#include <stdlib.h>

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

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_result pxl8_gfx_init(
    pxl8_gfx_ctx* ctx,
    pxl8_color_mode mode,
    pxl8_resolution resolution,
    const char* title,
    i32 window_width,
    i32 window_height
) {
    memset(ctx, 0, sizeof(pxl8_gfx_ctx));
    
    ctx->color_mode = mode;
    pxl8_gfx_get_resolution_dimensions(resolution, &ctx->framebuffer_width, &ctx->framebuffer_height);
    
    ctx->window = SDL_CreateWindow(
        title,
        window_width, window_height,
        SDL_WINDOW_RESIZABLE
    );
    
    if (!ctx->window) {
        pxl8_error("Failed to create window: %s", SDL_GetError());
        return PXL8_ERROR_SYSTEM_FAILURE;
    }
    
    ctx->renderer = SDL_CreateRenderer(ctx->window, NULL);
    if (!ctx->renderer) {
        pxl8_error("Failed to create renderer: %s", SDL_GetError());
        SDL_DestroyWindow(ctx->window);
        return PXL8_ERROR_SYSTEM_FAILURE;
    }
    
    SDL_SetRenderLogicalPresentation(ctx->renderer, 
                                     ctx->framebuffer_width, 
                                     ctx->framebuffer_height,
                                     SDL_LOGICAL_PRESENTATION_INTEGER_SCALE);
    
    i32 bytes_per_pixel = (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) ? 4 : 1;
    i32 fb_size = ctx->framebuffer_width * ctx->framebuffer_height * bytes_per_pixel;
    ctx->framebuffer = (u8*)SDL_calloc(1, fb_size);
    if (!ctx->framebuffer) {
        pxl8_error("Failed to allocate framebuffer");
        pxl8_gfx_shutdown(ctx);
        return PXL8_ERROR_OUT_OF_MEMORY;
    }
    
    ctx->framebuffer_texture = SDL_CreateTexture(
        ctx->renderer,
        SDL_PIXELFORMAT_RGBA32,
        SDL_TEXTUREACCESS_STREAMING,
        ctx->framebuffer_width,
        ctx->framebuffer_height
    );
    
    SDL_SetTextureScaleMode(ctx->framebuffer_texture, SDL_SCALEMODE_NEAREST);
    
    if (!ctx->framebuffer_texture) {
        pxl8_error("Failed to create framebuffer texture: %s", SDL_GetError());
        pxl8_gfx_shutdown(ctx);
        return PXL8_ERROR_SYSTEM_FAILURE;
    }
    
    ctx->palette_size = pxl8_get_palette_size(mode);
    if (ctx->palette_size > 0) {
        ctx->palette = (u32*)SDL_calloc(ctx->palette_size, sizeof(u32));
        if (!ctx->palette) {
            pxl8_error("Failed to allocate palette");
            pxl8_gfx_shutdown(ctx);
            return PXL8_ERROR_OUT_OF_MEMORY;
        }
        
        for (u32 i = 0; i < 256 && i < ctx->palette_size; i++) {
            u8 gray = (u8)(i * 255 / 255);
            ctx->palette[i] = 0xFF000000 | (gray << 16) | (gray << 8) | gray;
        }
    }
    
    ctx->viewport_x = 0;
    ctx->viewport_y = 0;
    ctx->viewport_width = ctx->framebuffer_width;
    ctx->viewport_height = ctx->framebuffer_height;
    
    ctx->initialized = true;
    return PXL8_OK;
}

void pxl8_gfx_shutdown(pxl8_gfx_ctx* ctx) {
    if (!ctx) return;
    
    if (ctx->framebuffer_texture) {
        SDL_DestroyTexture(ctx->framebuffer_texture);
        ctx->framebuffer_texture = NULL;
    }
    
    if (ctx->sprite_atlas_texture) {
        SDL_DestroyTexture(ctx->sprite_atlas_texture);
        ctx->sprite_atlas_texture = NULL;
    }
    
    if (ctx->renderer) {
        SDL_DestroyRenderer(ctx->renderer);
        ctx->renderer = NULL;
    }
    
    if (ctx->window) {
        SDL_DestroyWindow(ctx->window);
        ctx->window = NULL;
    }
    
    SDL_free(ctx->framebuffer);
    SDL_free(ctx->palette);
    SDL_free(ctx->atlas);
    SDL_free(ctx->atlas_entries);
    
    ctx->initialized = false;
}

// resource loading
pxl8_result pxl8_gfx_init_atlas(pxl8_gfx_ctx* ctx, u32 width, u32 height) {
    if (!ctx || !ctx->initialized) return PXL8_ERROR_INVALID_ARGUMENT;
    
    ctx->sprite_atlas_width = width;
    ctx->sprite_atlas_height = height;
    
    i32 bytes_per_pixel = (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) ? 4 : 1;
    ctx->atlas = (u8*)SDL_calloc(width * height, bytes_per_pixel);
    if (!ctx->atlas) {
        return PXL8_ERROR_OUT_OF_MEMORY;
    }
    
    ctx->sprite_atlas_texture = SDL_CreateTexture(
        ctx->renderer,
        SDL_PIXELFORMAT_RGBA32,
        SDL_TEXTUREACCESS_STREAMING,
        width,
        height
    );
    
    if (!ctx->sprite_atlas_texture) {
        SDL_free(ctx->atlas);
        ctx->atlas = NULL;
        return PXL8_ERROR_SYSTEM_FAILURE;
    }
    
    ctx->atlas_entries_cap = 256;
    ctx->atlas_entries = (pxl8_atlas_entry*)SDL_calloc(ctx->atlas_entries_cap, sizeof(pxl8_atlas_entry));
    if (!ctx->atlas_entries) {
        SDL_DestroyTexture(ctx->sprite_atlas_texture);
        SDL_free(ctx->atlas);
        return PXL8_ERROR_OUT_OF_MEMORY;
    }
    
    ctx->sprite_frames_per_row = width / 128;
    if (ctx->sprite_frames_per_row == 0) ctx->sprite_frames_per_row = 1;
    
    return PXL8_OK;
}

pxl8_result pxl8_gfx_load_sprite(pxl8_gfx_ctx* ctx, const char* path) {
    if (!ctx || !ctx->initialized || !path) return PXL8_ERROR_INVALID_ARGUMENT;
    if (!ctx->atlas) {
        pxl8_error("Atlas not initialized");
        return PXL8_ERROR_NOT_INITIALIZED;
    }
    
    for (u32 i = 0; i < ctx->atlas_entries_len; i++) {
        if (strcmp(ctx->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_free(&ase_file);
        return PXL8_ERROR_INVALID_FORMAT;
    }
    
    u32 sprite_w = ase_file.header.width;
    u32 sprite_h = ase_file.header.height;
    
    if (ctx->sprite_frame_width == 0 || ctx->sprite_frame_height == 0) {
        ctx->sprite_frame_width = sprite_w;
        ctx->sprite_frame_height = sprite_h;
    }
    
    u32 id = ctx->atlas_entries_len;
    u32 frames_per_row = ctx->sprite_frames_per_row;
    u32 atlas_x = (id % frames_per_row) * ctx->sprite_frame_width;
    u32 atlas_y = (id / frames_per_row) * ctx->sprite_frame_height;
    
    if (atlas_x + sprite_w > ctx->sprite_atlas_width || 
        atlas_y + sprite_h > ctx->sprite_atlas_height) {
        pxl8_error("Sprite doesn't fit in atlas");
        pxl8_ase_free(&ase_file);
        return PXL8_ERROR_INVALID_SIZE;
    }
    
    i32 bytes_per_pixel = (ctx->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) * ctx->sprite_atlas_width + (atlas_x + x);
            
            if (bytes_per_pixel == 4) {
                ((u32*)ctx->atlas)[atlas_idx] = ((u32*)ase_file.frames[0].pixels)[frame_idx];
            } else {
                ctx->atlas[atlas_idx] = ase_file.frames[0].pixels[frame_idx];
            }
        }
    }
    
    if (ctx->atlas_entries_len >= ctx->atlas_entries_cap) {
        ctx->atlas_entries_cap *= 2;
        pxl8_atlas_entry* new_entries = (pxl8_atlas_entry*)SDL_realloc(
            ctx->atlas_entries, 
            ctx->atlas_entries_cap * sizeof(pxl8_atlas_entry)
        );
        if (!new_entries) {
            pxl8_ase_free(&ase_file);
            return PXL8_ERROR_OUT_OF_MEMORY;
        }
        ctx->atlas_entries = new_entries;
    }
    
    ctx->atlas_entries[id].x = atlas_x;
    ctx->atlas_entries[id].y = atlas_y;
    ctx->atlas_entries[id].w = sprite_w;
    ctx->atlas_entries[id].h = sprite_h;
    ctx->atlas_entries[id].sprite_id = id;
    strncpy(ctx->atlas_entries[id].path, path, sizeof(ctx->atlas_entries[id].path) - 1);
    ctx->atlas_entries[id].path[sizeof(ctx->atlas_entries[id].path) - 1] = '\0';
    ctx->atlas_entries_len++;
    
    ctx->atlas_dirty = true;
    
    pxl8_ase_free(&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_ctx* ctx, const char* path) {
    if (!ctx || !ctx->initialized || !path) return PXL8_ERROR_INVALID_ARGUMENT;
    if (ctx->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_free(&ase_file);
        return PXL8_ERROR_INVALID_FORMAT;
    }
    
    u32 copy_size = (ase_file.palette.entry_count < ctx->palette_size) ? ase_file.palette.entry_count : ctx->palette_size;
    memcpy(ctx->palette, ase_file.palette.colors, copy_size * sizeof(u32));
    
    if (ctx->color_mode != PXL8_COLOR_MODE_HICOLOR && ctx->framebuffer_texture) {
        SDL_Color colors[256];
        for (u32 i = 0; i < 256 && i < ctx->palette_size; i++) {
            colors[i].r = (ctx->palette[i] >> 16) & 0xFF;
            colors[i].g = (ctx->palette[i] >> 8) & 0xFF;
            colors[i].b = ctx->palette[i] & 0xFF;
            colors[i].a = (ctx->palette[i] >> 24) & 0xFF;
        }
        SDL_SetPaletteColors(SDL_CreateSurfacePalette(NULL), colors, 0, 256);
    }
    
    pxl8_ase_free(&ase_file);
    pxl8_debug("Loaded palette with %u colors", copy_size);
    
    return PXL8_OK;
}

pxl8_result pxl8_gfx_load_font_atlas(pxl8_gfx_ctx* ctx) {
    (void)ctx;
    return PXL8_OK;
}

// rendering pipeline
void pxl8_gfx_upload_framebuffer(pxl8_gfx_ctx* ctx) {
    if (!ctx || !ctx->initialized || !ctx->framebuffer_texture) return;
    
    if (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        SDL_UpdateTexture(ctx->framebuffer_texture, NULL, ctx->framebuffer, 
                         ctx->framebuffer_width * 4);
    } else {
        static u32* rgba_buffer = NULL;
        static size_t buffer_size = 0;
        size_t needed_size = ctx->framebuffer_width * ctx->framebuffer_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 < ctx->framebuffer_width * ctx->framebuffer_height; i++) {
            u8 index = ctx->framebuffer[i];
            rgba_buffer[i] = (index < ctx->palette_size) ? ctx->palette[index] : 0xFF000000;
        }
        
        SDL_UpdateTexture(ctx->framebuffer_texture, NULL, rgba_buffer, 
                         ctx->framebuffer_width * 4);
    }
}

void pxl8_gfx_upload_atlas(pxl8_gfx_ctx* ctx) {
    if (!ctx || !ctx->initialized || !ctx->sprite_atlas_texture || !ctx->atlas_dirty) return;
    
    if (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        SDL_UpdateTexture(ctx->sprite_atlas_texture, NULL, ctx->atlas, 
                         ctx->sprite_atlas_width * 4);
    } else {
        u32* rgba_buffer = (u32*)SDL_malloc(ctx->sprite_atlas_width * ctx->sprite_atlas_height * 4);
        if (!rgba_buffer) return;
        
        for (u32 i = 0; i < ctx->sprite_atlas_width * ctx->sprite_atlas_height; i++) {
            u8 index = ctx->atlas[i];
            rgba_buffer[i] = (index < ctx->palette_size) ? ctx->palette[index] : 0x00000000;
        }
        
        SDL_UpdateTexture(ctx->sprite_atlas_texture, NULL, rgba_buffer, 
                         ctx->sprite_atlas_width * 4);
        SDL_free(rgba_buffer);
    }
    
    ctx->atlas_dirty = false;
    pxl8_debug("Atlas uploaded to GPU");
}

void pxl8_gfx_present(pxl8_gfx_ctx* ctx) {
    if (!ctx || !ctx->initialized) return;
    
    SDL_SetRenderDrawColor(ctx->renderer, 0, 0, 0, 255);
    SDL_RenderClear(ctx->renderer);
    
    if (ctx->framebuffer_texture) {
        SDL_RenderTexture(ctx->renderer, ctx->framebuffer_texture, NULL, NULL);
    }
    
    SDL_RenderPresent(ctx->renderer);
}

void pxl8_gfx_viewport(pxl8_gfx_ctx* ctx, i32 x, i32 y, i32 width, i32 height) {
    if (!ctx) return;
    ctx->viewport_x = x;
    ctx->viewport_y = y;
    ctx->viewport_width = width;
    ctx->viewport_height = height;
}

void pxl8_gfx_project(pxl8_gfx_ctx* ctx, f32 left, f32 right, f32 top, f32 bottom) {
    (void)ctx; (void)left; (void)right; (void)top; (void)bottom;
}

// drawing primitives
void pxl8_clr(pxl8_gfx_ctx* ctx, u32 color) {
    if (!ctx || !ctx->framebuffer) return;
    
    i32 bytes_per_pixel = (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) ? 4 : 1;
    i32 size = ctx->framebuffer_width * ctx->framebuffer_height;
    
    if (bytes_per_pixel == 4) {
        u32* fb32 = (u32*)ctx->framebuffer;
        for (i32 i = 0; i < size; i++) {
            fb32[i] = color;
        }
    } else {
        memset(ctx->framebuffer, color & 0xFF, size);
    }
}

void pxl8_pixel(pxl8_gfx_ctx* ctx, i32 x, i32 y, u32 color) {
    if (!ctx || !ctx->framebuffer) return;
    if (x < 0 || x >= ctx->framebuffer_width || y < 0 || y >= ctx->framebuffer_height) return;
    
    i32 idx = y * ctx->framebuffer_width + x;
    if (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        ((u32*)ctx->framebuffer)[idx] = color;
    } else {
        ctx->framebuffer[idx] = color & 0xFF;
    }
}

u32 pxl8_get_pixel(pxl8_gfx_ctx* ctx, i32 x, i32 y) {
    if (!ctx || !ctx->framebuffer) return 0;
    if (x < 0 || x >= ctx->framebuffer_width || y < 0 || y >= ctx->framebuffer_height) return 0;
    
    i32 idx = y * ctx->framebuffer_width + x;
    if (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
        return ((u32*)ctx->framebuffer)[idx];
    } else {
        return ctx->framebuffer[idx];
    }
}

void pxl8_line(pxl8_gfx_ctx* ctx, i32 x0, i32 y0, i32 x1, i32 y1, u32 color) {
    if (!ctx) 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(ctx, 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_ctx* ctx, i32 x, i32 y, i32 w, i32 h, u32 color) {
    if (!ctx) return;
    
    pxl8_line(ctx, x, y, x + w - 1, y, color);
    pxl8_line(ctx, x + w - 1, y, x + w - 1, y + h - 1, color);
    pxl8_line(ctx, x + w - 1, y + h - 1, x, y + h - 1, color);
    pxl8_line(ctx, x, y + h - 1, x, y, color);
}

void pxl8_rect_fill(pxl8_gfx_ctx* ctx, i32 x, i32 y, i32 w, i32 h, u32 color) {
    if (!ctx) return;
    
    for (i32 py = y; py < y + h; py++) {
        for (i32 px = x; px < x + w; px++) {
            pxl8_pixel(ctx, px, py, color);
        }
    }
}

void pxl8_circle(pxl8_gfx_ctx* ctx, i32 cx, i32 cy, i32 radius, u32 color) {
    if (!ctx) return;
    
    i32 x = radius;
    i32 y = 0;
    i32 err = 0;
    
    while (x >= y) {
        pxl8_pixel(ctx, cx + x, cy + y, color);
        pxl8_pixel(ctx, cx + y, cy + x, color);
        pxl8_pixel(ctx, cx - y, cy + x, color);
        pxl8_pixel(ctx, cx - x, cy + y, color);
        pxl8_pixel(ctx, cx - x, cy - y, color);
        pxl8_pixel(ctx, cx - y, cy - x, color);
        pxl8_pixel(ctx, cx + y, cy - x, color);
        pxl8_pixel(ctx, 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_ctx* ctx, i32 cx, i32 cy, i32 radius, u32 color) {
    if (!ctx) return;
    
    for (i32 y = -radius; y <= radius; y++) {
        for (i32 x = -radius; x <= radius; x++) {
            if (x * x + y * y <= radius * radius) {
                pxl8_pixel(ctx, cx + x, cy + y, color);
            }
        }
    }
}

void pxl8_text(pxl8_gfx_ctx* ctx, const char* text, i32 x, i32 y, u32 color) {
    if (!ctx || !text) return;
    (void)x; (void)y; (void)color;
}


void pxl8_sprite(pxl8_gfx_ctx* ctx, u32 sprite_id, i32 x, i32 y, i32 w, i32 h) {
    if (!ctx || !ctx->atlas || !ctx->framebuffer || sprite_id >= ctx->atlas_entries_len) return;
    
    pxl8_atlas_entry* entry = &ctx->atlas_entries[sprite_id];
    
    i32 clip_left = (x < 0) ? -x : 0;
    i32 clip_top = (y < 0) ? -y : 0;
    i32 clip_right = (x + w > ctx->framebuffer_width) ? x + w - ctx->framebuffer_width : 0;
    i32 clip_bottom = (y + h > ctx->framebuffer_height) ? y + h - ctx->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 = ctx->atlas + entry->y * ctx->sprite_atlas_width + entry->x;
        
        if (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
            pxl8_blit_simd_hicolor((u32*)ctx->framebuffer, ctx->framebuffer_width, 
                                   (const u32*)sprite_data, ctx->sprite_atlas_width, x, y, w, h);
        } else {
            pxl8_blit_simd_indexed(ctx->framebuffer, ctx->framebuffer_width, 
                                   sprite_data, ctx->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 * ctx->sprite_atlas_width + src_x;
                i32 dest_idx = (dest_y + py) * ctx->framebuffer_width + (dest_x + px);
                
                if (ctx->color_mode == PXL8_COLOR_MODE_HICOLOR) {
                    u32 pixel = ((u32*)ctx->atlas)[src_idx];
                    if (pixel & 0xFF000000) {
                        ((u32*)ctx->framebuffer)[dest_idx] = pixel;
                    }
                } else {
                    u8 pixel = ctx->atlas[src_idx];
                    if (pixel != 0) {
                        ctx->framebuffer[dest_idx] = pixel;
                    }
                }
            }
        }
    }
}

// palette effects
void pxl8_gfx_cycle_palette(pxl8_gfx_ctx* ctx, u8 start, u8 count, i32 step) {
    if (!ctx || !ctx->palette || count == 0) return;
    
    u32 temp[256];
    for (u8 i = 0; i < count; i++) {
        temp[i] = ctx->palette[start + i];
    }
    
    for (u8 i = 0; i < count; i++) {
        u8 src_idx = i;
        u8 dst_idx = (i + step) % count;
        ctx->palette[start + dst_idx] = temp[src_idx];
    }
}

void pxl8_gfx_swap_palette(pxl8_gfx_ctx* ctx, u8 start, u8 count, u32* new_colors) {
    if (!ctx || !ctx->palette || !new_colors) return;
    
    for (u8 i = 0; i < count && (start + i) < ctx->palette_size; i++) {
        ctx->palette[start + i] = new_colors[i];
    }
}

void pxl8_gfx_fade_palette(pxl8_gfx_ctx* ctx, u8 start, u8 count, f32 amount, u32 target_color) {
    if (!ctx || !ctx->palette || count == 0) return;
    
    if (amount < 0.0f) amount = 0.0f;
    if (amount > 1.0f) amount = 1.0f;
    
    u8 target_r = target_color & 0xFF;
    u8 target_g = (target_color >> 8) & 0xFF;
    u8 target_b = (target_color >> 16) & 0xFF;
    u8 target_a = (target_color >> 24) & 0xFF;
    
    for (u8 i = 0; i < count && (start + i) < ctx->palette_size; i++) {
        u32 current = ctx->palette[start + i];
        u8 cur_r = current & 0xFF;
        u8 cur_g = (current >> 8) & 0xFF;
        u8 cur_b = (current >> 16) & 0xFF;
        u8 cur_a = (current >> 24) & 0xFF;
        
        u8 new_r = cur_r + (i32)((target_r - cur_r) * amount);
        u8 new_g = cur_g + (i32)((target_g - cur_g) * amount);
        u8 new_b = cur_b + (i32)((target_b - cur_b) * amount);
        u8 new_a = cur_a + (i32)((target_a - cur_a) * amount);
        
        ctx->palette[start + i] = new_r | (new_g << 8) | (new_b << 16) | (new_a << 24);
    }
}

void pxl8_gfx_interpolate_palettes(pxl8_gfx_ctx* ctx, u32* palette1, u32* palette2, u8 start, u8 count, f32 t) {
    if (!ctx || !ctx->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) < ctx->palette_size; i++) {
        u32 col1 = palette1[i];
        u32 col2 = palette2[i];
        
        u8 r1 = col1 & 0xFF;
        u8 g1 = (col1 >> 8) & 0xFF;
        u8 b1 = (col1 >> 16) & 0xFF;
        u8 a1 = (col1 >> 24) & 0xFF;
        
        u8 r2 = col2 & 0xFF;
        u8 g2 = (col2 >> 8) & 0xFF;
        u8 b2 = (col2 >> 16) & 0xFF;
        u8 a2 = (col2 >> 24) & 0xFF;
        
        u8 r = r1 + (i32)((r2 - r1) * t);
        u8 g = g1 + (i32)((g2 - g1) * t);
        u8 b = b1 + (i32)((b2 - b1) * t);
        u8 a = a1 + (i32)((a2 - a1) * t);
        
        ctx->palette[start + i] = r | (g << 8) | (b << 16) | (a << 24);
    }
}

void pxl8_gfx_color_ramp(pxl8_gfx_ctx* ctx, u8 start, u8 count, u32 from_color, u32 to_color) {
    if (!ctx || !ctx->palette || count <= 1) return;
    
    u8 from_r = from_color & 0xFF;
    u8 from_g = (from_color >> 8) & 0xFF;
    u8 from_b = (from_color >> 16) & 0xFF;
    u8 from_a = (from_color >> 24) & 0xFF;
    
    u8 to_r = to_color & 0xFF;
    u8 to_g = (to_color >> 8) & 0xFF;
    u8 to_b = (to_color >> 16) & 0xFF;
    u8 to_a = (to_color >> 24) & 0xFF;
    
    for (u8 i = 0; i < count && (start + i) < ctx->palette_size; i++) {
        f32 t = (f32)i / (f32)(count - 1);
        
        u8 r = from_r + (i32)((to_r - from_r) * t);
        u8 g = from_g + (i32)((to_g - from_g) * t);
        u8 b = from_b + (i32)((to_b - from_b) * t);
        u8 a = from_a + (i32)((to_a - from_a) * t);
        
        ctx->palette[start + i] = r | (g << 8) | (b << 16) | (a << 24);
    }
}

void pxl8_gfx_process_effects(pxl8_gfx_ctx* ctx, pxl8_effects* effects, f32 dt) {
    if (!ctx || !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(ctx, cycle->start_index, count, 1);
        }
    }
}