[Bf-blender-cvs] [400e6f37b80] master: Cycles: reduce subsurface stack memory usage.
Brecht Van Lommel
noreply at git.blender.org
Thu Sep 28 15:19:50 CEST 2017
Commit: 400e6f37b80dde3910b8d7a9d5e619b778a1c1ff
Author: Brecht Van Lommel
Date: Wed Aug 23 03:57:27 2017 +0200
Branches: master
https://developer.blender.org/rB400e6f37b80dde3910b8d7a9d5e619b778a1c1ff
Cycles: reduce subsurface stack memory usage.
This is done by storing only a subset of PathRadiance, and by storing
direct light immediately in the main PathRadiance. Saves about 10% of
CUDA stack memory, and simplifies subsurface indirect ray code.
===================================================================
M intern/cycles/kernel/kernel_accumulate.h
M intern/cycles/kernel/kernel_bake.h
M intern/cycles/kernel/kernel_path.h
M intern/cycles/kernel/kernel_path_branched.h
M intern/cycles/kernel/kernel_path_subsurface.h
M intern/cycles/kernel/kernel_path_surface.h
M intern/cycles/kernel/kernel_path_volume.h
M intern/cycles/kernel/kernel_types.h
M intern/cycles/kernel/split/kernel_branched.h
M intern/cycles/kernel/split/kernel_do_volume.h
M intern/cycles/kernel/split/kernel_indirect_subsurface.h
M intern/cycles/kernel/split/kernel_next_iteration_setup.h
===================================================================
diff --git a/intern/cycles/kernel/kernel_accumulate.h b/intern/cycles/kernel/kernel_accumulate.h
index 5e604586557..ae5f6e5e070 100644
--- a/intern/cycles/kernel/kernel_accumulate.h
+++ b/intern/cycles/kernel/kernel_accumulate.h
@@ -181,7 +181,6 @@ ccl_device_inline void path_radiance_init(PathRadiance *L, int use_light_pass)
if(use_light_pass) {
L->indirect = make_float3(0.0f, 0.0f, 0.0f);
- L->direct_throughput = make_float3(0.0f, 0.0f, 0.0f);
L->direct_emission = make_float3(0.0f, 0.0f, 0.0f);
L->color_diffuse = make_float3(0.0f, 0.0f, 0.0f);
@@ -202,18 +201,19 @@ ccl_device_inline void path_radiance_init(PathRadiance *L, int use_light_pass)
L->indirect_subsurface = make_float3(0.0f, 0.0f, 0.0f);
L->indirect_scatter = make_float3(0.0f, 0.0f, 0.0f);
- L->path_diffuse = make_float3(0.0f, 0.0f, 0.0f);
- L->path_glossy = make_float3(0.0f, 0.0f, 0.0f);
- L->path_transmission = make_float3(0.0f, 0.0f, 0.0f);
- L->path_subsurface = make_float3(0.0f, 0.0f, 0.0f);
- L->path_scatter = make_float3(0.0f, 0.0f, 0.0f);
-
L->transparent = 0.0f;
L->emission = make_float3(0.0f, 0.0f, 0.0f);
L->background = make_float3(0.0f, 0.0f, 0.0f);
L->ao = make_float3(0.0f, 0.0f, 0.0f);
L->shadow = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
L->mist = 0.0f;
+
+ L->state.diffuse = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.glossy = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.transmission = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.subsurface = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.scatter = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.direct = make_float3(0.0f, 0.0f, 0.0f);
}
else
#endif
@@ -245,26 +245,34 @@ ccl_device_inline void path_radiance_init(PathRadiance *L, int use_light_pass)
#endif
}
-ccl_device_inline void path_radiance_bsdf_bounce(PathRadiance *L, ccl_addr_space float3 *throughput,
- BsdfEval *bsdf_eval, float bsdf_pdf, int bounce, int bsdf_label)
+ccl_device_inline void path_radiance_bsdf_bounce(
+ KernelGlobals *kg,
+ PathRadianceState *L_state,
+ ccl_addr_space float3 *throughput,
+ BsdfEval *bsdf_eval,
+ float bsdf_pdf, int bounce, int bsdf_label)
{
float inverse_pdf = 1.0f/bsdf_pdf;
#ifdef __PASSES__
- if(L->use_light_pass) {
+ if(kernel_data.film.use_light_pass) {
if(bounce == 0 && !(bsdf_label & LABEL_TRANSPARENT)) {
/* first on directly visible surface */
float3 value = *throughput*inverse_pdf;
- L->path_diffuse = bsdf_eval->diffuse*value;
- L->path_glossy = bsdf_eval->glossy*value;
- L->path_transmission = bsdf_eval->transmission*value;
- L->path_subsurface = bsdf_eval->subsurface*value;
- L->path_scatter = bsdf_eval->scatter*value;
-
- *throughput = L->path_diffuse + L->path_glossy + L->path_transmission + L->path_subsurface + L->path_scatter;
+ L_state->diffuse = bsdf_eval->diffuse*value;
+ L_state->glossy = bsdf_eval->glossy*value;
+ L_state->transmission = bsdf_eval->transmission*value;
+ L_state->subsurface = bsdf_eval->subsurface*value;
+ L_state->scatter = bsdf_eval->scatter*value;
+
+ *throughput = L_state->diffuse +
+ L_state->glossy +
+ L_state->transmission +
+ L_state->subsurface +
+ L_state->scatter;
- L->direct_throughput = *throughput;
+ L_state->direct = *throughput;
}
else {
/* transparent bounce before first hit, or indirectly visible through BSDF */
@@ -493,19 +501,19 @@ ccl_device_inline void path_radiance_sum_indirect(PathRadiance *L)
* only a single throughput further along the path, here we recover just
* the indirect path that is not influenced by any particular BSDF type */
if(L->use_light_pass) {
- L->direct_emission = safe_divide_color(L->direct_emission, L->direct_throughput);
- L->direct_diffuse += L->path_diffuse*L->direct_emission;
- L->direct_glossy += L->path_glossy*L->direct_emission;
- L->direct_transmission += L->path_transmission*L->direct_emission;
- L->direct_subsurface += L->path_subsurface*L->direct_emission;
- L->direct_scatter += L->path_scatter*L->direct_emission;
+ L->direct_emission = safe_divide_color(L->direct_emission, L->state.direct);
+ L->direct_diffuse += L->state.diffuse*L->direct_emission;
+ L->direct_glossy += L->state.glossy*L->direct_emission;
+ L->direct_transmission += L->state.transmission*L->direct_emission;
+ L->direct_subsurface += L->state.subsurface*L->direct_emission;
+ L->direct_scatter += L->state.scatter*L->direct_emission;
- L->indirect = safe_divide_color(L->indirect, L->direct_throughput);
- L->indirect_diffuse += L->path_diffuse*L->indirect;
- L->indirect_glossy += L->path_glossy*L->indirect;
- L->indirect_transmission += L->path_transmission*L->indirect;
- L->indirect_subsurface += L->path_subsurface*L->indirect;
- L->indirect_scatter += L->path_scatter*L->indirect;
+ L->indirect = safe_divide_color(L->indirect, L->state.direct);
+ L->indirect_diffuse += L->state.diffuse*L->indirect;
+ L->indirect_glossy += L->state.glossy*L->indirect;
+ L->indirect_transmission += L->state.transmission*L->indirect;
+ L->indirect_subsurface += L->state.subsurface*L->indirect;
+ L->indirect_scatter += L->state.scatter*L->indirect;
}
#endif
}
@@ -514,11 +522,11 @@ ccl_device_inline void path_radiance_reset_indirect(PathRadiance *L)
{
#ifdef __PASSES__
if(L->use_light_pass) {
- L->path_diffuse = make_float3(0.0f, 0.0f, 0.0f);
- L->path_glossy = make_float3(0.0f, 0.0f, 0.0f);
- L->path_transmission = make_float3(0.0f, 0.0f, 0.0f);
- L->path_subsurface = make_float3(0.0f, 0.0f, 0.0f);
- L->path_scatter = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.diffuse = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.glossy = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.transmission = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.subsurface = make_float3(0.0f, 0.0f, 0.0f);
+ L->state.scatter = make_float3(0.0f, 0.0f, 0.0f);
L->direct_emission = make_float3(0.0f, 0.0f, 0.0f);
L->indirect = make_float3(0.0f, 0.0f, 0.0f);
@@ -531,11 +539,7 @@ ccl_device_inline void path_radiance_copy_indirect(PathRadiance *L,
{
#ifdef __PASSES__
if(L->use_light_pass) {
- L->path_diffuse = L_src->path_diffuse;
- L->path_glossy = L_src->path_glossy;
- L->path_transmission = L_src->path_transmission;
- L->path_subsurface = L_src->path_subsurface;
- L->path_scatter = L_src->path_scatter;
+ L->state = L_src->state;
L->direct_emission = L_src->direct_emission;
L->indirect = L_src->indirect;
diff --git a/intern/cycles/kernel/kernel_bake.h b/intern/cycles/kernel/kernel_bake.h
index f06005c5072..4d89839c46c 100644
--- a/intern/cycles/kernel/kernel_bake.h
+++ b/intern/cycles/kernel/kernel_bake.h
@@ -103,7 +103,6 @@ ccl_device_inline void compute_light_pass(KernelGlobals *kg,
throughput,
&state,
&L_sample);
- kernel_path_subsurface_accum_indirect(&ss_indirect, &L_sample);
}
is_sss_sample = true;
}
@@ -114,7 +113,7 @@ ccl_device_inline void compute_light_pass(KernelGlobals *kg,
if(!is_sss_sample && (pass_filter & (BAKE_FILTER_DIRECT | BAKE_FILTER_INDIRECT))) {
kernel_path_surface_connect_light(kg, sd, &emission_sd, throughput, &state, &L_sample);
- if(kernel_path_surface_bounce(kg, sd, &throughput, &state, &L_sample, &ray)) {
+ if(kernel_path_surface_bounce(kg, sd, &throughput, &state, &L_sample.state, &ray)) {
#ifdef __LAMP_MIS__
state.ray_t = 0.0f;
#endif
diff --git a/intern/cycles/kernel/kernel_path.h b/intern/cycles/kernel/kernel_path.h
index d43d6374c13..793fede0deb 100644
--- a/intern/cycles/kernel/kernel_path.h
+++ b/intern/cycles/kernel/kernel_path.h
@@ -222,7 +222,7 @@ ccl_device_forceinline VolumeIntegrateResult kernel_path_volume(
kernel_volume_decoupled_free(kg, &volume_segment);
if(result == VOLUME_PATH_SCATTERED) {
- if(kernel_path_volume_bounce(kg, sd, throughput, state, L, ray))
+ if(kernel_path_volume_bounce(kg, sd, throughput, state, &L->state, ray))
return VOLUME_PATH_SCATTERED;
else
return VOLUME_PATH_MISSED;
@@ -244,7 +244,7 @@ ccl_device_forceinline VolumeIntegrateResult kernel_path_volume(
kernel_path_volume_connect_light(kg, sd, emission_sd, *throughput, state, L);
/* indirect light bounce */
- if(kernel_path_volume_bounce(kg, sd, throughput, state, L, ray))
+ if(kernel_path_volume_bounce(kg, sd, throughput, state, &L->state, ray))
return VOLUME_PATH_SCATTERED;
else
return VOLUME_PATH_MISSED;
@@ -519,7 +519,7 @@ ccl_device void kernel_path_indirect(KernelGlobals *kg,
}
#endif /* defined(__EMISSION__) */
- if(!kernel_path_surface_bounce(kg, sd, &throughput, state, L, ray))
+ if(!kernel_path_surface_bounce(kg, sd, &throughput, state, &L->state, ray))
break;
}
}
@@ -648,13 +648,11 @@ ccl_device_forceinline void kernel_path_integrate(
kernel_path_surface_connect_light(kg, &sd, emission_sd, throughput, state, L);
/* compute direct lighting and next bounce */
- if(!kernel_path_surface_bounce(kg, &sd, &throughput, state, L, ray))
+ if(!kernel_path_surface_bounce(kg, &sd, &throughput, state, &L->state, ray))
break;
}
#ifdef __SUBSURFACE__
- kernel_path_subsurface_accum_indirect(&ss_indirect, L);
-
/* Trace indirect subsurface rays by restarting the loop. this uses less
* stack memory than invoking kernel_path_indirect.
*/
diff --git a/intern/cycles/kernel/kernel_path_branched.h b/intern/cycles/kernel/kernel_path_branched.h
index 010988d2a02..6e0ec22d581 100644
--- a/intern/cycles/kernel/kernel_path_branched.h
+++ b/intern/cycles/kernel/kernel_path_branched.h
@@ -128,7 +128,7 @@ ccl_device_noinline void kernel_branched_path_surface_indirect_light(KernelGloba
num_samples,
&tp,
&ps,
- L,
+ &L->state,
&bsdf_ray,
sum_sample_weight))
{
@@ -350,7 +350,7 @@ ccl_device void kernel_branched_path_integrate(KernelGlobals *kg,
&sd,
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list