mratsim

40 ms·

391.
by mratsim 6y ago
What's the threat model? I've looked into the code. It uses Nettle's Curve25519 which seems to be implemented constant-time (assuming ECC mul is using "ecc-mul-g" from which is protected against cache timing attacks
392.
GitHub released their new notification UI to everyone(github.com)
2 points by mratsim 7y ago|0 comments
393.
by mratsim 7y ago
Sure, you might be interested in the research in loop nest scheduling I gathered here: https://github.com/numforge/laser/blob/master/research/autom... The most promising here being the Halide and Ti
394.
by mratsim 7y ago
LLVM properly generates ADC code when you use __addcarry_u64 contrary to GCC[1] which generates really dumb code with setc/mov in and out of carry. However __addcarryx_u64 which is supposed to generate ADCX and ADOX is not properly han
395.
by mratsim 7y ago
Constant-time division is super slow for sure but it's not needed for cryptography. Modulo is also similarly slow but rarely needed as well. For example in elliptic curve cryptography it's only needed at the very beginning of a co
396.
by mratsim 7y ago
I have an in-depth look on constant-time and the security implications of not being constant-time here: https://github.com/mratsim/constantine/wiki/Constant-time-ar...
397.
by mratsim 7y ago
Often that hints boils down to use assembly though. Constant-time is a pain to achieve and new compiler optimizations might make your code not constant-time anymore: https://www.cl.cam.ac.uk/~rja14/Papers/whatyouc.
398.
by mratsim 7y ago
Obviously you didn't read what I posted. The Quantum Transport is compute-intensive physics problem that has a lot of optimization research going behind it. One of the main bottleneck to solve this problem is Strided Matrix Multiplicat
399.
by mratsim 7y ago
Tree Style Tabs?
400.
by mratsim 7y ago
Cryptography. Want to do finite field computation on a 254-bit integer? Now you can (BN254, very popular for zero-knowledge proofs) 381-bit? you're covered. It's very perf critical and the field modulus bitsize and values are know
401.
by mratsim 7y ago
I mentioned theory and experts, you said enough theory. I switched to practical applications and walk the talk showing my code, and then you back off and want to go back to opinions. I see now that you want to back myself with experts since
402.
by mratsim 7y ago
It might be Internet and the issue of communicating emotions across but you sound quite taken by this issue. Anyway, I stand by what I say and I'm backed by my high performance code: - Writing matrix multiplication that is as fast as A
403.
by mratsim 7y ago
I did have one, but I was working for an American company.
404.
by mratsim 7y ago
I'm not playing word games. Template metaprogramming and constexpr doesn't help being faster in HPC or GPGPU, it helps reduce the redundancy of your code, for example if you want a generic algorithm on float, double, int, complex.
405.
by mratsim 7y ago
I didn't say C is faster than C++, I said C++ is not faster than C and so speed is probably not the main driver behind implementing a C++-->Assembly pipeline without intermediate C.
406.
by mratsim 7y ago
In embedded development C++ is often frowned upon because in the past people used C++ features that bloated the code. Regarding features that may not be properly optimized, there are exceptions and move semantics I guess, but exceptions are
407.
by mratsim 7y ago
I'm curious about the NHWC layout they mentioned. AFAIK CuDNN always had optimizations for NCHW and that was one of Tensorflow speed issue when they choose to default on NHWC, plus the related issues on writing transformation pipelines
408.
by mratsim 7y ago
Cryptographic software is at the very least audited, at the very best formally verified (i.e. proven bug-free). And now you can use proof assistants that are used to prove mathematical formulas to generate crypto code: https://gi
409.
by mratsim 7y ago
But books had exemptions and you could export crypto as long as it was printed on a book
410.
by mratsim 7y ago
Indeed. Time-to-market is an obvious one and esoteric platform support, and maybe also debuggability.
411.
by mratsim 7y ago
I would be surprised if optimization was the actual goal. C++ is not faster than C.
412.
by mratsim 7y ago
LLVM has a library and modular approach which makes it easier for people to contribute just in their area of expertise instead of having to find their way in the hundred of thousands of line of GCC.
413.
by mratsim 7y ago
I'm pretty sure it's the (lack of) package managers. The presence of a defacto package manager completely changes the library dynamics. I can't even link and test properly against GMP on Windows :/
414.
by mratsim 7y ago
I'm not sure what's your domain but for machine learning that will be slow: - Parallel matrix multiplication may be embarassingly parallel, you have a reduction step that is not trivial to parallelize across processes. Also you ne
415.
by mratsim 7y ago
Can Julia build static binaries now? Because for mobile devices you don't want to ship LLVM. Plus it's easier to package binaries than scripts which usually require Docker to deal with dependencies in a sane way.
416.
by mratsim 7y ago
That's not infantilizing. Asian countries have a very ingrained sense of "face" and honor. Face in China and honor in Japan are very very important.
417.
by mratsim 7y ago
And they can't even train in their country
418.
by mratsim 7y ago
And also the most probable, especially given the diet in some of the regions. That said, some countries in Africa also eat bats and pangolin and Ebola came.
419.
by mratsim 7y ago
I would have liked (2018) in the title since it's an article from March 2018. At first glance I interpreted that it was directly tested on the current strain.
420.
by mratsim 7y ago
Do you have a benchmark that measures this cost? I'm pretty sure the cost of checking this in a static language where the tag enums range is known at compile-time is: - 100x cheaper than cache-miss on a memory access - and 10000x cheap

More ›