danieldk

25 ms·

241.
by danieldk 6y ago
Also error-chain , which is more or less replaced by thiserror .
242.
by danieldk 6y ago
I hope that std::prelude::Result becomes anyhow::Result, including the default generic argument (i.e. Result<u32> just works But that's basically giving up on informative library-specific error types. It may a valid solution, b
243.
by danieldk 6y ago
proper foundation I got quite nervous as the plane appears to be resting on a pile of wood pellets.
244.
by danieldk 6y ago
"Windows programs for Linux"? Whoops, typo. Windows programs from Linux , I'll fix that. You can mount other host filesystems from inside the VM in any hypervisor I know of. In most hypervisors you can't do a no
245.
by danieldk 6y ago
But the filesystem integration is much better than a VM in macOS, being able to run Window programs from Linux, etc. Also, Windows has decent X11 servers that work well with retina displays, which can't be said of XQuartz.
246.
by danieldk 6y ago
This sounds promising, something closer to OSX would give apple competition for devs like myself that want a linux based system and a fluffy "it-just-works" gui I switched from macOS to Linux. GNOME and others are pretty good t
247.
by danieldk 6y ago
SiS also had 3D graphics chips. Tseng Labs was probably one of the earliest victims of the 3D GPUs wars, having developed but never completed the ET6300. PowerVR still exists and is seems successful in the mobile market (having provided App
248.
by danieldk 6y ago
This is probably anecdotal, but I can recall a huge amount of my collection pretty much instantly when someone asks me which records I have by artist X, or which records did I buy in 2019. I used to be the same, but then with CDs. I could
249.
by danieldk 6y ago
It is a tiresome process. I attempted to submit a particular package (and others before and after me). It was either met with a lot of red tape (IIRC document licenses for each file individually) and months of no visible progress when comme
250.
by danieldk 6y ago
Some Debian developers tried to get funding to finish the Etch release, but it was met with a lot of hostility: https://lists.debian.org/debian-project/2006/10/msg00260.htm... That doesn't mean that it c
251.
by danieldk 6y ago
You need an app? In The Netherlands they use whatever emergency warning system is built-in in iOS and Android. https://support.apple.com/en-us/HT202743
252.
by danieldk 6y ago
This is total nonsense. E.g. in The Netherlands we have had tests every first Monday of the month, since I can't remember how long (I am 37 now), with the exception of 1997-2003 (when it was done once a year).
253.
by danieldk 6y ago
De Telegraaf and het AD alone have lost about 800k subscriptions in the last two decades. The curves for the other newspapers are far less profound and some of them do really well in digital subscriptions (NRC, Volkskrant, FD). Would be i
254.
by danieldk 6y ago
Digital subscriptions of Dutch newspapers have grown steadily: https://www.svdj.nl/de-stand-van-de-nieuwsmedia/digitale-opl... (For non-native speakers, scroll to the first graph.)
255.
by danieldk 6y ago
Wait, what? The largest newspaper (Telegraaf) is definitely right wing. The second largest newspaper (AD) is also anything but left wing. I agree with public broadcasting, but that’s in part because the right wing has changed. Traditionally
256.
by danieldk 6y ago
Well, there is nothing that says you can't have proper dependency tracking just because something is statically linked. Didn't say so. It is just easier with dynamic linking, because you can see what libraries (and versions) a b
257.
by danieldk 6y ago
Dynamic linking doesn’t work unless you can live inside a distro maintainer’s special bubble for all your software. If you can exist in that bubble, great—I really like Debian for certain use-cases—but if you can't, the benefits of dy
258.
by danieldk 6y ago
I don't disagree that another language/runtime could reduce the cost. But why do you value developer time at 0 USD? Unless this is a hobby, the hourly rate of a developer is higher than that monthly Digital Ocean bill.
259.
by danieldk 6y ago
Also, pricing of Digital Ocean seem to be quite steep for larger machines. E.g., Hetzner offers AMD Ryzen 3600 dedicated servers with 64GB RAM and 2x 512GB SSD for 42.16 per month including VAT. Or if you really insist on a VPS, you can get
260.
by danieldk 6y ago
I benchmarked some of my large transformer networks the last few days and MKL is still 50% faster than OpenBLAS. What's even worse in real-world applications is that OpenBLAS misbehaves when an application uses threads. This is also de
261.
by danieldk 6y ago
It's early days. Maybe they are still working on tuning the kernels?
262.
by danieldk 6y ago
Indeed, I had three theories, from optimistic to cynical: 1. They are sharing increasingly much code between oneDNN (formerly MKL-DNN) and Intel MKL. So, it simply pays of from a development cost perspective. 2. Intel sees AArch64 as a seri
263.
by danieldk 6y ago
I do not agree with this. I work heavily with NLP models for production in the Legal domain (where my baseline is where a 8GB 1080 must predict more than 1000 words/sec). What kind of network are you using? I can do near-SoTA multi-t
264.
by danieldk 6y ago
It seems I have found the issue. We were both right. MKL now uses a Zen-optimized kernel for dgemm, but not (yet?) for sgemm. More details: https://github.com/pytorch/builder/issues/504
265.
by danieldk 6y ago
Found the discrepancy. I use single precision in PyTorch. When I benchmark sgemm, the SSE code path is selected. Conclusion: MKL detects Zen now, but currently only implements a Zen code path for dgemm and not for sgemm. To get good perform
266.
by danieldk 6y ago
The plot thickens. As I reported elsewhere in the thread, the slow code paths were selected on my machine, unless I override the mkl_serv_intel_cpu_true function to always return true. However, this was with PyTorch. I have now also compi
267.
by danieldk 6y ago
By the way, if you want make this permanent in a binary, there is no need to set LD_PRELOAD all the time. You could just add a DT_NEEDED entry to the dynamic section. E.g. something like: patchelf --add-needed libfakeintel.so your
268.
by danieldk 6y ago
Indeed works. A simple trace reveals that the function is called mkl_serv_intel_cpu_true() . Make a file with the following content: int mkl_serv_intel_cpu_true() { return 1; } Compile gcc -shared -o libfake.so fak
269.
by danieldk 6y ago
That's definitely possible (it probably checks that the manufacturer ID is GenuineIntel ), but nobody wants to distribute patched MKL versions, because it most likely violates the MKL license. It may even be easier to replace the func
270.
by danieldk 6y ago
I think MKL actually fixed Zen performance. That is, the workaround no longer makes any difference because it is no longer needed. Odd. I am trying on my 3700X and it is definitely not using AVX, FMA or AVX2 code paths. Intel MKL 2020 upd

More ›