On 20 August 2015 at 10:03, Stefan Dösinger stefan@codeweavers.com wrote:
@@ -6332,7 +6332,8 @@ static GLuint shader_glsl_generate_ffp_fragment_shader(struct shader_glsl_priv * }
if (settings->color_key_enabled)
shader_addline(buffer, "if (all(equal(tex0, color_key))) discard;\n");
shader_addline(buffer, "if (all(lessThan(abs(tex0 - color_key), vec4(%s)))) discard;\n",wined3d_color_key_precision);
...
+/* Normalization of B5G6R5_UNORM textures is horribly imprecise if we don't have
- GL_RGB_565 support. 1 / 256 (~0.0039) works in practice, but is awfully close
- to the next possible value in 8 bit formats. 1 / 384 is too precise for some
- 5 and 6 bit channel values at least on Nvidia.
- An exact comparison isn't reliable for any format, except for the normalized
- values 0.0 and 1.0. */
+const char *wined3d_color_key_precision = "0.003";
Shouldn't you just pass the color key as a range to the shader instead? E.g. in a 6 bpc format anything between 0.484f and 0.500f would get quantized to 0x1f. (Which, other floating point issues aside, would correspond to about 0.008 slop here. For a 5 bpc format you'd even have 0.016.)
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Am 2015-08-20 um 11:06 schrieb Henri Verbeet:
Shouldn't you just pass the color key as a range to the shader instead? E.g. in a 6 bpc format anything between 0.484f and 0.500f would get quantized to 0x1f. (Which, other floating point issues aside, would correspond to about 0.008 slop here. For a 5 bpc format you'd even have 0.016.)
That would mean two uniforms instead of one, but could help with range color keys. I'll investigate the idea.
Otoh the idea of re-testing this stuff on old those GPUs doesn't really appeal to me :-( .
On 20 August 2015 at 12:28, Stefan Dösinger stefandoesinger@gmail.com wrote:
Am 2015-08-20 um 11:06 schrieb Henri Verbeet:
Shouldn't you just pass the color key as a range to the shader instead? E.g. in a 6 bpc format anything between 0.484f and 0.500f would get quantized to 0x1f. (Which, other floating point issues aside, would correspond to about 0.008 slop here. For a 5 bpc format you'd even have 0.016.)
That would mean two uniforms instead of one, but could help with range color keys. I'll investigate the idea.
You could of course also derive the slop from the format's channel sizes, but it would need to be a vector because of formats like WINED3DFMT_B5G6R5_UNORM, still need an extra uniform, and not make range color keys any easier.
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
Am 2015-08-20 um 12:50 schrieb Henri Verbeet:
You could of course also derive the slop from the format's channel sizes, but it would need to be a vector because of formats like WINED3DFMT_B5G6R5_UNORM, still need an extra uniform, and not make range color keys any easier.
Ya, and the hardcoded 0.003 also needs a uniform, so we're not winning anything there. Besides, we have plenty of uniforms for blits and fixed function fragment processing.
-
Henri Verbeet -
Stefan Dösinger