blender
/
blender
129
398
Fork 572
Code Issues 6.2k Pull Requests 750 Projects 19 Wiki Activity

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.
This commit is contained in:
Brecht Van Lommel 2017-08-23 03:57:27 +02:00
parent 9ddee885ae
commit 400e6f37b8
12 changed files with 91 additions and 115 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

82
intern/cycles/kernel/kernel_accumulate.h
View File

@ -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;
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->path_diffuse + L->path_glossy + L->path_transmission + L->path_subsurface + L->path_scatter;
*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;

3
intern/cycles/kernel/kernel_bake.h
View File

@ -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

10
intern/cycles/kernel/kernel_path.h
View File

@ -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.
*/

6
intern/cycles/kernel/kernel_path_branched.h
View File

@ -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,
&tp,
&ps,
L,
&L->state,
&pray))
{
kernel_path_indirect(kg,
@ -405,7 +405,7 @@ ccl_device void kernel_branched_path_integrate(KernelGlobals *kg,
&sd,
&tp,
&ps,
L,
&L->state,
&pray))
{
kernel_path_indirect(kg,

50
intern/cycles/kernel/kernel_path_subsurface.h
View File

@ -43,7 +43,7 @@ bool kernel_path_subsurface_scatter(
* the second one should be converted to a diffuse BSDF to
* avoid this.
*/
kernel_assert(!ss_indirect->tracing);
kernel_assert(!(state->flag & PATH_RAY_DIFFUSE_ANCESTOR));
uint lcg_state = lcg_state_init_addrspace(state, 0x68bc21eb);
@ -56,7 +56,7 @@ bool kernel_path_subsurface_scatter(
bssrdf_u, bssrdf_v,
false);
# ifdef __VOLUME__
ss_indirect->need_update_volume_stack =
bool need_update_volume_stack =
kernel_data.integrator.use_volumes &&
sd->object_flag & SD_OBJECT_INTERSECTS_VOLUME;
# endif /* __VOLUME__ */
@ -75,28 +75,25 @@ bool kernel_path_subsurface_scatter(
sc,
false);
kernel_path_surface_connect_light(kg, sd, emission_sd, *throughput, state, L);
ccl_addr_space PathState *hit_state = &ss_indirect->state[ss_indirect->num_rays];
ccl_addr_space Ray *hit_ray = &ss_indirect->rays[ss_indirect->num_rays];
ccl_addr_space float3 *hit_tp = &ss_indirect->throughputs[ss_indirect->num_rays];
PathRadiance *hit_L = &ss_indirect->L[ss_indirect->num_rays];
PathRadianceState *hit_L_state = &ss_indirect->L_state[ss_indirect->num_rays];
*hit_state = *state;
*hit_ray = *ray;
*hit_tp = *throughput;
*hit_L_state = L->state;
hit_state->rng_offset += PRNG_BOUNCE_NUM;
path_radiance_init(hit_L, kernel_data.film.use_light_pass);
hit_L->direct_throughput = L->direct_throughput;
path_radiance_copy_indirect(hit_L, L);
kernel_path_surface_connect_light(kg, sd, emission_sd, *hit_tp, state, hit_L);
if(kernel_path_surface_bounce(kg,
sd,
hit_tp,
hit_state,
hit_L,
hit_L_state,
hit_ray))
{
# ifdef __LAMP_MIS__
@ -104,7 +101,7 @@ bool kernel_path_subsurface_scatter(
# endif /* __LAMP_MIS__ */
# ifdef __VOLUME__
if(ss_indirect->need_update_volume_stack) {
if(need_update_volume_stack) {
Ray volume_ray = *ray;
/* Setup ray from previous surface point to the new one. */
volume_ray.D = normalize_len(hit_ray->P - volume_ray.P,
@ -117,12 +114,8 @@ bool kernel_path_subsurface_scatter(
hit_state->volume_stack);
}
# endif /* __VOLUME__ */
path_radiance_reset_indirect(L);
ss_indirect->num_rays++;
}
else {
path_radiance_accum_sample(L, hit_L);
}
}
return true;
}
@ -132,23 +125,9 @@ bool kernel_path_subsurface_scatter(
ccl_device_inline void kernel_path_subsurface_init_indirect(
ccl_addr_space SubsurfaceIndirectRays *ss_indirect)
{
ss_indirect->tracing = false;
ss_indirect->num_rays = 0;
}
ccl_device void kernel_path_subsurface_accum_indirect(
ccl_addr_space SubsurfaceIndirectRays *ss_indirect,
PathRadiance *L)
{
if(ss_indirect->tracing) {
path_radiance_sum_indirect(L);
path_radiance_accum_sample(&ss_indirect->direct_L, L);
if(ss_indirect->num_rays == 0) {
*L = ss_indirect->direct_L;
}
}
}
ccl_device void kernel_path_subsurface_setup_indirect(
KernelGlobals *kg,
ccl_addr_space SubsurfaceIndirectRays *ss_indirect,
@ -157,20 +136,15 @@ ccl_device void kernel_path_subsurface_setup_indirect(
PathRadiance *L,
ccl_addr_space float3 *throughput)
{
if(!ss_indirect->tracing) {
ss_indirect->direct_L = *L;
}
ss_indirect->tracing = true;
/* Setup state, ray and throughput for indirect SSS rays. */
ss_indirect->num_rays--;
ccl_addr_space Ray *indirect_ray = &ss_indirect->rays[ss_indirect->num_rays];
PathRadiance *indirect_L = &ss_indirect->L[ss_indirect->num_rays];
path_radiance_sum_indirect(L);
path_radiance_reset_indirect(L);
*state = ss_indirect->state[ss_indirect->num_rays];
*ray = *indirect_ray;
*L = *indirect_L;
*ray = ss_indirect->rays[ss_indirect->num_rays];
L->state = ss_indirect->L_state[ss_indirect->num_rays];
*throughput = ss_indirect->throughputs[ss_indirect->num_rays];
state->rng_offset += ss_indirect->num_rays * PRNG_BOUNCE_NUM;

8
intern/cycles/kernel/kernel_path_surface.h
View File

@ -150,7 +150,7 @@ ccl_device bool kernel_branched_path_surface_bounce(
int num_samples,
ccl_addr_space float3 *throughput,
ccl_addr_space PathState *state,
PathRadiance *L,
PathRadianceState *L_state,
ccl_addr_space Ray *ray,
float sum_sample_weight)
{
@ -170,7 +170,7 @@ ccl_device bool kernel_branched_path_surface_bounce(
return false;
/* modify throughput */
path_radiance_bsdf_bounce(L, throughput, &bsdf_eval, bsdf_pdf, state->bounce, label);
path_radiance_bsdf_bounce(kg, L_state, throughput, &bsdf_eval, bsdf_pdf, state->bounce, label);
#ifdef __DENOISING_FEATURES__
state->denoising_feature_weight *= sc->sample_weight / (sum_sample_weight * num_samples);
@ -271,7 +271,7 @@ ccl_device bool kernel_path_surface_bounce(KernelGlobals *kg,
ShaderData *sd,
ccl_addr_space float3 *throughput,
ccl_addr_space PathState *state,
PathRadiance *L,
PathRadianceState *L_state,
ccl_addr_space Ray *ray)
{
/* no BSDF? we can stop here */
@ -292,7 +292,7 @@ ccl_device bool kernel_path_surface_bounce(KernelGlobals *kg,
return false;
/* modify throughput */
path_radiance_bsdf_bounce(L, throughput, &bsdf_eval, bsdf_pdf, state->bounce, label);
path_radiance_bsdf_bounce(kg, L_state, throughput, &bsdf_eval, bsdf_pdf, state->bounce, label);
/* set labels */
if(!(label & LABEL_TRANSPARENT)) {

4
intern/cycles/kernel/kernel_path_volume.h
View File

@ -68,7 +68,7 @@ bool kernel_path_volume_bounce(
ShaderData *sd,
ccl_addr_space float3 *throughput,
ccl_addr_space PathState *state,
PathRadiance *L,
PathRadianceState *L_state,
ccl_addr_space Ray *ray)
{
/* sample phase function */
@ -87,7 +87,7 @@ bool kernel_path_volume_bounce(
return false;
/* modify throughput */
path_radiance_bsdf_bounce(L, throughput, &phase_eval, phase_pdf, state->bounce, label);
path_radiance_bsdf_bounce(kg, L_state, throughput, &phase_eval, phase_pdf, state->bounce, label);
/* set labels */
state->ray_pdf = phase_pdf;

33
intern/cycles/kernel/kernel_types.h
View File

@ -466,6 +466,18 @@ typedef struct DebugData {
} DebugData;
#endif
typedef ccl_addr_space struct PathRadianceState {
#ifdef __PASSES__
float3 diffuse;
float3 glossy;
float3 transmission;
float3 subsurface;
float3 scatter;
float3 direct;
#endif
} PathRadianceState;
typedef ccl_addr_space struct PathRadiance {
#ifdef __PASSES__
int use_light_pass;
@ -478,7 +490,6 @@ typedef ccl_addr_space struct PathRadiance {
float3 ao;
float3 indirect;
float3 direct_throughput;
float3 direct_emission;
float3 color_diffuse;
@ -499,16 +510,12 @@ typedef ccl_addr_space struct PathRadiance {
float3 indirect_subsurface;
float3 indirect_scatter;
float3 path_diffuse;
float3 path_glossy;
float3 path_transmission;
float3 path_subsurface;
float3 path_scatter;
float4 shadow;
float mist;
#endif
PathRadianceState state;
#ifdef __SHADOW_TRICKS__
/* Total light reachable across the path, ignoring shadow blocked queries. */
float3 path_total;
@ -1032,8 +1039,7 @@ typedef struct PathState {
/* Subsurface */
/* Struct to gather multiple SSS hits. */
typedef struct SubsurfaceIntersection
{
typedef struct SubsurfaceIntersection {
Ray ray;
float3 weight[BSSRDF_MAX_HITS];
@ -1043,17 +1049,14 @@ typedef struct SubsurfaceIntersection
} SubsurfaceIntersection;
/* Struct to gather SSS indirect rays and delay tracing them. */
typedef struct SubsurfaceIndirectRays
{
bool need_update_volume_stack;
bool tracing;
typedef struct SubsurfaceIndirectRays {
PathState state[BSSRDF_MAX_HITS];
struct PathRadiance direct_L;
int num_rays;
struct Ray rays[BSSRDF_MAX_HITS];
float3 throughputs[BSSRDF_MAX_HITS];
struct PathRadiance L[BSSRDF_MAX_HITS];
struct PathRadianceState L_state[BSSRDF_MAX_HITS];
} SubsurfaceIndirectRays;
/* Constant Kernel Data

3
intern/cycles/kernel/split/kernel_branched.h
View File

@ -87,7 +87,6 @@ ccl_device_inline bool kernel_split_branched_indirect_start_shared(KernelGlobals
PathRadiance *inactive_L = &kernel_split_state.path_radiance[inactive_ray];
path_radiance_init(inactive_L, kernel_data.film.use_light_pass);
inactive_L->direct_throughput = L->direct_throughput;
path_radiance_copy_indirect(inactive_L, L);
ray_state[inactive_ray] = RAY_REGENERATED;
@ -176,7 +175,7 @@ ccl_device_noinline bool kernel_split_branched_path_surface_indirect_light_iter(
num_samples,
tp,
ps,
L,
&L->state,
bsdf_ray,
sum_sample_weight))
{

4
intern/cycles/kernel/split/kernel_do_volume.h
View File

@ -65,7 +65,7 @@ ccl_device_noinline bool kernel_split_branched_path_volume_indirect_light_iter(K
kernel_path_volume_connect_light(kg, sd, emission_sd, *tp, &branched_state->path_state, L);
/* indirect light bounce */
if(!kernel_path_volume_bounce(kg, sd, tp, ps, L, pray)) {
if(!kernel_path_volume_bounce(kg, sd, tp, ps, &L->state, pray)) {
continue;
}
@ -170,7 +170,7 @@ ccl_device void kernel_do_volume(KernelGlobals *kg)
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)) {
ASSIGN_RAY_STATE(ray_state, ray_index, RAY_REGENERATED);
}
else {

1
intern/cycles/kernel/split/kernel_indirect_subsurface.h
View File

@ -54,7 +54,6 @@ ccl_device void kernel_indirect_subsurface(KernelGlobals *kg)
#endif
if(IS_STATE(ray_state, ray_index, RAY_UPDATE_BUFFER)) {
ccl_addr_space SubsurfaceIndirectRays *ss_indirect = &kernel_split_state.ss_rays[ray_index];
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.

2
intern/cycles/kernel/split/kernel_next_iteration_setup.h
View File

@ -134,7 +134,7 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
if(!kernel_data.integrator.branched || IS_FLAG(ray_state, ray_index, RAY_BRANCHED_INDIRECT)) {
#endif
/* 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)) {
kernel_split_path_end(kg, ray_index);
}
#ifdef __BRANCHED_PATH__