[Libre-soc-dev] DCT/FFT augmentations

[Luke Kenneth Casson Leighton]

On Fri, Jun 25, 2021 at 11:08 AM Luke Kenneth Casson Leighton <lkcl at lkcl.net>
wrote:


> this is an interesting experiment all round, how to add new
> experimental (overlapping) opcodes
>

correction: it's an EXPENSIVE and costly lesson in why the rule was set:

     NO CHANGING BEHAVIOUR OF EXISTING OPCODES

* the SVP64Asm decoder is now "poisoned" with special exceptions, to
  recognise "sv.ldbr" with *three* arguments, and turn it into a v3.0B with
*two*.

* the "RC" of sv.ldbr RT, SVD(RA), RC if turning into Vectorised RC, well,
  there's not much point doing that, but it will literally be the first
time that
  a register RA/B/C will never have been left as scalar

the only reason i'm carrying on is because the gains as a strategic
algorithm (DCT, FFT) are so high.

l.