proper clear-plane support, including bitmap mode

GPU3D.cpp

@@ -143,6 +143,7 @@ FIFO<CmdFIFOEntry>* CmdPIPE;
 
 u32 NumCommands, CurCommand, ParamCount, TotalParams;
 
+u32 DispCnt;
 u32 GXStat;
 
 u32 ExecParams[32];
@@ -196,6 +197,8 @@ Polygon* CurPolygonRAM;
 u32 NumVertices, NumPolygons;
 u32 CurRAMBank;
 
+u32 ClearAttr1, ClearAttr2;
+
 u32 FlushRequest;
 u32 FlushAttributes, CurFlushAttributes;
 
@@ -229,6 +232,7 @@ void Reset()
     ParamCount = 0;
     TotalParams = 0;
 
+    DispCnt = 0;
     GXStat = 0;
 
     memset(ExecParams, 0, 32*4);
@@ -262,7 +266,12 @@ void Reset()
     NumVertices = 0;
     NumPolygons = 0;
 
+    ClearAttr1 = 0;
+    ClearAttr2 = 0;
+
     FlushRequest = 0;
+    FlushAttributes = 0;
+    CurFlushAttributes = 0;
 
     SoftRenderer::Reset();
 }
@@ -544,6 +553,8 @@ void SubmitPolygon()
     // if it's outside, check if the previous and next vertices are inside
     // if so, place a new vertex at the edge of the view volume
 
+    // TODO: optional culling of polygons that clip through the far plane
+
     // X clipping
 
     c = clipstart;
@@ -1414,6 +1425,12 @@ u8 Read8(u32 addr)
 
 u16 Read16(u32 addr)
 {
+    switch (addr)
+    {
+    case 0x04000060:
+        return DispCnt;
+    }
+
     return 0;
 }
 
@@ -1421,6 +1438,9 @@ u32 Read32(u32 addr)
 {
     switch (addr)
     {
+    case 0x04000060:
+        return DispCnt;
+
     case 0x04000320:
         return 46; // TODO, eventually
 
@@ -1453,18 +1473,46 @@ u32 Read32(u32 addr)
 
 void Write8(u32 addr, u8 val)
 {
-    //
 }
 
 void Write16(u32 addr, u16 val)
 {
-    //
+    switch (addr)
+    {
+    case 0x04000060:
+        DispCnt = val;
+        return;
+
+    case 0x04000350:
+        ClearAttr1 = (ClearAttr1 & 0xFFFF0000) | val;
+        return;
+    case 0x04000352:
+        ClearAttr1 = (ClearAttr1 & 0xFFFF) | (val << 16);
+        return;
+    case 0x04000354:
+        ClearAttr2 = (ClearAttr2 & 0xFFFF0000) | val;
+        return;
+    case 0x04000356:
+        ClearAttr2 = (ClearAttr2 & 0xFFFF) | (val << 16);
+        return;
+    }
 }
 
 void Write32(u32 addr, u32 val)
 {
     switch (addr)
     {
+    case 0x04000060:
+        DispCnt = val & 0xFFFF;
+        return;
+
+    case 0x04000350:
+        ClearAttr1 = val;
+        return;
+    case 0x04000354:
+        ClearAttr2 = val;
+        return;
+
     case 0x04000600:
         if (val & 0x8000) GXStat &= ~0x8000;
         val &= 0xC0000000;

GPU3D.h

@@ -53,7 +53,9 @@ typedef struct
 
 } Polygon;
 
+extern u32 DispCnt;
 extern s32 Viewport[4];
+extern u32 ClearAttr1, ClearAttr2;
 
 bool Init();
 void DeInit();

GPU3D_Soft.cpp

@@ -582,11 +582,56 @@ void RenderFrame(u32 attr, Vertex* vertices, Polygon* polygons, int npolys)
 {
     // TODO: render translucent polygons last
 
-    // TODO proper clear color/depth support!
-    for (int i = 0; i < 256*192; i++)
+    // TODO: fog, poly ID, other attributes
+
+    if (DispCnt & (1<<14))
     {
-        ColorBuffer[i] = 0x00000000;
-        DepthBuffer[i] = 0xFFFFFF;
+        u8 xoff = (ClearAttr2 >> 16) & 0xFF;
+        u8 yoff = (ClearAttr2 >> 24) & 0xFF;
+
+        for (int y = 0; y < 256*192; y += 256)
+        {
+            for (int x = 0; x < 256; x++)
+            {
+                u16 val2 = GPU::ReadVRAM_Texture<u16>(0x40000 + (yoff << 9) + (xoff << 1));
+                u16 val3 = GPU::ReadVRAM_Texture<u16>(0x60000 + (yoff << 9) + (xoff << 1));
+
+                // TODO: confirm color conversion
+                u32 r = (val2 << 1) & 0x3E; if (r) r++;
+                u32 g = (val2 >> 4) & 0x3E; if (g) g++;
+                u32 b = (val2 >> 9) & 0x3E; if (b) b++;
+                u32 a = (val2 & 0x8000) ? 0x1F000000 : 0;
+                u32 color = r | (g << 8) | (b << 16) | a;
+
+                u32 z = ((val3 & 0x7FFF) * 0x200) + 0x1FF;
+                if (z >= 0x10000 && z < 0xFFFFFF) z++;
+
+                ColorBuffer[y+x] = color;
+                DepthBuffer[y+x] = z;
+
+                xoff++;
+            }
+
+            yoff++;
+        }
+    }
+    else
+    {
+        // TODO: confirm color conversion
+        u32 r = (ClearAttr1 << 1) & 0x3E; if (r) r++;
+        u32 g = (ClearAttr1 >> 4) & 0x3E; if (g) g++;
+        u32 b = (ClearAttr1 >> 9) & 0x3E; if (b) b++;
+        u32 a = (ClearAttr1 >> 16) & 0x1F;
+        u32 color = r | (g << 8) | (b << 16) | (a << 24);
+
+        u32 z = ((ClearAttr2 & 0x7FFF) * 0x200) + 0x1FF;
+		if (z >= 0x10000 && z < 0xFFFFFF) z++;
+
+        for (int i = 0; i < 256*192; i++)
+        {
+            ColorBuffer[i] = color;
+            DepthBuffer[i] = z;
+        }
     }
 
     for (int i = 0; i < npolys; i++)

NDS.cpp

@@ -1295,7 +1295,7 @@ u16 ARM9IORead16(u32 addr)
     case 0x04000004: return GPU::DispStat[0];
     case 0x04000006: return GPU::VCount;
 
-    case 0x04000060: return 0;
+    case 0x04000060: return GPU3D::Read16(addr);
     case 0x04000064:
     case 0x04000066: return GPU::GPU2D_A->Read16(addr);
 
@@ -1382,6 +1382,7 @@ u32 ARM9IORead32(u32 addr)
     {
     case 0x04000004: return GPU::DispStat[0] | (GPU::VCount << 16);
 
+    case 0x04000060: return GPU3D::Read32(addr);
     case 0x04000064: return GPU::GPU2D_A->Read32(addr);
 
     case 0x040000B0: return DMAs[0]->SrcAddr;
@@ -1547,7 +1548,7 @@ void ARM9IOWrite16(u32 addr, u16 val)
     {
     case 0x04000004: GPU::SetDispStat(0, val); return;
 
-    case 0x04000060: return;
+    case 0x04000060: GPU3D::Write16(addr, val); return;
 
     case 0x040000B8: DMAs[0]->WriteCnt((DMAs[0]->Cnt & 0xFFFF0000) | val); return;
     case 0x040000BA: DMAs[0]->WriteCnt((DMAs[0]->Cnt & 0x0000FFFF) | (val << 16)); return;
@@ -1667,7 +1668,7 @@ void ARM9IOWrite32(u32 addr, u32 val)
 {
     switch (addr)
     {
-    case 0x04000060: return;
+    case 0x04000060: GPU3D::Write32(addr, val); return;
     case 0x04000064: GPU::GPU2D_A->Write32(addr, val); return;
 
     case 0x040000B0: DMAs[0]->SrcAddr = val; return;