ChillFish8 / cfavml

An unopinionated SIMD vector operation library for Rust, supporting no_std and no-alloc workloads.
Apache License 2.0
3 stars 1 forks source link

A difficult problem to solve: Computation with higher-bitwidth accumulators #5

Open ChillFish8 opened 3 months ago

ChillFish8 commented 3 months ago

In some situations like int8 dot product, we want to accumulate into a higher bitwidth accumulator, but how do we go about supporting this in a sane and logical way? Currently, the system is very simple T in == T out but if we want to start doing say accumulate to u32 then this becomes considerably harder...