1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
|
/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*- */
/*
* This file is part of the LibreOffice project.
*
* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/.
*
*/
#define LO_ARRAYSUM_SPACE AVX
#include "arraysum.hxx"
#include <arraysumfunctorinternal.hxx>
#include <tools/simd.hxx>
#include <tools/simdsupport.hxx>
#include <stdlib.h>
namespace sc::op
{
#ifdef LO_AVX_AVAILABLE // Old processors
using namespace AVX;
/** Kahan sum with AVX.
*/
static inline void sumAVX(__m256d& sum, __m256d& err, const __m256d& value)
{
const __m256d ANNULATE_SIGN_BIT
= _mm256_castsi256_pd(_mm256_set1_epi64x(0x7FFF'FFFF'FFFF'FFFF));
// Temporal parameter
__m256d t = _mm256_add_pd(sum, value);
// Absolute value of the total sum
__m256d asum = _mm256_and_pd(sum, ANNULATE_SIGN_BIT);
// Absolute value of the value to add
__m256d avalue = _mm256_and_pd(value, ANNULATE_SIGN_BIT);
// Compare the absolute values sum >= value
__m256d mask = _mm256_cmp_pd(asum, avalue, _CMP_GE_OQ);
// The following code has this form ( a - t + b)
// Case 1: a = sum b = value
// Case 2: a = value b = sum
__m256d a = _mm256_add_pd(_mm256_and_pd(mask, sum), _mm256_andnot_pd(mask, value));
__m256d b = _mm256_add_pd(_mm256_and_pd(mask, value), _mm256_andnot_pd(mask, sum));
err = _mm256_add_pd(err, _mm256_add_pd(_mm256_sub_pd(a, t), b));
// Store result
sum = t;
}
#endif
/** Execute Kahan sum with AVX.
*/
KahanSumSimple executeAVX(size_t& i, size_t nSize, const double* pCurrent)
{
#ifdef LO_AVX_AVAILABLE
// Make sure we don't fall out of bounds.
// This works by sums of 8 terms.
// So the 8'th term is i+7
// If we iterate until nSize won't fall out of bounds
if (nSize > i + 7)
{
// Setup sums and errors as 0
__m256d sum1 = _mm256_setzero_pd();
__m256d err1 = _mm256_setzero_pd();
__m256d sum2 = _mm256_setzero_pd();
__m256d err2 = _mm256_setzero_pd();
for (; i + 7 < nSize; i += 8)
{
// Kahan sum 1
__m256d load1 = _mm256_loadu_pd(pCurrent);
sumAVX(sum1, err1, load1);
pCurrent += 4;
// Kahan sum 2
__m256d load2 = _mm256_loadu_pd(pCurrent);
sumAVX(sum2, err2, load2);
pCurrent += 4;
}
// Now we combine pairwise summation with Kahan summation
// sum 1 + sum 2 -> sum 1
sumAVX(sum1, err1, sum2);
sumAVX(sum1, err1, err2);
// Store results
double sums[4];
double errs[4];
_mm256_storeu_pd(&sums[0], sum1);
_mm256_storeu_pd(&errs[0], err1);
// First Kahan & pairwise summation
// 0+1 1+2 -> 0, 2
sumNeumanierNormal(sums[0], errs[0], sums[1]);
sumNeumanierNormal(sums[2], errs[2], sums[3]);
sumNeumanierNormal(sums[0], errs[0], errs[1]);
sumNeumanierNormal(sums[2], errs[2], errs[3]);
// 0+2 -> 0
sumNeumanierNormal(sums[0], errs[0], sums[2]);
sumNeumanierNormal(sums[0], errs[0], errs[2]);
// Store result
return { sums[0], errs[0] };
}
return { 0.0, 0.0 };
#else
(void)i;
(void)nSize;
(void)pCurrent;
abort();
#endif
}
} // end namespace sc::op
/* vim:set shiftwidth=4 softtabstop=4 expandtab: */
|
