"""GPU reference encoding must preserve all PCM samples and edited transcripts."""
import base64
import io
import numpy as np
import soundfile as sf

from test_gradio_ui import studio, body
import server


def test_clone_and_retake_compression_preserve_every_pcm_sample(studio):
    # Include both PCM extremes, plus very quiet samples used by whispered voices.
    samples = np.array([-32768, -3000, -1, 0, 1, 3000, 32767] * 500, dtype='<i2')
    wav = io.BytesIO()
    sf.write(wav, samples, 44100, format='WAV', subtype='PCM_16')
    voice = server.Voice('test', 'Test', wav.getvalue(), 'Original words.', len(samples) / 44100)
    choice = server.VoiceChoice('clone', voice=voice, transcript='Edited transcript.')
    take = server.Take('take', choice, 'Earlier line.', 'Earlier line.', [], samples.tobytes(), 44100)
    request = server.engine_request(choice, 'Next line.', [take], body(studio))
    assert voice.wav == wav.getvalue()
    assert [r['text'] for r in request['references']] == ['Edited transcript.', 'Earlier line.']
    for reference in request['references']:
        assert reference['media_type'] == 'audio/flac'
        decoded, rate = sf.read(io.BytesIO(base64.b64decode(reference['data'])), dtype='int16')
        assert rate == 44100 and np.array_equal(samples, decoded)