/*
 * WMA compatible decoder
 *
 * Extracted 2009 from the mplayer source code 2009-02-10.
 *
 * Copyright (c) 2002 The FFmpeg Project
 *
 * Licensed under the GNU Lesser General Public License, see file COPYING.LIB.
 */

/** \file wmadec_filter.c paraslash's WMA decoder. */

/*
 * This decoder handles Microsoft Windows Media Audio data version 2.
 */

#include <math.h>

#include "para.h"
#include "error.h"
#include "list.h"
#include "string.h"
#include "sched.h"
#include "buffer_tree.h"
#include "filter.h"
#include "portable_io.h"
#include "bitstream.h"
#include "imdct.h"
#include "wma.h"
#include "wmadata.h"


/* size of blocks */
#define BLOCK_MIN_BITS 7
#define BLOCK_MAX_BITS 11
#define BLOCK_MAX_SIZE (1 << BLOCK_MAX_BITS)

#define BLOCK_NB_SIZES (BLOCK_MAX_BITS - BLOCK_MIN_BITS + 1)

/* XXX: is it a suitable value ? */
#define MAX_CODED_SUPERFRAME_SIZE 16384

#define MAX_CHANNELS 2
#define LSP_POW_BITS 7

struct private_wmadec_data {
	/** Information contained in the audio file header. */
	struct asf_header_info ahi;
	struct getbit_context gb;
	/** Depends on number of the bits per second and the frame length. */
	int byte_offset_bits;
	/** Only used if ahi->use_exp_vlc is true. */
	struct vlc exp_vlc;
	uint16_t exponent_bands[BLOCK_NB_SIZES][25];
	/** The index of the first coef in high band. */
	int high_band_start[BLOCK_NB_SIZES];
	/** Maximal number of coded coefficients. */
	int coefs_end[BLOCK_NB_SIZES];
	/* there are two possible tables for spectral coefficients */
	struct vlc coef_vlc[2];
	uint16_t *run_table[2];
	uint16_t *level_table[2];
	/** Frame length in samples. */
	int frame_len;
	/** log2 of frame_len. */
	int frame_len_bits;
	/** Number of block sizes, one if !ahi->use_variable_block_len. */
	int nb_block_sizes;
	/* Whether to update block lengths from getbit context. */
	bool reset_block_lengths;
	/** log2 of current block length. */
	int block_len_bits;
	/** log2 of next block length. */
	int next_block_len_bits;
	/** log2 of previous block length. */
	int prev_block_len_bits;
	/** Block length in samples. */
	int block_len;
	/** Current position in frame. */
	int block_pos;
	/** True if channel is coded. */
	uint8_t channel_coded[MAX_CHANNELS];
	/** log2 ratio frame/exp. length. */
	int exponents_bsize[MAX_CHANNELS];

	float exponents[MAX_CHANNELS][BLOCK_MAX_SIZE];
	float max_exponent[MAX_CHANNELS];
	int16_t coefs1[MAX_CHANNELS][BLOCK_MAX_SIZE];
	float coefs[MAX_CHANNELS][BLOCK_MAX_SIZE];
	float output[BLOCK_MAX_SIZE * 2];
	struct mdct_context *mdct_ctx[BLOCK_NB_SIZES];
	float *windows[BLOCK_NB_SIZES];
	/** Output buffer for one frame and the last for IMDCT windowing. */
	float frame_out[MAX_CHANNELS][BLOCK_MAX_SIZE * 2];
	/** Last frame info. */
	uint8_t last_superframe[MAX_CODED_SUPERFRAME_SIZE + 4];	/* padding added */
	int last_bitoffset;
	int last_superframe_len;
	/* lsp_to_curve tables */
	float lsp_cos_table[BLOCK_MAX_SIZE];
	float lsp_pow_e_table[256];
	float lsp_pow_m_table1[(1 << LSP_POW_BITS)];
	float lsp_pow_m_table2[(1 << LSP_POW_BITS)];
};

#define EXPVLCBITS 8
#define VLCBITS 9

/** \cond sine_windows */

#define SINE_WINDOW(x) static float sine_ ## x[x] __a_aligned(16)

SINE_WINDOW(128);
SINE_WINDOW(256);
SINE_WINDOW(512);
SINE_WINDOW(1024);
SINE_WINDOW(2048);
SINE_WINDOW(4096);

static float *sine_windows[6] = {
	sine_128, sine_256, sine_512, sine_1024, sine_2048, sine_4096
};
/** \endcond sine_windows */

/* Generate a sine window. */
static void sine_window_init(float *window, int n)
{
	int i;

	for (i = 0; i < n; i++)
		window[i] = sinf((i + 0.5) * (M_PI / (2.0 * n)));
}

static void init_coef_vlc(struct private_wmadec_data *pwd, int sidx, int didx)
{
	const struct coef_vlc_table *src = coef_vlcs + sidx;
	struct vlc *dst = pwd->coef_vlc + didx;
	int i, l, j, k, level, n = src->n;

	init_vlc(dst, VLCBITS, n, src->huffbits, src->huffcodes, 4);
	pwd->run_table[didx] = arr_alloc(n, sizeof(uint16_t));
	pwd->level_table[didx] = arr_alloc(n, sizeof(uint16_t));
	i = 2;
	level = 1;
	k = 0;
	while (i < n) {
		l = src->levels[k++];
		for (j = 0; j < l; j++) {
			pwd->run_table[didx][i] = j;
			pwd->level_table[didx][i] = level;
			i++;
		}
		level++;
	}
}

/* compute the scale factor band sizes for each MDCT block size */
static void compute_scale_factor_band_sizes(struct private_wmadec_data *pwd,
	float high_freq)
{
	struct asf_header_info *ahi = &pwd->ahi;
	int a, b, pos, lpos, k, block_len, i, j, n;
	const uint8_t *table;

	for (k = 0; k < pwd->nb_block_sizes; k++) {
		int exponent_size;

		block_len = pwd->frame_len >> k;
		table = NULL;
		a = pwd->frame_len_bits - BLOCK_MIN_BITS - k;
		if (a < 3) {
			if (ahi->sample_rate >= 44100)
				table = exponent_band_44100[a];
			else if (ahi->sample_rate >= 32000)
				table = exponent_band_32000[a];
			else if (ahi->sample_rate >= 22050)
				table = exponent_band_22050[a];
		}
		if (table) {
			n = *table++;
			for (i = 0; i < n; i++)
				pwd->exponent_bands[k][i] = table[i];
			exponent_size = n;
		} else {
			j = 0;
			lpos = 0;
			for (i = 0; i < 25; i++) {
				a = wma_critical_freqs[i];
				b = ahi->sample_rate;
				pos = ((block_len * 2 * a) + (b << 1)) / (4 * b);
				pos <<= 2;
				if (pos > block_len)
					pos = block_len;
				if (pos > lpos)
					pwd->exponent_bands[k][j++] = pos - lpos;
				if (pos >= block_len)
					break;
				lpos = pos;
			}
			exponent_size = j;
		}

		/* max number of coefs */
		pwd->coefs_end[k] = (pwd->frame_len - ((pwd->frame_len * 9) / 100)) >> k;
		/* high freq computation */
		pwd->high_band_start[k] = (int) ((block_len * 2 * high_freq)
			/ ahi->sample_rate + 0.5);
		n = exponent_size;
		pos = 0;
		for (i = 0; i < n; i++)
			pos += pwd->exponent_bands[k][i];
	}
}

static int wma_init(struct private_wmadec_data *pwd)
{
	int i;
	float bps1, high_freq;
	volatile float bps;
	int sample_rate1;
	int coef_vlc_table;
	struct asf_header_info *ahi = &pwd->ahi;
	int flags2 = ahi->flags2;

	if (ahi->sample_rate <= 0 || ahi->sample_rate > 50000
		|| ahi->channels <= 0 || ahi->channels > 8
		|| ahi->bit_rate <= 0)
		return -E_WMA_BAD_PARAMS;

	/* compute MDCT block size */
	if (ahi->sample_rate <= 16000)
		pwd->frame_len_bits = 9;
	else if (ahi->sample_rate <= 22050)
		pwd->frame_len_bits = 10;
	else
		pwd->frame_len_bits = 11;
	pwd->frame_len = 1 << pwd->frame_len_bits;
	if (pwd->ahi.use_variable_block_len) {
		int nb_max, nb;
		nb = ((flags2 >> 3) & 3) + 1;
		if ((ahi->bit_rate / ahi->channels) >= 32000)
			nb += 2;
		nb_max = pwd->frame_len_bits - BLOCK_MIN_BITS;
		if (nb > nb_max)
			nb = nb_max;
		pwd->nb_block_sizes = nb + 1;
	} else
		pwd->nb_block_sizes = 1;

	/* init rate dependent parameters */
	high_freq = ahi->sample_rate * 0.5;

	/* wma2 rates are normalized */
	sample_rate1 = ahi->sample_rate;
	if (sample_rate1 >= 44100)
		sample_rate1 = 44100;
	else if (sample_rate1 >= 22050)
		sample_rate1 = 22050;
	else if (sample_rate1 >= 16000)
		sample_rate1 = 16000;
	else if (sample_rate1 >= 11025)
		sample_rate1 = 11025;
	else if (sample_rate1 >= 8000)
		sample_rate1 = 8000;

	bps = (float) ahi->bit_rate / (float) (ahi->channels * ahi->sample_rate);
	pwd->byte_offset_bits = wma_log2((int) (bps * pwd->frame_len / 8.0 + 0.5)) + 2;
	/* Adjust bps and compute high frequency value. */
	bps1 = bps;
	if (ahi->channels == 2)
		bps1 = bps * 1.6;
	if (sample_rate1 == 44100) {
		if (bps1 < 0.61)
			high_freq = high_freq * 0.4;
	} else if (sample_rate1 == 22050) {
		if (bps1 < 1.16) {
			if (bps1 >= 0.72)
				high_freq = high_freq * 0.7;
			else
				high_freq = high_freq * 0.6;
		}
	} else if (sample_rate1 == 16000) {
		if (bps > 0.5)
			high_freq = high_freq * 0.5;
		else
			high_freq = high_freq * 0.3;
	} else if (sample_rate1 == 11025)
		high_freq = high_freq * 0.7;
	else if (sample_rate1 == 8000) {
		if (bps <= 0.625)
			high_freq = high_freq * 0.5;
		else if (bps <= 0.75)
			high_freq = high_freq * 0.65;
	} else {
		if (bps >= 0.8)
			high_freq = high_freq * 0.75;
		else if (bps >= 0.6)
			high_freq = high_freq * 0.6;
		else
			high_freq = high_freq * 0.5;
	}
	PARA_INFO_LOG("channels=%u sample_rate=%u "
		"bitrate=%u block_align=%d\n",
		ahi->channels, ahi->sample_rate,
		ahi->bit_rate, ahi->block_align);
	PARA_INFO_LOG("frame_len=%d, bps=%f bps1=%f "
		"high_freq=%f bitoffset=%d\n",
		pwd->frame_len, bps, bps1,
		high_freq, pwd->byte_offset_bits);
	PARA_INFO_LOG("use_exp_vlc=%d nb_block_sizes=%d\n",
		pwd->ahi.use_exp_vlc, pwd->nb_block_sizes);

	compute_scale_factor_band_sizes(pwd, high_freq);
	/* init MDCT windows : simple sinus window */
	for (i = 0; i < pwd->nb_block_sizes; i++) {
		int n;
		n = 1 << (pwd->frame_len_bits - i);
		sine_window_init(sine_windows[pwd->frame_len_bits - i - 7], n);
		pwd->windows[i] = sine_windows[pwd->frame_len_bits - i - 7];
	}

	pwd->reset_block_lengths = true;

	/* choose the VLC tables for the coefficients */
	coef_vlc_table = 4;
	if (ahi->sample_rate >= 32000) {
		if (bps1 < 0.72)
			coef_vlc_table = 0;
		else if (bps1 < 1.16)
			coef_vlc_table = 2;
	}
	init_coef_vlc(pwd, coef_vlc_table, 0);
	init_coef_vlc(pwd, coef_vlc_table + 1, 1);
	return 0;
}

static void wma_lsp_to_curve_init(struct private_wmadec_data *pwd)
{
	float wdel, a, b;
	int i, e, m;

	wdel = M_PI / pwd->frame_len;
	for (i = 0; i < pwd->frame_len; i++)
		pwd->lsp_cos_table[i] = 2.0f * cos(wdel * i);

	/* tables for x^-0.25 computation */
	for (i = 0; i < 256; i++) {
		e = i - 126;
		pwd->lsp_pow_e_table[i] = pow(2.0, e * -0.25);
	}

	/* These two tables are needed to avoid two operations in pow_m1_4. */
	b = 1.0;
	for (i = (1 << LSP_POW_BITS) - 1; i >= 0; i--) {
		m = (1 << LSP_POW_BITS) + i;
		a = (float) m *(0.5 / (1 << LSP_POW_BITS));
		a = pow(a, -0.25);
		pwd->lsp_pow_m_table1[i] = 2 * a - b;
		pwd->lsp_pow_m_table2[i] = b - a;
		b = a;
	}
}

static int wma_decode_init(char *initial_buf, int len, struct private_wmadec_data **result)
{
	struct private_wmadec_data *pwd;
	int ret, i;

	PARA_NOTICE_LOG("initial buf: %d bytes\n", len);
	pwd = zalloc(sizeof(*pwd));
	ret = read_asf_header(initial_buf, len, &pwd->ahi);
	if (ret <= 0) {
		free(pwd);
		return ret;
	}

	ret = wma_init(pwd);
	if (ret < 0)
		return ret;
	/* init MDCT */
	for (i = 0; i < pwd->nb_block_sizes; i++) {
		ret = imdct_init(pwd->frame_len_bits - i + 1, &pwd->mdct_ctx[i]);
		if (ret < 0)
			return ret;
	}
	if (pwd->ahi.use_exp_vlc) {
		PARA_INFO_LOG("using exp_vlc\n");
		init_vlc(&pwd->exp_vlc, EXPVLCBITS, sizeof(wma_scale_huffbits),
			wma_scale_huffbits, wma_scale_huffcodes, 4);
	} else {
		PARA_INFO_LOG("using curve\n");
		wma_lsp_to_curve_init(pwd);
	}
	*result = pwd;
	return pwd->ahi.header_len;
}

/**
 * compute x^-0.25 with an exponent and mantissa table. We use linear
 * interpolation to reduce the mantissa table size at a small speed
 * expense (linear interpolation approximately doubles the number of
 * bits of precision).
 */
static inline float pow_m1_4(struct private_wmadec_data *pwd, float x)
{
	union {
		float f;
		unsigned int v;
	} u, t;
	unsigned int e, m;
	float a, b;

	u.f = x;
	e = u.v >> 23;
	m = (u.v >> (23 - LSP_POW_BITS)) & ((1 << LSP_POW_BITS) - 1);
	/* build interpolation scale: 1 <= t < 2. */
	t.v = ((u.v << LSP_POW_BITS) & ((1 << 23) - 1)) | (127 << 23);
	a = pwd->lsp_pow_m_table1[m];
	b = pwd->lsp_pow_m_table2[m];
	return pwd->lsp_pow_e_table[e] * (a + b * t.f);
}

static void wma_lsp_to_curve(struct private_wmadec_data *pwd,
		float *out, float *val_max_ptr, int n, float *lsp)
{
	int i, j;
	float p, q, w, v, val_max;

	val_max = 0;
	for (i = 0; i < n; i++) {
		p = 0.5f;
		q = 0.5f;
		w = pwd->lsp_cos_table[i];
		for (j = 1; j < NB_LSP_COEFS; j += 2) {
			q *= w - lsp[j - 1];
			p *= w - lsp[j];
		}
		p *= p * (2.0f - w);
		q *= q * (2.0f + w);
		v = p + q;
		v = pow_m1_4(pwd, v);
		if (v > val_max)
			val_max = v;
		out[i] = v;
	}
	*val_max_ptr = val_max;
}

/* Decode exponents coded with LSP coefficients (same idea as Vorbis). */
static void decode_exp_lsp(struct private_wmadec_data *pwd, int ch)
{
	float lsp_coefs[NB_LSP_COEFS];
	int val, i;

	for (i = 0; i < NB_LSP_COEFS; i++) {
		if (i == 0 || i >= 8)
			val = get_bits(&pwd->gb, 3);
		else
			val = get_bits(&pwd->gb, 4);
		lsp_coefs[i] = wma_lsp_codebook[i][val];
	}

	wma_lsp_to_curve(pwd, pwd->exponents[ch], &pwd->max_exponent[ch],
		pwd->block_len, lsp_coefs);
}

/* Decode exponents coded with VLC codes. */
static int decode_exp_vlc(struct private_wmadec_data *pwd, int ch)
{
	int last_exp, n, code;
	const uint16_t *ptr, *band_ptr;
	float v, *q, max_scale, *q_end;

	band_ptr = pwd->exponent_bands[pwd->frame_len_bits - pwd->block_len_bits];
	ptr = band_ptr;
	q = pwd->exponents[ch];
	q_end = q + pwd->block_len;
	max_scale = 0;
	last_exp = 36;

	while (q < q_end) {
		code = get_vlc(&pwd->gb, &pwd->exp_vlc);
		if (code < 0)
			return code;
		/* NOTE: this offset is the same as MPEG4 AAC ! */
		last_exp += code - 60;
		/* XXX: use a table */
		v = pow(10, last_exp * (1.0 / 16.0));
		if (v > max_scale)
			max_scale = v;
		n = *ptr++;
		do {
			*q++ = v;
		} while (--n);
	}
	pwd->max_exponent[ch] = max_scale;
	return 0;
}

/* compute src0 * src1 + src2 */
static inline void vector_mult_add(float *dst, const float *src0, const float *src1,
		const float *src2, int len)
{
	int i;

	for (i = 0; i < len; i++)
		dst[i] = src0[i] * src1[i] + src2[i];
}

static inline void vector_mult_reverse(float *dst, const float *src0,
		const float *src1, int len)
{
	int i;

	src1 += len - 1;
	for (i = 0; i < len; i++)
		dst[i] = src0[i] * src1[-i];
}

/**
 * Apply MDCT window and add into output.
 *
 * We ensure that when the windows overlap their squared sum
 * is always 1 (MDCT reconstruction rule).
 */
static void wma_window(struct private_wmadec_data *pwd, float *out)
{
	float *in = pwd->output;
	int block_len, bsize, n;

	/* left part */
	if (pwd->block_len_bits <= pwd->prev_block_len_bits) {
		block_len = pwd->block_len;
		bsize = pwd->frame_len_bits - pwd->block_len_bits;
		vector_mult_add(out, in, pwd->windows[bsize], out, block_len);
	} else {
		block_len = 1 << pwd->prev_block_len_bits;
		n = (pwd->block_len - block_len) / 2;
		bsize = pwd->frame_len_bits - pwd->prev_block_len_bits;
		vector_mult_add(out + n, in + n, pwd->windows[bsize], out + n,
			block_len);
		memcpy(out + n + block_len, in + n + block_len,
			n * sizeof(float));
	}
	out += pwd->block_len;
	in += pwd->block_len;
	/* right part */
	if (pwd->block_len_bits <= pwd->next_block_len_bits) {
		block_len = pwd->block_len;
		bsize = pwd->frame_len_bits - pwd->block_len_bits;
		vector_mult_reverse(out, in, pwd->windows[bsize], block_len);
	} else {
		block_len = 1 << pwd->next_block_len_bits;
		n = (pwd->block_len - block_len) / 2;
		bsize = pwd->frame_len_bits - pwd->next_block_len_bits;
		memcpy(out, in, n * sizeof(float));
		vector_mult_reverse(out + n, in + n, pwd->windows[bsize],
			block_len);
		memset(out + n + block_len, 0, n * sizeof(float));
	}
}

static int wma_total_gain_to_bits(int total_gain)
{
	if (total_gain < 15)
		return 13;
	else if (total_gain < 32)
		return 12;
	else if (total_gain < 40)
		return 11;
	else if (total_gain < 45)
		return 10;
	else
		return 9;
}

static void compute_mdct_coefficients(struct private_wmadec_data *pwd,
		int bsize, int total_gain, int nb_coefs[MAX_CHANNELS])
{
	int ch;
	float mdct_norm = 1.0 / (pwd->block_len / 2);

	for (ch = 0; ch < pwd->ahi.channels; ch++) {
		int16_t *coefs1;
		float *coefs, *exponents, mult;
		int i, n, esize;

		if (!pwd->channel_coded[ch])
			continue;
		coefs1 = pwd->coefs1[ch];
		exponents = pwd->exponents[ch];
		esize = pwd->exponents_bsize[ch];
		mult = pow(10, total_gain * 0.05) / pwd->max_exponent[ch];
		mult *= mdct_norm;
		coefs = pwd->coefs[ch];
		/* XXX: optimize more */
		n = nb_coefs[ch];
		for (i = 0; i < n; i++)
			*coefs++ = coefs1[i] *
				exponents[i << bsize >> esize] * mult;
		n = pwd->block_len - pwd->coefs_end[bsize];
		for (i = 0; i < n; i++)
			*coefs++ = 0.0;
	}
}

/**
 * Returns 0 if OK, 1 if last block of frame, negative on uncorrectable
 * errors.
 */
static int wma_decode_block(struct private_wmadec_data *pwd)
{
	int ret, n, v, ch, code, bsize;
	int coef_nb_bits, total_gain;
	int nb_coefs[MAX_CHANNELS];
	bool ms_stereo = false; /* mid/side stereo mode */

	/* compute current block length */
	if (pwd->ahi.use_variable_block_len) {
		n = wma_log2(pwd->nb_block_sizes - 1) + 1;

		if (pwd->reset_block_lengths) {
			pwd->reset_block_lengths = false;
			v = get_bits(&pwd->gb, n);
			if (v >= pwd->nb_block_sizes)
				return -E_WMA_BLOCK_SIZE;
			pwd->prev_block_len_bits = pwd->frame_len_bits - v;
			v = get_bits(&pwd->gb, n);
			if (v >= pwd->nb_block_sizes)
				return -E_WMA_BLOCK_SIZE;
			pwd->block_len_bits = pwd->frame_len_bits - v;
		} else {
			/* update block lengths */
			pwd->prev_block_len_bits = pwd->block_len_bits;
			pwd->block_len_bits = pwd->next_block_len_bits;
		}
		v = get_bits(&pwd->gb, n);
		if (v >= pwd->nb_block_sizes)
			return -E_WMA_BLOCK_SIZE;
		pwd->next_block_len_bits = pwd->frame_len_bits - v;
	} else {
		/* fixed block len */
		pwd->next_block_len_bits = pwd->frame_len_bits;
		pwd->prev_block_len_bits = pwd->frame_len_bits;
		pwd->block_len_bits = pwd->frame_len_bits;
	}

	/* now check if the block length is coherent with the frame length */
	pwd->block_len = 1 << pwd->block_len_bits;
	if ((pwd->block_pos + pwd->block_len) > pwd->frame_len)
		return -E_INCOHERENT_BLOCK_LEN;

	if (pwd->ahi.channels == 2)
		ms_stereo = get_bit(&pwd->gb);
	v = 0;
	for (ch = 0; ch < pwd->ahi.channels; ch++) {
		int a = get_bit(&pwd->gb);
		pwd->channel_coded[ch] = a;
		v |= a;
	}

	bsize = pwd->frame_len_bits - pwd->block_len_bits;

	/* if no channel coded, no need to go further */
	/* XXX: fix potential framing problems */
	if (!v)
		goto next;

	/*
	 * Read total gain and extract corresponding number of bits for coef
	 * escape coding.
	 */
	total_gain = 1;
	for (;;) {
		int a = get_bits(&pwd->gb, 7);
		total_gain += a;
		if (a != 127)
			break;
	}

	coef_nb_bits = wma_total_gain_to_bits(total_gain);

	/* compute number of coefficients */
	n = pwd->coefs_end[bsize];
	for (ch = 0; ch < pwd->ahi.channels; ch++)
		nb_coefs[ch] = n;

	/* exponents can be reused in short blocks. */
	if ((pwd->block_len_bits == pwd->frame_len_bits) || get_bit(&pwd->gb)) {
		for (ch = 0; ch < pwd->ahi.channels; ch++) {
			if (pwd->channel_coded[ch]) {
				if (pwd->ahi.use_exp_vlc) {
					ret = decode_exp_vlc(pwd, ch);
					if (ret < 0)
						return ret;
				} else
					decode_exp_lsp(pwd, ch);
				pwd->exponents_bsize[ch] = bsize;
			}
		}
	}

	/* parse spectral coefficients : just RLE encoding */
	for (ch = 0; ch < pwd->ahi.channels; ch++) {
		struct vlc *coef_vlc;
		int level, run, tindex;
		int16_t *ptr, *eptr;
		const uint16_t *level_table, *run_table;

		if (!pwd->channel_coded[ch])
			continue;
		/*
		 * special VLC tables are used for ms stereo because there is
		 * potentially less energy there
		 */
		tindex = ch == 1 && ms_stereo;
		coef_vlc = &pwd->coef_vlc[tindex];
		run_table = pwd->run_table[tindex];
		level_table = pwd->level_table[tindex];
		/* XXX: optimize */
		ptr = &pwd->coefs1[ch][0];
		eptr = ptr + nb_coefs[ch];
		memset(ptr, 0, pwd->block_len * sizeof(int16_t));
		for (;;) {
			code = get_vlc(&pwd->gb, coef_vlc);
			if (code < 0)
				return code;
			if (code == 1) /* EOB */
				break;
			if (code == 0) { /* escape */
				level = get_bits(&pwd->gb, coef_nb_bits);
				/* reading block_len_bits would be better */
				run = get_bits(&pwd->gb, pwd->frame_len_bits);
			} else { /* normal code */
				run = run_table[code];
				level = level_table[code];
			}
			if (!get_bit(&pwd->gb))
				level = -level;
			ptr += run;
			if (ptr >= eptr) {
				PARA_ERROR_LOG("overflow in spectral RLE, ignoring\n");
				break;
			}
			*ptr++ = level;
			if (ptr >= eptr) /* EOB can be omitted */
				break;
		}
	}
	compute_mdct_coefficients(pwd, bsize, total_gain, nb_coefs);
	if (ms_stereo && pwd->channel_coded[1]) {
		float a, b;
		int i;
		/*
		 * Nominal case for ms stereo: we do it before mdct.
		 *
		 * No need to optimize this case because it should almost never
		 * happen.
		 */
		if (!pwd->channel_coded[0]) {
			PARA_NOTICE_LOG("rare ms-stereo\n");
			memset(pwd->coefs[0], 0, sizeof(float) * pwd->block_len);
			pwd->channel_coded[0] = 1;
		}
		for (i = 0; i < pwd->block_len; i++) {
			a = pwd->coefs[0][i];
			b = pwd->coefs[1][i];
			pwd->coefs[0][i] = a + b;
			pwd->coefs[1][i] = a - b;
		}
	}
next:
	for (ch = 0; ch < pwd->ahi.channels; ch++) {
		int n4, idx;

		n4 = pwd->block_len / 2;
		if (pwd->channel_coded[ch])
			imdct(pwd->mdct_ctx[bsize], pwd->output, pwd->coefs[ch]);
		else if (!(ms_stereo && ch == 1))
			memset(pwd->output, 0, sizeof(pwd->output));

		/* multiply by the window and add in the frame */
		idx = (pwd->frame_len / 2) + pwd->block_pos - n4;
		wma_window(pwd, &pwd->frame_out[ch][idx]);
	}

	/* update block number */
	pwd->block_pos += pwd->block_len;
	if (pwd->block_pos >= pwd->frame_len)
		return 1;
	else
		return 0;
}

/*
 * Clip a signed integer value into the -32768,32767 range.
 *
 * \param a The value to clip.
 *
 * \return The clipped value.
 */
static inline int16_t av_clip_int16(int a)
{
	if ((a + 32768) & ~65535)
		return (a >> 31) ^ 32767;
	else
		return a;
}

/* Decode a frame of frame_len samples. */
static int wma_decode_frame(struct private_wmadec_data *pwd, int16_t *samples)
{
	int ret, i, ch;
	int16_t *ptr;
	float *iptr;

	/* read each block */
	pwd->block_pos = 0;
	for (;;) {
		ret = wma_decode_block(pwd);
		if (ret < 0)
			return ret;
		if (ret)
			break;
	}

	/* convert frame to integer */
	for (ch = 0; ch < pwd->ahi.channels; ch++) {
		ptr = samples + ch;
		iptr = pwd->frame_out[ch];

		for (i = 0; i < pwd->frame_len; i++) {
			*ptr = av_clip_int16(lrintf(*iptr++));
			ptr += pwd->ahi.channels;
		}
		/* prepare for next block */
		memmove(&pwd->frame_out[ch][0], &pwd->frame_out[ch][pwd->frame_len],
			pwd->frame_len * sizeof(float));
	}
	return 0;
}

static int wma_decode_superframe(struct private_wmadec_data *pwd, void *out,
		int *out_size, const uint8_t *in)
{
	int ret, in_size = pwd->ahi.packet_size - WMA_FRAME_SKIP;
	int16_t *samples = out;

	init_get_bits(&pwd->gb, in, in_size);
	if (pwd->ahi.use_bit_reservoir) {
		int i, nb_frames, bit_offset, pos, len;
		uint8_t *q;

		/* read super frame header */
		skip_bits(&pwd->gb, 4);	/* super frame index */
		nb_frames = get_bits(&pwd->gb, 4) - 1;
		// PARA_DEBUG_LOG("have %d frames\n", nb_frames);
		ret = -E_WMA_OUTPUT_SPACE;
		if ((nb_frames + 1) * pwd->ahi.channels * pwd->frame_len
				* sizeof(int16_t) > *out_size)
			goto fail;

		bit_offset = get_bits(&pwd->gb, pwd->byte_offset_bits + 3);

		if (pwd->last_superframe_len > 0) {
			/* add bit_offset bits to last frame */
			ret = -E_WMA_BAD_SUPERFRAME;
			if ((pwd->last_superframe_len + ((bit_offset + 7) >> 3)) >
					MAX_CODED_SUPERFRAME_SIZE)
				goto fail;
			q = pwd->last_superframe + pwd->last_superframe_len;
			len = bit_offset;
			while (len > 7) {
				*q++ = get_bits(&pwd->gb, 8);
				len -= 8;
			}
			if (len > 0)
				*q++ = get_bits(&pwd->gb, len) << (8 - len);

			/* XXX: bit_offset bits into last frame */
			init_get_bits(&pwd->gb, pwd->last_superframe,
				MAX_CODED_SUPERFRAME_SIZE);
			/* skip unused bits */
			if (pwd->last_bitoffset > 0)
				skip_bits(&pwd->gb, pwd->last_bitoffset);
			/*
			 * This frame is stored in the last superframe and in
			 * the current one.
			 */
			ret = wma_decode_frame(pwd, samples);
			if (ret < 0)
				goto fail;
			samples += pwd->ahi.channels * pwd->frame_len;
		}

		/* read each frame starting from bit_offset */
		pos = bit_offset + 4 + 4 + pwd->byte_offset_bits + 3;
		init_get_bits(&pwd->gb, in + (pos >> 3),
			(MAX_CODED_SUPERFRAME_SIZE - (pos >> 3)));
		len = pos & 7;
		if (len > 0)
			skip_bits(&pwd->gb, len);

		pwd->reset_block_lengths = true;
		for (i = 0; i < nb_frames; i++) {
			ret = wma_decode_frame(pwd, samples);
			if (ret < 0)
				goto fail;
			samples += pwd->ahi.channels * pwd->frame_len;
		}

		/* we copy the end of the frame in the last frame buffer */
		pos = get_bits_count(&pwd->gb) +
			((bit_offset + 4 + 4 + pwd->byte_offset_bits + 3) & ~7);
		pwd->last_bitoffset = pos & 7;
		pos >>= 3;
		len = in_size - pos;
		ret = -E_WMA_BAD_SUPERFRAME;
		if (len > MAX_CODED_SUPERFRAME_SIZE || len < 0)
			goto fail;
		pwd->last_superframe_len = len;
		memcpy(pwd->last_superframe, in + pos, len);
	} else {
		PARA_DEBUG_LOG("not using bit reservoir\n");
		ret = -E_WMA_OUTPUT_SPACE;
		if (pwd->ahi.channels * pwd->frame_len * sizeof(int16_t) > *out_size)
			goto fail;
		/* single frame decode */
		ret = wma_decode_frame(pwd, samples);
		if (ret < 0)
			goto fail;
		samples += pwd->ahi.channels * pwd->frame_len;
	}
	PARA_DEBUG_LOG("frame_len: %d, block_len: %d, outbytes: %d, eaten: %d\n",
		pwd->frame_len, pwd->block_len,
		(int)((int8_t *)samples - (int8_t *)out), pwd->ahi.block_align);
	*out_size = (int8_t *)samples - (int8_t *)out;
	return pwd->ahi.block_align;
fail:
	/* reset the bit reservoir on errors */
	pwd->last_superframe_len = 0;
	return ret;
}

static void wmadec_close(struct filter_node *fn)
{
	struct private_wmadec_data *pwd = fn->private_data;
	int i;

	if (!pwd)
		return;
	for (i = 0; i < pwd->nb_block_sizes; i++)
		imdct_end(pwd->mdct_ctx[i]);
	if (pwd->ahi.use_exp_vlc)
		free_vlc(&pwd->exp_vlc);
	for (i = 0; i < 2; i++) {
		free_vlc(&pwd->coef_vlc[i]);
		free(pwd->run_table[i]);
		free(pwd->level_table[i]);
	}
	free(fn->private_data);
	fn->private_data = NULL;
}

static int wmadec_execute(const struct btr_node *btrn, const char *cmd,
		char **result)
{
	struct filter_node *fn = btr_context(btrn);
	struct private_wmadec_data *pwd = fn->private_data;

	return decoder_execute(cmd, pwd->ahi.sample_rate, pwd->ahi.channels,
		result);
}

#define WMA_OUTPUT_BUFFER_SIZE (128 * 1024)

static int wmadec_post_monitor(__a_unused struct sched *s, void *context)
{
	struct filter_node *fn = context;
	int ret, converted, out_size;
	struct private_wmadec_data *pwd = fn->private_data;
	struct btr_node *btrn = fn->btrn;
	size_t len;
	char *in, *out;

next_buffer:
	converted = 0;
	ret = btr_node_status(btrn, fn->min_iqs, BTR_NT_INTERNAL);
	if (ret < 0)
		goto err;
	if (ret == 0)
		return 0;
	btr_merge(btrn, fn->min_iqs);
	len = btr_next_buffer(btrn, &in);
	ret = -E_EOF;
	if (len < fn->min_iqs)
		goto err;
	if (!pwd) {
		ret = wma_decode_init(in, len, &pwd);
		if (ret < 0)
			goto err;
		if (ret == 0) {
			fn->min_iqs += 4096;
			goto next_buffer;
		}
		fn->min_iqs = 2 * pwd->ahi.packet_size;
		fn->private_data = pwd;
		converted = pwd->ahi.header_len;
		goto success;
	}
	fn->min_iqs = pwd->ahi.packet_size;
	if (fn->min_iqs > len)
		goto success;
	out_size = WMA_OUTPUT_BUFFER_SIZE;
	out = alloc(out_size);
	ret = wma_decode_superframe(pwd, out, &out_size,
		(uint8_t *)in + WMA_FRAME_SKIP);
	if (ret < 0) {
		free(out);
		goto err;
	}
	if (out_size > 0) {
		out = para_realloc(out, out_size);
		btr_add_output(out, out_size, btrn);
	} else
		free(out);
	converted += pwd->ahi.packet_size;
success:
	btr_consume(btrn, converted);
	return 0;
err:
	assert(ret < 0);
	btr_remove_node(&fn->btrn);
	return ret;
}

static void wmadec_open(struct filter_node *fn)
{
	fn->private_data = NULL;
	fn->min_iqs = 4096;
}

/** \cond doxygen_ignore */
const struct filter lsg_filter_cmd_com_wmadec_user_data = {
	.open = wmadec_open,
	.close = wmadec_close,
	.execute = wmadec_execute,
	.pre_monitor = generic_filter_pre_monitor,
	.post_monitor = wmadec_post_monitor,
};
/** \endcond */