Detailed Description
Definition at line 286 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
Data Fields | |
| guint | batch_size |
| max concurrent streams. More... | |
| guint | per_model_max_streams |
| gboolean | unified_batch |
| default cap for a model if unset More... | |
| gint | process_mode |
| one display batch (TRUE) vs per-role More... | |
| gboolean | gie_ids_from_config |
| 1 = pgie, 2 = sgie; forwarded to every inference element this stage builds More... | |
| gboolean | enable |
| gboolean | shard_compact |
| [multimodel] enable: 0 => bypass the inference bin (decode->display passthrough) More... | |
| gboolean | compact_idle_models_across_pools |
| auto-compact under-full shards of a model back into fewer shards when streams drain (default TRUE). More... | |
| gboolean | model_sharding |
| dedup-on-idle policy (default TRUE). More... | |
| guint | mux_width |
| auto-shard streams when per-model cap exceeded. More... | |
| guint | mux_height |
| mux surface resolution, all muxers (default 1920) More... | |
| gboolean | mux_live_source |
| mux surface resolution, all muxers (default 1080) More... | |
| guint | model_mux_push_timeout |
| guint | combined_mux_push_timeout |
| per-model mux flush us (default 10000 = 10ms) More... | |
| gboolean | attach_perf_metric |
| combined mux flush us (default 33333 ~= 30fps) More... | |
| guint | perf_metric_interval_sec |
| enable fps/latency measurement (default FALSE) More... | |
| gboolean | copy_primary_buffer |
| measurement window seconds (default 5, clamp >=1) More... | |
| gboolean | copy_shadow_buffer |
| copy on the PRIMARY lane (nvvideoconvert dp=1; def 0) More... | |
| guint | copy_buffer_pool_size |
| copy on the SHADOW lane (nvvideoconvert dp=1; def 1) More... | |
| gboolean | lane_leaky |
| copy-lane nvvideoconvert output-buffers (0 = element default). More... | |
| guint | gpu |
| [multimodel] gpu-id: the PIPELINE device – where the decoders, the shared input demux and the combined display mux live (default 0). More... | |
| gint | placement_policy |
| [multimodel] placement-policy (pack|spread): how unconstrained streams are placed across a multi-GPU instance group. More... | |
| gboolean | auto_gpu_scale |
| [multimodel] auto-gpu-scale (default 0): ELASTIC multi-GPU scale-out without any per-model gpu configuration. More... | |
| ModelCatalogEntry | models [MM_MAX_MODELS] |
| model registry: built from [MultiModelInference] default primary (+ optional shadow). More... | |
| guint | num_models |
| ModelVersionArtifact | versions [MM_MAX_MODELS *4] |
| PER-VERSION artifacts: optional engine/config overrides per (name, version), parsed from [model-<name>-<version>] blocks. More... | |
| guint | num_versions |
| GMutex | versions_lock |
| guards versions[] / num_versions – the ONLY ModelMuxConfig state mutable at RUNTIME: modelmux_config_register_version (model/load|update trigger) and modelmux_control_registry_unregister_version (OTA rollback) mutate on the main loop under api_control_lock, while GET model/status resolves placements on the CIVETWEB thread under ib->lock only. More... | |
| gchar * | default_primary |
| gchar * | default_primary_version |
| default primary model NAME (catalog ref) More... | |
| gchar * | default_shadow |
| default primary model VERSION (e.g. More... | |
| gchar * | default_shadow_version |
| default shadow model NAME (NULL=>none) More... | |
| gint | default_primary_gpu |
| default shadow model VERSION (NULL=>none) More... | |
| gint | default_shadow_gpu |
| requested instance placement of the runtime default (model/routing default{}.gpu); -1 = any instance. More... | |
| StreamModelBinding | bindings [MM_MAX_STREAMS_DEFAULT] |
| config-time per-sensor model bindings ([stream-model-<sensor_id>] groups). More... | |
| guint | num_bindings |
Field Documentation
◆ attach_perf_metric
| gboolean ModelMuxConfig::attach_perf_metric |
combined mux flush us (default 33333 ~= 30fps)
[multimodel] perf metrics: when attach_perf_metric is on, the plugin measures per-model (aggregate) and per-(stream,role) fps + inference latency and reports them additively in model/status. OFF (default) => zero overhead, model/status byte-identical to before.
Definition at line 341 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ auto_gpu_scale
| gboolean ModelMuxConfig::auto_gpu_scale |
[multimodel] auto-gpu-scale (default 0): ELASTIC multi-GPU scale-out without any per-model gpu configuration.
When every shard of a model is at capacity and a NEW shard must grow, pick the healthiest OTHER device instead of the model's own – but ONLY when that device is IDENTICAL to the model's (same compute capability + device name), so the existing TensorRT engine is legally reused (deserialize in seconds; a mismatched device would silently trigger a minutes-long rebuild). The overflow shard is UNPINNED: compaction migrates its streams back and reclaims it when load drops (scale-in). Heterogeneous boxes / single-GPU: behaves exactly as auto-gpu-scale=0.
Definition at line 388 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ batch_size
| guint ModelMuxConfig::batch_size |
max concurrent streams.
Definition at line 289 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ bindings
| StreamModelBinding ModelMuxConfig::bindings |
config-time per-sensor model bindings ([stream-model-<sensor_id>] groups).
Empty => every stream uses the defaults (fully backward-compatible).
Definition at line 428 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ combined_mux_push_timeout
| guint ModelMuxConfig::combined_mux_push_timeout |
per-model mux flush us (default 10000 = 10ms)
Definition at line 336 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ compact_idle_models_across_pools
| gboolean ModelMuxConfig::compact_idle_models_across_pools |
auto-compact under-full shards of a model back into fewer shards when streams drain (default TRUE).
Make-before-break, zero-drop. Set 0 to keep grown shards until they empty naturally.
Definition at line 309 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ copy_buffer_pool_size
| guint ModelMuxConfig::copy_buffer_pool_size |
copy on the SHADOW lane (nvvideoconvert dp=1; def 1)
Definition at line 352 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ copy_primary_buffer
| gboolean ModelMuxConfig::copy_primary_buffer |
measurement window seconds (default 5, clamp >=1)
[multimodel] buffer-copy-mode: which per-(stream,role) lane(s) DEEP-COPY the decoded frame via an nvvideoconvert (disable-passthrough=1) spliced BEFORE that lane's queue. A copy lane releases the upstream decoder buffer at fan-out, so a SLOW lane can no longer backpressure the decoder when per-model fps differ; the other (passthrough) role stays zero-copy (no element added). Config string none|primary|shadow|both -> these two derived flags. Default: shadow. "none" => byte-identical to before (no converter on any lane).
Definition at line 350 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ copy_shadow_buffer
| gboolean ModelMuxConfig::copy_shadow_buffer |
copy on the PRIMARY lane (nvvideoconvert dp=1; def 0)
Definition at line 351 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ default_primary
| gchar * ModelMuxConfig::default_primary |
Definition at line 416 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ default_primary_gpu
| gint ModelMuxConfig::default_primary_gpu |
default shadow model VERSION (NULL=>none)
Definition at line 420 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ default_primary_version
| gchar * ModelMuxConfig::default_primary_version |
default primary model NAME (catalog ref)
Definition at line 417 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ default_shadow
| gchar * ModelMuxConfig::default_shadow |
default primary model VERSION (e.g.
1)
Definition at line 418 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ default_shadow_gpu
| gint ModelMuxConfig::default_shadow_gpu |
requested instance placement of the runtime default (model/routing default{}.gpu); -1 = any instance.
Applied to every stream that RESOLVES to the default (add/bind/sweep).
Definition at line 424 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ default_shadow_version
| gchar * ModelMuxConfig::default_shadow_version |
default shadow model NAME (NULL=>none)
Definition at line 419 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ enable
| gboolean ModelMuxConfig::enable |
Definition at line 303 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ gie_ids_from_config
| gboolean ModelMuxConfig::gie_ids_from_config |
1 = pgie, 2 = sgie; forwarded to every inference element this stage builds
TRUE when every catalog model declares a DISTINCT gie-unique-id, so this stage adopts the configured ids (making operate-on-gie-id knowable from config, as with a plain nvinfer). FALSE when any two overlap – the stage then falls back to the historical monotonic allocator for ALL models, so ids stay unambiguous and an existing config keeps working unchanged. All-or-nothing on purpose: a mix of configured and allocated ids would be the least predictable outcome.
Definition at line 302 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ gpu
| guint ModelMuxConfig::gpu |
[multimodel] gpu-id: the PIPELINE device – where the decoders, the shared input demux and the combined display mux live (default 0).
A model bin ON this device carries NO migration elements (mux links straight to nvinfer); a bin pinned to ANOTHER device carries an nvdsxfer pair: xfer-in migrates the batched surfaces to the shard's device (nvinfer rejects cross-GPU input memory), xfer-out brings the batch back to THIS device so demux/display/OSD stay on the pipeline gpu. Cross-device placement requires CUDA peer access (NVLink/PCIe P2P) + the nvdsxfer factory; otherwise the placement is rejected at bin creation.
Definition at line 373 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ lane_leaky
| gboolean ModelMuxConfig::lane_leaky |
copy-lane nvvideoconvert output-buffers (0 = element default).
Size to the slow lane's depth to keep the copy from backpressuring the converter.
[multimodel] lane-leaky: leak policy of the per-(stream,role) in_q/out_q lane queues. TRUE (default, existing behaviour): leaky=downstream – a full lane drops its OLDEST buffer instead of blocking, so a stalled/slow model can never backpressure the shared tee + input demux and stall the OTHER streams (freshness over completeness; live sources keep moving when per-model fps mismatch). FALSE: non-leaky lanes guarantee COMPLETENESS (every frame reaches every routed model) but a slower A/B twin then backpressures the lane and can stall mismatched-fps siblings sharing the tee.
Definition at line 363 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ model_mux_push_timeout
| guint ModelMuxConfig::model_mux_push_timeout |
Definition at line 335 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ model_sharding
| gboolean ModelMuxConfig::model_sharding |
dedup-on-idle policy (default TRUE).
When a model goes idle as a lone redundant warm copy (no twin), TRUE moves it out of its role pool (primary/shadow) into the shared limbo pool so a later request from EITHER role reuses that one warm copy without a reload – fewer warm copies overall. FALSE keeps it role-tagged in its pool. The configured default model is never moved. Redundant twins are freed either way.
Definition at line 318 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ models
| ModelCatalogEntry ModelMuxConfig::models |
model registry: built from [MultiModelInference] default primary (+ optional shadow).
No per-camera mapping – every stream uses the default models. MODEL CATALOG: model NAME -> base config (+ auto-detected backend type), parsed from the [model-name-<idx>] blocks. A catalog entry is metadata only – it is NOT instantiated until a role ref (default_primary/shadow) or the model/load API loads a concrete (name, version) instance from it. version/engine_file on a catalog entry are unset (per-instance, runtime).
Definition at line 397 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ mux_height
| guint ModelMuxConfig::mux_height |
mux surface resolution, all muxers (default 1920)
Definition at line 333 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ mux_live_source
| gboolean ModelMuxConfig::mux_live_source |
mux surface resolution, all muxers (default 1080)
1=live/RTSP, 0=file – a source property (def TRUE)
Definition at line 334 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ mux_width
| guint ModelMuxConfig::mux_width |
auto-shard streams when per-model cap exceeded.
TRUE (default): overflow streams get a new shard of the same model (scale-out, existing behaviour). FALSE: streams beyond the per-model batch cap are silently ignored – matching nvinfer behaviour when more streams than batch-size arrive.
Definition at line 332 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ num_bindings
| guint ModelMuxConfig::num_bindings |
Definition at line 429 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ num_models
| guint ModelMuxConfig::num_models |
Definition at line 398 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ num_versions
| guint ModelMuxConfig::num_versions |
Definition at line 403 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ per_model_max_streams
| guint ModelMuxConfig::per_model_max_streams |
Definition at line 292 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ perf_metric_interval_sec
| guint ModelMuxConfig::perf_metric_interval_sec |
enable fps/latency measurement (default FALSE)
Definition at line 342 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ placement_policy
| gint ModelMuxConfig::placement_policy |
[multimodel] placement-policy (pack|spread): how unconstrained streams are placed across a multi-GPU instance group.
See ModelPlacementPolicy.
Definition at line 377 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ process_mode
| gint ModelMuxConfig::process_mode |
one display batch (TRUE) vs per-role
Definition at line 294 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ shard_compact
| gboolean ModelMuxConfig::shard_compact |
[multimodel] enable: 0 => bypass the inference bin (decode->display passthrough)
Definition at line 305 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ unified_batch
| gboolean ModelMuxConfig::unified_batch |
default cap for a model if unset
Definition at line 293 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ versions
| ModelVersionArtifact ModelMuxConfig::versions |
PER-VERSION artifacts: optional engine/config overrides per (name, version), parsed from [model-<name>-<version>] blocks.
Consulted when a (name,version) is instantiated (default, per-sensor, or load). Absent => the model's base catalog config/engine is used.
Definition at line 402 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
◆ versions_lock
| GMutex ModelMuxConfig::versions_lock |
guards versions[] / num_versions – the ONLY ModelMuxConfig state mutable at RUNTIME: modelmux_config_register_version (model/load|update trigger) and modelmux_control_registry_unregister_version (OTA rollback) mutate on the main loop under api_control_lock, while GET model/status resolves placements on the CIVETWEB thread under ib->lock only.
Readers may already hold api_control_lock and/or ib->lock, so the LOCK ORDER is (api_control_lock | ib->lock) -> versions_lock, and versions_lock NEVER wraps another lock: critical sections copy out what they need and release – never file I/O (modelmux_config_derive runs strictly after unlock). Initialised wherever the ModelMuxConfig is (re)zeroed (modelmux_config_parse / gst_modelmux_build defaults); cleared in modelmux_config_clear.
Definition at line 415 of file sources/gst-plugins/gst-nvmodelmux/gstnvmodelmux_priv.h.
The documentation for this struct was generated from the following file: