Skip to content

v0.2.0

Latest

Choose a tag to compare

@mastercyb mastercyb released this 14 Apr 07:34
· 20 commits to main since this release

honeycrisp v0.2.0

first release of the unified workspace. bare-metal Rust drivers for every compute unit on Apple Silicon — 34k lines of Rust, 134 source files, 282 tests, zero clippy warnings.

four independent repos (unimem, acpu, aruminium, rane) merged into a single cargo workspace with lockstep versioning. aruminium and rane were previously published at 0.1.0 as standalone crates — this release unifies them under the honeycrisp umbrella.

what ships

five crates. one cargo add for everything, or pick what you need:

cargo add honeycrisp          # all four drivers
cargo add acpu                # just CPU/AMX
cargo add aruminium           # just Metal GPU
cargo add rane                # just Neural Engine
cargo add unimem              # just memory

unimem — zero-copy memory (new to crates.io)

IOSurface-backed pinned buffers visible to CPU, GPU, AMX, and ANE through one allocation. four layers: Block (pinned IOSurface, ~20us create), Tape (bump allocator, ~1ns take), Grid (fixed-size tensor pool, ~10ns take/give), Layout (three-tape inference memory: weights/scratch/history).

acpu — CPU/AMX compute (new to crates.io)

pure Rust driver for Apple Silicon CPU. AMX matrix coprocessor via undocumented .word-encoded inline assembly. NEON vector engine for elementwise math, reductions, softmax, RoPE. hardware crypto (AES, SHA-256, PMULL). Goldilocks field arithmetic for ZK (Poseidon2, batch inverse, Merkle roots). PMU performance counters. P-core/E-core affinity.

1.2–10x faster than Apple Accelerate across 80 operations in 16 categories. specs/amx.md is the most complete open-source reference for AMX instruction encoding.

aruminium — Metal GPU (updated from 0.1.0)

direct objc_msgSend FFI to Metal.framework. pre-resolved IMP dispatch for hot-path inference loops — 1.79x faster pipelined throughput than objc2-metal. zero-copy buffer sharing via unimem::Block.

rane — Apple Neural Engine (updated from 0.1.0)

compile MIL programs, load into ANE SRAM, dispatch on hardware. bypasses CoreML entirely — three private frameworks via dlopen, reverse-engineered MIL bytecode format, IOSurface zero-copy I/O. 10-20x lower dispatch overhead than CoreML predict path.

honeycrisp — meta-crate (new)

re-exports all four. use honeycrisp::acpu, use honeycrisp::rane, etc.

what changed from 0.1.0

  • unified cargo workspace with lockstep versioning
  • honeycrisp meta-crate
  • unimem + acpu: first crates.io publish
  • all READMEs rewritten (experimental status, undocumented hardware documentation, performance vs Accelerate)
  • unimem README created from scratch
  • root README with architecture diagram, benchmark table, crates.io badges, combined code example
  • all crates.io metadata (homepage per crate, keywords, categories)
  • 28 clippy warnings resolved
  • stale API references fixed across all docs
  • CLAUDE.md excluded from published crates
  • bench crates marked publish = false

requirements

macOS on Apple Silicon (aarch64-apple-darwin). M1/M2/M3/M4.

experimental. API unstable.

license

cyber license: don't trust. don't fear. don't beg.