Fix cloth stability when in perfect rest shape.

The way cloth is coded, structural springs are only effective when stretched, while bending springs act only when shrunk. However, when cloth is exactly in its rest shape, neither have any effect, and effectively don't exist for the implicit solver.

This creates a stability problem in the initial frames of the simulation, especially considering that gravity seems to act so precisely that it doesn't disturb the strict equality of lengths, so in parts of the cloth this springless state can continue for quite a while.

Here is an example of things going haywire because of this and some suspicious logic in collision code acting together: {F314558}

Changing the condition so that structural springs are active even at exactly rest length fixes this test case. The use of >= is also supported by the original paper that the cloth implementation in blender is based on.

Reviewers: lukastoenne

Reviewed By: lukastoenne

Projects: #bf_blender

Differential Revision: https://developer.blender.org/D2028

source/blender/physics/intern/implicit_blender.c

@@ -1586,8 +1586,11 @@ bool BPH_mass_spring_force_spring_linear(Implicit_Data *data, int i, int j, floa
 	
 	// calculate elonglation
 	spring_length(data, i, j, extent, dir, &length, vel);
-	
-	if (length > restlen || no_compress) {
+
+	/* This code computes not only the force, but also its derivative.
+	   Zero derivative effectively disables the spring for the implicit solver.
+	   Thus length > restlen makes cloth unconstrained at the start of simulation. */
+	if ((length >= restlen && length > 0) || no_compress) {
 		float stretch_force, f[3], dfdx[3][3], dfdv[3][3];
 		
 		stretch_force = stiffness * (length - restlen);