diff options
Diffstat (limited to 'intern/cycles/integrator/path_trace_work_gpu.cpp')
-rw-r--r-- | intern/cycles/integrator/path_trace_work_gpu.cpp | 68 |
1 files changed, 47 insertions, 21 deletions
diff --git a/intern/cycles/integrator/path_trace_work_gpu.cpp b/intern/cycles/integrator/path_trace_work_gpu.cpp index 0acaeace4b0..ee250a6916b 100644 --- a/intern/cycles/integrator/path_trace_work_gpu.cpp +++ b/intern/cycles/integrator/path_trace_work_gpu.cpp @@ -181,27 +181,49 @@ void PathTraceWorkGPU::alloc_integrator_queue() void PathTraceWorkGPU::alloc_integrator_sorting() { - /* 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); - integrator_shader_sort_counter_.zero_to_device(); + /* 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_shader_raytrace_sort_counter_.alloc(max_shaders); - integrator_shader_raytrace_sort_counter_.zero_to_device(); + integrator_state_gpu_.sort_partition_divisor = (int)divide_up(max_num_paths_, + num_sort_partitions_); - integrator_shader_mnee_sort_counter_.alloc(max_shaders); - integrator_shader_mnee_sort_counter_.zero_to_device(); + /* Allocate arrays for shader sorting. */ + 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_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; + if (device_scene_->data.kernel_features & KERNEL_FEATURE_NODE_RAYTRACE) { + if (integrator_shader_raytrace_sort_counter_.size() < sort_buckets) { + 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; + } + } + + if (device_scene_->data.kernel_features & KERNEL_FEATURE_MNEE) { + if (integrator_shader_mnee_sort_counter_.size() < sort_buckets) { + 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; + } } } @@ -239,7 +261,7 @@ void PathTraceWorkGPU::init_execution() /* Copy to device side struct in constant memory. */ device_->const_copy_to( - "__integrator_state", &integrator_state_gpu_, sizeof(integrator_state_gpu_)); + "integrator_state", &integrator_state_gpu_, sizeof(integrator_state_gpu_)); } void PathTraceWorkGPU::render_samples(RenderStatistics &statistics, @@ -333,8 +355,12 @@ 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 +512,9 @@ 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); } |