author | Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com> |
Fri, 16 Apr 2010 16:46:38 +0300 | |
changeset 18 | 47c74d1534e1 |
permissions | -rw-r--r-- |
18
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
1 |
/* |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
2 |
* Copyright (c) 2005 |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
3 |
* Eric Anholt. All rights reserved. |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
4 |
* |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
5 |
* Redistribution and use in source and binary forms, with or without |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
6 |
* modification, are permitted provided that the following conditions |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
7 |
* are met: |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
8 |
* 1. Redistributions of source code must retain the above copyright |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
9 |
* notice, this list of conditions and the following disclaimer. |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
10 |
* 2. Redistributions in binary form must reproduce the above copyright |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
11 |
* notice, this list of conditions and the following disclaimer in the |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
12 |
* documentation and/or other materials provided with the distribution. |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
13 |
* |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
14 |
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
15 |
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
16 |
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
17 |
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
18 |
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
19 |
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
20 |
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
21 |
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
22 |
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
23 |
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
24 |
* SUCH DAMAGE. |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
25 |
*/ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
26 |
//Portions Copyright (c) 2008-2009 Nokia Corporation and/or its subsidiary(-ies). All rights reserved. |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
27 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
28 |
#ifdef HAVE_CONFIG_H |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
29 |
#include "config.h" |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
30 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
31 |
#include <liboil/liboilclasses.h> |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
32 |
#include <liboil/liboilfunction.h> |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
33 |
#include <emmintrin.h> |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
34 |
#include <xmmintrin.h> |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
35 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
36 |
#define SSE_FUNCTION __attribute__((force_align_arg_pointer)) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
37 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
38 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
39 |
add_f32_sse_unroll2 (float *dest, float *src1, float *src2, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
40 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
41 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
42 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
43 |
*dest++ = *src1++ + *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
44 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
45 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
46 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
47 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
48 |
xmm1 = _mm_loadu_ps(src2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
49 |
xmm0 = _mm_add_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
50 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
51 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
52 |
xmm1 = _mm_loadu_ps(src2 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
53 |
xmm0 = _mm_add_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
54 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
55 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
56 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
57 |
src2 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
58 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
59 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
60 |
*dest++ = *src1++ + *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
61 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
62 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
63 |
OIL_DEFINE_IMPL_FULL (add_f32_sse_unroll2, add_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
64 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
65 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
66 |
subtract_f32_sse_unroll2 (float *dest, float *src1, float *src2, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
67 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
68 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
69 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
70 |
*dest++ = *src1++ - *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
71 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
72 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
73 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
74 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
75 |
xmm1 = _mm_loadu_ps(src2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
76 |
xmm0 = _mm_sub_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
77 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
78 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
79 |
xmm1 = _mm_loadu_ps(src2 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
80 |
xmm0 = _mm_sub_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
81 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
82 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
83 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
84 |
src2 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
85 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
86 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
87 |
*dest++ = *src1++ - *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
88 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
89 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
90 |
OIL_DEFINE_IMPL_FULL (subtract_f32_sse_unroll2, subtract_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
91 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
92 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
93 |
multiply_f32_sse_unroll2 (float *dest, float *src1, float *src2, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
94 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
95 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
96 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
97 |
*dest++ = *src1++ * *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
98 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
99 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
100 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
101 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
102 |
xmm1 = _mm_loadu_ps(src2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
103 |
xmm0 = _mm_mul_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
104 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
105 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
106 |
xmm1 = _mm_loadu_ps(src2 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
107 |
xmm0 = _mm_mul_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
108 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
109 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
110 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
111 |
src2 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
112 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
113 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
114 |
*dest++ = *src1++ * *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
115 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
116 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
117 |
OIL_DEFINE_IMPL_FULL (multiply_f32_sse_unroll2, multiply_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
118 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
119 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
120 |
divide_f32_sse_unroll2 (float *dest, float *src1, float *src2, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
121 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
122 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
123 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
124 |
*dest++ = *src1++ / *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
125 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
126 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
127 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
128 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
129 |
xmm1 = _mm_loadu_ps(src2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
130 |
xmm0 = _mm_div_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
131 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
132 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
133 |
xmm1 = _mm_loadu_ps(src2 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
134 |
xmm0 = _mm_div_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
135 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
136 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
137 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
138 |
src2 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
139 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
140 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
141 |
*dest++ = *src1++ / *src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
142 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
143 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
144 |
OIL_DEFINE_IMPL_FULL (divide_f32_sse_unroll2, divide_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
145 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
146 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
147 |
minimum_f32_sse_unroll2 (float *dest, float *src1, float *src2, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
148 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
149 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
150 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
151 |
*dest++ = *src1 < *src2 ? *src1 : *src2; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
152 |
src1++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
153 |
src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
154 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
155 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
156 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
157 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
158 |
xmm1 = _mm_loadu_ps(src2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
159 |
xmm0 = _mm_min_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
160 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
161 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
162 |
xmm1 = _mm_loadu_ps(src2 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
163 |
xmm0 = _mm_min_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
164 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
165 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
166 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
167 |
src2 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
168 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
169 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
170 |
*dest++ = *src1 < *src2 ? *src1 : *src2; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
171 |
src1++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
172 |
src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
173 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
174 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
175 |
OIL_DEFINE_IMPL_FULL (minimum_f32_sse_unroll2, minimum_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
176 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
177 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
178 |
maximum_f32_sse_unroll2 (float *dest, float *src1, float *src2, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
179 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
180 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
181 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
182 |
*dest++ = *src1 > *src2 ? *src1 : *src2; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
183 |
src1++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
184 |
src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
185 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
186 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
187 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
188 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
189 |
xmm1 = _mm_loadu_ps(src2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
190 |
xmm0 = _mm_max_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
191 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
192 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
193 |
xmm1 = _mm_loadu_ps(src2 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
194 |
xmm0 = _mm_max_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
195 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
196 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
197 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
198 |
src2 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
199 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
200 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
201 |
*dest++ = *src1 > *src2 ? *src1 : *src2; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
202 |
src1++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
203 |
src2++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
204 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
205 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
206 |
OIL_DEFINE_IMPL_FULL (maximum_f32_sse_unroll2, maximum_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
207 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
208 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
209 |
inverse_f32_sse_unroll2 (float *dest, float *src1, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
210 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
211 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
212 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
213 |
*dest++ = 1.0 / *src1++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
214 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
215 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
216 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
217 |
/* While _mm_rcp_ps sounds promising, the results it gives are rather |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
218 |
* different from the 1.0 / src1 reference implementation, so do that. |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
219 |
*/ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
220 |
xmm0 = _mm_set_ps1(1.0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
221 |
xmm1 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
222 |
xmm0 = _mm_div_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
223 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
224 |
xmm0 = _mm_set_ps1(1.0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
225 |
xmm1 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
226 |
xmm0 = _mm_div_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
227 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
228 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
229 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
230 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
231 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
232 |
*dest++ = 1.0 / *src1++; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
233 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
234 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
235 |
OIL_DEFINE_IMPL_FULL (inverse_f32_sse_unroll2, inverse_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
236 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
237 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
238 |
negative_f32_sse_unroll2 (float *dest, float *src1, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
239 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
240 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
241 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
242 |
*dest++ = -(*src1++); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
243 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
244 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
245 |
__m128 xmm0, xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
246 |
xmm0 = _mm_setzero_ps(); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
247 |
xmm1 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
248 |
xmm0 = _mm_sub_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
249 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
250 |
xmm0 = _mm_setzero_ps(); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
251 |
xmm1 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
252 |
xmm0 = _mm_sub_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
253 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
254 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
255 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
256 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
257 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
258 |
*dest++ = -(*src1++); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
259 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
260 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
261 |
OIL_DEFINE_IMPL_FULL (negative_f32_sse_unroll2, negative_f32, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
262 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
263 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
264 |
scalaradd_f32_ns_sse_unroll2 (float *dest, float *src1, float *val, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
265 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
266 |
__m128 xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
267 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
268 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
269 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
270 |
*dest++ = *src1++ + *val; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
271 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
272 |
xmm1 = _mm_load_ps1(val); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
273 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
274 |
__m128 xmm0; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
275 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
276 |
xmm0 = _mm_add_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
277 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
278 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
279 |
xmm0 = _mm_add_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
280 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
281 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
282 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
283 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
284 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
285 |
*dest++ = *src1++ + *val; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
286 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
287 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
288 |
OIL_DEFINE_IMPL_FULL (scalaradd_f32_ns_sse_unroll2, scalaradd_f32_ns, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
289 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
290 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
291 |
scalarmultiply_f32_ns_sse_unroll2 (float *dest, float *src1, float *val, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
292 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
293 |
__m128 xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
294 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
295 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
296 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
297 |
*dest++ = *src1++ * *val; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
298 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
299 |
xmm1 = _mm_load_ps1(val); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
300 |
for (; n >= 8; n -= 8) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
301 |
__m128 xmm0; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
302 |
xmm0 = _mm_loadu_ps(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
303 |
xmm0 = _mm_mul_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
304 |
_mm_store_ps(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
305 |
xmm0 = _mm_loadu_ps(src1 + 4); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
306 |
xmm0 = _mm_mul_ps(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
307 |
_mm_store_ps(dest + 4, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
308 |
dest += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
309 |
src1 += 8; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
310 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
311 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
312 |
*dest++ = *src1++ * *val; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
313 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
314 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
315 |
OIL_DEFINE_IMPL_FULL (scalarmultiply_f32_ns_sse_unroll2, scalarmultiply_f32_ns, OIL_IMPL_FLAG_SSE); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
316 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
317 |
SSE_FUNCTION static void |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
318 |
scalarmultiply_f64_ns_sse2_unroll2 (double *dest, double *src1, double *val, int n) |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
319 |
{ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
320 |
__m128d xmm1; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
321 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
322 |
/* Initial operations to align the destination pointer */ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
323 |
for (; ((long)dest & 15) && (n > 0); n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
324 |
*dest++ = *src1++ * *val; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
325 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
326 |
xmm1 = _mm_load_pd1(val); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
327 |
for (; n >= 4; n -= 4) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
328 |
__m128d xmm0; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
329 |
xmm0 = _mm_loadu_pd(src1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
330 |
xmm0 = _mm_mul_pd(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
331 |
_mm_store_pd(dest, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
332 |
xmm0 = _mm_loadu_pd(src1 + 2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
333 |
xmm0 = _mm_mul_pd(xmm0, xmm1); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
334 |
_mm_store_pd(dest + 2, xmm0); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
335 |
dest += 4; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
336 |
src1 += 4; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
337 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
338 |
for (; n > 0; n--) { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
339 |
*dest++ = *src1++ * *val; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
340 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
341 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
342 |
OIL_DEFINE_IMPL_FULL (scalarmultiply_f64_ns_sse2_unroll2, scalarmultiply_f64_ns, OIL_IMPL_FLAG_SSE2); |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
343 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
344 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
345 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
346 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
347 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
348 |
OilFunctionImpl* __oil_function_impl_add_f32_sse_unroll2, add_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
349 |
return &_oil_function_impl_add_f32_sse_unroll2, add_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
350 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
351 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
352 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
353 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
354 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
355 |
OilFunctionImpl* __oil_function_impl_subtract_f32_sse_unroll2, subtract_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
356 |
return &_oil_function_impl_subtract_f32_sse_unroll2, subtract_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
357 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
358 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
359 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
360 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
361 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
362 |
OilFunctionImpl* __oil_function_impl_multiply_f32_sse_unroll2, multiply_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
363 |
return &_oil_function_impl_multiply_f32_sse_unroll2, multiply_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
364 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
365 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
366 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
367 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
368 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
369 |
OilFunctionImpl* __oil_function_impl_divide_f32_sse_unroll2, divide_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
370 |
return &_oil_function_impl_divide_f32_sse_unroll2, divide_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
371 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
372 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
373 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
374 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
375 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
376 |
OilFunctionImpl* __oil_function_impl_minimum_f32_sse_unroll2, minimum_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
377 |
return &_oil_function_impl_minimum_f32_sse_unroll2, minimum_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
378 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
379 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
380 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
381 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
382 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
383 |
OilFunctionImpl* __oil_function_impl_maximum_f32_sse_unroll2, maximum_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
384 |
return &_oil_function_impl_maximum_f32_sse_unroll2, maximum_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
385 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
386 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
387 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
388 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
389 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
390 |
OilFunctionImpl* __oil_function_impl_inverse_f32_sse_unroll2, inverse_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
391 |
return &_oil_function_impl_inverse_f32_sse_unroll2, inverse_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
392 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
393 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
394 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
395 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
396 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
397 |
OilFunctionImpl* __oil_function_impl_negative_f32_sse_unroll2, negative_f32() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
398 |
return &_oil_function_impl_negative_f32_sse_unroll2, negative_f32; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
399 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
400 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
401 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
402 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
403 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
404 |
OilFunctionImpl* __oil_function_impl_scalaradd_f32_ns_sse_unroll2, scalaradd_f32_ns() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
405 |
return &_oil_function_impl_scalaradd_f32_ns_sse_unroll2, scalaradd_f32_ns; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
406 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
407 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
408 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
409 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
410 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
411 |
OilFunctionImpl* __oil_function_impl_scalarmultiply_f32_ns_sse_unroll2, scalarmultiply_f32_ns() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
412 |
return &_oil_function_impl_scalarmultiply_f32_ns_sse_unroll2, scalarmultiply_f32_ns; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
413 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
414 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
415 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
416 |
#ifdef __SYMBIAN32__ |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
417 |
|
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
418 |
OilFunctionImpl* __oil_function_impl_scalarmultiply_f64_ns_sse2_unroll2, scalarmultiply_f64_ns() { |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
419 |
return &_oil_function_impl_scalarmultiply_f64_ns_sse2_unroll2, scalarmultiply_f64_ns; |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
420 |
} |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
421 |
#endif |
47c74d1534e1
Revision: 201011
Dremov Kirill (Nokia-D-MSW/Tampere) <kirill.dremov@nokia.com>
parents:
diff
changeset
|
422 |