Glossary

Terms and acronyms that appear across the framework's docs and code.

Framework concepts

Term	Definition
Neat (the framework)	The library this documentation describes — a C++/Python framework for building, validating, and running GStreamer-based AI inference pipelines on Modalix.
Node	The smallest building block of a pipeline. A typed wrapper over a GStreamer element. See include/builder/Node.h.
NodeGroup	A bundle of Nodes that travels as a unit (e.g., a model's preprocess).
Session	The assembly stage that turns Nodes into a runnable Run.
Run	A live, running pipeline. Push samples in, pull samples out.
Model	The simplified entry point for loading and running an MPK.
MPK	Model Pack — a .tar.gz / .mpk archive bundling a compiled model and everything it needs at runtime.
Sample	The framework's unit of pipeline data — wraps a tensor or encoded media plus metadata.
Tensor	The framework's typed view of a buffer of pixels / audio / inference results.
TensorBuffer	The underlying memory backing a Tensor — carries the (buffer_id, paddr, vaddr) triple.
Segment	A named memory region with allocator + access metadata. Buffers come from segments.

Hardware

Term	Definition
Modalix	SiMa's edge AI SoC platform. The framework targets it.
MLA	Machine Learning Accelerator — the SoC's main inference engine.
MLASHM	MLA Shared Memory — a low-latency memory region the MLA reads from.
EV74 / CVU	Compute Vision Unit — a SIMD-friendly DSP for preprocess / postprocess kernels.
A65	The application ARM cores running Linux.
APU	Audio Processing Unit.
M4	A Cortex-M4 used for low-level coordination (RPMsg, hardware sequencing).
TVM	Apache TVM — the framework's CPU-side fallback compiler for ops the MLA can't handle.
VCCM	A SoC-internal coherent memory region used by some accelerators.
OCM	On-chip memory used by the MLA.
DMS	Direct Memory Server — the SoC's memory controller / allocator service.
RPMsg	Remote-processor messaging — the IPC channel between A65 and M4.
IOMMU	I/O memory-management unit — maps physical to virtual for hardware.

Data formats

Term	Definition
BF16	Brain Float 16 — IEEE 754 binary16 with full FP32 exponent range. The MLA's preferred float dtype.
NV12	YUV 4:2:0 with Y plane + interleaved UV plane. The framework's default decoded-video format.
I420	YUV 4:2:0 with separate Y, U, V planes.
HWC / CHW	Tensor layouts (Height-Width-Channels vs. Channels-Height-Width).
Tessellation	The tile-shuffle that arranges a tensor into the geometry the MLA's input scratchpad expects. Pure layout — same bytes, different order.
Quantization	Mapping FP32 values to INT8 with a scale and zero-point.

Operational

Term	Definition
Generic Preproc	A preprocess upgrade that fuses arbitrary user-supplied transforms into the standard preprocess graph.
BoxDecode	A postprocess fusion that runs NMS / decode for detection models on the EV74.
DetectionMeta	The metadata struct attached to detection-model output samples by BoxDecode.
GstSimaMeta	The framework's GStreamer metadata struct, attached to every framework-managed buffer.
Route plan	The framework's compile-time decision about which processor runs each stage and which segments hold each buffer.
Repro launch string	The deterministic gst-launch text reproducer emitted by Session::describe().

See also

• Session::describe() — emits the repro launch string.
• The dtype contract.
• Memory model.
• Processor backends.