/*
 * meow_png_stream.c — M23 真缩解后端（二）：PNG 逐行 inflate + BMP 按行跳采样。
 *
 * 为什么不用 stb 的 PNG 路径：stb 先整幅 inflate + 反滤波成 raw，再展开成 RGBA 整幅，
 * 最后才降采样——1600×2400 PNG 实测 +26.2MB。这里改成「按行流式」：
 *   - IDAT 边到边 inflate，一次只持有一行（filter 字节 + rowbytes）与上一行；
 *   - 反滤波后**只把需要的列**写进目标行（最近邻，与 meow_img_downscale 同口径）；
 *   - 不需要的行整行丢弃（但仍要反滤波，filter 依赖上一行）；
 *   - 峰值 ≈ 两行缓冲 + zlib 窗口(32KB) + 目标 RGBA，与源图分辨率无关。
 * 支持：gray / gray+A / RGB / RGBA / 调色板，bit depth 1/2/4/8/16（16bit 取高字节降 8bit），
 *       tRNS 透明。不支持 → 返回 0 回退 stb：Adam7 交错、非 0 压缩/滤波方法、异常色型组合。
 *
 * BMP：文件流按目标行直接 seek 到源行（不读全图），行内跳列采样。支持 8bpp 调色板 /
 *      16bpp RGB555 / 24bpp / 32bpp（BI_RGB）；RLE、BITFIELDS、1/2/4bpp → 回退 stb。
 *
 * 许可证：MeowOS 自写（MIT，随 SDK 发布）。PNG/zlib 语义参考 RFC 2083 / zlib 手册；
 * BMP 布局参考 stb_image 的解析口径（同为本仓已有依赖，行为对齐）。
 */
#include "meow_jpeg.h"

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

/* ---------------- zlib ABI ---------------- */
#if defined(__has_include)
#  if __has_include(<zlib.h>)
#    include <zlib.h>
#    define MZP_HAVE_H 1
#  endif
#endif
#ifndef MZP_HAVE_H
/* 设备 musl sysroot 没有 zlib.h（meow_font.c 同此口径）：按 libz.a 的 ABI 声明最小 API。
 * 指针/unsigned long 交给编译器定尺寸，宿主 64 位与设备 32 位都成立。 */
typedef unsigned char  mzp_Bytef;
typedef unsigned int   mzp_uInt;
typedef unsigned long  mzp_uLong;
typedef void          *mzp_voidpf;
typedef struct mzp_stream_s {
    mzp_Bytef *next_in;  mzp_uInt avail_in;  mzp_uLong total_in;
    mzp_Bytef *next_out; mzp_uInt avail_out; mzp_uLong total_out;
    char *msg;
    void *state;
    void *zalloc;
    void *zfree;
    mzp_voidpf opaque;
    int data_type;
    mzp_uLong adler;
    mzp_uLong reserved;
} mzp_stream;
extern int inflateInit_(mzp_stream *, const char *, int);
extern int inflate(mzp_stream *, int);
extern int inflateEnd(mzp_stream *);
extern const char *zlibVersion(void);
#  define z_stream mzp_stream
#  define Z_OK 0
#  define Z_STREAM_END 1
#  define Z_NO_FLUSH 0
#endif

/* ---------------- 字节源 ---------------- */
typedef struct {
    FILE *f;
    const unsigned char *mem;
    size_t memlen, mempos;
} png_src;

static int ps_read(png_src *s, void *dst, size_t n)
{
    if (s->mem) {
        if (s->mempos + n > s->memlen) return 0;
        memcpy(dst, s->mem + s->mempos, n);
        s->mempos += n;
        return 1;
    }
    return fread(dst, 1, n, s->f) == n;
}

static int ps_skip(png_src *s, size_t n)
{
    if (s->mem) {
        if (s->mempos + n > s->memlen) return 0;
        s->mempos += n;
        return 1;
    }
    return fseek(s->f, (long)n, SEEK_CUR) == 0;
}

/* ================================================================== */
/* PNG                                                                 */
/* ================================================================== */
typedef struct {
    png_src src;
    z_stream z;
    int z_ready, z_end;
    unsigned w, h;
    int depth, color;
    int bits;                 /* 每像素位数 */
    size_t rowbytes;
    unsigned char *line;      /* 1 filter 字节 + rowbytes */
    unsigned char *prev;      /* 上一行的像素字节 */
    unsigned char *cur;       /* 本行的像素字节 */
    size_t need, got;
    unsigned char pal[256][3];
    unsigned char trns_a[256]; /* 调色板透明度 */
    int pal_n;
    int have_trns;
    unsigned trns[3];          /* gray / RGB 的 tRNS 值（已降到 8bit） */
    unsigned char *out;
    int tw, th;
    int next_ty;
    unsigned y;
} png_ctx;

/* 从行里取 px 列的 8bit 通道；返回通道数（写入 o[0..3] 的 RGB(A) 语义由调用方定）。 */
static void png_get_px(const png_ctx *c, const unsigned char *row, int x, unsigned char *o)
{
    int d = c->depth;
    if (d < 8) {
        int per = 8 / d;                     /* 每字节像素数 */
        int byte = x / per;
        int shift = 8 - d * (x % per + 1);
        int v = (row[byte] >> shift) & ((1 << d) - 1);
        if (c->color == 3) {                       /* 调色板：子字节值即索引 */
            o[0] = c->pal[v][0]; o[1] = c->pal[v][1]; o[2] = c->pal[v][2];
            o[3] = c->trns_a[v];
        } else {
            unsigned char g = (unsigned char)(v * 255 / ((1 << d) - 1));
            o[0] = o[1] = o[2] = g;
            /* tRNS 存的是原始采样值（0..2^d-1），不是展开后的 0..255 */
            o[3] = (c->color == 0 && c->have_trns && (unsigned)v == c->trns[0]) ? 0 : 255;
        }
        return;
    }
    if (d == 8) {
        if (c->color == 0) {                       /* gray */
            o[0] = o[1] = o[2] = row[x];
            o[3] = (c->have_trns && row[x] == c->trns[0]) ? 0 : 255;
        } else if (c->color == 2) {                /* RGB */
            o[0] = row[x * 3]; o[1] = row[x * 3 + 1]; o[2] = row[x * 3 + 2];
            o[3] = (c->have_trns && row[x * 3] == c->trns[0] &&
                    row[x * 3 + 1] == c->trns[1] && row[x * 3 + 2] == c->trns[2]) ? 0 : 255;
        } else if (c->color == 3) {                /* palette */
            int i = row[x];
            o[0] = c->pal[i][0]; o[1] = c->pal[i][1]; o[2] = c->pal[i][2];
            o[3] = c->trns_a[i];
        } else if (c->color == 4) {                /* gray + A */
            o[0] = o[1] = o[2] = row[x * 2];
            o[3] = row[x * 2 + 1];
        } else {                                   /* RGBA */
            o[0] = row[x * 4]; o[1] = row[x * 4 + 1];
            o[2] = row[x * 4 + 2]; o[3] = row[x * 4 + 3];
        }
        return;
    }
    /* 16bit：取高字节 */
    if (c->color == 0) {
        o[0] = o[1] = o[2] = row[x * 2];
        o[3] = (c->have_trns && row[x * 2] == c->trns[0]) ? 0 : 255;
    } else if (c->color == 2) {
        o[0] = row[x * 6]; o[1] = row[x * 6 + 2]; o[2] = row[x * 6 + 4];
        o[3] = (c->have_trns && row[x * 6] == c->trns[0] &&
                row[x * 6 + 2] == c->trns[1] && row[x * 6 + 4] == c->trns[2]) ? 0 : 255;
    } else if (c->color == 4) {
        o[0] = o[1] = o[2] = row[x * 4];
        o[3] = row[x * 4 + 2];
    } else {
        o[0] = row[x * 8]; o[1] = row[x * 8 + 2];
        o[2] = row[x * 8 + 4]; o[3] = row[x * 8 + 6];
    }
}

/* 一行反滤波完成后：按需写目标行。 */
static int png_emit_row(png_ctx *c)
{
    unsigned char ft = c->line[0];
    const unsigned char *src = c->line + 1;
    unsigned char *dst = c->cur;
    size_t x, bpp = (size_t)((c->bits + 7) / 8);
    if (bpp < 1) bpp = 1;
    if (ft > 4) return 0;
    for (x = 0; x < c->rowbytes; x++) {
        unsigned char left = (x >= bpp) ? dst[x - bpp] : 0;
        unsigned char up = c->prev ? c->prev[x] : 0;
        unsigned char ul = (c->prev && x >= bpp) ? c->prev[x - bpp] : 0;
        unsigned char v = src[x];
        switch (ft) {
        case 0: dst[x] = v; break;
        case 1: dst[x] = (unsigned char)(v + left); break;
        case 2: dst[x] = (unsigned char)(v + up); break;
        case 3: dst[x] = (unsigned char)(v + (((unsigned)left + (unsigned)up) >> 1)); break;
        default: {
            int p = (int)left + (int)up - (int)ul;
            int pa = p - (int)left; if (pa < 0) pa = -pa;
            int pb = p - (int)up;   if (pb < 0) pb = -pb;
            int pc = p - (int)ul;   if (pc < 0) pc = -pc;
            unsigned char pred = (pa <= pb && pa <= pc) ? left : (pb <= pc ? up : ul);
            dst[x] = (unsigned char)(v + pred);
            break;
        }
        }
    }
    if (c->next_ty < c->th) {
        long sy = (long)c->next_ty * (long)c->h / c->th;
        if (sy == (long)c->y) {
            unsigned char *o = c->out + (size_t)c->next_ty * c->tw * 4;
            int tx;
            for (tx = 0; tx < c->tw; tx++) {
                int sx = (int)((long)tx * (long)c->w / c->tw);
                unsigned char px[4];
                png_get_px(c, dst, sx, px);
                o[tx * 4] = px[0]; o[tx * 4 + 1] = px[1];
                o[tx * 4 + 2] = px[2]; o[tx * 4 + 3] = px[3];
            }
            c->next_ty++;
        }
    }
    /* 交换 prev/cur（prev 直接成为反滤波基准） */
    {
        unsigned char *t = c->prev;
        c->prev = c->cur;
        c->cur = t;
    }
    c->y++;
    return 1;
}

#define PNG_NEED_MORE 2

static int png_pump(png_ctx *c)
{
    while (c->got < c->need) {
        int ret;
        c->z.next_out = c->line + c->got;
        c->z.avail_out = (unsigned)(c->need - c->got);
        ret = inflate(&c->z, Z_NO_FLUSH);
        c->got = c->need - c->z.avail_out;
        if (ret == Z_STREAM_END) c->z_end = 1;
        else if (ret != Z_OK) return 0;
        if (c->got < c->need && (c->z.avail_in == 0 || c->z_end))
            return c->z_end ? 0 : PNG_NEED_MORE;
    }
    if (!png_emit_row(c)) return 0;
    c->got = 0;
    return 1;
}

static int png_bits_of(int color, int depth)
{
    switch (color) {
    case 0: return depth;                 /* gray */
    case 2: return depth * 3;             /* rgb */
    case 3: return depth;                 /* palette */
    case 4: return depth * 2;             /* gray + a */
    case 6: return depth * 4;             /* rgba */
    default: return 0;
    }
}

static int png_depth_ok(int color, int depth)
{
    if (color == 0) return depth == 1 || depth == 2 || depth == 4 || depth == 8 || depth == 16;
    if (color == 3) return depth == 1 || depth == 2 || depth == 4 || depth == 8;
    if (color == 2 || color == 4 || color == 6) return depth == 8 || depth == 16;
    return 0;
}

int meow_png_decode_stream(FILE *f, const unsigned char *mem, size_t len,
                           int maxdim, int *tw, int *th, unsigned char **rgba)
{
    png_ctx c;
    unsigned char sig[8], hdr[8];
    unsigned char ibuf[4096];
    int have_ihdr = 0, i, ok = 0;

    if (maxdim <= 0) return 0;
    memset(&c, 0, sizeof c);
    c.src.f = f;
    c.src.mem = mem;
    c.src.memlen = len;
    if (!ps_read(&c.src, sig, 8)) return 0;
    if (memcmp(sig, "\x89PNG\r\n\x1a\n", 8) != 0) return 0;

    while (ps_read(&c.src, hdr, 8)) {
        unsigned clen = ((unsigned)hdr[0] << 24) | ((unsigned)hdr[1] << 16) |
                        ((unsigned)hdr[2] << 8) | hdr[3];
        const unsigned char *type = hdr + 4;
        if (clen > 0x7FFFFFFFu) return 0;

        if (memcmp(type, "IHDR", 4) == 0) {
            unsigned char d[13];
            if (clen < 13 || have_ihdr) return 0;
            if (!ps_read(&c.src, d, 13)) return 0;
            c.w = ((unsigned)d[0] << 24) | ((unsigned)d[1] << 16) | ((unsigned)d[2] << 8) | d[3];
            c.h = ((unsigned)d[4] << 24) | ((unsigned)d[5] << 16) | ((unsigned)d[6] << 8) | d[7];
            c.depth = d[8];
            c.color = d[9];
            if (d[10] != 0 || d[11] != 0 || d[12] != 0) return 0;   /* 只支持非交错 + 标准压缩/滤波 */
            if (c.w == 0 || c.h == 0 || c.w > 65535u || c.h > 65535u) return 0;
            if (!png_depth_ok(c.color, c.depth)) return 0;
            c.bits = png_bits_of(c.color, c.depth);
            c.rowbytes = ((size_t)c.w * c.bits + 7) / 8;
            if (c.rowbytes == 0) return 0;
            /* 小图不动：与 meow_img_load_file_dim 的「旧入口」判定一致 */
            if ((long)(c.w > c.h ? c.w : c.h) <= maxdim) return 0;
            if (clen > 13 && !ps_skip(&c.src, clen - 13)) return 0;
            have_ihdr = 1;
            c.need = c.rowbytes + 1;
            c.line = (unsigned char *)malloc(c.need);
            c.prev = (unsigned char *)malloc(c.rowbytes);
            c.cur = (unsigned char *)malloc(c.rowbytes);
            if (!c.line || !c.prev || !c.cur) goto done;
            memset(c.prev, 0, c.rowbytes);
            memset(c.cur, 0, c.rowbytes);
            for (i = 0; i < 256; i++) c.trns_a[i] = 255;
            meow_img_target_dims((int)c.w, (int)c.h, maxdim, &c.tw, &c.th);
            c.out = (unsigned char *)malloc((size_t)c.tw * c.th * 4);
            if (!c.out) goto done;
            memset(c.out, 0, (size_t)c.tw * c.th * 4);
            memset(&c.z, 0, sizeof c.z);
            if (inflateInit_(&c.z, zlibVersion(), (int)sizeof c.z) != Z_OK) goto done;
            c.z_ready = 1;
        } else if (memcmp(type, "PLTE", 4) == 0) {
            unsigned char d[256 * 3];
            if (clen > sizeof d || clen % 3) return 0;
            if (!ps_read(&c.src, d, clen)) return 0;
            c.pal_n = (int)(clen / 3);
            memcpy(c.pal, d, clen);
        } else if (memcmp(type, "tRNS", 4) == 0) {
            unsigned char d[256];
            if (clen > sizeof d) return 0;
            if (!ps_read(&c.src, d, clen)) return 0;
            c.have_trns = 1;
            if (c.color == 3) {
                for (i = 0; i < (int)clen && i < 256; i++) c.trns_a[i] = d[i];
            } else if (c.color == 0 && clen >= 2) {
                /* 8bit/低位深：采样值存在低字节；16bit：与像素一样取高字节比较 */
                c.trns[0] = (c.depth == 16) ? d[0] : d[1];
            } else if (c.color == 2 && clen >= 6) {
                if (c.depth == 16) {
                    c.trns[0] = d[0]; c.trns[1] = d[2]; c.trns[2] = d[4];
                } else {
                    c.trns[0] = d[1]; c.trns[1] = d[3]; c.trns[2] = d[5];
                }
            }
        } else if (memcmp(type, "IDAT", 4) == 0) {
            unsigned remain = clen;
            if (!have_ihdr || !c.out) return 0;
            while (remain) {
                unsigned n = remain > sizeof ibuf ? (unsigned)sizeof ibuf : remain;
                if (!ps_read(&c.src, ibuf, n)) goto done;
                remain -= n;
                c.z.next_in = ibuf;
                c.z.avail_in = n;
                while (c.z.avail_in && c.next_ty < c.th) {
                    int r = png_pump(&c);
                    if (r == 0) goto done;
                    if (r == PNG_NEED_MORE) break;
                }
            }
        } else if (memcmp(type, "IEND", 4) == 0) {
            break;
        } else {
            if (clen && !ps_skip(&c.src, clen)) return 0;
        }
        /* 跳过 CRC */
        {
            unsigned char crc[4];
            if (!ps_read(&c.src, crc, 4)) break;
        }
        if (c.out && c.next_ty >= c.th) break;      /* 目标行都拿到了，不必读完 */
    }

    if (!have_ihdr || !c.out || c.next_ty < c.th) goto done;
    if (c.color == 3 && c.pal_n <= 0) goto done;
    *tw = c.tw;
    *th = c.th;
    *rgba = c.out;
    c.out = NULL;
    ok = 1;
done:
    if (c.z_ready) inflateEnd(&c.z);
    free(c.line);
    free(c.prev);
    free(c.cur);
    free(c.out);
    return ok;
}

/* ================================================================== */
/* BMP                                                                 */
/* ================================================================== */
static unsigned le16(const unsigned char *p) { return (unsigned)p[0] | ((unsigned)p[1] << 8); }
static unsigned le32(const unsigned char *p)
{
    return (unsigned)p[0] | ((unsigned)p[1] << 8) | ((unsigned)p[2] << 16) | ((unsigned)p[3] << 24);
}

int meow_bmp_decode_stream(FILE *f, const unsigned char *mem, size_t len,
                           int maxdim, int *tw, int *th, unsigned char **rgba)
{
    png_src src;
    unsigned char fh[14], dh[4];
    unsigned char dib[1024];
    unsigned dibsize, dataoff, bpp, compression, colors;
    long w, h, rowsize;
    int topdown, i;
    unsigned char pal[256][3];
    unsigned char *row = NULL, *out = NULL;
    int tw_ = 0, th_ = 0, ok = 0;

    if (maxdim <= 0) return 0;
    memset(&src, 0, sizeof src);
    src.f = f;
    src.mem = mem;
    src.memlen = len;
    if (!ps_read(&src, fh, 14)) return 0;
    if (fh[0] != 'B' || fh[1] != 'M') return 0;
    dataoff = le32(fh + 10);
    if (!ps_read(&src, dh, 4)) return 0;
    dibsize = le32(dh);
    if (dibsize < 40 || dibsize > sizeof dib) return 0;     /* BITMAPCOREHEADER/OS2 → stb */
    memcpy(dib, dh, 4);
    if (!ps_read(&src, dib + 4, dibsize - 4)) return 0;
    w = (long)(int)le32(dib + 4);
    h = (long)(int)le32(dib + 8);
    bpp = le16(dib + 14);
    compression = le32(dib + 16);
    colors = (dibsize >= 36) ? le32(dib + 32) : 0;
    if (compression != 0) return 0;                          /* RLE/BITFIELDS → stb */
    if (w <= 0 || h == 0) return 0;
    if (bpp != 8 && bpp != 16 && bpp != 24 && bpp != 32) return 0;
    topdown = h < 0;
    if (h < 0) h = -h;
    if ((long)(w > h ? w : h) <= maxdim) return 0;           /* 小图走旧入口 */

    if (bpp <= 8) {
        unsigned n = colors ? colors : 256;
        unsigned char e[4];
        if (n > 256) n = 256;
        for (i = 0; (unsigned)i < n; i++) {
            if (!ps_read(&src, e, 4)) return 0;
            pal[i][0] = e[2]; pal[i][1] = e[1]; pal[i][2] = e[0];
        }
        for (; i < 256; i++) pal[i][0] = pal[i][1] = pal[i][2] = 0;
    }

    rowsize = (w * bpp + 31) / 32 * 4;
    meow_img_target_dims((int)w, (int)h, maxdim, &tw_, &th_);
    row = (unsigned char *)malloc((size_t)rowsize);
    out = (unsigned char *)malloc((size_t)tw_ * th_ * 4);
    if (!row || !out) goto done;

    for (int ty = 0; ty < th_; ty++) {
        int sy = (int)((long)ty * h / th_);
        long file_row = topdown ? sy : (h - 1 - sy);
        long off = (long)dataoff + file_row * rowsize;
        unsigned char *o = out + (size_t)ty * tw_ * 4;
        if (src.mem) {
            if (off < 0 || (size_t)off + (size_t)rowsize > src.memlen) goto done;
            memcpy(row, src.mem + off, (size_t)rowsize);
        } else {
            if (fseek(src.f, off, SEEK_SET) != 0) goto done;
            if (fread(row, 1, (size_t)rowsize, src.f) != (size_t)rowsize) goto done;
        }
        for (int tx = 0; tx < tw_; tx++) {
            int sx = (int)((long)tx * w / tw_);
            unsigned char *p = o + (size_t)tx * 4;
            if (bpp == 24) {
                const unsigned char *q = row + sx * 3;
                p[0] = q[2]; p[1] = q[1]; p[2] = q[0]; p[3] = 255;
            } else if (bpp == 32) {
                const unsigned char *q = row + sx * 4;
                p[0] = q[2]; p[1] = q[1]; p[2] = q[0]; p[3] = 255;   /* BI_RGB：alpha 不作为语义 */
            } else if (bpp == 16) {
                unsigned v = le16(row + sx * 2);
                unsigned r = (v >> 10) & 31, g = (v >> 5) & 31, b = v & 31;
                p[0] = (unsigned char)((r << 3) | (r >> 2));
                p[1] = (unsigned char)((g << 3) | (g >> 2));
                p[2] = (unsigned char)((b << 3) | (b >> 2));
                p[3] = 255;
            } else {
                unsigned idx = row[sx];
                p[0] = pal[idx & 255][0]; p[1] = pal[idx & 255][1];
                p[2] = pal[idx & 255][2]; p[3] = 255;
            }
        }
    }
    *tw = tw_;
    *th = th_;
    *rgba = out;
    out = NULL;
    ok = 1;
done:
    free(row);
    free(out);
    return ok;
}