From b3b50bdb66c03fde51eca5fbc977e4dc5ed9fb9d Mon Sep 17 00:00:00 2001
From: tildearrow <me@tildearrow.org>
Date: Mon, 3 Mar 2025 01:45:42 -0500
Subject: [PATCH] ESFM: optimize osc buf

---
 src/engine/dispatch.h        | 24 ++++++++++++------
 src/engine/platform/esfm.cpp | 47 ++++++++++++++++++++++++++++++++++++
 src/engine/platform/esfm.h   |  3 +++
 src/gui/debugWindow.cpp      |  3 +--
 4 files changed, 68 insertions(+), 9 deletions(-)

diff --git a/src/engine/dispatch.h b/src/engine/dispatch.h
index 69f49c12b..e633ac822 100644
--- a/src/engine/dispatch.h
+++ b/src/engine/dispatch.h
@@ -427,18 +427,20 @@ struct DivSamplePos {
 constexpr size_t OSCBUF_PREC=(sizeof(size_t)>=8)?16:16;
 constexpr size_t OSCBUF_MASK=(UINTMAX_C(1)<<OSCBUF_PREC)-1;
 
+#define putSampleIKnowWhatIAmDoing(_ob,_pos,_val) \
+  _ob->data[_pos]=_val;
+
 // the actual output of all DivDispatchOscBuffer instanced runs at 65536Hz.
 struct DivDispatchOscBuffer {
   size_t rate;
   size_t rateMul;
   unsigned int needle;
   unsigned short readNeedle;
-  unsigned short followNeedle;
-  unsigned short lastSample;
+  //unsigned short lastSample;
   bool follow;
   short data[65536];
 
-  inline void putSample(size_t pos, short val) {
+  inline void putSample(const size_t pos, const short val) {
     unsigned short realPos=((needle+pos*rateMul)>>OSCBUF_PREC);
     if (val==-1) {
       data[realPos]=0xfffe;
@@ -447,6 +449,16 @@ struct DivDispatchOscBuffer {
     //lastSample=val;
     data[realPos]=val;
   }
+  /*
+  inline void putSampleIKnowWhatIAmDoing(const unsigned short pos, const short val) {
+    //unsigned short realPos=((needle+pos*rateMul)>>OSCBUF_PREC);
+    if (val==-1) {
+      data[pos]=0xfffe;
+      return;
+    }
+    //lastSample=val;
+    data[pos]=val;
+  }*/
   inline void begin(unsigned short len) {
     size_t calc=(len*rateMul);
     unsigned short start=needle>>16;
@@ -473,8 +485,7 @@ struct DivDispatchOscBuffer {
     memset(data,-1,65536*sizeof(short));
     needle=0;
     readNeedle=0;
-    followNeedle=0;
-    lastSample=0;
+    //lastSample=0;
   }
   void setRate(unsigned int r) {
     double rateMulD=65536.0/(double)r;
@@ -487,8 +498,7 @@ struct DivDispatchOscBuffer {
     rateMul(UINTMAX_C(1)<<OSCBUF_PREC),
     needle(0),
     readNeedle(0),
-    followNeedle(0),
-    lastSample(0),
+    //lastSample(0),
     follow(true) {
     memset(data,-1,65536*sizeof(short));
   }
diff --git a/src/engine/platform/esfm.cpp b/src/engine/platform/esfm.cpp
index bbce9d665..d05ebd188 100644
--- a/src/engine/platform/esfm.cpp
+++ b/src/engine/platform/esfm.cpp
@@ -66,6 +66,46 @@ void DivPlatformESFM::acquire(short** buf, size_t len) {
   }
 }
 
+void DivPlatformESFM::acquireDirect(blip_buffer_t** bb, size_t off, size_t len) {
+  thread_local short o[2];
+  unsigned int sharedNeedlePos=oscBuf[0]->needle;
+  for (int i=0; i<18; i++) {
+    oscBuf[i]->begin(len);
+  }
+  size_t pos=off;
+  for (size_t h=0; h<len; h++) {
+    if (!writes.empty()) {
+      QueuedWrite& w=writes.front();
+      ESFM_write_reg_buffered_fast(&chip,w.addr,w.val);
+      if (w.addr<ESFM_REG_POOL_SIZE) {
+        regPool[w.addr]=w.val;
+      }
+      writes.pop();
+    }
+
+    ESFM_generate(&chip,o);
+    const unsigned int shiftedNeedlePos=sharedNeedlePos>>OSCBUF_PREC;
+    for (int c=0; c<18; c++) {
+      putSampleIKnowWhatIAmDoing(oscBuf[c],shiftedNeedlePos,ESFM_get_channel_output_native(&chip,c));
+    }
+    sharedNeedlePos+=oscBuf[0]->rateMul;
+
+    if (o[0]!=oldOut[0]) {
+      blip_add_delta(bb[0],pos,oldOut[0]-o[0]);
+      oldOut[0]=o[0];
+    }
+    if (o[1]!=oldOut[1]) {
+      blip_add_delta(bb[1],pos,oldOut[1]-o[1]);
+      oldOut[1]=o[1];
+    }
+
+    pos++;
+  }
+  for (int i=0; i<18; i++) {
+    oscBuf[i]->end(len);
+  }
+}
+
 void DivPlatformESFM::tick(bool sysTick) {
   for (int i=0; i<18; i++) {
     chan[i].std.next();
@@ -1024,6 +1064,9 @@ void DivPlatformESFM::reset() {
     chan[i].vol=0x3f;
     chan[i].outVol=0x3f;
   }
+
+  oldOut[0]=0;
+  oldOut[1]=0;
 }
 
 int DivPlatformESFM::getOutputCount() {
@@ -1038,6 +1081,10 @@ bool DivPlatformESFM::keyOffAffectsPorta(int ch) {
   return false;
 }
 
+bool DivPlatformESFM::hasAcquireDirect() {
+  return true;
+}
+
 bool DivPlatformESFM::getLegacyAlwaysSetVolume() {
   return false;
 }
diff --git a/src/engine/platform/esfm.h b/src/engine/platform/esfm.h
index e9d5863b5..df7fcc068 100644
--- a/src/engine/platform/esfm.h
+++ b/src/engine/platform/esfm.h
@@ -120,6 +120,7 @@ class DivPlatformESFM: public DivDispatch {
     };
   FixedQueue<QueuedWrite,2048> writes;
   esfm_chip chip;
+  short oldOut[2];
   bool isFast;
 
   unsigned char regPool[ESFM_REG_POOL_SIZE];
@@ -180,6 +181,7 @@ class DivPlatformESFM: public DivDispatch {
 
   public:
     void acquire(short** buf, size_t len);
+    void acquireDirect(blip_buffer_t** bb, size_t off, size_t len);
     int dispatch(DivCommand c);
     void* getChanState(int chan);
     DivMacroInt* getChanMacroInt(int ch);
@@ -194,6 +196,7 @@ class DivPlatformESFM: public DivDispatch {
     void muteChannel(int ch, bool mute);
     bool keyOffAffectsArp(int ch);
     bool keyOffAffectsPorta(int ch);
+    bool hasAcquireDirect();
     bool getLegacyAlwaysSetVolume();
     void toggleRegisterDump(bool enable);
     void notifyInsChange(int ins);
diff --git a/src/gui/debugWindow.cpp b/src/gui/debugWindow.cpp
index 15f17f851..e6c8673a3 100644
--- a/src/gui/debugWindow.cpp
+++ b/src/gui/debugWindow.cpp
@@ -276,10 +276,9 @@ void FurnaceGUI::drawDebug() {
                 ImGui::Checkbox(fmt::sprintf("##%d_OSCFollow_%d",i,c).c_str(),&oscBuf->follow);
                 // address
                 ImGui::TableNextColumn();
-                int needle=oscBuf->follow?oscBuf->needle:oscBuf->followNeedle;
+                int needle=oscBuf->needle;
                 ImGui::BeginDisabled(oscBuf->follow);
                 if (ImGui::InputInt(fmt::sprintf("##%d_OSCFollowNeedle_%d",i,c).c_str(),&needle,1,100)) {
-                  oscBuf->followNeedle=MIN(MAX(needle,0),65535);
                 }
                 ImGui::EndDisabled();
                 // data