hardware

A bfloat16 matrix multiply-accumulate (MMA) unit written in Amaranth HDL.

It builds bottom-up from arithmetic primitives to a 4×4×4 MAC array (MMA) that computes D = A·B + C over bf16 matrices, accumulating in extended (26-bit mantissa) precision and rounding to bf16 only at the output.

Setup

uv sync

This installs the project editable, putting src/ on the import path so tests can from bfloat16 import ... directly.

Test

uv run pytest test/ -v              # all tests
uv run pytest test/ --vcd           # also dump .vcd waveforms

Lint

ruff check --fix && ruff format

Layout

src/

• bf16_mac.py (BF16_MAC) is the fused multiply-add core.
• pe_mac.py wraps it with a registered accumulator.
• mma.py (MMA) is the 16-PE array.

The rest are standalone arithmetic primitives (adders, aligner, normalizer, LZA, multiplier, rounder).

test/

• amaranth.sim benches.
• test_mma.py holds the single-rounding FMA reference model.

The per-primitive files cover the building blocks.