1 | /* Copyright (C) 2013-2014 Povilas Kanapickas <povilas@radix.lt> |
2 | |
3 | Distributed under the Boost Software License, Version 1.0. |
4 | (See accompanying file LICENSE_1_0.txt or copy at |
5 | http://www.boost.org/LICENSE_1_0.txt) |
6 | */ |
7 | |
8 | #ifndef LIBSIMDPP_SIMDPP_CORE_UNZIP_HI_H |
9 | #define LIBSIMDPP_SIMDPP_CORE_UNZIP_HI_H |
10 | |
11 | #ifndef LIBSIMDPP_SIMD_H |
12 | #error "This file must be included through simd.h" |
13 | #endif |
14 | |
15 | #include <simdpp/types.h> |
16 | #include <simdpp/detail/insn/unzip_hi.h> |
17 | |
18 | namespace simdpp { |
19 | namespace SIMDPP_ARCH_NAMESPACE { |
20 | |
21 | /** De-interleaves the even(higher) elements of two vectors |
22 | |
23 | For example, in case of int8x16: |
24 | |
25 | @code |
26 | | 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 | |
27 | r = [ a1 a3 a5 a7 a9 a11 a13 a15 b1 b3 b5 b7 b9 b11 b13 b15 ] |
28 | @endcode |
29 | |
30 | @par int8 |
31 | |
32 | @par 128-bit version: |
33 | @icost{SSE2-AVX2, 3} |
34 | |
35 | @par 256-bit version: |
36 | The lower and higher 128-bit halves are processed as if 128-bit instruction |
37 | was applied to each of them separately. |
38 | |
39 | @icost{SSE2-AVX, 6} |
40 | @icost{NEON, ALTIVEC, 2} |
41 | @icost{AVX2, 3} |
42 | |
43 | @par int16 |
44 | |
45 | @par 128-bit version: |
46 | @icost{SSE2-AVX2, 3} |
47 | |
48 | @par 256-bit version: |
49 | The lower and higher 128-bit halves are processed as if 128-bit instruction |
50 | was applied to each of them separately. |
51 | |
52 | @icost{SSE2-AVX, 6} |
53 | @icost{NEON, ALTIVEC, 2} |
54 | @icost{AVX2, 3} |
55 | |
56 | @par int32 |
57 | |
58 | @par 128-bit version: |
59 | @icost{ALTIVEC, 1-2} |
60 | |
61 | @par 256-bit version: |
62 | The lower and higher 128-bit halves are processed as if 128-bit instruction |
63 | was applied to each of them separately. |
64 | |
65 | @icost{ALTIVEC, 2-3} |
66 | @icost{SSE2-AVX, NEON, 2} |
67 | |
68 | @par int64 |
69 | |
70 | @par 128-bit version: |
71 | @icost{ALTIVEC, 1-2} |
72 | |
73 | @par 256-bit version: |
74 | The lower and higher 128-bit halves are processed as if 128-bit instruction |
75 | was applied to each of them separately. |
76 | |
77 | @icost{ALTIVEC, 2-3} |
78 | @icost{SSE2-AVX, NEON, 2} |
79 | |
80 | @par float32 |
81 | @par 256-bit version: |
82 | The lower and higher 128-bit halves are processed as if 128-bit instruction |
83 | was applied to each of them separately. |
84 | |
85 | @icost{SSE2-SSE4.1, NEON, ALTIVEC, 2} |
86 | |
87 | @par float64 |
88 | @par 128-bit version: |
89 | @novec{NEON, ALTIVEC} |
90 | |
91 | @par 256-bit version: |
92 | The lower and higher 128-bit halves are processed as if 128-bit instruction |
93 | was applied to each of them separately. |
94 | |
95 | @novec{NEON, ALTIVEC} |
96 | @icost{SSE2-AVX, 2} |
97 | */ |
98 | template<unsigned N, class V1, class V2> SIMDPP_INL |
99 | typename detail::get_expr2_nomask<V1, V2>::empty |
100 | unzip16_hi(const any_vec8<N,V1>& a, const any_vec8<N,V2>& b) |
101 | { |
102 | typename detail::get_expr2_nomask_nosign<V1, V2>::type ra, rb; |
103 | ra = a.wrapped().eval(); |
104 | rb = b.wrapped().eval(); |
105 | return detail::insn::i_unzip16_hi(ra, rb); |
106 | } |
107 | |
108 | template<unsigned N, class V1, class V2> SIMDPP_INL |
109 | typename detail::get_expr2_nomask<V1, V2>::empty |
110 | unzip8_hi(const any_vec16<N,V1>& a, const any_vec16<N,V2>& b) |
111 | { |
112 | typename detail::get_expr2_nomask_nosign<V1, V2>::type ra, rb; |
113 | ra = a.wrapped().eval(); |
114 | rb = b.wrapped().eval(); |
115 | return detail::insn::i_unzip8_hi(ra, rb); |
116 | } |
117 | |
118 | template<unsigned N, class V1, class V2> SIMDPP_INL |
119 | typename detail::get_expr2_nomask<V1, V2>::empty |
120 | unzip4_hi(const any_vec32<N,V1>& a, const any_vec32<N,V2>& b) |
121 | { |
122 | typename detail::get_expr2_nomask_nosign<V1, V2>::type ra, rb; |
123 | ra = a.wrapped().eval(); |
124 | rb = b.wrapped().eval(); |
125 | return detail::insn::i_unzip4_hi(ra, rb); |
126 | } |
127 | |
128 | template<unsigned N, class V1, class V2> SIMDPP_INL |
129 | typename detail::get_expr2_nomask<V1, V2>::empty |
130 | unzip2_hi(const any_vec64<N,V1>& a, const any_vec64<N,V2>& b) |
131 | { |
132 | typename detail::get_expr2_nomask_nosign<V1, V2>::type ra, rb; |
133 | ra = a.wrapped().eval(); |
134 | rb = b.wrapped().eval(); |
135 | return detail::insn::i_unzip2_hi(ra, rb); |
136 | } |
137 | |
138 | } // namespace SIMDPP_ARCH_NAMESPACE |
139 | } // namespace simdpp |
140 | |
141 | #endif |
142 | |
143 | |