nv2a: Use trunc in vertex rounding instead of floor

Xbox seems to truncate instead of flooring, which can be inferred from
interpolated depth buffer values.

hw/xbox/nv2a/pgraph/glsl/vsh-ff.c

@@ -474,7 +474,7 @@ GLSL_DEFINE(materialEmissionColor, GLSL_LTCTXA(NV_IGRAPH_XF_LTCTXA_CM_COL) ".xyz
     "  oPos.w = clampAwayZeroInf(oPos.w);\n"
     "  oPos.xy /= oPos.w;\n"
     "  oPos.xy += c[" stringify(NV_IGRAPH_XF_XFCTX_VPOFF) "].xy;\n"
-    "  oPos.xy = floor(oPos.xy * 16.0f) / 16.0f;\n"
+    "  oPos.xy = roundScreenCoords(oPos.xy);\n"
     "  oPos.xy = (2.0f * oPos.xy - surfaceSize) / surfaceSize;\n"
     "  oPos.xy *= oPos.w;\n"
     );

hw/xbox/nv2a/pgraph/glsl/vsh-prog.c

@@ -822,11 +822,9 @@ void pgraph_gen_vsh_prog_glsl(uint16_t version,
 
     mstring_append(body,
         /* The shaders leave the result in screen space, while OpenGL expects it
-         * in clip space. Xbox NV2A rasterizer appears to have 4 bit precision
-         * fixed point fractional part and to convert floating point coordinates
-         * by flooring.
+         * in clip space.
          */
-        "  oPos.xy = floor(oPos.xy * 16.0f) / 16.0f;\n"
+        "  oPos.xy = roundScreenCoords(oPos.xy);\n"
         "  oPos.xy = (2.0f * oPos.xy - surfaceSize) / surfaceSize;\n"
 
         "  oPos.z = oPos.z / clipRange.y;\n"

hw/xbox/nv2a/pgraph/glsl/vsh.c

@@ -85,6 +85,13 @@ MString *pgraph_gen_vsh_glsl(const ShaderState *state, bool prefix_outputs)
         "\n"
         "vec4 NaNToOne(vec4 src) {\n"
         "  return mix(src, vec4(1.0), isnan(src));\n"
+        "}\n"
+        "\n"
+        // Xbox NV2A rasterizer appears to have 4 bit precision fixed-point
+        // fractional part and to convert floating-point coordinates by
+        // by truncating (not flooring).
+        "vec2 roundScreenCoords(vec2 pos) {\n"
+        "  return trunc(pos * 16.0f) / 16.0f;\n"
         "}\n");
 
     pgraph_get_glsl_vtx_header(header, state->vulkan, state->smooth_shading,