244 lines
5.7 KiB
C
244 lines
5.7 KiB
C
/**
|
||
* @license Apache-2.0
|
||
*
|
||
* Copyright (c) 2020 The Stdlib Authors.
|
||
*
|
||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||
* you may not use this file except in compliance with the License.
|
||
* You may obtain a copy of the License at
|
||
*
|
||
* http://www.apache.org/licenses/LICENSE-2.0
|
||
*
|
||
* Unless required by applicable law or agreed to in writing, software
|
||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||
* See the License for the specific language governing permissions and
|
||
* limitations under the License.
|
||
*/
|
||
|
||
#include "stdlib/stats/base/snanvariancepn.h"
|
||
#include <stdint.h>
|
||
|
||
/**
|
||
* Computes the sum of single-precision floating-point strided array elements, ignoring `NaN` values and using pairwise summation.
|
||
*
|
||
* ## Method
|
||
*
|
||
* - This implementation uses pairwise summation, which accrues rounding error `O(log2 N)` instead of `O(N)`. The recursion depth is also `O(log2 N)`.
|
||
*
|
||
* ## References
|
||
*
|
||
* - 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).
|
||
*
|
||
* @private
|
||
* @param N number of indexed elements
|
||
* @param W two-element output array
|
||
* @param X input array
|
||
* @param stride stride length
|
||
* @return output value
|
||
*/
|
||
static void snansumpw( const int64_t N, double *W, const float *X, const int64_t stride ) {
|
||
int64_t ix;
|
||
float *xp1;
|
||
float *xp2;
|
||
float sum;
|
||
int64_t M;
|
||
int64_t n;
|
||
int64_t i;
|
||
float s0;
|
||
float s1;
|
||
float s2;
|
||
float s3;
|
||
float s4;
|
||
float s5;
|
||
float s6;
|
||
float s7;
|
||
float v;
|
||
|
||
if ( stride < 0 ) {
|
||
ix = (1-N) * stride;
|
||
} else {
|
||
ix = 0;
|
||
}
|
||
if ( N < 8 ) {
|
||
// Use simple summation...
|
||
sum = 0.0f;
|
||
n = 0;
|
||
for ( i = 0; i < N; i++ ) {
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
sum += X[ ix ];
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
}
|
||
W[ 0 ] += (double)sum;
|
||
W[ 1 ] += (double)n;
|
||
return;
|
||
}
|
||
// 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`.)
|
||
if ( N <= 128 ) {
|
||
// Sum a block with 8 accumulators (by loop unrolling, we lower the effective blocksize to 16)...
|
||
s0 = 0.0f;
|
||
s1 = 0.0f;
|
||
s2 = 0.0f;
|
||
s3 = 0.0f;
|
||
s4 = 0.0f;
|
||
s5 = 0.0f;
|
||
s6 = 0.0f;
|
||
s7 = 0.0f;
|
||
n = 0;
|
||
|
||
M = N % 8;
|
||
for ( i = 0; i < N-M; i += 8 ) {
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s0 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s1 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s2 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s3 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s4 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s5 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s6 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
s7 += v;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
}
|
||
// Pairwise sum the accumulators:
|
||
sum = ((s0+s1) + (s2+s3)) + ((s4+s5) + (s6+s7));
|
||
|
||
// Clean-up loop...
|
||
for (; i < N; i++ ) {
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
sum += X[ ix ];
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
}
|
||
W[ 0 ] += (double)sum;
|
||
W[ 1 ] += (double)n;
|
||
return;
|
||
}
|
||
// Recurse by dividing by two, but avoiding non-multiples of unroll factor...
|
||
n = N / 2;
|
||
n -= n % 8;
|
||
if ( stride < 0 ) {
|
||
xp1 = (float *)X + ( (n-N)*stride );
|
||
xp2 = (float *)X;
|
||
} else {
|
||
xp1 = (float *)X;
|
||
xp2 = (float *)X + ( n*stride );
|
||
}
|
||
snansumpw( n, W, xp1, stride );
|
||
snansumpw( N-n, W, xp2, stride );
|
||
}
|
||
|
||
/**
|
||
* Computes the variance of a single-precision floating-point strided array ignoring `NaN` values and using a two-pass algorithm.
|
||
*
|
||
* ## Method
|
||
*
|
||
* - This implementation uses a two-pass approach, as suggested by Neely (1966).
|
||
*
|
||
* ## References
|
||
*
|
||
* - 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).
|
||
* - 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).
|
||
*
|
||
* @param N number of indexed elements
|
||
* @param correction degrees of freedom adjustment
|
||
* @param X input array
|
||
* @param stride stride length
|
||
* @return output value
|
||
*/
|
||
float stdlib_strided_snanvariancepn( const int64_t N, const float correction, const float *X, const int64_t stride ) {
|
||
double W[] = { 0.0, 0.0 };
|
||
int64_t ix;
|
||
int64_t i;
|
||
double nc;
|
||
double dM;
|
||
double n;
|
||
float M2;
|
||
float mu;
|
||
float M;
|
||
float d;
|
||
float v;
|
||
|
||
if ( N <= 0 ) {
|
||
return 0.0f / 0.0f; // NaN
|
||
}
|
||
if ( N == 1 || stride == 0 ) {
|
||
v = X[ 0 ];
|
||
if ( v == v && (double)N-(double)correction > 0.0 ) {
|
||
return 0.0f;
|
||
}
|
||
return 0.0f / 0.0f; // NaN
|
||
}
|
||
// Compute an estimate for the mean...
|
||
snansumpw( N, W, X, stride );
|
||
n = W[ 1 ];
|
||
nc = n - (double)correction;
|
||
if ( nc <= 0.0 ) {
|
||
return 0.0f / 0.0f; // NaN
|
||
}
|
||
if ( stride < 0 ) {
|
||
ix = (1-N) * stride;
|
||
} else {
|
||
ix = 0;
|
||
}
|
||
mu = (float)( W[ 0 ] / n );
|
||
|
||
// Compute the variance...
|
||
M2 = 0.0f;
|
||
M = 0.0f;
|
||
for ( i = 0; i < N; i++ ) {
|
||
v = X[ ix ];
|
||
if ( v == v ) {
|
||
d = v - mu;
|
||
M2 += d * d;
|
||
M += d;
|
||
n += 1;
|
||
}
|
||
ix += stride;
|
||
}
|
||
dM = (double)M;
|
||
return (float)((double)M2/nc) - ( (float)(dM/n) * (float)(dM/nc) );
|
||
}
|