google / XNNPACK

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

Remove excessive unrolling from some microkernels #7524

Closed copybara-service[bot] closed 4 days ago

copybara-service[bot] commented 4 days ago

Remove excessive unrolling from some microkernels

It doesn't really make sense to unroll more than 4x in most cases. In the cases we used a higher unroll factor, there was a kernel that unrolls less than 4x that performs almost identically.