textures.

well, it's a start.

4 files changed, 133 insertions, 13 deletions

diff --git a/GPU.h b/GPU.h
index 683384f..a77f6c0 100644
--- a/GPU.h
+++ b/GPU.h
@@ -354,6 +354,35 @@ T ReadVRAM_OBJ(u32 addr)
 }
 
 
+template<typename T>
+T ReadVRAM_Texture(u32 addr)
+{
+    u32 ret = 0;
+    u32 mask = VRAMMap_Texture[(addr >> 17) & 0x3];
+
+    if (mask & (1<<0)) ret |= *(T*)&VRAM_A[addr & 0x1FFFF];
+    if (mask & (1<<1)) ret |= *(T*)&VRAM_B[addr & 0x1FFFF];
+    if (mask & (1<<2)) ret |= *(T*)&VRAM_C[addr & 0x1FFFF];
+    if (mask & (1<<3)) ret |= *(T*)&VRAM_D[addr & 0x1FFFF];
+
+    return ret;
+}
+
+template<typename T>
+T ReadVRAM_TexPal(u32 addr)
+{
+    u32 ret = 0;
+    if (addr >= 0x18000) return 0;
+    u32 mask = VRAMMap_TexPal[(addr >> 14) & 0x7];
+
+    if (mask & (1<<4)) ret |= *(T*)&VRAM_E[addr & 0xFFFF];
+    if (mask & (1<<5)) ret |= *(T*)&VRAM_F[addr & 0x3FFF];
+    if (mask & (1<<6)) ret |= *(T*)&VRAM_G[addr & 0x3FFF];
+
+    return ret;
+}
+
+
 void DisplaySwap(u32 val);
 
 void StartFrame();
diff --git a/GPU3D.cpp b/GPU3D.cpp
index 3c21cfa..827b351 100644
--- a/GPU3D.cpp
+++ b/GPU3D.cpp
@@ -1245,7 +1245,7 @@ void ExecuteCommand()
 
         case 0x22: // texcoord
             TexCoords[0] = ExecParams[0] & 0xFFFF;
-            TexCoords[1] = ExecParams[1] >> 16;
+            TexCoords[1] = ExecParams[0] >> 16;
             if ((TexParam >> 30) == 1)
             {
                 TexCoords[0] = (TexCoords[0]*TexMatrix[0] + TexCoords[1]*TexMatrix[4] + TexMatrix[8] + TexMatrix[12]) >> 12;
diff --git a/GPU3D_Soft.cpp b/GPU3D_Soft.cpp
index 6ea1ba4..5418980 100644
--- a/GPU3D_Soft.cpp
+++ b/GPU3D_Soft.cpp
@@ -19,7 +19,7 @@
 #include <stdio.h>
 #include <string.h>
 #include "NDS.h"
-#include "GPU3D.h"
+#include "GPU.h"
 
 
 namespace GPU3D
@@ -48,8 +48,64 @@ void Reset()
 }
 
 
-void RenderPixel(u32 attr, s32 x, s32 y, s32 z, u8 vr, u8 vg, u8 vb)
+void TextureLookup(u32 texparam, u32 texpal, s16 s, s16 t, u8* r, u8* g, u8* b)
 {
+    u32 vramaddr = (texparam & 0xFFFF) << 3;
+
+    u32 width = 8 << ((texparam >> 20) & 0x7);
+    u32 height = 8 << ((texparam >> 23) & 0x7);
+
+    s >>= 4;
+    t >>= 4;
+
+    // TODO: wraparound modes
+    s &= width-1;
+    t &= height-1;
+
+    switch ((texparam >> 26) & 0x7)
+    {
+    case 3: // 16-color
+        {
+            vramaddr += (((t * width) + s) >> 1);
+            u8 pixel = GPU::ReadVRAM_Texture<u8>(vramaddr);
+            if (s & 0x1) pixel >>= 4;
+            else         pixel &= 0xF;
+
+            texpal <<= 4;
+            u16 color = GPU::ReadVRAM_TexPal<u16>(texpal + (pixel<<1));
+
+            *r = (color << 1) & 0x3E; if (*r) *r++;
+            *g = (color >> 4) & 0x3E; if (*g) *g++;
+            *b = (color >> 9) & 0x3E; if (*b) *b++;
+        }
+        break;
+
+    case 4: // 256-color
+        {
+            vramaddr += ((t * width) + s);
+            u8 pixel = GPU::ReadVRAM_Texture<u8>(vramaddr);
+
+            texpal <<= 4;
+            u16 color = GPU::ReadVRAM_TexPal<u16>(texpal + (pixel<<1));
+
+            *r = (color << 1) & 0x3E; if (*r) *r++;
+            *g = (color >> 4) & 0x3E; if (*g) *g++;
+            *b = (color >> 9) & 0x3E; if (*b) *b++;
+        }
+        break;
+
+    default:
+        *r = (s)&0x3F;
+        *g = 0;
+        *b = (t)&0x3F;
+        break;
+    }
+}
+
+void RenderPixel(Polygon* polygon, s32 x, s32 y, s32 z, u8 vr, u8 vg, u8 vb, s16 s, s16 t)
+{
+    u32 attr = polygon->Attr;
+
     u32* depth = &DepthBuffer[(256*y) + x];
 
     bool passdepth = false;
@@ -65,10 +121,34 @@ void RenderPixel(u32 attr, s32 x, s32 y, s32 z, u8 vr, u8 vg, u8 vb)
 
     if (!passdepth) return;
 
+    u8 r, g, b;
+
+    if (((polygon->TexParam >> 26) & 0x7) != 0)
+    {
+        // TODO: also take DISP3DCNT into account
+
+        u8 tr, tg, tb;
+        TextureLookup(polygon->TexParam, polygon->TexPalette, s, t, &tr, &tg, &tb);
+
+        // TODO: other blending modes
+        /*r = ((tr+1) * (vr+1) - 1) >> 6;
+        g = ((tg+1) * (vg+1) - 1) >> 6;
+        b = ((tb+1) * (vb+1) - 1) >> 6;*/
+        r = tr;
+        g = tg;
+        b = tb;
+    }
+    else
+    {
+        r = vr;
+        g = vg;
+        b = vb;
+    }
+
     u8* pixel = &ColorBuffer[((256*y) + x) * 4];
-    pixel[0] = vr;
-    pixel[1] = vg;
-    pixel[2] = vb;
+    pixel[0] = r;
+    pixel[1] = g;
+    pixel[2] = b;
     pixel[3] = 31; // TODO: alpha
 
     // TODO: optional update for translucent pixels
@@ -151,6 +231,8 @@ void RenderPolygon(Polygon* polygon)
             ybot = vtx->FinalPosition[1];
             vbot = i;
         }
+
+        //printf("v%d: %d,%d\n", i, vtx->FinalPosition[0], vtx->FinalPosition[1]);
     }
 
     // draw, line per line
@@ -305,10 +387,16 @@ void RenderPolygon(Polygon* polygon)
         s32 gl = ((perspfactorl1 * vlcur->FinalColor[1]) + (perspfactorl2 * vlnext->FinalColor[1])) / (perspfactorl1 + perspfactorl2);
         s32 bl = ((perspfactorl1 * vlcur->FinalColor[2]) + (perspfactorl2 * vlnext->FinalColor[2])) / (perspfactorl1 + perspfactorl2);
 
+        s32 sl = ((perspfactorl1 * vlcur->TexCoords[0]) + (perspfactorl2 * vlnext->TexCoords[0])) / (perspfactorl1 + perspfactorl2);
+        s32 tl = ((perspfactorl1 * vlcur->TexCoords[1]) + (perspfactorl2 * vlnext->TexCoords[1])) / (perspfactorl1 + perspfactorl2);
+
         s32 rr = ((perspfactorr1 * vrcur->FinalColor[0]) + (perspfactorr2 * vrnext->FinalColor[0])) / (perspfactorr1 + perspfactorr2);
         s32 gr = ((perspfactorr1 * vrcur->FinalColor[1]) + (perspfactorr2 * vrnext->FinalColor[1])) / (perspfactorr1 + perspfactorr2);
         s32 br = ((perspfactorr1 * vrcur->FinalColor[2]) + (perspfactorr2 * vrnext->FinalColor[2])) / (perspfactorr1 + perspfactorr2);
 
+        s32 sr = ((perspfactorr1 * vrcur->TexCoords[0]) + (perspfactorr2 * vrnext->TexCoords[0])) / (perspfactorr1 + perspfactorr2);
+        s32 tr = ((perspfactorr1 * vrcur->TexCoords[1]) + (perspfactorr2 * vrnext->TexCoords[1])) / (perspfactorr1 + perspfactorr2);
+
         if (xr == xl) xr++;
         s32 xdiv = 0x1000 / (xr - xl);
 
@@ -340,7 +428,10 @@ void RenderPolygon(Polygon* polygon)
             u32 vg = ((perspfactor1 * gl) + (perspfactor2 * gr)) / (perspfactor1 + perspfactor2);
             u32 vb = ((perspfactor1 * bl) + (perspfactor2 * br)) / (perspfactor1 + perspfactor2);
 
-            RenderPixel(polygon->Attr, x, y, z, vr>>3, vg>>3, vb>>3);
+            s16 s = ((perspfactor1 * sl) + (perspfactor2 * sr)) / (perspfactor1 + perspfactor2);
+            s16 t = ((perspfactor1 * tl) + (perspfactor2 * tr)) / (perspfactor1 + perspfactor2);
+
+            RenderPixel(polygon, x, y, z, vr>>3, vg>>3, vb>>3, s, t);
         }
     }
 
diff --git a/melonDS.depend b/melonDS.depend
index 6ca48b0..67904b5 100644
--- a/melonDS.depend
+++ b/melonDS.depend
@@ -10,7 +10,7 @@
 
 1481161027 c:\documents\sources\melonds\types.h
 
-1488405583 source:c:\documents\sources\melonds\nds.cpp
+1488411597 source:c:\documents\sources\melonds\nds.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
@@ -87,7 +87,7 @@
 	"NDS.h"
 	"SPI.h"
 
-1488405475 source:c:\documents\sources\melonds\gpu2d.cpp
+1488405741 source:c:\documents\sources\melonds\gpu2d.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
@@ -125,7 +125,7 @@
 	"NDS.h"
 	"GPU.h"
 
-1488227363 c:\documents\sources\melonds\gpu.h
+1488410598 c:\documents\sources\melonds\gpu.h
 	"GPU2D.h"
 	"GPU3D.h"
 
@@ -148,16 +148,16 @@
 
 1488239485 c:\documents\sources\melonds\gpu3d.h
 
-1488239513 source:c:\documents\sources\melonds\gpu3d.cpp
+1488410045 source:c:\documents\sources\melonds\gpu3d.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
 	"GPU.h"
 	"FIFO.h"
 
-1488390544 source:c:\documents\sources\melonds\gpu3d_soft.cpp
+1488412022 source:c:\documents\sources\melonds\gpu3d_soft.cpp
 	<stdio.h>
 	<string.h>
 	"NDS.h"
-	"GPU3D.h"
+	"GPU.h"