pulp::gpu_audio::GpuNeuralRealtimeNode::Config Struct Reference

#include <gpu_neural_realtime_node.hpp>

Public Attributes
NeuralFamily	family = NeuralFamily::Unknown
	Must match the family the model was built for.
GpuNeuralModelHandle	model
std::uint32_t	channels = 1
	Must match the model's channel count.
std::uint32_t	lead_blocks = 2
	Fixed output delay in blocks; the CPU shadow is delayed by the same amount, so host delay compensation is family-independent.
std::uint32_t	capacity = 8
	GPU ring depth. Must exceed lead_blocks.
std::uint32_t	max_inflight = 0
	Maximum GPU blocks in flight; at most capacity - lead_blocks and at most the backend's own in-flight capacity.
std::uint32_t	prewarm_blocks = 0
	Silent blocks run through the GPU and the CPU shadow at prepare so both histories begin at the same stream position.
std::uint64_t	completion_service_wait_ns = 0
	Optional bounded completion-service budget for each non-RT worker pump.
MissPolicy	miss_policy = MissPolicy::CpuFallback
	Only CpuFallback is accepted: the shadow is built in.
ShadowPolicy	shadow_policy = ShadowPolicy::Auto
GpuNeuralModelHandle	shadow_model
	Optional cheaper model for the shadow.
std::uint32_t	shadow_queue_blocks = 16
	Worker queue depth in blocks, for both directions.
void *	shadow_audio_workgroup = nullptr
	Borrowed os_workgroup_t of the host's audio thread.
std::uint32_t	watchdog_blocks = 0
	GPU watchdog budget in callback blocks: a provider that completes nothing for this many blocks while work is in flight latches ProviderUnresponsive.

Member Data Documentation

capacity

std::uint32_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::capacity = 8

GPU ring depth. Must exceed lead_blocks.

channels

std::uint32_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::channels = 1

Must match the model's channel count.

completion_service_wait_ns

std::uint64_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::completion_service_wait_ns = 0

Optional bounded completion-service budget for each non-RT worker pump.

Zero preserves nonblocking service.

family

NeuralFamily pulp::gpu_audio::GpuNeuralRealtimeNode::Config::family = NeuralFamily::Unknown

Must match the family the model was built for.

lead_blocks

std::uint32_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::lead_blocks = 2

Fixed output delay in blocks; the CPU shadow is delayed by the same amount, so host delay compensation is family-independent.

max_inflight

std::uint32_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::max_inflight = 0

Maximum GPU blocks in flight; at most capacity - lead_blocks and at most the backend's own in-flight capacity.

Zero selects two when the backend admits two and the ring holds two blocks past the lead, else one: one block in flight cannot keep up when a block's round trip exceeds its period (64 frames on an M1 Max). The lead, and so the latency, is unchanged.

miss_policy

MissPolicy pulp::gpu_audio::GpuNeuralRealtimeNode::Config::miss_policy = MissPolicy::CpuFallback

Only CpuFallback is accepted: the shadow is built in.

model

GpuNeuralModelHandle pulp::gpu_audio::GpuNeuralRealtimeNode::Config::model

prewarm_blocks

std::uint32_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::prewarm_blocks = 0

Silent blocks run through the GPU and the CPU shadow at prepare so both histories begin at the same stream position.

shadow_audio_workgroup

void* pulp::gpu_audio::GpuNeuralRealtimeNode::Config::shadow_audio_workgroup = nullptr

Borrowed os_workgroup_t of the host's audio thread.

The Worker shadow thread joins it; without one it runs at USER_INTERACTIVE.

shadow_model

GpuNeuralModelHandle pulp::gpu_audio::GpuNeuralRealtimeNode::Config::shadow_model

Optional cheaper model for the shadow.

Never a default: status reports shadow_model_differs, and its output will not null against the GPU model. Binds like model.

shadow_policy

ShadowPolicy pulp::gpu_audio::GpuNeuralRealtimeNode::Config::shadow_policy = ShadowPolicy::Auto

shadow_queue_blocks

std::uint32_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::shadow_queue_blocks = 16

Worker queue depth in blocks, for both directions.

watchdog_blocks

std::uint32_t pulp::gpu_audio::GpuNeuralRealtimeNode::Config::watchdog_blocks = 0

GPU watchdog budget in callback blocks: a provider that completes nothing for this many blocks while work is in flight latches ProviderUnresponsive.

Zero selects 200 ms converted to blocks at prepare() (at 48 kHz: 150 blocks of 64 frames, 75 of 128, 38 of 256), or capacity - 1 if that is larger, and grows the ingress ring to hold the window and the lead. A non-zero value is used as given with the configured capacity. A provider that keeps completing, however late, is never unresponsive.

---

The documentation for this struct was generated from the following file:

• /home/runner/work/pulp/pulp/core/gpu_audio/include/pulp/gpu_audio/gpu_neural_realtime_node.hpp