diff options
author | Rostislav Pehlivanov <atomnuker@gmail.com> | 2015-09-12 13:04:27 +0100 |
---|---|---|
committer | Rostislav Pehlivanov <atomnuker@gmail.com> | 2015-09-12 13:04:27 +0100 |
commit | a83a8d70681a7093ff4bd8232b8f75fed3d6af48 (patch) | |
tree | a28f3bc5df2467018713f86169a622150fdce7a9 | |
parent | 33be39efe31326dc787efc54f5077e4c4b97896c (diff) | |
download | ffmpeg-a83a8d70681a7093ff4bd8232b8f75fed3d6af48.tar.gz |
aacenc_tns: redo coefficient quantization and decision making
This finally (and again) gets rid of basically everything the
specifications say about how TNS should be done. The main
problem used to be that a single filter was used for all
coefficients which despite being explicitly recommended by
the specifications usually sounds wrong, therefore it's
a corner case in the current TNS implementation.
This commit also changes the coefficient bit size, as apparently
it's better to use lower precision in case the windows are eight
short. This is apparently what fdk_aac uses, looking at the bit
stream and makes sense. Also the order when 8 SHORT windows happen
is important as 7 was too much and according to PSNR was worse
while 5 is just about correct.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
-rw-r--r-- | libavcodec/aacenc_tns.c | 60 | ||||
-rw-r--r-- | libavcodec/aacenc_tns.h | 3 |
2 files changed, 37 insertions, 26 deletions
diff --git a/libavcodec/aacenc_tns.c b/libavcodec/aacenc_tns.c index 637b81385e..da353d65c5 100644 --- a/libavcodec/aacenc_tns.c +++ b/libavcodec/aacenc_tns.c @@ -67,20 +67,6 @@ void ff_aac_encode_tns_info(AACEncContext *s, SingleChannelElement *sce) } } -static inline void quantize_coefs(double *coef, int *idx, float *lpc, int order) -{ - int i; - uint8_t u_coef; - const float *quant_arr = tns_tmp2_map[TNS_Q_BITS == 4]; - const double iqfac_p = ((1 << (TNS_Q_BITS-1)) - 0.5)/(M_PI/2.0); - const double iqfac_m = ((1 << (TNS_Q_BITS-1)) + 0.5)/(M_PI/2.0); - for (i = 0; i < order; i++) { - idx[i] = ceilf(asin(coef[i])*((coef[i] >= 0) ? iqfac_p : iqfac_m)); - u_coef = (idx[i])&(~(~0<<TNS_Q_BITS)); - lpc[i] = quant_arr[u_coef]; - } -} - /* Apply TNS filter */ void ff_aac_apply_tns(AACEncContext *s, SingleChannelElement *sce) { @@ -122,22 +108,41 @@ void ff_aac_apply_tns(AACEncContext *s, SingleChannelElement *sce) } } +/* + * c_bits - 1 if 4 bit coefficients, 0 if 3 bit coefficients + */ +static inline void quantize_coefs(double *coef, int *idx, float *lpc, int order, + int c_bits) +{ + int i; + const float *quant_arr = tns_tmp2_map[c_bits]; + for (i = 0; i < order; i++) { + idx[i] = quant_array_idx((float)coef[i], quant_arr, c_bits ? 16 : 8); + lpc[i] = quant_arr[idx[i]]; + } +} + +/* + * 3 bits per coefficient with 8 short windows + */ void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce) { TemporalNoiseShaping *tns = &sce->tns; int w, w2, g, count = 0; const int mmm = FFMIN(sce->ics.tns_max_bands, sce->ics.max_sfb); const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE; - const int order = is8 ? 7 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER; + const int c_bits = is8 ? TNS_Q_BITS_SHORT == 4 : TNS_Q_BITS == 4; int sfb_start = av_clip(tns_min_sfb[is8][s->samplerate_index], 0, mmm); int sfb_end = av_clip(sce->ics.num_swb, 0, mmm); for (w = 0; w < sce->ics.num_windows; w++) { - float e_ratio = 0.0f, threshold = 0.0f, spread = 0.0f, en[2] = {0.0, 0.0f}; - double gain = 0.0f, coefs[MAX_LPC_ORDER] = {0}; + int use_tns; + int order = is8 ? 5 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER; int coef_start = w*sce->ics.num_swb + sce->ics.swb_offset[sfb_start]; int coef_len = sce->ics.swb_offset[sfb_end] - sce->ics.swb_offset[sfb_start]; + float e_ratio = 0.0f, threshold = 0.0f, spread = 0.0f, en[2] = {0.0, 0.0f}; + double gain = 0.0f, coefs[MAX_LPC_ORDER] = {0}; for (g = 0; g < sce->ics.num_swb; g++) { if (w*16+g < sfb_start || w*16+g > sfb_end) @@ -149,22 +154,26 @@ void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce) else en[0] += band->energy; threshold += band->threshold; - spread += band->spread; + spread += band->spread; } } if (coef_len <= 0 || (sfb_end - sfb_start) <= 0) continue; - else - e_ratio = en[0]/en[1]; /* LPC */ gain = ff_lpc_calc_ref_coefs_f(&s->lpc, &sce->coeffs[coef_start], coef_len, order, coefs); - if (gain > TNS_GAIN_THRESHOLD_LOW && gain < TNS_GAIN_THRESHOLD_HIGH && - (en[0]+en[1]) > TNS_GAIN_THRESHOLD_LOW*threshold && - spread < TNS_SPREAD_THRESHOLD && order) { + if (!order || gain < TNS_GAIN_THRESHOLD_LOW || gain > TNS_GAIN_THRESHOLD_HIGH) + use_tns = 0; + else if ((en[0]+en[1]) < TNS_GAIN_THRESHOLD_LOW*threshold || spread < TNS_SPREAD_THRESHOLD) + use_tns = 0; + else + use_tns = 1; + + if (use_tns) { + e_ratio = en[0]/en[1]; if (is8 || order < 2 || (e_ratio > TNS_E_RATIO_LOW && e_ratio < TNS_E_RATIO_HIGH)) { tns->n_filt[w] = 1; for (g = 0; g < tns->n_filt[w]; g++) { @@ -172,7 +181,7 @@ void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce) tns->direction[w][g] = en[0] < en[1]; tns->order[w][g] = order; quantize_coefs(coefs, tns->coef_idx[w][g], tns->coef[w][g], - order); + order, c_bits); } } else { /* 2 filters due to energy disbalance */ tns->n_filt[w] = 2; @@ -183,12 +192,11 @@ void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce) (sfb_end - sfb_start) - tns->length[w][g-1]; quantize_coefs(&coefs[!g ? 0 : order - tns->order[w][g-1]], tns->coef_idx[w][g], tns->coef[w][g], - tns->order[w][g]); + tns->order[w][g], c_bits); } } count++; } } - sce->tns.present = !!count; } diff --git a/libavcodec/aacenc_tns.h b/libavcodec/aacenc_tns.h index f0e75569ae..46e4794965 100644 --- a/libavcodec/aacenc_tns.h +++ b/libavcodec/aacenc_tns.h @@ -33,6 +33,9 @@ /* Could be set to 3 to save an additional bit at the cost of little quality */ #define TNS_Q_BITS 4 +/* Coefficient resolution in short windows */ +#define TNS_Q_BITS_SHORT 3 + /* TNS will only be used if the LPC gain is within these margins */ #define TNS_GAIN_THRESHOLD_LOW 1.395f #define TNS_GAIN_THRESHOLD_HIGH 11.19f |