13:14 <somlo> _florent_, daveshah: tried the point-to-point 128-wide rocket<->litedram link with LiteDRAMAXI2Native buffer_depth as low as 2 (for both write and read), still won't fit on the versa5g (got an assert error when I tried buffer_depth = 1 :)

13:15 <somlo> this is just fyi, not actually asking for anything right now :)

13:23 <daveshah> Thanks for testing!

13:32 <_florent_> somlo: ok, maybe you can also test by changing buffered=True to buffered=False on w_buffer/r_buffer

13:48 <somlo> _florent_: 99%, now placing and routing, after which I'll try to start it up :)

13:53 <somlo> I'll have to use the trellisboard to measure if (and how much) removing buffering affects performance

13:53 <somlo> but my hunch is not much (if at all), since this is a point-to-point connection

14:05 <somlo> and it boots, and works fine, seems to have decent performance as well

14:07 <somlo> now, back to the trellisboard, where I can verify whether reduced buffering has any sort of impact on performance, before we decide if it's worth talking about exposing the true/false option through a LiteDRAMAXI2Native parameter...

14:28 <_florent_> somlo: ok good. The current w/r_buffer depth were a good compromise (resource usage vs performance) i found while testing LiteDRAM on DDR4 with continuous writes/reads, but that's possible it's not optimal for your use-case, so this could indeed be interesting to expose more parameters.

15:34 <somlo> didn't have patience to wait for nbench memory-specific results, but the integer tests feel marginally better on the original (unmodified) buffer depth

15:35 <somlo> so we'd only do it for area, and that also only for the versa, and only with the current state of nextpnr :)

15:35 <somlo> I'd say it isn't worth it, I'd rather wait for daveshah's next round of p&r improvements :)

15:36 <somlo> particularly since I'm now focusing on the trellisboard, which has sufficient room to spare

15:37 <somlo> in the mean time, litex+rocket fits fine on the versa when using the wishbone-based width converter, so it's not like one can't get a bitstream to boot 64bit linux if they really really want it...