[Libre-soc-dev] microwatt / libresoc dcache

[Luke Kenneth Casson Leighton]

On Tue, May 11, 2021 at 10:55 PM Benjamin Herrenschmidt
<benh at kernel.crashing.org> wrote:

> I'll let Paul comment on the TLB, that's his work, but as for delays,
> the biggest issues we have on FPGAs aren't so much logic delays but
> wire delays. "wires" in FPGAs are slow and BRAMs impose placement
> constraints, especially when you have large (N*64-bit) busses going
> in/out. Keep in mind that microwatt targets entry level FPGAs such as
> the little Artix in the Arty board with -1 speed grades.

ahh this was where the 40mhz limit came from, is that right?

all FPGAs and FPGA toolchains not being equal, you may be fascinated
to know that we ran into exactly the same resource utilisation in yosys
that's been widely reported for microwatt builds on ECP5 + nextpnr5 + yosys,
when we direct-ported dcache.vhdl into nmigen for use in Libre-SOC.

(it was so bad that it caused yosys to eat 34 GB of resident RAM and
had the loadavg well over 40 on my 8-core i9 laptop with 64 GB of RAM!)

on a practical note, then: i notice that microwatt has 4 "ways" in the TLB.
perhaps halving that to 2 for resource-constrained FPGAs (and providing
it as a compile-time option) would help?

(i did notice - cc'ing Anton - that there was a commit message about
doing exactly this, for sky130A).

Minerva (an open source RV32 core) has only a 2-way set
on its caches.
https://github.com/lambdaconcept/minerva/blob/0b5f6b2466367f262f9a16a83f9c86fc7f008edf/minerva/cache.py#L19

l.