[libre-riscv-dev] [Bug 316] bperm TODO

[bugzilla-daemon at libre-soc.org]

https://bugs.libre-soc.org/show_bug.cgi?id=316

--- Comment #13 from Luke Kenneth Casson Leighton <lkcl at lkcl.net> ---
ok.  right.  re-run the generate -t ilang, and yosys show top.

then trace "rs", note how there are *two* copies of 8-bit chunks of rs?
one that goes into the "lt" (less-than) comparison, and a *second* copy
that goes into the the big "group" (this is a massive multiplexer btw)

that's 2x copies of 8 8-bit chunks.

so let's trace it back.

what thing or things - related to rs - have an 8-bit width?

well, that will be index, won't it?

index = self.rs[8 * i:8 * i + 8]

now, look where index is used.

first time:

   with m.If(index < 64):

that's where the lt (less-than) comes in (times 8)

second time:

   (rb64[index])

this is really important to note the difference between a *python* variable
and a nmigen Abstract Syntax Tree code-fragment.

you have created - with the "index = self.rs[blah blah]" a *python* variable
assignment from a *nmigen AST*.

therefore, whereever you use that python variable it INSERTS A COPY OF THAT
NMIGEN AST... into more nmigen AST.

now, in this particular case, that's fine (it's not a pretty graph) but in
cases where you did something like this:

     index = Cat(rs[blah], 5) * i + rb[i] * 3, rs[i] & rb[i])

or anything hugely and massively complex, that would ALSO GET INSERTED TWICE
and you would end up with a huge duplication of gate logic as a result, which
yosys would be very unlikely to spot and optimise away.

the solution - therefore - the best practice - is never to assign anything
other
than the most basic of nmigen objects - Signals, Consts, Records etc. - to a
python variable

    OR

to actually create a Signal into which that nmigen AST fragment is assigned...
*and then use that signal*.

in this case that would be:

    index = ...
    idx = Signal(8, name="idx%d" % i, reset_less=True)
    m.d.comb += idx.eq(index)

and then

    m.If(idx < 64)

and also

    rb64[idx]

see what difference that makes to the graphviz, ok?

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