SIMD

Single instruction, multiple data (SIMD) is a form of “data parallelism,” as contrasted to “task parallelism” with Web Workers, because the emphasis is not really on program logic chunks being parallelized, but rather multiple bits of data being processed in parallel.

With SIMD, threads don’t provide the parallelism. Instead, modern CPUs provide SIMD capability with “vectors” of numbers — think: type specialized arrays — as well as instructions that can operate in parallel across all the numbers; these are low-level operations leveraging instruction-level parallelism.

The effort to expose SIMD capability to JavaScript is primarily spearheaded by Intel (https://01.org/node/1495), namely by Mohammad Haghighat (at the time of this writing), in cooperation with Firefox and Chrome teams. SIMD is on an early standards track with a good chance of making it into a future revision of JavaScript, likely in the ES7 timeframe.

SIMD JavaScript proposes to expose short vector types and APIs to JS code, which on those SIMD-enabled systems would map the operations directly through to the CPU equivalents, with fallback to non-parallelized operation “shims” on non-SIMD systems.

The performance benefits for data-intensive applications (signal analysis, matrix operations on graphics, etc.) with such parallel math processing are quite obvious!

Early proposal forms of the SIMD API at the time of this writing look like this:

var v1 = SIMD.float32x4( 3.14159, 21.0, 32.3, 55.55 );
var v2 = SIMD.float32x4( 2.1, 3.2, 4.3, 5.4 );
var v3 = SIMD.int32x4( 10, 101, 1001, 10001 );
var v4 = SIMD.int32x4( 10, 20, 30, 40 );
SIMD.float32x4.mul( v1, v2 );    // [ 6.597339, 67.2, 138.89, 299.97 ]
SIMD.int32x4.add( v3, v4 );        // [ 20, 121, 1031, 10041 ]

Shown here are two different vector data types, 32-bit floating-point numbers and 32-bit integer numbers. You can see that these vectors are sized exactly to four 32-bit elements, as this matches the SIMD vector sizes (128-bit) available in most modern CPUs. It’s also possible we may see an x8 (or larger!) version of these APIs in the future.

Besides mul() and add(), many other operations are likely to be included, such as sub(), div(), abs(), neg(), sqrt(), reciprocal(), reciprocalSqrt() (arithmetic), shuffle() (rearrange vector elements), and(), or(), xor(), not() (logical), equal(), greaterThan(), lessThan() (comparison), shiftLeft(), shiftRightLogical(), shiftRightArithmetic() (shifts), fromFloat32x4(), and fromInt32x4() (conversions).

Note: There’s an official “prollyfill” (hopeful, expectant, future-leaning polyfill) for the SIMD functionality available (https://github.com/johnmccutchan/ecmascript_simd), which illustrates a lot more of the planned SIMD capability than we’ve illustrated in this section.