Cycles: disable fast math flags, only use a subset.

Empty BVH nodes are set to NaN which must be preserved all the way to the
tnear <= tfar test which can then give false for empty nodes. This needs
strict semantices and careful argument ordering for min() and max(), so
the second argument is used if either of the arguments is NaN.

Fixes T52635: crash in BVH traversal with SSE4.1.

Differential Revision: https://developer.blender.org/D2828

1 files changed, 12 insertions, 12 deletions

diff --git a/intern/cycles/kernel/bvh/bvh_nodes.h b/intern/cycles/kernel/bvh/bvh_nodes.h
index 5f1dd434a44..6c33dad5426 100644
--- a/intern/cycles/kernel/bvh/bvh_nodes.h
+++ b/intern/cycles/kernel/bvh/bvh_nodes.h
@@ -52,8 +52,8 @@ ccl_device_forceinline int bvh_aligned_node_intersect(KernelGlobals *kg,
 	float c0hiy = (node1.z - P.y) * idir.y;
 	float c0loz = (node2.x - P.z) * idir.z;
 	float c0hiz = (node2.z - P.z) * idir.z;
-	float c0min = max4(min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz), 0.0f);
-	float c0max = min4(max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz), t);
+	float c0min = max4(0.0f, min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz));
+	float c0max = min4(t, max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz));
 
 	float c1lox = (node0.y - P.x) * idir.x;
 	float c1hix = (node0.w - P.x) * idir.x;
@@ -61,8 +61,8 @@ ccl_device_forceinline int bvh_aligned_node_intersect(KernelGlobals *kg,
 	float c1hiy = (node1.w - P.y) * idir.y;
 	float c1loz = (node2.y - P.z) * idir.z;
 	float c1hiz = (node2.w - P.z) * idir.z;
-	float c1min = max4(min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz), 0.0f);
-	float c1max = min4(max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz), t);
+	float c1min = max4(0.0f, min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz));
+	float c1max = min4(t, max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz));
 
 	dist[0] = c0min;
 	dist[1] = c1min;
@@ -101,8 +101,8 @@ ccl_device_forceinline int bvh_aligned_node_intersect_robust(KernelGlobals *kg,
 	float c0hiy = (node1.z - P.y) * idir.y;
 	float c0loz = (node2.x - P.z) * idir.z;
 	float c0hiz = (node2.z - P.z) * idir.z;
-	float c0min = max4(min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz), 0.0f);
-	float c0max = min4(max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz), t);
+	float c0min = max4(0.0f, min(c0lox, c0hix), min(c0loy, c0hiy), min(c0loz, c0hiz));
+	float c0max = min4(t, max(c0lox, c0hix), max(c0loy, c0hiy), max(c0loz, c0hiz));
 
 	float c1lox = (node0.y - P.x) * idir.x;
 	float c1hix = (node0.w - P.x) * idir.x;
@@ -110,8 +110,8 @@ ccl_device_forceinline int bvh_aligned_node_intersect_robust(KernelGlobals *kg,
 	float c1hiy = (node1.w - P.y) * idir.y;
 	float c1loz = (node2.y - P.z) * idir.z;
 	float c1hiz = (node2.w - P.z) * idir.z;
-	float c1min = max4(min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz), 0.0f);
-	float c1max = min4(max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz), t);
+	float c1min = max4(0.0f, min(c1lox, c1hix), min(c1loy, c1hiy), min(c1loz, c1hiz));
+	float c1max = min4(t, max(c1lox, c1hix), max(c1loy, c1hiy), max(c1loz, c1hiz));
 
 	if(difl != 0.0f) {
 		float hdiff = 1.0f + difl;
@@ -483,8 +483,8 @@ ccl_device_forceinline int bvh_unaligned_node_intersect(KernelGlobals *kg,
 	ssef tfar_y = max(lower_y, upper_y);
 	ssef tfar_z = max(lower_z, upper_z);
 
-	const ssef tnear = max4(tnear_x, tnear_y, tnear_z, isect_near);
-	const ssef tfar = min4(tfar_x, tfar_y, tfar_z, isect_far);
+	const ssef tnear = max4(isect_near, tnear_x, tnear_y, tnear_z);
+	const ssef tfar = min4(isect_far, tfar_x, tfar_y, tfar_z);
 	sseb vmask = tnear <= tfar;
 	dist[0] = tnear.f[0];
 	dist[1] = tnear.f[1];
@@ -545,8 +545,8 @@ ccl_device_forceinline int bvh_unaligned_node_intersect_robust(KernelGlobals *kg
 	ssef tfar_y = max(lower_y, upper_y);
 	ssef tfar_z = max(lower_z, upper_z);
 
-	const ssef tnear = max4(tnear_x, tnear_y, tnear_z, isect_near);
-	const ssef tfar = min4(tfar_x, tfar_y, tfar_z, isect_far);
+	const ssef tnear = max4(isect_near, tnear_x, tnear_y, tnear_z);
+	const ssef tfar = min4(isect_far, tfar_x, tfar_y, tfar_z);
 	sseb vmask;
 	if(difl != 0.0f) {
 		const float round_down = 1.0f - difl;