From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Sat, 21 Jul 2018 21:38:43 +0000 (+0200)
Subject: Hack around a convergence problem in the SOR code, by stopping the over-relaxation... 
X-Git-Tag: 1.8.0~76^2~205
X-Git-Url: https://git.sesse.net/?a=commitdiff_plain;h=37ed9122ae6d6e207439c73115810db47cdea490;p=nageru

Hack around a convergence problem in the SOR code, by stopping the over-relaxation and reverting to pure Gauss-Seidel.
---

diff --git a/sor.frag b/sor.frag
index 25f0ce5..6c5333f 100644
--- a/sor.frag
+++ b/sor.frag
@@ -45,7 +45,10 @@ void main()
 	b += smooth_d * textureOffset(diff_flow_tex, tc, ivec2( 0, -1)).xy;
 	b += smooth_u * textureOffset(diff_flow_tex, tc, ivec2( 0,  1)).xy;
 
-	const float omega = 1.6;
+	// FIXME: omega=1.6 seems to make our entire system diverge.
+	// Is this because we do Gauss-Seidel instead of Jacobi?
+	// Figure out what's going on.
+	const float omega = 1.0;
 	diff_flow = texture(diff_flow_tex, tc).xy;
 
 	// From https://en.wikipedia.org/wiki/Successive_over-relaxation.