some dma improvements

no idea if anything needs this, shouldn't hurt either

diff --git a/pico/carthw/svp/svp.c b/pico/carthw/svp/svp.c
index 1bccb3f..8861de0 100644 (file)
--- a/pico/carthw/svp/svp.c
+++ b/pico/carthw/svp/svp.c
@@ -91,23 +91,20 @@ static void PicoSVPLine(void)
 }
 
 
-static int PicoSVPDma(unsigned int source, int len, unsigned short **srcp, unsigned short **limitp)
+static int PicoSVPDma(unsigned int source, int len, unsigned short **base, unsigned int *mask)
 {
        if (source < Pico.romsize) // Rom
        {
-               source -= 2;
-               *srcp = (unsigned short *)(Pico.rom + (source&~1));
-               *limitp = (unsigned short *)(Pico.rom + Pico.romsize);
-               return 1;
+               *base = (unsigned short *)(Pico.rom + (source & 0xfe0000));
+               *mask = 0x1ffff;
+               return source - 2;
        }
        else if ((source & 0xfe0000) == 0x300000)
        {
                elprintf(EL_VDPDMA|EL_SVP, "SVP DmaSlow from %06x, len=%i", source, len);
-               source &= 0x1fffe;
-               source -= 2;
-               *srcp = (unsigned short *)(svp->dram + source);
-               *limitp = (unsigned short *)(svp->dram + sizeof(svp->dram));
-               return 1;
+               *base = (unsigned short *)svp->dram;
+               *mask = 0x1ffff;
+               return source - 2;
        }
        else
                elprintf(EL_VDPDMA|EL_SVP|EL_ANOMALY, "SVP FIXME unhandled DmaSlow from %06x, len=%i", source, len);

diff --git a/pico/memory.h b/pico/memory.h
index 4260a35..ae7ae50 100644 (file)
--- a/pico/memory.h
+++ b/pico/memory.h
@@ -133,17 +133,25 @@ void name(u32 a, u32 d)                         \
   }                                             \
 }
 
+#ifdef NEED_DMA_SOURCE // meh
+
 static __inline void *m68k_dma_source(u32 a)
 {
+  u8 *base;
   uptr v;
-  a &= 0x00fffffe;
   v = m68k_read16_map[a >> M68K_MEM_SHIFT];
-  if (map_flag_set(v))
-    return NULL;
+  if (map_flag_set(v)) {
+    if (a >= Pico.romsize) // Rom
+      return NULL;
+    base = Pico.rom;
+  }
   else
-    return (void *)((v << 1) + a);
+    base = (void *)(v << 1);
+  return base + (a & 0xfe0000);
 }
 
+#endif
+
 // 32x
 typedef struct {
   uptr addr; // stores (membase >> 1) or ((handler >> 1) | (1<<31))

diff --git a/pico/pico.c b/pico/pico.c
index 5951b59..23f7efb 100644 (file)
--- a/pico/pico.c
+++ b/pico/pico.c
@@ -276,7 +276,8 @@ PICO_INTERNAL int CheckDMA(void)
     Pico.m.dma_xfers -= xfers_can;\r
   }\r
 \r
-  elprintf(EL_VDPDMA, "~Dma %i op=%i can=%i burn=%i [%i]", Pico.m.dma_xfers, dma_op1, xfers_can, burn, SekCyclesDone());\r
+  elprintf(EL_VDPDMA, "~Dma %i op=%i can=%i burn=%i [%u]",\r
+    Pico.m.dma_xfers, dma_op1, xfers_can, burn, SekCyclesDone());\r
   //dprintf("~aim: %i, cnt: %i", SekCycleAim, SekCycleCnt);\r
   return burn;\r
 }\r

diff --git a/pico/pico_int.h b/pico/pico_int.h
index 5d7bfbc..794fcd1 100644 (file)
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -782,7 +782,7 @@ extern int line_base_cycles;
 PICO_INTERNAL_ASM void PicoVideoWrite(unsigned int a,unsigned short d);\r
 PICO_INTERNAL_ASM unsigned int PicoVideoRead(unsigned int a);\r
 PICO_INTERNAL_ASM unsigned int PicoVideoRead8(unsigned int a);\r
-extern int (*PicoDmaHook)(unsigned int source, int len, unsigned short **srcp, unsigned short **limitp);\r
+extern int (*PicoDmaHook)(unsigned int source, int len, unsigned short **base, unsigned int *mask);\r
 \r
 // misc.c\r
 PICO_INTERNAL_ASM void memcpy16(unsigned short *dest, unsigned short *src, int count);\r

diff --git a/pico/videoport.c b/pico/videoport.c
index 35ece48..3f26d58 100644 (file)
--- a/pico/videoport.c
+++ b/pico/videoport.c
@@ -8,6 +8,7 @@
  */\r
 \r
 #include "pico_int.h"\r
+#define NEED_DMA_SOURCE\r
 #include "memory.h"\r
 \r
 int line_base_cycles;\r
@@ -21,7 +22,7 @@ typedef unsigned int   u32;
 #define UTYPES_DEFINED\r
 #endif\r
 \r
-int (*PicoDmaHook)(unsigned int source, int len, unsigned short **srcp, unsigned short **limitp) = NULL;\r
+int (*PicoDmaHook)(unsigned int source, int len, unsigned short **base, unsigned int *mask) = NULL;\r
 \r
 static __inline void AutoIncrement(void)\r
 {\r
@@ -73,104 +74,92 @@ static int GetDmaLength(void)
   // 16-bit words to transfer:\r
   len =pvid->reg[0x13];\r
   len|=pvid->reg[0x14]<<8;\r
-  // Charles MacDonald:\r
-  if(!len) len = 0xffff;\r
+  len = ((len - 1) & 0xffff) + 1;\r
   return len;\r
 }\r
 \r
-static void DmaSlow(int len)\r
+static void DmaSlow(int len, unsigned int source)\r
 {\r
-  u16 *pd=0, *pdend, *r;\r
-  unsigned int a=Pico.video.addr, a2, d;\r
-  unsigned char inc=Pico.video.reg[0xf];\r
-  unsigned int source;\r
-\r
-  source =Pico.video.reg[0x15]<<1;\r
-  source|=Pico.video.reg[0x16]<<9;\r
-  source|=Pico.video.reg[0x17]<<17;\r
+  u32 inc = Pico.video.reg[0xf];\r
+  u32 a = Pico.video.addr;\r
+  u16 *r, *base = NULL;\r
+  u32 mask = 0x1ffff;\r
 \r
-  elprintf(EL_VDPDMA, "DmaSlow[%i] %06x->%04x len %i inc=%i blank %i [%i] @ %06x",\r
+  elprintf(EL_VDPDMA, "DmaSlow[%i] %06x->%04x len %i inc=%i blank %i [%u] @ %06x",\r
     Pico.video.type, source, a, len, inc, (Pico.video.status&8)||!(Pico.video.reg[1]&0x40),\r
     SekCyclesDone(), SekPc);\r
 \r
   Pico.m.dma_xfers += len;\r
+  if (Pico.m.dma_xfers < len) // lame 16bit var\r
+    Pico.m.dma_xfers = ~0;\r
   SekCyclesBurnRun(CheckDMA());\r
 \r
-  if ((source&0xe00000)==0xe00000) { // Ram\r
-    pd=(u16 *)(Pico.ram+(source&0xfffe));\r
-    pdend=(u16 *)(Pico.ram+0x10000);\r
+  if ((source & 0xe00000) == 0xe00000) { // Ram\r
+    base = (u16 *)Pico.ram;\r
+    mask = 0xffff;\r
   }\r
   else if (PicoAHW & PAHW_MCD)\r
   {\r
-    elprintf(EL_VDPDMA, "DmaSlow CD, r3=%02x", Pico_mcd->s68k_regs[3]);\r
-    if(source<0x20000) { // Bios area\r
-      pd=(u16 *)(Pico_mcd->bios+(source&~1));\r
-      pdend=(u16 *)(Pico_mcd->bios+0x20000);\r
-    } else if ((source&0xfc0000)==0x200000) { // Word Ram\r
-      source -= 2;\r
-      if (!(Pico_mcd->s68k_regs[3]&4)) { // 2M mode\r
-        pd=(u16 *)(Pico_mcd->word_ram2M+(source&0x3fffe));\r
-        pdend=(u16 *)(Pico_mcd->word_ram2M+0x40000);\r
+    u8 r3 = Pico_mcd->s68k_regs[3];\r
+    elprintf(EL_VDPDMA, "DmaSlow CD, r3=%02x", r3);\r
+    if (source < 0x20000) { // Bios area\r
+      base = (u16 *)Pico_mcd->bios;\r
+    } else if ((source & 0xfc0000) == 0x200000) { // Word Ram\r
+      if (!(r3 & 4)) { // 2M mode\r
+        base = (u16 *)(Pico_mcd->word_ram2M + (source & 0x20000));\r
       } else {\r
         if (source < 0x220000) { // 1M mode\r
-          int bank = Pico_mcd->s68k_regs[3]&1;\r
-          pd=(u16 *)(Pico_mcd->word_ram1M[bank]+(source&0x1fffe));\r
-          pdend=(u16 *)(Pico_mcd->word_ram1M[bank]+0x20000);\r
+          int bank = r3 & 1;\r
+          base = (u16 *)(Pico_mcd->word_ram1M[bank]);\r
         } else {\r
-          DmaSlowCell(source, a, len, inc);\r
+          DmaSlowCell(source - 2, a, len, inc);\r
           return;\r
         }\r
       }\r
-    } else if ((source&0xfe0000)==0x020000) { // Prg Ram\r
-      u8 *prg_ram = Pico_mcd->prg_ram_b[Pico_mcd->s68k_regs[3]>>6];\r
-      pd=(u16 *)(prg_ram+(source&0x1fffe));\r
-      pdend=(u16 *)(prg_ram+0x20000);\r
-    } else {\r
-      elprintf(EL_VDPDMA|EL_ANOMALY, "DmaSlow[%i] %06x->%04x: FIXME: unsupported src", Pico.video.type, source, a);\r
-      return;\r
+      source -= 2;\r
+    } else if ((source & 0xfe0000) == 0x020000) { // Prg Ram\r
+      base = (u16 *)Pico_mcd->prg_ram_b[r3 >> 6];\r
+      source -= 2; // XXX: test\r
     }\r
   }\r
   else\r
   {\r
     // if we have DmaHook, let it handle ROM because of possible DMA delay\r
-    if (PicoDmaHook && PicoDmaHook(source, len, &pd, &pdend));\r
-    else if (source<Pico.romsize) { // Rom\r
-      pd=m68k_dma_source(source);\r
-      pdend=(u16 *)(Pico.rom+Pico.romsize);\r
-    }\r
-    if (!pd) {\r
-      elprintf(EL_VDPDMA|EL_ANOMALY, "DmaSlow[%i] %06x->%04x: invalid src", Pico.video.type, source, a);\r
-      return;\r
-    }\r
+    u32 source2;\r
+    if (PicoDmaHook && (source2 = PicoDmaHook(source, len, &base, &mask)))\r
+      source = source2;\r
+    else // Rom\r
+      base = m68k_dma_source(source);\r
   }\r
-\r
-  // overflow protection, might break something..\r
-  if (len > pdend - pd) {\r
-    len = pdend - pd;\r
-    elprintf(EL_VDPDMA|EL_ANOMALY, "DmaSlow overflow");\r
+  if (!base) {\r
+    elprintf(EL_VDPDMA|EL_ANOMALY, "DmaSlow[%i] %06x->%04x: invalid src", Pico.video.type, source, a);\r
+    return;\r
   }\r
 \r
+  // operate in words\r
+  source >>= 1;\r
+  mask >>= 1;\r
+\r
   switch (Pico.video.type)\r
   {\r
     case 1: // vram\r
       r = Pico.vram;\r
-      if (inc == 2 && !(a&1) && a+len*2 < 0x10000)\r
+      if (inc == 2 && !(a & 1) && a + len * 2 < 0x10000\r
+          && !(((source + len - 1) ^ source) & ~mask))\r
       {\r
         // most used DMA mode\r
-        memcpy16(r + (a>>1), pd, len);\r
-        a += len*2;\r
+        memcpy((char *)r + a, base + (source & mask), len * 2);\r
+        a += len * 2;\r
       }\r
       else\r
       {\r
         for(; len; len--)\r
         {\r
-          d=*pd++;\r
-          if(a&1) d=(d<<8)|(d>>8);\r
-          r[a>>1] = (u16)d; // will drop the upper bits\r
+          u16 d = base[source++ & mask];\r
+          if(a & 1) d=(d<<8)|(d>>8);\r
+          r[a >> 1] = d;\r
           // AutoIncrement\r
-          a=(u16)(a+inc);\r
-          // didn't src overlap?\r
-          //if(pd >= pdend) pd-=0x8000; // should be good for RAM, bad for ROM\r
+          a = (u16)(a + inc);\r
         }\r
       }\r
       Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;\r
@@ -179,32 +168,22 @@ static void DmaSlow(int len)
     case 3: // cram\r
       Pico.m.dirtyPal = 1;\r
       r = Pico.cram;\r
-      for(a2=a&0x7f; len; len--)\r
+      for (; len; len--)\r
       {\r
-        r[a2>>1] = (u16)*pd++; // bit 0 is ignored\r
+        r[(a / 2) & 0x3f] = base[source++ & mask];\r
         // AutoIncrement\r
-        a2+=inc;\r
-        // didn't src overlap?\r
-        //if(pd >= pdend) pd-=0x8000;\r
-        // good dest?\r
-        if(a2 >= 0x80) break; // Todds Adventures in Slime World / Andre Agassi tennis\r
+        a += inc;\r
       }\r
-      a=(a&0xff00)|a2;\r
       break;\r
 \r
-    case 5: // vsram[a&0x003f]=d;\r
+    case 5: // vsram\r
       r = Pico.vsram;\r
-      for(a2=a&0x7f; len; len--)\r
+      for (; len; len--)\r
       {\r
-        r[a2>>1] = (u16)*pd++;\r
+        r[(a / 2) & 0x3f] = base[source++ & mask];\r
         // AutoIncrement\r
-        a2+=inc;\r
-        // didn't src overlap?\r
-        //if(pd >= pdend) pd-=0x8000;\r
-        // good dest?\r
-        if(a2 >= 0x80) break;\r
+        a += inc;\r
       }\r
-      a=(a&0xff00)|a2;\r
       break;\r
 \r
     default:\r
@@ -220,23 +199,21 @@ static void DmaCopy(int len)
 {\r
   u16 a=Pico.video.addr;\r
   unsigned char *vr = (unsigned char *) Pico.vram;\r
-  unsigned char *vrs;\r
   unsigned char inc=Pico.video.reg[0xf];\r
   int source;\r
   elprintf(EL_VDPDMA, "DmaCopy len %i [%i]", len, SekCyclesDone());\r
 \r
   Pico.m.dma_xfers += len;\r
+  if (Pico.m.dma_xfers < len)\r
+    Pico.m.dma_xfers = ~0;\r
   Pico.video.status |= 2; // dma busy\r
 \r
   source =Pico.video.reg[0x15];\r
   source|=Pico.video.reg[0x16]<<8;\r
-  vrs=vr+source;\r
-\r
-  if (source+len > 0x10000) len=0x10000-source; // clip??\r
 \r
   for (; len; len--)\r
   {\r
-    vr[a] = *vrs++;\r
+    vr[a] = vr[source++ & 0xffff];\r
     // AutoIncrement\r
     a=(u16)(a+inc);\r
   }\r
@@ -245,57 +222,85 @@ static void DmaCopy(int len)
   Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;\r
 }\r
 \r
-// check: Contra, Megaman\r
-// note: this is still inaccurate\r
-static void DmaFill(int data)\r
+static NOINLINE void DmaFill(int data)\r
 {\r
-  int len;\r
   unsigned short a=Pico.video.addr;\r
   unsigned char *vr=(unsigned char *) Pico.vram;\r
   unsigned char high = (unsigned char) (data >> 8);\r
   unsigned char inc=Pico.video.reg[0xf];\r
+  int source;\r
+  int len, l;\r
 \r
-  len=GetDmaLength();\r
+  len = GetDmaLength();\r
   elprintf(EL_VDPDMA, "DmaFill len %i inc %i [%i]", len, inc, SekCyclesDone());\r
 \r
   Pico.m.dma_xfers += len;\r
+  if (Pico.m.dma_xfers < len) // lame 16bit var\r
+    Pico.m.dma_xfers = ~0;\r
   Pico.video.status |= 2; // dma busy\r
 \r
-  // from Charles MacDonald's genvdp.txt:\r
-  // Write lower byte to address specified\r
-  vr[a] = (unsigned char) data;\r
-  a=(u16)(a+inc);\r
-\r
-  if (!inc) len=1;\r
-\r
-  for (; len; len--) {\r
-    // Write upper byte to adjacent address\r
-    // (here we are byteswapped, so address is already 'adjacent')\r
-    vr[a] = high;\r
+  switch (Pico.video.type)\r
+  {\r
+    case 1: // vram\r
+      for (l = len; l; l--) {\r
+        // Write upper byte to adjacent address\r
+        // (here we are byteswapped, so address is already 'adjacent')\r
+        vr[a] = high;\r
 \r
-    // Increment address register\r
-    a=(u16)(a+inc);\r
+        // Increment address register\r
+        a = (u16)(a + inc);\r
+      }\r
+      break;\r
+    case 3:   // cram\r
+    case 5: { // vsram\r
+      // TODO: needs fifo; anyone using these?\r
+      static int once;\r
+      if (!once++)\r
+        elprintf(EL_STATUS|EL_ANOMALY|EL_VDPDMA, "TODO: cram/vsram fill");\r
+    }\r
+    default:\r
+      a += len * inc;\r
+      break;\r
   }\r
+\r
   // remember addr\r
-  Pico.video.addr=a;\r
-  // update length\r
-  Pico.video.reg[0x13] = Pico.video.reg[0x14] = 0; // Dino Dini's Soccer (E) (by Haze)\r
+  Pico.video.addr = a;\r
+  // register update\r
+  Pico.video.reg[0x13] = Pico.video.reg[0x14] = 0;\r
+  source  = Pico.video.reg[0x15];\r
+  source |= Pico.video.reg[0x16] << 8;\r
+  source += len;\r
+  Pico.video.reg[0x15] = source;\r
+  Pico.video.reg[0x16] = source >> 8;\r
 \r
   Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;\r
 }\r
 \r
-static void CommandDma(void)\r
+static NOINLINE void CommandDma(void)\r
 {\r
   struct PicoVideo *pvid=&Pico.video;\r
-  int len=0,method=0;\r
+  u32 len, method;\r
+  u32 source;\r
 \r
   if ((pvid->reg[1]&0x10)==0) return; // DMA not enabled\r
 \r
-  len=GetDmaLength();\r
+  len = GetDmaLength();\r
+  source =Pico.video.reg[0x15];\r
+  source|=Pico.video.reg[0x16] << 8;\r
+  source|=Pico.video.reg[0x17] << 16;\r
 \r
   method=pvid->reg[0x17]>>6;\r
-  if (method< 2) DmaSlow(len); // 68000 to VDP\r
-  if (method==3) DmaCopy(len); // VRAM Copy\r
+  if (method < 2)\r
+    DmaSlow(len, source << 1); // 68000 to VDP\r
+  else if (method == 3)\r
+    DmaCopy(len); // VRAM Copy\r
+  else\r
+    return;\r
+\r
+  source += len;\r
+  Pico.video.reg[0x13] = Pico.video.reg[0x14] = 0;\r
+  Pico.video.reg[0x15] = source;\r
+  Pico.video.reg[0x16] = source >> 8;\r
 }\r
 \r
 static void CommandChange(void)\r
@@ -312,9 +317,6 @@ static void CommandChange(void)
   addr =(cmd>>16)&0x3fff;\r
   addr|=(cmd<<14)&0xc000;\r
   pvid->addr=(unsigned short)addr;\r
-\r
-  // Check for dma:\r
-  if (cmd&0x80) CommandDma();\r
 }\r
 \r
 static void DrawSync(int blank_on)\r
@@ -348,27 +350,23 @@ PICO_INTERNAL_ASM void PicoVideoWrite(unsigned int a,unsigned short d)
       pvid->pending=0;\r
     }\r
 \r
-    // If a DMA fill has been set up, do it\r
-    if ((pvid->command&0x80) && (pvid->reg[1]&0x10) && (pvid->reg[0x17]>>6)==2)\r
+    // preliminary FIFO emulation for Chaos Engine, The (E)\r
+    if (!(pvid->status&8) && (pvid->reg[1]&0x40) && !(PicoOpt&POPT_DIS_VDP_FIFO)) // active display?\r
     {\r
-      DmaFill(d);\r
-    }\r
-    else\r
-    {\r
-      // preliminary FIFO emulation for Chaos Engine, The (E)\r
-      if (!(pvid->status&8) && (pvid->reg[1]&0x40) && !(PicoOpt&POPT_DIS_VDP_FIFO)) // active display?\r
-      {\r
-        pvid->status&=~0x200; // FIFO no longer empty\r
-        pvid->lwrite_cnt++;\r
-        if (pvid->lwrite_cnt >= 4) pvid->status|=0x100; // FIFO full\r
-        if (pvid->lwrite_cnt >  4) {\r
-          SekCyclesBurnRun(32); // penalty // 488/12-8\r
-        }\r
-        elprintf(EL_ASVDP, "VDP data write: %04x [%06x] {%i} #%i @ %06x", d, Pico.video.addr,\r
-                 Pico.video.type, pvid->lwrite_cnt, SekPc);\r
+      pvid->status&=~0x200; // FIFO no longer empty\r
+      pvid->lwrite_cnt++;\r
+      if (pvid->lwrite_cnt >= 4) pvid->status|=0x100; // FIFO full\r
+      if (pvid->lwrite_cnt >  4) {\r
+        SekCyclesBurnRun(32); // penalty // 488/12-8\r
       }\r
-      VideoWrite(d);\r
+      elprintf(EL_ASVDP, "VDP data write: %04x [%06x] {%i} #%i @ %06x", d, Pico.video.addr,\r
+               Pico.video.type, pvid->lwrite_cnt, SekPc);\r
     }\r
+    VideoWrite(d);\r
+\r
+    if ((pvid->command&0x80) && (pvid->reg[1]&0x10) && (pvid->reg[0x17]>>6)==2)\r
+      DmaFill(d);\r
+\r
     return;\r
   }\r
 \r
@@ -376,12 +374,16 @@ PICO_INTERNAL_ASM void PicoVideoWrite(unsigned int a,unsigned short d)
   {\r
     if (pvid->pending)\r
     {\r
-      if (d & 0x80) DrawSync(0); // only need sync for DMA\r
       // Low word of command:\r
-      pvid->command&=0xffff0000;\r
-      pvid->command|=d;\r
-      pvid->pending=0;\r
+      pvid->command &= 0xffff0000;\r
+      pvid->command |= d;\r
+      pvid->pending = 0;\r
       CommandChange();\r
+      // Check for dma:\r
+      if (d & 0x80) {\r
+        DrawSync(0);\r
+        CommandDma();\r
+      }\r
     }\r
     else\r
     {\r