#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""《深海回音》Stable Audio 3 small-sfx 音效（批次 2，17-abyss.md §5.2 SA3 条目）。

生成（GPU，flock 串行；模型 gated 走 hf-mirror，见 env.d/sfx.sh）：
  python3 gen_sfx_sa3.py --emit   # 打印生成命令，原样执行
生成 12 条原始 wav 到 gen/staging/sfx_sa3/abyss_<i>.wav（顺序 = 下表 idx）。
本脚本随后做后处理：修整/去直流/8ms cos² 淡化（amb 床只去直流不淡化——游戏层
buffer.loop 循环，radio phone_ring/amb 先例）/ 响度定档 → ogg q4 + m4a 128k
（highpass=f=30，one-shot 前后垫 60/120ms 静音）→ 解码回读 QC：逐声道真峰值
<= -3dBFS、|直流| <= 0.001；超限降增益 + 首轮重编加 18kHz 低通，最多 3 轮补偿。

echo_metal_l/_r：文档 prompt 带 panned slightly left/right，但引擎无运行时 pan、
SA3 立体声分离度不可靠（dungeon 先例）——取「left」版生成源，mono 下混后 ffmpeg
pan 派生 _l/_r 预 pan 立体声（能量只在目标声道，脚本断言极性；dungeon side_* 工艺）。

prompt 均出自 17-abyss.md §5.2（英文原稿）；定档理由：
  spark_flash/landslide -4（D2/D4 危机拍，前置 2s 渐强同窗）
  echo_metal -5 / echo_groan -6 / surf_splash -6（信息音/叙事音，主层次）
  valve_close -5 / bell_toll -5（道具 payoff 音）
  echo_debris -7 / static_burst -8（垫在通话底下的纹理层）
  amb_* RMS -26 / 峰值帽 -8（环境床，escape/adio amb 先例）

用法：.venv-tts/bin/python web/abyss/gen/gen_sfx_sa3.py [--emit]
"""
import os
import subprocess
import sys

import numpy as np
import soundfile as sf

SR = 44100
HERE = os.path.dirname(os.path.abspath(__file__))
STAGING = os.path.join(HERE, "staging", "sfx_sa3")
DST = os.path.join(HERE, "..", "audio", "sfx")
SA3 = "/home/xsl/SoundGame/tools/stable-audio-3"
FFMPEG = "/home/xsl/SoundGame/tools/bin/ffmpeg"
SEED = 20260909

# (id, gen 秒, prompt, 模式 oneshot/amb, 定档 dBFS, trim 上限 s)
ITEMS = [
    ("echo_metal", 3.0, "Sonar echo returning from a steel submarine hull, resonant metallic ping with slow decay, panned slightly left, underwater, 2 seconds", "oneshot", -5.0, 2.4),
    ("echo_debris", 3.0, "Sonar echo returning from an underwater rubble and debris field, diffuse granular reverberation, 2 seconds", "oneshot", -7.0, 2.6),
    ("echo_groan", 5.0, "Stressed submarine hull groaning deep underwater, long metallic moan under immense pressure, 4 seconds", "oneshot", -6.0, 4.5),
    ("static_burst", 3.0, "Shortwave radio static burst with faint ghost of a distorted voice, 2 seconds", "oneshot", -8.0, 2.4),
    ("amb_deep", 12.0, "Deep ocean ambient loop, low pressure hum, distant current wash, occasional far creak, seamless loop", "amb", -26.0, 0),
    ("amb_deck", 12.0, "Night ship deck ambience, wind, gentle waves against hull, distant rigging creak and engine hum, seamless loop", "amb", -26.0, 0),
    ("breath_loop", 12.0, "Slow scuba regulator breathing loop, calm but slightly labored, underwater, seamless loop", "amb", -26.0, 0),
    ("spark_flash", 3.0, "Underwater cutting torch sparks with a brief muffled gas pop, 2 seconds", "oneshot", -4.0, 2.6),
    ("valve_close", 3.0, "Heavy submarine valve wheel shutting, metal rotation clunk and water seal, 2 seconds", "oneshot", -5.0, 2.4),
    ("landslide", 6.0, "Underwater landslide rumble, rolling sediment and rocks, ominous low roar building, 4 seconds", "oneshot", -4.0, 5.5),
    ("surf_splash", 3.0, "Diver breaking through the sea surface, splash and one gasp, 2 seconds", "oneshot", -6.0, 2.6),
    ("bell_toll", 4.0, "Old ship brass bell, single clean toll with warm resonance, 3 seconds", "oneshot", -5.0, 3.5),
]

# 事件窗（SA3 raw 为定长 latent 床，与请求时长无关；None=按 -50dBFS 活动窗自然修剪）。
# 首轮先自然修剪，实测事件位置偏离设计时再回填窗（radio arg_muffled 同法）。
WIN = {}


def emit():
    items = [it for it in ITEMS if it[3] != "skip"]
    durs = " ".join(str(d) for _, d, *_ in items)
    ps = " ".join(f'"{p}"' for _, _, p, *_ in items)
    print(f"cd {SA3} && flock /home/xsl/SoundGame/.gpu.lock .venv/bin/stable-audio "
          f"--model small-sfx --steps 8 --seed {SEED} "
          f"-p {ps} --duration {durs} -o {STAGING}/abyss.wav")


def trim_oneshot(y, floor_db=-50.0, head_ms=50, tail_ms=140, max_s=None):
    env = np.max(np.abs(y), axis=1)
    act = np.where(env > 10 ** (floor_db / 20))[0]
    a = max(0, act[0] - int(SR * head_ms / 1000))
    b = min(len(y), act[-1] + int(SR * tail_ms / 1000))
    if max_s:
        b = min(b, a + int(SR * max_s))
    return y[a:b]


def prep_oneshot(y, target_db, win=None, max_s=None):
    if win:
        a, b = int(SR * win[0]), int(SR * win[1])
        y = y[a:b]
        nf2 = max(1, int(SR * 0.050))                    # 窗切在持续声中，切点 50ms 缓出
        y[-nf2:] *= (0.5 * (1 + np.cos(np.pi * np.arange(nf2) / nf2)))[:, None]
    else:
        y = trim_oneshot(y, max_s=max_s)
    y = y - np.mean(y, axis=0, keepdims=True)          # 去直流
    nf = max(1, int(SR * 0.008))
    fade = 0.5 * (1 - np.cos(np.pi * np.arange(nf) / nf))
    y[:nf] *= fade[:, None]                             # 8ms cos² 淡入淡出
    y[-nf:] *= fade[::-1][:, None]
    y = y / (np.max(np.abs(y)) + 1e-9) * 10 ** (target_db / 20)
    pad_h = np.zeros((int(SR * 0.06), y.shape[1]))      # 编码器软着陆
    pad_t = np.zeros((int(SR * 0.12), y.shape[1]))
    return np.concatenate([pad_h, y, pad_t])


def prep_amb(y, rms_db, peak_db):
    """环境床：只去直流，不淡化（游戏层 buffer.loop 循环播放）；RMS 定档 + 峰值帽。"""
    y = y - np.mean(y, axis=0, keepdims=True)
    rms = 20 * np.log10(np.sqrt(np.mean(y ** 2)) + 1e-9)
    peak = 20 * np.log10(np.max(np.abs(y)) + 1e-9)
    s = min(10 ** ((rms_db - rms) / 20), 10 ** ((peak_db - peak) / 20))
    return y * s


def encode(name, sig):
    base = sig.astype(np.float64)
    vol, lpf, dc_adj = 0.0, False, 0.0
    tmp = os.path.join(STAGING, "_enc.wav")
    chk = os.path.join(STAGING, "_qc.wav")
    res = {}
    for attempt in range(4):  # 首发 + 最多 3 轮补偿（首轮重编加 18kHz 低通）
        x = base - dc_adj
        if vol < 0:
            x = x * (10 ** (vol / 20))
        sf.write(tmp, x.astype(np.float32), SR)
        af = "highpass=f=30,lowpass=f=18000" if lpf else "highpass=f=30"
        for ext, cmd in [("ogg", ["-c:a", "libvorbis", "-q:a", "4"]),
                         ("m4a", ["-c:a", "aac", "-b:a", "128k"])]:
            out = os.path.join(DST, f"{name}.{ext}")
            subprocess.run([FFMPEG, "-y", "-v", "error", "-i", tmp, "-af", af,
                            "-ar", "44100", *cmd, out], check=True)
            subprocess.run([FFMPEG, "-y", "-v", "error", "-i", out, "-c:a", "pcm_f32le", chk], check=True)
            y, sr = sf.read(chk, always_2d=True)
            tp = 20 * np.log10(np.max(np.abs(y)) + 1e-9)
            dcv = float(np.mean(y.mean(axis=1)))
            rms = 20 * np.log10(np.sqrt(np.mean(y ** 2)) + 1e-9)
            ok_tp, ok_dc = tp <= -3.0, abs(dcv) <= 0.001
            res[ext] = (tp, dcv, rms, y.shape[0] / sr)
            if not (ok_tp and ok_dc):
                if not ok_tp:
                    vol -= (tp + 3.0) + 0.25
                    lpf = True       # 首轮重编起加 18kHz 低通压 MDCT 过冲
                if not ok_dc:
                    dc_adj -= dcv
                break
        else:
            break
    else:
        raise AssertionError(f"{name}: 补偿 3 轮后仍超限 {res}")
    return res


def pan_derive(name, sig_mono):
    """mono 源 -> 预 pan 立体声 _l/_r（能量只在 L 或 R 声道），ffmpeg pan 派生
    （dungeon side_* 工艺复用：引擎 play() 无实时 pan）。"""
    src = os.path.join(STAGING, name + "_mono.wav")
    sf.write(src, sig_mono.astype(np.float32), SR)
    for side, out_ch in [("l", "c0"), ("r", "c1")]:
        out = os.path.join(STAGING, f"{name}_{side}.wav")
        subprocess.run([FFMPEG, "-y", "-v", "error", "-i", src,
                        "-af", f"pan=stereo|{out_ch}=c0", "-c:a", "pcm_f32le", out], check=True)
    # 极性断言（staging 源级）：能量只在声像侧声道
    for side, ch in [("l", 0), ("r", 1)]:
        y, _ = sf.read(os.path.join(STAGING, f"{name}_{side}.wav"), always_2d=True)
        assert np.max(np.abs(y[:, 1 - ch])) < 1e-6, f"{name}_{side}: 非目标声道有能量"
        assert np.max(np.abs(y[:, ch])) > 0.1, f"{name}_{side}: 目标声道无能量"


def main():
    os.makedirs(STAGING, exist_ok=True)
    if "--emit" in sys.argv:
        emit()
        return
    print(f"| {'id':14s} | {'时长s':>6s} | {'ogg峰dBFS':>9s} | {'m4a峰dBFS':>9s} | {'RMS dBFS':>8s} | 直流 |")
    idx = 0
    for name, _d, _p, mode, tgt, max_s in ITEMS:
        if mode == "skip":
            continue
        raw = os.path.join(STAGING, f"abyss_{idx}.wav")
        y, sr = sf.read(raw, always_2d=True, dtype="float64")
        assert sr == SR, f"{name}: 采样率 {sr}"
        if y.shape[1] == 1:
            y = np.stack([y[:, 0], y[:, 0]], axis=1)
        idx += 1
        if mode == "oneshot":
            sig = prep_oneshot(y, tgt, win=WIN.get(name), max_s=max_s)
        else:
            sig = prep_amb(y, tgt, -8.0)
        if name == "echo_metal":
            # mono 下混 -> pan 派生 _l/_r（保留 oneshot 修整/淡化/定档后的信号）
            mono = sig.mean(axis=1)
            pad_h = np.zeros((int(SR * 0.06), 2))
            pad_t = np.zeros((int(SR * 0.12), 2))
            pan_derive(name, mono)
            for side in ("l", "r"):
                s2, _ = sf.read(os.path.join(STAGING, f"{name}_{side}.wav"), always_2d=True)
                res = encode(f"echo_metal_{side}", np.concatenate([pad_h, s2, pad_t]))
                tp_o, dc, rms, dur = res["ogg"]
                print(f"| echo_metal_{side:<3s} | {dur:6.2f} | {tp_o:9.2f} | {res['m4a'][0]:9.2f} | {rms:8.2f} | {dc:+.4f} |")
            continue
        res = encode(name, sig)
        tp_o, dc, rms, dur = res["ogg"]
        tp_m = res["m4a"][0]
        print(f"| {name:14s} | {dur:6.2f} | {tp_o:9.2f} | {tp_m:9.2f} | {rms:8.2f} | {dc:+.4f} |")
    print("abyss SA3 sfx done -> web/abyss/audio/sfx/")


if __name__ == "__main__":
    main()