nv2a: Add ARL-bias to work around OpenGL float behaviour

hw/xbox/nv2a/nv2a_vsh.c

@@ -627,7 +627,16 @@ static const char* vsh_header =
     "#define ARL(dest, src) dest = _ARL(_in(src).x)\n"
     "int _ARL(float src)\n"
     "{\n"
-    "  return int(floor(src));\n"
+    "  /* Xbox GPU does specify rounding, OpenGL doesn't; so we need a bias.\n"
+    "   * Example: We probably want to floor 16.99.. to 17, not 16.\n"
+    "   * Source of error (why we get 16.99.. instead of 17.0) is typically\n"
+    "   * vertex-attributes being normalized from a byte value to float:\n"
+    "   *   17 / 255 = 0.06666.. so is this 0.06667 (ceil) or 0.06666 (floor)?\n"
+    "   * Which value we get depends on the host GPU.\n"
+    "   * If we multiply these rounded values by 255 later, we get:\n"
+    "   *   17.00 (ARL result = 17) or 16.99 (ARL result = 16).\n"
+    "   * We assume the intend was to get 17, so we add our bias to fix it. */\n"
+    "  return int(floor(src + 0.001));\n"
     "}\n"
     "\n"
     "#define SGE(dest, mask, src0, src1) dest.mask = _SGE(_in(src0), _in(src1)).mask\n"