typedef struct {
	const char *devname;
	uint samplerate;
	uint frames;
	uint bufsize;
	uint width;
} snd_devconfig;

enum snd_cmd {
	SND_CMD_STOP,
	SND_CMD_FLUSH,
	SND_CMD_QUERY,
	SND_CMD_RESET,
	SND_CMD_MKOPL,
	SND_CMD_VOICES,
	// SND_CMD_BANK,
	SND_CMD_PAUSE,
	SND_CMD_LOAD,
	SND_CMD_OFF,
	SND_CMD_DEBUG,
	SND_CMD_WAVECAP,
	SND_CMD_VOLUME
};

enum snd_stat {
	SND_STAT_RUN,
	SND_STAT_STOPPED
};

ulong snd_process() {
	ulong elapsed = 1000ULL;
	if(!sgt.paused) {
		if(sgt.mid.track && (sgt.playing ^ 1) && (sgt.nowait || !(sgt.chip) || !OPL3_Playing(sgt.chip))) {
			// flush ?
			if(sgt.mid.track) {
				free(sgt.mid.track);
				sgt.mid.track = NULL;
			}
			if(sgt.chip) {
				OPL3_Reset(sgt.chip);
				bank_reset(sgt.chip, sgt.bank);
			}
			sgt.mid_tick = 0;
			sgt.mid_time = 0ULL;
			sgt.stopped = 1;
		}
		else if(sgt.mid.track && sgt.playing && mid_tick(&sgt.mid, sgt.bank, sgt.chip)) {
			elapsed = ((ulong)sgt.mid.ticktime) * 1000ULL;
			sgt.mid_tick += 1;
			sgt.mid_time += ((ulong)sgt.mid.ticktime);
		}
		else if(sgt.playing) {
			sgt.playing = 0;
			if(sgt.bank)
				bank_alloff(sgt.bank);
		}
	}
	// sampling ...
	return elapsed;
}

void snd_generate(short *sample) {
	int mix[2];
	if(!sgt.paused && sgt.chip)
		OPL3_GenerateResampled(sgt.chip, &sample[SND_VOL_MUSIC*2]);
	else
		sample[SND_VOL_MUSIC*2+0] = sample[SND_VOL_MUSIC*2+1] = 0;
	// sampling ...
	sample[SND_VOL_SFX*2+0] = sample[SND_VOL_SFX*2+1] = 0;
	sample[SND_VOL_GUI*2+0] = sample[SND_VOL_GUI*2+1] = 0;
	mix[0] = mix[1] = 0;
	for(int z = 1; z < SND_VOLUMES; z++) {
		mix[0] += (short)(int)(((int)sample[z*2+0] * (int)sgt.volumes[z]) / 32767);
		mix[1] += (short)(int)(((int)sample[z*2+1] * (int)sgt.volumes[z]) / 32767);
	}
	sample[0] = (short)(int)((mix[0] * (int)sgt.volumes[0]) / 32767);
	sample[1] = (short)(int)((mix[1] * (int)sgt.volumes[0]) / 32767);
}

void snd_dispatch(void *buf, short op, short addr, uint param, uint rate) {
	switch(op) {
		case SND_CMD_PAUSE:
			sgt.paused = (byte)param;
			break;
		case SND_CMD_LOAD:
			// mid_cmlog();
			sgt.playing = 0;
			if(sgt.mid.track) {
				free(sgt.mid.track);
				sgt.mid.track = NULL;
			}
			if(buf && sgt.chip) {
				sgt.mid.tracks = param & 0xffff;
				sgt.mid.tpqn = (param >> 16) & 0x7fff;
				mid_settempo(&sgt.mid, MID_DEFTEMPO);
				sgt.mid.track = (mid_track *)buf;
				buf += (sizeof(mid_track) * sgt.mid.tracks);
				for(int z = 0; z < sgt.mid.tracks; z++) {
					sgt.mid.track[z].buffer = buf;
					buf += sgt.mid.track[z].size;
				}
				sgt.playing = 1;
				sgt.nowait = (byte)(param >> 31);
			}
			else if(buf) {
				free(buf);
			}
		case SND_CMD_RESET:
			if(sgt.chip) {
				OPL3_Reset(sgt.chip);
				bank_reset(sgt.chip, sgt.bank);
			}
			sgt.mid_tick = 0;
			sgt.mid_time = 0ULL;
			break;
		case SND_CMD_OFF:
			if(sgt.mid.track)
				bank_alloff(sgt.bank);
			break;
		case SND_CMD_MKOPL:
			if(sgt.chip) {
				free(sgt.bank);
				free(sgt.chip);
			}
			sgt.chip = param ? OPL3_Alloc(rate, param & 0xff) : NULL;
			sgt.bank = param ? bank_alloc(sgt.chip, sgt.ibank, (param >> 16) & 1, (param >> 17) & 1, (param >> 8) & 0xff) : NULL;
			sgt.stopped = param ? 1 : 0;
			break;
		case SND_CMD_VOICES:
			if(buf) {
				memcpy(sgt.ibank, buf, sizeof(bank_instr) * 256);
				free(buf);
			}
			// else {
				// memcpy(sgt.ibank, snd_banks[addr], sizeof(bank_instr) * 256);
			// }
			break;
		// case SND_CMD_BANK:
			// sgt.instr = addr >= 31 ? sgt.ibank : (bank_instr *)snd_banks[addr];
			// break;
		case SND_CMD_DEBUG:
			sgt.log_debug = (byte)param;
			break;
		case SND_CMD_WAVECAP:
			if(sgt.capture)
				wav_close(&sgt.wav);
			sgt.capture = buf && wav_open(&sgt.wav, (char *)buf, rate, 2, 2);
			if(buf)
				free(buf);
			break;
		case SND_CMD_VOLUME:
			sgt.volumes[addr] = (ushort)param;
			// snd_logd("V %d = %d", addr, param);
			break;
	}
}

uint snd_status(short addr, uint param) {
	switch(addr) {
		case SND_STAT_STOPPED:
			param = sgt.stopped;
			sgt.stopped = 0;
			return param;
		default:
			return 0;
	}
}

void *snd_loop(void *arg) {
	snd_devconfig *cfg = (snd_devconfig*)arg;
	aud_handle handle;
	uint pop = 0;
	short sample[2 * SND_VOLUMES];
	uint rate = strcmp(cfg->devname, "none") ? aud_open_dev(&handle, cfg->devname, cfg->samplerate, 2, (byte)cfg->width, cfg->frames, cfg->bufsize) : 0;
	ulong sampletime = rate ? (1000000000ULL / (ulong)rate) : 100000000000000000ULL;
	byte running = 1;
	ulong elapsed = 0;
	ulong samples = 0;
	sample[0] = sample[1] = 0;
	while(running) {
		pthread_mutex_lock(&sgt.lock);
		while(sgt.cmd_queue) {
			snd_cmd_t *cmd = &sgt.cmds[pop];
			switch(cmd->command) {
				case SND_CMD_STOP:
					running = 0;
					break;
				case SND_CMD_FLUSH:
					
					elapsed = 0;
					if(rate)
						aud_flush(&handle);
					break;
				case SND_CMD_QUERY:
					switch(cmd->address) {
						case SND_STAT_RUN:
							sgt.query = rate;
							break;
						default:
							sgt.query = snd_status(cmd->address, cmd->param);
							break;
					}
					sgt.waiting = 0;
					break;
				default:
					snd_dispatch(cmd->buffer, cmd->command, cmd->address, cmd->param, rate);
					break;
			}
			if(++pop == SND_QUEUE)
				pop = 0;
			sgt.cmd_queue -= 1;
		}
		pthread_mutex_unlock(&sgt.lock);
		if(sgt.interrupt)
			sgt.interrupt = 0;
		else
			elapsed += snd_process();
		// if((elapsed < sampletime) || !rate)
		// 	usleep(1000);
		while((elapsed >= sampletime) && !sgt.interrupt) {
			snd_generate(sample);
			if(rate)
				aud_push(&handle, sample);
			if(sgt.capture)
				wav_write16(&sgt.wav, sample, 1);
			samples++;
			elapsed -= sampletime;
		}
	}
	if(rate) {
		aud_flush(&handle);
		aud_close(&handle);
	}
	if(sgt.mid.track)
		free(sgt.mid.track);
	if(sgt.chip) {
		free(sgt.bank);
		free(sgt.chip);
	}
	sgt.query = rate;
	sgt.waiting = 1;
}

void snd_instruct(const void *buf, uint size, short op, short addr, uint param) {
	if(sgt.disabled)
		return;
	sgt.interrupt = 1;
	while(sgt.cmd_queue == SND_QUEUE) {
		;
	}
	pthread_mutex_lock(&sgt.lock);
	sgt.interrupt = 1;
	snd_cmd_t *cmd = &sgt.cmds[sgt.cmd_push];
	cmd->command = op;
	cmd->address = addr;
	cmd->param = param;
	if(buf) {
		cmd->buffer = malloc(size);
		memcpy(cmd->buffer, buf, size);
	}
	else {
		cmd->buffer = NULL;
	}
	if(++(sgt.cmd_push) == SND_QUEUE)
		sgt.cmd_push = 0;
	sgt.cmd_queue += 1;
	pthread_mutex_unlock(&sgt.lock);
}

uint snd_query(short addr, uint param) {
	if(sgt.disabled || sgt.waiting)
		return 0;
	sgt.waiting = 1;
	snd_instruct(NULL, 0, SND_CMD_QUERY, addr, param);
	while(sgt.waiting) {
		;
	}
	return sgt.query;
}

uint snd_init(const char *devname, uint samplerate, uint width, uint frames, uint bufsize, byte disabled) {
	if(disabled) {
		sgt.disabled = 1;
		return 0;
	}
	snd_devconfig snd_cfg;
	snd_cfg.devname = devname;
	snd_cfg.samplerate = samplerate;
	snd_cfg.width = ((width == 32) ? 32 : 16) / 8;
	snd_cfg.frames = frames;
	snd_cfg.bufsize = bufsize;
	pthread_mutex_init(&sgt.lock, NULL);
	pthread_mutex_init(&sgt.log_lock, NULL);
	pthread_create(&sgt.thread, NULL, snd_loop, &snd_cfg);
	return snd_query(SND_STAT_RUN, 0);
}

uint snd_end() {
	if(sgt.disabled)
		return 0;
	snd_instruct(NULL, 0, SND_CMD_STOP, 0, 0);
	while(sgt.waiting ^ 1) {
		;
	}
	pthread_mutex_destroy(&sgt.lock);
	pthread_mutex_destroy(&sgt.log_lock);
	return sgt.query;
}