11:05 <msh> has anyone looked at somehow combining packing and placing in nextpnr? experimenting here I see a fmax speedup (at a cost of slices) if I remove the FF packing in pack_remaining_luts. I assume some packed FFs are useful but some harm timing - after placing it'd be known. this is ecp5

11:09 <lofty> msh: you want to do exactly the opposite, really

11:20 <gatecat> msh: it's complicated; newer arches tend to place the two seperately but that doesn't always help matters

11:20 <gatecat> I had a go doing this for ECP5 a year or two ago and it didn't really help overall

11:21 <Sarayan> Theoretically you can jointly optimize the cost function (e.g. fmax) jointly, but I suspect the search space is insane

11:21 <Sarayan> (remove jointly to taste)

11:22 <Sarayan> it's probably a very fun and very complicated research field actually

11:23 <Sarayan> especially since you're optimizing a max, and that's always a pain because you can change a lot of stuff without changing a max

11:24 <Sarayan> and if you optimize a mean, or a mean square, you can end up with a very high max

11:24 <msh> hm right. my thought was to try greedily unpacking paired ff/luts near yhe critical path

11:24 <msh> lofty: what do you mean by opposite sorry?

11:25 <gatecat> msh: it's a big, PITA, patch but ultimately the fix would be to make FFs and LUTs separate bels like Nexus

11:25 <gatecat> and then manipulating them seperately is much easier

11:25 <lofty> And like Mistral! :p

11:25 <Sarayan> fwiw they seem to be separate in quartus

11:25 <lofty> Effectively the placer decides the packing

11:25 <gatecat> e.g. you could do https://github.com/YosysHQ/nextpnr/pull/863 but only if criticality is less then some unit

11:26 <msh> ah *nod*

11:28 <gatecat> my predecessor's attempt at implementing the former part was https://github.com/YosysHQ/nextpnr/tree/dave/ecp5-split-bels and https://github.com/YosysHQ/prjtrellis/tree/split_slice

11:30 <Sarayan> I know Micko is kinda afraid of the legal issues, but I dream of a llvm-of-fpga with yosys, nextpnr, mistral and its friends in one tree (you can add amaranth in there too, I wouldn't mind). It would be nice for experimenting with the fundamental algorithms while having the simplicity of use of a software compiler (at least the good ones)

11:31 <lofty> Just throw GHDL in there and now everything is GPL :P

11:32 <Sarayan> what's ghdl?

11:32 <Sarayan> (I wouldn't mind throwing an equivalent of gtkwave in there though)

11:33 <Sarayan> ghdl = vhdl frontend for that beast + cxxrtl?

12:14 <lofty> Sarayan: yeah

13:19 <msh> gatecat: those patches are "working" but don't give a benefit?

13:19 <gatecat> they worked a couple of years ago

13:20 <msh> *nod* thanks. I'll have a play around with them

13:20 <gatecat> have fun!

13:21 <gatecat> if they are at all promising then I'm happy to have a go at getting them back up to date

18:25 <Sarayan> hi, when I clean -dfx then do a make config-clang then make install, I'm getting a weird error:

18:25 <Sarayan> [ 99%] Building abc/abc-f6fa2dd

18:25 <Sarayan> ERROR: ABC contains local modifications! Set ABCREV=default in Yosys Makefile!

18:25 <Sarayan> make: *** [Makefile:745: abc/abc-f6fa2dd] Error 1

18:26 <Sarayan> I'm in sync with origin/master

18:27 <Sarayan> oh, it's a friggin' submodule