edgefirst-tensor

Zero-copy tensor memory management for edge AI applications.

This crate provides a unified interface for managing multi-dimensional arrays (tensors) with support for different memory backends optimized for ML inference pipelines.

Memory Types

Type	Description	Use Case
DMA	Linux DMA-BUF allocation	Hardware accelerators (GPU, NPU, video codecs)
SHM	POSIX shared memory	Inter-process communication, zero-copy IPC
Mem	Standard heap allocation	General purpose, maximum compatibility
PBO	OpenGL Pixel Buffer Object	GPU-accelerated image processing (created by ImageProcessor)

Features

Automatic memory selection - Tries DMA → SHM → Mem based on availability
Zero-copy sharing - Share tensors between processes via file descriptors
Memory mapping - Efficient CPU access to tensor data
ndarray integration - Optional conversion to/from ndarray::Array (feature: ndarray)

Quick Start

use edgefirst_tensor::{Tensor, TensorMemory, TensorTrait, TensorMapTrait};

// Create a tensor with automatic memory selection
let tensor = Tensor::<f32>::new(&[1, 3, 224, 224], None, None)?;
println!("Memory type: {:?}", tensor.memory());

// Create with explicit memory type
let dma_tensor = Tensor::<u8>::new(&[1920, 1080, 4], Some(TensorMemory::Dma), None)?;

// Map tensor for CPU access
let mut map = tensor.map()?;
map.as_mut_slice().fill(0.0);

// Share via file descriptor (Unix only)
#[cfg(unix)]
let fd = tensor.clone_fd()?;

Platform Support

Platform	DMA	SHM	Mem	PBO
Linux	Yes	Yes	Yes	Yes (with OpenGL)
macOS	No	Yes	Yes	No
Other Unix	No	Yes	Yes	No
Windows	No	No	Yes	No

Feature Flags

ndarray (default) - Enable ndarray integration for array conversions

Environment Variables

EDGEFIRST_TENSOR_FORCE_MEM - Set to 1 or true to force heap allocation

License

Licensed under the Apache License, Version 2.0. See LICENSE for details.

edgefirst-tensor 0.11.0