diff --git a/Cxbx.opt b/Cxbx.opt
index e5f0aa37b..5f8138841 100644
Binary files a/Cxbx.opt and b/Cxbx.opt differ
diff --git a/Doc/RemovedCode.txt b/Doc/RemovedCode.txt
index d41e1c7e9..2a2e1ea33 100644
--- a/Doc/RemovedCode.txt
+++ b/Doc/RemovedCode.txt
@@ -1,3 +1,50 @@
+>> nice code for dumping textures inside of push buffer stuff
+
+            /*
+            IDirect3DBaseTexture8 *pTexture = 0;
+
+            g_pD3DDevice8->GetTexture(0, &pTexture);
+    
+            if(pTexture != NULL)
+            {
+                static int dwDumpTexture = 0;
+
+                char szBuffer[256];
+
+                printf("Texture Time...\n");
+
+                switch(pTexture->GetType())
+                {
+                    case D3DRTYPE_TEXTURE:
+                    {
+                        printf("_TEXTURE\n");
+                        sprintf(szBuffer, "C:\\Aaron\\Textures\\PushBufferTex - %.03d (0x%.08X).bmp", dwDumpTexture++, pTexture);
+
+                        ((IDirect3DTexture8*)pTexture)->UnlockRect(0);
+
+                        D3DXSaveTextureToFile(szBuffer, D3DXIFF_BMP, pTexture, NULL);
+                    }
+                    break;
+
+                    case D3DRTYPE_CUBETEXTURE:
+                    {
+                        printf("_CUBETEXTURE\n");
+                        for(int face=0;face<6;face++)
+                        {
+                            sprintf(szBuffer, "C:\\Aaron\\Textures\\PushBufferTexCube%d - %.03d (0x%.08X).bmp", face, dwDumpTexture++, pTexture);
+
+                            ((IDirect3DCubeTexture8*)pTexture)->UnlockRect((D3DCUBEMAP_FACES)face, 0);
+                            
+                            D3DXSaveTextureToFile(szBuffer, D3DXIFF_BMP, pTexture, NULL);
+                        }
+                    }
+                    break;
+                }
+            }
+            //*/
+
+<<
+
 >>
 
     /* TODO: Use new handle wrapping code
diff --git a/Include/Cxbx.h b/Include/Cxbx.h
index cf5c0b3f2..45e46f2cb 100644
--- a/Include/Cxbx.h
+++ b/Include/Cxbx.h
@@ -68,15 +68,15 @@ typedef signed long    sint32;
 // define this to trace warnings
 #define _DEBUG_WARNINGS
 
-// define this to dump textures @ SetTexture
+// define these to dump textures
 //#define _DEBUG_DUMP_TEXTURE_SETTEXTURE "C:\\Aaron\\Textures\\"
 //#define _DEBUG_DUMP_TEXTURE_REGISTER   "C:\\Aaron\\Textures\\"
 
 // version information
 #ifndef _DEBUG_TRACE
-#define _CXBX_VERSION "0.8.0-Pre2"
+#define _CXBX_VERSION "0.8.0-Pre3"
 #else
-#define _CXBX_VERSION "0.8.0-Pre2-Trace"
+#define _CXBX_VERSION "0.8.0-Pre3-Trace"
 #endif
 
 // round dwValue to the nearest multiple of dwMult
diff --git a/Source/Win32/CxbxKrnl/EmuD3D8/PushBuffer.cpp b/Source/Win32/CxbxKrnl/EmuD3D8/PushBuffer.cpp
index a2263392e..8329a1462 100644
--- a/Source/Win32/CxbxKrnl/EmuD3D8/PushBuffer.cpp
+++ b/Source/Win32/CxbxKrnl/EmuD3D8/PushBuffer.cpp
@@ -162,8 +162,6 @@ extern void XTL::EmuExecutePushBufferRaw
 
             pVertexData = ++pdwPushData;
 
-            // TODO: debug print out of vertex data
-
             pdwPushData += dwCount;
 
             // retrieve vertex shader
@@ -180,7 +178,20 @@ extern void XTL::EmuExecutePushBufferRaw
                 dwVertexShader = -1;
             }
 
+            //
             // calculate stride
+            //
+
+            dwStride = 0;
+
+            if(!VshHandleIsVertexShader(dwVertexShader))
+            {
+                if(dwVertexShader & D3DFVF_XYZRHW) { dwStride += sizeof(FLOAT)*4; }
+                if(dwVertexShader & D3DFVF_DIFFUSE) { dwStride += sizeof(DWORD); }
+                if(dwVertexShader & D3DFVF_SPECULAR) { dwStride += sizeof(DWORD); }
+
+                dwStride += ((dwVertexShader & D3DFVF_TEXCOUNT_MASK) >> D3DFVF_TEXCOUNT_SHIFT)*sizeof(FLOAT)*2;
+            }
 
             /*
             // create cached vertex buffer only once, with maxed out size
@@ -208,10 +219,44 @@ extern void XTL::EmuExecutePushBufferRaw
             }
             */
 
+            #ifdef _DEBUG_TRACK_PB
+            if(bShowPB)
+            {
+                printf("NVPB_InlineVertexArray(...)\n");
+                printf("  dwCount : %d\n", dwCount);
+                printf("  dwVertexShader : 0x%08X\n", dwVertexShader);
+            }
+            #endif
+
             // render vertices
             if(dwVertexShader != -1)
             {
-                g_pD3DDevice8->DrawPrimitiveUP(PCPrimitiveType, dwCount / dwStride, pVertexData, dwStride);
+                UINT VertexCount = (dwCount*sizeof(DWORD)) / dwStride;
+                UINT PrimitiveCount = EmuD3DVertex2PrimitiveCount(XBPrimitiveType, VertexCount);
+
+                VertexPatchDesc VPDesc;
+
+                VPDesc.dwVertexCount = VertexCount;
+                VPDesc.PrimitiveType = XBPrimitiveType;
+                VPDesc.dwPrimitiveCount = PrimitiveCount;
+                VPDesc.dwOffset = 0;
+                VPDesc.pVertexStreamZeroData = pVertexData;
+                VPDesc.uiVertexStreamZeroStride = dwStride;
+                VPDesc.hVertexShader = dwVertexShader;
+
+                VertexPatcher VertPatch;
+
+                bool bPatched = VertPatch.Apply(&VPDesc);
+
+                g_pD3DDevice8->DrawPrimitiveUP
+                (
+                    PCPrimitiveType,
+                    VPDesc.dwPrimitiveCount,
+                    VPDesc.pVertexStreamZeroData,
+                    VPDesc.uiVertexStreamZeroStride
+                );
+
+                VertPatch.Restore();
             }
 
             pdwPushData--;
diff --git a/Source/Win32/CxbxKrnl/HLEDataBase/Xapi.1.0.4627.inl b/Source/Win32/CxbxKrnl/HLEDataBase/Xapi.1.0.4627.inl
index 82b926ea3..d6c6d8a8c 100644
--- a/Source/Win32/CxbxKrnl/HLEDataBase/Xapi.1.0.4627.inl
+++ b/Source/Win32/CxbxKrnl/HLEDataBase/Xapi.1.0.4627.inl
@@ -321,36 +321,6 @@ SOOVPA<12> XInputGetState_1_0_4928 =
     -1, // XRef Not Saved
     0,  // XRef Not Used
 
-    {
-        { 0x0B, 0x8B }, // (Offset,Value)-Pair #1
-        { 0x0C, 0x1D }, // (Offset,Value)-Pair #2
-        { 0x0D, 0xAC }, // (Offset,Value)-Pair #3
-
-        { 0x25, 0xF6 }, // (Offset,Value)-Pair #4
-        { 0x26, 0x46 }, // (Offset,Value)-Pair #5
-        { 0x27, 0x04 }, // (Offset,Value)-Pair #6
-        { 0x28, 0x02 }, // (Offset,Value)-Pair #7
-
-        { 0x39, 0xF3 }, // (Offset,Value)-Pair #8
-        { 0x3A, 0xAB }, // (Offset,Value)-Pair #9
-
-        { 0x90, 0xBF }, // (Offset,Value)-Pair #10
-        { 0x91, 0xE8 }, // (Offset,Value)-Pair #11
-        { 0x92, 0x7B }, // (Offset,Value)-Pair #12
-    }
-};
-
-// ******************************************************************
-// * XInputSetState
-// ******************************************************************
-SOOVPA<12> XInputSetState_1_0_4928 =
-{
-    0,  // Large == 0
-    12, // Count == 12
-
-    -1, // XRef Not Saved
-    0,  // XRef Not Used
-
     {
         { 0x0E, 0x8B }, // (Offset,Value)-Pair #1
         { 0x0F, 0x8A }, // (Offset,Value)-Pair #2
@@ -371,6 +341,36 @@ SOOVPA<12> XInputSetState_1_0_4928 =
     }
 };
 
+// ******************************************************************
+// * XInputGetCapabilities
+// ******************************************************************
+SOOVPA<12> XInputGetCapabilities_1_0_4928 =
+{
+    0,  // Large == 0
+    12, // Count == 12
+
+    -1, // XRef Not Saved
+    0,  // XRef Not Used
+
+    {
+        { 0x0B, 0x8B }, // (Offset,Value)-Pair #1
+        { 0x0C, 0x1D }, // (Offset,Value)-Pair #2
+        { 0x0D, 0xAC }, // (Offset,Value)-Pair #3
+
+        { 0x25, 0xF6 }, // (Offset,Value)-Pair #4
+        { 0x26, 0x46 }, // (Offset,Value)-Pair #5
+        { 0x27, 0x04 }, // (Offset,Value)-Pair #6
+        { 0x28, 0x02 }, // (Offset,Value)-Pair #7
+
+        { 0x39, 0xF3 }, // (Offset,Value)-Pair #8
+        { 0x3A, 0xAB }, // (Offset,Value)-Pair #9
+
+        { 0x90, 0xBF }, // (Offset,Value)-Pair #10
+        { 0x91, 0xE8 }, // (Offset,Value)-Pair #11
+        { 0x92, 0x7B }, // (Offset,Value)-Pair #12
+    }
+};
+
 // ******************************************************************
 // * XAPI_1_0_4627
 // ******************************************************************
@@ -614,14 +614,14 @@ OOVPATable XAPI_1_0_4627[] =
         "EmuXInputGetState"
         #endif
     },
-    // XInputSetState
+    // XInputGetCapabilities
     {
-        (OOVPA*)&XInputSetState_1_0_4928,
+        (OOVPA*)&XInputGetCapabilities_1_0_4928,
 
-        XTL::EmuXInputSetState,
+        XTL::EmuXInputGetCapabilities,
 
         #ifdef _DEBUG_TRACE
-        "EmuXInputSetState"
+        "EmuXInputGetCapabilities"
         #endif
     },
     // XInputSetState (* unchanged since 1.0.4361 *)