[FFmpeg-devel] AAC decoder round 7
Michael Niedermayer
michaelni
Tue Aug 12 20:05:44 CEST 2008
On Tue, Aug 12, 2008 at 06:45:35PM +0100, Robert Swain wrote:
> 2008/8/12 Michael Niedermayer <michaelni at gmx.at>:
[...]
> >> >> /**
> >> >> + * Decode spectral data; reference: table 4.50.
> >> >> + *
> >> >> + * @param band_type array of the used band type
> >> >> + * @param icoef array of quantized spectral data
> >> >> + *
> >> >> + * @return Returns error status. 0 - OK, !0 - error
> >> >> + */
> >> >> +static int decode_spectrum(AACContext * ac, int icoef[1024], GetBitContext * gb,
> >> >> + const IndividualChannelStream * ics, enum BandType band_type[120]) {
> >> >> + int i, k, g, idx = 0;
> >> >> + const uint16_t * offsets = ics->swb_offset;
> >> >> +
> >> >> + for (g = 0; g < ics->num_window_groups; g++) {
> >> >> + for (i = 0; i < ics->max_sfb; i++, idx++) {
> >> >> + const int cur_band_type = band_type[idx];
> >> >> + const int dim = cur_band_type >= FIRST_PAIR_BT ? 2 : 4;
> >> >> + const int is_cb_unsigned = IS_CODEBOOK_UNSIGNED(cur_band_type);
> >> >> + int group;
> >> >> + if (cur_band_type == ZERO_BT) {
> >> >> + for (group = 0; group < ics->group_len[g]; group++) {
> >> >> + memset(icoef + group * 128 + offsets[i], 0, (offsets[i+1] - offsets[i])*sizeof(int));
> >> >> + }
> >> >> + }else if (cur_band_type != NOISE_BT && cur_band_type != INTENSITY_BT2 && cur_band_type != INTENSITY_BT) {
> >> >> + for (group = 0; group < ics->group_len[g]; group++) {
> >> >> + for (k = offsets[i]; k < offsets[i+1]; k += dim) {
> >> >> + const int index = get_vlc2(gb, vlc_spectral[cur_band_type - 1].table, 6, 3);
> >> >> + const int coef_idx = (group << 7) + k;
> >> >> + const int8_t *vq_ptr;
> >> >> + int j;
> >> >> + if(index >= ff_aac_spectral_sizes[cur_band_type - 1]) {
> >> >> + av_log(ac->avccontext, AV_LOG_ERROR,
> >> >> + "Read beyond end of ff_aac_codebook_vectors[%d][]. index %d >= %d\n",
> >> >> + cur_band_type - 1, index, ff_aac_spectral_sizes[cur_band_type - 1]);
> >> >> + return -1;
> >> >> + }
> >> >> + vq_ptr = &ff_aac_codebook_vectors[cur_band_type - 1][index * dim];
> >> >> + if (is_cb_unsigned) {
> >> >> + for (j = 0; j < dim; j++)
> >> >> + if (vq_ptr[j])
> >> >> + icoef[coef_idx + j] = 1 - 2*get_bits1(gb);
> >> >> + }else {
> >> >> + for (j = 0; j < dim; j++)
> >> >> + icoef[coef_idx + j] = 1;
> >> >> + }
> >> >> + if (cur_band_type == ESC_BT) {
> >> >> + for (j = 0; j < 2; j++) {
> >> >> + if (vq_ptr[j] == 16) {
> >> >> + int n = 4;
> >> >> + /* The total length of escape_sequence must be < 22 bits according
> >> >> + to the specification (i.e. max is 11111111110xxxxxxxxxx). */
> >> >> + while (get_bits1(gb) && n < 15) n++;
> >> >> + if(n == 15) {
> >> >> + av_log(ac->avccontext, AV_LOG_ERROR, "error in spectral data, ESC overflow\n");
> >> >> + return -1;
> >> >> + }
> >> >> + icoef[coef_idx + j] *= (1<<n) + get_bits(gb, n);
> >> >> + }else
> >> >> + icoef[coef_idx + j] *= vq_ptr[j];
> >> >> + }
> >> >> + }else
> >> >> + for (j = 0; j < dim; j++)
> >> >> + icoef[coef_idx + j] *= vq_ptr[j];
> >> >> + }
> >> >> + }
> >> >> + }
> >> >> + }
> >> >> + icoef += ics->group_len[g]<<7;
> >> >> + }
> >> >> + return 0;
> >> >> +}
> >> >> +
> >> >> +/**
> >> >> * Add pulses with particular amplitudes to the quantized spectral data; reference: 4.6.3.3.
> >> >> *
> >> >> * @param pulse pointer to pulse data struct
> >> >> @@ -538,6 +780,46 @@
> >> >> }
> >> >>
> >> >> /**
> >> >> + * Dequantize and scale spectral data; reference: 4.6.3.3.
> >> >> + *
> >> >> + * @param icoef array of quantized spectral data
> >> >> + * @param band_type array of the used band type
> >> >> + * @param sf array of scalefactors or intensity stereo positions
> >> >> + * @param coef array of dequantized, scaled spectral data
> >> >> + */
> >> >> +static void dequant(AACContext * ac, float coef[1024], const int icoef[1024], float sf[120],
> >> >> + const IndividualChannelStream * ics, enum BandType band_type[120]) {
> >> >> + const uint16_t * offsets = ics->swb_offset;
> >> >> + const int c = 1024/ics->num_windows;
> >> >> + int g, i, group, k, idx = 0;
> >> >> +
> >> >> + for (g = 0; g < ics->num_windows; g++)
> >> >> + memset(coef + g * 128 + offsets[ics->max_sfb], 0, sizeof(float)*(c - offsets[ics->max_sfb]));
> >> >> +
> >> >> + for (g = 0; g < ics->num_window_groups; g++) {
> >> >> + for (i = 0; i < ics->max_sfb; i++, idx++) {
> >> >> + if (band_type[idx] == NOISE_BT) {
> >> >> + const float scale = sf[idx] / ((offsets[i+1] - offsets[i]) * PNS_MEAN_ENERGY);
> >> >> + for (group = 0; group < ics->group_len[g]; group++) {
> >> >> + for (k = offsets[i]; k < offsets[i+1]; k++) {
> >> >> + ac->random_state = lcg_random(ac->random_state);
> >> >> + coef[group*128+k] = ac->random_state * scale;
> >> >> + }
> >> >> + }
> >> >> + } else if (band_type[idx] != INTENSITY_BT && band_type[idx] != INTENSITY_BT2) {
> >> >> + for (group = 0; group < ics->group_len[g]; group++) {
> >> >> + for (k = offsets[i]; k < offsets[i+1]; k++) {
> >> >> + coef[group*128+k] = ivquant(icoef[group*128+k]) * sf[idx];
> >> >> + }
> >> >> + }
> >> >> + }
> >> >> + }
> >> >> + coef += ics->group_len[g]*128;
> >> >> + icoef += ics->group_len[g]*128;
> >> >> + }
> >> >> +}
> >> >> +
> >> >> +/**
> >> >
> >> > dequant and decode_spectrum() can be merged, especially the VQ tables can
> >> > already include ivquant().
> >> > and yes i know that this means add_pulses will be more tricky
> >>
> >> See attached (20080812-1422-merge_decspec_dequant.diff). It's not that
> >> nice to look at and svn diff makes it worse. Suggestions to improve it
> >> very welcome. Making add_pulse() recursive seemed like a good solution
> >> to avoid refactoring the main loops and conditions but maybe you have
> >> a better idea.
> >
> > yes i do :)
> > for(each pulse){
> > float c= coeff[ pos[i] ];
> > float ic= c / sqrtf(sqrtf(fabsf(c))) + amp[i];
> > coeff[ pos[i] ]= cbrtf(fabsf(ic)) * ic;
> > }
> >
> > This idea is of course based on the assumtation that there are few pulses
> > compared to the number of coefficients, which as far as i understand is true
> > as there are max 4 pulses IIRC
>
> You are correct. Good idea, though I don't know why I didn't think of
> doing it like that. :) See attached.
patch ok, but you still need to change the codebooks/vq to include ivquant()
[...]
--
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
If a bugfix only changes things apparently unrelated to the bug with no
further explanation, that is a good sign that the bugfix is wrong.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20080812/a44610f3/attachment.pgp>
More information about the ffmpeg-devel
mailing list