NaN (divide by zero) fix for issue #561 and #790 (#891)

Summary:
https://github.com/facebookresearch/pytorch3d/issues/561
https://github.com/facebookresearch/pytorch3d/issues/790
Divide by zero fix (NaN fix).  When perspective_correct=True, BarycentricPerspectiveCorrectionForward and BarycentricPerspectiveCorrectionBackward in ../csrc/utils/geometry_utils.cuh are called.  The denominator (denom) values should not be allowed to go to zero. I'm able to resolve this issue locally with this PR and submit it for the team's review.

Pull Request resolved: https://github.com/facebookresearch/pytorch3d/pull/891

Reviewed By: patricklabatut

Differential Revision: D31829695

Pulled By: bottler

fbshipit-source-id: a3517b8362f6e60d48c35731258d8ce261b1d912

pytorch3d/csrc/utils/geometry_utils.cuh

@@ -177,7 +177,7 @@ __device__ inline float3 BarycentricPerspectiveCorrectionForward(
   const float w0_top = bary.x * z1 * z2;
   const float w1_top = z0 * bary.y * z2;
   const float w2_top = z0 * z1 * bary.z;
-  const float denom = w0_top + w1_top + w2_top;
+  const float denom = fmaxf(w0_top + w1_top + w2_top, kEpsilon);
   const float w0 = w0_top / denom;
   const float w1 = w1_top / denom;
   const float w2 = w2_top / denom;
@@ -208,7 +208,7 @@ BarycentricPerspectiveCorrectionBackward(
   const float w0_top = bary.x * z1 * z2;
   const float w1_top = z0 * bary.y * z2;
   const float w2_top = z0 * z1 * bary.z;
-  const float denom = w0_top + w1_top + w2_top;
+  const float denom = fmaxf(w0_top + w1_top + w2_top, kEpsilon);
 
   // Now do backward pass
   const float grad_denom_top =

pytorch3d/csrc/utils/geometry_utils.h

@@ -198,7 +198,7 @@ inline vec3<T> BarycentricPerspectiveCorrectionForward(
   const T w0_top = bary.x * z1 * z2;
   const T w1_top = bary.y * z0 * z2;
   const T w2_top = bary.z * z0 * z1;
-  const T denom = w0_top + w1_top + w2_top;
+  const T denom = std::max<T>(w0_top + w1_top + w2_top, kEpsilon);
   const T w0 = w0_top / denom;
   const T w1 = w1_top / denom;
   const T w2 = w2_top / denom;
@@ -229,7 +229,7 @@ inline std::tuple<vec3<T>, T, T, T> BarycentricPerspectiveCorrectionBackward(
   const T w0_top = bary.x * z1 * z2;
   const T w1_top = bary.y * z0 * z2;
   const T w2_top = bary.z * z0 * z1;
-  const T denom = w0_top + w1_top + w2_top;
+  const T denom = std::max<T>(w0_top + w1_top + w2_top, kEpsilon);
 
   // Now do backward pass
   const T grad_denom_top =