HNHacker News
TopNewBestAskShowJobs

banachtarski

2,796 karma · joined July 16, 2012

programming for ~20 years and counting. math background. generalist with some emphasis on machine learning, graphics, physics simulation, compilers, and server technology. I dabble in electronics and FPGAs.
submissionscomments
banachtarski··on C-for-Metal: High Performance SIMD Programming on Intel GPUs
From an ergonomics point of view, it would be nice to have pointers directly into a descriptor heap, but this is coming already with SM 6.6 any day now. If I'm allowed to request hardware capabilities, I would like a non-divergent warp jmp instruction based on a function pointer table (much like is done with DXR-class raytracing hardware/shaders) to eliminate pipelining introduced only to mitigate shader permutation explosion.
banachtarski··on C-for-Metal: High Performance SIMD Programming on Intel GPUs
I'm not a hardware engineer, but I am a GPU-focused graphics engineer.

> C-style fine-level control over a GPU device as though it were a CPU.

Personally, I think this is a fool's errand, and this has nothing to do with my desire for job security or anything. When I look at how code in the ML world is written for a GPU for example, it's really easy to see why it's so slow. The CPU and GPU architectures are fundamentally different. Different pipelining architecture, scalar instead of vector, 32/64-wide instruction dispatches, etc. HLSL/GLSL and other such shader languages are perfectly "high level" with other needed intrinsics needed to perform relevant warp level barriers, wave broadcasts/ballots/queries, use LDS storage, execute device level barriers, etc. This isn't to say that high level shader language improvements aren't welcome, but that trying to emulate a CPU is an unfortunate goal.

banachtarski··on Class Action Lawsuit Filed Against Robinhood
Framed another way, can a broker prevent the purchase or sale of stock to bias the asset's movement in a particular direction to benefit themselves or a friendly third party?

Of course not. This is any number of SEC violations

banachtarski··on Class Action Lawsuit Filed Against Robinhood
You're not a lawyer, which is why you don't understand that ToS writing never protects you from actual criminality, which in this case appears to be a clear case of market manipulation (restricting trades to only sales and not buys? how much more blatant can you get).
banachtarski··on An interactive review of the Oklab perceptual color space
I think the issue is you've described something qualitatively without giving the needed quantitative description of how the non euclidean space should be specified. Implicit surface with some parameterization? Family of geodesics, selecting a curve with a parameter per curve? How do artists specify where in the color space they are? How is your coordinate mapping defined?
banachtarski··on Zig in 30 Minutes
epoll and io_uring (the latter of which did not exist when K&R was written) have nothing to do with C though
banachtarski··on Building a simple neural net in Java
Batch size is an orthogonal concept. The reason your memory increases so much is because with larger batches, you need to store more intermediate results before running your optimizer. This would be true with a pure CPU implementation of an SGD (stochastic gradient descent) optimizer as well.
banachtarski··on Building a simple neural net in Java
> Matrix multiplication is an embarrassingly parallel task.

(GPU programmer take). This is a common misconception. In reality, matrix multiplication is _extremely tricky_ to implement a fast parallel algorithm for because the number of useful instructions you can perform per byte is limited and you end up getting memory bound very quickly, so you start needing to think about properly employing LDS, managing bank conflicts, etc. For many applications that actually need to run fast, matrices are great for sorting out the theory, but then it's a good time to come up with something that will actually run fast after the paper problem is sorted out.

banachtarski··on A computer science study plan to become a software engineer
Computer science courses these days do in fact teach about memory coherency and MESI protocols with respect to CPU cache management and architecture. You'd learn about pipelining as well, and scalar architectures for that matter. I guess my point is that even if you didn't, if you wanted a head start, it's sufficient to know at least the basics of what microarchitectures are, and how to reason about a computational model of complexity, even if you need to adjust it later (for a particular console, or GPU, or architecture).

As for quaternions, I can derive them (no really, I've written a ton of code in this space, although I prefer the GA formulation), but I want to know, does the candidate understand why they're mathematically advantageous (blending properties, linearity, etc). Can they productively read a paper that uses it and apply the technique, etc.

banachtarski··on A computer science study plan to become a software engineer
Oh yea, calculus and linear algebra is all washed up and the industry has moved way past it. /s

"I don't think any of that is true" is a strong statement. Nobody is saying that a toy renderer in school resembles a AAA rendering engine, but the point is that the thought process needed is the same, and I'm still going to expect a certain degree of rigor from any candidate. Also, the parent post is speaking more about the theory and fundamentals in general, not the direct contents of any particular coursework. FWIW, I am not formally trained in graphics or CS (my training was math).

banachtarski··on A computer science study plan to become a software engineer
I think the claim is easily sustained with just empirical evidence. If you want to do a career in embedded, or graphics, or robotics, or maybe flight control, etc., you're going to need the math and the comp sci in droves. HN tends to have a crowd skewed towards certain disciplines, which is why every time the topic of a whiteboard interview shows up, most of the comments voice a strong opposition. In my line of work (graphics), the whiteboard is used all the time, both in and out of the interview.
banachtarski··on A Beginner's Garden of Chess Openings (2002)
I grew up playing chess competitively and used to make this argument all the time.

Since then however, I've relaxed my stance. In reality, endgame theory is in some sense, far more complicated than many opening systems, because their are so many degrees of freedom. Yes there are 32 pieces at the start of the game, but prior to development and initial piece placement, there are usually far fewer candidate moves. I think it's a good idea for a beginner to learn at least a few opening systems after getting the "gist" of the endgame like king opposition and opposite color bishops etc.

banachtarski··on Diet may help reduce cognitive decline
The problem is that the entire premise of the study is flawed, whether three questions were asked, or a hundred. Asking a hundred questions simply means it’s more likely your study will “succeed” in the sense of finding something to report that aligns with your agenda. Statistically, take any sample size. If you measure enough factors, you are bound to find some bias in your sample size.

Instead, a better study would consider actual science. Microbiology, biochemistry, etc and form measurable hypotheses instead of this spray-and-pray pseudoscience

banachtarski··on Cydia, the original app store, sues Apple on antitrust grounds
I’d rather have that than the current situation on mobile where because of the 30% cut, every game is littered with ads or pay to win tactics to increase their margins. The situation for games on desktop and console and handheld portable consoles is leaps and bounds better than mobile.
banachtarski··on Diet may help reduce cognitive decline
Ask participants about 24 things and you are bound to find correlations that have no actual basis in fact for that sample size. They might have well have asked about their favorite colors and music while they were at it. This is an obviously bad study funded by industries that are not impartial, and really has no business being posted here.
banachtarski··on Cydia, the original app store, sues Apple on antitrust grounds
Competition is key though. Plenty of publishers know that ditching established installers and app stores comes with a price, and so publishing or at least co-publishing on steam/epic/gog/etc is pretty important.
banachtarski··on Linus Torvalds' good taste argument for linked lists, explained
What you are espousing is known as “defensive programming” which is very much a bad thing (tm).
banachtarski··on Choose Boring Technology
I'm not going to throw out a wildly inaccurate number, but based on pure market capitalization, I'm willing to bet the majority of software driving the primary products sold by companies globally is not just CRUD apps.
banachtarski··on Is the preprocessor still needed in C++? (2017)
I should clarify. My comment is a statement about the ability for the compiler to reliably cache compiled object files in an incrementally compiling situation.
banachtarski··on Is the preprocessor still needed in C++? (2017)
It should absolutely pass the sniff test.

The cost isn't the cost of parsing. It's the cost of compiling something you've compiled before. If you change a header that is included in N translation units, you compile N translation units, even if you didn't fundamentally change the header contents in a way that would effect the final object files.

banachtarski··on Is the preprocessor still needed in C++? (2017)
This is the reason yes, and why macro encapsulation by modules is so important
banachtarski··on Is the preprocessor still needed in C++? (2017)
The preprocessor is still needed to implement a routine that allocates memory on the stack in a cross platform way.
banachtarski··on Is the preprocessor still needed in C++? (2017)
The reason I don't agree with this is because the preprocessor is the main killer of compiler throughput in a large project, preventing a number of optimizations that would otherwise be possible.
banachtarski··on Choose Boring Technology
Maybe in your universe
banachtarski··on Controversial Programming Opinions (2012)
My (slightly controversial) pet peeve is the abuse of the notion that "premature optimization is the root of all evil." Sure, let's not prematurely optimize, but let's also not use that as an excuse to make horrible decisions or unforced errors also.
banachtarski··on Show HN: NN-512 – Generate standalone C code for neural nets
AVX-512 causes just as much if not more throttling in my experience? In what case are you finding that AVX-512 runs at lower voltages than AVX2?
banachtarski··on Geometric Algebra (2012)
Right, as I said, composition is faster, not application.
banachtarski··on Raytracing in One Weekend (2016)
Normals are part of trigonometry.

Eigenvectors won't show up in an elementary path/ray tracer.

You need the most barebone understanding of LA for xform composition/application.

banachtarski··on Geometric Algebra (2012)
Quaternions are used in animation for blending and composing transforms for a reason. You don't have to trust me, but yes... tried for maybe ten years and counting from doing graphics and animation work.
banachtarski··on Geometric Algebra (2012)
It depends. For composing multiple transforms, composing versors aka quaternions/dual quaternions is cheaper than a full 4x4 multiply. For a single application, a 4x4 will be cheaper.
← PreviousPage 2 of 34Next →