[Bf-blender-cvs] [e7ad962a40f] temp-cycles-opencl-staging: Cycles: Faster split branched path tracing by sharing samples with inactive threads
Mai Lavelle
noreply at git.blender.org
Thu Jun 8 11:40:07 CEST 2017
Commit: e7ad962a40f12280ab0b40026d839e2562bfbbc2
Author: Mai Lavelle
Date: Mon May 29 21:35:27 2017 -0400
Branches: temp-cycles-opencl-staging
https://developer.blender.org/rBe7ad962a40f12280ab0b40026d839e2562bfbbc2
Cycles: Faster split branched path tracing by sharing samples with inactive threads
Unlike regular path tracing, branched path tracing is usually used with lower
sample counts, at least for primary rays. This means that are less samples for
the GPU to work on in parallel and rendering is slower. As there is less work
overall there is also more inactive threads during rendering with BPT. This
patch makes use of those inactive rays to render branched samples in parallel
with other samples.
Each thread that is preparing for a branched sample will attempt to find an
inactive thread and if one is found the state for the sample is copied to that
thread. Potentially, if there are enough inactive threads, 100s of branched
samples could be generated from the same originating thread and ran in
parallel giving large speed ups.
Gives 70% faster render for pavillion midday scene. 20-60% faster on BMW
with car paint replaced with SSS/volumes.
===================================================================
M intern/cycles/device/device_split_kernel.cpp
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_next_iteration_setup.h
M intern/cycles/kernel/split/kernel_scene_intersect.h
M intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
M intern/cycles/kernel/split/kernel_split_common.h
M intern/cycles/kernel/split/kernel_split_data_types.h
M intern/cycles/kernel/split/kernel_subsurface_scatter.h
===================================================================
diff --git a/intern/cycles/device/device_split_kernel.cpp b/intern/cycles/device/device_split_kernel.cpp
index bb289a51912..d2b3a89fa98 100644
--- a/intern/cycles/device/device_split_kernel.cpp
+++ b/intern/cycles/device/device_split_kernel.cpp
@@ -259,6 +259,7 @@ bool DeviceSplitKernel::path_trace(DeviceTask *task,
ENQUEUE_SPLIT_KERNEL(direct_lighting, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(shadow_blocked_ao, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(shadow_blocked_dl, global_size, local_size);
+ ENQUEUE_SPLIT_KERNEL(enqueue_inactive, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(next_iteration_setup, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(indirect_subsurface, global_size, local_size);
ENQUEUE_SPLIT_KERNEL(queue_enqueue, global_size, local_size);
diff --git a/intern/cycles/kernel/kernel_types.h b/intern/cycles/kernel/kernel_types.h
index c9860e8d181..2c3859bf7c1 100644
--- a/intern/cycles/kernel/kernel_types.h
+++ b/intern/cycles/kernel/kernel_types.h
@@ -1431,6 +1431,9 @@ enum RayState {
RAY_BRANCHED_VOLUME_INDIRECT = (1 << 5),
RAY_BRANCHED_SUBSURFACE_INDIRECT = (1 << 6),
RAY_BRANCHED_INDIRECT = (RAY_BRANCHED_LIGHT_INDIRECT | RAY_BRANCHED_VOLUME_INDIRECT | RAY_BRANCHED_SUBSURFACE_INDIRECT),
+
+ /* Ray is evaluating an iteration of an indirect loop for another thread */
+ RAY_BRANCHED_INDIRECT_SHARED = (1 << 7),
};
#define ASSIGN_RAY_STATE(ray_state, ray_index, state) (ray_state[ray_index] = ((ray_state[ray_index] & RAY_FLAG_MASK) | state))
diff --git a/intern/cycles/kernel/split/kernel_branched.h b/intern/cycles/kernel/split/kernel_branched.h
index dc74a2ada53..e2762a85fc8 100644
--- a/intern/cycles/kernel/split/kernel_branched.h
+++ b/intern/cycles/kernel/split/kernel_branched.h
@@ -63,12 +63,49 @@ ccl_device_inline void kernel_split_branched_path_indirect_loop_end(KernelGlobal
REMOVE_RAY_FLAG(kernel_split_state.ray_state, ray_index, RAY_BRANCHED_INDIRECT);
}
+ccl_device_inline bool kernel_split_branched_indirect_start_shared(KernelGlobals *kg, int ray_index)
+{
+ ccl_global char *ray_state = kernel_split_state.ray_state;
+
+ int inactive_ray = dequeue_ray_index(QUEUE_INACTIVE_RAYS,
+ kernel_split_state.queue_data, kernel_split_params.queue_size, kernel_split_params.queue_index);
+
+ if(!IS_STATE(ray_state, inactive_ray, RAY_INACTIVE)) {
+ return false;
+ }
+
+#define SPLIT_DATA_ENTRY(type, name, num) \
+ kernel_split_state.name[inactive_ray] = kernel_split_state.name[ray_index];
+ SPLIT_DATA_ENTRIES_BRANCHED_SHARED
+#undef SPLIT_DATA_ENTRY
+
+ kernel_split_state.branched_state[inactive_ray].shared_sample_count = 0;
+ kernel_split_state.branched_state[inactive_ray].original_ray = ray_index;
+ kernel_split_state.branched_state[inactive_ray].waiting_on_shared_samples = false;
+
+ PathRadiance *L = &kernel_split_state.path_radiance[ray_index];
+ 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;
+ ADD_RAY_FLAG(ray_state, inactive_ray, RAY_BRANCHED_INDIRECT_SHARED);
+ ADD_RAY_FLAG(ray_state, inactive_ray, IS_FLAG(ray_state, ray_index, RAY_BRANCHED_INDIRECT));
+
+ atomic_fetch_and_inc_uint32((ccl_global uint*)&kernel_split_state.branched_state[ray_index].shared_sample_count);
+
+ return true;
+}
+
/* bounce off surface and integrate indirect light */
ccl_device_noinline bool kernel_split_branched_path_surface_indirect_light_iter(KernelGlobals *kg,
int ray_index,
float num_samples_adjust,
ShaderData *saved_sd,
- bool reset_path_state)
+ bool reset_path_state,
+ bool wait_for_shared)
{
SplitBranchedState *branched_state = &kernel_split_state.branched_state[ray_index];
@@ -155,12 +192,25 @@ ccl_device_noinline bool kernel_split_branched_path_surface_indirect_light_iter(
/* start the indirect path */
*tp *= num_samples_inv;
+ if(kernel_split_branched_indirect_start_shared(kg, ray_index)) {
+ continue;
+ }
+
return true;
}
branched_state->next_sample = 0;
}
+ branched_state->next_closure = sd->num_closure;
+
+ if(wait_for_shared) {
+ branched_state->waiting_on_shared_samples = (branched_state->shared_sample_count > 0);
+ if(branched_state->waiting_on_shared_samples) {
+ return true;
+ }
+ }
+
return false;
}
diff --git a/intern/cycles/kernel/split/kernel_do_volume.h b/intern/cycles/kernel/split/kernel_do_volume.h
index 694b777f429..9f8dd2392d9 100644
--- a/intern/cycles/kernel/split/kernel_do_volume.h
+++ b/intern/cycles/kernel/split/kernel_do_volume.h
@@ -75,11 +75,30 @@ ccl_device_noinline bool kernel_split_branched_path_volume_indirect_light_iter(K
branched_state->next_sample = j+1;
branched_state->num_samples = num_samples;
+ /* Attempting to share too many samples is slow for volumes as it causes us to
+ * loop here more and have many calls to kernel_volume_integrate which evaluates
+ * shaders. The many expensive shader evaluations cause the work load to become
+ * unbalanced and many threads to become idle in this kernel. Limiting the
+ * number of shared samples here helps quite a lot.
+ */
+ if(branched_state->shared_sample_count < 2) {
+ if(kernel_split_branched_indirect_start_shared(kg, ray_index)) {
+ continue;
+ }
+ }
+
return true;
}
# endif
}
+ branched_state->next_sample = num_samples;
+
+ branched_state->waiting_on_shared_samples = (branched_state->shared_sample_count > 0);
+ if(branched_state->waiting_on_shared_samples) {
+ return true;
+ }
+
kernel_split_branched_path_indirect_loop_end(kg, ray_index);
/* todo: avoid this calculation using decoupled ray marching */
diff --git a/intern/cycles/kernel/split/kernel_next_iteration_setup.h b/intern/cycles/kernel/split/kernel_next_iteration_setup.h
index 71017fed19e..7758e35fd32 100644
--- a/intern/cycles/kernel/split/kernel_next_iteration_setup.h
+++ b/intern/cycles/kernel/split/kernel_next_iteration_setup.h
@@ -147,6 +147,7 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
ray_index,
1.0f,
&kernel_split_state.branched_state[ray_index].sd,
+ true,
true))
{
ASSIGN_RAY_STATE(ray_state, ray_index, RAY_REGENERATED);
@@ -193,6 +194,7 @@ ccl_device void kernel_next_iteration_setup(KernelGlobals *kg,
ray_index,
1.0f,
&kernel_split_state.branched_state[ray_index].sd,
+ true,
true))
{
ASSIGN_RAY_STATE(ray_state, ray_index, RAY_REGENERATED);
diff --git a/intern/cycles/kernel/split/kernel_scene_intersect.h b/intern/cycles/kernel/split/kernel_scene_intersect.h
index 5dc94caec85..45984ca509b 100644
--- a/intern/cycles/kernel/split/kernel_scene_intersect.h
+++ b/intern/cycles/kernel/split/kernel_scene_intersect.h
@@ -43,11 +43,21 @@ ccl_device void kernel_scene_intersect(KernelGlobals *kg)
}
/* All regenerated rays become active here */
- if(IS_STATE(kernel_split_state.ray_state, ray_index, RAY_REGENERATED))
- ASSIGN_RAY_STATE(kernel_split_state.ray_state, ray_index, RAY_ACTIVE);
+ if(IS_STATE(kernel_split_state.ray_state, ray_index, RAY_REGENERATED)) {
+#ifdef __BRANCHED_PATH__
+ if(kernel_split_state.branched_state[ray_index].waiting_on_shared_samples) {
+ kernel_split_path_end(kg, ray_index);
+ }
+ else
+#endif /* __BRANCHED_PATH__ */
+ {
+ ASSIGN_RAY_STATE(kernel_split_state.ray_state, ray_index, RAY_ACTIVE);
+ }
+ }
- if(!IS_STATE(kernel_split_state.ray_state, ray_index, RAY_ACTIVE))
+ if(!IS_STATE(kernel_split_state.ray_state, ray_index, RAY_ACTIVE)) {
return;
+ }
#ifdef __KERNEL_DEBUG__
DebugData *debug_data = &kernel_split_state.debug_data[ray_index];
diff --git a/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h b/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
index 386fbbc4d09..78e61709b01 100644
--- a/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
+++ b/intern/cycles/kernel/split/kernel_shadow_blocked_dl.h
@@ -29,6 +29,14 @@ ccl_device void kernel_shadow_blocked_dl(KernelGlobals *kg)
kernel_split_state.queue_data, kernel_split_params.queue_size, 1);
}
+#ifdef __BRANCHED_PATH__
+ /* TODO(mai): move this somewhere else? */
+ if(thread_index == 0) {
+ /* Clear QUEUE_INACTIVE_RAYS before next kernel. */
+ kernel_split_params.queue_index[QUEUE_INACTIVE_RAYS] = 0;
+ }
+#endif /* __BRANCHED_PATH__ */
+
if(ray_index == QUEUE_EMPTY_SLOT)
return;
diff --git a/intern/cycles/kernel/split/kernel_split_common.h b/intern/cycles/kernel/split/kernel_split_common.h
index 57f070d51e0..0286569a8a5 100644
--- a/intern/cycles/kernel/split/kernel_split_common.h
+++ b/intern/cycles/kernel/split/kernel_split_common.h
@@ -56,7 +56,20 @@ ccl_device_inline void kernel_split_path_end(KernelGlobals *kg, int ray_index)
ccl_global char *ray_state = kernel_split_state.ray_state;
#ifdef __BRANCHED_PATH__
- if(IS_FLAG(ray_state, ray_index, RAY_BRANCHED_LIGHT_INDIRECT)) {
+ if(IS_FLAG(ray_state, ray_index, RAY_BRANCHED_INDIRECT_SHARED)) {
+ int orig_ray = kernel_split_state.branched_state[ray_index].original_ray;
+
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list