Cr8-xyz/blender
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Cycles CUDA: reduce branched path stack memory by sharing indirect ShaderData.

Browse Source 
Saves about 15% for the branched path kernel.
This commit is contained in:
Brecht Van Lommel
2016-05-24 22:28:03 +02:00
parent 7928030eff
commit b49185df99
3 changed files with 45 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
intern/cycles/kernel/kernel_path.h
View File

@@ -53,6 +53,7 @@
CCL_NAMESPACE_BEGIN
ccl_device void kernel_path_indirect(KernelGlobals *kg,
ShaderData *sd,
ShaderData *emission_sd,
RNG *rng,
Ray *ray,
@@ -61,9 +62,6 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
PathState *state,
PathRadiance *L)
{
/* shader data memory used for both volumes and surfaces, saves stack space */
ShaderData sd;
/* path iteration */
for(;;) {
/* intersect scene */
@@ -121,12 +119,12 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
VolumeSegment volume_segment;
shader_setup_from_volume(kg,
&sd,
sd,
&volume_ray);
kernel_volume_decoupled_record(kg,
state,
&volume_ray,
&sd,
sd,
&volume_segment,
heterogeneous);
@@ -149,7 +147,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
/* direct light sampling */
kernel_branched_path_volume_connect_light(kg,
rng,
&sd,
sd,
emission_sd,
throughput,
state,
@@ -167,7 +165,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
result = kernel_volume_decoupled_scatter(kg,
state,
&volume_ray,
&sd,
sd,
&throughput,
rphase,
rscatter,
@@ -182,7 +180,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
if(result == VOLUME_PATH_SCATTERED) {
if(kernel_path_volume_bounce(kg,
rng,
&sd,
sd,
&throughput,
state,
L,
@@ -203,14 +201,14 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
{
/* integrate along volume segment with distance sampling */
VolumeIntegrateResult result = kernel_volume_integrate(
kg, state, &sd, &volume_ray, L, &throughput, rng, heterogeneous);
kg, state, sd, &volume_ray, L, &throughput, rng, heterogeneous);
# ifdef __VOLUME_SCATTER__
if(result == VOLUME_PATH_SCATTERED) {
/* direct lighting */
kernel_path_volume_connect_light(kg,
rng,
&sd,
sd,
emission_sd,
throughput,
state,
@@ -219,7 +217,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
/* indirect light bounce */
if(kernel_path_volume_bounce(kg,
rng,
&sd,
sd,
&throughput,
state,
L,
@@ -251,13 +249,13 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
/* setup shading */
shader_setup_from_ray(kg,
&sd,
sd,
&isect,
ray);
float rbsdf = path_state_rng_1D_for_decision(kg, rng, state, PRNG_BSDF);
shader_eval_surface(kg, &sd, state, rbsdf, state->flag, SHADER_CONTEXT_INDIRECT);
shader_eval_surface(kg, sd, state, rbsdf, state->flag, SHADER_CONTEXT_INDIRECT);
#ifdef __BRANCHED_PATH__
shader_merge_closures(&sd);
shader_merge_closures(sd);
#endif
/* blurring of bsdf after bounces, for rays that have a small likelihood
@@ -267,15 +265,15 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
if(blur_pdf < 1.0f) {
float blur_roughness = sqrtf(1.0f - blur_pdf)*0.5f;
shader_bsdf_blur(kg, &sd, blur_roughness);
shader_bsdf_blur(kg, sd, blur_roughness);
}
}
#ifdef __EMISSION__
/* emission */
if(sd.flag & SD_EMISSION) {
if(sd->flag & SD_EMISSION) {
float3 emission = indirect_primitive_emission(kg,
&sd,
sd,
isect.t,
state->flag,
state->ray_pdf);
@@ -305,30 +303,30 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
#ifdef __AO__
/* ambient occlusion */
if(kernel_data.integrator.use_ambient_occlusion || (sd.flag & SD_AO)) {
if(kernel_data.integrator.use_ambient_occlusion || (sd->flag & SD_AO)) {
float bsdf_u, bsdf_v;
path_state_rng_2D(kg, rng, state, PRNG_BSDF_U, &bsdf_u, &bsdf_v);
float ao_factor = kernel_data.background.ao_factor;
float3 ao_N;
float3 ao_bsdf = shader_bsdf_ao(kg, &sd, ao_factor, &ao_N);
float3 ao_bsdf = shader_bsdf_ao(kg, sd, ao_factor, &ao_N);
float3 ao_D;
float ao_pdf;
float3 ao_alpha = make_float3(0.0f, 0.0f, 0.0f);
sample_cos_hemisphere(ao_N, bsdf_u, bsdf_v, &ao_D, &ao_pdf);
if(dot(sd.Ng, ao_D) > 0.0f && ao_pdf != 0.0f) {
if(dot(sd->Ng, ao_D) > 0.0f && ao_pdf != 0.0f) {
Ray light_ray;
float3 ao_shadow;
light_ray.P = ray_offset(sd.P, sd.Ng);
light_ray.P = ray_offset(sd->P, sd->Ng);
light_ray.D = ao_D;
light_ray.t = kernel_data.background.ao_distance;
# ifdef __OBJECT_MOTION__
light_ray.time = sd.time;
light_ray.time = sd->time;
# endif
light_ray.dP = sd.dP;
light_ray.dP = sd->dP;
light_ray.dD = differential3_zero();
if(!shadow_blocked(kg, emission_sd, state, &light_ray, &ao_shadow)) {
@@ -346,9 +344,9 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
#ifdef __SUBSURFACE__
/* bssrdf scatter to a different location on the same object, replacing
* the closures with a diffuse BSDF */
if(sd.flag & SD_BSSRDF) {
if(sd->flag & SD_BSSRDF) {
float bssrdf_probability;
ShaderClosure *sc = subsurface_scatter_pick_closure(kg, &sd, &bssrdf_probability);
ShaderClosure *sc = subsurface_scatter_pick_closure(kg, sd, &bssrdf_probability);
/* modify throughput for picking bssrdf or bsdf */
throughput *= bssrdf_probability;
@@ -364,7 +362,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
PRNG_BSDF_U,
&bssrdf_u, &bssrdf_v);
subsurface_scatter_step(kg,
&sd,
sd,
state,
state->flag,
sc,
@@ -380,7 +378,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
int all = kernel_data.integrator.sample_all_lights_indirect;
kernel_branched_path_surface_connect_light(kg,
rng,
&sd,
sd,
emission_sd,
state,
throughput,
@@ -390,7 +388,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
}
#endif
if(!kernel_path_surface_bounce(kg, rng, &sd, &throughput, state, L, ray))
if(!kernel_path_surface_bounce(kg, rng, sd, &throughput, state, L, ray))
break;
}
}