A High-Level Lua API for GPU Parallelism | Hacker News Reader

A High-Level Lua API for GPU Parallelism

willcrichton.github.io

71 pointsdevy17 comments

While the project is interesting, the main benchmark is very misleading. They claim that terracuda matrix multiplication is faster than the one written in cuda. This can be only true if it is compared to NAIVE matrix multiplication in cuda, which can be confirmed by the time that is reported in seconds, which is orders of magnitude slower than the "real" cuda matrix multiplication.

Why is that benchmark completely meaningles? Because cuda matrix multiplication is so optimized that no one with a bit of sense would roll out his own, and even if he does, that implementation would have to use very specialized and ugly code - can this be done in terracuda, and is it simpler than in C?

Addidionally, the main issue in gpu kernels is not the ugliness of c itself, but the need to optimize algorithms to the hardware of the gpu, thus making the code much more verbose and ugly than the naive c code. That's why terracuda and the likes look fantastic when you compare them with naive C implementations, but they try to solve the wrong problem IMO.

Project author here. Your first claim is true, and you can verify it here: https://github.com/willcrichton/terracuda/blob/master/matrix...

However, I would dispute that it makes the benchmark meaningless. I agree that no one will write the hand-optimized matrix multiply routine in Terra, but I think that misunderstands the use case of the language. Terracuda is intended to make everyday GPGPU computations super easy, so you can speed up simple scripting code that would otherwise be order of magnitudes slower. Look at the other examples in the repo: if you wanted to write a simple hash function, Mandelbrot generator, or renderer, Terracuda makes it simple to accelerate Lua code.

There will always be room for super optimized kernels for common functions like matrix multiplication, but I believe there's a strong need for a language that handles everything else.

My point is that while it can sometimes accelerate lua code, it will also (i'd say) decelerate lua code, when the algorithm is not trivially parallelizable (and that is the majority of cases). So, to get the gpu boost, or evento understand how to implement the algorithm properly, I'd have to drop to the GPU specifics. The main question is then: do terracuda helps with that or will it be the obstacle? (it is not a rethorical question; while I suspect it will be the obstacle, I am interested to hear what is your experience with such cases)

I strongly doubt writing Terracuda GPU code would ever decelerate your original application, although it might not be as fast as a hand-written CUDA program (and that's an important distinction!). Terracuda doesn't actually lack that many GPU specifics--really it just doesn't provide access to shared memory constructs which are important for hand-tuned matrix mult kernels and blocked algorithms. And that's in the NVPTX API so it's not a true obstacle, just something I haven't gotten around to writing yet.

But have you tried? I've seen a good share of naive kernels that are slower on a monster gpu than on the cpu. My guess is that most of the code written by users that do not have a clue about the specifics of how gpus work (which is the typical person that terracuda might appeal to) would fall for that trap.

> although it might not be as fast as a hand-written CUDA program

If you're dealing with larger pieces of code at once (that is, if you have a bigger view of what the programmer wants to run, which you should have), one interesting thing you could do is "exploratory compiling". Try different compilation step sequences and transformations, if you have enough invariants in the code. Machine time is cheap and this could yield some interesting results. Of course, the more constrained the language model is, the more opportunities for this.

(There's been some interesting results in this area, like superoptimizers for local optimization, but larger scale compilation would probably use something like genetic algorithms, as you're unlikely to have the opportunity for exhaustive searches.)

> Because cuda matrix multiplication is so optimized that no one with a bit of sense would roll out his own

Really? What if you have, for example, special matrices, not general ones? What's going to be faster, generating a matrix and then running the fastest general matrix multiplication code money can buy on it, or, even a somewhat slower code, but one that can instead use some extra invariants?

> Addidionally, the main issue in gpu kernels is not the ugliness of c itself, but the need to optimize algorithms to the hardware of the gpu, thus making the code much more verbose and ugly than the naive c code.

But that sounds like a compiler issue to be solved. And then you don't have to do it by hand, you let the machine find it out. (And I can't help but think that C is terrible for array based programming anyway. Granted, pure Lua might not be great either for the same reasons.)

While these things are desirable, please point to: 1. a solution such as you describe that might be used today 2. how would a typical user of lua (a GPU non-expert) program those special matrices 3. even if he/she can code the algorithm, how can it be done from lua instead of C

This is great. Playing around with Terra for a few weeks now in my spare time. And it's a great language.

As a side note. I'm not sure this is usable in software other than Scientific Computing kernels, as relying solely on CUDA rules out AMD and Intel GPUs, if i'm not mistaken.

Though I guess one could use it as an inspiration for implementing something similar for OpenCL or Vulkan.

Awesome project. Terra seems so powerful, it's great to see more applications for it.

Sidenote -- I hate to be the guy complaining about web design, but the Raleway font is very hard to read (Retina MacBook). Here's a screenshot: http://i.imgur.com/9DkVEaQ.png

The screenshot is at 2x resolution so the readability problem is not quite obvious, but look at the vertical lines on the "m" letters -- the line width is all over the place even within a single letter. Raleway's miserable kerning doesn't help things either.

Totally valid complaint. I've updated it to something that's hopefully more readable.

Thank you! Much better now.

I love stuff like this, but I'm wondering whether it's actually useful for solving any real-world GPGPU problems.

My understanding is that the hardest part of moving stuff to the GPU is moving data around to and from the GPU without killing the computational performance gains by data dependencies, and to transform your problem to something that doesn't need a lot of control flow or rich/complex data structures. Does Terracuda address these things?

Good question. For your first point, Terracuda doesn't do anything particularly intelligent with movement between host/device memories--it's mostly there just to hide that away from you. However, empirically, if your baseline is Lua or Python, almost anything you write in CUDA is going to be more efficient so long as you have enough data to parallelize over.

As for the second point, that is still mostly left to the programmer. GPU code doesn't necessarily preclude complex data structures, however. For one of my benchmarks, we used a software renderer that generates a quadtree on the fly and accesses that during the shading phase (https://github.com/willcrichton/terracuda/blob/master/render...).

On typical consumer hardware you really want to batch up your processing in a way that gives you the best computation/data ratio to prevent performance being killed by data latencies, I don't think any kind of solution exists that can do that for you automatically/transparently. For the usual suspects like signal processing stuff and such, these problems are relatively easy to overcome because you typically have large amounts of data that can be processed at once or in in blocks.

What would make your project even more interesting is to see how it would perform on an heterogeneous system architecture (HSA), ie: what AMD is working towards. In an HSA, the CPU and GPU share the same memory and address space and a program can basically move processing between the two at close to no cost (no memory ever needs to be copied or mapped in/out GPU address space). For problems that allow some kind of pipelining to absorb start-up latencies caused by data dependencies, a combination of Lua + Terracuda would be pretty awesome, allowing you to move even small/short-running computational tasks to the GPU.

Newbie question: Would this Terracuda system be the easiest way to get started with GPU/CUDA programming for someone who hasn't done it before or are there easier approaches out there for running code in parallel on the GPU?

(I have worked with Lua for personal gamedev projects before)

Ideally this would be your easiest starting point for GPU programming, but I haven't this particular course project for a year and a half so it would need some polishing before it's ready for prime time. Honestly, I would just dive into the CUDA docs and mess around with their toolchain. Some resources:

Lecture on "GPU Architecture and CUDA Programming": http://15418.courses.cs.cmu.edu/spring2016/lecture/gpuarch

CUDA C Programming Guide: http://docs.nvidia.com/cuda/pdf/CUDA_C_Programming_Guide.pdf

Reply on news.ycombinator.com