wined3d: Improve the unsigned emulation of WINED3DFMT_R5G5_SNORM_L6_UNORM.

dlls/wined3d/utils.c

@@ -294,25 +294,29 @@ static void convert_r5g5_snorm_l6_unorm(const BYTE *src, BYTE *dst, UINT src_row
         UINT dst_row_pitch, UINT dst_slice_pitch, UINT width, UINT height, UINT depth)
 {
     unsigned int x, y, z;
-    const WORD *Source;
+    unsigned char r_in, g_in, l_in;
+    const unsigned short *texel_in;
+    unsigned short *texel_out;
 
+    /* Emulating signed 5 bit values with unsigned 5 bit values has some precision problems by design:
+     * E.g. the signed input value 0 becomes 16. GL normalizes it to 16 / 31 = 0.516. We convert it
+     * back to a signed value by subtracting 0.5 and multiplying by 2.0. The resulting value is
+     * ((16 / 31) - 0.5) * 2.0 = 0.032, which is quite different from the intended result 0.000. */
     for (z = 0; z < depth; z++)
     {
         for (y = 0; y < height; y++)
         {
-            unsigned short *Dest_s = (unsigned short *) (dst + z * dst_slice_pitch + y * dst_row_pitch);
-            Source = (const WORD *)(src + z * src_slice_pitch + y * src_row_pitch);
+            texel_out = (unsigned short *) (dst + z * dst_slice_pitch + y * dst_row_pitch);
+            texel_in = (const unsigned short *)(src + z * src_slice_pitch + y * src_row_pitch);
             for (x = 0; x < width; x++ )
             {
-                short color = (*Source++);
-                unsigned char l = ((color >> 10) & 0xfc);
-                        short v = ((color >>  5) & 0x3e);
-                        short u = ((color      ) & 0x1f);
-                short v_conv = v + 16;
-                short u_conv = u + 16;
-
-                *Dest_s = ((v_conv << 11) & 0xf800) | ((l << 5) & 0x7e0) | (u_conv & 0x1f);
-                Dest_s += 1;
+                l_in = (*texel_in & 0xfc00) >> 10;
+                g_in = (*texel_in & 0x03e0) >> 5;
+                r_in = *texel_in & 0x001f;
+
+                *texel_out = ((r_in + 16) << 11) | (l_in << 5) | (g_in + 16);
+                texel_out++;
+                texel_in++;
             }
         }
     }