From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Sat, 25 Jan 2014 12:57:41 +0000 (+0100)
Subject: Don't dither alpha.
X-Git-Tag: 1.0~54
X-Git-Url: https://git.sesse.net/?p=movit;a=commitdiff_plain;h=88ee2edc41d536c7057424df185fc6b49f993896;ds=sidebyside

Don't dither alpha.

For alpha, correct rounding is more important to us than overall
spectral properties, simply because 255 -> 255 is such an important
case.
---

diff --git a/dither_effect.frag b/dither_effect.frag
index f9c6ad1..6994e84 100644
--- a/dither_effect.frag
+++ b/dither_effect.frag
@@ -3,10 +3,13 @@ uniform vec2 PREFIX(tc_scale);
 uniform float PREFIX(round_fac), PREFIX(inv_round_fac);
 
 vec4 FUNCNAME(vec2 tc) {
-	// We also choose to dither alpha, just in case.
-	// Maybe it should in theory have a separate dither,
-	// but I doubt it matters much.
-	vec4 result = INPUT(tc) + texture2D(PREFIX(dither_tex), tc * PREFIX(tc_scale)).xxxx;
+	vec4 result = INPUT(tc);
+
+	// Don't dither alpha; the case of alpha=255 (1.0) is very important to us,
+	// and if there's any inaccuracy earlier in the chain so that it becomes e.g.
+	// 254.8, it's better to just get it rounded off than to dither and have it
+	// possibly get down to 254. This is not the case for the color components.
+	result.rgb += texture2D(PREFIX(dither_tex), tc * PREFIX(tc_scale)).xxx;
 
 	// NEED_EXPLICIT_ROUND will be #defined to 1 if the GPU has inaccurate
 	// fp32 -> int8 framebuffer rounding, and 0 otherwise.