mastodon.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
The original server operated by the Mastodon gGmbH non-profit

Administered by:

Server stats:

364K
active users

Got to the point of building LLVM 15.0.6 on the VisionFive 2 - it's going like a rocket compared to doing the same in QEMU (even though I emulated many more cores with QEMU!)

Graham Markall

CPU Temperature seems to be sat at about 65 deg C, 8GB seems to be plenty of RAM for building with 4 processes. Seems to have got 25% of the way through the build in 30 minutes. Note sure how to check what speed the cores are running at / if they're throttled.

Better than I'd expected, will probably redo this build once I have a heatsink / fan and NVMe to see how much faster it could be.

Spoke too soon, ld was oom-killed linking libLTO.so :-(

Between the various attempts to build (with 4 jobs, then 2) the total times for building LLVM 15 on the VF 2 were:

real: 280 minutes
user: 790 minutes.

(i.e. 4 hours 40 minutes of waiting). Will try again with 3 jobs when I have the NVMe drive and heatsink / fan, to see if that finishes in one go and how fast it goes.

Ran again with NVMe drive and 3 jobs:

real 230m55.985s
user 671m8.705s
sys 20m51.997s

Not bad, LLVM build in under 4 hours, and no OOM!

According to sensors, temp mostly hovered around 70C:

$ sensors
120e0000.tmon-isa-0000
Adapter: ISA adapter
temp1: +69.6 C

Didn't appear to throttle, /sys/devices/system/cpu/cpufreq/policy0/cpuinfo_cur_freq always reported 1500000 - however with 4 cores going I could see it reach 85 and start to throttle... Seems the min freq is 375MHz.

@gmarkall Out of curiosity, are you only building llvm for riscv?

@conor Yeah, I just want to work on JIT so I don't need anything else.

I put together a proof of concept that Numba could JIT Python on RISC-V with Qemu a couple of weeks ago: github.com/gmarkall/numba/comm

Since I got that far, I felt it was justified to invest in some real hardware to make a serious attempt at RISC-V support in Numba / llvmlite in early 2023, hence getting the VF2.

GitHubA couple of cheeky hacks to get successful jit on riscv · gmarkall/numba@7c8c5e7We disable optimisation (because some segfault occurs during opt passes, maybe related to the fact I haven't properly updated llvmlite for llvm 15). This causes some issue for NRT so I've d...
@gmarkall Ahh, figured. My internal gauge of LLVM build times comes from running https://github.com/ClangBuiltLinux/tc-build/build-llvm.py which builds all archs by default. 2 hours (with headroom to improve) for the single target build sounds promising in terms of being able to do native building of things.