philips_supernote-obsidian-plugin/node_modules/ml-convolution

convolution

Convolution using the FFT or standard algorithm.

Installation

$ npm install --save ml-convolution

Usage

import {directConvolution} from 'ml-convolution';

const input = [0, 1, 2, 3];
const kernel = [-1, 1, -1];

const output = directConvolution(input, kernel);
// [0, -1, -1, -2, 1, -3]

API Documentation

Benchmark

With small kernels, direct convolution is usually faster:
Current results suggest that from a kernel size of 512, fft convolution should be used.

Data x Kernel	fft [ops/s]	direct [ops/s]
128 x 16	17640	256329
128 x 32	17590	138211
128 x 128	17295	35486
128 x 512	4515	9308
128 x 1024	2157	4695
512 x 16	4527	66017
512 x 32	4556	35356
512 x 128	3848	8134
512 x 512	3217	2311
512 x 1024	2079	1072
2048 x 16	681	15864
2048 x 32	977	8749
2048 x 128	1098	2164
2048 x 512	1089	591
2048 x 1024	1088	298
4096 x 16	473	8082
4096 x 32	404	4272
4096 x 128	489	1128
4096 x 512	510	296
4096 x 1024	506	149
16384 x 16	80	2112
16384 x 32	72	1107
16384 x 128	78	282
16384 x 512	79	74
16384 x 1024	77	37
65536 x 16	18	491
65536 x 32	18	270
65536 x 128	18	70
65536 x 512	17	18
65536 x 1024	19	9
262144 x 16	4	125
262144 x 32	4	67
262144 x 128	4	17
262144 x 512	4	5
262144 x 1024	4	2
1048576 x 16	1	31
1048576 x 32	1	15
1048576 x 128	1	4
1048576 x 512	1	1
1048576 x 1024	1	1

License

MIT