diff options
Diffstat (limited to 'intern/cycles/device/device_split_kernel.cpp')
| -rw-r--r-- | intern/cycles/device/device_split_kernel.cpp | 35 |
1 files changed, 30 insertions, 5 deletions
diff --git a/intern/cycles/device/device_split_kernel.cpp b/intern/cycles/device/device_split_kernel.cpp index 981ec74fe56..d2b3a89fa98 100644 --- a/intern/cycles/device/device_split_kernel.cpp +++ b/intern/cycles/device/device_split_kernel.cpp @@ -19,6 +19,7 @@ #include "kernel/kernel_types.h" #include "kernel/split/kernel_split_data_types.h" +#include "util/util_logging.h" #include "util/util_time.h" CCL_NAMESPACE_BEGIN @@ -38,12 +39,15 @@ DeviceSplitKernel::DeviceSplitKernel(Device *device) : device(device) kernel_do_volume = NULL; kernel_queue_enqueue = NULL; kernel_indirect_background = NULL; + kernel_shader_setup = NULL; + kernel_shader_sort = NULL; kernel_shader_eval = NULL; kernel_holdout_emission_blurring_pathtermination_ao = NULL; kernel_subsurface_scatter = NULL; kernel_direct_lighting = NULL; kernel_shadow_blocked_ao = NULL; kernel_shadow_blocked_dl = NULL; + kernel_enqueue_inactive = NULL; kernel_next_iteration_setup = NULL; kernel_indirect_subsurface = NULL; kernel_buffer_update = NULL; @@ -63,12 +67,15 @@ DeviceSplitKernel::~DeviceSplitKernel() delete kernel_do_volume; delete kernel_queue_enqueue; delete kernel_indirect_background; + delete kernel_shader_setup; + delete kernel_shader_sort; delete kernel_shader_eval; delete kernel_holdout_emission_blurring_pathtermination_ao; delete kernel_subsurface_scatter; delete kernel_direct_lighting; delete kernel_shadow_blocked_ao; delete kernel_shadow_blocked_dl; + delete kernel_enqueue_inactive; delete kernel_next_iteration_setup; delete kernel_indirect_subsurface; delete kernel_buffer_update; @@ -88,12 +95,15 @@ bool DeviceSplitKernel::load_kernels(const DeviceRequestedFeatures& requested_fe LOAD_KERNEL(do_volume); LOAD_KERNEL(queue_enqueue); LOAD_KERNEL(indirect_background); + LOAD_KERNEL(shader_setup); + LOAD_KERNEL(shader_sort); LOAD_KERNEL(shader_eval); LOAD_KERNEL(holdout_emission_blurring_pathtermination_ao); LOAD_KERNEL(subsurface_scatter); LOAD_KERNEL(direct_lighting); LOAD_KERNEL(shadow_blocked_ao); LOAD_KERNEL(shadow_blocked_dl); + LOAD_KERNEL(enqueue_inactive); LOAD_KERNEL(next_iteration_setup); LOAD_KERNEL(indirect_subsurface); LOAD_KERNEL(buffer_update); @@ -108,6 +118,9 @@ bool DeviceSplitKernel::load_kernels(const DeviceRequestedFeatures& requested_fe size_t DeviceSplitKernel::max_elements_for_max_buffer_size(device_memory& kg, device_memory& data, uint64_t max_buffer_size) { uint64_t size_per_element = state_buffer_size(kg, data, 1024) / 1024; + VLOG(1) << "Split state element size: " + << string_human_readable_number(size_per_element) << " bytes. (" + << string_human_readable_size(size_per_element) << ")."; return max_buffer_size / size_per_element; } @@ -156,13 +169,13 @@ bool DeviceSplitKernel::path_trace(DeviceTask *task, unsigned int max_work_groups = num_global_elements / work_pool_size + 1; /* Allocate work_pool_wgs memory. */ - work_pool_wgs.resize(max_work_groups * sizeof(unsigned int)); + work_pool_wgs.resize(max_work_groups); device->mem_alloc("work_pool_wgs", work_pool_wgs, MEM_READ_WRITE); - queue_index.resize(NUM_QUEUES * sizeof(int)); + queue_index.resize(NUM_QUEUES); device->mem_alloc("queue_index", queue_index, MEM_READ_WRITE); - use_queues_flag.resize(sizeof(char)); + use_queues_flag.resize(1); device->mem_alloc("use_queues_flag", use_queues_flag, MEM_READ_WRITE); ray_state.resize(num_global_elements); @@ -227,6 +240,7 @@ bool DeviceSplitKernel::path_trace(DeviceTask *task, ENQUEUE_SPLIT_KERNEL(path_init, global_size, local_size); bool activeRaysAvailable = true; + double cancel_time = DBL_MAX; while(activeRaysAvailable) { /* Do path-iteration in host [Enqueue Path-iteration kernels. */ @@ -236,18 +250,29 @@ bool DeviceSplitKernel::path_trace(DeviceTask *task, ENQUEUE_SPLIT_KERNEL(do_volume, global_size, local_size); ENQUEUE_SPLIT_KERNEL(queue_enqueue, global_size, local_size); ENQUEUE_SPLIT_KERNEL(indirect_background, global_size, local_size); + ENQUEUE_SPLIT_KERNEL(shader_setup, global_size, local_size); + ENQUEUE_SPLIT_KERNEL(shader_sort, global_size, local_size); ENQUEUE_SPLIT_KERNEL(shader_eval, global_size, local_size); ENQUEUE_SPLIT_KERNEL(holdout_emission_blurring_pathtermination_ao, global_size, local_size); ENQUEUE_SPLIT_KERNEL(subsurface_scatter, global_size, local_size); + ENQUEUE_SPLIT_KERNEL(queue_enqueue, global_size, local_size); 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); ENQUEUE_SPLIT_KERNEL(buffer_update, global_size, local_size); - if(task->get_cancel()) { + if(task->get_cancel() && cancel_time == DBL_MAX) { + /* Wait up to twice as many seconds for current samples to finish + * to avoid artifacts in render result from ending too soon. + */ + cancel_time = time_dt() + 2.0 * time_multiplier; + } + + if(time_dt() > cancel_time) { return true; } } @@ -271,7 +296,7 @@ bool DeviceSplitKernel::path_trace(DeviceTask *task, } } - if(task->get_cancel()) { + if(time_dt() > cancel_time) { return true; } } |