diff --git a/src/engine/dispatch.h b/src/engine/dispatch.h
index 12831b5fb..127ffc18c 100644
--- a/src/engine/dispatch.h
+++ b/src/engine/dispatch.h
@@ -1,6 +1,8 @@
 #ifndef _DISPATCH_H
 #define _DISPATCH_H
 
+#include <stdint.h>
+
 #define ONE_SEMITONE 2200
 
 enum DivDispatchCmds {
@@ -83,7 +85,7 @@ class DivDispatch {
      * the engine shall resample to the output rate.
      */
     int rate;
-    virtual void acquire(int& l, int& r);
+    virtual void acquire(short** buf, size_t start, size_t len);
     virtual int dispatch(DivCommand c);
     virtual void tick();
 
diff --git a/src/engine/engine.cpp b/src/engine/engine.cpp
index 360d5d5b6..6d663664e 100644
--- a/src/engine/engine.cpp
+++ b/src/engine/engine.cpp
@@ -738,6 +738,10 @@ bool DivEngine::init() {
   bbOut[0]=new short[got.bufsize];
   bbOut[1]=new short[got.bufsize];
 
+  bbIn[0]=new short[32768];
+  bbIn[1]=new short[32768];
+  bbInLen=32768;
+
   for (int i=0; i<64; i++) {
     vibTable[i]=127*sin(((double)i/64.0)*(2*M_PI));
   }
diff --git a/src/engine/engine.h b/src/engine/engine.h
index 47c17ff62..79f0d948d 100644
--- a/src/engine/engine.h
+++ b/src/engine/engine.h
@@ -67,7 +67,9 @@ class DivEngine {
   short vibTable[64];
 
   blip_buffer_t* bb[2];
+  size_t bbInLen;
   int temp[2], prevSample[2];
+  short* bbIn[2];
   short* bbOut[2];
 
   int dispatchCmd(DivCommand c);
@@ -117,6 +119,7 @@ class DivEngine {
       cmdsPerSecond(0),
       view(DIV_STATUS_PATTERN),
       audioEngine(DIV_AUDIO_SDL),
+      bbInLen(0),
       temp{0,0},
       prevSample{0,0} {}
 };
diff --git a/src/engine/platform/abstract.cpp b/src/engine/platform/abstract.cpp
index 344e1c230..52a100756 100644
--- a/src/engine/platform/abstract.cpp
+++ b/src/engine/platform/abstract.cpp
@@ -1,8 +1,6 @@
 #include "../dispatch.h"
 
-void DivDispatch::acquire(int& l, int& r) {
-  l=0;
-  r=0;
+void DivDispatch::acquire(short** buf, size_t start, size_t len) {
 }
 
 void DivDispatch::tick() {
diff --git a/src/engine/platform/c64.cpp b/src/engine/platform/c64.cpp
index f5052f0cc..11fb8cb63 100644
--- a/src/engine/platform/c64.cpp
+++ b/src/engine/platform/c64.cpp
@@ -4,10 +4,12 @@
 
 #define FREQ_BASE 277.0f
 
-void DivPlatformC64::acquire(int& l, int& r) {
-  sid.clock();
-  l=sid.output();
-  r=l;
+void DivPlatformC64::acquire(short** buf, size_t start, size_t len) {
+  for (size_t i=start; i<start+len; i++) {
+    sid.clock();
+    buf[0][i]=sid.output();
+    buf[1][i]=buf[0][i];
+  }
 }
 
 void DivPlatformC64::tick() {
diff --git a/src/engine/platform/c64.h b/src/engine/platform/c64.h
index 5aab80223..a7b45fe9f 100644
--- a/src/engine/platform/c64.h
+++ b/src/engine/platform/c64.h
@@ -39,7 +39,7 @@ class DivPlatformC64: public DivDispatch {
 
   void updateWave();
   public:
-    void acquire(int& l, int& r);
+    void acquire(short** buf, size_t start, size_t len);
     int dispatch(DivCommand c);
     void tick();
     int init(DivEngine* parent, int channels, int sugRate);
diff --git a/src/engine/platform/dummy.cpp b/src/engine/platform/dummy.cpp
index ef30bae3e..972ed659b 100644
--- a/src/engine/platform/dummy.cpp
+++ b/src/engine/platform/dummy.cpp
@@ -2,16 +2,17 @@
 #include <stdio.h>
 #include <math.h>
 
-void DivPlatformDummy::acquire(int& l, int& r) {
-  l=0;
-  for (unsigned char i=0; i<chans; i++) {
-    if (chan[i].active) {
-      l+=((chan[i].pos>=0x8000)?chan[i].vol:-chan[i].vol)*chan[i].amp;
-      
-      chan[i].pos+=chan[i].freq;
+void DivPlatformDummy::acquire(short** buf, size_t start, size_t len) {
+  for (size_t i=start; i<start+len; i++) {
+    buf[0][i]=0;
+    for (unsigned char j=0; j<chans; j++) {
+      if (chan[j].active) {
+        buf[0][i]+=((chan[j].pos>=0x8000)?chan[j].vol:-chan[j].vol)*chan[j].amp;
+        chan[j].pos+=chan[j].freq;
+      }
     }
+    buf[1][i]=buf[0][i];
   }
-  r=l;
 }
 
 void DivPlatformDummy::tick() {
diff --git a/src/engine/platform/dummy.h b/src/engine/platform/dummy.h
index 333389d90..f220e35a5 100644
--- a/src/engine/platform/dummy.h
+++ b/src/engine/platform/dummy.h
@@ -15,7 +15,7 @@ class DivPlatformDummy: public DivDispatch {
   Channel chan[17];
   unsigned char chans;
   public:
-    void acquire(int& l, int& r);
+    void acquire(short** buf, size_t start, size_t len);
     int dispatch(DivCommand c);
     void tick();
     int init(DivEngine* parent, int channels, int sugRate);
diff --git a/src/engine/platform/gb.cpp b/src/engine/platform/gb.cpp
index 36e234f52..90385cd24 100644
--- a/src/engine/platform/gb.cpp
+++ b/src/engine/platform/gb.cpp
@@ -7,10 +7,12 @@
 
 #define FREQ_BASE 8015.85f
 
-void DivPlatformGB::acquire(int& l, int& r) {
-  GB_advance_cycles(gb,16);
-  l=gb->apu_output.final_sample.left<<2;
-  r=gb->apu_output.final_sample.right<<2;
+void DivPlatformGB::acquire(short** buf, size_t start, size_t len) {
+  for (int i=start; i<start+len; i++) {
+    GB_advance_cycles(gb,16);
+    buf[0][i]=gb->apu_output.final_sample.left<<2;
+    buf[1][i]=gb->apu_output.final_sample.right<<2;
+  }
 }
 
 void DivPlatformGB::updateWave() {
diff --git a/src/engine/platform/gb.h b/src/engine/platform/gb.h
index 404017d44..19e4f519d 100644
--- a/src/engine/platform/gb.h
+++ b/src/engine/platform/gb.h
@@ -36,7 +36,7 @@ class DivPlatformGB: public DivDispatch {
   GB_gameboy_t* gb;
   void updateWave();
   public:
-    void acquire(int& l, int& r);
+    void acquire(short** buf, size_t start, size_t len);
     int dispatch(DivCommand c);
     void tick();
     int init(DivEngine* parent, int channels, int sugRate);
diff --git a/src/engine/platform/genesis.cpp b/src/engine/platform/genesis.cpp
index 7bbad5c20..6aa1b0862 100644
--- a/src/engine/platform/genesis.cpp
+++ b/src/engine/platform/genesis.cpp
@@ -12,59 +12,60 @@ static unsigned char konOffs[6]={
   0, 1, 2, 4, 5, 6
 };
 
-void DivPlatformGenesis::acquire(int& l, int& r) {
+void DivPlatformGenesis::acquire(short** buf, size_t start, size_t len) {
   static short o[2];
   static int os[2];
 
-  if (dacMode && dacSample!=-1) {
-    dacPeriod-=6;
-    if (dacPeriod<1) {
-      DivSample* s=parent->song.sample[dacSample];
-      if (s->depth==8) {
-        writes.emplace(0x2a,(unsigned char)s->rendData[dacPos++]+0x80);
-      } else {
-        writes.emplace(0x2a,((unsigned short)s->rendData[dacPos++]+0x8000)>>8);
+  for (size_t h=start; h<start+len; h++) {
+    if (dacMode && dacSample!=-1) {
+      dacPeriod-=6;
+      if (dacPeriod<1) {
+        DivSample* s=parent->song.sample[dacSample];
+        if (s->depth==8) {
+          writes.emplace(0x2a,(unsigned char)s->rendData[dacPos++]+0x80);
+        } else {
+          writes.emplace(0x2a,((unsigned short)s->rendData[dacPos++]+0x8000)>>8);
+        }
+        if (dacPos>=s->rendLength) {
+          dacSample=-1;
+        }
+        dacPeriod+=dacRate;
       }
-      if (dacPos>=s->rendLength) {
-        dacSample=-1;
-      }
-      dacPeriod+=dacRate;
     }
-  }
-
-  os[0]=0; os[1]=0;
-  for (int i=0; i<6; i++) {
-    if (!writes.empty() && --delay<0) {
-      delay=0;
-      QueuedWrite& w=writes.front();
-      if (w.addrOrVal) {
-        OPN2_Write(&fm,0x1+((w.addr>>8)<<1),w.val);
-        //printf("write: %x = %.2x\n",w.addr,w.val);
-        lastBusy=0;
-        writes.pop();
-      } else {
-        lastBusy++;
-        if (fm.write_busy==0) {
-          //printf("busycounter: %d\n",lastBusy);
-          OPN2_Write(&fm,0x0+((w.addr>>8)<<1),w.addr);
-          w.addrOrVal=true;
+  
+    os[0]=0; os[1]=0;
+    for (int i=0; i<6; i++) {
+      if (!writes.empty() && --delay<0) {
+        delay=0;
+        QueuedWrite& w=writes.front();
+        if (w.addrOrVal) {
+          OPN2_Write(&fm,0x1+((w.addr>>8)<<1),w.val);
+          //printf("write: %x = %.2x\n",w.addr,w.val);
+          lastBusy=0;
+          writes.pop();
+        } else {
+          lastBusy++;
+          if (fm.write_busy==0) {
+            //printf("busycounter: %d\n",lastBusy);
+            OPN2_Write(&fm,0x0+((w.addr>>8)<<1),w.addr);
+            w.addrOrVal=true;
+          }
         }
       }
-    }
+      
+      OPN2_Clock(&fm,o); os[0]+=o[0]; os[1]+=o[1];
+      //OPN2_Write(&fm,0,0);
+      }
     
-    OPN2_Clock(&fm,o); os[0]+=o[0]; os[1]+=o[1];
-    //OPN2_Write(&fm,0,0);
-  }
+    psgClocks+=223722;
+    while (psgClocks>=rate) {
+      psgOut=(psg.acquireOne()*3)>>3;
+      psgClocks-=rate;
+    }
   
-  psgClocks+=223722;
-  while (psgClocks>=rate) {
-    psg.acquire(psgOut,psgOut);
-    psgClocks-=rate;
-    psgOut=(psgOut>>2)+(psgOut>>3);
+    buf[0][h]=(os[0]<<5)+psgOut;
+    buf[1][h]=(os[1]<<5)+psgOut;
   }
-
-  l=(os[0]<<5)+psgOut;
-  r=(os[1]<<5)+psgOut;
 }
 
 void DivPlatformGenesis::tick() {
diff --git a/src/engine/platform/genesis.h b/src/engine/platform/genesis.h
index 62dfcc58f..522c002fb 100644
--- a/src/engine/platform/genesis.h
+++ b/src/engine/platform/genesis.h
@@ -48,7 +48,7 @@ class DivPlatformGenesis: public DivDispatch {
     int toFreq(int freq);
   
   public:
-    void acquire(int& l, int& r);
+    void acquire(short** buf, size_t start, size_t len);
     int dispatch(DivCommand c);
     void tick();
     int init(DivEngine* parent, int channels, int sugRate);
diff --git a/src/engine/platform/nes.cpp b/src/engine/platform/nes.cpp
index 504a6f98b..185d48ae0 100644
--- a/src/engine/platform/nes.cpp
+++ b/src/engine/platform/nes.cpp
@@ -5,30 +5,32 @@
 
 #define FREQ_BASE 3424.0f
 
-void DivPlatformNES::acquire(int& l, int& r) {
-  if (dacSample!=-1) {
-    dacPeriod+=dacRate;
-    if (dacPeriod>=rate) {
-      DivSample* s=parent->song.sample[dacSample];
-      if (s->depth==8) {
-        apu_wr_reg(0x4011,((unsigned char)s->rendData[dacPos++]+0x80)>>1);
-      } else {
-        apu_wr_reg(0x4011,((unsigned short)s->rendData[dacPos++]+0x8000)>>9);
+void DivPlatformNES::acquire(short** buf, size_t start, size_t len) {
+  for (size_t i=start; i<start+len; i++) {
+    if (dacSample!=-1) {
+      dacPeriod+=dacRate;
+      if (dacPeriod>=rate) {
+        DivSample* s=parent->song.sample[dacSample];
+        if (s->depth==8) {
+          apu_wr_reg(0x4011,((unsigned char)s->rendData[dacPos++]+0x80)>>1);
+        } else {
+          apu_wr_reg(0x4011,((unsigned short)s->rendData[dacPos++]+0x8000)>>9);
+        }
+        if (dacPos>=s->rendLength) {
+          dacSample=-1;
+        }
+        dacPeriod-=rate;
       }
-      if (dacPos>=s->rendLength) {
-        dacSample=-1;
-      }
-      dacPeriod-=rate;
     }
-  }
-
-  apu_tick(NULL);
-  apu.odd_cycle=!apu.odd_cycle;
-  if (apu.clocked) {
-    apu.clocked=false;
-    l=(pulse_output()+tnd_output())*30;
-    r=l;
-    //printf("output value: %d\n",l);
+  
+    apu_tick(NULL);
+    apu.odd_cycle=!apu.odd_cycle;
+    if (apu.clocked) {
+      apu.clocked=false;
+      buf[0][i]=(pulse_output()+tnd_output())*30;
+      buf[1][i]=buf[0][i];
+      //printf("output value: %d\n",l);
+    }
   }
 }
 
diff --git a/src/engine/platform/nes.h b/src/engine/platform/nes.h
index eacfc21ea..2e1859ac4 100644
--- a/src/engine/platform/nes.h
+++ b/src/engine/platform/nes.h
@@ -36,7 +36,7 @@ class DivPlatformNES: public DivDispatch {
 
   void updateWave();
   public:
-    void acquire(int& l, int& r);
+    void acquire(short** buf, size_t start, size_t len);
     int dispatch(DivCommand c);
     void tick();
     int init(DivEngine* parent, int channels, int sugRate);
diff --git a/src/engine/platform/pce.cpp b/src/engine/platform/pce.cpp
index 57db7b5d0..ff6f7c602 100644
--- a/src/engine/platform/pce.cpp
+++ b/src/engine/platform/pce.cpp
@@ -13,41 +13,43 @@
 
 #define FREQ_BASE 1712.0f*2
 
-void DivPlatformPCE::acquire(int& l, int& r) {
-  // PCM part
-  for (int i=0; i<6; i++) {
-    if (chan[i].pcm && chan[i].dacSample!=-1) {
-      if (--chan[i].dacPeriod<1) {
-        DivSample* s=parent->song.sample[chan[i].dacSample];
-        chWrite(i,0x07,0);
-        if (s->depth==8) {
-          chWrite(i,0x04,0xdf);
-          chWrite(i,0x06,(((unsigned char)s->rendData[chan[i].dacPos++]+0x80)>>3));
-        } else {
-          chWrite(i,0x04,0xdf);
-          chWrite(i,0x06,(((unsigned short)s->rendData[chan[i].dacPos++]+0x8000)>>11));
+void DivPlatformPCE::acquire(short** buf, size_t start, size_t len) {
+  for (size_t h=start; h<start+len; h++) {
+    // PCM part
+    for (int i=0; i<6; i++) {
+      if (chan[i].pcm && chan[i].dacSample!=-1) {
+        if (--chan[i].dacPeriod<1) {
+          DivSample* s=parent->song.sample[chan[i].dacSample];
+          chWrite(i,0x07,0);
+          if (s->depth==8) {
+            chWrite(i,0x04,0xdf);
+            chWrite(i,0x06,(((unsigned char)s->rendData[chan[i].dacPos++]+0x80)>>3));
+          } else {
+            chWrite(i,0x04,0xdf);
+            chWrite(i,0x06,(((unsigned short)s->rendData[chan[i].dacPos++]+0x8000)>>11));
+          }
+          if (chan[i].dacPos>=s->rendLength) {
+            chan[i].dacSample=-1;
+          }
+          chan[i].dacPeriod=chan[i].dacRate;
         }
-        if (chan[i].dacPos>=s->rendLength) {
-          chan[i].dacSample=-1;
-        }
-        chan[i].dacPeriod=chan[i].dacRate;
       }
     }
-  }
-
-  // PCE part
-  while (!writes.empty()) {
-    QueuedWrite w=writes.front();
-    pce->Write(cycles,w.addr,w.val);
-    writes.pop();
-  }
-  tempL=0; tempR=0;
-  pce->Update(4);
-  pce->ResetTS(0);
   
-  //printf("tempL: %d tempR: %d\n",tempL,tempR);
-  l=tempL;
-  r=tempR;
+    // PCE part
+    while (!writes.empty()) {
+      QueuedWrite w=writes.front();
+      pce->Write(cycles,w.addr,w.val);
+      writes.pop();
+    }
+    tempL=0; tempR=0;
+    pce->Update(4);
+    pce->ResetTS(0);
+    
+    //printf("tempL: %d tempR: %d\n",tempL,tempR);
+    buf[0][h]=tempL;
+    buf[1][h]=tempR;
+  }
 }
 
 void DivPlatformPCE::updateWave(int ch) {
diff --git a/src/engine/platform/pce.h b/src/engine/platform/pce.h
index 689c04955..2775dc0bb 100644
--- a/src/engine/platform/pce.h
+++ b/src/engine/platform/pce.h
@@ -50,7 +50,7 @@ class DivPlatformPCE: public DivDispatch {
   PCE_PSG* pce;
   void updateWave(int ch);
   public:
-    void acquire(int& l, int& r);
+    void acquire(short** buf, size_t start, size_t len);
     int dispatch(DivCommand c);
     void tick();
     int init(DivEngine* parent, int channels, int sugRate);
diff --git a/src/engine/platform/sms.cpp b/src/engine/platform/sms.cpp
index b88415229..99cc2a89f 100644
--- a/src/engine/platform/sms.cpp
+++ b/src/engine/platform/sms.cpp
@@ -2,11 +2,15 @@
 #include "../engine.h"
 #include <math.h>
 
-void DivPlatformSMS::acquire(int& l, int& r) {
+void DivPlatformSMS::acquire(short** buf, size_t start, size_t len) {
+  sn->sound_stream_update(buf[0]+start,len);
+  memcpy(buf[1]+start,buf[0]+start,sizeof(short)*len);
+}
+
+int DivPlatformSMS::acquireOne() {
   short v;
   sn->sound_stream_update(&v,1);
-  l=v;
-  r=v;
+  return v;
 }
 
 void DivPlatformSMS::tick() {
diff --git a/src/engine/platform/sms.h b/src/engine/platform/sms.h
index c6e6f261e..a390ca909 100644
--- a/src/engine/platform/sms.h
+++ b/src/engine/platform/sms.h
@@ -30,7 +30,8 @@ class DivPlatformSMS: public DivDispatch {
   bool updateSNMode;
   sn76496_device* sn;
   public:
-    void acquire(int& l, int& r);
+    int acquireOne();
+    void acquire(short** buf, size_t start, size_t len);
     int dispatch(DivCommand c);
     void tick();
     int init(DivEngine* parent, int channels, int sugRate);
diff --git a/src/engine/playback.cpp b/src/engine/playback.cpp
index 33bf419b6..d98339b13 100644
--- a/src/engine/playback.cpp
+++ b/src/engine/playback.cpp
@@ -604,15 +604,39 @@ void DivEngine::nextTick() {
 
 void DivEngine::nextBuf(float** in, float** out, int inChans, int outChans, unsigned int size) {
   size_t runtotal=blip_clocks_needed(bb[0],size);
-  for (size_t i=0; i<runtotal; i++) {
-    if (--cycles<=0) {
-      nextTick();
-    }
-    dispatch->acquire(temp[0],temp[1]);
 
+  if (runtotal>bbInLen) {
+    delete bbIn[0];
+    delete bbIn[1];
+    bbIn[0]=new short[runtotal+256];
+    bbIn[1]=new short[runtotal+256];
+    bbInLen=runtotal+256;
+  }
+
+  size_t runLeft=runtotal;
+  size_t runPos=0;
+  while (runLeft) {
+    if (runLeft>=cycles) {
+      runLeft-=cycles;
+      dispatch->acquire(bbIn,runPos,cycles);
+      runPos+=cycles;
+      nextTick();
+    } else {
+      dispatch->acquire(bbIn,runPos,runLeft);
+      cycles-=runLeft;
+      break;
+    }
+  }
+
+  for (size_t i=0; i<runtotal; i++) {
+    temp[0]=bbIn[0][i];
     blip_add_delta(bb[0],i,temp[0]-prevSample[0]);
-    blip_add_delta(bb[1],i,temp[1]-prevSample[1]);
     prevSample[0]=temp[0];
+  }
+
+  for (size_t i=0; i<runtotal; i++) {
+    temp[1]=bbIn[1][i];
+    blip_add_delta(bb[1],i,temp[1]-prevSample[1]);
     prevSample[1]=temp[1];
   }