Commit FS#9318 - MP3 synthesis filter on COP. Loads the MP3 synth filer on to the CoProcessor on all PortalPlayer devices, resulting in an ~90% speedup according to test_codec on the Sansa. Real world improvement is somewhat less, but still considerable. Allows MP3 decoding at 30MHz without boosting, or use of more DSP/EQ with less boosting/skipping, thus improving battery life. Minor changes to mpegplayer to retain compatibility with libmad changes. Should be no significant changes for other targets or codecs.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@18557 a1c6a512-1295-4272-9138-f99709370657

6 files changed, 36 insertions, 27 deletions

diff --git a/apps/codecs/libmad/frame.c b/apps/codecs/libmad/frame.c
index 91cf2f9..2c7fdca 100644
--- a/apps/codecs/libmad/frame.c
+++ b/apps/codecs/libmad/frame.c
@@ -31,6 +31,7 @@
 # include "timer.h"
 # include "layer12.h"
 # include "layer3.h"
+# include "../lib/codeclib.h"
 
 static
 unsigned long const bitrate_table[5][15] = {
@@ -467,7 +468,9 @@ int mad_frame_decode(struct mad_frame *frame, struct mad_stream *stream)
 
     mad_bit_finish(&next_frame);
   }
-
+    
+ 
+    
   return 0;
 
  fail:
@@ -485,8 +488,8 @@ void mad_frame_mute(struct mad_frame *frame)
 
   for (s = 0; s < 36; ++s) {
     for (sb = 0; sb < 32; ++sb) {
-      frame->sbsample[0][s][sb] =
-      frame->sbsample[1][s][sb] = 0;
+      (*frame->sbsample)[0][s][sb] =
+      (*frame->sbsample)[1][s][sb] = 0;
     }
   }
 
diff --git a/apps/codecs/libmad/frame.h b/apps/codecs/libmad/frame.h
index dce573d..d2d6dca 100644
--- a/apps/codecs/libmad/frame.h
+++ b/apps/codecs/libmad/frame.h
@@ -65,12 +65,15 @@ struct mad_header {
 };
 
 struct mad_frame {
-  struct mad_header header;		/* MPEG audio header */
-
-  int options;				/* decoding options (from stream) */
-
-  mad_fixed_t sbsample[2][36][32];	/* synthesis subband filter samples */
-  mad_fixed_t (*overlap)[2][32][18];	/* Layer III block overlap data */
+  struct mad_header header;     /* MPEG audio header */
+  int options;              /* decoding options (from stream) */
+
+  mad_fixed_t (*sbsample)[2][36][32];   /* synthesis subband filter samples */
+  mad_fixed_t (*sbsample_prev)[2][36][32];  /* synthesis subband filter samples 
+                                               from previous frame only needed
+                                               when synthesis is on cop */
+                                               
+  mad_fixed_t (*overlap)[2][32][18];    /* Layer III block overlap data */
 };
 
 # define MAD_NCHANNELS(header)		((header)->mode ? 2 : 1)
diff --git a/apps/codecs/libmad/layer12.c b/apps/codecs/libmad/layer12.c
index c71c005..ccac392 100644
--- a/apps/codecs/libmad/layer12.c
+++ b/apps/codecs/libmad/layer12.c
@@ -185,7 +185,7 @@ int mad_layer_I(struct mad_stream *stream, struct mad_frame *frame)
     for (sb = 0; sb < bound; ++sb) {
       for (ch = 0; ch < nch; ++ch) {
 	nb = allocation[ch][sb];
-	frame->sbsample[ch][s][sb] = nb ?
+	(*frame->sbsample)[ch][s][sb] = nb ?
 	  mad_f_mul(I_sample(&stream->ptr, nb),
 		    sf_table[scalefactor[ch][sb]]) : 0;
       }
@@ -198,13 +198,13 @@ int mad_layer_I(struct mad_stream *stream, struct mad_frame *frame)
 	sample = I_sample(&stream->ptr, nb);
 
 	for (ch = 0; ch < nch; ++ch) {
-	  frame->sbsample[ch][s][sb] =
+	  (*frame->sbsample)[ch][s][sb] =
 	    mad_f_mul(sample, sf_table[scalefactor[ch][sb]]);
 	}
       }
       else {
 	for (ch = 0; ch < nch; ++ch)
-	  frame->sbsample[ch][s][sb] = 0;
+	  (*frame->sbsample)[ch][s][sb] = 0;
       }
     }
   }
@@ -492,13 +492,13 @@ int mad_layer_II(struct mad_stream *stream, struct mad_frame *frame)
 	  II_samples(&stream->ptr, &qc_table[index], samples);
 
 	  for (s = 0; s < 3; ++s) {
-	    frame->sbsample[ch][3 * gr + s][sb] =
+	    (*frame->sbsample)[ch][3 * gr + s][sb] =
 	      mad_f_mul(samples[s], sf_table[scalefactor[ch][sb][gr / 4]]);
 	  }
 	}
 	else {
 	  for (s = 0; s < 3; ++s)
-	    frame->sbsample[ch][3 * gr + s][sb] = 0;
+	    (*frame->sbsample)[ch][3 * gr + s][sb] = 0;
 	}
       }
     }
@@ -512,7 +512,7 @@ int mad_layer_II(struct mad_stream *stream, struct mad_frame *frame)
 
 	for (ch = 0; ch < nch; ++ch) {
 	  for (s = 0; s < 3; ++s) {
-	    frame->sbsample[ch][3 * gr + s][sb] =
+	    (*frame->sbsample)[ch][3 * gr + s][sb] =
 	      mad_f_mul(samples[s], sf_table[scalefactor[ch][sb][gr / 4]]);
 	  }
 	}
@@ -520,7 +520,7 @@ int mad_layer_II(struct mad_stream *stream, struct mad_frame *frame)
       else {
 	for (ch = 0; ch < nch; ++ch) {
 	  for (s = 0; s < 3; ++s)
-	    frame->sbsample[ch][3 * gr + s][sb] = 0;
+	    (*frame->sbsample)[ch][3 * gr + s][sb] = 0;
 	}
       }
     }
@@ -528,7 +528,7 @@ int mad_layer_II(struct mad_stream *stream, struct mad_frame *frame)
     for (ch = 0; ch < nch; ++ch) {
       for (s = 0; s < 3; ++s) {
 	for (sb = sblimit; sb < 32; ++sb)
-	  frame->sbsample[ch][3 * gr + s][sb] = 0;
+	  (*frame->sbsample)[ch][3 * gr + s][sb] = 0;
       }
     }
   }
diff --git a/apps/codecs/libmad/layer3.c b/apps/codecs/libmad/layer3.c
index a95927e..0a53086 100644
--- a/apps/codecs/libmad/layer3.c
+++ b/apps/codecs/libmad/layer3.c
@@ -3112,7 +3112,7 @@ enum mad_error III_decode(struct mad_bitptr *ptr, struct mad_frame *frame,
 
     for (ch = 0; ch < nch; ++ch) {
       struct channel const *channel = &granule->ch[ch];
-      mad_fixed_t (*sample)[32] = &frame->sbsample[ch][18 * gr];
+      mad_fixed_t (*sample)[32] = &((*frame->sbsample)[ch][18 * gr]);
       unsigned int sb, l, i, sblimit;
       mad_fixed_t output[36];
 
diff --git a/apps/codecs/libmad/mad.h b/apps/codecs/libmad/mad.h
index f5d8f1d..52a74d1 100644
--- a/apps/codecs/libmad/mad.h
+++ b/apps/codecs/libmad/mad.h
@@ -777,10 +777,13 @@ struct mad_header {
 
 struct mad_frame {
   struct mad_header header;		/* MPEG audio header */
-
   int options;				/* decoding options (from stream) */
 
-  mad_fixed_t sbsample[2][36][32];	/* synthesis subband filter samples */
+  mad_fixed_t (*sbsample)[2][36][32];	/* synthesis subband filter samples */
+  mad_fixed_t (*sbsample_prev)[2][36][32];  /* synthesis subband filter samples 
+                                               from previous frame only needed
+                                               when synthesis is on cop */
+                                               
   mad_fixed_t (*overlap)[2][32][18];	/* Layer III block overlap data */
 };
 
diff --git a/apps/codecs/libmad/synth.c b/apps/codecs/libmad/synth.c
index c023f01..b1a8491 100644
--- a/apps/codecs/libmad/synth.c
+++ b/apps/codecs/libmad/synth.c
@@ -592,14 +592,14 @@ void synth_full(struct mad_synth *synth, struct mad_frame const *frame,
   int sb;
   unsigned int phase, ch, s, p;
   mad_fixed_t *pcm, (*filter)[2][2][16][8];
-  mad_fixed_t const (*sbsample)[36][32];
+  mad_fixed_t (*sbsample)[36][32];
   mad_fixed_t (*fe)[8], (*fx)[8], (*fo)[8];
   mad_fixed_t const (*D0ptr)[32];
   mad_fixed_t const (*D1ptr)[32];
   mad_fixed64hi_t hi0, hi1;
 
   for (ch = 0; ch < nch; ++ch) {
-    sbsample = &frame->sbsample[ch];
+    sbsample = &*frame->sbsample_prev[ch];
     filter   = &synth->filter[ch];
     phase    = synth->phase;
     pcm      = synth->pcm.samples[ch];
@@ -1053,7 +1053,7 @@ void synth_full(struct mad_synth *synth, struct mad_frame const *frame,
   int          p;
   unsigned int phase, ch, s;
   mad_fixed_t *pcm, (*filter)[2][2][16][8];
-  mad_fixed_t const (*sbsample)[36][32];
+  mad_fixed_t  (*sbsample)[36][32];
   mad_fixed_t (*fe)[8], (*fx)[8], (*fo)[8];
   mad_fixed_t const (*D0ptr)[32], *ptr;
   mad_fixed_t const (*D1ptr)[32];
@@ -1061,7 +1061,7 @@ void synth_full(struct mad_synth *synth, struct mad_frame const *frame,
   mad_fixed64lo_t lo;
 
   for (ch = 0; ch < nch; ++ch) {
-    sbsample = &frame->sbsample[ch];
+    sbsample = &(*frame->sbsample_prev)[ch];
     filter   = &synth->filter[ch];
     phase    = synth->phase;
     pcm      = synth->pcm.samples[ch];
@@ -1202,7 +1202,7 @@ void synth_full(struct mad_synth *synth, struct mad_frame const *frame,
   mad_fixed64lo_t lo;
 
   for (ch = 0; ch < nch; ++ch) {
-    sbsample = &frame->sbsample[ch];
+    sbsample = &frame->sbsample_prev[ch];
     filter   = &synth->filter[ch];
     phase    = synth->phase;
     pcm      = synth->pcm.samples[ch];
@@ -1403,14 +1403,14 @@ void synth_half(struct mad_synth *synth, struct mad_frame const *frame,
 {
   unsigned int phase, ch, s, sb, pe, po;
   mad_fixed_t *pcm1, *pcm2, (*filter)[2][2][16][8];
-  mad_fixed_t const (*sbsample)[36][32];
+  mad_fixed_t  (*sbsample)[36][32];
   register mad_fixed_t (*fe)[8], (*fx)[8], (*fo)[8];
   register mad_fixed_t const (*Dptr)[32], *ptr;
   register mad_fixed64hi_t hi;
   register mad_fixed64lo_t lo;
 
   for (ch = 0; ch < nch; ++ch) {
-    sbsample = &frame->sbsample[ch];
+    sbsample = &(*frame->sbsample_prev)[ch];
     filter   = &synth->filter[ch];
     phase    = synth->phase;
     pcm1     = synth->pcm.samples[ch];