time-to-botec

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

sdsnansumpw.c (4329B)


      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/blas/ext/base/sdsnansumpw.h"
     20 #include "stdlib/math/base/assert/is_nanf.h"
     21 #include <stdint.h>
     22 
     23 /**
     24 * Computes the sum of single-precision floating-point strided array elements, ignoring `NaN` values and using pairwise summation with extended accumulation.
     25 *
     26 * ## Method
     27 *
     28 * -   This implementation uses pairwise summation, which accrues rounding error `O(log2 N)` instead of `O(N)`. The recursion depth is also `O(log2 N)`.
     29 *
     30 * ## References
     31 *
     32 * -   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).
     33 *
     34 * @param N       number of indexed elements
     35 * @param X       input array
     36 * @param stride  stride length
     37 * @return        output value
     38 */
     39 float stdlib_strided_sdsnansumpw( const int64_t N, const float *X, const int64_t stride ) {
     40 	float *xp1;
     41 	float *xp2;
     42 	double sum;
     43 	int64_t ix;
     44 	int64_t M;
     45 	int64_t n;
     46 	int64_t i;
     47 	double s0;
     48 	double s1;
     49 	double s2;
     50 	double s3;
     51 	double s4;
     52 	double s5;
     53 	double s6;
     54 	double s7;
     55 
     56 	if ( N <= 0 ) {
     57 		return 0.0f;
     58 	}
     59 	if ( N == 1 || stride == 0 ) {
     60 		if ( stdlib_base_is_nanf( X[ 0 ] ) ) {
     61 			return 0.0f;
     62 		}
     63 		return X[ 0 ];
     64 	}
     65 	if ( stride < 0 ) {
     66 		ix = (1-N) * stride;
     67 	} else {
     68 		ix = 0;
     69 	}
     70 	if ( N < 8 ) {
     71 		// Use simple summation...
     72 		sum = 0.0;
     73 		for ( i = 0; i < N; i++ ) {
     74 			if ( !stdlib_base_is_nanf( X[ ix ] ) ) {
     75 				sum += X[ ix ];
     76 			}
     77 			ix += stride;
     78 		}
     79 		return sum;
     80 	}
     81 	// 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`.)
     82 	if ( N <= 128 ) {
     83 		// Sum a block with 8 accumulators (by loop unrolling, we lower the effective blocksize to 16)...
     84 		s0 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     85 		ix += stride;
     86 		s1 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     87 		ix += stride;
     88 		s2 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     89 		ix += stride;
     90 		s3 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     91 		ix += stride;
     92 		s4 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     93 		ix += stride;
     94 		s5 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     95 		ix += stride;
     96 		s6 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     97 		ix += stride;
     98 		s7 = ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
     99 		ix += stride;
    100 
    101 		M = N % 8;
    102 		for ( i = 8; i < N-M; i += 8 ) {
    103 			s0 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    104 			ix += stride;
    105 			s1 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    106 			ix += stride;
    107 			s2 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    108 			ix += stride;
    109 			s3 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    110 			ix += stride;
    111 			s4 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    112 			ix += stride;
    113 			s5 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    114 			ix += stride;
    115 			s6 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    116 			ix += stride;
    117 			s7 += ( stdlib_base_is_nanf( X[ ix ] ) ) ? 0.0 : X[ ix ];
    118 			ix += stride;
    119 		}
    120 		// Pairwise sum the accumulators:
    121 		sum = ((s0+s1) + (s2+s3)) + ((s4+s5) + (s6+s7));
    122 
    123 		// Clean-up loop...
    124 		for (; i < N; i++ ) {
    125 			if ( !stdlib_base_is_nanf( X[ ix ] ) ) {
    126 				sum += X[ ix ];
    127 			}
    128 			ix += stride;
    129 		}
    130 		return sum;
    131 	}
    132 	// Recurse by dividing by two, but avoiding non-multiples of unroll factor...
    133 	n = N / 2;
    134 	n -= n % 8;
    135 	if ( stride < 0 ) {
    136 		xp1 = (float *)X + ( (n-N)*stride );
    137 		xp2 = (float *)X;
    138 	} else {
    139 		xp1 = (float *)X;
    140 		xp2 = (float *)X + ( n*stride );
    141 	}
    142 	return stdlib_strided_sdsnansumpw( n, xp1, stride ) + stdlib_strided_sdsnansumpw( N-n, xp2, stride );
    143 }