include "pxl8_render.h"
#include "pxl8_shader_registry.h"
#include "pxl8_sys.h"
#include "pxl8_types.h"

void pxl8_renderer_update_stats(pxl8_renderer* r, f32 dt);
pxl8_gfx_stats* pxl8_renderer_get_stats(pxl8_renderer* r);

#define PXL8_MAX_TARGET_STACK 8

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

typedef struct pxl8_target_entry {
    pxl8_gfx_texture color;
    pxl8_gfx_texture depth;
    pxl8_gfx_texture light;
} pxl8_target_entry;

#define PXL8_MAX_FRAME_RESOURCES 512
#define PXL8_STREAM_VB_SIZE (256 * 1024)
#define PXL8_STREAM_IB_SIZE (512 * 1024)

typedef struct pxl8_frame_resources {
    pxl8_gfx_texture textures[PXL8_MAX_FRAME_RESOURCES];
    pxl8_gfx_buffer buffers[PXL8_MAX_FRAME_RESOURCES];
    pxl8_gfx_pipeline pipelines[PXL8_MAX_FRAME_RESOURCES];
    pxl8_gfx_bindings bindings[PXL8_MAX_FRAME_RESOURCES];
    u32 texture_count;
    u32 buffer_count;
    u32 pipeline_count;
    u32 bindings_count;
} pxl8_frame_resources;

struct pxl8_gfx {
    pxl8_atlas* atlas;
    pxl8_renderer* renderer;
    pxl8_gfx_texture color_target;
    pxl8_gfx_texture depth_target;
    pxl8_gfx_cmdbuf* cmdbuf;
    pxl8_target_entry target_stack[PXL8_MAX_TARGET_STACK];
    u32 target_stack_depth;
    const pxl8_bsp* bsp;
    pxl8_colormap* colormap;
    i32 framebuffer_height;
    i32 framebuffer_width;
    pxl8_frustum frustum;
    const pxl8_hal* hal;
    bool initialized;
    u32* output;
    pxl8_palette* palette;
    pxl8_palette_cube* palette_cube;
    pxl8_gfx_pass frame_pass;
    pxl8_frame_resources frame_res;
    pxl8_pixel_mode pixel_mode;
    void* platform_data;
    pxl8_sprite_cache_entry* sprite_cache;
    u32 sprite_cache_capacity;
    u32 sprite_cache_count;
    pxl8_viewport viewport;
    pxl8_mat4 view_proj;
    pxl8_3d_frame frame;

    pxl8_gfx_buffer stream_vb;
    pxl8_gfx_buffer stream_ib;
    u32 stream_vb_capacity;
    u32 stream_ib_capacity;
    u32 stream_vb_offset;
    u32 stream_ib_offset;
    bool wireframe;
};

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_pixel_mode pxl8_gfx_get_pixel_mode(pxl8_gfx* gfx) {
    return gfx ? gfx->pixel_mode : PXL8_PIXEL_INDEXED;
}

u8* pxl8_gfx_get_framebuffer_indexed(pxl8_gfx* gfx) {
    if (!gfx || gfx->pixel_mode == PXL8_PIXEL_HICOLOR) return NULL;
    return (u8*)pxl8_texture_get_data(gfx->renderer, gfx->color_target);
}

u16* pxl8_gfx_get_framebuffer_hicolor(pxl8_gfx* gfx) {
    if (!gfx || gfx->pixel_mode != PXL8_PIXEL_HICOLOR) return NULL;
    return (u16*)pxl8_texture_get_data(gfx->renderer, gfx->color_target);
}

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

u32* pxl8_gfx_get_output(pxl8_gfx* gfx) {
    if (!gfx) return NULL;
    return gfx->output;
}

void pxl8_gfx_resolve(pxl8_gfx* gfx) {
    if (!gfx || !gfx->initialized) return;
    if (gfx->pixel_mode != PXL8_PIXEL_INDEXED) return;

    const u32* pal = gfx->palette ? pxl8_palette_colors(gfx->palette) : NULL;
    if (!pal) {
        pxl8_error("resolve: no palette!");
        return;
    }
    pxl8_resolve_to_rgba(
        gfx->renderer,
        gfx->color_target,
        pal,
        gfx->output
    );
}

const u32* pxl8_gfx_palette_colors(pxl8_gfx* gfx) {
    if (!gfx || !gfx->palette) return NULL;
    return pxl8_palette_colors(gfx->palette);
}

u16* pxl8_gfx_get_zbuffer(pxl8_gfx* gfx) {
    if (!gfx) return NULL;
    return (u16*)pxl8_texture_get_data(gfx->renderer, gfx->depth_target);
}

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

pxl8_palette* pxl8_gfx_palette(pxl8_gfx* gfx) {
    return gfx ? gfx->palette : NULL;
}

pxl8_colormap* pxl8_gfx_colormap(pxl8_gfx* gfx) {
    return gfx ? gfx->colormap : NULL;
}

u8 pxl8_gfx_find_color(pxl8_gfx* gfx, u32 color) {
    if (!gfx || !gfx->palette) return 0;
    if (color <= 0xFFFFFF) color = (color << 8) | 0xFF;
    u8 r = (color >> 24) & 0xFF;
    u8 g = (color >> 16) & 0xFF;
    u8 b = (color >> 8) & 0xFF;
    return pxl8_palette_find_closest(gfx->palette, r, g, b);
}

void pxl8_gfx_set_palette_colors(pxl8_gfx* gfx, const u32* colors, u16 count) {
    if (!gfx || !gfx->palette || !colors) return;
    pxl8_set_palette(gfx->palette, colors, count);
}

i32 pxl8_gfx_load_palette(pxl8_gfx* gfx, const char* filepath) {
    if (!gfx || !gfx->palette || !filepath) return -1;
    pxl8_result result = pxl8_palette_load_ase(gfx->palette, filepath);
    if (result != PXL8_OK) return (i32)result;
    if (gfx->pixel_mode != PXL8_PIXEL_HICOLOR) {
        if (gfx->colormap) {
            pxl8_colormap_generate(gfx->colormap, pxl8_palette_colors(gfx->palette));
        }
    }
    return 0;
}

pxl8_gfx* pxl8_gfx_create(
    const pxl8_hal* hal,
    void* platform_data,
    pxl8_pixel_mode mode,
    pxl8_resolution resolution
) {
    pxl8_shader_registry_init();

    pxl8_gfx* gfx = (pxl8_gfx*)pxl8_calloc(1, sizeof(pxl8_gfx));
    if (!gfx) {
        pxl8_error("Failed to allocate graphics context");
        return NULL;
    }

    gfx->hal = hal;
    gfx->platform_data = platform_data;

    gfx->pixel_mode = mode;
    pxl8_size size = pxl8_get_resolution_dimensions(resolution);
    gfx->framebuffer_width = size.w;
    gfx->framebuffer_height = size.h;

    if (!gfx->platform_data) {
        pxl8_error("Platform data cannot be NULL");
        pxl8_free(gfx);
        return NULL;
    }

    if (mode != PXL8_PIXEL_HICOLOR) {
        gfx->palette = pxl8_palette_create();
    }

    gfx->renderer = pxl8_renderer_create(gfx->framebuffer_width, gfx->framebuffer_height);
    if (!gfx->renderer) {
        pxl8_error("Failed to create renderer");
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    pxl8_gfx_texture_desc color_desc = {
        .width = (u32)gfx->framebuffer_width,
        .height = (u32)gfx->framebuffer_height,
        .format = PXL8_GFX_FORMAT_INDEXED8,
        .render_target = true,
    };
    gfx->color_target = pxl8_create_texture(gfx->renderer, &color_desc);

    pxl8_gfx_texture_desc depth_desc = {
        .width = (u32)gfx->framebuffer_width,
        .height = (u32)gfx->framebuffer_height,
        .format = PXL8_GFX_FORMAT_DEPTH16,
        .render_target = true,
    };
    gfx->depth_target = pxl8_create_texture(gfx->renderer, &depth_desc);

    u32 pixel_count = (u32)gfx->framebuffer_width * (u32)gfx->framebuffer_height;
    gfx->output = pxl8_calloc(pixel_count, sizeof(u32));
    if (!gfx->output) {
        pxl8_error("Failed to allocate output buffer");
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    gfx->cmdbuf = pxl8_cmdbuf_create(1024);
    if (!gfx->cmdbuf) {
        pxl8_error("Failed to create command buffer");
        pxl8_gfx_destroy(gfx);
        return NULL;
    }

    if (mode != PXL8_PIXEL_HICOLOR) {
        gfx->colormap = pxl8_calloc(1, sizeof(pxl8_colormap));
    }

    gfx->target_stack[0] = (pxl8_target_entry){
        .color = gfx->color_target,
        .depth = gfx->depth_target,
    };
    gfx->target_stack_depth = 1;

    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;

    pxl8_gfx_buffer_desc vb_desc = {
        .capacity = PXL8_STREAM_VB_SIZE,
        .type = PXL8_GFX_BUFFER_VERTEX,
        .usage = PXL8_GFX_USAGE_STREAM,
    };
    gfx->stream_vb = pxl8_create_buffer(gfx->renderer, &vb_desc);
    gfx->stream_vb_capacity = PXL8_STREAM_VB_SIZE;
    gfx->stream_vb_offset = 0;

    pxl8_gfx_buffer_desc ib_desc = {
        .capacity = PXL8_STREAM_IB_SIZE,
        .type = PXL8_GFX_BUFFER_INDEX,
        .usage = PXL8_GFX_USAGE_STREAM,
    };
    gfx->stream_ib = pxl8_create_buffer(gfx->renderer, &ib_desc);
    gfx->stream_ib_capacity = PXL8_STREAM_IB_SIZE;
    gfx->stream_ib_offset = 0;

    gfx->initialized = true;
    return gfx;
}

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

    pxl8_atlas_destroy(gfx->atlas);
    pxl8_cmdbuf_destroy(gfx->cmdbuf);
    pxl8_free(gfx->colormap);
    pxl8_free(gfx->output);
    pxl8_palette_cube_destroy(gfx->palette_cube);
    pxl8_palette_destroy(gfx->palette);
    pxl8_free(gfx->sprite_cache);
    pxl8_renderer_destroy(gfx->renderer);

    pxl8_free(gfx);
}

static pxl8_result pxl8_gfx_ensure_atlas(pxl8_gfx* gfx) {
    if (gfx->atlas) return PXL8_OK;
    gfx->atlas = pxl8_atlas_create(PXL8_DEFAULT_ATLAS_SIZE, PXL8_DEFAULT_ATLAS_SIZE, gfx->pixel_mode);
    return gfx->atlas ? PXL8_OK : PXL8_ERROR_OUT_OF_MEMORY;
}

void pxl8_gfx_clear_textures(pxl8_gfx* gfx) {
    if (!gfx) return;

    if (gfx->atlas) {
        pxl8_atlas_clear(gfx->atlas, 0);
    }

    if (gfx->sprite_cache) {
        gfx->sprite_cache_count = 0;
    }
}

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;

    pxl8_result result = pxl8_gfx_ensure_atlas(gfx);
    if (result != PXL8_OK) return result;

    u32 texture_id = pxl8_atlas_add_texture(gfx->atlas, pixels, width, height, gfx->pixel_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 = PXL8_DEFAULT_SPRITE_CACHE_CAPACITY;
        gfx->sprite_cache = (pxl8_sprite_cache_entry*)pxl8_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;
        }
    }

    pxl8_result result = pxl8_gfx_ensure_atlas(gfx);
    if (result != PXL8_OK) return result;

    pxl8_ase_file ase_file;
    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->pixel_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) {
        u32 new_capacity = gfx->sprite_cache_capacity * 2;
        pxl8_sprite_cache_entry* new_cache = (pxl8_sprite_cache_entry*)pxl8_realloc(
            gfx->sprite_cache,
            new_capacity * sizeof(pxl8_sprite_cache_entry)
        );
        if (!new_cache) return PXL8_ERROR_OUT_OF_MEMORY;
        gfx->sprite_cache = new_cache;
        gfx->sprite_cache_capacity = new_capacity;
    }

    pxl8_sprite_cache_entry* entry = &gfx->sprite_cache[gfx->sprite_cache_count++];
    entry->active = true;
    entry->sprite_id = sprite_id;
    pxl8_strncpy(entry->path, path, sizeof(entry->path));

    return sprite_id;
}

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

    if (pxl8_gfx_ensure_atlas(gfx) != PXL8_OK) return NULL;
    return gfx->atlas;
}

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

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

    if (gfx->pixel_mode == PXL8_PIXEL_INDEXED) {
        pxl8_gfx_resolve(gfx);
        gfx->hal->upload_texture(
            gfx->platform_data,
            gfx->output,
            gfx->framebuffer_width,
            gfx->framebuffer_height,
            PXL8_PIXEL_RGBA,
            NULL
        );
        return;
    }

    u32* colors = gfx->palette ? pxl8_palette_colors(gfx->palette) : NULL;
    u8* framebuffer = (u8*)pxl8_texture_get_data(gfx->renderer, gfx->color_target);
    gfx->hal->upload_texture(
        gfx->platform_data,
        framebuffer,
        gfx->framebuffer_width,
        gfx->framebuffer_height,
        gfx->pixel_mode,
        colors
    );
}

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;
}

static pxl8_gfx_texture gfx_current_color(pxl8_gfx* gfx) {
    if (gfx->target_stack_depth == 0) return gfx->color_target;
    return gfx->target_stack[gfx->target_stack_depth - 1].color;
}

static pxl8_gfx_texture gfx_current_depth(pxl8_gfx* gfx) {
    if (gfx->target_stack_depth == 0) return gfx->depth_target;
    return gfx->target_stack[gfx->target_stack_depth - 1].depth;
}

static void gfx_composite_over(pxl8_gfx* gfx, const pxl8_target_entry* src, pxl8_target_entry* dst) {
    if (!gfx || !src || !dst) return;

    u8* src_color = (u8*)pxl8_texture_get_data(gfx->renderer, src->color);
    u8* dst_color = (u8*)pxl8_texture_get_data(gfx->renderer, dst->color);
    if (!src_color || !dst_color) return;

    u32 width = pxl8_texture_get_width(gfx->renderer, src->color);
    u32 height = pxl8_texture_get_height(gfx->renderer, src->color);
    if (width == 0 || height == 0) return;

    u32 dst_width = pxl8_texture_get_width(gfx->renderer, dst->color);
    u32 dst_height = pxl8_texture_get_height(gfx->renderer, dst->color);
    if (dst_width != width || dst_height != height) return;

    u32* src_light = (u32*)pxl8_texture_get_data(gfx->renderer, src->light);
    u32* dst_light = (u32*)pxl8_texture_get_data(gfx->renderer, dst->light);

    u32 count = width * height;
    for (u32 i = 0; i < count; i++) {
        u8 sc = src_color[i];
        if (sc == 0) continue;

        dst_color[i] = sc;
        if (dst_light) {
            dst_light[i] = src_light ? src_light[i] : 0;
        }
    }
}

void pxl8_2d_clear(pxl8_gfx* gfx, u32 color) {
    if (!gfx) return;
    pxl8_clear(gfx->renderer, gfx_current_color(gfx), (u8)color);
}

void pxl8_2d_pixel(pxl8_gfx* gfx, i32 x, i32 y, u32 color) {
    if (!gfx) return;
    pxl8_draw_pixel(gfx->renderer, gfx_current_color(gfx), x, y, (u8)color);
}

u32 pxl8_2d_get_pixel(pxl8_gfx* gfx, i32 x, i32 y) {
    if (!gfx) return 0;
    return pxl8_get_pixel(gfx->renderer, gfx_current_color(gfx), x, y);
}

void pxl8_2d_line(pxl8_gfx* gfx, i32 x0, i32 y0, i32 x1, i32 y1, u32 color) {
    if (!gfx) return;
    pxl8_draw_line(gfx->renderer, gfx_current_color(gfx), x0, y0, x1, y1, (u8)color);
}

void pxl8_2d_rect(pxl8_gfx* gfx, i32 x, i32 y, i32 w, i32 h, u32 color) {
    if (!gfx) return;
    pxl8_draw_rect(gfx->renderer, gfx_current_color(gfx), x, y, w, h, (u8)color);
}

void pxl8_2d_rect_fill(pxl8_gfx* gfx, i32 x, i32 y, i32 w, i32 h, u32 color) {
    if (!gfx) return;
    pxl8_draw_rect_fill(gfx->renderer, gfx_current_color(gfx), x, y, w, h, (u8)color);
}

void pxl8_2d_circle(pxl8_gfx* gfx, i32 cx, i32 cy, i32 radius, u32 color) {
    if (!gfx) return;
    pxl8_draw_circle(gfx->renderer, gfx_current_color(gfx), cx, cy, radius, (u8)color);
}

void pxl8_2d_circle_fill(pxl8_gfx* gfx, i32 cx, i32 cy, i32 radius, u32 color) {
    if (!gfx) return;
    pxl8_draw_circle_fill(gfx->renderer, gfx_current_color(gfx), cx, cy, radius, (u8)color);
}

void pxl8_2d_text(pxl8_gfx* gfx, const char* text, i32 x, i32 y, u32 color) {
    if (!gfx || !text) return;

    pxl8_gfx_texture target = gfx_current_color(gfx);
    u8* framebuffer = (u8*)pxl8_texture_get_data(gfx->renderer, target);
    i32 fb_width = (i32)pxl8_texture_get_width(gfx->renderer, target);
    i32 fb_height = (i32)pxl8_texture_get_height(gfx->renderer, target);

    if (!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 >= fb_width || py < 0 || py >= fb_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 idx = py * fb_width + px;
                    framebuffer[idx] = (u8)color;
                }
            }
        }

        cursor_x += font->default_width;
    }
}


void pxl8_2d_sprite(pxl8_gfx* gfx, u32 sprite_id, i32 x, i32 y, i32 w, i32 h, bool flip_x, bool flip_y) {
    if (!gfx || !gfx->atlas) return;

    pxl8_gfx_texture target = gfx_current_color(gfx);
    u8* framebuffer = (u8*)pxl8_texture_get_data(gfx->renderer, target);
    i32 fb_width = (i32)pxl8_texture_get_width(gfx->renderer, target);
    i32 fb_height = (i32)pxl8_texture_get_height(gfx->renderer, target);

    if (!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 > fb_width) ? x + w - fb_width : 0;
    i32 clip_bottom = (y + h > fb_height) ? y + h - fb_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);
    bool is_flipped = flip_x || flip_y;

    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 && !is_flipped) {
        const u8* sprite_data = atlas_pixels + entry->y * atlas_width + entry->x;
        pxl8_blit_indexed(framebuffer, fb_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 local_x = (px + clip_left) * entry->w / w;
                i32 local_y = (py + clip_top) * entry->h / h;

                i32 src_x = flip_x ? entry->x + entry->w - 1 - local_x : entry->x + local_x;
                i32 src_y = flip_y ? entry->y + entry->h - 1 - local_y : entry->y + local_y;

                i32 src_idx = src_y * atlas_width + src_x;
                i32 dest_idx = (dest_y + py) * fb_width + (dest_x + px);

                framebuffer[dest_idx] = pxl8_blend_indexed(atlas_pixels[src_idx], framebuffer[dest_idx]);
            }
        }
    }
}

void pxl8_gfx_update(pxl8_gfx* gfx, f32 dt) {
    if (!gfx) return;

    if (gfx->palette) {
        u16 delta_ticks = (u16)(dt * 1000.0f);
        pxl8_palette_tick(gfx->palette, delta_ticks);
    }
}

pxl8_3d_frame pxl8_3d_frame_from_camera(const pxl8_3d_camera* camera, const pxl8_shader_uniforms* uniforms) {
    pxl8_3d_frame frame = {0};
    if (!camera) return frame;

    frame.view = pxl8_3d_camera_get_view(camera);
    frame.projection = pxl8_3d_camera_get_projection(camera);
    frame.camera_pos = pxl8_3d_camera_get_position(camera);
    frame.camera_dir = pxl8_3d_camera_get_forward(camera);
    frame.near_clip = 1.0f;
    frame.far_clip = 4096.0f;

    if (uniforms) {
        frame.uniforms = *uniforms;
    }

    return frame;
}

void pxl8_3d_set_bsp(pxl8_gfx* gfx, const pxl8_bsp* bsp) {
    if (!gfx) return;
    gfx->bsp = bsp;
}

void pxl8_3d_begin_frame(pxl8_gfx* gfx, const pxl8_3d_camera* camera, const pxl8_lights* lights, const pxl8_shader_uniforms* uniforms) {
    if (!gfx || !camera) return;

    pxl8_3d_frame frame = pxl8_3d_frame_from_camera(camera, uniforms);
    frame.bsp = gfx->bsp;
    frame.uniforms.camera_pos = frame.camera_pos;
    frame.uniforms.lights = lights ? pxl8_lights_data(lights) : NULL;
    frame.uniforms.lights_count = lights ? pxl8_lights_count(lights) : 0;

    pxl8_mat4 vp = pxl8_mat4_multiply(frame.projection, frame.view);

    gfx->frustum = pxl8_frustum_from_matrix(vp);
    gfx->view_proj = vp;
    gfx->frame = frame;

    gfx->frame_res.texture_count = 0;
    gfx->frame_res.buffer_count = 0;
    gfx->frame_res.pipeline_count = 0;
    gfx->frame_res.bindings_count = 0;

    gfx->stream_vb_offset = 0;
    gfx->stream_ib_offset = 0;

    pxl8_cmdbuf_reset(gfx->cmdbuf);

    pxl8_gfx_pass_desc pass_desc = {
        .color = { .texture = gfx_current_color(gfx), .load = PXL8_GFX_LOAD_CLEAR, .clear_value = 0 },
        .depth = { .texture = gfx_current_depth(gfx), .load = PXL8_GFX_LOAD_CLEAR, .clear_value = 0xFFFF },
    };
    gfx->frame_pass = pxl8_create_pass(gfx->renderer, &pass_desc);
    pxl8_begin_pass(gfx->cmdbuf, gfx->frame_pass);
}

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

const pxl8_mat4* pxl8_3d_get_view_proj(pxl8_gfx* gfx) {
    if (!gfx) return NULL;
    return &gfx->view_proj;
}

u32 pxl8_3d_project_points(pxl8_gfx* gfx, const pxl8_vec3* in, pxl8_vec3* out, u32 count, const pxl8_mat4* transform) {
    if (!gfx || !in || !out) return 0;

    pxl8_mat4 mvp = transform ? pxl8_mat4_multiply(gfx->view_proj, *transform) : gfx->view_proj;

    f32 hw = (f32)gfx->framebuffer_width * 0.5f;
    f32 hh = (f32)gfx->framebuffer_height * 0.5f;
    u32 visible = 0;

    for (u32 i = 0; i < count; i++) {
        pxl8_vec4 clip = pxl8_mat4_multiply_vec4(mvp, (pxl8_vec4){in[i].x, in[i].y, in[i].z, 1.0f});

        if (clip.w <= 0.0f) {
            out[i].z = -1.0f;
            continue;
        }

        f32 inv_w = 1.0f / clip.w;
        out[i].x = hw + clip.x * inv_w * hw;
        out[i].y = hh - clip.y * inv_w * hh;
        out[i].z = clip.w;
        visible++;
    }

    return visible;
}

void pxl8_3d_clear(pxl8_gfx* gfx, u8 color) {
    if (!gfx) return;
    pxl8_clear(gfx->renderer, gfx_current_color(gfx), color);
}

void pxl8_3d_clear_depth(pxl8_gfx* gfx) {
    if (!gfx) return;
    pxl8_cmdbuf_clear_depth(gfx->cmdbuf, gfx_current_depth(gfx));
}

void pxl8_3d_draw_line(pxl8_gfx* gfx, pxl8_vec3 v0, pxl8_vec3 v1, u8 color) {
    if (!gfx) return;

    pxl8_vec4 c0 = pxl8_mat4_multiply_vec4(gfx->view_proj, (pxl8_vec4){v0.x, v0.y, v0.z, 1.0f});
    pxl8_vec4 c1 = pxl8_mat4_multiply_vec4(gfx->view_proj, (pxl8_vec4){v1.x, v1.y, v1.z, 1.0f});

    if (c0.w <= 0.0f && c1.w <= 0.0f) return;

    f32 hw = (f32)gfx->framebuffer_width * 0.5f;
    f32 hh = (f32)gfx->framebuffer_height * 0.5f;

    i32 x0 = (i32)(hw + c0.x / c0.w * hw);
    i32 y0 = (i32)(hh - c0.y / c0.w * hh);
    i32 x1 = (i32)(hw + c1.x / c1.w * hw);
    i32 y1 = (i32)(hh - c1.y / c1.w * hh);

    pxl8_draw_line(gfx->renderer, gfx_current_color(gfx), x0, y0, x1, y1, color);
}

void pxl8_3d_draw_mesh(pxl8_gfx* gfx, const pxl8_mesh* mesh, const pxl8_mat4* model, const pxl8_gfx_material* material) {
    if (!gfx || !mesh || !model || !material) return;
    if (!pxl8_gfx_handle_valid(gfx->frame_pass)) return;

    pxl8_frame_resources* res = &gfx->frame_res;
    bool is_wireframe = gfx->wireframe;

    const char* shader_name = "lit";
    if (material->emissive || (!material->dynamic_lighting && !material->per_pixel)) {
        shader_name = "unlit";
    }

    pxl8_gfx_pipeline_desc pipe_desc = {
        .blend = {
            .enabled = false,
            .src = PXL8_GFX_BLEND_ONE,
            .dst = PXL8_GFX_BLEND_ZERO,
            .alpha_test = false,
            .alpha_ref = 0,
        },
        .depth = { .test = true, .write = true, .compare = PXL8_GFX_COMPARE_LESS },
        .dither = material->dither,
        .double_sided = material->double_sided,
        .emissive = material->emissive,
        .rasterizer = { .cull = PXL8_GFX_CULL_BACK, .fill = is_wireframe ? PXL8_GFX_FILL_WIREFRAME : PXL8_GFX_FILL_SOLID },
        .shader = pxl8_shader_registry_get(shader_name),
    };

    switch (material->blend_mode) {
        case PXL8_BLEND_ALPHA_TEST:
            pipe_desc.blend.alpha_test = true;
            pipe_desc.blend.alpha_ref = material->alpha;
            break;
        case PXL8_BLEND_ALPHA:
            pipe_desc.blend.enabled = true;
            pipe_desc.blend.src = PXL8_GFX_BLEND_SRC_ALPHA;
            pipe_desc.blend.dst = PXL8_GFX_BLEND_INV_SRC_ALPHA;
            break;
        case PXL8_BLEND_ADDITIVE:
            pipe_desc.blend.enabled = true;
            pipe_desc.blend.src = PXL8_GFX_BLEND_ONE;
            pipe_desc.blend.dst = PXL8_GFX_BLEND_ONE;
            break;
        case PXL8_BLEND_OPAQUE:
        default:
            break;
    }
    pxl8_gfx_pipeline pipeline = pxl8_create_pipeline(gfx->renderer, &pipe_desc);
    if (res->pipeline_count < PXL8_MAX_FRAME_RESOURCES) {
        res->pipelines[res->pipeline_count++] = pipeline;
    }

    pxl8_gfx_bindings_desc bind_desc = {
        .atlas = gfx->atlas,
        .colormap = gfx->colormap,
        .palette = gfx->palette ? pxl8_palette_colors(gfx->palette) : NULL,
        .texture_id = material->texture_id,
    };
    pxl8_gfx_bindings bindings = pxl8_create_bindings(gfx->renderer, &bind_desc);
    if (res->bindings_count < PXL8_MAX_FRAME_RESOURCES) {
        res->bindings[res->bindings_count++] = bindings;
    }

    u32 vb_size = mesh->vertex_count * sizeof(pxl8_vertex);
    u32 ib_size = mesh->index_count * sizeof(u16);

    u32 base_vertex = 0;
    u32 first_index = 0;
    pxl8_gfx_buffer vb, ib;

    bool use_stream = pxl8_gfx_handle_valid(gfx->stream_vb) &&
                      pxl8_gfx_handle_valid(gfx->stream_ib) &&
                      (gfx->stream_vb_offset + vb_size <= gfx->stream_vb_capacity) &&
                      (gfx->stream_ib_offset + ib_size <= gfx->stream_ib_capacity);

    if (use_stream) {
        pxl8_gfx_range vb_data = { .ptr = mesh->vertices, .size = vb_size };
        pxl8_gfx_range ib_data = { .ptr = mesh->indices, .size = ib_size };
        i32 vb_offset = pxl8_append_buffer(gfx->renderer, gfx->stream_vb, &vb_data);
        i32 ib_offset = pxl8_append_buffer(gfx->renderer, gfx->stream_ib, &ib_data);

        if (vb_offset >= 0 && ib_offset >= 0) {
            gfx->stream_vb_offset += vb_size;
            gfx->stream_ib_offset += ib_size;

            base_vertex = (u32)vb_offset / sizeof(pxl8_vertex);
            first_index = (u32)ib_offset / sizeof(u16);

            vb = gfx->stream_vb;
            ib = gfx->stream_ib;
        } else {
            use_stream = false;
        }
    }

    if (!use_stream) {
        pxl8_gfx_buffer_desc vb_desc = {
            .type = PXL8_GFX_BUFFER_VERTEX,
            .data = { .ptr = mesh->vertices, .size = vb_size },
        };
        vb = pxl8_create_buffer(gfx->renderer, &vb_desc);
        if (res->buffer_count < PXL8_MAX_FRAME_RESOURCES) {
            res->buffers[res->buffer_count++] = vb;
        }

        pxl8_gfx_buffer_desc ib_desc = {
            .type = PXL8_GFX_BUFFER_INDEX,
            .data = { .ptr = mesh->indices, .size = ib_size },
        };
        ib = pxl8_create_buffer(gfx->renderer, &ib_desc);
        if (res->buffer_count < PXL8_MAX_FRAME_RESOURCES) {
            res->buffers[res->buffer_count++] = ib;
        }
    }

    pxl8_gfx_cmd_draw_params draw_params = {
        .model = *model,
        .view = gfx->frame.view,
        .projection = gfx->frame.projection,
        .shader = gfx->frame.uniforms,
    };

    pxl8_set_pipeline(gfx->cmdbuf, pipeline);
    pxl8_set_bindings(gfx->cmdbuf, bindings);
    pxl8_set_draw_params(gfx->cmdbuf, &draw_params);
    pxl8_draw(gfx->cmdbuf, vb, ib, first_index, mesh->index_count, base_vertex);
}

void pxl8_3d_end_frame(pxl8_gfx* gfx) {
    if (!gfx) { pxl8_error("end_frame: gfx is NULL"); return; }
    if (!pxl8_gfx_handle_valid(gfx->frame_pass)) { pxl8_error("end_frame: frame_pass invalid (id=%u)", gfx->frame_pass.id); return; }

    pxl8_end_pass(gfx->cmdbuf);
    pxl8_gfx_submit(gfx->renderer, gfx->cmdbuf);

    pxl8_frame_resources* res = &gfx->frame_res;

    for (u32 i = 0; i < res->buffer_count; i++) {
        pxl8_destroy_buffer(gfx->renderer, res->buffers[i]);
    }
    for (u32 i = 0; i < res->pipeline_count; i++) {
        pxl8_destroy_pipeline(gfx->renderer, res->pipelines[i]);
    }
    for (u32 i = 0; i < res->bindings_count; i++) {
        pxl8_destroy_bindings(gfx->renderer, res->bindings[i]);
    }
    for (u32 i = 0; i < res->texture_count; i++) {
        pxl8_destroy_texture(gfx->renderer, res->textures[i]);
    }

    pxl8_destroy_pass(gfx->renderer, gfx->frame_pass);
    gfx->frame_pass = (pxl8_gfx_pass){PXL8_GFX_INVALID_ID};

    res->buffer_count = 0;
    res->pipeline_count = 0;
    res->bindings_count = 0;
    res->texture_count = 0;
}

u8* pxl8_3d_get_framebuffer(pxl8_gfx* gfx) {
    if (!gfx) return NULL;
    return (u8*)pxl8_texture_get_data(gfx->renderer, gfx_current_color(gfx));
}

bool pxl8_gfx_push_target(pxl8_gfx* gfx) {
    if (!gfx || gfx->target_stack_depth >= PXL8_MAX_TARGET_STACK) return false;

    pxl8_gfx_texture_desc color_desc = {
        .width = (u32)gfx->framebuffer_width,
        .height = (u32)gfx->framebuffer_height,
        .format = PXL8_GFX_FORMAT_INDEXED8,
        .render_target = true,
    };
    pxl8_gfx_texture new_color = pxl8_create_texture(gfx->renderer, &color_desc);

    pxl8_gfx_texture_desc depth_desc = {
        .width = (u32)gfx->framebuffer_width,
        .height = (u32)gfx->framebuffer_height,
        .format = PXL8_GFX_FORMAT_DEPTH16,
        .render_target = true,
    };
    pxl8_gfx_texture new_depth = pxl8_create_texture(gfx->renderer, &depth_desc);

    gfx->target_stack[gfx->target_stack_depth] = (pxl8_target_entry){
        .color = new_color,
        .depth = new_depth,
    };
    gfx->target_stack_depth++;

    return true;
}

void pxl8_gfx_pop_target(pxl8_gfx* gfx) {
    if (!gfx || gfx->target_stack_depth <= 1) return;

    u32 top = gfx->target_stack_depth - 1;
    pxl8_target_entry* entry = &gfx->target_stack[top];
    pxl8_target_entry* parent = &gfx->target_stack[top - 1];

    gfx_composite_over(gfx, entry, parent);

    gfx->target_stack_depth--;

    pxl8_destroy_texture(gfx->renderer, entry->color);
    pxl8_destroy_texture(gfx->renderer, entry->depth);
    pxl8_destroy_texture(gfx->renderer, entry->light);
}

void pxl8_gfx_ensure_blend_tables(pxl8_gfx* gfx) {
    if (!gfx || !gfx->palette) return;

    if (!gfx->palette_cube) {
        gfx->palette_cube = pxl8_palette_cube_create(gfx->palette);
    }
}

void pxl8_gfx_blend_tables_update(pxl8_gfx* gfx) {
    if (!gfx || !gfx->palette) return;

    pxl8_gfx_ensure_blend_tables(gfx);

    if (gfx->palette_cube) {
        pxl8_palette_cube_rebuild(gfx->palette_cube, gfx->palette);
    }
}

void pxl8_gfx_colormap_update(pxl8_gfx* gfx) {
    if (!gfx || !gfx->palette || !gfx->colormap) return;
    u32* colors = pxl8_palette_colors(gfx->palette);
    if (colors) {
        pxl8_colormap_generate(gfx->colormap, colors);
    }
}

u8 pxl8_gfx_find_closest_color(pxl8_gfx* gfx, u8 r, u8 g, u8 b) {
    if (!gfx || !gfx->palette) return 0;
    return pxl8_palette_find_closest(gfx->palette, r, g, b);
}

u8 pxl8_gfx_ui_color(pxl8_gfx* gfx, u8 index) {
    if (!gfx || !gfx->palette || index >= PXL8_UI_PALETTE_SIZE) return 0;
    u32 abgr = pxl8_ui_palette[index];
    u8 r = (abgr >> 0) & 0xFF;
    u8 g = (abgr >> 8) & 0xFF;
    u8 b = (abgr >> 16) & 0xFF;
    return pxl8_palette_find_closest(gfx->palette, r, g, b);
}

void pxl8_gfx_set_wireframe(pxl8_gfx* gfx, bool enabled) {
    if (gfx) gfx->wireframe = enabled;
}

bool pxl8_gfx_get_wireframe(const pxl8_gfx* gfx) {
    return gfx ? gfx->wireframe : false;
}

void pxl8_gfx_apply_effect(pxl8_gfx* gfx, pxl8_gfx_effect effect, const void* params, u32 count) {
    if (!gfx || !params || count == 0) return;

    if (effect == PXL8_GFX_EFFECT_GLOWS) {
        u8* fb = (u8*)pxl8_texture_get_data(gfx->renderer, gfx_current_color(gfx));
        if (!fb || !gfx->colormap) return;

        u16* zb = (u16*)pxl8_texture_get_data(gfx->renderer, gfx_current_depth(gfx));
        i32 w = gfx->framebuffer_width;
        i32 h = gfx->framebuffer_height;
        const pxl8_glow* glows = (const pxl8_glow*)params;

        for (u32 i = 0; i < count; i++) {
            const pxl8_glow* g = &glows[i];
            i32 cx = g->x;
            i32 cy = g->y;
            i32 r = (i32)g->radius;
            if (r <= 0) continue;

            bool depth_test = g->depth > 0 && zb;
            f32 base_intensity = (f32)g->intensity / 255.0f;
            f32 inv_r = 1.0f / (f32)r;

            i32 x0 = cx - r;
            i32 y0 = cy - r;
            i32 x1 = cx + r;
            i32 y1 = cy + r;

            if (g->shape == PXL8_GLOW_SHAFT) {
                i32 sh = (i32)g->height;
                if (sh <= 0) sh = 1;
                y0 = cy - sh;
                y1 = cy + sh;
            }

            if (x0 < 0) x0 = 0;
            if (y0 < 0) y0 = 0;
            if (x1 >= w) x1 = w - 1;
            if (y1 >= h) y1 = h - 1;

            for (i32 py = y0; py <= y1; py++) {
                i32 dy = py - cy;
                for (i32 px = x0; px <= x1; px++) {
                    i32 idx = py * w + px;

                    if (depth_test && zb[idx] < g->depth) continue;

                    u8 src = fb[idx];
                    if (src == PXL8_TRANSPARENT) continue;

                    i32 dx = px - cx;
                    f32 falloff;

                    switch (g->shape) {
                    case PXL8_GLOW_CIRCLE: {
                        f32 dist_sq = (f32)(dx * dx + dy * dy);
                        f32 norm = dist_sq * inv_r * inv_r;
                        if (norm >= 1.0f) continue;
                        falloff = 1.0f - norm;
                        falloff *= falloff;
                        break;
                    }
                    case PXL8_GLOW_DIAMOND: {
                        f32 manhattan = (f32)(abs(dx) + abs(dy));
                        f32 norm = manhattan * inv_r;
                        if (norm >= 1.0f) continue;
                        falloff = 1.0f - norm;
                        falloff *= falloff;
                        break;
                    }
                    case PXL8_GLOW_SHAFT: {
                        f32 abs_dx = (f32)abs(dx);
                        if (abs_dx >= (f32)r) continue;
                        f32 x_falloff = 1.0f - abs_dx * inv_r;
                        i32 sh = (i32)g->height;
                        if (sh <= 0) sh = 1;
                        f32 abs_dy = (f32)abs(dy);
                        f32 y_falloff = 1.0f - abs_dy / (f32)sh;
                        falloff = x_falloff * y_falloff;
                        falloff *= falloff;
                        break;
                    }
                    default: continue;
                    }

                    f32 brightness = falloff * base_intensity;
                    f32 raw = brightness * 255.0f;
                    if (raw > 255.0f) raw = 255.0f;
                    u8 light_u8 = (u8)raw;
                    if (light_u8 < 4) continue;

                    fb[idx] = pxl8_colormap_additive_dithered(
                        gfx->colormap, src, g->color,
                        light_u8, (u32)px, (u32)py
                    );
                }
            }
        }
    }
}

u8 pxl8_gfx_get_ambient(const pxl8_gfx* gfx) {
    return gfx ? gfx->frame.uniforms.ambient : 0;
}

const pxl8_gfx_stats* pxl8_gfx_get_stats(const pxl8_gfx* gfx) {
    return gfx ? pxl8_renderer_get_stats(gfx->renderer) : NULL;
}

void pxl8_gfx_update_stats(pxl8_gfx* gfx, f32 dt) {
    if (gfx) pxl8_renderer_update_stats(gfx->renderer, dt);
}