prepare for possible major optimization

by just forwarding an output buffer to the dispatch and begin/length,
the number of calls may be reduced which improves performance.

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();
 

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));
   }

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} {}
 };

src/engine/platform/abstract.cpp

@@ -1,8 +1,6 @@
 #include "../dispatch.h"
 
-void DivDispatch::acquire(int& l, int& r) {
+void DivDispatch::acquire(short** buf, size_t start, size_t len) {
-  l=0;
-  r=0;
 }
 
 void DivDispatch::tick() {

src/engine/platform/c64.cpp

@@ -4,10 +4,12 @@
 
 #define FREQ_BASE 277.0f
 
-void DivPlatformC64::acquire(int& l, int& r) {
+void DivPlatformC64::acquire(short** buf, size_t start, size_t len) {
+  for (size_t i=start; i<start+len; i++) {
     sid.clock();
-  l=sid.output();
+    buf[0][i]=sid.output();
-  r=l;
+    buf[1][i]=buf[0][i];
+  }
 }
 
 void DivPlatformC64::tick() {

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);

src/engine/platform/dummy.cpp

@@ -2,16 +2,17 @@
 #include <stdio.h>
 #include <math.h>
 
-void DivPlatformDummy::acquire(int& l, int& r) {
+void DivPlatformDummy::acquire(short** buf, size_t start, size_t len) {
-  l=0;
+  for (size_t i=start; i<start+len; i++) {
-  for (unsigned char i=0; i<chans; i++) {
+    buf[0][i]=0;
-    if (chan[i].active) {
+    for (unsigned char j=0; j<chans; j++) {
-      l+=((chan[i].pos>=0x8000)?chan[i].vol:-chan[i].vol)*chan[i].amp;
+      if (chan[j].active) {
-      
+        buf[0][i]+=((chan[j].pos>=0x8000)?chan[j].vol:-chan[j].vol)*chan[j].amp;
-      chan[i].pos+=chan[i].freq;
+        chan[j].pos+=chan[j].freq;
       }
     }
-  r=l;
+    buf[1][i]=buf[0][i];
+  }
 }
 
 void DivPlatformDummy::tick() {

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);

src/engine/platform/gb.cpp

@@ -7,10 +7,12 @@
 
 #define FREQ_BASE 8015.85f
 
-void DivPlatformGB::acquire(int& l, int& r) {
+void DivPlatformGB::acquire(short** buf, size_t start, size_t len) {
+  for (int i=start; i<start+len; i++) {
     GB_advance_cycles(gb,16);
-  l=gb->apu_output.final_sample.left<<2;
+    buf[0][i]=gb->apu_output.final_sample.left<<2;
-  r=gb->apu_output.final_sample.right<<2;
+    buf[1][i]=gb->apu_output.final_sample.right<<2;
+  }
 }
 
 void DivPlatformGB::updateWave() {

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);

src/engine/platform/genesis.cpp

@@ -12,10 +12,11 @@ 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];
 
+  for (size_t h=start; h<start+len; h++) {
     if (dacMode && dacSample!=-1) {
       dacPeriod-=6;
       if (dacPeriod<1) {
@@ -58,13 +59,13 @@ void DivPlatformGenesis::acquire(int& l, int& r) {
     
     psgClocks+=223722;
     while (psgClocks>=rate) {
-    psg.acquire(psgOut,psgOut);
+      psgOut=(psg.acquireOne()*3)>>3;
       psgClocks-=rate;
-    psgOut=(psgOut>>2)+(psgOut>>3);
     }
   
-  l=(os[0]<<5)+psgOut;
+    buf[0][h]=(os[0]<<5)+psgOut;
-  r=(os[1]<<5)+psgOut;
+    buf[1][h]=(os[1]<<5)+psgOut;
+  }
 }
 
 void DivPlatformGenesis::tick() {

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);

src/engine/platform/nes.cpp

@@ -5,7 +5,8 @@
 
 #define FREQ_BASE 3424.0f
 
-void DivPlatformNES::acquire(int& l, int& r) {
+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) {
@@ -26,10 +27,11 @@ void DivPlatformNES::acquire(int& l, int& r) {
     apu.odd_cycle=!apu.odd_cycle;
     if (apu.clocked) {
       apu.clocked=false;
-    l=(pulse_output()+tnd_output())*30;
+      buf[0][i]=(pulse_output()+tnd_output())*30;
-    r=l;
+      buf[1][i]=buf[0][i];
       //printf("output value: %d\n",l);
     }
+  }
 }
 
 static int dacRates[6]={

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);

src/engine/platform/pce.cpp

@@ -13,7 +13,8 @@
 
 #define FREQ_BASE 1712.0f*2
 
-void DivPlatformPCE::acquire(int& l, int& r) {
+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) {
@@ -46,8 +47,9 @@ void DivPlatformPCE::acquire(int& l, int& r) {
     pce->ResetTS(0);
     
     //printf("tempL: %d tempR: %d\n",tempL,tempR);
-  l=tempL;
+    buf[0][h]=tempL;
-  r=tempR;
+    buf[1][h]=tempR;
+  }
 }
 
 void DivPlatformPCE::updateWave(int ch) {

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);

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;
+  return v;
-  r=v;
 }
 
 void DivPlatformSMS::tick() {

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);

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];
   }