[Bf-blender-cvs] [943e73b07e2] master: Cycles: decouple shadow paths from main path on GPU
Brecht Van Lommel
noreply at git.blender.org
Tue Oct 19 15:30:40 CEST 2021
Commit: 943e73b07e26d64c04ccb7d8f656e3818a57cca0
Author: Brecht Van Lommel
Date: Sun Oct 17 16:22:20 2021 +0200
Branches: master
https://developer.blender.org/rB943e73b07e26d64c04ccb7d8f656e3818a57cca0
Cycles: decouple shadow paths from main path on GPU
The motivation for this is twofold. It improves performance (5-10% on most
benchmark scenes), and will help to bring back transparency support for the
ambient occlusion pass.
* Duplicate some members from the main path state in the shadow path state.
* Add shadow paths incrementally to the array similar to what we do for
the shadow catchers.
* For the scheduling, allow running shade surface and shade volume kernels
as long as there is enough space in the shadow paths array. If not, execute
shadow kernels until it is empty.
* Add IntegratorShadowState and ConstIntegratorShadowState typedefs that
can be different between CPU and GPU. For GPU both main and shadow paths
juse have an integer for SoA access. Bt with CPU it's a different pointer
type so we get type safety checks in code shared between CPU and GPU.
* For CPU, add a separate IntegratorShadowStateCPU struct embedded in
IntegratorShadowState.
* Update various functions to take the shadow state, and make SVM take either
type of state using templates.
Differential Revision: https://developer.blender.org/D12889
===================================================================
M intern/cycles/integrator/path_trace_work_gpu.cpp
M intern/cycles/integrator/path_trace_work_gpu.h
M intern/cycles/kernel/CMakeLists.txt
M intern/cycles/kernel/device/cpu/kernel_arch_impl.h
M intern/cycles/kernel/device/gpu/kernel.h
M intern/cycles/kernel/integrator/integrator_intersect_closest.h
M intern/cycles/kernel/integrator/integrator_intersect_shadow.h
M intern/cycles/kernel/integrator/integrator_megakernel.h
M intern/cycles/kernel/integrator/integrator_shade_shadow.h
M intern/cycles/kernel/integrator/integrator_shade_surface.h
M intern/cycles/kernel/integrator/integrator_shade_volume.h
A intern/cycles/kernel/integrator/integrator_shadow_state_template.h
M intern/cycles/kernel/integrator/integrator_state.h
M intern/cycles/kernel/integrator/integrator_state_flow.h
M intern/cycles/kernel/integrator/integrator_state_template.h
M intern/cycles/kernel/integrator/integrator_state_util.h
M intern/cycles/kernel/integrator/integrator_volume_stack.h
M intern/cycles/kernel/kernel_accumulate.h
M intern/cycles/kernel/kernel_path_state.h
M intern/cycles/kernel/kernel_shader.h
M intern/cycles/kernel/kernel_shadow_catcher.h
M intern/cycles/kernel/kernel_types.h
M intern/cycles/kernel/osl/osl_services.cpp
M intern/cycles/kernel/osl/osl_shader.cpp
M intern/cycles/kernel/osl/osl_shader.h
M intern/cycles/kernel/svm/svm.h
M intern/cycles/kernel/svm/svm_ao.h
M intern/cycles/kernel/svm/svm_aov.h
M intern/cycles/kernel/svm/svm_bevel.h
M intern/cycles/kernel/svm/svm_light_path.h
===================================================================
diff --git a/intern/cycles/integrator/path_trace_work_gpu.cpp b/intern/cycles/integrator/path_trace_work_gpu.cpp
index bc380f269ad..18aa5dda70d 100644
--- a/intern/cycles/integrator/path_trace_work_gpu.cpp
+++ b/intern/cycles/integrator/path_trace_work_gpu.cpp
@@ -52,7 +52,11 @@ static size_t estimate_single_state_size()
* For until then use common value. Currently this size is only used for logging, but is weak to
* rely on this. */
#define KERNEL_STRUCT_VOLUME_STACK_SIZE 4
+
#include "kernel/integrator/integrator_state_template.h"
+
+#include "kernel/integrator/integrator_shadow_state_template.h"
+
#undef KERNEL_STRUCT_BEGIN
#undef KERNEL_STRUCT_MEMBER
#undef KERNEL_STRUCT_ARRAY_MEMBER
@@ -74,6 +78,8 @@ PathTraceWorkGPU::PathTraceWorkGPU(Device *device,
integrator_shader_sort_counter_(device, "integrator_shader_sort_counter", MEM_READ_WRITE),
integrator_shader_raytrace_sort_counter_(
device, "integrator_shader_raytrace_sort_counter", MEM_READ_WRITE),
+ integrator_next_shadow_path_index_(
+ device, "integrator_next_shadow_path_index", MEM_READ_WRITE),
integrator_next_shadow_catcher_path_index_(
device, "integrator_next_shadow_catcher_path_index", MEM_READ_WRITE),
queued_paths_(device, "queued_paths", MEM_READ_WRITE),
@@ -138,7 +144,11 @@ void PathTraceWorkGPU::alloc_integrator_soa()
} \
}
#define KERNEL_STRUCT_VOLUME_STACK_SIZE (integrator_state_soa_volume_stack_size_)
+
#include "kernel/integrator/integrator_state_template.h"
+
+#include "kernel/integrator/integrator_shadow_state_template.h"
+
#undef KERNEL_STRUCT_BEGIN
#undef KERNEL_STRUCT_MEMBER
#undef KERNEL_STRUCT_ARRAY_MEMBER
@@ -199,16 +209,22 @@ void PathTraceWorkGPU::alloc_integrator_sorting()
void PathTraceWorkGPU::alloc_integrator_path_split()
{
- if (integrator_next_shadow_catcher_path_index_.size() != 0) {
- return;
+ if (integrator_next_shadow_path_index_.size() == 0) {
+ integrator_next_shadow_path_index_.alloc(1);
+ integrator_next_shadow_path_index_.zero_to_device();
+
+ integrator_state_gpu_.next_shadow_path_index =
+ (int *)integrator_next_shadow_path_index_.device_pointer;
}
- integrator_next_shadow_catcher_path_index_.alloc(1);
- /* TODO(sergey): Use queue? */
- integrator_next_shadow_catcher_path_index_.zero_to_device();
+ if (integrator_next_shadow_catcher_path_index_.size() == 0) {
+ integrator_next_shadow_catcher_path_index_.alloc(1);
+ integrator_next_shadow_path_index_.data()[0] = 0;
+ integrator_next_shadow_catcher_path_index_.zero_to_device();
- integrator_state_gpu_.next_shadow_catcher_path_index =
- (int *)integrator_next_shadow_catcher_path_index_.device_pointer;
+ integrator_state_gpu_.next_shadow_catcher_path_index =
+ (int *)integrator_next_shadow_catcher_path_index_.device_pointer;
+ }
}
void PathTraceWorkGPU::alloc_work_memory()
@@ -341,27 +357,45 @@ bool PathTraceWorkGPU::enqueue_path_iteration()
return false;
}
- /* Finish shadows before potentially adding more shadow rays. We can only
- * store one shadow ray in the integrator state.
+ /* If the number of shadow kernels dropped to zero, set the next shadow path
+ * index to zero as well.
*
- * When there is a shadow catcher in the scene finish shadow rays before invoking intersect
- * closest kernel since so that the shadow paths are writing to the pre-split state. */
- if (kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE ||
- kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE ||
- kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_VOLUME ||
- (has_shadow_catcher() && kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_CLOSEST)) {
- if (queue_counter->num_queued[DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW]) {
- enqueue_path_iteration(DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW);
- return true;
+ * TODO: use shadow path compaction to lower it more often instead of letting
+ * it fill up entirely? */
+ const int num_queued_shadow =
+ queue_counter->num_queued[DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW] +
+ queue_counter->num_queued[DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW];
+ if (num_queued_shadow == 0) {
+ if (integrator_next_shadow_path_index_.data()[0] != 0) {
+ integrator_next_shadow_path_index_.data()[0] = 0;
+ queue_->copy_to_device(integrator_next_shadow_path_index_);
}
- else if (queue_counter->num_queued[DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW]) {
- enqueue_path_iteration(DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW);
- return true;
+ }
+
+ /* For kernels that add shadow paths, check if there is enough space available.
+ * If not, schedule shadow kernels first to clear out the shadow paths. */
+ if (kernel_creates_shadow_paths(kernel)) {
+ if (max_num_paths_ - integrator_next_shadow_path_index_.data()[0] <
+ queue_counter->num_queued[kernel]) {
+ if (queue_counter->num_queued[DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW]) {
+ enqueue_path_iteration(DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW);
+ return true;
+ }
+ else if (queue_counter->num_queued[DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW]) {
+ enqueue_path_iteration(DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW);
+ return true;
+ }
}
}
/* Schedule kernel with maximum number of queued items. */
enqueue_path_iteration(kernel);
+
+ /* Update next shadow path index for kernels that can add shadow paths. */
+ if (kernel_creates_shadow_paths(kernel)) {
+ queue_->copy_from_device(integrator_next_shadow_path_index_);
+ }
+
return true;
}
@@ -370,13 +404,12 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel)
void *d_path_index = (void *)NULL;
/* Create array of path indices for which this kernel is queued to be executed. */
- int work_size = max_active_path_index_;
+ int work_size = kernel_max_active_path_index(kernel);
IntegratorQueueCounter *queue_counter = integrator_queue_counter_.data();
int num_queued = queue_counter->num_queued[kernel];
- if (kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE ||
- kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE) {
+ if (kernel_uses_sorting(kernel)) {
/* Compute array of active paths, sorted by shader. */
work_size = num_queued;
d_path_index = (void *)queued_paths_.device_pointer;
@@ -387,8 +420,7 @@ void PathTraceWorkGPU::enqueue_path_iteration(DeviceKernel kernel)
work_size = num_queued;
d_path_index = (void *)queued_paths_.device_pointer;
- if (kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW ||
- kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW) {
+ if (kernel_is_shadow_path(kernel)) {
/* Compute array of active shadow paths for specific kernel. */
compute_queued_paths(DEVICE_KERNEL_INTEGRATOR_QUEUED_SHADOW_PATHS_ARRAY, kernel);
}
@@ -452,7 +484,7 @@ void PathTraceWorkGPU::compute_sorted_queued_paths(DeviceKernel kernel, DeviceKe
{
/* TODO: this could be smaller for terminated paths based on amount of work we want
* to schedule. */
- const int work_size = max_active_path_index_;
+ const int work_size = kernel_max_active_path_index(queued_kernel);
void *d_queued_paths = (void *)queued_paths_.device_pointer;
void *d_num_queued_paths = (void *)num_queued_paths_.device_pointer;
@@ -481,7 +513,7 @@ void PathTraceWorkGPU::compute_queued_paths(DeviceKernel kernel, DeviceKernel qu
int d_queued_kernel = queued_kernel;
/* Launch kernel to fill the active paths arrays. */
- const int work_size = max_active_path_index_;
+ const int work_size = kernel_max_active_path_index(queued_kernel);
void *d_queued_paths = (void *)queued_paths_.device_pointer;
void *d_num_queued_paths = (void *)num_queued_paths_.device_pointer;
void *args[] = {
@@ -981,4 +1013,29 @@ int PathTraceWorkGPU::shadow_catcher_count_possible_splits()
return num_queued_paths_.data()[0];
}
+bool PathTraceWorkGPU::kernel_uses_sorting(DeviceKernel kernel)
+{
+ return (kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE ||
+ kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE);
+}
+
+bool PathTraceWorkGPU::kernel_creates_shadow_paths(DeviceKernel kernel)
+{
+ return (kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE ||
+ kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE ||
+ kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_VOLUME);
+}
+
+bool PathTraceWorkGPU::kernel_is_shadow_path(DeviceKernel kernel)
+{
+ return (kernel == DEVICE_KERNEL_INTEGRATOR_INTERSECT_SHADOW ||
+ kernel == DEVICE_KERNEL_INTEGRATOR_SHADE_SHADOW);
+}
+
+int PathTraceWorkGPU::kernel_max_active_path_index(DeviceKernel kernel)
+{
+ return (kernel_is_shadow_path(kernel)) ? integrator_next_shadow_path_index_.data()[0] :
+ max_active_path_index_;
+}
+
CCL_NAMESPACE_END
diff --git a/intern/cycles/integrator/path_trace_work_gpu.h b/intern/cycles/integrator/path_trace_work_gpu.h
index e66851cc8d8..dd2c1c197ae 100644
--- a/intern/cycles/integrator/path_trace_work_gpu.h
+++ b/intern/cycles/integrator/path_trace_work_gpu.h
@@ -113,6 +113,12 @@ class PathTraceWorkGPU : public PathTraceWork {
/* Count how many currently scheduled paths can still split. */
int shadow_catcher_count_possible_splits();
+ /* Kernel properties. */
+ bool kernel_uses_sorting(DeviceKernel kernel);
+ bool kernel_creates_shadow_paths(DeviceKernel kernel);
+ bool kernel_is_shadow_path(DeviceKernel kernel);
+ int kernel_max_active_path_index(DeviceKernel kernel);
+
/* Integrator queue. */
unique_ptr<DeviceQueue> queue_;
@@ -131,6 +137,7 @@ class PathTraceWorkGPU : public PathTraceWork {
device_vector<int> integrator_shader_sort_counter_;
device_vector<int> integrator_shader_raytrace_sort_counter_;
/* Path split. */
+ device_vector<int> integrator_next_shadow_path_index_;
device_vector<int> integrator_next_shadow_catcher_path_index_;
/* Temporary buffer to get an array of queued path for a particular kernel. */
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index e0d48361650..7357c5804ed 100644
--- a/intern/cycles/k
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list