Cycles: simplify handling of ray differentials

* Store compact ray differentials in ShaderData and compute full differentials
  on demand. This reduces register pressure on the GPU.
* Remove BSDF differential code that was effectively doing nothing as the
  differential orientation was discarded when making it compact.

This gives a 1-5% speedup with RTX A6000 + OptiX in our benchmarks, with the
bigger speedups in simpler scenes.

Renders appear to be identical except for the Both displacement option that
does both displacement and bump.

Differential Revision: https://developer.blender.org/D15677

1 files changed, 4 insertions, 4 deletions

diff --git a/intern/cycles/kernel/types.h b/intern/cycles/kernel/types.h
index 252c27a40f0..59ea6c64be7 100644
--- a/intern/cycles/kernel/types.h
+++ b/intern/cycles/kernel/types.h
@@ -873,10 +873,10 @@ typedef struct ccl_align(16) ShaderData
   float ray_length;
 
 #ifdef __RAY_DIFFERENTIALS__
-  /* differential of P. these are orthogonal to Ng, not N */
-  differential3 dP;
-  /* differential of I */
-  differential3 dI;
+  /* Radius of differential of P. */
+  float dP;
+  /* Radius of differential of I. */
+  float dI;
   /* differential of u, v */
   differential du;
   differential dv;