Web assembly?

[Quafadas]

Are any parts of the library compiled to web assembly? I was thinking in particular of the blas module…

I could find no reference in the docs, so I’m just checking I’m not missing a trick here…

[kgryte]

@Quafadas We do support compiling to WebAssembly, but we are not currently shipping WebAssembly binaries. What is your use case and what BLAS functionality do you need?

[Quafadas]

It’s a curiosity now that anything else . I have an application which does some basic linear algebra (arrays only) and I was thinking about converting to wasm to see if there was a runtime benefit.

As things stand I think I have only level 1 functions in use. I wouldn’t want to absorb a lot of time.

Is that enough of a picture or should I provide more detail?

[Quafadas]

This is exactly the sorts of issues I want to explore. Hmmm.

Currently, I don’t have an example I’m in a position to share. Which isn’t a great response - I’m sort of planning the experiment at the moment. I’ll go quiet for a couple of weeks - if it’s okay, I’m going to @ you when i have something concrete… bear with me!

[kgryte]

This is a bit dated at this point, but you may be interested in https://www.youtube.com/watch?v=iJL59lh4IJA&list=PLfMzBWSH11xa-iNnQG2555lgi4574nZOh&index=37.

[kgryte]

Unrelated to this discussion, I just did some performance comparisons of JS to other languages. For the benchmarks, I computed the sum using built-in language features (e.g., for Python, no NumPy; for C, no SIMD; etc). Here are the results, where the y-axis (Rate) is normalized relative to C and where languages are grouped along the x-axis according to array length. In short, using vanilla JS, V8, at least, can achieve near native performance once array sizes increase in size. All this to say, I'll be curious what you find during your WebAssembly experiments.

[Quafadas]

Gosh! That’s a cool picture. If that picture generalises, then I’m wasting my time!

My arrays are more or less Float64Array of size 10,000. I guess I’m wondering about the extent that sum could be aggressively optimised as it’s a common operation. Only way would be to measure my workflow I guess.

Could things like simd potential inn wasm make a difference? Although presumably, that’s more custom work?

To me, this seems to would imply that only parrallism , perhaps webworkers could help. And that is not to do with wasm or stdlib. Hmmmm.

[kgryte]

@Quafadas Quick follow-up. Here is a benchmark with the BLAS level 1 routine daxpy. For the Wasm benchmark, computation is repeated on the same vectors allocated on the WebAssembly module heap. For the Wasm (copy) benchmark, data is copied into and out of the WebAssembly module heap (e.g., as might be the case when scripting or working with data coming from dynamic sources).

As can be seen from the chart, unless the FLOPS of the routine warrants the cost of marshalling data to and from Wasm memory, one can achieve equivalent, if not better, performance in just JavaScript.

...and this doesn't even address the increased complexity associated with memory fragmentation and unbounded growth which may occur when using Wasm modules.

Depends on your use case, but getting optimal Wasm performance can be a bit tricky to realize. I ran similar benchmarks on dgemm and there it was worth the perf penalty of marshalling data to and from the heap, as matmul is O(N^3) and copying data is O(N^2), where N corresponds to the number of elements along each dimension of a square matrix.