[Feature] Math Ops bitwise

[3Rr0RHACK3R]

Feature

Math Ops Using Bitwise

Description

I Think that when a calculation involves a integer , the math engine should calculate it using bitwise techniques . (I Don't Know whether this is already implemented) and this is mainly Because Users Might Have to Buy GPU's to run models properly. CPU Users Struggle a lot

[akleine]

That's probably correct, but don't expect much from this kind of low-level optimization.

As a former assembler programmer, I've looked at ggml's important function ggml_vec_dot_f32(). In this special CPU case (ARMv8 @ Android and Raspberry Pi 5) I have noticed that code produced by compiler (cc and NDK) is so very well and after some manual edits I've got a speed gain of less than 5%.
So at all I think that work on low-level optimization is wasted time. But perhaps you should ask again at ggml repository.