Repository navigation
[SIMD] Support optimal 4x4 matrix multiplication. #64238
Description
I wanted to share where we stand on being able to support optimal 4x4 matrix multiplication via our typed_data SIMD APIs so it can be used by e.g. Flutter for close to optimal performance.
- As of today, Flutter uses scalar matrix multiplication (matmul).
- There were some efforts to use a vectorized matmul (Investigate using SIMD to improve performance of common Matrix operations flutter/flutter#60538) but they moved nowhere mostly due to other SIMD code gen issues that are no longer an issue More efficient SIMD signMask for ARM #41950.
- Vectorized matmul is roughly 1.5x~ faster. With shuffle optimizations that are on my radar (SIMD tracking issue. #64170) we should be able to make it 2.5x~.
- However, to get close to the optimum we'd need fused multiply adds (fma)
- That's where the Relaxed SIMD proposal comes in [dart2wasm, typed_data] Support for Relaxed SIMD. #64216 as it exposes fused multiply adds.
- However, to get optimal 4x4 matmul, we need something known as the
indexed fmlawhich allows us to not splat + fma, but use a single chain of indexed fmla instructions for implementing 4x4 matmul.
Wasm SIMD doesn't support indexed fmla, which costs us roughly 30% compared to splat+fma (which is still a nice 5x~ perf gains over scalar! cc @kevmoo).
Optimal matmul gives use almost 6x faster 4x4 matmul on the VM. Given that matmul is a very important operation in Flutter, squeezing out the last 30%~ seems worthwhile.
Activity
kevmoo commentedon Sep 8, 2026
I LOVE that you're digging in here @modulovalue !
Some notes:
vector_mathwas designed on the typed_data bits - as they were many years agovector_mathis now owned by the Flutter team (more or less)☝️ Is product hand waving, speculation, and wishful thinking. But I'm happy to engage with you on it!
mraleph commentedon Sep 9, 2026
It would be nice to quantify this by some sort of Flutter benchmark, which is not a microbenchmark (e.g. it does actual UI rendering and demonstrates that Matrix multiplication is on the hot path). 30% of 50% is "sure, it's worthwhile", 30% of 10% is "maybe, it's worthwhile" and 30% of 5% is pretty much "it's irrelevant".