| 1 | /* Copyright (C) 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_MAKE_UINT_H |
| 9 | #define LIBSIMDPP_SIMDPP_CORE_MAKE_UINT_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/make_const.h> |
| 17 | |
| 18 | namespace simdpp { |
| 19 | namespace SIMDPP_ARCH_NAMESPACE { |
| 20 | |
| 21 | /** Creates a vector from unsigned integer values known at compile-time. |
| 22 | The result of this function may be assigned or converted to a vector of any |
| 23 | type: standard conversions are used to convert the arguments. All |
| 24 | conversions and other overhead is performed at compile-time thus even if the |
| 25 | minimal optimization level is selected, the function results in a simple |
| 26 | load from memory. |
| 27 | |
| 28 | The function is not guaranteed to have adequate performance if the |
| 29 | arguments are not known at compile-time. |
| 30 | |
| 31 | If the vector has fewer elements than the number of the parameters this |
| 32 | function accepts then the extra values are discarded. |
| 33 | |
| 34 | Note that per C++ rules negative values are sign-extended to fill entire |
| 35 | element before being converted to unsigned type thus e.g. it's safe to use |
| 36 | -1 to fill element with ones. |
| 37 | |
| 38 | @par 1 parameter version |
| 39 | @code |
| 40 | | 0 1 2 3 ... n | |
| 41 | r = [ v0 v0 v0 v0 ... v0 ] |
| 42 | @endcode |
| 43 | |
| 44 | @par 2 parameters version |
| 45 | @code |
| 46 | | 0 1 2 3 ... n | |
| 47 | r = [ v0 v1 v0 v1 ... v1 ] |
| 48 | @endcode |
| 49 | |
| 50 | @par 4 parameters version |
| 51 | @code |
| 52 | | 0 1 2 3 ... n | |
| 53 | r = [ v0 v1 v2 v3 ... v3 ] |
| 54 | @endcode |
| 55 | |
| 56 | @par 8 parameters version |
| 57 | @code |
| 58 | | 0 1 .. 7 8 ... n | |
| 59 | r = [ v0 v1 .. v7 v0 ... v7 ] |
| 60 | @endcode |
| 61 | */ |
| 62 | // FIXME: return empty expr |
| 63 | SIMDPP_INL expr_vec_make_const<uint64_t,1> make_uint(uint64_t v0) |
| 64 | { |
| 65 | expr_vec_make_const<uint64_t,1> a; |
| 66 | a.a[0] = v0; |
| 67 | return a; |
| 68 | } |
| 69 | |
| 70 | static SIMDPP_INL |
| 71 | expr_vec_make_const<uint64_t,2> make_uint(uint64_t v0, uint64_t v1) |
| 72 | { |
| 73 | expr_vec_make_const<uint64_t,2> a; |
| 74 | a.a[0] = v0; a.a[1] = v1; |
| 75 | return a; |
| 76 | } |
| 77 | |
| 78 | static SIMDPP_INL |
| 79 | expr_vec_make_const<uint64_t,4> |
| 80 | make_uint(uint64_t v0, uint64_t v1, uint64_t v2, uint64_t v3) |
| 81 | { |
| 82 | expr_vec_make_const<uint64_t,4> a; |
| 83 | a.a[0] = v0; a.a[1] = v1; a.a[2] = v2; a.a[3] = v3; |
| 84 | return a; |
| 85 | } |
| 86 | |
| 87 | static SIMDPP_INL |
| 88 | expr_vec_make_const<uint64_t,8> |
| 89 | make_uint(uint64_t v0, uint64_t v1, uint64_t v2, uint64_t v3, |
| 90 | uint64_t v4, uint64_t v5, uint64_t v6, uint64_t v7) |
| 91 | { |
| 92 | expr_vec_make_const<uint64_t,8> a; |
| 93 | a.a[0] = v0; a.a[1] = v1; a.a[2] = v2; a.a[3] = v3; |
| 94 | a.a[4] = v4; a.a[5] = v5; a.a[6] = v6; a.a[7] = v7; |
| 95 | return a; |
| 96 | } |
| 97 | |
| 98 | static SIMDPP_INL |
| 99 | expr_vec_make_const<uint64_t,16> |
| 100 | make_uint(uint64_t v0, uint64_t v1, uint64_t v2, uint64_t v3, |
| 101 | uint64_t v4, uint64_t v5, uint64_t v6, uint64_t v7, |
| 102 | uint64_t v8, uint64_t v9, uint64_t v10, uint64_t v11, |
| 103 | uint64_t v12, uint64_t v13, uint64_t v14, uint64_t v15) |
| 104 | { |
| 105 | expr_vec_make_const<uint64_t,16> a; |
| 106 | a.a[0] = v0; a.a[1] = v1; a.a[2] = v2; a.a[3] = v3; |
| 107 | a.a[4] = v4; a.a[5] = v5; a.a[6] = v6; a.a[7] = v7; |
| 108 | a.a[8] = v8; a.a[9] = v9; a.a[10] = v10; a.a[11] = v11; |
| 109 | a.a[12] = v12; a.a[13] = v13; a.a[14] = v14; a.a[15] = v15; |
| 110 | return a; |
| 111 | } |
| 112 | |
| 113 | template<class V> SIMDPP_INL |
| 114 | V make_uint(uint64_t v0) |
| 115 | { |
| 116 | static_assert(is_vector<V>::value && !is_mask<V>::value, |
| 117 | "V must be a non-mask vector" ); |
| 118 | expr_vec_make_const<uint64_t,1> a; |
| 119 | a.a[0] = v0; |
| 120 | return detail::insn::i_make_const_any<V>(a); |
| 121 | } |
| 122 | |
| 123 | template<class V> SIMDPP_INL |
| 124 | V make_uint(uint64_t v0, uint64_t v1) |
| 125 | { |
| 126 | static_assert(is_vector<V>::value && !is_mask<V>::value, |
| 127 | "V must be a non-mask vector" ); |
| 128 | expr_vec_make_const<uint64_t,2> a; |
| 129 | a.a[0] = v0; a.a[1] = v1; |
| 130 | return detail::insn::i_make_const_any<V>(a); |
| 131 | } |
| 132 | |
| 133 | template<class V> SIMDPP_INL |
| 134 | V make_uint(uint64_t v0, uint64_t v1, uint64_t v2, uint64_t v3) |
| 135 | { |
| 136 | static_assert(is_vector<V>::value && !is_mask<V>::value, |
| 137 | "V must be a non-mask vector" ); |
| 138 | expr_vec_make_const<uint64_t,4> a; |
| 139 | a.a[0] = v0; a.a[1] = v1; a.a[2] = v2; a.a[3] = v3; |
| 140 | return detail::insn::i_make_const_any<V>(a); |
| 141 | } |
| 142 | |
| 143 | template<class V> SIMDPP_INL |
| 144 | V make_uint(uint64_t v0, uint64_t v1, uint64_t v2, uint64_t v3, |
| 145 | uint64_t v4, uint64_t v5, uint64_t v6, uint64_t v7) |
| 146 | { |
| 147 | static_assert(is_vector<V>::value && !is_mask<V>::value, |
| 148 | "V must be a non-mask vector" ); |
| 149 | expr_vec_make_const<uint64_t,8> a; |
| 150 | a.a[0] = v0; a.a[1] = v1; a.a[2] = v2; a.a[3] = v3; |
| 151 | a.a[4] = v4; a.a[5] = v5; a.a[6] = v6; a.a[7] = v7; |
| 152 | return detail::insn::i_make_const_any<V>(a); |
| 153 | } |
| 154 | |
| 155 | template<class V> SIMDPP_INL |
| 156 | V make_uint(uint64_t v0, uint64_t v1, uint64_t v2, uint64_t v3, |
| 157 | uint64_t v4, uint64_t v5, uint64_t v6, uint64_t v7, |
| 158 | uint64_t v8, uint64_t v9, uint64_t v10, uint64_t v11, |
| 159 | uint64_t v12, uint64_t v13, uint64_t v14, uint64_t v15) |
| 160 | { |
| 161 | static_assert(is_vector<V>::value && !is_mask<V>::value, |
| 162 | "V must be a non-mask vector" ); |
| 163 | expr_vec_make_const<uint64_t,16> a; |
| 164 | a.a[0] = v0; a.a[1] = v1; a.a[2] = v2; a.a[3] = v3; |
| 165 | a.a[4] = v4; a.a[5] = v5; a.a[6] = v6; a.a[7] = v7; |
| 166 | a.a[8] = v8; a.a[9] = v9; a.a[10] = v10; a.a[11] = v11; |
| 167 | a.a[12] = v12; a.a[13] = v13; a.a[14] = v14; a.a[15] = v15; |
| 168 | return detail::insn::i_make_const_any<V>(a); |
| 169 | } |
| 170 | |
| 171 | /// Creates a vector initialized to zero |
| 172 | SIMDPP_INL expr_vec_make_const<uint64_t,1> make_zero() |
| 173 | { |
| 174 | return make_uint(0); |
| 175 | } |
| 176 | |
| 177 | template<class V> SIMDPP_INL |
| 178 | V make_zero() |
| 179 | { |
| 180 | return make_uint<V>(0); |
| 181 | } |
| 182 | |
| 183 | /// Creates a vector initialized to ones |
| 184 | SIMDPP_INL expr_vec_make_ones make_ones() |
| 185 | { |
| 186 | return expr_vec_make_ones(); |
| 187 | } |
| 188 | |
| 189 | template<class V> SIMDPP_INL |
| 190 | V make_ones() |
| 191 | { |
| 192 | return (V) make_ones(); |
| 193 | } |
| 194 | |
| 195 | } // namespace SIMDPP_ARCH_NAMESPACE |
| 196 | } // namespace simdpp |
| 197 | |
| 198 | #endif |
| 199 | |
| 200 | |