3 files changed, 89 insertions, 12 deletions

diff --git a/src/GPU3D.cpp b/src/GPU3D.cpp
index 9179d46..7ebe956 100644
--- a/src/GPU3D.cpp
+++ b/src/GPU3D.cpp
@@ -179,7 +179,7 @@ u16 RenderToonTable[32];
 u16 RenderEdgeTable[8];
 
 u32 RenderFogColor, RenderFogOffset;
-u8 RenderFogDensityTable[32];
+u8 RenderFogDensityTable[34];
 
 u32 RenderClearAttr1, RenderClearAttr2;
 
@@ -1815,7 +1815,9 @@ void VBlank()
 
         RenderFogColor = FogColor;
         RenderFogOffset = FogOffset;
-        memcpy(RenderFogDensityTable, FogDensityTable, 32);
+        RenderFogDensityTable[0] = FogDensityTable[0];
+        memcpy(&RenderFogDensityTable[1], FogDensityTable, 32);
+        RenderFogDensityTable[33] = FogDensityTable[31];
 
         RenderClearAttr1 = ClearAttr1;
         RenderClearAttr2 = ClearAttr2;
@@ -1988,7 +1990,7 @@ void Write8(u32 addr, u8 val)
 
     if (addr >= 0x04000360 && addr < 0x04000380)
     {
-        FogDensityTable[addr - 0x04000360] = val;
+        FogDensityTable[addr - 0x04000360] = val & 0x7F;
         return;
     }
 
@@ -2031,7 +2033,7 @@ void Write16(u32 addr, u16 val)
         FogColor = (FogColor & 0xFFFF) | (val << 16);
         return;
     case 0x0400035C:
-        FogOffset = val;
+        FogOffset = val & 0x7FFF;
         return;
     }
 
@@ -2044,8 +2046,8 @@ void Write16(u32 addr, u16 val)
     if (addr >= 0x04000360 && addr < 0x04000380)
     {
         addr -= 0x04000360;
-        FogDensityTable[addr] = val & 0xFF;
-        FogDensityTable[addr+1] = val >> 8;
+        FogDensityTable[addr] = val & 0x7F;
+        FogDensityTable[addr+1] = (val >> 8) & 0x7F;
         return;
     }
 
@@ -2085,7 +2087,7 @@ void Write32(u32 addr, u32 val)
         FogColor = val;
         return;
     case 0x0400035C:
-        FogOffset = val;
+        FogOffset = val & 0x7FFF;
         return;
 
     case 0x04000600:
@@ -2129,10 +2131,10 @@ void Write32(u32 addr, u32 val)
     if (addr >= 0x04000360 && addr < 0x04000380)
     {
         addr -= 0x04000360;
-        FogDensityTable[addr] = val & 0xFF;
-        FogDensityTable[addr+1] = (val >> 8) & 0xFF;
-        FogDensityTable[addr+2] = (val >> 16) & 0xFF;
-        FogDensityTable[addr+3] = val >> 24;
+        FogDensityTable[addr] = val & 0x7F;
+        FogDensityTable[addr+1] = (val >> 8) & 0x7F;
+        FogDensityTable[addr+2] = (val >> 16) & 0x7F;
+        FogDensityTable[addr+3] = (val >> 24) & 0x7F;
         return;
     }
 
diff --git a/src/GPU3D.h b/src/GPU3D.h
index 6c1b43c..1f98e8f 100644
--- a/src/GPU3D.h
+++ b/src/GPU3D.h
@@ -78,7 +78,7 @@ extern u16 RenderToonTable[32];
 extern u16 RenderEdgeTable[8];
 
 extern u32 RenderFogColor, RenderFogOffset;
-extern u8 RenderFogDensityTable[32];
+extern u8 RenderFogDensityTable[34];
 
 extern u32 RenderClearAttr1, RenderClearAttr2;
 
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index fe97b5f..e7906a9 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -1027,6 +1027,9 @@ void RenderPolygonScanline(RendererPolygon* rp, s32 y)
             if ((dstattr & 0x7F000000) == (attr & 0x7F000000))
                 continue;
 
+            if (!(dstattr & (1<<15)))
+                attr &= ~(1<<15);
+
             u32 dstcolor = ColorBuffer[pixeladdr];
             u32 dstalpha = dstcolor >> 24;
 
@@ -1106,6 +1109,78 @@ void RenderScanline(s32 y, int npolys)
         if (y >= polygon->YTop && (y < polygon->YBottom || (y == polygon->YTop && polygon->YBottom == polygon->YTop)))
             RenderPolygonScanline(rp, y);
     }
+
+    if (RenderDispCnt & (1<<7))
+    {
+        // fog
+
+        // hardware testing shows that the fog step is 0x80000>>SHIFT
+        // basically, the depth values used in GBAtek need to be
+        // multiplied by 0x200 to match Z-buffer values
+
+        bool fogcolor = !(RenderDispCnt & (1<<6));
+        u32 fogshift = (RenderDispCnt >> 8) & 0xF;
+        u32 fogoffset = RenderFogOffset * 0x200;
+
+        u32 fogR = (RenderFogColor << 1) & 0x3E; if (fogR) fogR++;
+        u32 fogG = (RenderFogColor >> 4) & 0x3E; if (fogG) fogG++;
+        u32 fogB = (RenderFogColor >> 9) & 0x3E; if (fogB) fogB++;
+        u32 fogA = (RenderFogColor >> 16) & 0x1F;
+
+        for (int x = 0; x < 256; x++)
+        {
+            u32 pixeladdr = (y*256) + x;
+
+            u32 attr = AttrBuffer[pixeladdr];
+            if (!(attr & (1<<15))) continue;
+
+            u32 z = DepthBuffer[pixeladdr];
+            u32 densityid, densityfrac;
+            if (z < fogoffset)
+            {
+                densityid = 0;
+                densityfrac = 0;
+            }
+            else
+            {
+                z = (z - fogoffset) << fogshift;
+                densityid = z >> 19;
+                if (densityid >= 32)
+                {
+                    densityid = 32;
+                    densityfrac = 0;
+                }
+                else
+                    densityfrac = z & 0x7FFFF;
+            }
+
+            // checkme
+            u32 density =
+                ((RenderFogDensityTable[densityid] * (0x80000-densityfrac)) +
+                 (RenderFogDensityTable[densityid+1] * densityfrac)) >> 19;
+            if (density >= 127) density = 128;
+
+            u32 srccolor = ColorBuffer[pixeladdr];
+            u32 srcR = srccolor & 0x3F;
+            u32 srcG = (srccolor >> 8) & 0x3F;
+            u32 srcB = (srccolor >> 16) & 0x3F;
+            u32 srcA = (srccolor >> 24) & 0x1F;
+
+            if (fogcolor)
+            {
+                srcR = ((fogR * density) + (srcR * (128-density))) >> 7;
+                srcG = ((fogG * density) + (srcG * (128-density))) >> 7;
+                srcB = ((fogB * density) + (srcB * (128-density))) >> 7;
+            }
+
+            if (densityid > 0)
+                srcA = ((fogA * density) + (srcA * (128-density))) >> 7;
+            else
+                srcA = ((0x1F * density) + (srcA * (128-density))) >> 7; // checkme
+
+            ColorBuffer[pixeladdr] = srcR | (srcG << 8) | (srcB << 16) | (srcA << 24);
+        }
+    }
 }
 
 void ClearBuffers()