kannonier8/blender
1
0
Fork 0
forked from bartvdbraak/blender
Code Pull Requests Activity

Fix Cycles NaN assert in random walk SSS due to very small throughput

Browse Source 
Now terminate if there are many bounces and the throughput gets so small
that we get precision issues.
This commit is contained in:
Brecht Van Lommel 2021-03-19 13:14:56 +01:00
parent d235f6a48e
commit a7f4270748
2 changed files with 21 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -640,6 +640,12 @@ ccl_device_noinline
/* If we hit the surface, we are done. */ /* If we hit the surface, we are done. */
break; break;
} }
else if (throughput.x < VOLUME_THROUGHPUT_EPSILON &&
throughput.y < VOLUME_THROUGHPUT_EPSILON &&
throughput.z < VOLUME_THROUGHPUT_EPSILON) {
/* Avoid unnecessary work and precision issue when throughput gets really small. */
break;
}
} }
kernel_assert(isfinite_safe(throughput.x) && isfinite_safe(throughput.y) && kernel_assert(isfinite_safe(throughput.x) && isfinite_safe(throughput.y) &&

24
intern/cycles/kernel/kernel_volume.h
View File

@ -16,6 +16,12 @@
CCL_NAMESPACE_BEGIN CCL_NAMESPACE_BEGIN
/* Ignore paths that have volume throughput below this value, to avoid unnecessary work
* and precision issues.
* todo: this value could be tweaked or turned into a probability to avoid unnecessary
* work in volumes and subsurface scattering. */
#define VOLUME_THROUGHPUT_EPSILON 1e-6f
/* Events for probalistic scattering */ /* Events for probalistic scattering */
typedef enum VolumeIntegrateResult { typedef enum VolumeIntegrateResult {
@ -226,7 +232,6 @@ ccl_device void kernel_volume_shadow_heterogeneous(KernelGlobals *kg,
const float object_step_size) const float object_step_size)
{ {
float3 tp = *throughput; float3 tp = *throughput;
const float tp_eps = 1e-6f; /* todo: this is likely not the right value */
/* Prepare for stepping. /* Prepare for stepping.
* For shadows we do not offset all segments, since the starting point is * For shadows we do not offset all segments, since the starting point is
@ -254,13 +259,15 @@ ccl_device void kernel_volume_shadow_heterogeneous(KernelGlobals *kg,
/* compute attenuation over segment */ /* compute attenuation over segment */
if (volume_shader_extinction_sample(kg, sd, state, new_P, &sigma_t)) { if (volume_shader_extinction_sample(kg, sd, state, new_P, &sigma_t)) {
/* Compute expf() only for every Nth step, to save some calculations /* Compute expf() only for every Nth step, to save some calculations
* because exp(a)*exp(b) = exp(a+b), also do a quick tp_eps check then. */ * because exp(a)*exp(b) = exp(a+b), also do a quick VOLUME_THROUGHPUT_EPSILON
* check then. */
sum += (-sigma_t * dt); sum += (-sigma_t * dt);
if ((i & 0x07) == 0) { /* ToDo: Other interval? */ if ((i & 0x07) == 0) { /* ToDo: Other interval? */
tp = *throughput * exp3(sum); tp = *throughput * exp3(sum);
/* stop if nearly all light is blocked */ /* stop if nearly all light is blocked */
if (tp.x < tp_eps && tp.y < tp_eps && tp.z < tp_eps) if (tp.x < VOLUME_THROUGHPUT_EPSILON && tp.y < VOLUME_THROUGHPUT_EPSILON &&
tp.z < VOLUME_THROUGHPUT_EPSILON)
break; break;
} }
} }
@ -572,7 +579,6 @@ kernel_volume_integrate_heterogeneous_distance(KernelGlobals *kg,
const float object_step_size) const float object_step_size)
{ {
float3 tp = *throughput; float3 tp = *throughput;
const float tp_eps = 1e-6f; /* todo: this is likely not the right value */
/* Prepare for stepping. /* Prepare for stepping.
* Using a different step offset for the first step avoids banding artifacts. */ * Using a different step offset for the first step avoids banding artifacts. */
@ -669,7 +675,8 @@ kernel_volume_integrate_heterogeneous_distance(KernelGlobals *kg,
tp = new_tp; tp = new_tp;
/* stop if nearly all light blocked */ /* stop if nearly all light blocked */
if (tp.x < tp_eps && tp.y < tp_eps && tp.z < tp_eps) { if (tp.x < VOLUME_THROUGHPUT_EPSILON && tp.y < VOLUME_THROUGHPUT_EPSILON &&
tp.z < VOLUME_THROUGHPUT_EPSILON) {
tp = zero_float3(); tp = zero_float3();
break; break;
} }
@ -770,8 +777,6 @@ ccl_device void kernel_volume_decoupled_record(KernelGlobals *kg,
VolumeSegment *segment, VolumeSegment *segment,
const float object_step_size) const float object_step_size)
{ {
const float tp_eps = 1e-6f; /* todo: this is likely not the right value */
/* prepare for volume stepping */ /* prepare for volume stepping */
int max_steps; int max_steps;
float step_size, step_shade_offset, steps_offset; float step_size, step_shade_offset, steps_offset;
@ -895,8 +900,9 @@ ccl_device void kernel_volume_decoupled_record(KernelGlobals *kg,
break; break;
/* stop if nearly all light blocked */ /* stop if nearly all light blocked */
if (accum_transmittance.x < tp_eps && accum_transmittance.y < tp_eps && if (accum_transmittance.x < VOLUME_THROUGHPUT_EPSILON &&
accum_transmittance.z < tp_eps) accum_transmittance.y < VOLUME_THROUGHPUT_EPSILON &&
accum_transmittance.z < VOLUME_THROUGHPUT_EPSILON)
break; break;
} }