time-to-botec

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

dnanvariancepn.c (5741B)


      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 #include "stdlib/stats/base/dnanvariancepn.h"
     20 #include <stdint.h>
     21 
     22 /**
     23 * Computes the sum of double-precision floating-point strided array elements, ignoring `NaN` values and using pairwise summation.
     24 *
     25 * ## Method
     26 *
     27 * -   This implementation uses pairwise summation, which accrues rounding error `O(log2 N)` instead of `O(N)`. The recursion depth is also `O(log2 N)`.
     28 *
     29 * ## References
     30 *
     31 * -   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).
     32 *
     33 * @private
     34 * @param N       number of indexed elements
     35 * @param W       two-element output array
     36 * @param X       input array
     37 * @param stride  stride length
     38 * @return        output value
     39 */
     40 static void dnansumpw( const int64_t N, double *W, const double *X, const int64_t stride ) {
     41 	double *xp1;
     42 	double *xp2;
     43 	double sum;
     44 	int64_t ix;
     45 	int64_t M;
     46 	int64_t n;
     47 	int64_t i;
     48 	double s0;
     49 	double s1;
     50 	double s2;
     51 	double s3;
     52 	double s4;
     53 	double s5;
     54 	double s6;
     55 	double s7;
     56 	double v;
     57 
     58 	if ( stride < 0 ) {
     59 		ix = (1-N) * stride;
     60 	} else {
     61 		ix = 0;
     62 	}
     63 	if ( N < 8 ) {
     64 		// Use simple summation...
     65 		sum = 0.0;
     66 		n = 0;
     67 		for ( i = 0; i < N; i++ ) {
     68 			v = X[ ix ];
     69 			if ( v == v ) {
     70 				sum += X[ ix ];
     71 				n += 1;
     72 			}
     73 			ix += stride;
     74 		}
     75 		W[ 0 ] += sum;
     76 		W[ 1 ] += n;
     77 		return;
     78 	}
     79 	// Blocksize for pairwise summation: 128 (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`.)
     80 	if ( N <= 128 ) {
     81 		// Sum a block with 8 accumulators (by loop unrolling, we lower the effective blocksize to 16)...
     82 		s0 = 0.0;
     83 		s1 = 0.0;
     84 		s2 = 0.0;
     85 		s3 = 0.0;
     86 		s4 = 0.0;
     87 		s5 = 0.0;
     88 		s6 = 0.0;
     89 		s7 = 0.0;
     90 		n = 0;
     91 
     92 		M = N % 8;
     93 		for ( i = 0; i < N-M; i += 8 ) {
     94 			v = X[ ix ];
     95 			if ( v == v ) {
     96 				s0 += v;
     97 				n += 1;
     98 			}
     99 			ix += stride;
    100 			v = X[ ix ];
    101 			if ( v == v ) {
    102 				s1 += v;
    103 				n += 1;
    104 			}
    105 			ix += stride;
    106 			v = X[ ix ];
    107 			if ( v == v ) {
    108 				s2 += v;
    109 				n += 1;
    110 			}
    111 			ix += stride;
    112 			v = X[ ix ];
    113 			if ( v == v ) {
    114 				s3 += v;
    115 				n += 1;
    116 			}
    117 			ix += stride;
    118 			v = X[ ix ];
    119 			if ( v == v ) {
    120 				s4 += v;
    121 				n += 1;
    122 			}
    123 			ix += stride;
    124 			v = X[ ix ];
    125 			if ( v == v ) {
    126 				s5 += v;
    127 				n += 1;
    128 			}
    129 			ix += stride;
    130 			v = X[ ix ];
    131 			if ( v == v ) {
    132 				s6 += v;
    133 				n += 1;
    134 			}
    135 			ix += stride;
    136 			v = X[ ix ];
    137 			if ( v == v ) {
    138 				s7 += v;
    139 				n += 1;
    140 			}
    141 			ix += stride;
    142 		}
    143 		// Pairwise sum the accumulators:
    144 		sum = ((s0+s1) + (s2+s3)) + ((s4+s5) + (s6+s7));
    145 
    146 		// Clean-up loop...
    147 		for (; i < N; i++ ) {
    148 			v = X[ ix ];
    149 			if ( v == v ) {
    150 				sum += X[ ix ];
    151 				n += 1;
    152 			}
    153 			ix += stride;
    154 		}
    155 		W[ 0 ] += sum;
    156 		W[ 1 ] += n;
    157 		return;
    158 	}
    159 	// Recurse by dividing by two, but avoiding non-multiples of unroll factor...
    160 	n = N / 2;
    161 	n -= n % 8;
    162 	if ( stride < 0 ) {
    163 		xp1 = (double *)X + ( (n-N)*stride );
    164 		xp2 = (double *)X;
    165 	} else {
    166 		xp1 = (double *)X;
    167 		xp2 = (double *)X + ( n*stride );
    168 	}
    169 	dnansumpw( n, W, xp1, stride );
    170 	dnansumpw( N-n, W, xp2, stride );
    171 }
    172 
    173 /**
    174 * Computes the variance of a double-precision floating-point strided array ignoring `NaN` values and using a two-pass algorithm.
    175 *
    176 * ## Method
    177 *
    178 * -   This implementation uses a two-pass approach, as suggested by Neely (1966).
    179 *
    180 * ## References
    181 *
    182 * -   Neely, Peter M. 1966. "Comparison of Several Algorithms for Computation of Means, Standard Deviations and Correlation Coefficients." _Communications of the ACM_ 9 (7). Association for Computing Machinery: 496–99. doi:[10.1145/365719.365958](https://doi.org/10.1145/365719.365958).
    183 * -   Schubert, Erich, and Michael Gertz. 2018. "Numerically Stable Parallel Computation of (Co-)Variance." In _Proceedings of the 30th International Conference on Scientific and Statistical Database Management_. New York, NY, USA: Association for Computing Machinery. doi:[10.1145/3221269.3223036](https://doi.org/10.1145/3221269.3223036).
    184 *
    185 * @param N           number of indexed elements
    186 * @param correction  degrees of freedom adjustment
    187 * @param X           input array
    188 * @param stride      stride length
    189 * @return            output value
    190 */
    191 double stdlib_strided_dnanvariancepn( const int64_t N, const double correction, const double *X, const int64_t stride ) {
    192 	double W[] = { 0.0, 0.0 };
    193 	int64_t ix;
    194 	int64_t i;
    195 	double mu;
    196 	double M2;
    197 	double nc;
    198 	double M;
    199 	double n;
    200 	double d;
    201 	double v;
    202 
    203 	if ( N <= 0 ) {
    204 		return 0.0 / 0.0; // NaN
    205 	}
    206 	if ( N == 1 || stride == 0 ) {
    207 		v = X[ 0 ];
    208 		if ( v == v && (double)N-correction > 0.0 ) {
    209 			return 0.0;
    210 		}
    211 		return 0.0 / 0.0; // NaN
    212 	}
    213 	// Compute an estimate for the mean...
    214 	dnansumpw( N, W, X, stride );
    215 	n = W[ 1 ];
    216 	nc = n - correction;
    217 	if ( nc <= 0.0 ) {
    218 		return 0.0 / 0.0; // NaN
    219 	}
    220 	if ( stride < 0 ) {
    221 		ix = (1-N) * stride;
    222 	} else {
    223 		ix = 0;
    224 	}
    225 	mu = W[ 0 ] / n;
    226 
    227 	// Compute the variance...
    228 	M2 = 0.0;
    229 	M = 0.0;
    230 	for ( i = 0; i < N; i++ ) {
    231 		v = X[ ix ];
    232 		if ( v == v ) {
    233 			d = v - mu;
    234 			M2 += d * d;
    235 			M += d;
    236 		}
    237 		ix += stride;
    238 	}
    239 	return (M2/nc) - ((M/n)*(M/nc));
    240 }