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

9 files changed, 181 insertions, 33 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];
diff --git a/apps/codecs/mpa.c b/apps/codecs/mpa.c
index 7ac96fb..c916aef 100644
--- a/apps/codecs/mpa.c
+++ b/apps/codecs/mpa.c
@@ -25,13 +25,34 @@
 
 CODEC_HEADER
 
+#if (CONFIG_CPU == PP5024 || CONFIG_CPU == PP5022 || CONFIG_CPU == PP5020 \
+    || CONFIG_CPU == PP5002) && !defined(MPEGPLAYER)
+#define MPA_SYNTH_ON_COP
+#endif
+
 struct mad_stream stream IBSS_ATTR;
 struct mad_frame frame IBSS_ATTR;
 struct mad_synth synth IBSS_ATTR;
 
+#ifdef MPA_SYNTH_ON_COP
+volatile short synth_running IBSS_ATTR;  /*synthesis is running*/
+volatile short die IBSS_ATTR;            /*thread should die*/
+
+#if (CONFIG_CPU == PP5024) || (CONFIG_CPU == PP5022)
+mad_fixed_t sbsample_prev[2][36][32] IBSS_ATTR;
+#else
+mad_fixed_t sbsample_prev[2][36][32] SHAREDBSS_ATTR; 
+#endif
+
+struct semaphore synth_done_sem IBSS_ATTR;
+struct semaphore synth_pending_sem IBSS_ATTR;
+#endif
+
 #define INPUT_CHUNK_SIZE   8192
 
 mad_fixed_t mad_frame_overlap[2][32][18] IBSS_ATTR;
+mad_fixed_t sbsample[2][36][32] IBSS_ATTR;
+
 unsigned char mad_main_data[MAD_BUFFER_MDLEN] IBSS_ATTR;
 /* TODO: what latency does layer 1 have? */
 int mpeg_latency[3] = { 0, 481, 529 };
@@ -43,6 +64,14 @@ void init_mad(void)
     ci->memset(&frame, 0, sizeof(struct mad_frame));
     ci->memset(&synth, 0, sizeof(struct mad_synth));
 
+#ifdef MPA_SYNTH_ON_COP
+    frame.sbsample_prev = &sbsample_prev;
+#else
+    frame.sbsample_prev = &sbsample;
+#endif
+
+    frame.sbsample=&sbsample;
+
     mad_stream_init(&stream);
     mad_frame_init(&frame);
     mad_synth_init(&synth);
@@ -51,6 +80,8 @@ void init_mad(void)
     ci->memset(mad_frame_overlap, 0, sizeof(mad_frame_overlap));
     frame.overlap = &mad_frame_overlap;
     stream.main_data = &mad_main_data;
+    
+    
 }
 
 int get_file_pos(int newtime)
@@ -159,6 +190,90 @@ static void set_elapsed(struct mp3entry* id3)
     }
 }
 
+#ifdef MPA_SYNTH_ON_COP
+
+/*
+ * Run the synthesis filter on the COProcessor 
+ */
+
+static int mad_synth_thread_stack[DEFAULT_STACK_SIZE/sizeof(int)/2] IBSS_ATTR;
+
+static const unsigned char * const mad_synth_thread_name = "mp3dec";
+static struct thread_entry *mad_synth_thread_p;
+
+
+static void mad_synth_thread(void){
+
+    while(1){
+        ci->semaphore_wait(&synth_pending_sem);
+        
+        if(die){
+            die=0;
+            invalidate_icache();
+            return ;
+        }   
+        synth_running = 1;
+        mad_synth_frame(&synth, &frame);
+        synth_running = 0;
+        ci->semaphore_release(&synth_done_sem);        
+    }    
+}
+
+
+static int mad_synth_thread_wait_pcm(void){
+    ci->semaphore_wait(&synth_done_sem);
+    return 0;  
+}
+
+static void mad_synth_thread_ready(void){
+    mad_fixed_t (*temp)[2][36][32];
+    while(1){
+        /*check if synth is currently running before changing its inputs! */
+        if(!synth_running){    
+            /*circular buffer that holds 2 frames' samples*/
+            temp=frame.sbsample;
+            frame.sbsample = frame.sbsample_prev;
+            frame.sbsample_prev=temp;
+            
+            ci->semaphore_release(&synth_pending_sem);
+            return ;
+        }
+        ci->yield(); /*synth thread currently running, wait for it*/
+    }
+}
+
+static int mad_synth_thread_create(void){
+       synth_running=0;
+       die=0;
+           
+       ci->semaphore_init(&synth_done_sem, 1, 0);
+       ci->semaphore_init(&synth_pending_sem, 1, 0);
+       
+       mad_synth_thread_p = ci->create_thread(mad_synth_thread, 
+                            mad_synth_thread_stack,
+                               sizeof(mad_synth_thread_stack), 0,
+                               mad_synth_thread_name 
+                               IF_PRIO(, PRIORITY_PLAYBACK), COP);
+    
+        if (mad_synth_thread_p == NULL)
+            return false;
+    
+        return true;    
+    
+}
+#else
+static void mad_synth_thread_ready(void){
+	 mad_synth_frame(&synth, &frame);
+}
+static int mad_synth_thread_create(void){
+	return 0;
+}
+static int mad_synth_thread_wait_pcm(void){
+	return 0;
+}
+
+#endif
+
 /* this is the codec entry point */
 enum codec_status codec_main(void)
 {
@@ -180,8 +295,12 @@ enum codec_status codec_main(void)
     /* Create a decoder instance */
 
     ci->configure(DSP_SET_SAMPLE_DEPTH, MAD_F_FRACBITS);
-
+    
 next_track:
+
+    /*does nothing on 1 processor systems*/
+    mad_synth_thread_create();
+
     status = CODEC_OK;
 
     /* Reinitializing seems to be necessary to avoid playback quircks when seeking. */
@@ -300,6 +419,9 @@ next_track:
            data (not the one just decoded above). When we exit the decoding
            loop we will need to process the final frame that was decoded. */
         if (framelength > 0) {
+            
+            mad_synth_thread_wait_pcm();
+            
             /* In case of a mono file, the second array will be ignored. */
             ci->pcmbuf_insert(&synth.pcm.samples[0][samples_to_skip],
                               &synth.pcm.samples[1][samples_to_skip],
@@ -308,8 +430,9 @@ next_track:
             /* Only skip samples for the first frame added. */
             samples_to_skip = 0;
         }
-
-        mad_synth_frame(&synth, &frame);
+        
+        mad_synth_thread_ready();
+        //mad_synth_frame(&synth, &frame);
 
         /* Check if sample rate and stereo settings changed in this frame. */
         if (frame.header.samplerate != current_frequency) {
@@ -345,11 +468,20 @@ next_track:
 
     /* Finish the remaining decoded frame.
        Cut the required samples from the end. */
-    if (framelength > stop_skip)
+    if (framelength > stop_skip){
+        mad_synth_thread_wait_pcm();
         ci->pcmbuf_insert(synth.pcm.samples[0], synth.pcm.samples[1],
                           framelength - stop_skip);
-
+}
+#ifdef MPA_SYNTH_ON_COP
+    /*mop up COP thread*/
+    die=1;
+    ci->semaphore_release(&synth_pending_sem);
+    ci->thread_wait(mad_synth_thread_p);
+    invalidate_icache();
     stream.error = 0;
+#endif
+    
 
     if (ci->request_next_track())
         goto next_track;
diff --git a/apps/plugins/mpegplayer/Makefile b/apps/plugins/mpegplayer/Makefile
index 66a6142..7f90cbc 100644
--- a/apps/plugins/mpegplayer/Makefile
+++ b/apps/plugins/mpegplayer/Makefile
@@ -10,7 +10,7 @@
 INCLUDES = -I$(APPSDIR) -I.. -I. $(TARGET_INC) -I$(FIRMDIR)/include -I$(FIRMDIR)/export \
  -I$(FIRMDIR)/common -I$(FIRMDIR)/drivers -I$(APPSDIR)/plugins/lib -I$(OUTDIR) -I$(BUILDDIR)
 CFLAGS = $(INCLUDES) $(GCCOPTS) -O2 $(TARGET) $(EXTRA_DEFINES) \
- -DTARGET_ID=$(TARGET_ID) -DMEM=${MEMORYSIZE} -DPLUGIN
+ -DTARGET_ID=$(TARGET_ID) -DMEM=${MEMORYSIZE} -DPLUGIN -DMPEGPLAYER
 
 ifdef APPEXTRA
    INCLUDES += $(patsubst %,-I$(APPSDIR)/%,$(subst :, ,$(APPEXTRA)))
diff --git a/apps/plugins/mpegplayer/audio_thread.c b/apps/plugins/mpegplayer/audio_thread.c
index 838dcad..3515815 100644
--- a/apps/plugins/mpegplayer/audio_thread.c
+++ b/apps/plugins/mpegplayer/audio_thread.c
@@ -55,6 +55,9 @@ static struct mad_stream stream IBSS_ATTR;
 static struct mad_frame  frame IBSS_ATTR;
 static struct mad_synth  synth IBSS_ATTR;
 
+/*sbsample buffer for mad_frame*/
+mad_fixed_t sbsample[2][36][32];
+
 /* 2567 bytes */
 static unsigned char mad_main_data[MAD_BUFFER_MDLEN];
 
@@ -229,6 +232,10 @@ static int audio_buffer(struct stream *str, enum stream_parse_mode type)
 /* Initialise libmad */
 static void init_mad(void)
 {
+    /*init the sbsample buffer*/
+    frame.sbsample = &sbsample;
+    frame.sbsample_prev = &sbsample;
+    
     mad_stream_init(&stream);
     mad_frame_init(&frame);
     mad_synth_init(&synth);