leah2 has quit [Ping timeout: 258 seconds]
Entei[m] has joined #fedora-riscv
leah2 has joined #fedora-riscv
iooi has quit [Quit: iooi]
<
davidlt[m]>
Morning 🌄
<
davidlt[m]>
I will attempt to quickly cook new disk images with v6.3.11 today before v6.4 arrives.
<
davidlt[m]>
BTW, Firefox finally compiled 🔥
drewfustini has quit [Ping timeout: 240 seconds]
drewfustini has joined #fedora-riscv
esv_ has joined #fedora-riscv
esv_ has quit [Ping timeout: 246 seconds]
<
davidlt[m]>
Have no recommendations right now, just started watching some.
<
davidlt[m]>
This is nothing new, but some might still don't know this.
<
davidlt[m]>
It will explain why not all PCIe cards will work on riscv64 systems, because PCIe also incl. x86_64 binary firmware.
<
davidlt[m]>
Basically you need to emulate x86_64 on your host (aarch64, riscv64) [thanks to QEMU TCG] to run that early firmware blobs on the card.
<
davidlt[m]>
That's of course on UEFI firmware (Tianocore EDK2).
<
conchuod>
Jeremy Bennett is a good speaker
<
davidlt[m]>
It's nice to see RISE being mentioned in videos :)
<
davidlt[m]>
I want to know if they picked baseline ISA for Android.
<
davidlt[m]>
He said it!
<
davidlt[m]>
RVA22 mandatory or optional extensions ;)
<
davidlt[m]>
At around ~4:30 in the video.
<
conchuod>
davidlt[m]: Is it worth watching though?
<
davidlt[m]>
Well, I listen to them instead of music right now.
<
davidlt[m]>
Watching not really.
<
davidlt[m]>
I was thinking Whisper AI + GPT4 for summary would be enough :D
<
conchuod>
All the titles are very $marketingstatement
<
davidlt[m]>
Yes they are.
<
davidlt[m]>
That's why I have never been at RISCV Summit (any).
<
davidlt[m]>
The amount of really interesting content is low. The most exciting part is meeting folks there, I guess.
<
conchuod>
I am not sure that "exciting" is the word for that, certainly not for 97.5% of the attendees.
<
conchuod>
The LLVM talk from Alex should be good
<
davidlt[m]>
at ~17 minute mark he again talks about RVA22 + probably vector + vector crypto
<
davidlt[m]>
explains kinda why scalar crypto is not an option in RVA23
<
davidlt[m]>
Actually he as a few places with "probably". So they thinking a lot about RVA22.
<
conchuod>
Ye, Alex's one is worth it if you aren't super up to date on LLVM RISC-V stuff
<
davidlt[m]>
Ah next slide, what's next after rva22 + vector + vector crypto :)
<
davidlt[m]>
That looks like RVA23 at the 1st glance.
<
davidlt[m]>
rwmjones: are you up for doing more compile benchmarks on VF2?
<
davidlt[m]>
From Mark H. talk:
<
davidlt[m]>
This kinda shows that Android Features are post RVA23 :)
<
davidlt[m]>
What the hell is "Public Name X"? :)
<
davidlt[m]>
Isn't RVAXY already that name?
<
rwmjones>
davidlt[m]: hey
<
davidlt[m]>
It basically shows that post-RVA23 (RVA24?) is a major thing.
<
rwmjones>
kernel huh :-?
<
rwmjones>
actually maybe llvm is worse
<
rwmjones>
anyway yes, I'll set one of them off
<
davidlt[m]>
Those both take similar time to build.
<
davidlt[m]>
The last one will be GCC ;) But that's later.
<
rwmjones>
looks like about 14 hours?
<
rwmjones>
15 hours for llvm
<
davidlt[m]>
Yeah, so should be 8-10 hours on VF2.
<
davidlt[m]>
At least I expect that.
<
rwmjones>
I'm just doing $ time rpmbuild --rebuild llvm-16.0.5-1.0.riscv64.fc38.src.rpm
<
rwmjones>
cmake will automatically do parallel builds IIRC?
<
davidlt[m]>
rwmjones: just remember to disable tmpfs for /tmp
<
rwmjones>
yeah I did that
<
rwmjones>
btw, "interesting" TCG bug:
<
rwmjones>
I'm not sure to what extent it affects riscv because its to do with self-modifying code
<
rwmjones>
& (I assume) riscv forces you to explicitly flush the i-cache
masami has joined #fedora-riscv
kalev has quit [Quit: leaving]
masami has quit [Quit: Leaving]
<
rwmjones>
llvm is @ 730/2932
<
davidlt[m]>
Do you monitor memory usage?
<
rwmjones>
not explicitly but ...
<
rwmjones>
total used free shared buff/cache available
<
rwmjones>
Mem: 7925 687 4079 17 3158 7136
<
rwmjones>
Swap: 0 0 0
zsun has joined #fedora-riscv
zsun has quit [Quit: Leaving.]
<
conchuod>
> & (I assume) riscv forces you to explicitly flush the i-cache
<
conchuod>
Far from an expert on these kind of things rwmjones , but for Linux at least we do icache maint. for code patching.
<
conchuod>
The ISA is pretty bad about these kinds of things though, there's been a few "discussions" about it on LKML ;)
<
rwmjones>
conchuod: yeah indeed we do, we do a "self-iret" thing
<
rwmjones>
(or SERIALIZE for intel cpus that support it, but I think that's very recent)
<
rwmjones>
however the problem here is really with TCG itself
<
rwmjones>
another core (TCG thread) is retranslating the code before the write happened
<
rwmjones>
it sees the old memory contents and so translates it to the old code (using int3)
<
rwmjones>
and then because we never invalidate the TB again, we always run the wrong code
<
rwmjones>
the iret doesn't really help (unless we invalidate every TB which would be slow)
<
rwmjones>
talking about it though, I might see if a patch which does invalidate every TB on iret actually fixes the problem ...
<
rwmjones>
it will at least confirm the theory is correct
zsun has joined #fedora-riscv
<
rwmjones>
the answer is that calling tb_flush (on sync_core) absolutely kills performance, so much so that the guest barely makes progress
zsun has quit [Quit: Leaving.]
<
conchuod>
No progress, no races.
<
conchuod>
Seems like a fix to me
dtometzki is now known as dtometzki__
esv has quit [Remote host closed the connection]
iooi has joined #fedora-riscv
<
rwmjones>
llvm is @
<
rwmjones>
2603/2932
esv has joined #fedora-riscv
<
davidlt[m]>
Heh, this actually might finish today.
<
sorear>
a little bit concerned every time I see a slide that implies the author thinks profiles will be monotonic forever
iooi has quit [Read error: Connection reset by peer]
iooi has joined #fedora-riscv
<
rwmjones>
llvm is doing RPM stuff ...
<
rwmjones>
Requires(rpmlib): rpmlib(CompressedFileNames) <= 3.0.4-1 rpmlib(FileDigests) <= 4.6.0-1 rpmlib(PayloadFilesHavePrefix) <= 4.0-1
<
rwmjones>
Recommends: llvm-debugsource(riscv-64) = 16.0.5-1.0.riscv64.fc38
<
rwmjones>
Checking for unpackaged file(s): /usr/lib/rpm/check-files /home/rjones/rpmbuild/BUILDROOT/llvm-16.0.5-1.0.riscv64.fc38.riscv64
<
rwmjones>
writing rpms ..
<
rwmjones>
llvm-16.0.5-1.0.riscv64.fc38:
<
rwmjones>
real601m41.466s
<
rwmjones>
user2193m5.753s
<
rwmjones>
sys95m42.081s
<
rwmjones>
I'm doing the kernel now ...
<
rwmjones>
$ time rpmbuild --rebuild kernel-6.3.11-200.0.riscv64.fc38.src.rpm
<
rwmjones>
601m is 10 hours
<
rwmjones>
but my feeling is that it was very fast apart from the serialized RPM bits at the end
<
rwmjones>
(ie amdahl's law, again)
iooi has quit [Read error: Connection reset by peer]
iooi has joined #fedora-riscv