liuliu

56 ms·

721.
by liuliu 6y ago
I am really curious to see people's take on Kotlin Native's compiler. From what I can see with Swift compiler development, there are tons of bugs when compiling down a high-level language. Of course, part of these Swift problems c
722.
by liuliu 6y ago
JNI is so expensive when cross the language barrier, you have to be conscious to constantly minimize your API exposure. This may be helpful in long-term, but not helpful for onboarding purposes. On the other hand, if you can lower Kotlin to
723.
by liuliu 6y ago
It is hard. Without a common background (raised with similar upbringing, same religions), your implicit rules can diverge greatly with the others. Not to mention born in different era will play a big role in this. Some people simply don&#x2
724.
by liuliu 6y ago
Would this allow people to host out-of-tree LLVM-based project easier? Last time I checked (shamefully, probably 2 or 3 years ago), LLVM still heavily prefer in-tree projects (creating a sub-directory and CMakeList.txt file) for source-code
725.
by liuliu 6y ago
For cc_library, I think now you can add include paths so `#include "somethingUnderYourPath"` will be easier.
726.
by liuliu 6y ago
I am trying Swift in the hope of integrating developer speed while with not too terrible performance. Julia would be another interesting choice in this space. Of course, as you get to nanoseconds, any language choice is irrelevant.
727.
by liuliu 6y ago
It is arguable whether wealth should be within right to privacy or not. As for items inheritably with illiquid market value, it is also arguable whether we should assess their ongoing market value at all. Real-estate on the other hand, woul
728.
by liuliu 6y ago
Starlark is OK (it is very similar to Python except removed non-deterministic part). But the part where no type-hinting kicks in is when you try to read the actual underlying macros / functions the others provided. It is much harder to
729.
by liuliu 6y ago
Latency fundamentally limited by the model processing single frame, all-in-all, probably somewhere around 10 to 15ms depends on your input size (assuming VGA type of input). This is a great article talks about system-engineering for the vis
730.
by liuliu 6y ago
Agreed. But doing SIMD optimizations could quickly become 2-3 weeks project for me just for one platform. Adding NEON would be another 1 or 2 weeks. Need much more dedication I guess. Your work on simdjson is awesome! Thanks for saving the
731.
by liuliu 6y ago
> how else would you multithread this? Doing two pass parsing: https://liuliu.me/eyes/loading-csv-file-at-the-speed-limit-o...
732.
by liuliu 6y ago
For 1), it is simple, but not trivial to parallelize csv parsing correctly, hence the yesterday’s post. Otherwise, agree on 2 and 3. Object wrapping, number decoding can be slow on the whole “file to dataframe object” path, but trivially pa
733.
by liuliu 6y ago
Yes, probably. With hindsight, it is probably a mistake to use mmap. I probably can do better to just read file myself, since I have to make a mirror buffer later for some data manipulation anyway.
734.
by liuliu 6y ago
Re csv2 code: if you look at the benchmark, they really just loop through each cell, without accessing the content: https://github.com/p-ranav/csv2/blob/master/benchmark/main.c... Thus, the code tri
735.
by liuliu 6y ago
Just updated https://github.com/liuliu/dflat/ with Apple's atomics. Seems to be just some naming changes from glessard ones (when doing the change), otherwise a pretty smooth transition.
736.
by liuliu 6y ago
This is great news. So far I have been using https://github.com/glessard/swift-atomics for projects. It sort of works, have a nice API, but without SE-0282, it is always "try with your own risks" kind of thin
737.
by liuliu 6y ago
The training data is around 40GiB: https://d4mucfpksywv.cloudfront.net/better-language-models/l... Consider the compression ratio for such text, and you need 2-word / 3-word to 1000-word dictionary, and at 1000-wo
738.
by liuliu 6y ago
You need to disable inlining. __builtin_frame_address can give surprising addresses when inline enabled. The code itself is likely not be wrong, but you won't confined to the lexical scope as if you see in the code (the defer can be tr
739.
by liuliu 6y ago
What about drag gesture? I.e. if you want to enable a particular widgets such as image display / text box to be draggable? How the event should be passed in and handled? These gestures seems not internal to the widgets, but also mutate
740.
by liuliu 6y ago
Seems to be good speedup overall relative to 2080 Ti (including FP16: see relatives 2080 Ti v Titan: https://lambdalabs.com/blog/2080-ti-deep-learning-benchmarks... ). This suggests we should see another Titan card that
741.
by liuliu 6y ago
I don't have access to see whether they have fine-tuning API. Do you have any links explain the said fine-tuning? It is certainly surprising given there is no fine-tuning experiment mentioned in the GPT-3 paper. Weights loading is emba
742.
by liuliu 6y ago
Yes, that would work. But that would be a departure from existing imgui paradigm where the output of imgui library is a 2d bitmap. At that point, it gets really close to the vDOM territory ... As for the stable identities, yes, it is import
743.
by liuliu 6y ago
You need tiny bit of memory for activations if you don't want fine-tuning. I think for GPT-3, fine-tuning is out of window. But it is reasonable to expect inference takes less than a minute with single 3090 and fast enough SSD.
744.
by liuliu 6y ago
It is a long post with a lot of good points. One minor thing I would argue: the "setter support" in languages is not really a plus for observer patterns. In many such languages support "setter" and use such to naively im
745.
by liuliu 6y ago
And current crop of imgui libraries already maintaining states cross updates (after all, that is how you can implement things as simple as drag & drop). However, the reason you need to have a retained render object tree somewhere for ac
746.
by liuliu 6y ago
Not really. But I did this: https://liuliu.me/eyes/write-cross-platform-gui-in-swift-lik... Unlike most comments here, I felt that Swift was a Python replacement to me.
747.
by liuliu 6y ago
The number on RTX Titan for fp16 is actually a bit foreign to me (3.5x faster than fp32). Not sure what's going on (maybe they tweaked batch size to fit more in GPU memory?). Need to wait more comprehensive ones. (The older numbers loo
748.
by liuliu 6y ago
I think that I understand how most AC_* macros I use internally works. Cannot claim for AX_* ones. I do write every line of configure.ac myself though :D
749.
by liuliu 6y ago
One thing I am a bit upset is how Apple approaches SwiftUI / Combine, as these two to me, are foundational pieces to the Swift language ecosystem. Particularly: 1. There is no way to bundle SwiftUI / Combine to the app so you can
750.
by liuliu 6y ago
Yes. I have an alternative DL framework to play whatever I want with :)

More ›