;============================================================
;   iron_gltf.hsp — glTF 2.0 ローダ (最小実装、pure HSP)
;
;   Khronos glTF 2.0 の .gltf (+ .bin sidecar) を読み込み、
;   1 つ目の mesh.primitives[0] から位置 / 法線 / UV / indices を
;   配列展開する最小実装。
;
;   対応:
;     - .gltf (JSON) + 同ディレクトリの .bin (URI は相対パス)
;     - .glb (バイナリコンテナ、JSON + BIN chunk)
;     - POSITION / NORMAL / TEXCOORD_0 / indices
;     - indices componentType 5121 (u8) / 5123 (u16) / 5125 (u32)
;     - 頂点データ componentType 5126 (float) のみ
;
;   対応外:
;     - data URI (base64 埋め込み、.gltf 内に JSON と一緒に入ってるケース)
;     - スキン / アニメーション / マテリアル詳細
;     - 複数 mesh / primitives (最初の 1 組のみ)
;     - 行列変換 / スケール
;
;   API:
;     gltf_load "model.gltf", var_verts, var_norms, var_uvs, var_tris,
;               var n_verts, var n_tris, var has_norm, var has_uv
;         verts: double[3 * n_verts]
;         norms: double[3 * n_verts]
;         uvs:   double[2 * n_verts]
;         tris:  int[3 * n_tris]
;         戻り値: 0=OK / <0=エラー
;
;   依存: iron_json.hsp (hspjson.dll)
;============================================================
#ifndef __iron_gltf_hsp__
#define __iron_gltf_hsp__

#include "iron_json.hsp"

#ifndef GLTF_MAX_VERTS
#define GLTF_MAX_VERTS 524288
#endif
#ifndef GLTF_MAX_TRIS
#define GLTF_MAX_TRIS 524288
#endif

#module iron_gltf

;---------------------------------------------------------
; internal: buf から offset で 4-byte little-endian float を取り出す
;---------------------------------------------------------
#defcfunc _gltf_float var buf, int ofs, \
	local _bits, local _sign, local _exp, local _mant, \
	local _mantd, local _e, local _v
	_bits = lpeek(buf, ofs)
	_sign = (_bits >> 31) & 1
	_exp  = (_bits >> 23) & 0xFF
	_mant = _bits & 0x7FFFFF
	if _exp = 0 {
		if _mant = 0 : return 0.0
		return 0.0
	}
	if _exp = 0xFF : return 0.0
	_mantd = 1.0 + (1.0 * _mant) / 8388608.0
	_e = _exp - 127
	_v = _mantd
	if _e >= 0 {
		repeat _e : _v *= 2.0 : loop
	} else {
		repeat 0 - _e : _v *= 0.5 : loop
	}
	if _sign = 1 : _v = 0.0 - _v
	return _v

#defcfunc _gltf_u16 var buf, int ofs
	return wpeek(buf, ofs)
#defcfunc _gltf_u32 var buf, int ofs
	return lpeek(buf, ofs)
#defcfunc _gltf_u8 var buf, int ofs
	return peek(buf, ofs)

;---------------------------------------------------------
; internal: data URI ("data:*/*;base64,XXX") のデコード
;   v_out にバイナリ書き込み、v_len にバイト数
;---------------------------------------------------------
#deffunc _gltf_decode_data_uri str uri, var v_out, var v_len, \
	local _comma, local _b64, local _c, local _val, \
	local _bits, local _bufbits, local _write_pos, local _len
	_comma = instr(uri, 0, ",")
	if _comma < 0 : v_len = 0 : return
	_b64 = strmid(uri, _comma + 1, strlen(uri) - _comma - 1)
	; 上限: base64 は 4 文字で 3 バイト → 出力バッファを確保
	sdim v_out, strlen(_b64) * 3 / 4 + 16
	_bits = 0 : _bufbits = 0 : _write_pos = 0
	repeat strlen(_b64)
		_c = peek(_b64, cnt)
		if _c = '=' : break
		_val = -1
		if (_c >= 'A') & (_c <= 'Z') : _val = _c - 'A'
		if (_c >= 'a') & (_c <= 'z') : _val = 26 + _c - 'a'
		if (_c >= '0') & (_c <= '9') : _val = 52 + _c - '0'
		if _c = '+' : _val = 62
		if _c = '/' : _val = 63
		if _c = '-' : _val = 62           ; URL-safe variant
		if _c = '_' : _val = 63
		if _val < 0 : continue            ; 空白等を読み飛ばし
		_bits = (_bits << 6) | _val
		_bufbits = _bufbits + 6
		if _bufbits >= 8 {
			_bufbits = _bufbits - 8
			poke v_out, _write_pos, (_bits >> _bufbits) & 0xFF
			_write_pos++
		}
	loop
	v_len = _write_pos
	return

;---------------------------------------------------------
; internal: ディレクトリを取り出して URI と結合
;---------------------------------------------------------
#deffunc _gltf_join_path str gltf_path, str uri, var v_out, \
	local _slash1, local _slash2, local _dir
	sdim v_out, 1024
	; 絶対パスらしきものは素通し
	if strlen(uri) >= 2 {
		if peek(uri, 1) = ':' : v_out = uri : return
	}
	if peek(uri, 0) = '/' : v_out = uri : return
	; gltf_path の最終 / or \ までを取り出す
	_slash1 = -1 : _slash2 = -1
	repeat strlen(gltf_path)
		if peek(gltf_path, cnt) = '/'  : _slash1 = cnt
		if peek(gltf_path, cnt) = '\\' : _slash2 = cnt
	loop
	if _slash2 > _slash1 : _slash1 = _slash2
	if _slash1 < 0 : v_out = uri : return
	_dir = strmid(gltf_path, 0, _slash1 + 1)
	v_out = _dir + uri
	return

;---------------------------------------------------------
; gltf_load メイン
;---------------------------------------------------------
#deffunc gltf_load str path, var v_verts, var v_norms, var v_uvs, var v_tris, \
	var v_nv, var v_nt, var v_has_norm, var v_has_uv, \
	local _size, local _file_buf, local _json_text, local _hid, \
	local _bin_uri, local _bin_path, local _bin_buf, local _bin_size, \
	local _is_glb, local _magic, local _total_len, \
	local _cp, local _chunk_len, local _chunk_type, \
	local _pos_acc, local _norm_acc, local _uv_acc, local _idx_acc, \
	local _pos_bv, local _norm_bv, local _uv_bv, local _idx_bv, \
	local _pos_off, local _pos_len, local _norm_off, local _norm_len, \
	local _uv_off, local _uv_len, local _idx_off, local _idx_len, \
	local _pos_count, local _idx_count, local _idx_ct, \
	local _bv_offset, local _bv_length, local _i
	exist path
	_size = strsize
	if _size <= 0 : return -1
	sdim _file_buf, _size + 16
	bload path, _file_buf, _size

	; --- GLB 判別 ---
	; Header (12 bytes):
	;   0-3  magic "glTF" (0x46546C67, little-endian u32 = 0x46546C67)
	;   4-7  version (u32, 通常 2)
	;   8-11 total length (u32)
	; その後 chunk 列 (each: 4 length u32 + 4 type u32 + data)
	;   chunk 0: type "JSON" (0x4E4F534A)
	;   chunk 1: type "BIN\0" (0x004E4942) — オプション
	_is_glb = 0
	if _size >= 12 {
		_magic = lpeek(_file_buf, 0)
		if _magic = 0x46546C67 : _is_glb = 1
	}

	if _is_glb {
		_total_len = lpeek(_file_buf, 8)
		if _total_len > _size : _total_len = _size
		_cp = 12
		sdim _json_text, 1
		sdim _bin_buf, 1
		_bin_size = 0
		repeat
			if _cp + 8 > _total_len : break
			_chunk_len  = lpeek(_file_buf, _cp)
			_chunk_type = lpeek(_file_buf, _cp + 4)
			_cp = _cp + 8
			if _chunk_type = 0x4E4F534A {       ; "JSON"
				sdim _json_text, _chunk_len + 16
				memcpy _json_text, _file_buf, _chunk_len, 0, _cp
				poke _json_text, _chunk_len, 0
			}
			if _chunk_type = 0x004E4942 {       ; "BIN\0"
				sdim _bin_buf, _chunk_len + 16
				memcpy _bin_buf, _file_buf, _chunk_len, 0, _cp
				_bin_size = _chunk_len
			}
			_cp = _cp + _chunk_len
			if _cp >= _total_len : break
		loop
		if strlen(_json_text) = 0 : return -6
		if _bin_size = 0 : return -7
		_hid = json_load(_json_text)
		if _hid < 0 : return -2
	} else {
		; --- 従来の .gltf + 外部 .bin / data URI ---
		_hid = json_load(_file_buf)
		if _hid < 0 : return -2
		_bin_uri = json_str(_hid, "buffers[0].uri")
		if strlen(_bin_uri) = 0 { json_release _hid : return -3 }
		; data URI ("data:application/octet-stream;base64,...") を検出
		if instr(_bin_uri, 0, "data:") = 0 {
			_bin_size = 0
			sdim _bin_buf, 1
			_gltf_decode_data_uri _bin_uri, _bin_buf, _bin_size
			if _bin_size <= 0 { json_release _hid : return -4 }
		} else {
			_gltf_join_path path, _bin_uri, _bin_path
			exist _bin_path
			_bin_size = strsize
			if _bin_size <= 0 { json_release _hid : return -4 }
			sdim _bin_buf, _bin_size + 16
			bload _bin_path, _bin_buf, _bin_size
		}
	}
	; meshes[0].primitives[0].attributes.POSITION / NORMAL / TEXCOORD_0 / indices
	_pos_acc = json_int(_hid, "meshes[0].primitives[0].attributes.POSITION")
	_norm_acc = json_int(_hid, "meshes[0].primitives[0].attributes.NORMAL")
	_uv_acc = json_int(_hid, "meshes[0].primitives[0].attributes.TEXCOORD_0")
	_idx_acc = json_int(_hid, "meshes[0].primitives[0].indices")
	v_has_norm = 0 : v_has_uv = 0
	if _pos_acc < 0 { json_release _hid : return -5 }
	_pos_bv   = json_int(_hid, "accessors[" + _pos_acc + "].bufferView")
	_pos_count= json_int(_hid, "accessors[" + _pos_acc + "].count")
	_pos_off  = json_int(_hid, "bufferViews[" + _pos_bv + "].byteOffset")
	if _pos_off < 0 : _pos_off = 0
	ddim v_verts, GLTF_MAX_VERTS * 3
	ddim v_norms, GLTF_MAX_VERTS * 3
	ddim v_uvs,   GLTF_MAX_VERTS * 2
	dim  v_tris,  GLTF_MAX_TRIS * 3
	; --- 頂点 ---
	repeat _pos_count
		if cnt >= GLTF_MAX_VERTS : break
		v_verts(cnt * 3 + 0) = _gltf_float(_bin_buf, _pos_off + cnt * 12 + 0)
		v_verts(cnt * 3 + 1) = _gltf_float(_bin_buf, _pos_off + cnt * 12 + 4)
		v_verts(cnt * 3 + 2) = _gltf_float(_bin_buf, _pos_off + cnt * 12 + 8)
	loop
	v_nv = _pos_count
	; --- 法線 ---
	if _norm_acc >= 0 {
		_norm_bv  = json_int(_hid, "accessors[" + _norm_acc + "].bufferView")
		_norm_off = json_int(_hid, "bufferViews[" + _norm_bv + "].byteOffset")
		if _norm_off < 0 : _norm_off = 0
		repeat _pos_count
			if cnt >= GLTF_MAX_VERTS : break
			v_norms(cnt * 3 + 0) = _gltf_float(_bin_buf, _norm_off + cnt * 12 + 0)
			v_norms(cnt * 3 + 1) = _gltf_float(_bin_buf, _norm_off + cnt * 12 + 4)
			v_norms(cnt * 3 + 2) = _gltf_float(_bin_buf, _norm_off + cnt * 12 + 8)
		loop
		v_has_norm = 1
	}
	; --- UV ---
	if _uv_acc >= 0 {
		_uv_bv  = json_int(_hid, "accessors[" + _uv_acc + "].bufferView")
		_uv_off = json_int(_hid, "bufferViews[" + _uv_bv + "].byteOffset")
		if _uv_off < 0 : _uv_off = 0
		repeat _pos_count
			if cnt >= GLTF_MAX_VERTS : break
			v_uvs(cnt * 2 + 0) = _gltf_float(_bin_buf, _uv_off + cnt * 8 + 0)
			v_uvs(cnt * 2 + 1) = _gltf_float(_bin_buf, _uv_off + cnt * 8 + 4)
		loop
		v_has_uv = 1
	}
	; --- indices ---
	v_nt = 0
	if _idx_acc >= 0 {
		_idx_bv   = json_int(_hid, "accessors[" + _idx_acc + "].bufferView")
		_idx_count= json_int(_hid, "accessors[" + _idx_acc + "].count")
		_idx_ct   = json_int(_hid, "accessors[" + _idx_acc + "].componentType")
		_idx_off  = json_int(_hid, "bufferViews[" + _idx_bv + "].byteOffset")
		if _idx_off < 0 : _idx_off = 0
		; indices は 3 つずつ 1 triangle
		_i = 0
		repeat _idx_count / 3
			if _i + 3 > _idx_count : break
			if (cnt * 3) >= (GLTF_MAX_TRIS * 3) : break
			if _idx_ct = 5121 {
				v_tris(cnt * 3 + 0) = _gltf_u8(_bin_buf, _idx_off + _i + 0)
				v_tris(cnt * 3 + 1) = _gltf_u8(_bin_buf, _idx_off + _i + 1)
				v_tris(cnt * 3 + 2) = _gltf_u8(_bin_buf, _idx_off + _i + 2)
				_i = _i + 3
			}
			if _idx_ct = 5123 {
				v_tris(cnt * 3 + 0) = _gltf_u16(_bin_buf, _idx_off + _i + 0)
				v_tris(cnt * 3 + 1) = _gltf_u16(_bin_buf, _idx_off + _i + 2)
				v_tris(cnt * 3 + 2) = _gltf_u16(_bin_buf, _idx_off + _i + 4)
				_i = _i + 6
			}
			if _idx_ct = 5125 {
				v_tris(cnt * 3 + 0) = _gltf_u32(_bin_buf, _idx_off + _i + 0)
				v_tris(cnt * 3 + 1) = _gltf_u32(_bin_buf, _idx_off + _i + 4)
				v_tris(cnt * 3 + 2) = _gltf_u32(_bin_buf, _idx_off + _i + 8)
				_i = _i + 12
			}
		loop
		v_nt = _idx_count / 3
	}
	json_release _hid
	return 0

#global
#endif