1. 9

    1. 1

      I’m not dealing with this area much, so I’m curious - given how often 3x3 and 4x4 matrix multiplication, inversion, and small dot products are used, is there some good reason they’re not available as single simd/avx instructions? Is there a chance they will be? Or is the software/hardware combined approach good enough? (I.e. is it the optimum time sacrifice vs silicon saving)

      1. 2

        Intel just announced something like that, https://en.wikichip.org/wiki/x86/amx

        I suppose something similar could be achieved by reusing GPU function blocks (where all kinds of matrix math should be available) in AMD’s Heterogeneous System Architecture. ARM also adopted it, but apparently software support for using that functionality isn’t very good on ARM or AMD.