Cycles: GPU code generation optimizations for direct lighting

Use a single loop to iterate over all lights, reducing divergence and amount
of code to generate. Moving ray intersection calls out of conditionals will
also help the Optix compiler.

Ref D5363

1 files changed, 34 insertions, 27 deletions

diff --git a/intern/cycles/kernel/kernel_light.h b/intern/cycles/kernel/kernel_light.h
index 758e91159b6..ce908ce0fe2 100644
--- a/intern/cycles/kernel/kernel_light.h
+++ b/intern/cycles/kernel/kernel_light.h
@@ -1076,7 +1076,7 @@ ccl_device int light_distribution_sample(KernelGlobals *kg, float *randu)
   int len = kernel_data.integrator.num_distribution + 1;
   float r = *randu;
 
-  while (len > 0) {
+  do {
     int half_len = len >> 1;
     int middle = first + half_len;
 
@@ -1087,7 +1087,7 @@ ccl_device int light_distribution_sample(KernelGlobals *kg, float *randu)
       first = middle + 1;
       len = len - half_len - 1;
     }
-  }
+  } while (len > 0);
 
   /* Clamping should not be needed but float rounding errors seem to
    * make this fail on rare occasions. */
@@ -1104,42 +1104,49 @@ ccl_device int light_distribution_sample(KernelGlobals *kg, float *randu)
 
 /* Generic Light */
 
-ccl_device bool light_select_reached_max_bounces(KernelGlobals *kg, int index, int bounce)
+ccl_device_inline bool light_select_reached_max_bounces(KernelGlobals *kg, int index, int bounce)
 {
   return (bounce > kernel_tex_fetch(__lights, index).max_bounces);
 }
 
-ccl_device_noinline bool light_sample(
-    KernelGlobals *kg, float randu, float randv, float time, float3 P, int bounce, LightSample *ls)
+ccl_device_noinline bool light_sample(KernelGlobals *kg,
+                                      int lamp,
+                                      float randu,
+                                      float randv,
+                                      float time,
+                                      float3 P,
+                                      int bounce,
+                                      LightSample *ls)
 {
-  /* sample index */
-  int index = light_distribution_sample(kg, &randu);
-
-  /* fetch light data */
-  const ccl_global KernelLightDistribution *kdistribution = &kernel_tex_fetch(__light_distribution,
-                                                                              index);
-  int prim = kdistribution->prim;
-
-  if (prim >= 0) {
-    int object = kdistribution->mesh_light.object_id;
-    int shader_flag = kdistribution->mesh_light.shader_flag;
+  if (lamp < 0) {
+    /* sample index */
+    int index = light_distribution_sample(kg, &randu);
+
+    /* fetch light data */
+    const ccl_global KernelLightDistribution *kdistribution = &kernel_tex_fetch(
+        __light_distribution, index);
+    int prim = kdistribution->prim;
+
+    if (prim >= 0) {
+      int object = kdistribution->mesh_light.object_id;
+      int shader_flag = kdistribution->mesh_light.shader_flag;
+
+      triangle_light_sample(kg, prim, object, randu, randv, time, ls, P);
+      ls->shader |= shader_flag;
+      return (ls->pdf > 0.0f);
+    }
 
-    triangle_light_sample(kg, prim, object, randu, randv, time, ls, P);
-    ls->shader |= shader_flag;
-    return (ls->pdf > 0.0f);
+    lamp = -prim - 1;
   }
-  else {
-    int lamp = -prim - 1;
 
-    if (UNLIKELY(light_select_reached_max_bounces(kg, lamp, bounce))) {
-      return false;
-    }
-
-    return lamp_light_sample(kg, lamp, randu, randv, P, ls);
+  if (UNLIKELY(light_select_reached_max_bounces(kg, lamp, bounce))) {
+    return false;
   }
+
+  return lamp_light_sample(kg, lamp, randu, randv, P, ls);
 }
 
-ccl_device int light_select_num_samples(KernelGlobals *kg, int index)
+ccl_device_inline int light_select_num_samples(KernelGlobals *kg, int index)
 {
   return kernel_tex_fetch(__lights, index).samples;
 }