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

7 files changed, 47 insertions, 39 deletions

diff --git a/intern/cycles/kernel/svm/attribute.h b/intern/cycles/kernel/svm/attribute.h
index a3609d8b4b0..5f0d1609f08 100644
--- a/intern/cycles/kernel/svm/attribute.h
+++ b/intern/cycles/kernel/svm/attribute.h
@@ -140,6 +140,16 @@ ccl_device_noinline void svm_node_attr(KernelGlobals kg,
   }
 }
 
+ccl_device_forceinline float3 svm_node_bump_P_dx(const ccl_private ShaderData *sd)
+{
+  return sd->P + differential_from_compact(sd->Ng, sd->dP).dx;
+}
+
+ccl_device_forceinline float3 svm_node_bump_P_dy(const ccl_private ShaderData *sd)
+{
+  return sd->P + differential_from_compact(sd->Ng, sd->dP).dy;
+}
+
 ccl_device_noinline void svm_node_attr_bump_dx(KernelGlobals kg,
                                                ccl_private ShaderData *sd,
                                                ccl_private float *stack,
@@ -167,7 +177,7 @@ ccl_device_noinline void svm_node_attr_bump_dx(KernelGlobals kg,
 
   if (node.y == ATTR_STD_GENERATED && desc.element == ATTR_ELEMENT_NONE) {
     /* No generated attribute, fall back to object coordinates. */
-    float3 f = sd->P + sd->dP.dx;
+    float3 f = svm_node_bump_P_dx(sd);
     if (sd->object != OBJECT_NONE) {
       object_inverse_position_transform(kg, sd, &f);
     }
@@ -265,7 +275,7 @@ ccl_device_noinline void svm_node_attr_bump_dy(KernelGlobals kg,
 
   if (node.y == ATTR_STD_GENERATED && desc.element == ATTR_ELEMENT_NONE) {
     /* No generated attribute, fall back to object coordinates. */
-    float3 f = sd->P + sd->dP.dy;
+    float3 f = svm_node_bump_P_dy(sd);
     if (sd->object != OBJECT_NONE) {
       object_inverse_position_transform(kg, sd, &f);
     }
diff --git a/intern/cycles/kernel/svm/bump.h b/intern/cycles/kernel/svm/bump.h
index 566c45f5f25..1009a6a4241 100644
--- a/intern/cycles/kernel/svm/bump.h
+++ b/intern/cycles/kernel/svm/bump.h
@@ -14,23 +14,21 @@ ccl_device_noinline void svm_node_enter_bump_eval(KernelGlobals kg,
 {
   /* save state */
   stack_store_float3(stack, offset + 0, sd->P);
-  stack_store_float3(stack, offset + 3, sd->dP.dx);
-  stack_store_float3(stack, offset + 6, sd->dP.dy);
+  stack_store_float(stack, offset + 3, sd->dP);
 
   /* set state as if undisplaced */
   const AttributeDescriptor desc = find_attribute(kg, sd, ATTR_STD_POSITION_UNDISPLACED);
 
   if (desc.offset != ATTR_STD_NOT_FOUND) {
-    float3 P, dPdx, dPdy;
-    P = primitive_surface_attribute_float3(kg, sd, desc, &dPdx, &dPdy);
+    differential3 dP;
+    float3 P = primitive_surface_attribute_float3(kg, sd, desc, &dP.dx, &dP.dy);
 
     object_position_transform(kg, sd, &P);
-    object_dir_transform(kg, sd, &dPdx);
-    object_dir_transform(kg, sd, &dPdy);
+    object_dir_transform(kg, sd, &dP.dx);
+    object_dir_transform(kg, sd, &dP.dy);
 
     sd->P = P;
-    sd->dP.dx = dPdx;
-    sd->dP.dy = dPdy;
+    sd->dP = differential_make_compact(dP);
   }
 }
 
@@ -41,8 +39,7 @@ ccl_device_noinline void svm_node_leave_bump_eval(KernelGlobals kg,
 {
   /* restore state */
   sd->P = stack_load_float3(stack, offset + 0);
-  sd->dP.dx = stack_load_float3(stack, offset + 3);
-  sd->dP.dy = stack_load_float3(stack, offset + 6);
+  sd->dP = stack_load_float(stack, offset + 3);
 }
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/kernel/svm/displace.h b/intern/cycles/kernel/svm/displace.h
index 128023263fd..230f8c73820 100644
--- a/intern/cycles/kernel/svm/displace.h
+++ b/intern/cycles/kernel/svm/displace.h
@@ -24,18 +24,17 @@ ccl_device_noinline void svm_node_set_bump(KernelGlobals kg,
     float3 normal_in = stack_valid(normal_offset) ? stack_load_float3(stack, normal_offset) :
                                                     sd->N;
 
-    float3 dPdx = sd->dP.dx;
-    float3 dPdy = sd->dP.dy;
+    differential3 dP = differential_from_compact(sd->Ng, sd->dP);
 
     if (use_object_space) {
       object_inverse_normal_transform(kg, sd, &normal_in);
-      object_inverse_dir_transform(kg, sd, &dPdx);
-      object_inverse_dir_transform(kg, sd, &dPdy);
+      object_inverse_dir_transform(kg, sd, &dP.dx);
+      object_inverse_dir_transform(kg, sd, &dP.dy);
     }
 
     /* get surface tangents from normal */
-    float3 Rx = cross(dPdy, normal_in);
-    float3 Ry = cross(normal_in, dPdx);
+    float3 Rx = cross(dP.dy, normal_in);
+    float3 Ry = cross(normal_in, dP.dx);
 
     /* get bump values */
     uint c_offset, x_offset, y_offset, strength_offset;
@@ -46,7 +45,7 @@ ccl_device_noinline void svm_node_set_bump(KernelGlobals kg,
     float h_y = stack_load_float(stack, y_offset);
 
     /* compute surface gradient and determinant */
-    float det = dot(dPdx, Rx);
+    float det = dot(dP.dx, Rx);
     float3 surfgrad = (h_x - h_c) * Rx + (h_y - h_c) * Ry;
 
     float absdet = fabsf(det);
diff --git a/intern/cycles/kernel/svm/geometry.h b/intern/cycles/kernel/svm/geometry.h
index bbefdcfa755..cbd87d84409 100644
--- a/intern/cycles/kernel/svm/geometry.h
+++ b/intern/cycles/kernel/svm/geometry.h
@@ -54,7 +54,7 @@ ccl_device_noinline void svm_node_geometry_bump_dx(KernelGlobals kg,
 
   switch (type) {
     case NODE_GEOM_P:
-      data = sd->P + sd->dP.dx;
+      data = svm_node_bump_P_dx(sd);
       break;
     case NODE_GEOM_uv:
       data = make_float3(1.0f - sd->u - sd->du.dx - sd->v - sd->dv.dx, sd->u + sd->du.dx, 0.0f);
@@ -81,7 +81,7 @@ ccl_device_noinline void svm_node_geometry_bump_dy(KernelGlobals kg,
 
   switch (type) {
     case NODE_GEOM_P:
-      data = sd->P + sd->dP.dy;
+      data = svm_node_bump_P_dy(sd);
       break;
     case NODE_GEOM_uv:
       data = make_float3(1.0f - sd->u - sd->du.dy - sd->v - sd->dv.dy, sd->u + sd->du.dy, 0.0f);
diff --git a/intern/cycles/kernel/svm/tex_coord.h b/intern/cycles/kernel/svm/tex_coord.h
index 2a0130e11d4..8154c542e6f 100644
--- a/intern/cycles/kernel/svm/tex_coord.h
+++ b/intern/cycles/kernel/svm/tex_coord.h
@@ -106,7 +106,7 @@ ccl_device_noinline int svm_node_tex_coord_bump_dx(KernelGlobals kg,
 
   switch (type) {
     case NODE_TEXCO_OBJECT: {
-      data = sd->P + sd->dP.dx;
+      data = svm_node_bump_P_dx(sd);
       if (node.w == 0) {
         if (sd->object != OBJECT_NONE) {
           object_inverse_position_transform(kg, sd, &data);
@@ -130,9 +130,9 @@ ccl_device_noinline int svm_node_tex_coord_bump_dx(KernelGlobals kg,
       Transform tfm = kernel_data.cam.worldtocamera;
 
       if (sd->object != OBJECT_NONE)
-        data = transform_point(&tfm, sd->P + sd->dP.dx);
+        data = transform_point(&tfm, svm_node_bump_P_dx(sd));
       else
-        data = transform_point(&tfm, sd->P + sd->dP.dx + camera_position(kg));
+        data = transform_point(&tfm, svm_node_bump_P_dx(sd) + camera_position(kg));
       break;
     }
     case NODE_TEXCO_WINDOW: {
@@ -140,7 +140,7 @@ ccl_device_noinline int svm_node_tex_coord_bump_dx(KernelGlobals kg,
           kernel_data.cam.type == CAMERA_ORTHOGRAPHIC)
         data = camera_world_to_ndc(kg, sd, sd->ray_P);
       else
-        data = camera_world_to_ndc(kg, sd, sd->P + sd->dP.dx);
+        data = camera_world_to_ndc(kg, sd, svm_node_bump_P_dx(sd));
       data.z = 0.0f;
       break;
     }
@@ -160,7 +160,7 @@ ccl_device_noinline int svm_node_tex_coord_bump_dx(KernelGlobals kg,
       break;
     }
     case NODE_TEXCO_VOLUME_GENERATED: {
-      data = sd->P + sd->dP.dx;
+      data = svm_node_bump_P_dx(sd);
 
 #  ifdef __VOLUME__
       if (sd->object != OBJECT_NONE)
@@ -191,7 +191,7 @@ ccl_device_noinline int svm_node_tex_coord_bump_dy(KernelGlobals kg,
 
   switch (type) {
     case NODE_TEXCO_OBJECT: {
-      data = sd->P + sd->dP.dy;
+      data = svm_node_bump_P_dy(sd);
       if (node.w == 0) {
         if (sd->object != OBJECT_NONE) {
           object_inverse_position_transform(kg, sd, &data);
@@ -215,9 +215,9 @@ ccl_device_noinline int svm_node_tex_coord_bump_dy(KernelGlobals kg,
       Transform tfm = kernel_data.cam.worldtocamera;
 
       if (sd->object != OBJECT_NONE)
-        data = transform_point(&tfm, sd->P + sd->dP.dy);
+        data = transform_point(&tfm, svm_node_bump_P_dy(sd));
       else
-        data = transform_point(&tfm, sd->P + sd->dP.dy + camera_position(kg));
+        data = transform_point(&tfm, svm_node_bump_P_dy(sd) + camera_position(kg));
       break;
     }
     case NODE_TEXCO_WINDOW: {
@@ -225,7 +225,7 @@ ccl_device_noinline int svm_node_tex_coord_bump_dy(KernelGlobals kg,
           kernel_data.cam.type == CAMERA_ORTHOGRAPHIC)
         data = camera_world_to_ndc(kg, sd, sd->ray_P);
       else
-        data = camera_world_to_ndc(kg, sd, sd->P + sd->dP.dy);
+        data = camera_world_to_ndc(kg, sd, svm_node_bump_P_dy(sd));
       data.z = 0.0f;
       break;
     }
@@ -245,7 +245,7 @@ ccl_device_noinline int svm_node_tex_coord_bump_dy(KernelGlobals kg,
       break;
     }
     case NODE_TEXCO_VOLUME_GENERATED: {
-      data = sd->P + sd->dP.dy;
+      data = svm_node_bump_P_dy(sd);
 
 #  ifdef __VOLUME__
       if (sd->object != OBJECT_NONE)
diff --git a/intern/cycles/kernel/svm/types.h b/intern/cycles/kernel/svm/types.h
index 12d0ec141e6..98dfe6a4375 100644
--- a/intern/cycles/kernel/svm/types.h
+++ b/intern/cycles/kernel/svm/types.h
@@ -12,7 +12,7 @@ CCL_NAMESPACE_BEGIN
 /* SVM stack offsets with this value indicate that it's not on the stack */
 #define SVM_STACK_INVALID 255
 
-#define SVM_BUMP_EVAL_STATE_SIZE 9
+#define SVM_BUMP_EVAL_STATE_SIZE 4
 
 /* Nodes */
 
diff --git a/intern/cycles/kernel/svm/wireframe.h b/intern/cycles/kernel/svm/wireframe.h
index e5fe08e5d04..91fadf4cfc4 100644
--- a/intern/cycles/kernel/svm/wireframe.h
+++ b/intern/cycles/kernel/svm/wireframe.h
@@ -14,6 +14,7 @@ CCL_NAMESPACE_BEGIN
 
 ccl_device_inline float wireframe(KernelGlobals kg,
                                   ccl_private ShaderData *sd,
+                                  const differential3 dP,
                                   float size,
                                   int pixel_size,
                                   ccl_private float3 *P)
@@ -46,8 +47,8 @@ ccl_device_inline float wireframe(KernelGlobals kg,
     if (pixel_size) {
       // Project the derivatives of P to the viewing plane defined
       // by I so we have a measure of how big is a pixel at this point
-      float pixelwidth_x = len(sd->dP.dx - dot(sd->dP.dx, sd->I) * sd->I);
-      float pixelwidth_y = len(sd->dP.dy - dot(sd->dP.dy, sd->I) * sd->I);
+      float pixelwidth_x = len(dP.dx - dot(dP.dx, sd->I) * sd->I);
+      float pixelwidth_y = len(dP.dy - dot(dP.dy, sd->I) * sd->I);
       // Take the average of both axis' length
       pixelwidth = (pixelwidth_x + pixelwidth_y) * 0.5f;
     }
@@ -86,16 +87,17 @@ ccl_device_noinline void svm_node_wireframe(KernelGlobals kg,
   int pixel_size = (int)use_pixel_size;
 
   /* Calculate wireframe */
-  float f = wireframe(kg, sd, size, pixel_size, &sd->P);
+  const differential3 dP = differential_from_compact(sd->Ng, sd->dP);
+  float f = wireframe(kg, sd, dP, size, pixel_size, &sd->P);
 
   /* TODO(sergey): Think of faster way to calculate derivatives. */
   if (bump_offset == NODE_BUMP_OFFSET_DX) {
-    float3 Px = sd->P - sd->dP.dx;
-    f += (f - wireframe(kg, sd, size, pixel_size, &Px)) / len(sd->dP.dx);
+    float3 Px = sd->P - dP.dx;
+    f += (f - wireframe(kg, sd, dP, size, pixel_size, &Px)) / len(dP.dx);
   }
   else if (bump_offset == NODE_BUMP_OFFSET_DY) {
-    float3 Py = sd->P - sd->dP.dy;
-    f += (f - wireframe(kg, sd, size, pixel_size, &Py)) / len(sd->dP.dy);
+    float3 Py = sd->P - dP.dy;
+    f += (f - wireframe(kg, sd, dP, size, pixel_size, &Py)) / len(dP.dy);
   }
 
   if (stack_valid(out_fac))