Fork me on GitHub
#uncomplicate
<
2017-08-11
>
sophiago15:08:24

Hi. I have some questions about transitioning a project to use Neanderthal. Mainly, it seems like it doesn't currently support sparse vectors? That's a big deal for me because the library I'd like to use it for is all sparse using nested n-tuples where the first element of is the coefficient, i.e. `[[2 1 1] [3 1 0] [5 0 1] [7 0 0]]` == 2xy + 3x + 5y + 7. I have functions to convert to sparse form, but I mainly use that for Taylor series whereas with multivariate polynomials I'd have to generate matrices. But obviously my own arithmetic functions with vanilla Clojure vectors are becoming quite a bottleneck.

blueberry15:08:45

@sophiago You're right; it doesn't currently support sparse vectors.

sophiago15:08:41

Hmm. So I'd probably need to test whether I get a benefit from converting them back and forth. Multiplying vectors right now is quite slow for me.

blueberry15:08:23

Myltiplying vectors? You are refering to dot product?

blueberry15:08:45

Converting them would be even slower.

sophiago15:08:25

I'm just referring to multiplying two vectors to get a third.

blueberry15:08:40

Elementwise?

sophiago15:08:05

Do you have vector-wise addition, subtraction, and division as well? I only see `mm` and `scal` which could be helpful. I'm assuming the mutable versions are faster?

blueberry15:08:06

Yes. Not only that, but arbitrary primitive functions too. There are articles at neanderthal website that cover that in detail.

blueberry15:08:37

although when you do that, you do not really use linear algebra operations, but only use vectors/matrices as glorified primitive arrays. Might help you, but is a naive way of doing these things, and might not give you the speed that you might hope for.

sophiago15:08:58

Hmm, ok. My next question was going to be about speeding up scalar math (almost all longs, some ratios...not sure if that would be an issue vs. doubles) by overriding the functions with yours.

sophiago15:08:07

`magnitude` from your math library might also be useful to me, except I assume it only works on scalars? In that case I can at least use `sqrt` vs. Java math I guess.

blueberry15:08:27

I am not sure what exactly are you trying to do, but the best way is to try and see. I'm skeptical though.

sophiago15:08:04

I have pretty good documentation of where I'm at so far. It might be easier for you to take a glance than for me to keep explaining it: https://github.com/Sophia-Gold/Madhava-v2

blueberry15:08:10

If you want speed, you have to work with primitive floats/doubles and that's it...

sophiago15:08:13

Yup, I'm fine switching to doubles. They mostly only show up in integration and division anyway and it makes sense for performance.

sophiago15:08:57

But I'd like to see how much using your library is even possible for me now anyway. I wonder if I could hack on it to add sparse vectors like how mine are formatted and, if so, how difficult that would be?

blueberry16:08:51

Depends on the results you want. In my opinion, it is relatively easy (but not trivial) to create some implementation. It is completely different to create something that is competitive with best available tools.

blueberry16:08:46

But the best way is to try and see where you're at and then decide how to proceed...

sophiago16:08:49

Right. I have little experience with MKL or BLAS and the whole point is to have their raw speed.

blueberry16:08:24

Not only that. The point is also to have access to their algorithms, of which there are many.

blueberry16:08:47

Reimplementig that in (slow) Java/Clojure code is also a huge task.

sophiago16:08:31

I think that for me I wouldn't be using a ton of them. It's just basic arithmetic of multivariate polynomials encoded as sparse Clojure vectors is horribly slow.

blueberry16:08:57

Why don't you try this then: Use matrices, and use dense matrices (ge). If your data is not really, really sparse, that won't waste much space, but can give you decent speedup.

blueberry16:08:41

And, obviously, use operations on matrices, do not loop over collections of vectors.

sophiago16:08:53

I'd still need to convert back and forth because of the differentiation and vector calculus functions.

blueberry16:08:22

But those operations can also be done on matrices, right?

blueberry16:08:34

Why keeping all this in vectors at all?

blueberry16:08:47

Remember: data movement is typically slower than computation. Whatever you might gain with whatever cool method, you'll loose in these conversions.

sophiago16:08:36

I have to generate and transform many partials. Using dense matrices vs. sparse vectors makes the size grow exponentially. And they're still all functions I have to write by hand...and already fine tuned for sparse form.

sophiago16:08:00

I'm actually wondering if you think Vectorz makes sense considering how small most of my my vectors are quite small and it says it works with sparse form?

sophiago16:08:56

Part of what would factor into that decision is I'm unsure how Neanderthal works for scalar math, i.e. if I can just drop in the functions as substitutes for the Clojure ones in my existing code.

blueberry16:08:11

I don't know. From this, I don't see why you need any linear algebra library at all. They are useful when you can describe parts of your operations with linear algebra operations. If you can't, then you are using them as fancy arrays, which will bring more complexity for only small superficial benefit, if all...

blueberry16:08:59

if any, I wanted to say...

sophiago16:08:43

I don't need any linear algebra. I would like two things: 1. Faster scalar math in all my functions, 2. Fast arithmetic of sparse vectors as I have them encoded above.

blueberry16:08:57

Both of those things are fastest when called with plain old primitive Clojure functions. Neanderthal can't help you with that, and I doubt any other library can.

sophiago16:08:36

Damn, ok. Yeah I looked up Vectorz's sparse implementation and it really doesn't seem like what I want.

blueberry16:08:38

The point of Neanderthal is to avoid working with scalars.

blueberry16:08:30

and to avoid writing your loops

sophiago16:08:32

Well, mainly scalar operations inside vectors to gain a tiny bit of speed. But my primary concern is just stuff like multiplication and addition of vectors being a huge bottleneck right now.

blueberry16:08:26

The thing is that you need really large vectors to see the speedup over plain clojure functions.

sophiago16:08:32

Like everything I have listed under arithmetic in the README is extremely slow compared to everything else, but used in many of the vector calculus functions.

blueberry16:08:45

It seems your problem is in the algorithm itself, not he implementation

sophiago16:08:54

Well, working with polynomials is necessarily going to be slow...especially when you get to things like division. I assumed I could just farm it out to FORTRAN at this point, though.

blueberry16:08:32

It will be equally slow. Calling native functions with scalar arguments is not (much) faster than calling primitive Java functions.

blueberry16:08:54

since primitive Java methods are already very, very fast. A nanosecond or four, but not more. That's what you get in C.

sophiago16:08:20

Well, I'm not even using Java arrays right now. Maybe that would be somewhere to start.

sophiago16:08:25

It's hard to integrate with the rest of the code, though. I may want to consider just taking the bottleneck arithmetic functions and writing custom Java implementations of them.

sophiago16:08:21

Oddly, I don't get any speedup using `vector-of`.

blueberry16:08:02

It's not odd to me, and I could bet you won't see any improvement from writing arithmetic functions in Java. Clojure functions are already at the speed of Java functions if written properly. For example, Neanderthal is completely implemented in Clojure...

blueberry16:08:38

The issue in your code is, it seems to me, completely unrelated to the specific language and platform, but in choice of datastructures and memory access - thus fundamentaly algorithms themselves.

sophiago16:08:01

I think I could probably refactor the arithmetic functions to use transients and possibly find some optimizations otherwise. It seems like what I want are the custom tuple implementations Zach Tellman wrote years ago, but the problem is they require protocols for everything so I think that's why they were abandoned.

sophiago16:08:40

Anyway, I did assume I could find a native library that would be much faster so didn't tune everything in the arithmetic.clj file much. I should revisit that. But to be clear, are you saying I should reconsider using nested Clojure vectors? Because I don't see much of an alternative other than custom tuples or something from Java, maybe LinkedLists.

blueberry17:08:24

All these things are quite slow compared to typical high-speed computation code. If you are chasing that kind of speed, none of those will help you much. I would hit the papers dealing with the implementations of those things you need, to see what is possible and what people use first.

blueberry17:08:31

Bottom line: no library in this area can help you if you do not know specifically what you want to do.

sophiago17:08:52

I do know exactly what I want to do. I assumed sparse BLAS level one included these functions, but now that I'm looking it seems I was wrong. Apache Commons Math does, but I'm not sure it's worth switching to vs. just optimizing my Clojure code further.

blueberry17:08:51

They may contain those functions, but it is questionable whether they will be faster than the stuff you'd write with primitive Clojure functions. That is especially true with sparse vectors. They are applicable when your data is really sparse. And, Level 1 functions are more convenience functions to support working with more demanding stuff, that they can speed up code.

sophiago17:08:33

I mean I did just look...level one really doesn't have as much in it as I assumed. I think the smartest thing is to just focus on refactoring my Clojure functions as you suggested. They were essentially written in one pass without the thought I put to the rest of the library.

sophiago17:08:45

Anyway, thanks for your input ðŸ™‚

blueberry18:08:23

you're welcome. I hope that would be of some help.