Module tensor_arena 

Arena allocator for inference pipeline tensor memory management.

Inference pipelines allocate many small tensors. An arena allocator avoids per-tensor heap allocation by bump-allocating from pre-allocated slabs. This module provides TensorArena, ArenaRegion, ArenaSlice, ArenaStats, and ArenaError.

Design

The arena is organized as a list of fixed-size ArenaRegion slabs. Each region is a Vec<u8> with a bump pointer (offset). Allocations are always 8-byte aligned. When a region is full a new one is created.

Lifetimes are avoided by representing allocated memory as (start, end) byte ranges within a region’s slab, wrapped in ArenaSlice.

Example

use ipfrs_tensorlogic::tensor_arena::{TensorArena, ArenaError};

let mut arena = TensorArena::new(1024 * 1024); // 1 MB regions

// Allocate space for 4 f32 values (16 bytes)
let slice = arena.allocate(4 * 4);

// Write and read back
slice.write_f32(&mut arena, &[1.0, 2.0, 3.0, 4.0]).expect("example: should succeed in docs");
let values = slice.read_f32(&arena);
assert_eq!(values, &[1.0f32, 2.0, 3.0, 4.0]);

// Reset for reuse (no deallocation)
arena.reset_all();

Structs

ArenaRegion

A contiguous slab of memory with a bump pointer.

ArenaSlice

A handle to a contiguous byte range inside a TensorArena region.

ArenaStats

Cumulative statistics for a TensorArena.

TensorArena

Bump-allocating arena for inference-pipeline tensors.

Enums

ArenaError

Errors that can occur during arena operations.