GA20: acquireDirect optimizations, part 1

buggy, slow and glitchy
output is good but chan osc is now broken

from 1.5 to 0.4 seconds

src/engine/platform/ga20.cpp

@@ -52,39 +52,23 @@ inline void DivPlatformGA20::chWrite(unsigned char ch, unsigned int addr, unsign
 }
 
 void DivPlatformGA20::acquire(short** buf, size_t len) {
-  if (ga20BufLen<len) {
-    ga20BufLen=len;
-    for (int i=0; i<4; i++) {
-      delete[] ga20Buf[i];
-      ga20Buf[i]=new short[ga20BufLen];
-    }
-  }
+  thread_local short ga20Buf[4];
 
   for (int i=0; i<4; i++) {
     oscBuf[i]->begin(len);
   }
 
   for (size_t h=0; h<len; h++) {
-    if ((--delay)<=0) {
-      delay=MAX(0,delay);
     if (!writes.empty()) {
       QueuedWrite& w=writes.front();
       ga20.write(w.addr,w.val);
       regPool[w.addr]=w.val;
       writes.pop();
-        delay=1;
     }
-    }
-    short *buffer[4]={
-      &ga20Buf[0][h],
-      &ga20Buf[1][h],
-      &ga20Buf[2][h],
-      &ga20Buf[3][h]
-    };
-    ga20.sound_stream_update(buffer,1);
-    buf[0][h]=(signed int)(ga20Buf[0][h]+ga20Buf[1][h]+ga20Buf[2][h]+ga20Buf[3][h])>>2;
+    ga20.sound_stream_update(ga20Buf,1);
+    buf[0][h]=(signed int)(ga20Buf[0]+ga20Buf[1]+ga20Buf[2]+ga20Buf[3])>>2;
     for (int i=0; i<4; i++) {
-      oscBuf[i]->putSample(h,ga20Buf[i][h]>>1);
+      oscBuf[i]->putSample(h,ga20Buf[i]>>1);
     }
   }
 
@@ -94,43 +78,42 @@ void DivPlatformGA20::acquire(short** buf, size_t len) {
 }
 
 void DivPlatformGA20::acquireDirect(blip_buffer_t** bb, size_t off, size_t len) {
-  if (ga20BufLen<len) {
-    ga20BufLen=len;
-    for (int i=0; i<4; i++) {
-      delete[] ga20Buf[i];
-      ga20Buf[i]=new short[ga20BufLen];
-    }
-  }
+  thread_local short ga20Buf[4];
 
   for (int i=0; i<4; i++) {
     oscBuf[i]->begin(len);
   }
 
   for (size_t h=off; h<off+len; h++) {
-    if ((--delay)<=0) {
-      delay=MAX(0,delay);
+    size_t advance=0;
     if (!writes.empty()) {
       QueuedWrite& w=writes.front();
       ga20.write(w.addr,w.val);
       regPool[w.addr]=w.val;
       writes.pop();
-        delay=1;
+    } else {
+      // heuristic
+      advance=off+len-h-1;
+
+      for (int i=0; i<4; i++) {
+        if (!ga20.m_channel[i].play) continue;
+        if (ga20.m_channel[i].hot) {
+          advance=0;
+          break;
+        }
+        const size_t newAdvance=ga20.m_channel[i].counter-ga20.m_channel[i].rate-1;
+        if (newAdvance<advance) advance=newAdvance;
       }
     }
-    short *buffer[4]={
-      &ga20Buf[0][h],
-      &ga20Buf[1][h],
-      &ga20Buf[2][h],
-      &ga20Buf[3][h]
-    };
-    ga20.sound_stream_update(buffer,1);
-    const int out=(signed int)(ga20Buf[0][h]+ga20Buf[1][h]+ga20Buf[2][h]+ga20Buf[3][h])>>2;
+    ga20.sound_stream_update(ga20Buf,advance+1);
+    h+=advance;
+    const int out=(signed int)(ga20Buf[0]+ga20Buf[1]+ga20Buf[2]+ga20Buf[3])>>2;
     if (out!=oldOut) {
       blip_add_delta(bb[0],h,out-oldOut);
       oldOut=out;
     }
     for (int i=0; i<4; i++) {
-      oscBuf[i]->putSample(h,ga20Buf[i][h]>>1);
+      oscBuf[i]->putSample(h,ga20Buf[i]>>1);
     }
   }
 
@@ -420,7 +403,6 @@ void DivPlatformGA20::reset() {
   writes.clear();
   memset(regPool,0,32);
   ga20.device_reset();
-  delay=0;
   oldOut=0;
   for (int i=0; i<4; i++) {
     chan[i]=DivPlatformGA20::Channel();
@@ -561,7 +543,6 @@ int DivPlatformGA20::init(DivEngine* p, int channels, int sugRate, const DivConf
   }
   sampleMem=new unsigned char[getSampleMemCapacity()];
   sampleMemLen=0;
-  delay=0;
   setFlags(flags);
   ga20BufLen=65536;
   for (int i=0; i<4; i++) ga20Buf[i]=new short[ga20BufLen];

src/engine/platform/ga20.h

@@ -58,7 +58,6 @@ class DivPlatformGA20: public DivDispatch, public iremga20_intf {
   unsigned int sampleOffGA20[256];
   bool sampleLoaded[256];
 
-  int delay;
   int oldOut;
 
   short* ga20Buf[4];

src/engine/platform/sound/ga20/iremga20.cpp

@@ -89,9 +89,7 @@ void iremga20_device::device_reset()
 //  sound_stream_update - handle a stream update
 //-------------------------------------------------
 
-void iremga20_device::sound_stream_update(short** outputs, int len)
-{
-	for (int i = 0; i < len; i++)
+void iremga20_device::sound_stream_update(short* outputs, int len)
 {
 	for (int j = 0; j < 4; j++)
 	{
@@ -106,7 +104,7 @@ void iremga20_device::sound_stream_update(short** outputs, int len)
     ch.hot=false;
 				  ch.output = ch.mute ? 0 : (ch.sample - 0x80) * (s32)ch.volume;
   }
-					ch.counter--;
+				ch.counter-=len;
 				if (ch.counter <= ch.rate)
 				{
 					ch.pos++;
@@ -116,8 +114,7 @@ void iremga20_device::sound_stream_update(short** outputs, int len)
 				}
 			}
 		}
-			outputs[j][i] = ch.output;
-		}
+		outputs[j] = ch.output;
 	}
 }
 

src/engine/platform/sound/ga20/iremga20.h

@@ -6,7 +6,7 @@
 
 	DISCLAIMER
 	- This file is modified for suitable in furnace.
-	- modified by cam900
+	- modified by cam900 and tildearrow
 
 *********************************************************/
 #ifndef MAME_SOUND_IREMGA20_H
@@ -50,10 +50,9 @@ public:
 	void device_reset();
 
 	// sound stream update overrides
-	void sound_stream_update(short** outputs, int len);
+	void sound_stream_update(short* outputs, int len);
   
 
-private:
 	struct channel_def
 	{
 		channel_def() :

src/gui/debug.cpp

@@ -521,7 +521,6 @@ void putDispatchChip(void* data, int type) {
       DivPlatformGA20* ch=(DivPlatformGA20*)data;
       ImGui::Text("> GA20");
       COMMON_CHIP_DEBUG;
-      ImGui::Text("- delay: %.2x",ch->delay);
       COMMON_CHIP_DEBUG_BOOL;
       break;
     }