[Libre-soc-dev] daily kan-ban update 25nov2021

lkcl luke.leighton at gmail.com
Thu Nov 25 18:32:39 GMT 2021


On November 25, 2021 4:26:27 PM UTC, Tobias Platen
<libre-soc at platen-software.de> wrote:
>today: more work on LoadStore1 and its test case

thanks tobias. apologies i cannot help focus on this right now,
myself, what i think should happen here is to invite you to be part of
the NGI POINTER team and we can put some EUR your way, simply out of
recognition of sheer persistence and committment, given that NLnet
"milestones" do not quite fit.

this is a bit long:

as i wrote yesterday, an instruction "addi 5,5,0x20" creates a
write-hazard *with itself* and consequently waits for infinity.

the solution is to add an FU-FU Dependency Matrix:

  "FU num {row} depends for its result on FU num {col}"

and in the FU-FU Matrix it is FORBIDDEN to create a dependency when
{row} == {col}, because that is precisely the scenario that is covered
for "addi write=5, read=5".  it is fine for an instruction to depend
on reg 5 for ANOTHER instruction, of either read-depend or
write-depend: just not on itself.

this distinction, "depend on everything but yourself" is what is
missing from core.py and it will be a LOT of work to sort that out.

therefore, i am looking at the score6600 multi code from now on,
getting it working again.  currently, there is a timing issue after
some changes to MultiCompUnit: src data arrives *one clock late*
(because it is fired by regfile), i have to sort that.

l.



More information about the Libre-soc-dev mailing list