Implementing typed arrays

[mateogianolio]

Should we use a fixed array type or give the user the ability to choose (see https://developer.mozilla.org/en-US/docs/Web/JavaScript/Typed_arrays)?

I'm thinking the best would be to use Float64Array, which is the equivalent of a C double.

Like this:

javascript
var buffer = new ArrayBuffer(length * 8);
var values = new Float64Array(buffer);

[metabench]

I think allowing the user to choose the type would be best. Hopefully some code could allow for this to work without code repetition and at a minimal performance cost.

[metabench]

We should implement this first on Vector. Matrix uses Vector a lot, so Matrix will get sped up.

I'm not that sure how Typed Arrays would be used to optimize Matrix but it will become clearer as time goes on. Perhaps we should just use unsigned integer Typed Arrays as indexes to the vectors.

[mateogianolio]

My latest commits provide a smooth transition to typed arrays. All methods are still working as they should and they have become significantly faster, which can be seen by comparing benchmark results between branches master and opt-js.

[mateogianolio]

I also created an opt-c branch with implementations and benchmarks for vector addition and subtraction in C++, also showing very promising increases in performance.

If you want to test it:

$ git checkout opt-c
$ node-gyp configure build
$ node src/benchmarks/add.js
$ node src/benchmarks/subtract.js

[metabench]

Oh that's very nice indeed. Any idea on the % performance increases?

[metabench]

I noticed it was a small change to Matrix.

I was thinking about optimizing Matrix, and I think that a matrix processor would run fastest if it had a typed array that it used for all it's data. However, what I'm thinking of is a substantially different architecture. I was thinking that doing it that way would be a different implementation, and Vector could extend Matrix as a single-dimensional case.

[mateogianolio]

I ran your benchmarks (once) with the following results and with the default type of Float64Array:

opt-js

• add - 43 ms
• constructor - 35 ms
• equals - 18 ms
• ones - 32 ms
• scale - 39 ms
• subtract - 34 ms
• zeros - 31 ms

master

• add - 55 ms
• constructor - 26 ms
• equals - 11 ms
• ones - 167 ms
• scale - 50 ms
• subtract - 53 ms
• zeros - 159 ms

delta

• add -12 ms
• constructor +9 ms
• equals +7 ms
• ones -135 ms
• scale -11 ms
• subtract -19 ms
• zeros -128 ms

Most notable improvements in Vector.ones and Vector.zeros.

We should probably start comparing performance with other libraries like Sylvester for example.