Investigate WebAssembly integration

[mateogianolio]

WebAssembly has reached Browser Preview and offers high performance typed array manipulation. It also makes it possible run C code in the browser, which might be very interesting for this project. We should investigate in which ways we can interface with WASM once it's released for general users.

[mateogianolio]

https://github.com/mdn/webassembly-examples

[Khalid-Nowaf]

I'm using victorious in the browser now, it would be awesome if I could leverage wasm integration and use BLAS in the browser!

[mateogianolio]

https://github.com/maierfelix/glmw

[mateogianolio]

https://github.com/AssemblyScript/assemblyscript

[metabench]

To what extent is it worth working on our own lower level implementations, or using existing libraries such as blas and glmw?

AssemblyScript looks very useful. Maybe it would be worth me coding in TypeScript to try it.

Since we already have (or soon will have) Vectorious in TypeScript, AssemblyScript looks like a good route to greater performance. I'm interested in that, as well as porting some JS code to C++ or Rust.

[mateogianolio]

Check this out. I did a quick test with the the saxpy routine in TypeScript, WASM, C (Node.js addon) and C with system linked BLAS (nblas).

https://github.com/mateogianolio/native-wasm-c

[MaxGraey]

Unfortunately cost of calls wasm <-> host is pretty significant so using bench tools like benchmark.js is not good choice. See my notes: mateogianolio/native-wasm-c#1 (comment)

[MaxGraey]

Also pretty interesting results you could got if you slightly modify part of algorithm (and changes this for every versions as well) from:

for (i = m; i < n; i += 4) {
   y[i + 0] += alpha * x[i + 0];
   y[i + 1] += alpha * x[i + 1];
   y[i + 2] += alpha * x[i + 2];
   y[i + 3] += alpha * x[i + 3];
}

to

for (; i < n; i += 4) {
   y[i + 0] += alpha * x[i + 0];
   y[i + 1] += alpha * x[i + 1];
   y[i + 2] += alpha * x[i + 2];
   y[i + 3] += alpha * x[i + 3];
}

Results:

ts x 221,134,839 ops/sec ±0.21% (96 runs sampled)
empty x 57,229,521 ops/sec ±0.42% (92 runs sampled)
wasm x 15.26 ops/sec ±0.72% (42 runs sampled)
c x 85.46 ops/sec ±1.00% (72 runs sampled)
vectorious x 98.55 ops/sec ±2.22% (71 runs sampled)

[mateogianolio]

Moved to discussions.