|
sl@0
|
1 |
/*
|
|
sl@0
|
2 |
* Copyright (c) 2009 Nokia Corporation and/or its subsidiary(-ies).
|
|
sl@0
|
3 |
* All rights reserved.
|
|
sl@0
|
4 |
* This component and the accompanying materials are made available
|
|
sl@0
|
5 |
* under the terms of "Eclipse Public License v1.0"
|
|
sl@0
|
6 |
* which accompanies this distribution, and is available
|
|
sl@0
|
7 |
* at the URL "http://www.eclipse.org/legal/epl-v10.html".
|
|
sl@0
|
8 |
*
|
|
sl@0
|
9 |
* Initial Contributors:
|
|
sl@0
|
10 |
* Nokia Corporation - initial contribution.
|
|
sl@0
|
11 |
*
|
|
sl@0
|
12 |
* Contributors:
|
|
sl@0
|
13 |
*
|
|
sl@0
|
14 |
* Description:
|
|
sl@0
|
15 |
*
|
|
sl@0
|
16 |
*/
|
|
sl@0
|
17 |
|
|
sl@0
|
18 |
#ifdef HAVE_CONFIG_H
|
|
sl@0
|
19 |
#include "config.h"
|
|
sl@0
|
20 |
#endif
|
|
sl@0
|
21 |
#include <liboil/liboilclasses.h>
|
|
sl@0
|
22 |
#include <liboil/liboilfunction.h>
|
|
sl@0
|
23 |
#include <emmintrin.h>
|
|
sl@0
|
24 |
|
|
sl@0
|
25 |
#define SSE_FUNCTION __attribute__((force_align_arg_pointer))
|
|
sl@0
|
26 |
|
|
sl@0
|
27 |
#define MULTSUM_SSE2_NSTRIDED(i) { \
|
|
sl@0
|
28 |
t1 = _mm_load_pd(&OIL_GET(src1, i, double)); \
|
|
sl@0
|
29 |
t2 = _mm_load_pd(&OIL_GET(src2, i, double)); \
|
|
sl@0
|
30 |
t1 = _mm_mul_pd(t1,t2); \
|
|
sl@0
|
31 |
sum.reg = _mm_add_pd(sum.reg,t1); \
|
|
sl@0
|
32 |
}
|
|
sl@0
|
33 |
#define MULTSUM_SSE2_NSTRIDEDP(i) { \
|
|
sl@0
|
34 |
t1 = _mm_load_pd(&OIL_GET(src1, i*sstr1, double)); \
|
|
sl@0
|
35 |
t2 = _mm_loadl_pd(t2, &OIL_GET(src2, i*sstr2, double)); \
|
|
sl@0
|
36 |
t2 = _mm_loadh_pd(t2, &OIL_GET(src2, (i+1)*sstr2, double)); \
|
|
sl@0
|
37 |
t1 = _mm_mul_pd(t1,t2); \
|
|
sl@0
|
38 |
sum.reg = _mm_add_pd(sum.reg,t1); \
|
|
sl@0
|
39 |
}
|
|
sl@0
|
40 |
#define MULTSUM_SSE2_STRIDED(i) { \
|
|
sl@0
|
41 |
t1 = _mm_loadl_pd(t1, &OIL_GET(src1, i*sstr1, double)); \
|
|
sl@0
|
42 |
t1 = _mm_loadh_pd(t1, &OIL_GET(src1, (i+1)*sstr1, double)); \
|
|
sl@0
|
43 |
t2 = _mm_loadl_pd(t2, &OIL_GET(src2, i*sstr2, double)); \
|
|
sl@0
|
44 |
t2 = _mm_loadh_pd(t2, &OIL_GET(src2, (i+1)*sstr2, double)); \
|
|
sl@0
|
45 |
t1 = _mm_mul_pd(t1,t2); \
|
|
sl@0
|
46 |
sum.reg = _mm_add_pd(sum.reg,t1); \
|
|
sl@0
|
47 |
}
|
|
sl@0
|
48 |
|
|
sl@0
|
49 |
|
|
sl@0
|
50 |
#ifdef ENABLE_BROKEN_IMPLS
|
|
sl@0
|
51 |
SSE_FUNCTION static void
|
|
sl@0
|
52 |
multsum_f64_sse2_unroll4(double *dest,
|
|
sl@0
|
53 |
const double *src1, int sstr1,
|
|
sl@0
|
54 |
const double *src2, int sstr2,
|
|
sl@0
|
55 |
int n)
|
|
sl@0
|
56 |
{
|
|
sl@0
|
57 |
__m128d t1, t2;
|
|
sl@0
|
58 |
union {
|
|
sl@0
|
59 |
__m128d reg;
|
|
sl@0
|
60 |
double vals[2];
|
|
sl@0
|
61 |
} sum;
|
|
sl@0
|
62 |
int i = 0;
|
|
sl@0
|
63 |
|
|
sl@0
|
64 |
sum.reg = _mm_setzero_pd();
|
|
sl@0
|
65 |
while (i < n-3) {
|
|
sl@0
|
66 |
MULTSUM_SSE2_STRIDED(0);
|
|
sl@0
|
67 |
MULTSUM_SSE2_STRIDED(2);
|
|
sl@0
|
68 |
|
|
sl@0
|
69 |
OIL_INCREMENT(src1, 4*sstr1);
|
|
sl@0
|
70 |
OIL_INCREMENT(src2, 4*sstr2);
|
|
sl@0
|
71 |
i += 4;
|
|
sl@0
|
72 |
}
|
|
sl@0
|
73 |
while (i < n-1) {
|
|
sl@0
|
74 |
MULTSUM_SSE2_STRIDED(0);
|
|
sl@0
|
75 |
|
|
sl@0
|
76 |
OIL_INCREMENT(src1, 2*sstr1);
|
|
sl@0
|
77 |
OIL_INCREMENT(src2, 2*sstr2);
|
|
sl@0
|
78 |
i+=2;
|
|
sl@0
|
79 |
}
|
|
sl@0
|
80 |
*dest = sum.vals[0] + sum.vals[1];
|
|
sl@0
|
81 |
if (i < n) {
|
|
sl@0
|
82 |
*dest += (OIL_GET(src1,0,double)*OIL_GET(src2,0,double));
|
|
sl@0
|
83 |
}
|
|
sl@0
|
84 |
}
|
|
sl@0
|
85 |
OIL_DEFINE_IMPL_FULL (multsum_f64_sse2_unroll4, multsum_f64, OIL_IMPL_FLAG_SSE2);
|
|
sl@0
|
86 |
#endif
|
|
sl@0
|
87 |
|
|
sl@0
|
88 |
|
|
sl@0
|
89 |
|
|
sl@0
|
90 |
#ifdef __SYMBIAN32__
|
|
sl@0
|
91 |
|
|
sl@0
|
92 |
OilFunctionImpl* __oil_function_impl_multsum_f64_sse2_unroll4, multsum_f64() {
|
|
sl@0
|
93 |
return &_oil_function_impl_multsum_f64_sse2_unroll4, multsum_f64;
|
|
sl@0
|
94 |
}
|
|
sl@0
|
95 |
#endif
|
|
sl@0
|
96 |
|