In T54768#496256, @Philipp Oeser (lichtwerk) wrote:

Thanx for investigating.
Could you please attach a .blend as well so someone can reproduce with your geometry?

I have attached a test scene.

Not really sure this is the solution to all issues in question (havent checked any code and your first example pic seems unrelated to the example scene?)

BUT: in the provided scene, there is a gap between faces (right where you get black pixels).
If I close this gap (by slightly moving one edge up in Z and merge vertices), those black pixels are gone...

Not sure this is a bug?
If this is unrelated, please provide another example (without gaps in the geometry, where this issue also shows...)

Marking as incomplete for the time being...

Dear @Philipp Oeser (lichtwerk) , thanks for your answer. Yes, you are right in some places I have gaps in the mesh, that's why positions were wrong. But in most cases 3D positions are almost correct, I get deviations from correct positions in ~cm or ~mm range. After careful debugging, I have noticed, this happens due to place in file "blender/intern/cycles/kernel/geom/geom_triangle_intersect.h", line 221:

if(det != 0.0f) {

		/* If determinant is zero it means ray lies in the plane of
		 * the triangle. It is possible in theory due to watertight
		 * nature of triangle intersection. For such cases we simply
		 * don't refine intersection hoping it'll go all fine.
		 */
		float rt = dot(edge2, qvec) / det;
		P = P + D*rt;

}

Mainly, I am interested in sentence:

For such cases we simply don't refine intersection hoping it'll go all fine.

In my case it doesnt go fine. And I am getting wrong(almost correct) 3D positions with wrong barycentric coordinates.

Noticed that function ray_triangle_intersect() in file "\cycles\util\util_math_intersect.h" returns sometimes wrong barycentric coordinates (u + v > 1, which is wrong), hence it renders incorrect 3D position for current pixel. Wrong barycentric coordinates are there due to fact that in line 182 (const float inv_den = 1.0f / den;) happens division by very small number.

Question:
• It is clear that in function ray_triangle_intersect() you are getting sometimes not really correct values, due to division on very small number, that's why you use function triangle_refine() to shoot ray second time to refine position. But it is not clear why it is not refined completely? Division on very small numbers still is not handled properly.
• How one can get correct barycentric coordinates?