Formerly arm-playground.
Neural models involve a lot of matrix multiplies. Speeding up the matrix multiplication provides us with fast neural network inference on the client machine.
The matrix-multiplications as compiled by emscripten from C++ sources into WebAssembly target turned out to be too slow.
From Bugzilla/1746631#c1:
- Wasm Gemm : 95 wps
- Wormhole : 390 wps (+310% to Wasm Gemm)
- Native Firefox gemm
- SSSE3 : 490 wps (+25% to Wormhole, +415% to Wasm Gemm)
- AVX2 : 560 wps (+43% to Wormhole, +489% to Wasm Gemm)
While the original support was intended for x86-64 based architectures through kpu/intgemm, and the above-mentioned interface is written closely bound to kpu/intgemm, an immediate demand surfaced for support for ARM.
A target-agnostic matrix multiply interface suitable for WASM was checked in by collaborators at Mozilla for purposes of speeding up translations in browsermt/marian-dev#49. This repository takes the interface provided above and attempts to provide an implementation for the interface using google/ruy in the short term.
The library target which uses either ruy or intgemm depending on the platform
is made available as moz_intgemm. Most of the sources are in
MozIntGemm directory.
The interface is available for testing closer to the real use case only on the WASM platform. This creates difficulty in doing end-to-end checks closer to the original task.
Fortunately, to test correctness, all we have to do is assert that a certain sequence of operations used through kpu/intgemm implementation gives the same output when used through google/ruy. A translation workload for marian is a composition of the functions specified in the Mozilla interface.
Testing for correctness happens on the x86 backend, that both ruy and intgemm support. When intgemm is turned off on ARM (where it lacks support) and ruy allowed to take over, we thus obtain an ARM backend for bergamot-translator's matrix multiplies, w.r.t integration with firefox and fast local translation extension.
There is also a slow reference implementation vs fast SIMD implementation using NEON intrinsics which covers source-code added here as preprocessing functions (quantize, unquantize, transpose).
The additions here are MIT Licensed. Respective licenses provided by original source in submodules inside 3rd-party/ apply.