We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Lobsters·Jun 28·5 min read
huff12 - a 12-stream Huffman decoder for Apple Silicon
TL;DR
Decoding multiple Huffman bitstreams efficiently is challenging, especially on Apple Silicon. The huff12 tool was developed as a 12-stream Huffman decoder optimized for Apple M1 and M4 chips.
✦ Why It Matters
Engineers can leverage huff12 for high-speed Huffman decoding on Apple Silicon, improving performance in data-intensive applications.
Key Takeaways
- Utilize huff12 for efficient multi-stream Huffman decoding.
- Benchmark performance on Apple M1 and M4 for optimal results.
- Consider SIMD codecs for even faster decoding speeds.
- Ensure byte-exact output for data integrity in applications.
Full Summary
Efficiently decoding multiple Huffman bitstreams is crucial for high-performance applications, particularly on modern processors like Apple's M1 and M4. The huff12 tool was created as a hand-tuned AArch64 kernel capable of decoding 12 independent Huffman bitstreams simultaneously, with all but the last stream processed in a forward manner.
It was benchmarked to sustain decoding speeds of approximately 5.1 GB/s on an M4 chip and 3.5 GB/s on an M1 Max, achieving around 0.86 cycles per byte at 4.4 GHz. While it decodes data about twice as fast as zstd's Huffman implementation, it is still slower than SIMD-based codecs like pivco-huffman, which can outperform it by 1.1 to 1.8 times.
The output from huff12 is verified to be byte-exact, ensuring accuracy in data processing. This development highlights the potential for optimizing Huffman decoding on wide ARM CPUs, providing a strong scalar baseline for future enhancements.