google / XNNPACK

High-efficiency floating-point neural network inference operators for mobile, server, and Web
Other
1.89k stars 376 forks source link

Remove u32-f32-cvt kernels #7439

Closed copybara-service[bot] closed 2 weeks ago

copybara-service[bot] commented 2 weeks ago

Remove u32-f32-cvt kernels

We don't really properly support uint32 datatypes. This is only used for converting reduction outputs back to qs8. However, we don't do this via uint32 anywhere else (e.g. gemms), and I don't think we can properly support it.

Using s32 instead just means accumulators are effectively 31 bits instead of 32, which seems insignificant.