Allocation-free shadow diagnostics. More...
#include <gpu_neural_realtime_node.hpp>
Public Attributes | |
| ShadowPolicy | requested = ShadowPolicy::Auto |
| float | inline_cost_per_channel_pct = -1.0f |
| One block of the shadow model for one channel, as a percentage of the block period, timed on the host thread at prepare(). | |
| float | inline_cost_total_pct = -1.0f |
| The per-channel cost times the channel count: what Inline would spend on the audio thread. | |
| std::uint64_t | shadow_input_dropped = 0 |
| Worker inputs dropped because the shadow queue was full. | |
| std::uint64_t | shadow_resyncs = 0 |
| Times the Worker restarted the model's state after a dropped input. | |
| bool | workgroup_joined = false |
Allocation-free shadow diagnostics.
| float pulp::gpu_audio::GpuNeuralRealtimeNode::ShadowStatus::inline_cost_per_channel_pct = -1.0f |
One block of the shadow model for one channel, as a percentage of the block period, timed on the host thread at prepare().
Negative when no shadow was prepared.
| float pulp::gpu_audio::GpuNeuralRealtimeNode::ShadowStatus::inline_cost_total_pct = -1.0f |
The per-channel cost times the channel count: what Inline would spend on the audio thread.
This is what Auto compares.
| ShadowPolicy pulp::gpu_audio::GpuNeuralRealtimeNode::ShadowStatus::requested = ShadowPolicy::Auto |
| std::uint64_t pulp::gpu_audio::GpuNeuralRealtimeNode::ShadowStatus::shadow_input_dropped = 0 |
Worker inputs dropped because the shadow queue was full.
| std::uint64_t pulp::gpu_audio::GpuNeuralRealtimeNode::ShadowStatus::shadow_resyncs = 0 |
Times the Worker restarted the model's state after a dropped input.
| bool pulp::gpu_audio::GpuNeuralRealtimeNode::ShadowStatus::workgroup_joined = false |