features: optimized DMA

Run DMA syncs only when needed, instead of running them on
VDP steps. Thanks to this, the optimizeSteps option is no
longer needed, since we can optimize all cases safely.

higan/emulator/scheduler-synchro/thread.cpp

@@ -2,7 +2,7 @@ Thread::~Thread() {
   destroy();
 }
 
-auto Thread::active() const -> bool { return true; } // Todo: what does this mean if threads don't run
+auto Thread::active() const -> bool { return _active; }
 auto Thread::handle() const -> thread_handle_t* { return _handle; }
 auto Thread::frequency() const -> uintmax { return _frequency; }
 auto Thread::scalar() const -> uintmax { return _scalar; }
@@ -69,7 +69,9 @@ auto Thread::synchronize(Thread& thread, P&&... p) -> void {
     //disable synchronization for auxiliary threads during scheduler synchronization.
     //synchronization can begin inside of this while loop.
     if(scheduler.synchronizing()) break;
+    thread._active = true;
     thread.main();
+    thread._active = false;
   }
   thread._handle->parent = _parent;
 

higan/emulator/scheduler-synchro/thread.hpp

@@ -48,6 +48,7 @@ struct Thread {
   uintmax _frequency = 0;
   uintmax _scalar = 0;
   uintmax _clock = 0;
+  bool _active;
 
   friend class Scheduler;
 };

higan/md/interface/interface.cpp

@@ -5,11 +5,7 @@ namespace higan::MegaDrive {
 Interface* interface = nullptr;
 
 auto MegaDriveInterface::configure(string name, double value) -> void {
-  if(name.match("vdp/optimizeSteps")) {
-#if defined(PROFILE_PERFORMANCE)
-    vdp.optimizeSteps = (bool) value;
-#endif
-  } else if(name.match("vdp/skipframe")) {
+  if(name.match("vdp/skipframe")) {
     vdp.isSkipping = (bool) value;
   } else if(name.match("cpu/overclock")) {
     cpu.overclock = value / 7.0;

higan/md/vdp-performance/dma.cpp

@@ -1,19 +1,28 @@
+auto VDP::DMA::poll() -> void {
+  static bool locked = false;
+  if(locked) return;
+  locked = true;
+  if(cpu.active()) cpu.synchronize(apu, vdp);
+  if(apu.active()) apu.synchronize(cpu, vdp);
+  while(run());
+  locked = false;
+}
+
 auto VDP::DMA::run() -> bool {
   if(!io.enable || io.wait) return false;
-
   if(!vdp.io.command.bit(5)) return false;
   if(io.mode <= 1) return load(), true;
   if(io.mode == 2) return fill(), true;
   if(!vdp.io.command.bit(4)) return false;
   if(io.mode == 3) return copy(), true;
-
-  unreachable;
+  return false;
 }
 
 auto VDP::DMA::load() -> void {
   active = 1;
 
-  auto data = cpu.read(1, 1, io.mode.bit(0) << 23 | io.source << 1);
+  auto address = io.mode.bit(0) << 23 | io.source << 1;
+  auto data = cpu.read(1, 1, address);
   vdp.writeDataPort(data);
 
   io.source.bit(0,15)++;
@@ -51,4 +60,4 @@ auto VDP::DMA::copy() -> void {
 auto VDP::DMA::power() -> void {
   active = 0;
   io = {};
-}
+}

higan/md/vdp-performance/io.cpp

@@ -91,6 +91,7 @@ auto VDP::writeDataPort(uint16 data) -> void {
     if(io.address.bit(0)) data = data >> 8 | data << 8;
     vram.write(address, data);
     io.address += io.dataIncrement;
+    dma.poll();
     return;
   }
 
@@ -148,6 +149,7 @@ auto VDP::writeControlPort(uint16 data) -> void {
 
     if(!dma.io.enable) io.command.bit(5) = 0;
     if(dma.io.mode == 3) dma.io.wait = false;
+    dma.poll();
     return;
   }
 
@@ -182,6 +184,7 @@ auto VDP::writeControlPort(uint16 data) -> void {
     io.displayEnable = data.bit(6);
     vram.mode = data.bit(7);
     if(!dma.io.enable) io.command.bit(5) = 0;
+    dma.poll();
     return;
   }
 
@@ -322,6 +325,7 @@ auto VDP::writeControlPort(uint16 data) -> void {
     dma.io.source.bit(16,21) = data.bit(0,5);
     dma.io.mode = data.bit(6,7);
     dma.io.wait = dma.io.mode.bit(1);
+    dma.poll();
     return;
   }
 
@@ -331,4 +335,4 @@ auto VDP::writeControlPort(uint16 data) -> void {
   }
 
   }
-}
+}

higan/md/vdp-performance/vdp.cpp

@@ -38,19 +38,8 @@ auto VDP::unload() -> void {
 }
 
 static uint counter = 0;
-static uint dmaCounter = 0;
 
 auto VDP::main() -> void {
-#if defined(SCHEDULER_SYNCHRO)
-  if (dmaCounter > 0) {
-    dma.run();
-    Thread::step(1);
-    Thread::synchronize(apu);
-    dmaCounter--;
-    return;
-  }
-#endif
-
   switch(counter++) {
     case 1:
       //H = 0
@@ -126,29 +115,7 @@ auto VDP::main() -> void {
 
 auto VDP::step(uint clocks) -> void {
   state.hcounter += clocks;
-
-  if(optimizeSteps && (!dma.io.enable || dma.io.wait)) {
-    dma.active = 0;
-    Thread::step(clocks);
-#if defined(SCHEDULER_SYNCHRO)
-    Thread::synchronize(apu);
-#else
-    Thread::synchronize(cpu, apu);
-#endif
-  } else {
-#if defined(SCHEDULER_SYNCHRO)
-    dmaCounter = clocks - 1;
-    dma.run();
-    Thread::step(1);
-    Thread::synchronize(apu);
-#else
-    while(clocks--) {
-      dma.run();
-      Thread::step(1);
-      Thread::synchronize(cpu, apu);
-    }
-#endif
-  }
+  Thread::step(clocks);
 }
 
 auto VDP::refresh() -> void {

higan/md/vdp-performance/vdp.hpp

@@ -8,7 +8,6 @@ struct VDP : Thread {
   double overclock = 0.5;
 
   bool hasRendered = false;
-  bool optimizeSteps = false;
   bool isSkipping = false;
   bool skip;
 
@@ -89,6 +88,7 @@ struct VDP : Thread {
 
   struct DMA {
     //dma.cpp
+    alwaysinline auto poll() -> void;
     alwaysinline auto run() -> bool;
     alwaysinline auto load() -> void;
     alwaysinline auto fill() -> void;

higan/target-web/shell.html

@@ -922,8 +922,6 @@
                 <div class="values">
                   <input type="checkbox" name="md-skipframe" id="md-skipframe" />
                   <label for="md-skipframe">Skip frame</label>
-                  <input type="checkbox" name="md-optimizeSteps" id="md-optimizeSteps" />
-                  <label for="md-optimizeSteps">VDP - Optimize DMA Steps</label>
                   <select name="md-cpu-overclock" id="md-cpu-overclock">
                     <option value="0.25">0.25</option>
                     <option value="0.5">0.5</option>
@@ -1457,9 +1455,6 @@
               case 'skipframe':
                 Module.configure('vdp/skipframe', value);
                 return true;
-              case 'optimizeSteps':
-                Module.configure('vdp/optimizeSteps', value);
-                return true;
             }
             return false;
         }