| 1 | /* Copyright (C) 2011-2014 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_BROADCAST_W_H |
| 9 | #define LIBSIMDPP_SIMDPP_DETAIL_INSN_BROADCAST_W_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/splat_n.h> |
| 17 | #include <simdpp/detail/extract128.h> |
| 18 | #include <simdpp/detail/insn/shuffle128.h> |
| 19 | #include <simdpp/detail/shuffle/shuffle_mask.h> |
| 20 | |
| 21 | namespace simdpp { |
| 22 | namespace SIMDPP_ARCH_NAMESPACE { |
| 23 | namespace detail { |
| 24 | namespace insn { |
| 25 | |
| 26 | // ----------------------------------------------------------------------------- |
| 27 | |
| 28 | template<unsigned s> SIMDPP_INL |
| 29 | uint8x16 i_splat(const uint8x16& a) |
| 30 | { |
| 31 | return i_splat16<s>(a); |
| 32 | } |
| 33 | |
| 34 | #if SIMDPP_USE_AVX2 |
| 35 | template<unsigned s> SIMDPP_INL |
| 36 | uint8x32 i_splat(const uint8x32& a) |
| 37 | { |
| 38 | static_assert(s < 32, "Access out of bounds" ); |
| 39 | uint8x16 lo; |
| 40 | lo = s < 16 ? detail::extract128<0>(a) : detail::extract128<1>(a); |
| 41 | lo = move16_l<s % 16>(lo); |
| 42 | return _mm256_broadcastb_epi8(lo.native()); |
| 43 | } |
| 44 | #endif |
| 45 | |
| 46 | #if SIMDPP_USE_AVX512BW |
| 47 | template<unsigned s> SIMDPP_INL |
| 48 | uint8<64> i_splat(const uint8<64>& a) |
| 49 | { |
| 50 | static_assert(s < 64, "Access out of bounds" ); |
| 51 | uint8<16> lo; |
| 52 | lo = detail::extract128<s / 16>(a); |
| 53 | lo = move16_l<s % 16>(lo); |
| 54 | return _mm512_broadcastb_epi8(lo.native()); |
| 55 | } |
| 56 | #endif |
| 57 | |
| 58 | // ----------------------------------------------------------------------------- |
| 59 | |
| 60 | template<unsigned s> SIMDPP_INL |
| 61 | uint16x8 i_splat(const uint16x8& a) |
| 62 | { |
| 63 | return i_splat8<s>(a); |
| 64 | } |
| 65 | |
| 66 | #if SIMDPP_USE_AVX2 |
| 67 | template<unsigned s> SIMDPP_INL |
| 68 | uint16x16 i_splat(const uint16x16& a) |
| 69 | { |
| 70 | static_assert(s < 16, "Access out of bounds" ); |
| 71 | uint16x8 lo; |
| 72 | lo = s < 8 ? detail::extract128<0>(a) : detail::extract128<1>(a); |
| 73 | lo = move8_l<s % 8>(lo); |
| 74 | return _mm256_broadcastw_epi16(lo.native()); |
| 75 | } |
| 76 | #endif |
| 77 | |
| 78 | #if SIMDPP_USE_AVX512BW |
| 79 | template<unsigned s> SIMDPP_INL |
| 80 | uint16<32> i_splat(const uint16<32>& a) |
| 81 | { |
| 82 | static_assert(s < 32, "Access out of bounds" ); |
| 83 | uint16<8> lo; |
| 84 | lo = detail::extract128<s / 8>(a); |
| 85 | lo = move8_l<s % 8>(lo); |
| 86 | return _mm512_broadcastw_epi16(lo.native()); |
| 87 | } |
| 88 | #endif |
| 89 | |
| 90 | // ----------------------------------------------------------------------------- |
| 91 | |
| 92 | template<unsigned s> SIMDPP_INL |
| 93 | uint32x4 i_splat(const uint32x4& a) |
| 94 | { |
| 95 | return i_splat4<s>(a); |
| 96 | } |
| 97 | |
| 98 | #if SIMDPP_USE_AVX2 |
| 99 | template<unsigned s> SIMDPP_INL |
| 100 | uint32x8 i_splat(const uint32x8& ca) |
| 101 | { |
| 102 | static_assert(s < 8, "Access out of bounds" ); |
| 103 | uint32<8> a = ca; |
| 104 | a = permute4<s%4,s%4,s%4,s%4>(a); |
| 105 | a = detail::shuffle1_128<s/4, s/4>(a, a); |
| 106 | return a; |
| 107 | } |
| 108 | #endif |
| 109 | |
| 110 | #if SIMDPP_USE_AVX512F |
| 111 | template<unsigned s> SIMDPP_INL |
| 112 | uint32<16> i_splat(const uint32<16>& ca) |
| 113 | { |
| 114 | static_assert(s < 16, "Access out of bounds" ); |
| 115 | uint32<16> a = ca; |
| 116 | a = permute4<s%4,s%4,s%4,s%4>(a); |
| 117 | a = detail::shuffle2_128<s/4, s/4, s/4, s/4>(a, a); |
| 118 | return a; |
| 119 | } |
| 120 | #endif |
| 121 | |
| 122 | // ----------------------------------------------------------------------------- |
| 123 | |
| 124 | template<unsigned s> SIMDPP_INL |
| 125 | uint64x2 i_splat(const uint64x2& a) |
| 126 | { |
| 127 | return i_splat2<s>(a); |
| 128 | } |
| 129 | |
| 130 | #if SIMDPP_USE_AVX2 |
| 131 | template<unsigned s> SIMDPP_INL |
| 132 | uint64x4 i_splat(const uint64x4& a) |
| 133 | { |
| 134 | static_assert(s < 4, "Access out of bounds" ); |
| 135 | return permute4<s,s,s,s>(a); |
| 136 | } |
| 137 | #endif |
| 138 | |
| 139 | #if SIMDPP_USE_AVX512F |
| 140 | template<unsigned s> SIMDPP_INL |
| 141 | uint64<8> i_splat(const uint64<8>& ca) |
| 142 | { |
| 143 | static_assert(s < 8, "Access out of bounds" ); |
| 144 | uint64<8> a = ca; |
| 145 | a = permute2<s%2,s%2>(a); |
| 146 | a = _mm512_shuffle_i64x2(a.native(), a.native(), |
| 147 | SIMDPP_SHUFFLE_MASK_4x4(s/2, s/2, s/2, s/2)); // TODO extract |
| 148 | return a; |
| 149 | } |
| 150 | #endif |
| 151 | |
| 152 | // ----------------------------------------------------------------------------- |
| 153 | |
| 154 | template<unsigned s> SIMDPP_INL |
| 155 | float32x4 i_splat(const float32x4& a) |
| 156 | { |
| 157 | return i_splat4<s>(a); |
| 158 | } |
| 159 | |
| 160 | #if SIMDPP_USE_AVX |
| 161 | template<unsigned s> SIMDPP_INL |
| 162 | float32x8 i_splat(const float32x8& ca) |
| 163 | { |
| 164 | static_assert(s < 8, "Access out of bounds" ); |
| 165 | float32<8> a = ca; |
| 166 | a = shuffle1_128<s/4,s/4>(a, a); |
| 167 | return permute4<s%4,s%4,s%4,s%4>(a); |
| 168 | } |
| 169 | #endif |
| 170 | |
| 171 | #if SIMDPP_USE_AVX512F |
| 172 | template<unsigned s> SIMDPP_INL |
| 173 | float32<16> i_splat(const float32<16>& ca) |
| 174 | { |
| 175 | static_assert(s < 16, "Access out of bounds" ); |
| 176 | float32<16> a = ca; |
| 177 | a = permute4<s%4,s%4,s%4,s%4>(a); |
| 178 | a = _mm512_shuffle_f32x4(a.native(), a.native(), |
| 179 | SIMDPP_SHUFFLE_MASK_4x4(s/4, s/4, s/4, s/4)); |
| 180 | return a; |
| 181 | } |
| 182 | #endif |
| 183 | |
| 184 | // ----------------------------------------------------------------------------- |
| 185 | |
| 186 | template<unsigned s> SIMDPP_INL |
| 187 | float64x2 i_splat(const float64x2& a) |
| 188 | { |
| 189 | return i_splat2<s>(a); |
| 190 | } |
| 191 | |
| 192 | #if SIMDPP_USE_AVX |
| 193 | template<unsigned s> SIMDPP_INL |
| 194 | float64x4 i_splat(const float64x4& a) |
| 195 | { |
| 196 | static_assert(s < 4, "Access out of bounds" ); |
| 197 | #if SIMDPP_USE_AVX2 |
| 198 | return permute4<s,s,s,s>(a); |
| 199 | #else // SIMDPP_USE_AVX |
| 200 | float64<4> b; |
| 201 | b = detail::shuffle1_128<s/2,s/2>(a, a); |
| 202 | b = permute2<s%2,s%2>(b); |
| 203 | return b; |
| 204 | #endif |
| 205 | } |
| 206 | #endif |
| 207 | |
| 208 | #if SIMDPP_USE_AVX512F |
| 209 | template<unsigned s> SIMDPP_INL |
| 210 | float64<8> i_splat(const float64<8>& ca) |
| 211 | { |
| 212 | static_assert(s < 8, "Access out of bounds" ); |
| 213 | float64<8> a = ca; |
| 214 | a = permute2<s%2,s%2>(a); |
| 215 | a = _mm512_shuffle_f64x2(a.native(), a.native(), |
| 216 | SIMDPP_SHUFFLE_MASK_4x4(s/2, s/2, s/2, s/2)); // TODO extract |
| 217 | return a; |
| 218 | } |
| 219 | #endif |
| 220 | |
| 221 | // ----------------------------------------------------------------------------- |
| 222 | |
| 223 | template<unsigned s, class V> SIMDPP_INL |
| 224 | V i_splat(const V& a) |
| 225 | { |
| 226 | static_assert(s < V::length, "Access out of bounds" ); |
| 227 | |
| 228 | using U = typename V::base_vector_type; |
| 229 | U one = a.vec(s / U::length); |
| 230 | |
| 231 | one = i_splat<s % U::length>(one); |
| 232 | |
| 233 | V r; |
| 234 | for (unsigned i = 0; i < V::vec_length; ++i) { |
| 235 | r.vec(i) = one; |
| 236 | } |
| 237 | return r; |
| 238 | } |
| 239 | |
| 240 | } // namespace insn |
| 241 | } // namespace detail |
| 242 | } // namespace SIMDPP_ARCH_NAMESPACE |
| 243 | } // namespace simdpp |
| 244 | |
| 245 | #endif |
| 246 | |
| 247 | |