[Jak3] Adjust overlord SPU dma to avoid audio hangs (#3804)

Change jak 3 SPU DMA to run the interrupt handler "immediately" (or at
least before `DMA_SendToSPUAndSync` returns).

This fixes an issue where audio can hang during fast cutscene playback.
I'm hoping it fixes more issues with looping/stuck audio as well, but
this needs more testing.

I originally wanted to do it this way, but thought that it didn't work -
from Jak 2 it seemed like things broke if the DMA was too fast. But, at
least for Jak 3, everything seems to work like this. This will remove a
huge source of non-deterministic timing in audio stuff and hopefully
make things easier to debug. It also means that a large portion of
streaming audio code will never have to run - from the game's point of
view there's always the next SPU buffer available.

If this works well, I might revisit jak 2 as well.

Co-authored-by: water111 <awaterford1111445@gmail.com>

game/overlord/jak3/dma.cpp

@@ -48,6 +48,7 @@ struct DmaInterruptHandlerHack {
   sceSdTransIntrHandler cb = nullptr;
   void* data;
   int countdown = 0;
+  bool pending = false;
 } g_DmaInterruptHack;
 
 }  // namespace
@@ -87,10 +88,20 @@ void set_dma_intr_handler_hack(s32 chan, sceSdTransIntrHandler cb, void* data) {
   g_DmaInterruptHack.cb = cb;
   g_DmaInterruptHack.data = data;
   g_DmaInterruptHack.countdown = 10;
+  g_DmaInterruptHack.pending = true;
 }
 
 int SPUDmaIntr(int channel, void* userdata);
 
+void complete_dma_now() {
+  if (g_DmaInterruptHack.pending) {
+    int chan = g_DmaInterruptHack.chan;
+    void* data = g_DmaInterruptHack.data;
+    g_DmaInterruptHack = {};
+    SPUDmaIntr(chan, data);
+  }
+}
+
 void dma_intr_hack() {
   if (g_DmaInterruptHack.countdown) {
     g_DmaInterruptHack.countdown--;
@@ -498,12 +509,26 @@ int DMA_SendToSPUAndSync(const u8* iop_mem,
     }
   }
 
-  // kick off dma, if we decided not to queue.
+  // Note on DMA interrupts.
+  // The DMA completion interrupt handler function may start more DMA transfers.
+  // If the second transfer's completion interrupt runs before the first transfer's completion
+  // interrupt returns, things break. This wasn't an issue on the real PS2 since the DMA takes
+  // longer. On PC, this means that we can't just call the completion handler from the DMA start
+  // function. Instead, put it at the end of this function.
+
+  // kick off dma, if we decided not to queue. This copies data immediately to the SPU buffer, but
+  // doesn't run the completion interrupt.
   if (!defer) {
     g_bSpuDmaBusy = true;
     set_dma_intr_handler_hack(g_nSpuDmaChannel, SPUDmaIntr, user_data);
     voice_trans_wrapper(g_nSpuDmaChannel, 0, iop_mem, spu_addr, length);
   }
+
+  // run completion interrupts. the interrupt may start another DMA transfer, which should also
+  // finish here.
+  while (g_DmaInterruptHack.pending) {
+    complete_dma_now();
+  }
   return ret;
 }
 

game/overlord/jak3/iso.cpp

@@ -821,7 +821,7 @@ u32 ISOThread() {
   // ISOFileDef* file_def = nullptr;
 
   while (true) {
-    dma_intr_hack();
+    // dma_intr_hack();
     // Part 1: Handle incoming messages from the user:
 
     int poll_result = PollMbx((MsgPacket**)&mbx_cmd, g_nISOMbx);