@Alexander Monakov
Relays (1)
- wss://relay.ditto.pub – read & write
Recent Notes
@nprofile1q... I guess the main issue is there's no standard way to arrange a dedicated process group for the two processes, otherwise you could make each of them send a signal to the entire group, like this?
sh -c 'exec setpgid sh -c "(sleep 1; kill -- -$$) & (sleep 2; kill -- -$$)"'
sh -c 'exec setpgid sh -c "(sleep 1; kill -- -$$) & (sleep 2; kill -- -$$)"'
Why do they call it a ring if it has two corners
I wonder how io_uring plays with that... does it hook into ptrace to report uring syscalls like normal syscalls?
Or does it have its own bespoke monitoring interface, like netlink?
@nprofile1q... it doesn't, but submission (io_uring_enter) is visible to strace, at which point it decodes the contents of the ring buffer
a sufficiently advanced userspace may enqueue more commands after the submission though, or even alter an enqueued command (racing against the kernel reading the buffer)
a sufficiently advanced userspace may enqueue more commands after the submission though, or even alter an enqueued command (racing against the kernel reading the buffer)
1
nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqsld6cd53kj8pkz9svnnt54um57nvwrlx9sgya7xhvhl5a5a4f3sqdxhcxx
ext4
@nprofile1q... what do you think about trying f2fs on the next one?
nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqsld6cd53kj8pkz9svnnt54um57nvwrlx9sgya7xhvhl5a5a4f3sqdxhcxx
(used rate(...[1h]) and assumed it worked like increase(...[1h]))
after fixing the q...
@nprofile1q... no idea then. perhaps its wear leveling algorithm is not strong enough? what filesystem type did you use?
1
Actually, less than a year - this one lasted 8 months.
What am I doing wrong?
@nprofile1q... How much write traffic was it taking?
1
nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqsld6cd53kj8pkz9svnnt54um57nvwrlx9sgya7xhvhl5a5a4f3sqdxhcxx https://developer.download.nvidia.com/compute/cuda/repos/
have fun
@nprofile1q... thanks again!! I meant nvidia's page https://developer.nvidia.com/cuda-toolkit-archive only has browsable doc links back to 8.0, earlier than that also needs downloading/unpacking.
1
nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqsld6cd53kj8pkz9svnnt54um57nvwrlx9sgya7xhvhl5a5a4f3sqdxhcxx "FMAD/FMAD32/FMAD32I Single-precision floating point multiply-add"
indeed it did cal...
@nprofile1q... oh, thanks! my recollection could be from the Tesla generation then. Did you have it locally? Nvidia hosts online docs only back to 8.0.
1
nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqsld6cd53kj8pkz9svnnt54um57nvwrlx9sgya7xhvhl5a5a4f3sqdxhcxx but yeah the third one will be used primarily by layered implementations or llvmpipe, ...
@nprofile1q... I thought that nvidia's cuobjdump used ffma and fmad, but apparently that's a false memory and their machine ISA never had fmad...
Anyway, the ffma vs fmad naming would be a significant source of confusion to me. I'd rather try to have the discriminating letter in the same position for all variants, e.g. ffma / fuma / ftma for fused / unfused / tentative (weak) variants respectively.
Anyway, the ffma vs fmad naming would be a significant source of confusion to me. I'd rather try to have the discriminating letter in the same position for all variants, e.g. ffma / fuma / ftma for fused / unfused / tentative (weak) variants respectively.
2
nostr:nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpqsld6cd53kj8pkz9svnnt54um57nvwrlx9sgya7xhvhl5a5a4f3sqdxhcxx this entire topic is such a disaster..
I'm adding two new opcodes to mesa/nir, becaus...
@nprofile1q... Godspeed! ffma is required, fmad becomes necessary as soon as you have hardware with unfused mul-add, and the third will match LLVM fmuladd intrinsic, afaict. So the plan sounds alright to me. Do you need suggestions for naming the third Schroedinger's FMA?
1
me popping into your chat: what's your fmad vs ffma situation today?
@nprofile1q... I'll be posting some fma patches for GCC in May, hopefully
1