[Libre-soc-dev] MP3 DCT36

[Luke Kenneth Casson Leighton]

hm.  REMAP combined with Context Propagation is two quite large tasks.
if it's ok with you Lauri let's leave REMAP for the "optimised"
version and go in the basic version with some use of LD/ST (if
needed).

REMAP itself involves setting up up to four "Shape" SPRs (four because
you can apply separate ones to dest, src1 src2 and src3).  this would
be where, instead of "i += 4" in the DCT36 algorithm, anything
referencing vec[i] would have instead a REMAP to jump i by 4-4-4-4

however it is not possible to put which registers shoild be REMAPped
into the SVP64 prefix.... we need *another* prefix and that's where
"Context Propagation" comes in.

otherwise we end up with 64+32... 96 bit instructions!

* 32 for the base
* 32 for the SVP64 prefix
* 32 for the REMAP prefix.

i *might* add that. just because it is a LOT simpler than Context
Propagation, which is a hardware-levrl compression algorithm, in
effect (!)

l.