archive/blender-archive
Archived
1
1
Fork 0
Code Pull Requests Packages Activity

Code cleanup: move rng into path state.

Browse Source 
Also pass by value and don't write back now that it is just a hash for seeding
and no longer an LCG state. Together this makes CUDA a tiny bit faster in my
tests, but mainly simplifies code.
This commit is contained in:
Brecht Van Lommel
2017-08-19 04:11:25 +02:00
parent 1cc4033df8
commit cfa8b762e2
28 changed files with 192 additions and 271 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
intern/cycles/kernel/split/kernel_branched.h
View File

@@ -110,7 +110,6 @@ ccl_device_noinline bool kernel_split_branched_path_surface_indirect_light_iter(
SplitBranchedState *branched_state = &kernel_split_state.branched_state[ray_index];
ShaderData *sd = saved_sd;
RNG rng = kernel_split_state.rng[ray_index];
PathRadiance *L = &kernel_split_state.path_radiance[ray_index];
float3 throughput = branched_state->throughput;
ccl_global PathState *ps = &kernel_split_state.path_state[ray_index];
@@ -157,20 +156,20 @@ ccl_device_noinline bool kernel_split_branched_path_surface_indirect_light_iter(
num_samples = ceil_to_int(num_samples_adjust*num_samples);
float num_samples_inv = num_samples_adjust/num_samples;
RNG bsdf_rng = cmj_hash(rng, i);
for(int j = branched_state->next_sample; j < num_samples; j++) {
if(reset_path_state) {
*ps = branched_state->path_state;
}
ps->rng_hash = cmj_hash(branched_state->path_state.rng_hash, i);
ccl_global float3 *tp = &kernel_split_state.throughput[ray_index];
*tp = throughput;
ccl_global Ray *bsdf_ray = &kernel_split_state.ray[ray_index];
if(!kernel_branched_path_surface_bounce(kg,
&bsdf_rng,
sd,
sc,
j,
@@ -184,6 +183,8 @@ ccl_device_noinline bool kernel_split_branched_path_surface_indirect_light_iter(
continue;
}
ps->rng_hash = branched_state->path_state.rng_hash;
/* update state for next iteration */
branched_state->next_closure = i;
branched_state->next_sample = j+1;