TensorBuffer Struct

Storage handle for a tensor — opaque container for one of four backing memory kinds. More...

Declaration

struct simaai::neat::TensorBuffer { ... }

Included Headers

#include <TensorCore.h>

Public Member Functions Index

Mapping	map (MapMode mode) const
	Map the buffer for read/write access; returns a scoped Mapping. More...

Public Member Attributes Index

StorageKind	kind = StorageKind::Unknown
	How the storage was acquired (CPU-owned, external, GstSample, device). More...
Device	device {}
	Where the buffer is accessible. More...
std::size_t	size_bytes = 0
	Total backing-memory size in bytes. More...
std::shared_ptr< void >	holder
	Lifetime guard (shared_ptr that owns/refcounts the underlying memory). More...
void *	data = ...
	Optional direct pointer (some storage kinds set this; others rely on map_fn). More...
std::function< Mapping(MapMode)>	map_fn
	Custom map function (set for non-trivial storage kinds). More...
std::uint64_t	sima_mem_target_flags = ...
	SIMA memory target flags (advanced; for accelerator-aware allocators). More...
std::uint64_t	sima_mem_flags = 0
	SIMA memory flags (cache, alignment, etc.). More...
std::vector< Segment >	sima_segments
	Named segments for multi-region buffers (composite formats, packed outputs). More...

Description

Storage handle for a tensor — opaque container for one of four backing memory kinds.

TensorBuffer is the abstraction over the four StorageKinds. The map_fn callback unifies the access path: regardless of where the bytes live (CPU heap, GStreamer pool, accelerator scratch), map(mode) returns a Mapping you can read or write. The holder shared_ptr keeps the underlying memory alive for the lifetime of the buffer (and any Mapping derived from it).

Multi-segment buffers (sima_segments) carry several named regions in one allocation — used for composite formats like NV12 (Y + UV) and packed multi-tensor MLA outputs.

Definition at line 427 of file TensorCore.h.

Public Member Functions

map()

Mapping simaai::neat::TensorBuffer::map (MapMode mode)

inline

Map the buffer for read/write access; returns a scoped Mapping.

If the storage has a custom map_fn, calls it. Otherwise, returns a Mapping wrapping the bare data pointer. Always sets keepalive to the buffer's holder so the Mapping safely outlives buffer destruction.

Definition at line 451 of file TensorCore.h.

451 Mapping map(MapMode mode) const {

452 Mapping mapping;

453 if (map_fn) {

454 mapping = map_fn(mode);

455 } else {

456 mapping.data = data;

457 mapping.size_bytes = size_bytes;

458 }

459 if (!mapping.keepalive) {

460 mapping.keepalive = holder;

461 }

462 return mapping;

463 }

Public Member Attributes

data

void* simaai::neat::TensorBuffer::data

Optional direct pointer (some storage kinds set this; others rely on map_fn).

Initialiser

= nullptr

Definition at line 434 of file TensorCore.h.

434 void* data =

device

Device simaai::neat::TensorBuffer::device {}

Where the buffer is accessible.

Definition at line 430 of file TensorCore.h.

430 Device device{};

holder

std::shared_ptr<void> simaai::neat::TensorBuffer::holder

Lifetime guard (shared_ptr that owns/refcounts the underlying memory).

Definition at line 433 of file TensorCore.h.

433 holder;

kind

StorageKind simaai::neat::TensorBuffer::kind = StorageKind::Unknown

How the storage was acquired (CPU-owned, external, GstSample, device).

Definition at line 428 of file TensorCore.h.

428 StorageKind kind = StorageKind::Unknown;

map_fn

std::function<Mapping(MapMode)> simaai::neat::TensorBuffer::map_fn

Custom map function (set for non-trivial storage kinds).

Definition at line 437 of file TensorCore.h.

437 map_fn;

sima_mem_flags

std::uint64_t simaai::neat::TensorBuffer::sima_mem_flags = 0

SIMA memory flags (cache, alignment, etc.).

Definition at line 440 of file TensorCore.h.

440 std::uint64_t sima_mem_flags = 0;

sima_mem_target_flags

std::uint64_t simaai::neat::TensorBuffer::sima_mem_target_flags

SIMA memory target flags (advanced; for accelerator-aware allocators).

Initialiser

= 0

Definition at line 438 of file TensorCore.h.

438 std::uint64_t sima_mem_target_flags =

sima_segments

std::vector<Segment> simaai::neat::TensorBuffer::sima_segments

Named segments for multi-region buffers (composite formats, packed outputs).

Definition at line 441 of file TensorCore.h.

441 std::vector<Segment> sima_segments;

size_bytes

std::size_t simaai::neat::TensorBuffer::size_bytes = 0

Total backing-memory size in bytes.

Definition at line 431 of file TensorCore.h.

431 std::size_t size_bytes = 0;

---

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

• TensorCore.h

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.