1 files changed, 143 insertions, 79 deletions

diff --git a/apps/codecs/libtremor/block.c b/apps/codecs/libtremor/block.c
index e609fc44f7..eb087e12a9 100644
--- a/apps/codecs/libtremor/block.c
+++ b/apps/codecs/libtremor/block.c
@@ -36,6 +36,13 @@ static int ilog(unsigned int v){
   return(ret);
 }
 
+static ogg_int32_t* _pcmp [CHANNELS] IBSS_ATTR;
+static ogg_int32_t* _pcmbp[CHANNELS] IBSS_ATTR;
+static ogg_int32_t* _pcmret[CHANNELS] IBSS_ATTR;
+/* if true, we have both pcm buffers in iram and we use a bufferflip.
+   if false, we have one in iram and one in mem, and we use a memcpy */
+static bool iram_pcm_doublebuffer IBSS_ATTR;
+
 /* pcm accumulator examples (not exhaustive):
 
  <-------------- lW ---------------->
@@ -145,18 +152,44 @@ int vorbis_block_clear(vorbis_block *vb){
 
 static int _vds_init(vorbis_dsp_state *v,vorbis_info *vi){
   int i;
+  long b_size[2];
+  LOOKUP_TNC *iramposw;
+  ogg_int32_t *internal_pcm=NULL;
+  
   codec_setup_info *ci=(codec_setup_info *)vi->codec_setup;
   private_state *b=NULL;
 
   memset(v,0,sizeof(*v));
+  v->reset_pcmb=true;
   b=(private_state *)(v->backend_state=_ogg_calloc(1,sizeof(*b)));
 
   v->vi=vi;
   b->modebits=ilog(ci->modes);
-
+  
+  /* allocate IRAM buffer for the PCM data generated by synthesis */
+  iram_malloc_init();
+  v->iram_pcm=(ogg_int32_t *)iram_malloc(vi->channels*ci->blocksizes[1]*sizeof(ogg_int32_t));
+  if(v->iram_pcm != NULL) v->iram_pcm_storage=ci->blocksizes[1];
+  else v->iram_pcm_storage=0;
+  
+  v->centerW=0;
+  
   /* Vorbis I uses only window type 0 */
-  b->window[0]=_vorbis_window(0,ci->blocksizes[0]/2);
-  b->window[1]=_vorbis_window(0,ci->blocksizes[1]/2);
+  b_size[0]=ci->blocksizes[0]/2;
+  b_size[1]=ci->blocksizes[1]/2;
+  b->window[0]=_vorbis_window(0,b_size[0]);
+  b->window[1]=_vorbis_window(0,b_size[1]);
+  
+  /* allocate IRAM buffer for window tables too, if sufficient iram available */
+  /* give preference to the larger window over the smaller window
+     (on the assumption that both windows are equally likely used) */
+  for(i=1; i>=0; i--){
+    iramposw=(LOOKUP_TNC *)iram_malloc(b_size[i]*sizeof(LOOKUP_TNC));
+    if(iramposw!=NULL) {
+      memcpy(iramposw, b->window[i], b_size[i]*sizeof(LOOKUP_TNC));
+      b->window[i]=iramposw;
+    }
+  }
 
   /* finish the codebooks */
   if(!ci->fullbooks){
@@ -169,14 +202,34 @@ static int _vds_init(vorbis_dsp_state *v,vorbis_info *vi){
     }
   }
 
+  /* if we can get away with it, put a double buffer into IRAM too, so that
+     overlap-add runs iram-to-iram and we avoid needing to memcpy */
   v->pcm_storage=ci->blocksizes[1];
-  v->pcm=(ogg_int32_t **)_ogg_malloc(vi->channels*sizeof(*v->pcm));
-  v->pcmb=(ogg_int32_t **)_ogg_malloc(vi->channels*sizeof(*v->pcmb));
-  v->pcmret=(ogg_int32_t **)_ogg_malloc(vi->channels*sizeof(*v->pcmret));
-
-  for(i=0;i<vi->channels;i++)
-    v->pcm[i]=(ogg_int32_t *)_ogg_calloc(v->pcm_storage,sizeof(*v->pcm[i]));
-    
+  v->pcm=_pcmp;
+  v->pcmret=_pcmret;
+  v->pcmb=_pcmbp;
+
+  _pcmp[0]=NULL;
+  _pcmp[1]=NULL;
+  _pcmbp[0]=NULL;
+  _pcmbp[1]=NULL;
+  
+  if(NULL != (internal_pcm = iram_malloc(vi->channels*v->pcm_storage*sizeof(ogg_int32_t))))
+  {
+    /* one-time initialisation at codec start or on switch from
+       blocksizes greater than IRAM_PCM_END to sizes that fit */
+    for(i=0;i<vi->channels;i++)
+      v->pcm[i]=&internal_pcm[i*v->pcm_storage];
+    iram_pcm_doublebuffer = true;
+  }
+  else
+  {
+    /* one-time initialisation at codec start or on switch from
+       blocksizes that fit in IRAM_PCM_END to those that don't */
+    for(i=0;i<vi->channels;i++)
+      v->pcm[i]=(ogg_int32_t *)_ogg_calloc(v->pcm_storage,sizeof(*v->pcm[i])); 
+    iram_pcm_doublebuffer = false;
+  }  
 
   /* all 1 (large block) or 0 (small block) */
   /* explicitly set for the sake of clarity */
@@ -203,13 +256,17 @@ int vorbis_synthesis_restart(vorbis_dsp_state *v){
   ci=vi->codec_setup;
   if(!ci)return -1;
 
-  v->centerW=ci->blocksizes[1]/2;
-  v->pcm_current=v->centerW;
+  v->centerW=0;
+  v->pcm_current=0;
   
   v->pcm_returned=-1;
   v->granulepos=-1;
   v->sequence=-1;
   ((private_state *)(v->backend_state))->sample_count=-1;
+  
+  /* indicate to synthesis code that buffer pointers no longer valid
+     (if we're using double pcm buffer) and will need to reset them */
+  v->reset_pcmb = true;
 
   return(0);
 }
@@ -228,11 +285,12 @@ void vorbis_dsp_clear(vorbis_dsp_state *v){
     codec_setup_info *ci=(codec_setup_info *)(vi?vi->codec_setup:NULL);
     private_state *b=(private_state *)v->backend_state;
 
-    if(v->pcm){
-      for(i=0;i<vi->channels;i++)
-        if(v->pcm[i])_ogg_free(v->pcm[i]);
-      _ogg_free(v->pcm);
-      if(v->pcmret)_ogg_free(v->pcmret);
+    if(!iram_pcm_doublebuffer)
+    {
+      if(v->pcm){
+        for(i=0;i<vi->channels;i++)
+          if(v->pcm[i])_ogg_free(v->pcm[i]);
+      } 
     }
 
     /* free mode lookups; these are actually vorbis_look_mapping structs */
@@ -258,7 +316,7 @@ void vorbis_dsp_clear(vorbis_dsp_state *v){
    calling (as it relies on the previous block). */
 
 int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb)
-    ICODE_ATTR_TREMOR_NOT_MDCT;
+    ICODE_ATTR;
 int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
   vorbis_info *vi=v->vi;
   codec_setup_info *ci=(codec_setup_info *)vi->codec_setup;
@@ -278,85 +336,91 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
   }
 
   v->sequence=vb->sequence;
+  int n=ci->blocksizes[v->W]/2;
+  int ln=ci->blocksizes[v->lW]/2;
   
-  if(vb->pcm){  /* no pcm to process if vorbis_synthesis_trackonly 
-                   was called on block */
-    int n=ci->blocksizes[v->W]/2;
+  if(LIKELY(vb->pcm)){ /* no pcm to process if vorbis_synthesis_trackonly 
+                          was called on block */
+    int prevCenter;
     int n0=ci->blocksizes[0]/2;
     int n1=ci->blocksizes[1]/2;
-    
-    int thisCenter;
-    int prevCenter;
-    
-    if(v->centerW){
-      thisCenter=n1;
-      prevCenter=0;
-    }else{
-      thisCenter=0;
-      prevCenter=n1;
+
+    if(iram_pcm_doublebuffer)
+    {
+      prevCenter = ln;
+    }
+    else 
+    {
+      prevCenter = v->centerW;
+      v->centerW = n1 - v->centerW;
     }
-    
-    /* v->pcm is now used like a two-stage double buffer.  We don't want
-       to have to constantly shift *or* adjust memory usage.  Don't
-       accept a new block until the old is shifted out */
     
     /* overlap/add PCM */
-    
-    for(j=0;j<vi->channels;j++){
-      /* the overlap/add section */
-      if(v->lW){
-        if(v->W){
-          /* large/large */
-          ogg_int32_t *pcm=v->pcm[j]+prevCenter;
-          ogg_int32_t *p=vb->pcm[j];
-          vect_add(p, pcm, n1);
-          v->pcmb[j]=p;
-        }else{
-          /* large/small */
-          ogg_int32_t *pcm=v->pcm[j]+prevCenter+n1/2-n0/2;
-          ogg_int32_t *p=vb->pcm[j];
-          vect_add(pcm, p, n0);
-          v->pcmb[j]=v->pcm[j]+prevCenter;
+    /* nb nothing to overlap with on first block so don't bother */ 
+    if(LIKELY(v->pcm_returned!=-1))
+    {    
+      for(j=0;j<vi->channels;j++)
+      {
+        ogg_int32_t *pcm=v->pcm[j]+prevCenter;
+        ogg_int32_t *p=vb->pcm[j];
+      
+        /* the overlap/add section */
+        if(v->lW == v->W)
+        {
+            /* large/large or small/small */
+            vect_add_right_left(pcm,p,n);
+            v->pcmb[j]=pcm;
         }
-      }else{
-        if(v->W){
-          /* small/large */
-          ogg_int32_t *pcm=v->pcm[j]+prevCenter;
-          ogg_int32_t *p=vb->pcm[j]+n1/2-n0/2;
-          vect_add(p, pcm, n0);
-          v->pcmb[j]=p;
-        }else{
-          /* small/small */
-          ogg_int32_t *pcm=v->pcm[j]+prevCenter;
-          ogg_int32_t *p=vb->pcm[j];
-          vect_add(p, pcm, n0);
-          v->pcmb[j]=p;
+        else if (!v->W)
+        {
+            /* large/small */
+            vect_add_right_left(pcm + (n1-n0)/2, p, n0);
+            v->pcmb[j]=pcm;
+        }
+        else
+        {
+            /* small/large */
+            p += (n1-n0)/2;
+            vect_add_left_right(p,pcm,n0);
+            v->pcmb[j]=p;
         }
       }
-      
-      /* the copy section */
+    }
+
+    /* the copy section */
+    if(iram_pcm_doublebuffer)
+    {
+      /* just flip the pointers over as we have a double buffer in iram */
+      ogg_int32_t *p;
+      p=v->pcm[0];
+      v->pcm[0]=vb->pcm[0];
+      vb->pcm[0] = p;
+      p=v->pcm[1];
+      v->pcm[1]=vb->pcm[1];
+      vb->pcm[1] = p;
+    }
+    else
+    {        
+      for(j=0;j<vi->channels;j++)
       {
-        ogg_int32_t *pcm=v->pcm[j]+thisCenter;
-        ogg_int32_t *p=vb->pcm[j]+n;
-        vect_copy(pcm, p, n);
+        /* at best only vb->pcm is in iram, and that's where we do the
+           synthesis, so we copy out the right-hand subframe of last
+           synthesis into (noniram) local buffer so we can still do
+           synth in iram */
+        vect_copy(v->pcm[j]+v->centerW, vb->pcm[j]+n, n);
       }
     }
     
-    if(v->centerW)
-      v->centerW=0;
-    else
-      v->centerW=n1;
-    
     /* deal with initial packet state; we do this using the explicit
        pcm_returned==-1 flag otherwise we're sensitive to first block
        being short or long */
 
     if(v->pcm_returned==-1){
-      v->pcm_returned=thisCenter;
-      v->pcm_current=thisCenter;
+      v->pcm_returned=0;
+      v->pcm_current=0;
     }else{
       v->pcm_returned=0;
-      v->pcm_current=ci->blocksizes[v->lW]/4+ci->blocksizes[v->W]/4;
+      v->pcm_current=(n+ln)/2;
     }
 
   }
@@ -375,7 +439,7 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
   if(b->sample_count==-1){
     b->sample_count=0;
   }else{
-    b->sample_count+=ci->blocksizes[v->lW]/4+ci->blocksizes[v->W]/4;
+    b->sample_count+=(n+ln)/2;
   }
     
   if(v->granulepos==-1){
@@ -406,7 +470,7 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
       
     }
   }else{
-    v->granulepos+=ci->blocksizes[v->lW]/4+ci->blocksizes[v->W]/4;
+    v->granulepos+=(n+ln)/2;
     if(vb->granulepos!=-1 && v->granulepos!=vb->granulepos){
       
       if(v->granulepos>vb->granulepos){