V3.1 : 31-Jan-2022 by wavemotion-dave

* Now using more VRAM for bank swapping and partial DMA transfer to speed up large games. * Reverted part of the sound core to improve sound quality. * File selection cleanups to scroll less fast and be generally easier to see.
wavemotion-dave · Jan 31, 2022 · 639544a · 639544a
1 parent 3f0b07d
commit 639544a
Show file tree

Hide file tree

Showing 10 changed files with 96 additions and 67 deletions.
diff --git a/A7800DS.nds b/A7800DS.nds
diff --git a/Makefile b/Makefile
@@ -1,4 +1,4 @@
-VERSION=3.0
+VERSION=3.1
 TARGNAME=A7800DS
 
 #---------------------------------------------------------------------------------

diff --git a/arm9/gfx/pdev_bg0.png b/arm9/gfx/pdev_bg0.png
diff --git a/arm9/source/a7800utils.c b/arm9/source/a7800utils.c
@@ -96,40 +96,50 @@ static void DumpDebugData(void)
 #endif
 }
 
-
-u16 sampleExtender[256] = {0};
-u16 *aptr __attribute__((section(".dtcm"))) = (u16*)((u32)&sound_buffer[0] + 0xA000000); 
-u16 *bptr __attribute__((section(".dtcm"))) = (u16*)((u32)&sound_buffer[2] + 0xA000000); 
-
 u16 myTiaBufIdx __attribute__((section(".dtcm"))) = 0;
-void VsoundHandler(void) 
+u8* snd_ptr __attribute__((section(".dtcm"))) = (u8*)((u32)sound_buffer + 0xA000000);
+u8* snd_sta __attribute__((section(".dtcm"))) = (u8*)((u32)sound_buffer + 0xA000000);
+u8* snd_end __attribute__((section(".dtcm"))) = (u8*)((u32)sound_buffer + 0xA000000 + SNDLENGTH);
+ITCM_CODE void VsoundHandler(void) 
 {
   extern unsigned char tia_buffer[];
   extern u16 tiaBufIdx;
 
-  // If there is a fresh sample... 
-  if (myTiaBufIdx != tiaBufIdx)
+  for (u8 i=0; i<4; i++)
   {
-      u16 sample = sampleExtender[tia_buffer[myTiaBufIdx]];
-      myTiaBufIdx = (myTiaBufIdx+1) & (SNDLENGTH-1);
-      *aptr = sample;
-      *bptr = sample;
+      // If there is a fresh sample... 
+      if (myTiaBufIdx != tiaBufIdx)
+      {
+          lastSample = tia_buffer[myTiaBufIdx];
+          myTiaBufIdx = (myTiaBufIdx+1) & (SNDLENGTH-1);
+      }
+      *snd_ptr++ = lastSample;
+      if (snd_ptr == snd_end)
+      {
+          snd_ptr = snd_sta;
+      }
   }
 }
 
 u16 myPokeyBufIdx __attribute__((section(".dtcm"))) = 0;
-void VsoundHandler_Pokey(void)
+ITCM_CODE void VsoundHandler_Pokey(void)
  {
   extern unsigned char pokey_buffer[];
   extern u16 pokeyBufIdx;
 
-  // If there is a fresh Pokey sample... 
-  if (myPokeyBufIdx != pokeyBufIdx)
+  for (u8 i=0; i<2; i++)
   {
-      u16 sample = sampleExtender[pokey_buffer[myPokeyBufIdx]];
-      myPokeyBufIdx = (myPokeyBufIdx+1) & (SNDLENGTH-1);
-      *aptr = sample;
-      *bptr = sample;
+      // If there is a fresh Pokey sample... 
+      if (myPokeyBufIdx != pokeyBufIdx)
+      {
+          lastPokeySample = pokey_buffer[myPokeyBufIdx];
+          myPokeyBufIdx = (myPokeyBufIdx+1) & (SNDLENGTH-1);
+      }
+      *snd_ptr++ = lastPokeySample;
+      if (snd_ptr == snd_end)
+      {
+          snd_ptr = snd_sta;
+      }
   }
 }
 
@@ -489,7 +499,8 @@ void dsDisplayFiles(unsigned int NoDebGame,u32 ucSel)
 unsigned int dsWaitForRom(void) 
 {
   bool bDone=false, bRet=false;
-  u32 ucHaut=0x00, ucBas=0x00,ucSHaut=0x00, ucSBas=0x00,romSelected= 0, firstRomDisplay=0,nbRomPerPage, uNbRSPage, uLenFic=0,ucFlip=0, ucFlop=0;
+  u32 ucHaut=0x00, ucBas=0x00,ucSHaut=0x00, ucSBas=0x00,romSelected= 0, firstRomDisplay=0,nbRomPerPage, uNbRSPage, uLenFic=0;
+  s32 ucFlip=0, ucFlop=0;
   char szName[64];
 
   decompress(bgFileSelTiles, bgGetGfxPtr(bg0b), LZ77Vram);
@@ -511,6 +522,9 @@ unsigned int dsWaitForRom(void)
   dsDisplayFiles(firstRomDisplay,romSelected);
   while (!bDone) {
     if (keysCurrent() & KEY_UP) {
+      ucFlip = -50;
+      ucFlop = 0;
+      uLenFic = 0;  
       if (!ucHaut) {
         ucFicAct = (ucFicAct>0 ? ucFicAct-1 : countpro-1);
         if (romSelected>uNbRSPage) { romSelected -= 1; }
@@ -536,6 +550,9 @@ unsigned int dsWaitForRom(void)
       ucHaut = 0;
     }  
     if (keysCurrent() & KEY_DOWN) {
+      ucFlip = -50;
+      ucFlop = 0;
+      uLenFic = 0;
       if (!ucBas) {
         ucFicAct = (ucFicAct< countpro-1 ? ucFicAct+1 : 0);
         if (romSelected<uNbRSPage-1) { romSelected += 1; }
@@ -630,12 +647,12 @@ unsigned int dsWaitForRom(void)
     // Scroll la selection courante
     if (strlen(proromlist[ucFicAct].filename) > 29) {
       ucFlip++;
-      if (ucFlip >= 8) {
+      if (ucFlip >= 20) {
         ucFlip = 0;
         uLenFic++;
         if ((uLenFic+29)>strlen(proromlist[ucFicAct].filename)) {
           ucFlop++;
-          if (ucFlop >= 8) {
+          if (ucFlop >= 20) {
             uLenFic=0;
             ucFlop = 0;
           }
@@ -733,57 +750,42 @@ void dsInstallSoundEmuFIFO(void)
     {
         return;    // We've been asked to not install sound... 
     }
-
-    for (int i=0; i<256; i++)
-    {
-        sampleExtender[i] = (i << 8);
-    }
 
     FifoMessage msg;
     msg.SoundPlay.data = &sound_buffer;
-    msg.SoundPlay.freq = (isDSiMode() ? 48000: 32000);
+    msg.SoundPlay.freq = (cartridge_pokey ? SOUND_FREQ+21:(SOUND_FREQ*2)+21);
     msg.SoundPlay.volume = 127;
     msg.SoundPlay.pan = 64;
     msg.SoundPlay.loop = 1;
-    msg.SoundPlay.format = ((1)<<4) | SoundFormat_16Bit;
+    msg.SoundPlay.format = ((1)<<4) | SoundFormat_8Bit;
     msg.SoundPlay.loopPoint = 0;
-    msg.SoundPlay.dataSize = 4 >> 2;
+    msg.SoundPlay.dataSize = SNDLENGTH >> 2;
     msg.type = EMUARM7_PLAY_SND;
     fifoSendDatamsg(FIFO_USER_01, sizeof(msg), (u8*)&msg);
 
     if (isDS_LITE)
     {
-        aptr = (u16*)((u32)&sound_buffer[0] + 0x00400000); 
-        bptr = (u16*)((u32)&sound_buffer[2] + 0x00400000); 
+        snd_ptr = (u8*)((u32)sound_buffer + 0x00400000);
+        snd_sta = (u8*)((u32)sound_buffer + 0x00400000);
+        snd_end = (u8*)((u32)sound_buffer + 0x00400000 + SNDLENGTH);
     }
     else
     {
-        aptr = (u16*)((u32)&sound_buffer[0] + 0xA000000); 
-        bptr = (u16*)((u32)&sound_buffer[2] + 0xA000000); 
+        snd_ptr = (u8*)((u32)sound_buffer + 0xA000000);
+        snd_sta = (u8*)((u32)sound_buffer + 0xA000000);
+        snd_end = (u8*)((u32)sound_buffer + 0xA000000 + SNDLENGTH);
     }
 
+    TIMER2_DATA = TIMER_FREQ(SOUND_FREQ/2);
+    TIMER2_CR = TIMER_DIV_1 | TIMER_IRQ_REQ | TIMER_ENABLE;	     
     if (cartridge_pokey)
-    {
-        TIMER2_DATA = TIMER_FREQ(SOUND_FREQ+3);
-        TIMER2_CR = TIMER_DIV_1 | TIMER_IRQ_REQ | TIMER_ENABLE;	     
         irqSet(IRQ_TIMER2, VsoundHandler_Pokey);  
-    }
     else
-    {
-        if (isDSiMode())
-        {
-            TIMER2_DATA = TIMER_FREQ((SOUND_FREQ*2)+3); // For the DSi we can run at the full 31KHz.
-        }
-        else
-        {
-            TIMER2_DATA = TIMER_FREQ((SOUND_FREQ)+3);   // For the older DS-LITE we only run at 15.5KHz
-        }
-        TIMER2_CR = TIMER_DIV_1 | TIMER_IRQ_REQ | TIMER_ENABLE;	     
         irqSet(IRQ_TIMER2, VsoundHandler);
-    }
     irqEnable(IRQ_TIMER2);  
 }
 
+
 void dsMainLoop(void) 
 {
   static int lcd_swap_counter=0;

diff --git a/arm9/source/emu/Cartridge.c b/arm9/source/emu/Cartridge.c
@@ -48,7 +48,7 @@ uint cartridge_diff1 = DIFF_A;
 uint cartridge_diff2 = DIFF_A;
 
 extern int debug[];  
-static byte* cartridge_buffer __attribute__((section(".dtcm"))) = NULL;
+static byte cartridge_buffer[512 * 1024] ALIGN(32);
 static uint cartridge_size __attribute__((section(".dtcm"))) = 0;
 static uint maxbank = 9;
 
@@ -93,9 +93,13 @@ inline static void cartridge_WriteBank(word address, byte bank)
   {
     last_bank = bank;
     uint offset = cartridge_GetBank(bank) * 16384;
-    if(offset < cartridge_size) {
-      memory_WriteROMFast(address, (16384/4), cartridge_buffer + offset);
-      cartridge_bank = bank;
+    if(offset < cartridge_size) 
+    {
+        if (offset < (256*1024))    // If we are in fast VRAM memory
+            memory_WriteROM_DMA((u32*)&memory_ram[address], (u32*)(0x06860000 + offset), 16384);
+        else    // Normal memory
+            memory_WriteROMFast(address, (16384/4), cartridge_buffer + offset);
+        cartridge_bank = bank;
     }
   }
 }
@@ -201,7 +205,7 @@ static bool _cartridge_Load(const byte* data, uint size)
   uint offset = 0;
   if(cartridge_HasHeader(header)) 
   {
-    cartridge_ReadHeader(header);
+    cartridge_ReadHeader(header);   // cartridge_size will get filled in from this
     size -= 128;
     offset = 128;
   }
@@ -211,10 +215,15 @@ static bool _cartridge_Load(const byte* data, uint size)
     cartridge_size = size;
   }
 
-  if (cartridge_size <= (144 * 1024))
-    cartridge_buffer = (byte *) 0x06860000;   // If smaller than 144k (98% of all carts are), we can use the VRAM buffer since it's faster to move stuff around...
-  else   
-    cartridge_buffer = (byte *) malloc(cartridge_size); // Otherwise allocate memory 
+  // -----------------------------------------------------------------------------
+  // Copy up to 256K bytes of cart into the fast memory - used for bank swap only
+  // -----------------------------------------------------------------------------
+  u32 *fast_mem = (u32*)0x06860000;
+  memcpy(fast_mem, &data[offset], (256 * 1024));
+
+  // ----------------------------------------------------  
+  // And copy the full ROM into the main cart area...
+  // ----------------------------------------------------  
   for(index = 0; index < cartridge_size; index++) {
     cartridge_buffer[index] = data[index + offset];
   }
@@ -410,13 +419,7 @@ void cartridge_Release( )
   {
     // Snap out the High Score SRAM (if used)
     cartridge_SaveHighScoreSram();
-
-    if ((u32)cartridge_buffer != 0x06860000)
-    {
-        free(cartridge_buffer);
-    }
     cartridge_size = 0;
-    cartridge_buffer = NULL;
 
     // These values need to be reset so that moving between carts works
     // consistently. This seems to be a ProSystem emulator bug.

diff --git a/arm9/source/emu/Database.c b/arm9/source/emu/Database.c
@@ -119,7 +119,8 @@ Database_Entry game_list[] = {
   {"220121f771fc4b98cef97dc040e8d378",  "Ninja Golf",                       CT_SUPROM,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_NO,  10, 22, 270,  220}, // title=Ninja Golf
   {"74569571a208f8b0b1ccfb22d7c914e1",  "One On One",                       CT_NORMAL,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_NO,  0,  12, 256,  220}, // title=One On One
   {"5013b69cb05b21a1194ce48517df7bfc",  "Pac-Man Collection",               CT_NORMAL, POKEY_AT_4000, JOY, JOY,  DIFF_B,  DIFF_B, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_YES, 0,  19, 256,  216}, // title=Pac-Man Collection (homebrew)
-  {"a59d362e3a391ff1482131aa0818ad3e",  "Pac-Man XM",                       CT_SUPCAR,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_YES, 9,  25, 278,  220}, // title=Pac-Man Collection 40th Anniversary Edition (homebrew)
+  {"a59d362e3a391ff1482131aa0818ad3e",  "Pac-Man 40th",                     CT_SUPCAR,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_YES, 9,  25, 278,  220}, // title=Pac-Man Collection 40th Anniversary Edition (homebrew)
+  {"xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",  "PMC",                              CT_SUPCAR,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_YES, 9,  25, 278,  220}, // title=Pac-Man Collection 40th Anniversary Edition (homebrew)
   {"1a5207870dec6fae9111cb747e20d8e3",  "Pete Rose Baseball",               CT_NORMAL,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_NO,  0,  22, 256,  220}, // title=Pete Rose Baseball
   {"ec206c8db4316eb1ebce9fc960da7d8f",  "Pit Fighter",                      CT_SUPROM,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_NO,  0,  22, 256,  220}, // title=Pit Fighter                     
   {"33aea1e2b6634a1dec8c7006d9afda22",  "Planet Smashers",                  CT_SUPROM,    POKEY_NONE, JOY, JOY,  DIFF_A,  DIFF_A, NTSC,  STEAL_CYCLE,    USES_WSYNC, HSC_NO,  0,  18, 256,  226}, // title=Planet Smashers

diff --git a/arm9/source/emu/Memory.c b/arm9/source/emu/Memory.c
@@ -250,6 +250,23 @@ ITCM_CODE void memory_WriteROMFast(word address, word size, const byte* data)
   }
 }
 
+// ----------------------------------------------------------------------------
+// WriteROMFast (assumes memory_rom[] already set properly)
+// size is already in multiples of u32 dwords
+// ----------------------------------------------------------------------------
+ITCM_CODE void memory_WriteROM_DMA(u32 *dest, u32 *src, u32 size)
+{
+    // -------------------------------------------------------------------------
+    // DMA is holding up the bus for sound so we do a mix of DMA and memcpy()
+    // which is a slight improvement but keeps the sound engine mostly running.
+    // -------------------------------------------------------------------------
+    DC_InvalidateRange(dest, 4096);
+    dmaCopyWordsAsynch(3, src, dest, 4096);
+    memcpy(dest+(4096>>2), src+(4096>>2), size-4096);
+    while (dmaBusy(3)) asm("nop");
+}
+
+
 // ----------------------------------------------------------------------------
 // ClearROM
 // ----------------------------------------------------------------------------

diff --git a/arm9/source/emu/Memory.h b/arm9/source/emu/Memory.h
@@ -40,6 +40,7 @@ extern ITCM_CODE void memory_Write(word address, byte data);
 
 extern void memory_WriteROM(word address, word size, const byte* data);
 extern void memory_WriteROMFast(word address, word size, const byte* data);
+extern void memory_WriteROM_DMA(u32 *dest, u32 *src, u32 size);
 extern void memory_WriteROM32(word address, word size, const byte* data);
 extern void memory_ClearROM(word address, word size);
 extern byte memory_ram[MEMORY_SIZE];

diff --git a/arm9/source/emu/ProSystem.c b/arm9/source/emu/ProSystem.c
@@ -132,7 +132,7 @@ ITCM_CODE void prosystem_ExecuteFrame(const byte* input)
         // --------------------------------------------------------------------
         pokey_Process(1);
         pokey_Scanline();
-    } else tia_Process(isDSiMode() ? 2:1); // If all we have to deal with is the TIA, we can do so at 31KHz unless we are older DS-LITE in which case we run at 15.5KHz
+    } else tia_Process(2); // If all we have to deal with is the TIA, we can do so at 31KHz
   }  
 }
 

diff --git a/readme.txt b/readme.txt
@@ -37,6 +37,11 @@ Philosophy :
 --------------------------------------------------------------------------------
 History :
 --------------------------------------------------------------------------------
+V3.1 : 31-Jan-2022 by wavemotion-dave
+  * Now using more VRAM for bank swapping and partial DMA transfer to speed up large games.
+  * Reverted part of the sound core to improve sound quality.
+  * File selection cleanups to scroll less fast and be generally easier to see.
+
 V3.0 : 06-Nov-2021 by wavemotion-dave
   * Reworked sound output core so it's now zinger-free!
   * Refresh of bottom screen - improved font and button debounce.