diff options
Diffstat (limited to 'intern/cycles/integrator')
| -rw-r--r-- | intern/cycles/integrator/path_trace.cpp | 2 | ||||
| -rw-r--r-- | intern/cycles/integrator/path_trace_work_gpu.cpp | 60 | ||||
| -rw-r--r-- | intern/cycles/integrator/path_trace_work_gpu.h | 3 |
3 files changed, 45 insertions, 20 deletions
diff --git a/intern/cycles/integrator/path_trace.cpp b/intern/cycles/integrator/path_trace.cpp index 6912bf928cd..ed278821b46 100644 --- a/intern/cycles/integrator/path_trace.cpp +++ b/intern/cycles/integrator/path_trace.cpp @@ -373,7 +373,7 @@ void PathTrace::path_trace(RenderWork &render_work) work_balance_infos_[i].time_spent += work_time; work_balance_infos_[i].occupancy = statistics.occupancy; - VLOG_WORK << "Rendered " << num_samples << " samples in " << work_time << " seconds (" + VLOG_INFO << "Rendered " << num_samples << " samples in " << work_time << " seconds (" << work_time / num_samples << " seconds per sample), occupancy: " << statistics.occupancy; }); diff --git a/intern/cycles/integrator/path_trace_work_gpu.cpp b/intern/cycles/integrator/path_trace_work_gpu.cpp index e262c252ce3..fa313f6460a 100644 --- a/intern/cycles/integrator/path_trace_work_gpu.cpp +++ b/intern/cycles/integrator/path_trace_work_gpu.cpp @@ -181,27 +181,45 @@ void PathTraceWorkGPU::alloc_integrator_queue() void PathTraceWorkGPU::alloc_integrator_sorting() { + /* Compute sort partitions, to balance between memory locality and coherence. + * Sort partitioning becomes less effective when more shaders are in the wavefront. In lieu of a + * more sophisticated heuristic we simply disable sort partitioning if the shader count is high. + */ + num_sort_partitions_ = 1; + if (device_scene_->data.max_shaders < 300) { + const int num_elements = queue_->num_sort_partition_elements(); + if (num_elements) { + num_sort_partitions_ = max(max_num_paths_ / num_elements, 1); + } + } + + integrator_state_gpu_.sort_partition_divisor = (int)divide_up(max_num_paths_, + num_sort_partitions_); + /* Allocate arrays for shader sorting. */ - const int max_shaders = device_scene_->data.max_shaders; - if (integrator_shader_sort_counter_.size() < max_shaders) { - integrator_shader_sort_counter_.alloc(max_shaders); + const int sort_buckets = device_scene_->data.max_shaders * num_sort_partitions_; + if (integrator_shader_sort_counter_.size() < sort_buckets) { + integrator_shader_sort_counter_.alloc(sort_buckets); integrator_shader_sort_counter_.zero_to_device(); + integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE] = + (int *)integrator_shader_sort_counter_.device_pointer; - integrator_shader_raytrace_sort_counter_.alloc(max_shaders); - integrator_shader_raytrace_sort_counter_.zero_to_device(); + if (device_scene_->data.kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) { + integrator_shader_raytrace_sort_counter_.alloc(sort_buckets); + integrator_shader_raytrace_sort_counter_.zero_to_device(); + integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE] = + (int *)integrator_shader_raytrace_sort_counter_.device_pointer; + } - integrator_shader_mnee_sort_counter_.alloc(max_shaders); - integrator_shader_mnee_sort_counter_.zero_to_device(); + if (device_scene_->data.kernel_features & KERNEL_FEATURE_MNEE) { + integrator_shader_mnee_sort_counter_.alloc(sort_buckets); + integrator_shader_mnee_sort_counter_.zero_to_device(); + integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_MNEE] = + (int *)integrator_shader_mnee_sort_counter_.device_pointer; + } - integrator_shader_sort_prefix_sum_.alloc(max_shaders); + integrator_shader_sort_prefix_sum_.alloc(sort_buckets); integrator_shader_sort_prefix_sum_.zero_to_device(); - - integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE] = - (int *)integrator_shader_sort_counter_.device_pointer; - integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE] = - (int *)integrator_shader_raytrace_sort_counter_.device_pointer; - integrator_state_gpu_.sort_key_counter[DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_MNEE] = - (int *)integrator_shader_mnee_sort_counter_.device_pointer; } } @@ -333,8 +351,12 @@ void PathTraceWorkGPU::enqueue_reset() queue_->enqueue(DEVICE_KERNEL_INTEGRATOR_RESET, max_num_paths_, args); queue_->zero_to_device(integrator_queue_counter_); queue_->zero_to_device(integrator_shader_sort_counter_); - queue_->zero_to_device(integrator_shader_raytrace_sort_counter_); - queue_->zero_to_device(integrator_shader_mnee_sort_counter_); + if (device_scene_->data.kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) { + queue_->zero_to_device(integrator_shader_raytrace_sort_counter_); + } + if (device_scene_->data.kernel_features & KERNEL_FEATURE_MNEE) { + queue_->zero_to_device(integrator_shader_mnee_sort_counter_); + } /* Tiles enqueue need to know number of active paths, which is based on this counter. Zero the * counter on the host side because `zero_to_device()` is not doing it. */ @@ -486,9 +508,9 @@ void PathTraceWorkGPU::compute_sorted_queued_paths(DeviceKernel kernel, /* Compute prefix sum of number of active paths with each shader. */ { const int work_size = 1; - int max_shaders = device_scene_->data.max_shaders; + int sort_buckets = device_scene_->data.max_shaders * num_sort_partitions_; - DeviceKernelArguments args(&d_counter, &d_prefix_sum, &max_shaders); + DeviceKernelArguments args(&d_counter, &d_prefix_sum, &sort_buckets); queue_->enqueue(DEVICE_KERNEL_PREFIX_SUM, work_size, args); } diff --git a/intern/cycles/integrator/path_trace_work_gpu.h b/intern/cycles/integrator/path_trace_work_gpu.h index 4c10a221a30..a805258d1b5 100644 --- a/intern/cycles/integrator/path_trace_work_gpu.h +++ b/intern/cycles/integrator/path_trace_work_gpu.h @@ -156,6 +156,9 @@ class PathTraceWorkGPU : public PathTraceWork { bool interop_use_checked_ = false; bool interop_use_ = false; + /* Number of partitions to sort state indices into prior to material sort. */ + int num_sort_partitions_; + /* Maximum number of concurrent integrator states. */ int max_num_paths_; |