sl@0
|
1 |
/*
|
sl@0
|
2 |
* Copyright (c) 2005
|
sl@0
|
3 |
* Eric Anholt. All rights reserved.
|
sl@0
|
4 |
*
|
sl@0
|
5 |
* Redistribution and use in source and binary forms, with or without
|
sl@0
|
6 |
* modification, are permitted provided that the following conditions
|
sl@0
|
7 |
* are met:
|
sl@0
|
8 |
* 1. Redistributions of source code must retain the above copyright
|
sl@0
|
9 |
* notice, this list of conditions and the following disclaimer.
|
sl@0
|
10 |
* 2. Redistributions in binary form must reproduce the above copyright
|
sl@0
|
11 |
* notice, this list of conditions and the following disclaimer in the
|
sl@0
|
12 |
* documentation and/or other materials provided with the distribution.
|
sl@0
|
13 |
*
|
sl@0
|
14 |
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND
|
sl@0
|
15 |
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
sl@0
|
16 |
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
sl@0
|
17 |
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE
|
sl@0
|
18 |
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
sl@0
|
19 |
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
|
sl@0
|
20 |
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
|
sl@0
|
21 |
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
sl@0
|
22 |
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
|
sl@0
|
23 |
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
|
sl@0
|
24 |
* SUCH DAMAGE.
|
sl@0
|
25 |
*/
|
sl@0
|
26 |
//Portions Copyright (c) 2008-2009 Nokia Corporation and/or its subsidiary(-ies). All rights reserved.
|
sl@0
|
27 |
|
sl@0
|
28 |
#ifdef HAVE_CONFIG_H
|
sl@0
|
29 |
#include "config.h"
|
sl@0
|
30 |
#endif
|
sl@0
|
31 |
#include <liboil/liboilclasses.h>
|
sl@0
|
32 |
#include <liboil/liboilfunction.h>
|
sl@0
|
33 |
#include <emmintrin.h>
|
sl@0
|
34 |
|
sl@0
|
35 |
#define SSE_FUNCTION __attribute__((force_align_arg_pointer))
|
sl@0
|
36 |
|
sl@0
|
37 |
SSE_FUNCTION static void
|
sl@0
|
38 |
splat_u32_ns_sse (uint32_t *dest, const uint32_t *param, int n)
|
sl@0
|
39 |
{
|
sl@0
|
40 |
__m128i v;
|
sl@0
|
41 |
|
sl@0
|
42 |
v = _mm_set1_epi32(*param);
|
sl@0
|
43 |
|
sl@0
|
44 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
45 |
*dest++ = *param;
|
sl@0
|
46 |
}
|
sl@0
|
47 |
for (; n >= 4; n -= 4) {
|
sl@0
|
48 |
_mm_store_si128((__m128i *)dest, v);
|
sl@0
|
49 |
dest += 4;
|
sl@0
|
50 |
}
|
sl@0
|
51 |
for (; n > 0; n--) {
|
sl@0
|
52 |
*dest++ = *param;
|
sl@0
|
53 |
}
|
sl@0
|
54 |
}
|
sl@0
|
55 |
OIL_DEFINE_IMPL_FULL (splat_u32_ns_sse, splat_u32_ns, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
56 |
|
sl@0
|
57 |
SSE_FUNCTION static void
|
sl@0
|
58 |
splat_u32_ns_sse_unroll2 (uint32_t *dest, const uint32_t *param, int n)
|
sl@0
|
59 |
{
|
sl@0
|
60 |
__m128i v;
|
sl@0
|
61 |
|
sl@0
|
62 |
v = _mm_set1_epi32(*param);
|
sl@0
|
63 |
|
sl@0
|
64 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
65 |
*dest++ = *param;
|
sl@0
|
66 |
}
|
sl@0
|
67 |
for (; n >= 8; n -= 8) {
|
sl@0
|
68 |
_mm_store_si128((__m128i *)dest, v);
|
sl@0
|
69 |
_mm_store_si128((__m128i *)(dest + 4), v);
|
sl@0
|
70 |
dest += 8;
|
sl@0
|
71 |
}
|
sl@0
|
72 |
for (; n >= 4; n -= 4) {
|
sl@0
|
73 |
_mm_store_si128((__m128i *)dest, v);
|
sl@0
|
74 |
dest += 4;
|
sl@0
|
75 |
}
|
sl@0
|
76 |
for (; n > 0; n--) {
|
sl@0
|
77 |
*dest++ = *param;
|
sl@0
|
78 |
}
|
sl@0
|
79 |
}
|
sl@0
|
80 |
OIL_DEFINE_IMPL_FULL (splat_u32_ns_sse_unroll2, splat_u32_ns, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
81 |
|
sl@0
|
82 |
SSE_FUNCTION static void
|
sl@0
|
83 |
splat_u8_ns_sse (uint8_t *dest, const uint8_t *param, int n)
|
sl@0
|
84 |
{
|
sl@0
|
85 |
__m128i v;
|
sl@0
|
86 |
|
sl@0
|
87 |
v = _mm_set1_epi8(*param);
|
sl@0
|
88 |
|
sl@0
|
89 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
90 |
*dest++ = *param;
|
sl@0
|
91 |
}
|
sl@0
|
92 |
for (; n >= 16; n -= 16) {
|
sl@0
|
93 |
_mm_store_si128((__m128i *)dest, v);
|
sl@0
|
94 |
dest += 16;
|
sl@0
|
95 |
}
|
sl@0
|
96 |
for (; n > 0; n--) {
|
sl@0
|
97 |
*dest++ = *param;
|
sl@0
|
98 |
}
|
sl@0
|
99 |
}
|
sl@0
|
100 |
OIL_DEFINE_IMPL_FULL (splat_u8_ns_sse, splat_u8_ns, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
101 |
|
sl@0
|
102 |
SSE_FUNCTION static void
|
sl@0
|
103 |
splat_u8_ns_sse_unroll2 (uint8_t *dest, const uint8_t *param, int n)
|
sl@0
|
104 |
{
|
sl@0
|
105 |
__m128i v;
|
sl@0
|
106 |
|
sl@0
|
107 |
v = _mm_set1_epi8(*param);
|
sl@0
|
108 |
|
sl@0
|
109 |
for (; ((long)dest & 15) && (n > 0); n--) {
|
sl@0
|
110 |
*dest++ = *param;
|
sl@0
|
111 |
}
|
sl@0
|
112 |
for (; n >= 32; n -= 32) {
|
sl@0
|
113 |
_mm_store_si128((__m128i *)dest, v);
|
sl@0
|
114 |
_mm_store_si128((__m128i *)(dest + 16), v);
|
sl@0
|
115 |
dest += 32;
|
sl@0
|
116 |
}
|
sl@0
|
117 |
for (; n >= 16; n -= 16) {
|
sl@0
|
118 |
_mm_store_si128((__m128i *)dest, v);
|
sl@0
|
119 |
dest += 16;
|
sl@0
|
120 |
}
|
sl@0
|
121 |
for (; n > 0; n--) {
|
sl@0
|
122 |
*dest++ = *param;
|
sl@0
|
123 |
}
|
sl@0
|
124 |
}
|
sl@0
|
125 |
OIL_DEFINE_IMPL_FULL (splat_u8_ns_sse_unroll2, splat_u8_ns, OIL_IMPL_FLAG_SSE2);
|
sl@0
|
126 |
|
sl@0
|
127 |
|
sl@0
|
128 |
#ifdef __SYMBIAN32__
|
sl@0
|
129 |
|
sl@0
|
130 |
OilFunctionImpl* __oil_function_impl_splat_u32_ns_sse, splat_u32_ns() {
|
sl@0
|
131 |
return &_oil_function_impl_splat_u32_ns_sse, splat_u32_ns;
|
sl@0
|
132 |
}
|
sl@0
|
133 |
#endif
|
sl@0
|
134 |
|
sl@0
|
135 |
#ifdef __SYMBIAN32__
|
sl@0
|
136 |
|
sl@0
|
137 |
OilFunctionImpl* __oil_function_impl_splat_u32_ns_sse_unroll2, splat_u32_ns() {
|
sl@0
|
138 |
return &_oil_function_impl_splat_u32_ns_sse_unroll2, splat_u32_ns;
|
sl@0
|
139 |
}
|
sl@0
|
140 |
#endif
|
sl@0
|
141 |
|
sl@0
|
142 |
#ifdef __SYMBIAN32__
|
sl@0
|
143 |
|
sl@0
|
144 |
OilFunctionImpl* __oil_function_impl_splat_u8_ns_sse, splat_u8_ns() {
|
sl@0
|
145 |
return &_oil_function_impl_splat_u8_ns_sse, splat_u8_ns;
|
sl@0
|
146 |
}
|
sl@0
|
147 |
#endif
|
sl@0
|
148 |
|
sl@0
|
149 |
#ifdef __SYMBIAN32__
|
sl@0
|
150 |
|
sl@0
|
151 |
OilFunctionImpl* __oil_function_impl_splat_u8_ns_sse_unroll2, splat_u8_ns() {
|
sl@0
|
152 |
return &_oil_function_impl_splat_u8_ns_sse_unroll2, splat_u8_ns;
|
sl@0
|
153 |
}
|
sl@0
|
154 |
#endif
|
sl@0
|
155 |
|