time-to-botec

Benchmark sampling in different programming languages
Log | Files | Refs | README

ndarray.js (3970B)

      1 /**
      2 * @license Apache-2.0
      3 *
      4 * Copyright (c) 2020 The Stdlib Authors.
      5 *
      6 * Licensed under the Apache License, Version 2.0 (the "License");
      7 * you may not use this file except in compliance with the License.
      8 * You may obtain a copy of the License at
      9 *
     10 *    http://www.apache.org/licenses/LICENSE-2.0
     11 *
     12 * Unless required by applicable law or agreed to in writing, software
     13 * distributed under the License is distributed on an "AS IS" BASIS,
     14 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     15 * See the License for the specific language governing permissions and
     16 * limitations under the License.
     17 */
     18 
     19 'use strict';
     20 
     21 // MODULES //
     22 
     23 var float64ToFloat32 = require( '@stdlib/number/float64/base/to-float32' );
     24 var floor = require( '@stdlib/math/base/special/floor' );
     25 
     26 
     27 // VARIABLES //
     28 
     29 // Blocksize for pairwise summation (NOTE: decreasing the blocksize decreases rounding error as more pairs are summed, but also decreases performance. Because the inner loop is unrolled eight times, the blocksize is effectively `16`.):
     30 var BLOCKSIZE = 128;
     31 
     32 
     33 // MAIN //
     34 
     35 /**
     36 * Computes the sum of single-precision floating-point strided array elements using pairwise summation.
     37 *
     38 * ## Method
     39 *
     40 * -   This implementation uses pairwise summation, which accrues rounding error `O(log2 N)` instead of `O(N)`. The recursion depth is also `O(log2 N)`.
     41 *
     42 * ## References
     43 *
     44 * -   Higham, Nicholas J. 1993. "The Accuracy of Floating Point Summation." _SIAM Journal on Scientific Computing_ 14 (4): 783–99. doi:[10.1137/0914050](https://doi.org/10.1137/0914050).
     45 *
     46 * @param {PositiveInteger} N - number of indexed elements
     47 * @param {Float32Array} x - input array
     48 * @param {integer} stride - stride length
     49 * @param {NonNegativeInteger} offset - starting index
     50 * @returns {number} sum
     51 *
     52 * @example
     53 * var Float32Array = require( '@stdlib/array/float32' );
     54 * var floor = require( '@stdlib/math/base/special/floor' );
     55 *
     56 * var x = new Float32Array( [ 2.0, 1.0, 2.0, -2.0, -2.0, 2.0, 3.0, 4.0 ] );
     57 * var N = floor( x.length / 2 );
     58 *
     59 * var v = ssumpw( N, x, 2, 1 );
     60 * // returns 5.0
     61 */
     62 function ssumpw( N, x, stride, offset ) {
     63 	var ix;
     64 	var s0;
     65 	var s1;
     66 	var s2;
     67 	var s3;
     68 	var s4;
     69 	var s5;
     70 	var s6;
     71 	var s7;
     72 	var M;
     73 	var s;
     74 	var n;
     75 	var i;
     76 
     77 	if ( N <= 0 ) {
     78 		return 0.0;
     79 	}
     80 	if ( N === 1 || stride === 0 ) {
     81 		return x[ offset ];
     82 	}
     83 	ix = offset;
     84 	if ( N < 8 ) {
     85 		// Use simple summation...
     86 		s = 0.0;
     87 		for ( i = 0; i < N; i++ ) {
     88 			s = float64ToFloat32( s + x[ ix ] );
     89 			ix += stride;
     90 		}
     91 		return s;
     92 	}
     93 	if ( N <= BLOCKSIZE ) {
     94 		// Sum a block with 8 accumulators (by loop unrolling, we lower the effective blocksize to 16)...
     95 		s0 = x[ ix ];
     96 		s1 = x[ ix+stride ];
     97 		s2 = x[ ix+(2*stride) ];
     98 		s3 = x[ ix+(3*stride) ];
     99 		s4 = x[ ix+(4*stride) ];
    100 		s5 = x[ ix+(5*stride) ];
    101 		s6 = x[ ix+(6*stride) ];
    102 		s7 = x[ ix+(7*stride) ];
    103 		ix += 8 * stride;
    104 
    105 		M = N % 8;
    106 		for ( i = 8; i < N-M; i += 8 ) {
    107 			s0 = float64ToFloat32( s0 + x[ ix ] );
    108 			s1 = float64ToFloat32( s1 + x[ ix+stride ] );
    109 			s2 = float64ToFloat32( s2 + x[ ix+(2*stride) ] );
    110 			s3 = float64ToFloat32( s3 + x[ ix+(3*stride) ] );
    111 			s4 = float64ToFloat32( s4 + x[ ix+(4*stride) ] );
    112 			s5 = float64ToFloat32( s5 + x[ ix+(5*stride) ] );
    113 			s6 = float64ToFloat32( s6 + x[ ix+(6*stride) ] );
    114 			s7 = float64ToFloat32( s7 + x[ ix+(7*stride) ] );
    115 			ix += 8 * stride;
    116 		}
    117 		// Pairwise sum the accumulators:
    118 		s = float64ToFloat32( float64ToFloat32( float64ToFloat32(s0+s1) + float64ToFloat32(s2+s3) ) + float64ToFloat32( float64ToFloat32(s4+s5) + float64ToFloat32(s6+s7) ) ); // eslint-disable-line max-len
    119 
    120 		// Clean-up loop...
    121 		for ( i; i < N; i++ ) {
    122 			s = float64ToFloat32( s + x[ ix ] );
    123 			ix += stride;
    124 		}
    125 		return s;
    126 	}
    127 	// Recurse by dividing by two, but avoiding non-multiples of unroll factor...
    128 	n = floor( N/2 );
    129 	n -= n % 8;
    130 	return float64ToFloat32( ssumpw( n, x, stride, ix ) + ssumpw( N-n, x, stride, ix+(n*stride) ) ); // eslint-disable-line max-len
    131 }
    132 
    133 
    134 // EXPORTS //
    135 
    136 module.exports = ssumpw;