[Libre-soc-dev] bigmul REMAP

[Luke Kenneth Casson Leighton]

jacob came up with a great idea last week, which is to
do x+y as a general-purpose mode that is a peer of the
matrix mode. this massively increases the capability and
power of the "mathematical toolbox" to cover far more
general purpose and scientific algorithms than previously
envisioned.

an example is the AV1 CODEC's Schedule for performing
motion estimation, which uses the following Schedule
in a rhombic shape:
https://bugs.libre-soc.org/show_bug.cgi?id=1155#c28

this (and many other) pattern may be easily reproduced,
and is a justifiable high priority.

also the bigmul patterns may easily be reproduced, including
another idea by jacob that allows us to do full vector looping
of biginteger arithmetic in something mad like 5 instructions
(normally in SIMD ISAs this would be hundreds of loop-unrolled
inline assembler instructions).

there is a way to go however we should easily hit the
cryptoprimitives deadline and also demo it in ed25519,
even if it is just a "glimpse" (focus on one function
rather than attempt to do all of them).

l.


-- 
---
geometry: without it life is pointless
the fibonacci series: easy as 1 1 2 3