* start GPU shit: draw BG0 16-color.

* fix a few DMA bugs.
* fix a bug where some code could still run right after a halt instruction.
* fix VRAM mapping, too. I'm an idiot.

ARM.cpp

@@ -228,7 +228,7 @@ s32 ARM::Execute(s32 cycles)
     if (Halted)
     {
         if (NDS::HaltInterrupted(Num))
-            Halted = false;
+            Halted = 0;
         else
             return cycles;
     }
@@ -283,6 +283,7 @@ s32 ARM::Execute(s32 cycles)
         }
 
         // TODO optimize this shit!!!
+        if (Halted) return cycles;
         if (NDS::HaltInterrupted(Num))
         {
             if (NDS::IME[Num]&1)

DMA.cpp

@@ -63,20 +63,20 @@ void DMA::WriteCnt(u32 val)
         CurSrcAddr = SrcAddr;
         CurDstAddr = DstAddr;
 
-        switch (Cnt & 0x00060000)
+        switch (Cnt & 0x00600000)
         {
         case 0x00000000: DstAddrInc = 1; break;
-        case 0x00020000: DstAddrInc = -1; break;
+        case 0x00200000: DstAddrInc = -1; break;
-        case 0x00040000: DstAddrInc = 0; break;
+        case 0x00400000: DstAddrInc = 0; break;
-        case 0x00060000: DstAddrInc = 1; break;
+        case 0x00600000: DstAddrInc = 1; break;
         }
 
-        switch (Cnt & 0x00180000)
+        switch (Cnt & 0x01800000)
         {
         case 0x00000000: SrcAddrInc = 1; break;
-        case 0x00080000: SrcAddrInc = -1; break;
+        case 0x00800000: SrcAddrInc = -1; break;
-        case 0x00100000: SrcAddrInc = 0; break;
+        case 0x01000000: SrcAddrInc = 0; break;
-        case 0x00180000: SrcAddrInc = 1; printf("BAD DMA SRC INC MODE 3\n"); break;
+        case 0x01800000: SrcAddrInc = 1; printf("BAD DMA SRC INC MODE 3\n"); break;
         }
 
         if (CPU == 0)
@@ -106,10 +106,10 @@ void DMA::Start()
     if ((Cnt & 0x00060000) == 0x00060000)
         CurDstAddr = DstAddr;
 
-    printf("ARM%d DMA%d %08X->%08X %d units %dbit\n", CPU?7:9, Num, CurSrcAddr, CurDstAddr, RemCount, (Cnt & 0x04000000)?16:32);
+    printf("ARM%d DMA%d %08X %08X->%08X %d bytes %dbit\n", CPU?7:9, Num, Cnt, CurSrcAddr, CurDstAddr, RemCount*((Cnt & 0x04000000)?2:4), (Cnt & 0x04000000)?16:32);
 
     // TODO: NOT MAKE THE DMA INSTANT!!
-    if (Cnt & 0x04000000)
+    if (!(Cnt & 0x04000000))
     {
         u16 (*readfn)(u32) = CPU ? NDS::ARM7Read16 : NDS::ARM9Read16;
         void (*writefn)(u32,u16) = CPU ? NDS::ARM7Write16 : NDS::ARM9Write16;

GPU.cpp

@@ -108,7 +108,7 @@ void Reset()
     GPU2D_A->Reset();
     GPU2D_B->Reset();
 
-    GPU2D_A->SetFramebuffer(&Framebuffer[0*192]);
+    GPU2D_A->SetFramebuffer(&Framebuffer[256*0]);
     GPU2D_B->SetFramebuffer(&Framebuffer[256*192]);
 }
 
@@ -682,24 +682,6 @@ void MapVRAM_I(u32 bank, u8 cnt)
 }
 
 
-void DrawScanline(u32 screen, u32 line)
-{
-    u16* dst = &Framebuffer[256 * ((192*screen) + line)];
-
-    if (screen==0)
-    {
-        u16* src = &((u16*)VRAM_A)[256*line];
-        for (int i = 0; i < 256; i++)
-            dst[i] = src[i];
-    }
-    else
-    {
-        for (int i = 0; i < 256; i++)
-            dst[i] = 0x7FFF;
-    }
-}
-
-
 void StartFrame()
 {
     StartScanline(0);
@@ -730,8 +712,8 @@ void StartScanline(u32 line)
     if (line < 192)
     {
         // draw
-        DrawScanline(0, line);
+        GPU2D_A->DrawScanline(line);
-        DrawScanline(1, line);
+        GPU2D_B->DrawScanline(line);
 
         NDS::ScheduleEvent(LINE_CYCLES, StartScanline, line+1);
     }

GPU.h

@@ -34,6 +34,7 @@ extern u8 VRAMSTAT;
 extern u8 Palette[2*1024];
 extern u8 OAM[2*1024];
 
+extern u8* VRAM[9];
 extern u8* VRAM_ABG[128];
 extern u8* VRAM_AOBJ[128];
 extern u8* VRAM_BBG[128];

GPU2D.cpp

@@ -17,9 +17,9 @@
 */
 
 #include <stdio.h>
+#include <string.h>
 #include "NDS.h"
 #include "GPU.h"
-#include "GPU2D.h"
 
 
 GPU2D::GPU2D(u32 num)
@@ -33,7 +33,8 @@ GPU2D::~GPU2D()
 
 void GPU2D::Reset()
 {
-    //
+    DispCnt = 0;
+    memset(BGCnt, 0, 4*2);
 }
 
 void GPU2D::SetFramebuffer(u16* buf)
@@ -47,21 +48,31 @@ void GPU2D::SetFramebuffer(u16* buf)
 u8 GPU2D::Read8(u32 addr)
 {
     printf("!! GPU2D READ8 %08X\n", addr);
+    return 0;
 }
 
 u16 GPU2D::Read16(u32 addr)
 {
     switch (addr & 0x00000FFF)
     {
-        //
+    case 0x000: return DispCnt&0xFFFF;
+    case 0x002: return DispCnt>>16;
+
+    case 0x008: return BGCnt[0];
+    case 0x00A: return BGCnt[1];
+    case 0x00C: return BGCnt[2];
+    case 0x00E: return BGCnt[3];
     }
+
+    printf("unknown GPU read16 %08X\n", addr);
+    return 0;
 }
 
 u32 GPU2D::Read32(u32 addr)
 {
     switch (addr & 0x00000FFF)
     {
-        //
+    case 0x000: return DispCnt;
     }
 
     return Read16(addr) | (Read16(addr+2) << 16);
@@ -76,15 +87,29 @@ void GPU2D::Write16(u32 addr, u16 val)
 {
     switch (addr & 0x00000FFF)
     {
-        //
+    case 0x000:
+        DispCnt = (DispCnt & 0xFFFF0000) | val;
+        return;
+    case 0x002:
+        DispCnt = (DispCnt & 0x0000FFFF) | (val << 16);
+        return;
+
+    case 0x008: BGCnt[0] = val; return;
+    case 0x00A: BGCnt[1] = val; return;
+    case 0x00C: BGCnt[2] = val; return;
+    case 0x00E: BGCnt[3] = val; return;
     }
+
+    printf("unknown GPU write16 %08X %04X\n", addr, val);
 }
 
 void GPU2D::Write32(u32 addr, u32 val)
 {
     switch (addr & 0x00000FFF)
     {
-        //
+    case 0x000:
+        DispCnt = val;
+        return;
     }
 
     Write16(addr, val&0xFFFF);
@@ -94,5 +119,148 @@ void GPU2D::Write32(u32 addr, u32 val)
 
 void GPU2D::DrawScanline(u32 line)
 {
-    //
+    u16* dst = &Framebuffer[256*line];
+
+    u32 dispmode = DispCnt >> 16;
+    dispmode &= (Num ? 0x1 : 0x3);
+
+    switch (dispmode)
+    {
+    case 0: // screen off
+        {
+            for (int i = 0; i < 256>>1; i++)
+                ((u32*)dst)[i] = 0x7FFF7FFF;
+        }
+        break;
+
+    case 1: // regular display
+        {
+            DrawScanline_Mode1(line, dst);
+        }
+        break;
+
+    case 2: // VRAM display
+        {
+            u32* vram = (u32*)GPU::VRAM[(DispCnt >> 18) & 0x3];
+            vram = &vram[line << 7];
+
+            for (int i = 0; i < 256>>1; i++)
+                ((u32*)dst)[i] = vram[i];
+        }
+        break;
+
+    case 3: // FIFO display
+        {
+            // uh, is there even anything that uses this?
+        }
+        break;
+    }
+}
+
+void GPU2D::DrawScanline_Mode1(u32 line, u16* dst)
+{
+    for (int i = 0; i < 256>>1; i++)
+        ((u32*)dst)[i] = 0; // TODO: backdrop
+
+    switch (DispCnt & 0x7)
+    {
+    case 0:
+        for (int i = 3; i >= 0; i--)
+        {
+            // TODO other BGs
+            if ((BGCnt[0] & 0x3) == i)
+            {
+                DrawBG_Text_4bpp(line, dst, 0);
+                // todo: sprites
+            }
+        }
+        break;
+    }
+}
+
+// char   06218000
+// screen 06208000
+void GPU2D::DrawBG_Text_4bpp(u32 line, u16* dst, u32 bgnum)
+{
+    u16 bgcnt = BGCnt[bgnum];
+
+    u8* tileset;
+    u16* tilemap;
+    u16* pal;
+
+    // TODO scroll
+    u16 xoff = 0;
+    u16 yoff = line;
+
+    u32 widexmask = (bgcnt & 0x4000) ? 0x100 : 0;
+    //u32 ymask = (bgcnt & 0x8000) ? 0x1FF : 0xFF;
+
+    if (Num)
+    {
+        tileset = (u8*)GPU::VRAM_BBG[((bgcnt & 0x000C) >> 2)];
+        tilemap = (u16*)GPU::VRAM_BBG[((bgcnt & 0x1800) >> 11)];
+        tilemap += ((bgcnt & 0x0700) << 2);
+
+        pal = (u16*)&GPU::Palette[0x400];
+    }
+    else
+    {
+        tileset = (u8*)GPU::VRAM_ABG[((DispCnt & 0x07000000) >> 22) + ((bgcnt & 0x003C) >> 2)];
+        tilemap = (u16*)GPU::VRAM_ABG[((DispCnt & 0x38000000) >> 27) + ((bgcnt & 0x1800) >> 11)];
+        tilemap += ((bgcnt & 0x0700) << 2);
+
+        pal = (u16*)&GPU::Palette[0];
+    }
+
+    // adjust Y position in tilemap
+    if (bgcnt & 0x8000)
+    {
+        tilemap += ((yoff & 0x1F8) << 2);
+        if (bgcnt & 0x4000)
+            tilemap += ((yoff & 0x100) << 2);
+    }
+    else
+        tilemap += ((yoff & 0xF8) << 2);
+
+    u16 curtile;
+    u16* curpal;
+    u8* pixels;
+
+    // preload shit as needed
+    if (xoff & 0x7)
+    {
+        // load a new tile
+        curtile = tilemap[((xoff & 0xFF) >> 3) + ((xoff & widexmask) << 2)];
+        curpal = pal + ((curtile & 0xF000) >> 8);
+        pixels = tileset + ((curtile & 0x01FF) << 5) + ((yoff & 0x7) << 2);
+        pixels += ((xoff & 0x7) >> 1);
+    }
+
+    for (int i = 0; i < 256; i++)
+    {
+        if (!(xoff & 0x7))
+        {
+            // load a new tile
+            curtile = tilemap[((xoff & 0xFF) >> 3) + ((xoff & widexmask) << 2)];
+            curpal = pal + ((curtile & 0xF000) >> 8);
+            pixels = tileset + ((curtile & 0x01FF) << 5) + ((yoff & 0x7) << 2);
+        }
+
+        // draw pixel
+        u8 color;
+        if (xoff & 0x1)
+        {
+            color = *pixels >> 4;
+            pixels++;
+        }
+        else
+        {
+            color = *pixels & 0x0F;
+        }
+        //color = (i >> 4) + ((line >> 4) << 4);
+        //if (Num) color = 0;
+        dst[i] = curpal[color];
+
+        xoff++;
+    }
 }

GPU2D.h

@@ -41,6 +41,12 @@ public:
 private:
     u32 Num;
     u16* Framebuffer;
+
+    u32 DispCnt;
+    u16 BGCnt[4];
+
+    void DrawScanline_Mode1(u32 line, u16* dst);
+    void DrawBG_Text_4bpp(u32 line, u16* dst, u32 num);
 };
 
 #endif

NDS.cpp

@@ -484,10 +484,9 @@ void MapSharedWRAM(u8 val)
 void TriggerIRQ(u32 cpu, u32 irq)
 {
     irq = 1 << irq;
-    //if (!(IE[cpu] & irq)) return;
-
     IF[cpu] |= irq;
 
+    // this is redundant
     if (!(IME[cpu] & 0x1)) return;
     (cpu?ARM7:ARM9)->TriggerIRQ();
 }
@@ -772,8 +771,8 @@ u8 ARM9Read8(u32 addr)
             switch (addr & 0x00E00000)
             {
             case 0x00000000: vram = GPU::VRAM_ABG[chunk]; break;
-            case 0x00200000: vram = GPU::VRAM_AOBJ[chunk]; break;
+            case 0x00200000: vram = GPU::VRAM_BBG[chunk]; break;
-            case 0x00400000: vram = GPU::VRAM_BBG[chunk]; break;
+            case 0x00400000: vram = GPU::VRAM_AOBJ[chunk]; break;
             case 0x00600000: vram = GPU::VRAM_BOBJ[chunk]; break;
             case 0x00800000: vram = GPU::VRAM_LCD[chunk]; break;
             }
@@ -784,6 +783,10 @@ u8 ARM9Read8(u32 addr)
 
     case 0x07000000:
         return *(u8*)&GPU::OAM[addr & 0x7FF];
+
+    case 0x08000000:
+    case 0x09000000:
+        return 0xFF;
     }
 
     printf("unknown arm9 read8 %08X\n", addr);
@@ -827,8 +830,8 @@ u16 ARM9Read16(u32 addr)
             switch (addr & 0x00E00000)
             {
             case 0x00000000: vram = GPU::VRAM_ABG[chunk]; break;
-            case 0x00200000: vram = GPU::VRAM_AOBJ[chunk]; break;
+            case 0x00200000: vram = GPU::VRAM_BBG[chunk]; break;
-            case 0x00400000: vram = GPU::VRAM_BBG[chunk]; break;
+            case 0x00400000: vram = GPU::VRAM_AOBJ[chunk]; break;
             case 0x00600000: vram = GPU::VRAM_BOBJ[chunk]; break;
             case 0x00800000: vram = GPU::VRAM_LCD[chunk]; break;
             }
@@ -839,6 +842,10 @@ u16 ARM9Read16(u32 addr)
 
     case 0x07000000:
         return *(u16*)&GPU::OAM[addr & 0x7FF];
+
+    case 0x08000000:
+    case 0x09000000:
+        return 0xFFFF;
     }
 
     printf("unknown arm9 read16 %08X\n", addr);
@@ -897,8 +904,8 @@ u32 ARM9Read32(u32 addr)
             switch (addr & 0x00E00000)
             {
             case 0x00000000: vram = GPU::VRAM_ABG[chunk]; break;
-            case 0x00200000: vram = GPU::VRAM_AOBJ[chunk]; break;
+            case 0x00200000: vram = GPU::VRAM_BBG[chunk]; break;
-            case 0x00400000: vram = GPU::VRAM_BBG[chunk]; break;
+            case 0x00400000: vram = GPU::VRAM_AOBJ[chunk]; break;
             case 0x00600000: vram = GPU::VRAM_BOBJ[chunk]; break;
             case 0x00800000: vram = GPU::VRAM_LCD[chunk]; break;
             }
@@ -909,6 +916,10 @@ u32 ARM9Read32(u32 addr)
 
     case 0x07000000:
         return *(u32*)&GPU::OAM[addr & 0x7FF];
+
+    case 0x08000000:
+    case 0x09000000:
+        return 0xFFFFFFFF;
     }
 
     printf("unknown arm9 read32 %08X | %08X %08X %08X\n", addr, ARM9->R[15], ARM9->R[12], ARM9Read32(0x027FF820));
@@ -990,8 +1001,8 @@ void ARM9Write16(u32 addr, u16 val)
             switch (addr & 0x00E00000)
             {
             case 0x00000000: vram = GPU::VRAM_ABG[chunk]; break;
-            case 0x00200000: vram = GPU::VRAM_AOBJ[chunk]; break;
+            case 0x00200000: vram = GPU::VRAM_BBG[chunk]; break;
-            case 0x00400000: vram = GPU::VRAM_BBG[chunk]; break;
+            case 0x00400000: vram = GPU::VRAM_AOBJ[chunk]; break;
             case 0x00600000: vram = GPU::VRAM_BOBJ[chunk]; break;
             case 0x00800000: vram = GPU::VRAM_LCD[chunk]; break;
             }
@@ -1047,8 +1058,8 @@ void ARM9Write32(u32 addr, u32 val)
             switch (addr & 0x00E00000)
             {
             case 0x00000000: vram = GPU::VRAM_ABG[chunk]; break;
-            case 0x00200000: vram = GPU::VRAM_AOBJ[chunk]; break;
+            case 0x00200000: vram = GPU::VRAM_BBG[chunk]; break;
-            case 0x00400000: vram = GPU::VRAM_BBG[chunk]; break;
+            case 0x00400000: vram = GPU::VRAM_AOBJ[chunk]; break;
             case 0x00600000: vram = GPU::VRAM_BOBJ[chunk]; break;
             case 0x00800000: vram = GPU::VRAM_LCD[chunk]; break;
             }
@@ -1396,7 +1407,7 @@ u16 ARM9IORead16(u32 addr)
         return GPU::GPU2D_B->Read16(addr);
     }
 
-    printf("unknown ARM9 IO read16 %08X\n", addr);
+    printf("unknown ARM9 IO read16 %08X %08X\n", addr, ARM9->R[15]);
     return 0;
 }
 
@@ -1792,7 +1803,7 @@ u16 ARM7IORead16(u32 addr)
     case 0x04000504: return _soundbias;
     }
 
-    printf("unknown ARM7 IO read16 %08X\n", addr);
+    printf("unknown ARM7 IO read16 %08X %08X\n", addr, ARM9->R[15]);
     return 0;
 }
 

NDS.h

@@ -76,6 +76,8 @@ typedef struct
 
 // hax
 extern u32 IME[2];
+extern u32 IE[2];
+extern u32 IF[2];
 extern Timer Timers[8];
 
 extern u32 ARM9ITCMSize;

main.cpp

@@ -34,6 +34,12 @@ LRESULT CALLBACK derpo(HWND window, UINT msg, WPARAM wparam, LPARAM lparam)
     {
     case WM_CLOSE:
         printf("close\n");
+        {
+            FILE* f = fopen("vram.bin", "wb");
+            for (int i = 0; i < 128; i++)
+                fwrite(GPU::VRAM_BBG[i], 16384, 1, f);
+            fclose(f);
+        }
         PostQuitMessage(0);
         return 0;
 
@@ -46,6 +52,7 @@ LRESULT CALLBACK derpo(HWND window, UINT msg, WPARAM wparam, LPARAM lparam)
         case VK_DOWN:   NDS::PressKey(7); break;
         case VK_LEFT:   NDS::PressKey(5); break;
         case VK_RIGHT:  NDS::PressKey(4); break;
+        case 'P': NDS::PressKey(16+6); break;
         }
         return 0;
 
@@ -58,6 +65,7 @@ LRESULT CALLBACK derpo(HWND window, UINT msg, WPARAM wparam, LPARAM lparam)
         case VK_DOWN:   NDS::ReleaseKey(7); break;
         case VK_LEFT:   NDS::ReleaseKey(5); break;
         case VK_RIGHT:  NDS::ReleaseKey(4); break;
+        case 'P': NDS::ReleaseKey(16+6); break;
         }
         return 0;
 

melonDS.depend

@@ -1,16 +1,16 @@
 # depslib dependency file v1.0
-1484707090 source:c:\documents\sources\melonds\main.cpp
+1484751556 source:c:\documents\sources\melonds\main.cpp
 	<stdio.h>
 	<windows.h>
 	"NDS.h"
 	"GPU.h"
 
-1484615598 c:\documents\sources\melonds\nds.h
+1484752883 c:\documents\sources\melonds\nds.h
 	"types.h"
 
 1481161027 c:\documents\sources\melonds\types.h
 
-1484708424 source:c:\documents\sources\melonds\nds.cpp
+1484758082 source:c:\documents\sources\melonds\nds.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
@@ -22,7 +22,7 @@
 	"SPI.h"
 	"Wifi.h"
 
-1484705354 source:c:\documents\sources\melonds\arm.cpp
+1484754523 source:c:\documents\sources\melonds\arm.cpp
 	<stdio.h>
 	"NDS.h"
 	"ARM.h"
@@ -38,7 +38,7 @@
 	"types.h"
 	"ARM.h"
 
-1484705417 source:c:\documents\sources\melonds\arminterpreter.cpp
+1484752824 source:c:\documents\sources\melonds\arminterpreter.cpp
 	<stdio.h>
 	"NDS.h"
 	"CP15.h"
@@ -68,7 +68,7 @@
 
 1481037554 c:\documents\sources\melonds\cp15.h
 
-1484616493 source:c:\documents\sources\melonds\cp15.cpp
+1484754534 source:c:\documents\sources\melonds\cp15.cpp
 	<stdio.h>
 	"NDS.h"
 	"ARM.h"
@@ -81,13 +81,13 @@
 	"NDS.h"
 	"SPI.h"
 
-1484708331 source:c:\documents\sources\melonds\gpu2d.cpp
+1484758294 source:c:\documents\sources\melonds\gpu2d.cpp
 	<stdio.h>
+	<string.h>
 	"NDS.h"
 	"GPU.h"
-	"GPU2D.h"
 
-1484707825 c:\documents\sources\melonds\gpu2d.h
+1484711212 c:\documents\sources\melonds\gpu2d.h
 
 1481040524 c:\documents\sources\melonds\wifi.h
 
@@ -103,7 +103,7 @@
 1484612398 c:\documents\sources\melonds\fifo.h
 	"types.h"
 
-1484706353 source:c:\documents\sources\melonds\dma.cpp
+1484756268 source:c:\documents\sources\melonds\dma.cpp
 	<stdio.h>
 	"NDS.h"
 	"DMA.h"
@@ -111,12 +111,12 @@
 1484698068 c:\documents\sources\melonds\dma.h
 	"types.h"
 
-1484708259 source:c:\documents\sources\melonds\gpu.cpp
+1484753188 source:c:\documents\sources\melonds\gpu.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
 	"GPU.h"
 
-1484708251 c:\documents\sources\melonds\gpu.h
+1484709419 c:\documents\sources\melonds\gpu.h
 	"GPU2D.h"