package com.kazeia.tts
// Engine TTS Qwen3 in-process (libkazeia_tts.so). Load une fois, synthesize N fois.
// Le bridge natif est dans dist/jni/kazeia_tts_jni.cpp ; build via dist/build_kazeia_tts.sh
// ou dist/CMakeLists.txt.
//
// Empreinte RAM typique : ~3.5 GB (talker f32 ~1.6 GB + CP f16 mixte ~250 MB +
// decoder ~325 MB + vocab Qwen3 ~50 MB + fixtures text_embed/tp_* ~1.2 GB).

class TtsJni {
    external fun nativeLoad(
        talkerGguf: String,
        vocabGguf: String,
        dumpDir: String,
        useHtp: Boolean,
        nThreads: Int,
        cpUseCache: Boolean
    ): Long

    external fun nativeSynthesize(
        handle: Long,
        text: String,
        outWavPath: String,
        maxSteps: Int,
        seed: Int,
        cpTemp: Float, cpTopK: Int, cpTopP: Float, cpRepPenalty: Float,
        talkerTemp: Float, talkerTopK: Int, talkerTopP: Float, talkerRepPenalty: Float
    ): IntArray   // [err, frames, total_ms, prefill_ms, talker_ms, cp_ms, decoder_ms]

    external fun nativeFree(handle: Long)

    companion object { init { System.loadLibrary("kazeia_tts") } }
}

data class TtsResult(
    val frames: Int,
    val audioMs: Int,
    val totalMs: Int,
    val prefillMs: Int,
    val talkerMs: Int,
    val cpMs: Int,
    val decoderMs: Int
)

// Configuration sampling. Défauts validés sur fixture FR (cf project_tts_session_28may).
data class TtsSampling(
    val cpTemp: Float = 0.9f, val cpTopK: Int = 50, val cpTopP: Float = 1.0f, val cpRepPenalty: Float = 1.05f,
    val talkerTemp: Float = 0.9f, val talkerTopK: Int = 50, val talkerTopP: Float = 1.0f, val talkerRepPenalty: Float = 1.05f
)

class TtsEngine(
    talkerGguf: String,
    vocabGguf: String,
    dumpDir: String,
    useHtp: Boolean = false,
    nThreads: Int = 6,
    cpUseCache: Boolean = true
) {
    private val jni = TtsJni()
    private val h = jni.nativeLoad(talkerGguf, vocabGguf, dumpDir, useHtp, nThreads, cpUseCache)
    init { require(h != 0L) { "TtsEngine: load FAILED (talker=$talkerGguf, vocab=$vocabGguf, dump=$dumpDir)" } }

    @Throws(IllegalStateException::class)
    fun synthesize(
        text: String,
        outWavPath: String,
        maxSteps: Int = 256,
        seed: Int = 42,
        sampling: TtsSampling = TtsSampling()
    ): TtsResult {
        val r = jni.nativeSynthesize(
            h, text, outWavPath, maxSteps, seed,
            sampling.cpTemp, sampling.cpTopK, sampling.cpTopP, sampling.cpRepPenalty,
            sampling.talkerTemp, sampling.talkerTopK, sampling.talkerTopP, sampling.talkerRepPenalty
        )
        if (r[0] != 0) error("TtsEngine.synthesize err=${r[0]} for text=\"$text\"")
        val frames = r[1]
        return TtsResult(
            frames    = frames,
            audioMs   = frames * 1000 / 12,   // 12 frames/s = 80 ms par frame
            totalMs   = r[2],
            prefillMs = r[3],
            talkerMs  = r[4],
            cpMs      = r[5],
            decoderMs = r[6]
        )
    }

    fun release() { jni.nativeFree(h) }
}