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_DETAIL_INSN_SHUFFLE_ZBYTES16_H |
9 | #define LIBSIMDPP_SIMDPP_DETAIL_INSN_SHUFFLE_ZBYTES16_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/core/bit_andnot.h> |
17 | #include <simdpp/core/bit_or.h> |
18 | #include <simdpp/core/blend.h> |
19 | #include <simdpp/core/cmp_lt.h> |
20 | #include <simdpp/core/shuffle_bytes16.h> |
21 | #include <simdpp/core/i_add.h> |
22 | #include <simdpp/detail/not_implemented.h> |
23 | |
24 | namespace simdpp { |
25 | namespace SIMDPP_ARCH_NAMESPACE { |
26 | namespace detail { |
27 | namespace insn { |
28 | |
29 | #if _MSC_VER |
30 | #pragma warning(push) |
31 | #pragma warning(disable: 4800) |
32 | #endif |
33 | |
34 | static SIMDPP_INL |
35 | uint8x16 i_shuffle_zbytes16(const uint8x16& a, const uint8x16& b, const uint8x16& mask) |
36 | { |
37 | #if SIMDPP_USE_NULL |
38 | uint8x16 r; |
39 | |
40 | for (unsigned i = 0; i < 16; i++) { |
41 | unsigned j = mask.el(i) & 0x0f; |
42 | unsigned which = mask.el(i) < 0x10; |
43 | bool zero = mask.el(i) & 0x80; |
44 | r.el(i) = zero ? 0 : (which ? a.el(j) : b.el(j)); |
45 | } |
46 | return r; |
47 | #elif SIMDPP_USE_XOP |
48 | return _mm_perm_epi8(a.native(), b.native(), mask.native()); |
49 | #elif SIMDPP_USE_SSE4_1 |
50 | uint8<16> sel = _mm_slli_epi16(mask.native(), 3); |
51 | |
52 | uint8<16> ai = _mm_shuffle_epi8(a.native(), mask.native()); |
53 | uint8<16> bi = _mm_shuffle_epi8(b.native(), mask.native()); |
54 | return _mm_blendv_epi8(ai.native(), bi.native(), sel.native()); |
55 | #elif SIMDPP_USE_SSSE3 |
56 | mask_int8<16> select_a = cmp_lt((int8<16>) bit_and(mask, 0x1f), 0x10); |
57 | |
58 | uint8<16> ai = _mm_shuffle_epi8(a.native(), mask.native()); |
59 | uint8<16> bi = _mm_shuffle_epi8(b.native(), mask.native()); |
60 | return blend(ai, bi, select_a); |
61 | #elif SIMDPP_USE_NEON |
62 | return shuffle_bytes16(a, b, mask); |
63 | #elif SIMDPP_USE_ALTIVEC |
64 | int8x16 a0 = a, b0 = b; |
65 | int8x16 zero_mask = mask; |
66 | zero_mask = shift_r<7>(zero_mask); // shift in the sign bit |
67 | a0 = i_shuffle_bytes16(a0, b0, mask); |
68 | a0 = bit_andnot(a0, zero_mask); |
69 | return a0; |
70 | #elif SIMDPP_USE_MSA |
71 | return (v16u8) __msa_vshf_b((v16i8) mask.native(), |
72 | (v16i8) b.native(), |
73 | (v16i8) a.native()); |
74 | #else |
75 | return SIMDPP_NOT_IMPLEMENTED3(a, b, mask); |
76 | #endif |
77 | } |
78 | |
79 | #if _MSC_VER |
80 | #pragma warning(pop) |
81 | #endif |
82 | |
83 | #if SIMDPP_USE_AVX2 |
84 | static SIMDPP_INL |
85 | uint8x32 i_shuffle_zbytes16(const uint8x32& a, const uint8x32& b, const uint8x32& mask) |
86 | { |
87 | int8x32 sel, ai, bi, r; |
88 | sel = mask; |
89 | sel = _mm256_slli_epi16(sel.native(), 3); |
90 | |
91 | ai = _mm256_shuffle_epi8(a.native(), mask.native()); |
92 | bi = _mm256_shuffle_epi8(b.native(), mask.native()); |
93 | r = _mm256_blendv_epi8(ai.native(), bi.native(), sel.native()); |
94 | return r; |
95 | } |
96 | #endif |
97 | |
98 | #if SIMDPP_USE_AVX512BW |
99 | SIMDPP_INL uint8<64> i_shuffle_zbytes16(const uint8<64>& a, const uint8<64>& b, const uint8<64>& mask) |
100 | { |
101 | uint8<64> sel_mask, ai, bi, r; |
102 | sel_mask = make_uint(0x10); |
103 | __mmask64 sel = _mm512_test_epi8_mask(mask.native(), sel_mask.native()); |
104 | |
105 | ai = _mm512_shuffle_epi8(a.native(), mask.native()); |
106 | bi = _mm512_shuffle_epi8(b.native(), mask.native()); |
107 | r = _mm512_mask_blend_epi8(sel, ai.native(), bi.native()); |
108 | return r; |
109 | } |
110 | #endif |
111 | |
112 | template<unsigned N> SIMDPP_INL |
113 | uint8<N> i_shuffle_zbytes16(const uint8<N>& a, const uint8<N>& b, const uint8<N>& mask) |
114 | { |
115 | SIMDPP_VEC_ARRAY_IMPL3(uint8<N>, i_shuffle_zbytes16, a, b, mask); |
116 | } |
117 | |
118 | template<unsigned N> SIMDPP_INL |
119 | uint16<N> i_shuffle_zbytes16(const uint16<N>& a, const uint16<N>& b, const uint16<N>& mask) |
120 | { |
121 | return (uint16<N>) i_shuffle_zbytes16(uint8<N*2>(a), uint8<N*2>(b), uint8<N*2>(mask)); |
122 | } |
123 | template<unsigned N> SIMDPP_INL |
124 | uint32<N> i_shuffle_zbytes16(const uint32<N>& a, const uint32<N>& b, const uint32<N>& mask) |
125 | { |
126 | return (uint32<N>) i_shuffle_zbytes16(uint8<N*4>(a), uint8<N*4>(b), uint8<N*4>(mask)); |
127 | } |
128 | template<unsigned N> SIMDPP_INL |
129 | uint64<N> i_shuffle_zbytes16(const uint64<N>& a, const uint64<N>& b, const uint64<N>& mask) |
130 | { |
131 | return (uint64<N>) i_shuffle_zbytes16(uint8<N*8>(a), uint8<N*8>(b), uint8<N*8>(mask)); |
132 | } |
133 | template<unsigned N> SIMDPP_INL |
134 | float32<N> i_shuffle_zbytes16(const float32<N>& a, const float32<N>& b, const uint32<N>& mask) |
135 | { |
136 | return float32<N>(i_shuffle_zbytes16(uint32<N>(a), uint32<N>(b), mask)); |
137 | } |
138 | template<unsigned N> SIMDPP_INL |
139 | float64<N> i_shuffle_zbytes16(const float64<N>& a, const float64<N>& b, const uint64<N>& mask) |
140 | { |
141 | return float64<N>(i_shuffle_zbytes16(uint64<N>(a), uint64<N>(b), mask)); |
142 | } |
143 | |
144 | |
145 | } // namespace insn |
146 | } // namespace detail |
147 | } // namespace SIMDPP_ARCH_NAMESPACE |
148 | } // namespace simdpp |
149 | |
150 | #endif |
151 | |
152 | |