VP8: idct_mb optimizations
[ffmpeg.git] / libavcodec / aaccoder.c
index f6b7c2a..8063fb6 100644 (file)
@@ -517,12 +517,12 @@ static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
 
 /** Return the minimum scalefactor where the quantized coef does not clip. */
 static av_always_inline uint8_t coef2minsf(float coef) {
-    return av_clip_uint8(log2(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
+    return av_clip_uint8(log2f(coef)*4 - 69 + SCALE_ONE_POS - SCALE_DIV_512);
 }
 
 /** Return the maximum scalefactor where the quantized coef is not zero. */
 static av_always_inline uint8_t coef2maxsf(float coef) {
-    return av_clip_uint8(log2(coef)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
+    return av_clip_uint8(log2f(coef)*4 +  6 + SCALE_ONE_POS - SCALE_DIV_512);
 }
 
 typedef struct TrellisPath {
@@ -572,7 +572,7 @@ static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
         int q0low  = q0;
         int q1high = q1;
         //minimum scalefactor index is when maximum nonzero coefficient after quantizing is not clipped
-        int qnrg = av_clip_uint8(log2(sqrt(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
+        int qnrg = av_clip_uint8(log2f(sqrtf(qnrgf/qcnt))*4 - 31 + SCALE_ONE_POS - SCALE_DIV_512);
         q1 = qnrg + 30;
         q0 = qnrg - 30;
     //av_log(NULL, AV_LOG_ERROR, "q0 %d, q1 %d\n", q0, q1);
@@ -697,6 +697,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
     int start = 0, i, w, w2, g;
     int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate / avctx->channels;
     float dists[128], uplims[128];
+    float maxvals[128];
     int fflag, minscaler;
     int its  = 0;
     int allz = 0;
@@ -731,13 +732,23 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
                 continue;
             }
-            sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2(uplims[w*16+g]/minthr)*4,59);
+            sce->sf_idx[w*16+g] = SCALE_ONE_POS + FFMIN(log2f(uplims[w*16+g]/minthr)*4,59);
         }
     }
 
     if (!allz)
         return;
     abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+
+    for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
+        start = w*128;
+        for (g = 0;  g < sce->ics.num_swb; g++) {
+            const float *scaled = s->scoefs + start;
+            maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
+            start += sce->ics.swb_sizes[g];
+        }
+    }
+
     //perform two-loop search
     //outer loop - improve quality
     do {
@@ -763,7 +774,7 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                         continue;
                     }
                     minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
-                    cb = find_min_book(find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled), sce->sf_idx[w*16+g]);
+                    cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
                     for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
                         int b;
                         dist += quantize_band_cost(s, coefs + w2*128,
@@ -795,27 +806,26 @@ static void search_for_quantizers_twoloop(AVCodecContext *avctx,
                         sce->sf_idx[i] -= qstep;
             }
             qstep >>= 1;
-            if (!qstep && tbits > destbits*1.02)
+            if (!qstep && tbits > destbits*1.02 && sce->sf_idx[0] < 217)
                 qstep = 1;
-            if (sce->sf_idx[0] >= 217)
-                break;
         } while (qstep);
 
         fflag = 0;
         minscaler = av_clip(minscaler, 60, 255 - SCALE_MAX_DIFF);
         for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
-            start = w*128;
             for (g = 0; g < sce->ics.num_swb; g++) {
                 int prevsc = sce->sf_idx[w*16+g];
-                const float *scaled = s->scoefs + start;
-                if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60)
+                if (dists[w*16+g] > uplims[w*16+g] && sce->sf_idx[w*16+g] > 60) {
+                    if (find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1))
                     sce->sf_idx[w*16+g]--;
+                    else //Try to make sure there is some energy in every band
+                        sce->sf_idx[w*16+g]-=2;
+                }
                 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF);
                 sce->sf_idx[w*16+g] = FFMIN(sce->sf_idx[w*16+g], 219);
                 if (sce->sf_idx[w*16+g] != prevsc)
                     fflag = 1;
-                sce->band_type[w*16+g] = find_min_book(find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled), sce->sf_idx[w*16+g]);
-                start += sce->ics.swb_sizes[g];
+                sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
             }
         }
         its++;
@@ -925,7 +935,7 @@ static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
                 continue;
             }
             sce->zeroes[w*16+g] = 0;
-            scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2(1/maxq[w*16+g])*16/3, 60, 218);
+            scf  = prev_scf = av_clip(SCALE_ONE_POS - SCALE_DIV_512 - log2f(1/maxq[w*16+g])*16/3, 60, 218);
             step = 16;
             for (;;) {
                 float dist = 0.0f;
@@ -954,7 +964,7 @@ static void search_for_quantizers_faac(AVCodecContext *avctx, AACEncContext *s,
                 if (curdiff <= 1.0f)
                     step = 0;
                 else
-                    step = log2(curdiff);
+                    step = log2f(curdiff);
                 if (dist > uplim[w*16+g])
                     step = -step;
                 scf += step;
@@ -1007,7 +1017,7 @@ static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
                     sce->sf_idx[(w+w2)*16+g] = 218;
                     sce->zeroes[(w+w2)*16+g] = 1;
                 } else {
-                    sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2(band->threshold), 80, 218);
+                    sce->sf_idx[(w+w2)*16+g] = av_clip(SCALE_ONE_POS - SCALE_DIV_512 + log2f(band->threshold), 80, 218);
                     sce->zeroes[(w+w2)*16+g] = 0;
                 }
                 minq = FFMIN(minq, sce->sf_idx[(w+w2)*16+g]);