edgefirst-tensor

Zero-copy tensor memory for edge AI inference pipelines — DMA-BUF, IOSurface, AHardwareBuffer, OpenGL PBO, POSIX shared memory and heap behind one Python API.

Part of the EdgeFirst HAL

edgefirst-tensor is one of five Python packages built from the EdgeFirst Hardware Abstraction Layer.

The EdgeFirstAI/hal repository is the home for all of them — source, issue tracker, architecture documentation and release notes.

This package is the foundation the other four build on; installing any of them installs this one.

Installation

pip install edgefirst-tensor

Requires Python 3.8 or newer and NumPy. Wheels are published for Linux (x86_64, aarch64), macOS (arm64), and Windows (x86_64).

The _codec, _image, _decoder and _tracker extensions locate libedgefirst_tensor.so via DT_RUNPATH=$ORIGIN/../tensor. That assumes every edgefirst.* package lands in the same site-packages tree, which pip normally guarantees. A split layout (pip install --target, some vendored trees) will fail at import with libedgefirst_tensor.so.0: cannot open shared object file. On Windows the library is edgefirst_tensor.dll in that same edgefirst/tensor/ directory and there is no rpath: each sibling package's __init__.py registers the directory with os.add_dll_directory() before loading its extension, because Python 3.8+ does not consult PATH for extension-module DLLs.

Packages install under the PEP 420 edgefirst.* namespace, so imports are edgefirst.tensor, edgefirst.codec, and so on. No package ships an edgefirst/__init__.py — a single one would shadow the namespace and hide its siblings.

Building from source

edgefirst-tensor carries the shared libedgefirst_tensor the codec, image and decoder wheels link against, so these four are built with the dynamic tensor backend:

make wheel            # or, per package, from its directory:
cd crates/python-tensor && maturin build --release

The crate's .cargo/config.toml selects the dynamic backend; a build from the repository root without --config .cargo/tensor-dynamic.toml fails with a message naming the fix. edgefirst.tensor.__tensor_backend__ reports "dynamic" for every wheel built this way.

Quick start

Allocate an image tensor, fill it through a mapped host view, and read it back as NumPy:

import numpy as np
from edgefirst.tensor import Tensor, PixelFormat

# Allocate once; a real pipeline reuses the tensor every frame.
# mem=None selects the best backend available on the platform.
tensor = Tensor.image(1920, 1080, PixelFormat.Rgb, None, "readwrite")

with tensor.map() as view:
    frame = np.asarray(view)  # shape, dtype and strides all carried
    frame[:] = 128

print(tensor.shape, tensor.format, tensor.dtype)

map() returns a HostView implementing the buffer protocol, so np.asarray wraps the tensor's memory without copying — and because the view publishes shape, dtype and the real row stride, no manual reshape is needed and a pitch-aligned DMA buffer is read correctly rather than sheared. The view is released when the with block exits.

The map also owns its cache-coherency bracket, and access chooses the direction. The default "readwrite" flushes the whole buffer on release; a reader does not need that, and on a non-coherent Arm DMA-BUF backing skipping it is a per-frame saving:

with tensor.map("read") as view:
    frame = np.asarray(view)  # read-only view, not writable

pin_host() is the exception: it is deliberately decoupled from coherency so a pinned address can survive across convert() calls, which is why it pairs with an explicit cpu_access() bracket instead.

Handing memory to an external runtime

pin_host() returns a stable host address that outlives any map guard and carries no borrow of the tensor, so a pinned buffer can be given to an inference runtime (TFLite custom allocations, ONNX Runtime external tensors) while your frame loop keeps writing to it:

pin = tensor.pin_host("readwrite")
print(hex(pin.ptr), pin.len, pin.alignment)

# pin.ptr stays valid for the lifetime of `pin` — across re-maps and across
# ImageProcessor.convert() calls — so an external runtime can hold on to it.
pin.release()

What this package provides

Interoperability

Each edgefirst-* package is a separate PyO3 extension module, so edgefirst.tensor.Tensor and, say, edgefirst.codec.Tensor are different Python classes even though they wrap the same Rust type — a known PyO3 limitation (issue #1444). A tensor still crosses package boundaries safely, through the __edgefirst_tensor__ capsule protocol every Tensor implements:

# CORRECT — works regardless of which edgefirst.* package produced obj
if hasattr(obj, "__edgefirst_tensor__"):
    ...

# WRONG — always False for a tensor from a sibling package
if isinstance(obj, edgefirst.image.Tensor):
    ...

edgefirst.tensor.EdgeFirstTensorExportable is a typing.Protocol you can annotate a cross-package parameter with, so a type checker accepts a tensor from any edgefirst.* package. See crates/python-common/INTEROP.md for the full protocol — capsule names, lifetime and ownership rules, and versioning.

Versioning and changelog

All four edgefirst-* packages are versioned and released together with the HAL itself, so a given version number refers to the same source tree in every language. Because of that there is no per-package changelog: release notes for every version live in the single CHANGELOG.md in the hal repository, which follows Keep a Changelog and Semantic Versioning.

Links

• Changelog — release notes for all packages
• Source — the hal monorepo
• Issue tracker
• Package documentation — the underlying Rust crate, including the memory backend matrix and CUDA mapping
• EdgeFirst

License

Apache-2.0