H.264: Add optimizations to predict x86 assembly.

[ffmpeg] / libavcodec / qdm2.c

diff --git a/libavcodec/qdm2.c b/libavcodec/qdm2.c
index f74cfd92581bcb5e0969df2309cdbcc9a6217eb1..2a45495c7cdc7200a75b52eabe09ff2b27d5dde0 100644 (file)
--- a/libavcodec/qdm2.c
+++ b/libavcodec/qdm2.c
@@ -26,6 +26,7 @@
  * @file
  * QDM2 decoder
  * @author Ewald Snel, Benjamin Larsson, Alex Beregszaszi, Roberto Togni
+ *
  * The decoder is not perfect yet, there are still some distortions
  * especially on files encoded with 16 or 8 subbands.
  */
@@ -172,9 +173,10 @@ typedef struct {
 
     /// Synthesis filter
     MPADSPContext mpadsp;
-    DECLARE_ALIGNED(16, MPA_INT, synth_buf)[MPA_MAX_CHANNELS][512*2];
+    DECLARE_ALIGNED(32, float, synth_buf)[MPA_MAX_CHANNELS][512*2];
     int synth_buf_offset[MPA_MAX_CHANNELS];
-    DECLARE_ALIGNED(16, int32_t, sb_samples)[MPA_MAX_CHANNELS][128][SBLIMIT];
+    DECLARE_ALIGNED(32, float, sb_samples)[MPA_MAX_CHANNELS][128][SBLIMIT];
+    DECLARE_ALIGNED(32, float, samples)[MPA_MAX_CHANNELS * MPA_FRAME_SIZE];
 
     /// Mixed temporary data used in decoding
     float tone_level[MPA_MAX_CHANNELS][30][64];
@@ -331,11 +333,6 @@ static av_cold void qdm2_init_vlc(void)
     }
 }
 
-
-/* for floating point to fixed point conversion */
-static const float f2i_scale = (float) (1 << (FRAC_BITS - 15));
-
-
 static int qdm2_get_vlc (GetBitContext *gb, VLC *vlc, int flag, int depth)
 {
     int value;
@@ -484,8 +481,8 @@ static void build_sb_samples_from_noise (QDM2Context *q, int sb)
 
     for (ch = 0; ch < q->nb_channels; ch++)
         for (j = 0; j < 64; j++) {
-            q->sb_samples[ch][j * 2][sb] = (int32_t)(f2i_scale * SB_DITHERING_NOISE(sb,q->noise_idx) * q->tone_level[ch][sb][j] + .5);
-            q->sb_samples[ch][j * 2 + 1][sb] = (int32_t)(f2i_scale * SB_DITHERING_NOISE(sb,q->noise_idx) * q->tone_level[ch][sb][j] + .5);
+            q->sb_samples[ch][j * 2][sb] = SB_DITHERING_NOISE(sb,q->noise_idx) * q->tone_level[ch][sb][j];
+            q->sb_samples[ch][j * 2 + 1][sb] = SB_DITHERING_NOISE(sb,q->noise_idx) * q->tone_level[ch][sb][j];
         }
 }
 
@@ -925,11 +922,11 @@ static void synthfilt_build_sb_samples (QDM2Context *q, GetBitContext *gb, int l
                     for (chs = 0; chs < q->nb_channels; chs++)
                         for (k = 0; k < run; k++)
                             if ((j + k) < 128)
-                                q->sb_samples[chs][j + k][sb] = (int32_t)(f2i_scale * q->tone_level[chs][sb][((j + k)/2)] * tmp[k][chs] + .5);
+                                q->sb_samples[chs][j + k][sb] = q->tone_level[chs][sb][((j + k)/2)] * tmp[k][chs];
                 } else {
                     for (k = 0; k < run; k++)
                         if ((j + k) < 128)
-                            q->sb_samples[ch][j + k][sb] = (int32_t)(f2i_scale * q->tone_level[ch][sb][(j + k)/2] * samples[k] + .5);
+                            q->sb_samples[ch][j + k][sb] = q->tone_level[ch][sb][(j + k)/2] * samples[k];
                 }
 
                 j += run;
@@ -1603,7 +1600,6 @@ static void qdm2_calculate_fft (QDM2Context *q, int channel, int sub_packet)
  */
 static void qdm2_synthesis_filter (QDM2Context *q, int index)
 {
-    OUT_INT samples[MPA_MAX_CHANNELS * MPA_FRAME_SIZE];
     int i, k, ch, sb_used, sub_sampling, dither_state = 0;
 
     /* copy sb_samples */
@@ -1615,12 +1611,12 @@ static void qdm2_synthesis_filter (QDM2Context *q, int index)
                 q->sb_samples[ch][(8 * index) + i][k] = 0;
 
     for (ch = 0; ch < q->nb_channels; ch++) {
-        OUT_INT *samples_ptr = samples + ch;
+        float *samples_ptr = q->samples + ch;
 
         for (i = 0; i < 8; i++) {
-            ff_mpa_synth_filter_fixed(&q->mpadsp,
+            ff_mpa_synth_filter_float(&q->mpadsp,
                 q->synth_buf[ch], &(q->synth_buf_offset[ch]),
-                ff_mpa_synth_window_fixed, &dither_state,
+                ff_mpa_synth_window_float, &dither_state,
                 samples_ptr, q->nb_channels,
                 q->sb_samples[ch][(8 * index) + i]);
             samples_ptr += 32 * q->nb_channels;
@@ -1632,7 +1628,7 @@ static void qdm2_synthesis_filter (QDM2Context *q, int index)
 
     for (ch = 0; ch < q->channels; ch++)
         for (i = 0; i < q->frame_size; i++)
-            q->output_buffer[q->channels * i + ch] += (float)(samples[q->nb_channels * sub_sampling * i + ch] >> (sizeof(OUT_INT)*8-16));
+            q->output_buffer[q->channels * i + ch] += (1 << 23) * q->samples[q->nb_channels * sub_sampling * i + ch];
 }
 
 
@@ -1649,7 +1645,7 @@ static av_cold void qdm2_init(QDM2Context *q) {
     initialized = 1;
 
     qdm2_init_vlc();
-    ff_mpa_synth_init_fixed(ff_mpa_synth_window_fixed);
+    ff_mpa_synth_init_float(ff_mpa_synth_window_float);
     softclip_table_init();
     rnd_table_init();
     init_noise_samples();