Merge commit 'a1a143adb0fd11c474221431417cff25db7d920f'
[ffmpeg.git] / libavcodec / aacenc_tns.c
index cf75b8a..2ffe1f8 100644 (file)
  * @author Rostislav Pehlivanov ( atomnuker gmail com )
  */
 
+#include "libavutil/libm.h"
 #include "aacenc.h"
 #include "aacenc_tns.h"
 #include "aactab.h"
 #include "aacenc_utils.h"
 #include "aacenc_quantization.h"
 
-static inline int compress_coef(int *coefs, int num)
+/* 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_IS8 4
+
+/* We really need the bits we save here elsewhere */
+#define TNS_ENABLE_COEF_COMPRESSION
+
+/* TNS will only be used if the LPC gain is within these margins */
+#define TNS_GAIN_THRESHOLD_LOW      1.4f
+#define TNS_GAIN_THRESHOLD_HIGH     1.16f*TNS_GAIN_THRESHOLD_LOW
+
+static inline int compress_coeffs(int *coef, int order, int c_bits)
 {
-    int i, c = 0;
-    for (i = 0; i < num; i++)
-        c += coefs[i] < 4 || coefs[i] > 11;
-    return c == num;
+    int i;
+    const int low_idx   = c_bits ?  4 : 2;
+    const int shift_val = c_bits ?  8 : 4;
+    const int high_idx  = c_bits ? 11 : 5;
+#ifndef TNS_ENABLE_COEF_COMPRESSION
+    return 0;
+#endif /* TNS_ENABLE_COEF_COMPRESSION */
+    for (i = 0; i < order; i++)
+        if (coef[i] >= low_idx && coef[i] <= high_idx)
+            return 0;
+    for (i = 0; i < order; i++)
+        coef[i] -= (coef[i] > high_idx) ? shift_val : 0;
+    return 1;
 }
 
 /**
  * Encode TNS data.
- * Coefficient compression saves a single bit per coefficient.
+ * Coefficient compression is simply not lossless as it should be
+ * on any decoder tested and as such is not active.
  */
 void ff_aac_encode_tns_info(AACEncContext *s, SingleChannelElement *sce)
 {
-    int i, w, filt, coef_len, coef_compress;
+    TemporalNoiseShaping *tns = &sce->tns;
+    int i, w, filt, coef_compress = 0, coef_len;
     const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
+    const int c_bits = is8 ? TNS_Q_BITS_IS8 == 4 : TNS_Q_BITS == 4;
 
     if (!sce->tns.present)
         return;
 
     for (i = 0; i < sce->ics.num_windows; i++) {
         put_bits(&s->pb, 2 - is8, sce->tns.n_filt[i]);
-        if (sce->tns.n_filt[i]) {
-            put_bits(&s->pb, 1, 1);
-            for (filt = 0; filt < sce->tns.n_filt[i]; filt++) {
-                put_bits(&s->pb, 6 - 2 * is8, sce->tns.length[i][filt]);
-                put_bits(&s->pb, 5 - 2 * is8, sce->tns.order[i][filt]);
-                if (sce->tns.order[i][filt]) {
-                    coef_compress = compress_coef(sce->tns.coef_idx[i][filt],
-                                                  sce->tns.order[i][filt]);
-                    put_bits(&s->pb, 1, !!sce->tns.direction[i][filt]);
-                    put_bits(&s->pb, 1, !!coef_compress);
-                    coef_len = 4 - coef_compress;
-                    for (w = 0; w < sce->tns.order[i][filt]; w++)
-                        put_bits(&s->pb, coef_len, sce->tns.coef_idx[i][filt][w]);
-                }
-            }
-        }
-    }
-}
-
-static void process_tns_coeffs(TemporalNoiseShaping *tns, double *coef_raw,
-                               int *order_p, int w, int filt)
-{
-    int i, j, order = *order_p;
-    int *idx = tns->coef_idx[w][filt];
-    float *lpc = tns->coef[w][filt];
-    float temp[TNS_MAX_ORDER] = {0.0f}, out[TNS_MAX_ORDER] = {0.0f};
-
-    if (!order)
-        return;
-
-    /* Not what the specs say, but it's better */
-    for (i = 0; i < order; i++) {
-        idx[i] = quant_array_idx(coef_raw[i], tns_tmp2_map_0_4, 16);
-        lpc[i] = tns_tmp2_map_0_4[idx[i]];
-    }
-
-    /* Trim any coeff less than 0.1f from the end */
-    for (i = order-1; i > -1; i--) {
-        lpc[i] = (fabs(lpc[i]) > 0.1f) ? lpc[i] : 0.0f;
-        if (lpc[i] != 0.0 ) {
-            order = i;
-            break;
-        }
-    }
-    order = av_clip(order, 0, TNS_MAX_ORDER - 1);
-    *order_p = order;
-    if (!order)
-        return;
-
-    /* Step up procedure, convert to LPC coeffs */
-    out[0] = 1.0f;
-    for (i = 1; i <= order; i++) {
-        for (j = 1; j < i; j++) {
-            temp[j] = out[j] + lpc[i]*out[i-j];
-        }
-        for (j = 1; j <= i; j++) {
-            out[j] = temp[j];
+        if (!tns->n_filt[i])
+            continue;
+        put_bits(&s->pb, 1, c_bits);
+        for (filt = 0; filt < tns->n_filt[i]; filt++) {
+            put_bits(&s->pb, 6 - 2 * is8, tns->length[i][filt]);
+            put_bits(&s->pb, 5 - 2 * is8, tns->order[i][filt]);
+            if (!tns->order[i][filt])
+                continue;
+            put_bits(&s->pb, 1, tns->direction[i][filt]);
+            coef_compress = compress_coeffs(tns->coef_idx[i][filt],
+                                            tns->order[i][filt], c_bits);
+            put_bits(&s->pb, 1, coef_compress);
+            coef_len = c_bits + 3 - coef_compress;
+            for (w = 0; w < tns->order[i][filt]; w++)
+                put_bits(&s->pb, coef_len, tns->coef_idx[i][filt][w]);
         }
-        out[i] = lpc[i-1];
     }
-    memcpy(lpc, out, TNS_MAX_ORDER*sizeof(float));
 }
 
 /* Apply TNS filter */
-void ff_aac_apply_tns(SingleChannelElement *sce)
+void ff_aac_apply_tns(AACEncContext *s, SingleChannelElement *sce)
 {
-    const int mmm = FFMIN(sce->ics.tns_max_bands, sce->ics.max_sfb);
-    float *coef = sce->pcoeffs;
     TemporalNoiseShaping *tns = &sce->tns;
-    int w, filt, m, i;
-    int bottom, top, order, start, end, size, inc;
-    float *lpc, tmp[TNS_MAX_ORDER+1];
+    IndividualChannelStream *ics = &sce->ics;
+    int w, filt, m, i, top, order, bottom, start, end, size, inc;
+    const int mmm = FFMIN(ics->tns_max_bands, ics->max_sfb);
+    float lpc[TNS_MAX_ORDER];
 
-    for (w = 0; w < sce->ics.num_windows; w++) {
-        bottom = sce->ics.num_swb;
+    for (w = 0; w < ics->num_windows; w++) {
+        bottom = ics->num_swb;
         for (filt = 0; filt < tns->n_filt[w]; filt++) {
             top    = bottom;
             bottom = FFMAX(0, top - tns->length[w][filt]);
             order  = tns->order[w][filt];
-            lpc    = tns->coef[w][filt];
-            if (!order)
+            if (order == 0)
                 continue;
 
-            start = sce->ics.swb_offset[FFMIN(bottom, mmm)];
-            end   = sce->ics.swb_offset[FFMIN(   top, mmm)];
+            // tns_decode_coef
+            compute_lpc_coefs(tns->coef[w][filt], order, lpc, 0, 0, 0);
+
+            start = ics->swb_offset[FFMIN(bottom, mmm)];
+            end   = ics->swb_offset[FFMIN(   top, mmm)];
             if ((size = end - start) <= 0)
                 continue;
             if (tns->direction[w][filt]) {
@@ -148,83 +131,85 @@ void ff_aac_apply_tns(SingleChannelElement *sce)
             }
             start += w * 128;
 
-            if (!sce->ics.ltp.present) {
-                // ar filter
-                for (m = 0; m < size; m++, start += inc)
-                    for (i = 1; i <= FFMIN(m, order); i++)
-                        coef[start] += coef[start - i * inc]*lpc[i - 1];
-            } else {
-                // ma filter
-                for (m = 0; m < size; m++, start += inc) {
-                    tmp[0] = coef[start];
-                    for (i = 1; i <= FFMIN(m, order); i++)
-                        coef[start] += tmp[i]*lpc[i - 1];
-                    for (i = order; i > 0; i--)
-                        tmp[i] = tmp[i - 1];
+            /* AR filter */
+            for (m = 0; m < size; m++, start += inc) {
+                for (i = 1; i <= FFMIN(m, order); i++) {
+                    sce->coeffs[start] += lpc[i-1]*sce->pcoeffs[start - i*inc];
                 }
             }
         }
     }
 }
 
+/*
+ * 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(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, g, w2, prev_end_sfb = 0, count = 0;
+    int w, g, count = 0;
+    double gain, coefs[MAX_LPC_ORDER];
+    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 tns_max_order = is8 ? 7 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER;
+    const int c_bits = is8 ? TNS_Q_BITS_IS8 == 4 : TNS_Q_BITS == 4;
+    const int sfb_start = av_clip(tns_min_sfb[is8][s->samplerate_index], 0, mmm);
+    const int sfb_end   = av_clip(sce->ics.num_swb, 0, mmm);
+    const int order = is8 ? 7 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER;
+    const int slant = sce->ics.window_sequence[0] == LONG_STOP_SEQUENCE  ? 1 :
+                      sce->ics.window_sequence[0] == LONG_START_SEQUENCE ? 0 : 2;
+    const int sfb_len = sfb_end - sfb_start;
+    const int coef_len = sce->ics.swb_offset[sfb_end] - sce->ics.swb_offset[sfb_start];
+
+    if (coef_len <= 0 || sfb_len <= 0) {
+        sce->tns.present = 0;
+        return;
+    }
 
     for (w = 0; w < sce->ics.num_windows; w++) {
-        int order = 0, filters = 1;
-        int sfb_start = 0, sfb_len = 0;
-        int coef_start = 0, coef_len = 0;
-        float energy = 0.0f, threshold = 0.0f;
-        double coefs[MAX_LPC_ORDER][MAX_LPC_ORDER] = {{0}};
-        for (g = 0;  g < sce->ics.num_swb; g++) {
-            if (!sfb_start && w*16+g > TNS_LOW_LIMIT && w*16+g > prev_end_sfb) {
-                sfb_start = w*16+g;
-                coef_start =  sce->ics.swb_offset[sfb_start];
-            }
-            if (sfb_start) {
-                for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
-                    FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
-                    if (!sfb_len && band->energy < band->threshold*1.3f) {
-                        sfb_len = (w+w2)*16+g - sfb_start;
-                        prev_end_sfb = sfb_start + sfb_len;
-                        coef_len = sce->ics.swb_offset[sfb_start + sfb_len] - coef_start;
-                        break;
-                    }
-                    energy += band->energy;
-                    threshold += band->threshold;
-                }
-                if (!sfb_len) {
-                    sfb_len = (w+1)*16+g - sfb_start - 1;
-                    coef_len = sce->ics.swb_offset[sfb_start + sfb_len] - coef_start;
-                }
-            }
+        float en[2] = {0.0f, 0.0f};
+        int oc_start = 0, os_start = 0;
+        int coef_start = sce->ics.swb_offset[sfb_start];
+
+        for (g = sfb_start; g < sce->ics.num_swb && g <= sfb_end; g++) {
+            FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[w*16+g];
+            if (g > sfb_start + (sfb_len/2))
+                en[1] += band->energy;
+            else
+                en[0] += band->energy;
         }
 
-        if (sfb_len <= 0 || coef_len <= 0)
+        /* LPC */
+        gain = ff_lpc_calc_ref_coefs_f(&s->lpc, &sce->coeffs[w*128 + coef_start],
+                                       coef_len, order, coefs);
+
+        if (!order || !isfinite(gain) || gain < TNS_GAIN_THRESHOLD_LOW || gain > TNS_GAIN_THRESHOLD_HIGH)
             continue;
-        if (coef_start + coef_len >= 1024)
-            coef_len = 1024 - coef_start;
 
-        /* LPC */
-        order = ff_lpc_calc_levinson(&s->lpc, &sce->coeffs[coef_start], coef_len,
-                                     coefs, 0, tns_max_order, ORDER_METHOD_LOG);
-
-        if (energy > threshold) {
-            int direction = 0;
-            tns->n_filt[w] = filters++;
-            for (g = 0; g < tns->n_filt[w]; g++) {
-                process_tns_coeffs(tns, coefs[order], &order, w, g);
-                tns->order[w][g]     = order;
-                tns->length[w][g]    = sfb_len;
-                tns->direction[w][g] = direction;
-            }
-            count++;
+        tns->n_filt[w] = is8 ? 1 : order != TNS_MAX_ORDER ? 2 : 3;
+        for (g = 0; g < tns->n_filt[w]; g++) {
+            tns->direction[w][g] = slant != 2 ? slant : en[g] < en[!g];
+            tns->order[w][g] = g < tns->n_filt[w] ? order/tns->n_filt[w] : order - oc_start;
+            tns->length[w][g] = g < tns->n_filt[w] ? sfb_len/tns->n_filt[w] : sfb_len - os_start;
+            quantize_coefs(&coefs[oc_start], tns->coef_idx[w][g], tns->coef[w][g],
+                            tns->order[w][g], c_bits);
+            oc_start += tns->order[w][g];
+            os_start += tns->length[w][g];
         }
+        count++;
     }
-
     sce->tns.present = !!count;
 }