1 files changed, 46 insertions, 135 deletions

diff --git a/apps/codecs/libtremor/block.c b/apps/codecs/libtremor/block.c
index d678719cbe..b7ce5adc4e 100644
--- a/apps/codecs/libtremor/block.c
+++ b/apps/codecs/libtremor/block.c
@@ -25,6 +25,7 @@
 #include "window.h"
 #include "registry.h"
 #include "misc.h"
+#include "ffmpeg_render_line.h"
 //#include <codecs/lib/codeclib.h>
 
 static int ilog(unsigned int v){
@@ -37,11 +38,10 @@ 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;
 /* save original pointers returned by malloc so we can free it easily */
-static ogg_int32_t* pcm_copy[CHANNELS] = {NULL};
+static ogg_int32_t* malloc_pointers[3] = {NULL};
 
 /* pcm accumulator examples (not exhaustive):
 
@@ -153,14 +153,13 @@ int vorbis_block_clear(vorbis_block *vb){
 static int _vds_init(vorbis_dsp_state *v,vorbis_info *vi){
   int i;
   long b_size[2];
-  
+
   codec_setup_info *ci=(codec_setup_info *)vi->codec_setup;
   private_state *b=NULL;
 
   if(ci==NULL) return 1;
 
   memset(v,0,sizeof(*v));
-  v->reset_pcmb=true;
   b=(private_state *)(v->backend_state=_ogg_calloc(1,sizeof(*b)));
 
   v->vi=vi;
@@ -169,16 +168,42 @@ static int _vds_init(vorbis_dsp_state *v,vorbis_info *vi){
 #ifdef TREMOR_USE_IRAM
   /* allocate IRAM buffer for the PCM data generated by synthesis */
   iram_malloc_init();
-  v->first_pcm = iram_malloc(vi->channels*ci->blocksizes[1]*sizeof(ogg_int32_t));
-  /* when can't allocate IRAM buffer, allocate normal RAM buffer */
-  if(v->first_pcm == NULL)
+
+  v->floors = iram_malloc(vi->channels*ci->blocksizes[1]/2*sizeof(ogg_int32_t));
+  v->residues[0] = iram_malloc(vi->channels*ci->blocksizes[1]/2*sizeof(ogg_int32_t));
+  /* 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->residues[1] = iram_malloc(vi->channels*ci->blocksizes[1]/2*sizeof(ogg_int32_t));
+  if (v->residues[1] == NULL)
+    v->saved = iram_malloc(vi->channels*ci->blocksizes[1]/4*sizeof(ogg_int32_t));
+
 #endif
-  {
-    pcm_copy[0] = _ogg_malloc(vi->channels*ci->blocksizes[1]*sizeof(ogg_int32_t));
-    v->first_pcm = pcm_copy[0];
+
+  if (v->residues[0] == NULL) {
+    malloc_pointers[0] = _ogg_malloc(vi->channels*ci->blocksizes[1]/2*sizeof(ogg_int32_t));
+    v->residues[0] = malloc_pointers[0];
+  }
+
+  if (v->residues[1] == NULL && v->saved == NULL) {
+    malloc_pointers[1] = _ogg_malloc(vi->channels*ci->blocksizes[1]/2*sizeof(ogg_int32_t));
+    v->residues[1] = malloc_pointers[1];
   }
 
-  v->centerW=0;
+  if (v->floors == NULL) {
+    malloc_pointers[2] = _ogg_malloc(vi->channels*ci->blocksizes[1]/2*sizeof(ogg_int32_t));
+    v->floors = malloc_pointers[2];
+  }
+
+  /* needed for the first overlap/add */
+  if (v->saved) {
+    memset(v->saved, 0, vi->channels*ci->blocksizes[1]/4*sizeof(ogg_int32_t));
+    for (i = 0; i < vi->channels; i++)
+      v->saved_ptr[i] = v->saved + i*ci->blocksizes[1]/4;
+  } else {
+    memset(v->residues[1], 0, vi->channels*ci->blocksizes[1]/2*sizeof(ogg_int32_t));
+    for (i = 0; i < vi->channels; i++)
+      v->saved_ptr[i] = v->residues[1] + i*ci->blocksizes[1]/2;
+  }
 
   /* Vorbis I uses only window type 0 */
   b_size[0]=ci->blocksizes[0]/2;
@@ -214,37 +239,13 @@ 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=_pcmp;
   v->pcmret=_pcmret;
   v->pcmb=_pcmbp;
 
-  _pcmp[0]=NULL;
-  _pcmp[1]=NULL;
   _pcmbp[0]=NULL;
   _pcmbp[1]=NULL;
 
-#ifdef TREMOR_USE_IRAM
-  if(NULL != (v->iram_double_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]=&v->iram_double_pcm[i*v->pcm_storage];
-  }
-  else
-#endif
-  {
-    /* one-time initialisation at codec start or on switch from
-       blocksizes that fit in IRAM_PCM_END to those that don't */
-    /* save copy of the pointer so we can free it easily later */
-    pcm_copy[1] = _ogg_calloc(vi->channels*v->pcm_storage,sizeof(*v->pcm[i]));
-    for(i=0;i<vi->channels;i++)
-      v->pcm[i] = pcm_copy[1]+i*v->pcm_storage;
-  }
-
   /* all 1 (large block) or 0 (small block) */
   /* explicitly set for the sake of clarity */
   v->lW=0; /* previous window size */
@@ -274,35 +275,19 @@ abort_books:
 int vorbis_synthesis_restart(vorbis_dsp_state *v){
   vorbis_info *vi=v->vi;
   codec_setup_info *ci;
-  int i;
 
   if(!v->backend_state)return -1;
   if(!vi)return -1;
   ci=vi->codec_setup;
   if(!ci)return -1;
 
-  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;
-  /* also reset our copy of the double buffer pointers if we have one */
-#ifdef TREMOR_USE_IRAM
-  if(v->iram_double_pcm)
-  {
-    for(i=0;i<vi->channels;i++)
-      v->pcm[i]=&v->iram_double_pcm[i*v->pcm_storage];
-  }
-#else
-  for(i=0;i<vi->channels;i++)
-    v->pcm[i] = pcm_copy[1]+i*v->pcm_storage;
-#endif
+
   return(0);
 }
 
@@ -323,11 +308,10 @@ void vorbis_dsp_clear(vorbis_dsp_state *v){
     if(vi != NULL)
     {
       /* pcm buffer came from oggmalloc rather than iram */
-      for(i=0;i<2;i++)
-        if(pcm_copy[i])
-        {
-          _ogg_free(pcm_copy[i]);
-          pcm_copy[i] = NULL;
+      for(i=0;i<3;i++)
+        if(malloc_pointers[i]) {
+          _ogg_free(malloc_pointers[i]);
+          malloc_pointers[i] = NULL;
         }
     }
 
@@ -359,10 +343,6 @@ 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;
   private_state *b=v->backend_state;
-  int j;
-#ifdef TREMOR_USE_IRAM
-  bool iram_pcm_doublebuffer = (NULL != v->iram_double_pcm);
-#endif
 
   if(v->pcm_current>v->pcm_returned  && v->pcm_returned!=-1)return(OV_EINVAL);
 
@@ -380,79 +360,11 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
   int n=ci->blocksizes[v->W]/2;
   int ln=ci->blocksizes[v->lW]/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;
-
-#ifdef TREMOR_USE_IRAM
-    if(!iram_pcm_doublebuffer)
-    {
-      prevCenter = v->centerW;
-      v->centerW = n1 - v->centerW;
-    }
-    else
-#endif
-      prevCenter = ln;
-
-    /* overlap/add PCM */
-    /* 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)
-        {
-            /* 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;
-        }
-      }
-    }
-#ifdef TREMOR_USE_IRAM
-    /* the copy section */
-    if(!iram_pcm_doublebuffer)
-    {        
-      for(j=0;j<vi->channels;j++)
-      {
-        /* 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);
-      }
-    }
-    else
-#endif
-    {
-      /* 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;
-    }
- 
+  if(LIKELY(vb->pcmend != 0)){ /* no pcm to process if vorbis_synthesis_trackonly 
+                                  was called on block */
+    window_overlap_add(ci->blocksizes[v->W], ci->blocksizes[v->lW],
+                       ci->blocksizes[0], ci->blocksizes[1], vi->channels,
+                       b->window[v->W & v->lW], v);
     /* 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 */
@@ -464,7 +376,6 @@ int vorbis_synthesis_blockin(vorbis_dsp_state *v,vorbis_block *vb){
       v->pcm_returned=0;
       v->pcm_current=(n+ln)/2;
     }
-
   }
 
   /* track the frame number... This is for convenience, but also