Differential D12570 Diff 42091 intern/cycles/integrator/path_trace_work_gpu.cpp

Changeset View

Standalone View

intern/cycles/integrator/path_trace_work_gpu.cpp

Show First 20 Lines • Show All 174 Lines • ▼ Show 20 Lines
{		{
queue_->init_execution();		queue_->init_execution();

/* Copy to device side struct in constant memory. */		/* Copy to device side struct in constant memory. */
device_->const_copy_to(		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(int start_sample, int samples_num)		void PathTraceWorkGPU::render_samples(RenderStatistics &statistics,
		int start_sample,
		int samples_num)
{		{
/* Limit number of states for the tile and rely on a greedy scheduling of tiles. This allows to		/* Limit number of states for the tile and rely on a greedy scheduling of tiles. This allows to
* add more work (because tiles are smaller, so there is higher chance that more paths will		* add more work (because tiles are smaller, so there is higher chance that more paths will
* become busy after adding new tiles). This is especially important for the shadow catcher which		* become busy after adding new tiles). This is especially important for the shadow catcher which
* schedules work in halves of available number of paths. */		* schedules work in halves of available number of paths. */
work_tile_scheduler_.set_max_num_path_states(max_num_paths_ / 8);		work_tile_scheduler_.set_max_num_path_states(max_num_paths_ / 8);

work_tile_scheduler_.reset(effective_buffer_params_, start_sample, samples_num);		work_tile_scheduler_.reset(effective_buffer_params_, start_sample, samples_num);

enqueue_reset();		enqueue_reset();

		int num_iterations = 0;
		uint64_t num_busy_accum = 0;

/* TODO: set a hard limit in case of undetected kernel failures? */		/* TODO: set a hard limit in case of undetected kernel failures? */
while (true) {		while (true) {
/* Enqueue work from the scheduler, on start or when there are not enough		/* Enqueue work from the scheduler, on start or when there are not enough
* paths to keep the device occupied. */		* paths to keep the device occupied. */
bool finished;		bool finished;
if (enqueue_work_tiles(finished)) {		if (enqueue_work_tiles(finished)) {
/* Copy stats from the device. */		/* Copy stats from the device. */
queue_->copy_from_device(integrator_queue_counter_);		queue_->copy_from_device(integrator_queue_counter_);
Show All 20 Lines	if (enqueue_path_iteration()) {
if (!queue_->synchronize()) {		if (!queue_->synchronize()) {
break; /* Stop on error. */		break; /* Stop on error. */
}		}
}		}

if (is_cancel_requested()) {		if (is_cancel_requested()) {
break;		break;
}		}

		num_busy_accum += get_num_active_paths();
		++num_iterations;
}		}

		statistics.occupancy = static_cast<float>(num_busy_accum) / num_iterations / max_num_paths_;
}		}

DeviceKernel PathTraceWorkGPU::get_most_queued_kernel() const		DeviceKernel PathTraceWorkGPU::get_most_queued_kernel() const
{		{
const IntegratorQueueCounter *queue_counter = integrator_queue_counter_.data();		const IntegratorQueueCounter *queue_counter = integrator_queue_counter_.data();

int max_num_queued = 0;		int max_num_queued = 0;
DeviceKernel kernel = DEVICE_KERNEL_NUM;		DeviceKernel kernel = DEVICE_KERNEL_NUM;
▲ Show 20 Lines • Show All 684 Lines • Show Last 20 Lines