This project is part of the @thi.ng/umbrella monorepo.
WebAssembly SIMD vector operations for array/batch processing, written in AssemblyScript. These functions use the CPU's vector instructions to process 128bit words at once, which is the equivalent width of a 4D vector with 4x 32bit components. Several of the provided functions can also be used to process 2D vectors.
See /assembly for sources:
abs4_f32add4_f32addn4_f32clamp4_f32clampn4_f32div4_f32divn4_f32dot2_f32_aos(2)dot4_f32_aosdot4_f32_soainvsqrt4_f32madd4_f32maddn4_f32mag2_f32_aosmag4_f32_aosmagsq2_f32_aosmagsq4_f32_aosmax4_f32min4_f32mix4_f32mixn4_f32msub4_f32msubn4_f32mul4_f32muln4_f32mul_m22v2_aos(2)mul_m23v2_aos(2)mul_m44v4_aosneg4_f32normalize2_f32_aos(2)normalize4_f32_aossqrt4_f32sub4_f32subn4_f32sum4_f32swizzle4_32(f32 and u32)
(2) 2x vec2 per iteration
Also see src/api.ts for documentation about the exposed TS/JS API...
ALPHA - bleeding edge / work-in-progress
Search or submit any issues for this package
The WebAssembly SIMD spec is still WIP and (at the time of writing) only partially implemented and hidden behind feature flags. Currently only fully tested (& testable for me) on Node 14.6+.
- SIMD implementation status
- Node (v12.10+):
node --experimental-wasm-simd - Chrome: Enable SIMD support via chrome://flags
Due to the opcode renumbering of SIMD operations proposed in April 2020, the WASM module will only work on engines released after 2020-05-21 when that change was committed to the WASM spec. For NodeJS this means only v14.6.0 or newer will be supported. This was an external change and outside our control...
- @thi.ng/malloc - ArrayBuffer based malloc() impl for hybrid JS/WASM use cases, based on thi.ng/tinyalloc
- @thi.ng/soa - SOA & AOS memory mapped structured views with optional & extensible serialization
- @thi.ng/vectors - Optimized 2d/3d/4d and arbitrary length vector operations
- @thi.ng/vector-pools - Data structures for managing & working with strided, memory mapped vectors
yarn add @thi.ng/simdES module import:
<script type="module" src="https://cdn.skypack.dev/@thi.ng/simd"></script>For Node.js REPL:
# with flag only for < v16
node --experimental-repl-await
> const simd = await import("@thi.ng/simd");
Package sizes (gzipped, pre-treeshake): ESM: 2.44 KB
import { init } from "@thi.ng/simd";
// the WASM module doesn't specify any own memory and it must be provided by user
// the returned object contains all available vector functions & memory views
// (an error will be thrown if WASM isn't available or SIMD unsupported)
const simd = init(new WebAssembly.Memory({ initial: 1 }));
// input data: 3x vec4 buffers
const a = simd.f32.subarray(0, 4);
const b = simd.f32.subarray(4, 16);
const out = simd.f32.subarray(16, 18);
a.set([1, 2, 3, 4])
b.set([10, 20, 30, 40, 40, 30, 20, 10]);
// compute dot products: dot(A[i], B[i])
// by using 0 as stride for A, all dot products are using the same vec
simd.dot4_f32_aos(
out.byteOffset, // output addr / pointer
a.byteOffset, // vector A addr
b.byteOffset, // vector B addr
2, // number of vectors to process
1, // output stride (floats)
0, // A stride (floats)
4 // B stride (floats)
);
// results for [dot(a0, b0), dot(a0, b1)]
out
// [300, 200]
// mat4 * vec4 matrix-vector multiplies
const mat = simd.f32.subarray(0, 16);
const points = simd.f32.subarray(16, 24);
// mat4 (col major)
mat.set([
10, 0, 0, 0,
0, 20, 0, 0,
0, 0, 30, 0,
100, 200, 300, 1
]);
// vec4 array
points.set([
1, 2, 3, 1,
4, 5, 6, 1,
]);
simd.mul_m44v4_aos(
points.byteOffset, // output addr / pointer
mat.byteOffset, // mat4 addr
points.byteOffset, // vec4 addr
2, // number of vectors to process
4, // output stride (float)
4 // vec stride (float)
);
// transformed points
points
// [110, 240, 390, 1, 140, 300, 480, 1]Karsten Schmidt
If this project contributes to an academic publication, please cite it as:
@misc{thing-simd,
title = "@thi.ng/simd",
author = "Karsten Schmidt",
note = "https://thi.ng/simd",
year = 2019
}© 2019 - 2021 Karsten Schmidt // Apache Software License 2.0