[Libre-soc-bugs] [Bug 782] add galois field bitmanip instructions

[bugzilla-daemon at libre-soc.org]

https://bugs.libre-soc.org/show_bug.cgi?id=782

--- Comment #67 from Luke Kenneth Casson Leighton <lkcl at lkcl.net> ---
(In reply to Jacob Lifshay from comment #65)
> (In reply to Luke Kenneth Casson Leighton from comment #63)
> > (In reply to Jacob Lifshay from comment #61)
> > > we could change the madd ops to be 3-arg rather than 4:
> > 
> > i already explained clearly and non-negotiably the damaging
> > consequences and cost of that.
> 
> I meant the non-twin madd ops (which aren't used for fft iirc).

ah yes, so sorry, yes those are used in Horizontal-First FFT REMAP
mode.  that was related to complex numbers, where the "usual"
single-for-loop-on-an-instruction would only work if we had actual
complex-mul-and-add [which would be complete overkill and swamp
us with work]

instead by using Horizontal-First it was possible to use *four*
separate and distinct mul-and-adds with intermediary variables
(as scalars)

there was also iirc some interaction between fmadd.. it turned
out to be possible to usefully set up all five REMAPs (RA RB RC, RT TS)
and by a beautiful/horrendous coincidence only some of those
are used by some operations (fmadds?) and then the next instruction
used others, without needing a 2nd REMAP in between.

having worked that all out in eight severely-intensive weeks,
i don't want to go anywhere near it, again.

-- 
You are receiving this mail because:
You are on the CC list for the bug.