diff --git a/doc/7-systems/pokey.md b/doc/7-systems/pokey.md
index 8bbb04b78..257b85d53 100644
--- a/doc/7-systems/pokey.md
+++ b/doc/7-systems/pokey.md
@@ -33,7 +33,7 @@ a sound and input chip developed by Atari for their 8-bit computers (Atari 400,
     - use for PWM effects (not automatic!).
   - bit 0: 15KHz mode.
 - `12xx`: **toggle two-tone mode.**
-  - when enabled, channel 2 modulates channel 1. I don't know how, but it does.
+  - when enabled, channel 2 modulates channel 1 in an oscillator sync-like manner.
   - only on ASAP core.
 
 ## info
diff --git a/src/engine/brrUtils.c b/src/engine/brrUtils.c
index 38073fb4b..ff62e5641 100644
--- a/src/engine/brrUtils.c
+++ b/src/engine/brrUtils.c
@@ -256,7 +256,7 @@ long brrEncode(short* buf, unsigned char* out, long len, long loopStart, unsigne
     total+=9;
   }
   // encode loop block
-  if (loopStart>=0) {
+  if (loopStart>=0 && loopStart<len) {
     long p=loopStart;
     for (int i=0; i<17; i++) {
       if (p>=len) {
diff --git a/src/engine/cmdStream.cpp b/src/engine/cmdStream.cpp
index f1605426e..21ed84b5c 100644
--- a/src/engine/cmdStream.cpp
+++ b/src/engine/cmdStream.cpp
@@ -682,6 +682,11 @@ bool DivCSPlayer::init() {
       chan[i].readPos=chan[i].startPos;
     }
   }
+  
+  // read stack sizes
+  for (unsigned int i=0; i<fileChans; i++) {
+    chan[i].callStackSize=stream.readC();
+  }
 
   // initialize state
   for (int i=0; i<e->getTotalChannelCount(); i++) {
diff --git a/src/engine/cmdStream.h b/src/engine/cmdStream.h
index 326134dbd..459a49d62 100644
--- a/src/engine/cmdStream.h
+++ b/src/engine/cmdStream.h
@@ -41,7 +41,7 @@ struct DivCSChannelState {
   unsigned char arp, arpStage, arpTicks;
 
   unsigned int callStack[DIV_MAX_CSSTACK];
-  unsigned char callStackPos;
+  unsigned char callStackPos, callStackSize;
 
   unsigned int trace[DIV_MAX_CSTRACE];
   unsigned char tracePos;
@@ -67,6 +67,7 @@ struct DivCSChannelState {
     arpStage(0),
     arpTicks(0),
     callStackPos(0),
+    callStackSize(0),
     tracePos(0) {
     for (int i=0; i<DIV_MAX_CSTRACE; i++) {
       trace[i]=0;
diff --git a/src/engine/platform/mmc5.cpp b/src/engine/platform/mmc5.cpp
index 672891bb4..fcb8dbdc6 100644
--- a/src/engine/platform/mmc5.cpp
+++ b/src/engine/platform/mmc5.cpp
@@ -377,6 +377,10 @@ void DivPlatformMMC5::forceIns() {
   for (int i=0; i<3; i++) {
     chan[i].insChanged=true;
     chan[i].prevFreq=65535;
+    if (i<2) {
+      // TODO: implement envelope mode
+      rWrite(0x5000+i*4,(0x30)|(chan[i].active?chan[i].outVol:0)|((chan[i].duty&3)<<6));
+    }
   }
 }
 
@@ -429,6 +433,10 @@ void DivPlatformMMC5::reset() {
 
   rWrite(0x5015,0x03);
   rWrite(0x5010,0x00);
+
+  for (int i=0; i<2; i++) {
+    rWrite(0x5000+i*4,(0x30)|0|((chan[i].duty&3)<<6));
+  }
 }
 
 bool DivPlatformMMC5::keyOffAffectsArp(int ch) {
diff --git a/src/engine/platform/opl.cpp b/src/engine/platform/opl.cpp
index 762bde308..5a8f81a4e 100644
--- a/src/engine/platform/opl.cpp
+++ b/src/engine/platform/opl.cpp
@@ -37,6 +37,9 @@
 #define PCM_CHECK(ch) ((chipType==4) && (ch>=pcmChanOffs))
 #define PCM_REG(ch) (ch-pcmChanOffs)
 
+// check if PCM in RAM (and size is <= 2MB) - 4MB uses whole sample area
+#define PCM_IN_RAM (ramSize<=0x200000)
+
 // N = invalid
 #define N 255
 
@@ -1466,7 +1469,7 @@ void DivPlatformOPL::tick(bool sysTick) {
         ctrl|=(chan[i].active?0x80:0)|(chan[i].damp?0x40:0)|(chan[i].lfoReset?0x20:0)|(chan[i].ch?0x10:0)|(isMuted[i]?8:(chan[i].pan&0xf));
         int waveNum=chan[i].sample;
         if (waveNum>=0) {
-          if (ramSize<=0x200000) {
+          if (PCM_IN_RAM) {
             waveNum=CLAMP(waveNum,0,0x7f)|0x180;
           }
           if (chan[i].keyOn) {
@@ -2941,7 +2944,7 @@ void DivPlatformOPL::reset() {
     if (chipType==4) {
       immWrite(0x105,3);
       // Reset wavetable header
-      immWrite(0x202,(ramSize<=0x200000)?0x10:0x00);
+      immWrite(0x202,PCM_IN_RAM?0x10:0x00);
       // initialize mixer volume
       fmMixL=7;
       fmMixR=7;
@@ -3209,7 +3212,7 @@ void DivPlatformOPL::setFlags(const DivConfig& flags) {
       pcm.setClockFrequency(chipClock);
       rate=chipClock/768;
       chipRateBase=chipClock/684;
-      immWrite(0x202,(ramSize<=0x200000)?0x10:0x00);
+      immWrite(0x202,PCM_IN_RAM?0x10:0x00);
       break;
     case 759:
       rate=48000;
@@ -3232,7 +3235,7 @@ const void* DivPlatformOPL::getSampleMem(int index) {
 
 size_t DivPlatformOPL::getSampleMemCapacity(int index) {
   return (index==0 && pcmChanOffs>=0)?
-          ((ramSize<=0x200000)?0x200000+ramSize:ramSize):
+          (PCM_IN_RAM?0x200000+ramSize:ramSize):
           ((index==0 && adpcmChan>=0)?262144:0);
 }
 
@@ -3269,10 +3272,16 @@ void DivPlatformOPL::renderSamples(int sysID) {
   memCompo.name="Sample Memory";
 
   if (pcmChanOffs>=0) { // OPL4 PCM
-    size_t memPos=((ramSize<=0x200000)?0x200600:0x1800);
-    const int maxSample=(ramSize<=0x200000)?127:511;
+    size_t memPos=(PCM_IN_RAM?0x200600:0x1800);
+    const int maxSample=PCM_IN_RAM?128:512;
     int sampleCount=parent->song.sampleLen;
-    if (sampleCount>maxSample) sampleCount=maxSample;
+    if (sampleCount>maxSample) {
+      // mark the rest as unavailable
+      for (int i=maxSample; i<sampleCount; i++) {
+        sampleLoaded[i]=false;
+      }
+      sampleCount=maxSample;
+    }
     for (int i=0; i<sampleCount; i++) {
       DivSample* s=parent->song.sample[i];
       if (!s->renderOn[0][sysID]) {
@@ -3335,7 +3344,7 @@ void DivPlatformOPL::renderSamples(int sysID) {
     // instrument table
     for (int i=0; i<sampleCount; i++) {
       DivSample* s=parent->song.sample[i];
-      unsigned int insAddr=(i*12)+((ramSize<=0x200000)?0x200000:0);
+      unsigned int insAddr=(i*12)+(PCM_IN_RAM?0x200000:0);
       unsigned char bitDepth;
       int endPos=CLAMP(s->isLoopable()?s->loopEnd:(s->samples+1),1,0x10000);
       int loop=s->isLoopable()?CLAMP(s->loopStart,0,endPos-2):(endPos-2);
@@ -3366,12 +3375,12 @@ void DivPlatformOPL::renderSamples(int sysID) {
       pcmMem[6+insAddr]=(~(endPos-1))&0xff;
       // on MultiPCM this consists of instrument params, but on OPL4 this is not used
       pcmMem[7+insAddr]=0; // LFO, VIB
-      pcmMem[8+insAddr]=(0xf << 4) | (0xf << 0); // AR, D1R
+      pcmMem[8+insAddr]=(0xf<<4)|(0xf<<0); // AR, D1R
       pcmMem[9+insAddr]=0; // DL, D2R
-      pcmMem[10+insAddr]=(0xf << 4) | (0xf << 0); // RC, RR
+      pcmMem[10+insAddr]=(0xf<<4)|(0xf<<0); // RC, RR
       pcmMem[11+insAddr]=0; // AM
     }
-    if (ramSize<=0x200000) {
+    if (PCM_IN_RAM) {
       memCompo.entries.push_back(DivMemoryEntry(DIV_MEMORY_RESERVED,"ROM data",0,0,0x200000));
     }
 
diff --git a/src/engine/sample.cpp b/src/engine/sample.cpp
index 2f7d50617..054d2bc4e 100644
--- a/src/engine/sample.cpp
+++ b/src/engine/sample.cpp
@@ -1479,7 +1479,8 @@ void DivSample::render(unsigned int formatMask) {
     }
   }
   if (NOT_IN_FORMAT(DIV_SAMPLE_DEPTH_BRR)) { // BRR
-    int sampleCount=loop?loopEnd:samples;
+    int sampleCount=isLoopable()?loopEnd:samples;
+    if (sampleCount>(int)samples) sampleCount=samples;
     if (!initInternal(DIV_SAMPLE_DEPTH_BRR,sampleCount)) return;
     brrEncode(data16,dataBRR,sampleCount,loop?loopStart:-1,brrEmphasis,brrNoFilter);
   }
diff --git a/src/gui/csPlayer.cpp b/src/gui/csPlayer.cpp
index c84494dda..e438a4d60 100644
--- a/src/gui/csPlayer.cpp
+++ b/src/gui/csPlayer.cpp
@@ -525,6 +525,72 @@ void FurnaceGUI::drawCSPlayer() {
           ImGui::PopFont();
           ImGui::EndTabItem();
         }
+        if (ImGui::BeginTabItem(_("Data Access Visualizer"))) {
+          ImGui::Text("%d bytes",(int)cs->getDataLen());
+
+          ImGui::PushFont(patFont);
+          if (ImGui::BeginTable("CSHexPos",chans,ImGuiTableFlags_SizingStretchSame)) {
+            ImGui::TableNextRow();
+            for (int i=0; i<chans; i++) {
+              ImGui::TableNextColumn();
+              ImGui::Text("%d",i);
+            }
+            ImGui::TableNextRow();
+            for (int i=0; i<chans; i++) {
+              DivCSChannelState* state=cs->getChanState(i);
+              ImGui::TableNextColumn();
+              ImGui::Text("$%.4x",state->readPos);
+            }
+            ImGui::EndTable();
+          }
+          ImGui::PopFont();
+
+          if (csTex==NULL || !rend->isTextureValid(csTex)) {
+            logD("recreating command stream data texture.");
+            csTex=rend->createTexture(true,256,256,false,GUI_TEXFORMAT_ABGR32);
+            if (csTex==NULL) {
+              logE("error while creating command stream data texture! %s",SDL_GetError());
+            }
+          }
+          if (csTex!=NULL) {
+            unsigned int* dataT=NULL;
+            int pitch=0;
+            if (!rend->lockTexture(csTex,(void**)&dataT,&pitch)) {
+              logE("error while locking command stream data texture! %s",SDL_GetError());
+            } else {
+              unsigned short* accessTS=cs->getDataAccess();
+              unsigned int csTick=cs->getCurTick();
+              const float fadeTime=64.0f;
+              size_t bufSize=cs->getDataLen();
+              if (bufSize>65536) bufSize=65536;
+
+              for (size_t i=0; i<bufSize; i++) {
+                float cellAlpha=(float)(fadeTime-(((short)(csTick&0xffff))-(short)accessTS[i]))/fadeTime;
+                if (cellAlpha>0.0f) {
+                  dataT[i]=ImGui::GetColorU32(ImGuiCol_HeaderActive,cellAlpha);
+                } else {
+                  dataT[i]=0;
+                }
+              }
+              for (size_t i=bufSize; i<65536; i++) {
+                dataT[i]=0;
+              }
+
+              for (int i=0; i<e->getTotalChannelCount(); i++) {
+                unsigned int pos=cs->getChanState(i)->readPos;
+                if (pos<65536) {
+                  ImVec4 col=ImVec4(1.0f,1.0f,1.0f,1.0f);
+                  ImGui::ColorConvertHSVtoRGB((float)i/(float)e->getTotalChannelCount(),0.8f,1.0f,col.x,col.y,col.z);
+                  dataT[pos]=ImGui::GetColorU32(col);
+                }
+              }
+              rend->unlockTexture(csTex);
+            }
+
+            ImGui::Image(rend->getTextureID(csTex),ImVec2(768.0*dpiScale,768.0*dpiScale));
+          }
+          ImGui::EndTabItem();
+        }
         if (ImGui::BeginTabItem(_("Stream Info"))) {
           ImGui::Text("%d bytes",(int)cs->getDataLen());
           ImGui::Text("%u channels",cs->getFileChans());
@@ -538,6 +604,11 @@ void FurnaceGUI::drawCSPlayer() {
             ImGui::SameLine();
             ImGui::Text("%d",cs->getFastCmds()[i]);
           }
+          ImGui::Text("stack sizes:");
+          for (unsigned int i=0; i<cs->getFileChans(); i++) {
+            ImGui::SameLine();
+            ImGui::Text("%d",cs->getChanState(i)->callStackSize);
+          }
           ImGui::Text("ticks: %u",cs->getCurTick());
           ImGui::EndTabItem();
         }
diff --git a/src/gui/gui.h b/src/gui/gui.h
index c2a4c818d..12d31de8b 100644
--- a/src/gui/gui.h
+++ b/src/gui/gui.h
@@ -1668,6 +1668,8 @@ class FurnaceGUI {
   int sampleTexW, sampleTexH;
   bool updateSampleTex;
 
+  FurnaceGUITexture* csTex;
+
   String workingDir, fileName, clipboard, warnString, errorString, lastError, curFileName, nextFile, sysSearchQuery, newSongQuery, paletteQuery, sampleBankSearchQuery;
   String workingDirSong, workingDirIns, workingDirWave, workingDirSample, workingDirAudioExport;
   String workingDirVGMExport, workingDirROMExport;
diff --git a/src/gui/sampleEdit.cpp b/src/gui/sampleEdit.cpp
index 240d0c8bd..e71f7f6b1 100644
--- a/src/gui/sampleEdit.cpp
+++ b/src/gui/sampleEdit.cpp
@@ -2318,6 +2318,32 @@ void FurnaceGUI::drawSampleEdit() {
               }
             }
           }
+          if (displayLoopHintsNDSA) {
+            if (sampleZoom<0.5) {
+              for (int i=0; i<(int)(sampleZoom*avail.x); i++) {
+                if (((i+samplePos)&7)==0) {
+                  ImVec2 p1=ImVec2(rectMin.x+((float)i/sampleZoom),rectMin.y);
+                  ImVec2 p2=p1;
+                  p2.y=rectMax.y;
+
+                  dl->AddLine(p1,p2,ImGui::GetColorU32(uiColors[GUI_COLOR_SAMPLE_LOOP_HINT]));
+                }
+              }
+            }
+          }
+          if (displayLoopHintsNDS8) {
+            if (sampleZoom<0.375) {
+              for (int i=0; i<(int)(sampleZoom*avail.x); i++) {
+                if (((i+samplePos)&3)==0) {
+                  ImVec2 p1=ImVec2(rectMin.x+((float)i/sampleZoom),rectMin.y);
+                  ImVec2 p2=p1;
+                  p2.y=rectMax.y;
+
+                  dl->AddLine(p1,p2,ImGui::GetColorU32(uiColors[GUI_COLOR_SAMPLE_LOOP_HINT]));
+                }
+              }
+            }
+          }
           if (displayLoopHintsAmiga) {
             if (sampleZoom<0.25) {
               for (int i=0; i<(int)(sampleZoom*avail.x); i++) {
diff --git a/src/gui/tutorial.cpp b/src/gui/tutorial.cpp
index 9b238be85..e06c52f6d 100644
--- a/src/gui/tutorial.cpp
+++ b/src/gui/tutorial.cpp
@@ -1313,7 +1313,7 @@ void FurnaceCV::buildStage(int which) {
     curStage=NULL;
   }
 
-  if (which>19 || which==4 || which==7 || which==9 || which==11 || which==13 || which==16 || which==17) {
+  if (which>18 || which==4 || which==7 || which==9 || which==11 || which==13 || which==16 || which==17) {
     stageWidth=80;
     stageHeight=56;
   } else {
@@ -1346,26 +1346,7 @@ void FurnaceCV::buildStage(int which) {
   memset(busy,0,28*40*sizeof(bool));
 
   // special stages
-  if ((which%10)==9) {
-    // vortex
-    for (int i=0; i<20+(which>>2); i++) {
-      int tries=0;
-      while (tries<20) {
-        int x=rand()%(stageWidth>>1);
-        int y=rand()%(stageHeight>>1);
-        int finalX=x<<4;
-        int finalY=y<<4;
-        if (busy[y][x]) {
-          tries++;
-          continue;
-        }
-        createObject<FurnaceCVEnemyVortex>(finalX,finalY);
-        createObject<FurnaceCVFurBallMedium>(finalX-4,finalY-4);
-        busy[y][x]=true;
-        break;
-      }
-    }
-  } else if ((which%10)==19) {
+  if ((which%20)==19) {
     for (int i=0; i<20+(which>>2); i++) {
       int tries=0;
       while (tries<20) {
@@ -1387,6 +1368,25 @@ void FurnaceCV::buildStage(int which) {
         break;
       }
     }
+  } else if ((which%10)==9) {
+    // vortex
+    for (int i=0; i<20+(which>>2); i++) {
+      int tries=0;
+      while (tries<20) {
+        int x=rand()%(stageWidth>>1);
+        int y=rand()%(stageHeight>>1);
+        int finalX=x<<4;
+        int finalY=y<<4;
+        if (busy[y][x]) {
+          tries++;
+          continue;
+        }
+        createObject<FurnaceCVEnemyVortex>(finalX,finalY);
+        createObject<FurnaceCVFurBallMedium>(finalX-4,finalY-4);
+        busy[y][x]=true;
+        break;
+      }
+    }
   } else {
     // large
     if (which>=2) for (int i=0; i<(rand()%3)+which-2; i++) {
@@ -1675,6 +1675,7 @@ void FurnaceCV::render(unsigned char joyIn) {
       lives+=lifeBank;
       respawnTime=1;
       lifeBank=0;
+      score=0;
       gameOver=false;
     }