| 1 | // This file is part of Eigen, a lightweight C++ template library
|
|---|
| 2 | // for linear algebra.
|
|---|
| 3 | //
|
|---|
| 4 | // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
|
|---|
| 5 | //
|
|---|
| 6 | // This Source Code Form is subject to the terms of the Mozilla
|
|---|
| 7 | // Public License v. 2.0. If a copy of the MPL was not distributed
|
|---|
| 8 | // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
|
|---|
| 9 |
|
|---|
| 10 | #ifndef EIGEN_COMPLEX_SSE_H
|
|---|
| 11 | #define EIGEN_COMPLEX_SSE_H
|
|---|
| 12 |
|
|---|
| 13 | namespace Eigen {
|
|---|
| 14 |
|
|---|
| 15 | namespace internal {
|
|---|
| 16 |
|
|---|
| 17 | //---------- float ----------
|
|---|
| 18 | struct Packet2cf
|
|---|
| 19 | {
|
|---|
| 20 | EIGEN_STRONG_INLINE Packet2cf() {}
|
|---|
| 21 | EIGEN_STRONG_INLINE explicit Packet2cf(const __m128& a) : v(a) {}
|
|---|
| 22 | __m128 v;
|
|---|
| 23 | };
|
|---|
| 24 |
|
|---|
| 25 | template<> struct packet_traits<std::complex<float> > : default_packet_traits
|
|---|
| 26 | {
|
|---|
| 27 | typedef Packet2cf type;
|
|---|
| 28 | enum {
|
|---|
| 29 | Vectorizable = 1,
|
|---|
| 30 | AlignedOnScalar = 1,
|
|---|
| 31 | size = 2,
|
|---|
| 32 |
|
|---|
| 33 | HasAdd = 1,
|
|---|
| 34 | HasSub = 1,
|
|---|
| 35 | HasMul = 1,
|
|---|
| 36 | HasDiv = 1,
|
|---|
| 37 | HasNegate = 1,
|
|---|
| 38 | HasAbs = 0,
|
|---|
| 39 | HasAbs2 = 0,
|
|---|
| 40 | HasMin = 0,
|
|---|
| 41 | HasMax = 0,
|
|---|
| 42 | HasSetLinear = 0
|
|---|
| 43 | };
|
|---|
| 44 | };
|
|---|
| 45 |
|
|---|
| 46 | template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; };
|
|---|
| 47 |
|
|---|
| 48 | template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); }
|
|---|
| 49 | template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); }
|
|---|
| 50 | template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a)
|
|---|
| 51 | {
|
|---|
| 52 | const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
|
|---|
| 53 | return Packet2cf(_mm_xor_ps(a.v,mask));
|
|---|
| 54 | }
|
|---|
| 55 | template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
|
|---|
| 56 | {
|
|---|
| 57 | const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
|
|---|
| 58 | return Packet2cf(_mm_xor_ps(a.v,mask));
|
|---|
| 59 | }
|
|---|
| 60 |
|
|---|
| 61 | template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
|
|---|
| 62 | {
|
|---|
| 63 | // TODO optimize it for SSE3 and 4
|
|---|
| 64 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 65 | return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
|
|---|
| 66 | _mm_mul_ps(_mm_movehdup_ps(a.v),
|
|---|
| 67 | vec4f_swizzle1(b.v, 1, 0, 3, 2))));
|
|---|
| 68 | // return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
|
|---|
| 69 | // _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
|
|---|
| 70 | // vec4f_swizzle1(b.v, 1, 0, 3, 2))));
|
|---|
| 71 | #else
|
|---|
| 72 | const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000));
|
|---|
| 73 | return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
|
|---|
| 74 | _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
|
|---|
| 75 | vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
|
|---|
| 76 | #endif
|
|---|
| 77 | }
|
|---|
| 78 |
|
|---|
| 79 | template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); }
|
|---|
| 80 | template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); }
|
|---|
| 81 | template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); }
|
|---|
| 82 | template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); }
|
|---|
| 83 |
|
|---|
| 84 | template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&numext::real_ref(*from))); }
|
|---|
| 85 | template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&numext::real_ref(*from))); }
|
|---|
| 86 |
|
|---|
| 87 | template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from)
|
|---|
| 88 | {
|
|---|
| 89 | Packet2cf res;
|
|---|
| 90 | #if EIGEN_GNUC_AT_MOST(4,2)
|
|---|
| 91 | // Workaround annoying "may be used uninitialized in this function" warning with gcc 4.2
|
|---|
| 92 | res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from));
|
|---|
| 93 | #elif EIGEN_GNUC_AT_LEAST(4,6)
|
|---|
| 94 | // Suppress annoying "may be used uninitialized in this function" warning with gcc >= 4.6
|
|---|
| 95 | #pragma GCC diagnostic push
|
|---|
| 96 | #pragma GCC diagnostic ignored "-Wuninitialized"
|
|---|
| 97 | res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
|
|---|
| 98 | #pragma GCC diagnostic pop
|
|---|
| 99 | #else
|
|---|
| 100 | res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
|
|---|
| 101 | #endif
|
|---|
| 102 | return Packet2cf(_mm_movelh_ps(res.v,res.v));
|
|---|
| 103 | }
|
|---|
| 104 |
|
|---|
| 105 | template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
|
|---|
| 106 |
|
|---|
| 107 | template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); }
|
|---|
| 108 | template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); }
|
|---|
| 109 |
|
|---|
| 110 | template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
|
|---|
| 111 |
|
|---|
| 112 | template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a)
|
|---|
| 113 | {
|
|---|
| 114 | #if EIGEN_GNUC_AT_MOST(4,3)
|
|---|
| 115 | // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares...
|
|---|
| 116 | // This workaround also fix invalid code generation with gcc 4.3
|
|---|
| 117 | EIGEN_ALIGN16 std::complex<float> res[2];
|
|---|
| 118 | _mm_store_ps((float*)res, a.v);
|
|---|
| 119 | return res[0];
|
|---|
| 120 | #else
|
|---|
| 121 | std::complex<float> res;
|
|---|
| 122 | _mm_storel_pi((__m64*)&res, a.v);
|
|---|
| 123 | return res;
|
|---|
| 124 | #endif
|
|---|
| 125 | }
|
|---|
| 126 |
|
|---|
| 127 | template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(_mm_castps_pd(a.v)))); }
|
|---|
| 128 |
|
|---|
| 129 | template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
|
|---|
| 130 | {
|
|---|
| 131 | return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v))));
|
|---|
| 132 | }
|
|---|
| 133 |
|
|---|
| 134 | template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
|
|---|
| 135 | {
|
|---|
| 136 | return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v)));
|
|---|
| 137 | }
|
|---|
| 138 |
|
|---|
| 139 | template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
|
|---|
| 140 | {
|
|---|
| 141 | return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v))));
|
|---|
| 142 | }
|
|---|
| 143 |
|
|---|
| 144 | template<int Offset>
|
|---|
| 145 | struct palign_impl<Offset,Packet2cf>
|
|---|
| 146 | {
|
|---|
| 147 | static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
|
|---|
| 148 | {
|
|---|
| 149 | if (Offset==1)
|
|---|
| 150 | {
|
|---|
| 151 | first.v = _mm_movehl_ps(first.v, first.v);
|
|---|
| 152 | first.v = _mm_movelh_ps(first.v, second.v);
|
|---|
| 153 | }
|
|---|
| 154 | }
|
|---|
| 155 | };
|
|---|
| 156 |
|
|---|
| 157 | template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
|
|---|
| 158 | {
|
|---|
| 159 | EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
|
|---|
| 160 | { return padd(pmul(x,y),c); }
|
|---|
| 161 |
|
|---|
| 162 | EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
|
|---|
| 163 | {
|
|---|
| 164 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 165 | return internal::pmul(a, pconj(b));
|
|---|
| 166 | #else
|
|---|
| 167 | const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
|
|---|
| 168 | return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
|
|---|
| 169 | _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
|
|---|
| 170 | vec4f_swizzle1(b.v, 1, 0, 3, 2))));
|
|---|
| 171 | #endif
|
|---|
| 172 | }
|
|---|
| 173 | };
|
|---|
| 174 |
|
|---|
| 175 | template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
|
|---|
| 176 | {
|
|---|
| 177 | EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
|
|---|
| 178 | { return padd(pmul(x,y),c); }
|
|---|
| 179 |
|
|---|
| 180 | EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
|
|---|
| 181 | {
|
|---|
| 182 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 183 | return internal::pmul(pconj(a), b);
|
|---|
| 184 | #else
|
|---|
| 185 | const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
|
|---|
| 186 | return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
|
|---|
| 187 | _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
|
|---|
| 188 | vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
|
|---|
| 189 | #endif
|
|---|
| 190 | }
|
|---|
| 191 | };
|
|---|
| 192 |
|
|---|
| 193 | template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
|
|---|
| 194 | {
|
|---|
| 195 | EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
|
|---|
| 196 | { return padd(pmul(x,y),c); }
|
|---|
| 197 |
|
|---|
| 198 | EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
|
|---|
| 199 | {
|
|---|
| 200 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 201 | return pconj(internal::pmul(a, b));
|
|---|
| 202 | #else
|
|---|
| 203 | const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
|
|---|
| 204 | return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
|
|---|
| 205 | _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
|
|---|
| 206 | vec4f_swizzle1(b.v, 1, 0, 3, 2))));
|
|---|
| 207 | #endif
|
|---|
| 208 | }
|
|---|
| 209 | };
|
|---|
| 210 |
|
|---|
| 211 | template<> struct conj_helper<Packet4f, Packet2cf, false,false>
|
|---|
| 212 | {
|
|---|
| 213 | EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
|
|---|
| 214 | { return padd(c, pmul(x,y)); }
|
|---|
| 215 |
|
|---|
| 216 | EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
|
|---|
| 217 | { return Packet2cf(Eigen::internal::pmul(x, y.v)); }
|
|---|
| 218 | };
|
|---|
| 219 |
|
|---|
| 220 | template<> struct conj_helper<Packet2cf, Packet4f, false,false>
|
|---|
| 221 | {
|
|---|
| 222 | EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
|
|---|
| 223 | { return padd(c, pmul(x,y)); }
|
|---|
| 224 |
|
|---|
| 225 | EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
|
|---|
| 226 | { return Packet2cf(Eigen::internal::pmul(x.v, y)); }
|
|---|
| 227 | };
|
|---|
| 228 |
|
|---|
| 229 | template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
|
|---|
| 230 | {
|
|---|
| 231 | // TODO optimize it for SSE3 and 4
|
|---|
| 232 | Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
|
|---|
| 233 | __m128 s = _mm_mul_ps(b.v,b.v);
|
|---|
| 234 | return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1)))));
|
|---|
| 235 | }
|
|---|
| 236 |
|
|---|
| 237 | EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
|
|---|
| 238 | {
|
|---|
| 239 | return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
|
|---|
| 240 | }
|
|---|
| 241 |
|
|---|
| 242 |
|
|---|
| 243 | //---------- double ----------
|
|---|
| 244 | struct Packet1cd
|
|---|
| 245 | {
|
|---|
| 246 | EIGEN_STRONG_INLINE Packet1cd() {}
|
|---|
| 247 | EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {}
|
|---|
| 248 | __m128d v;
|
|---|
| 249 | };
|
|---|
| 250 |
|
|---|
| 251 | template<> struct packet_traits<std::complex<double> > : default_packet_traits
|
|---|
| 252 | {
|
|---|
| 253 | typedef Packet1cd type;
|
|---|
| 254 | enum {
|
|---|
| 255 | Vectorizable = 1,
|
|---|
| 256 | AlignedOnScalar = 0,
|
|---|
| 257 | size = 1,
|
|---|
| 258 |
|
|---|
| 259 | HasAdd = 1,
|
|---|
| 260 | HasSub = 1,
|
|---|
| 261 | HasMul = 1,
|
|---|
| 262 | HasDiv = 1,
|
|---|
| 263 | HasNegate = 1,
|
|---|
| 264 | HasAbs = 0,
|
|---|
| 265 | HasAbs2 = 0,
|
|---|
| 266 | HasMin = 0,
|
|---|
| 267 | HasMax = 0,
|
|---|
| 268 | HasSetLinear = 0
|
|---|
| 269 | };
|
|---|
| 270 | };
|
|---|
| 271 |
|
|---|
| 272 | template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1}; };
|
|---|
| 273 |
|
|---|
| 274 | template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); }
|
|---|
| 275 | template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); }
|
|---|
| 276 | template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(a.v)); }
|
|---|
| 277 | template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
|
|---|
| 278 | {
|
|---|
| 279 | const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
|
|---|
| 280 | return Packet1cd(_mm_xor_pd(a.v,mask));
|
|---|
| 281 | }
|
|---|
| 282 |
|
|---|
| 283 | template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
|
|---|
| 284 | {
|
|---|
| 285 | // TODO optimize it for SSE3 and 4
|
|---|
| 286 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 287 | return Packet1cd(_mm_addsub_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
|
|---|
| 288 | _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
|
|---|
| 289 | vec2d_swizzle1(b.v, 1, 0))));
|
|---|
| 290 | #else
|
|---|
| 291 | const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
|
|---|
| 292 | return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
|
|---|
| 293 | _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
|
|---|
| 294 | vec2d_swizzle1(b.v, 1, 0)), mask)));
|
|---|
| 295 | #endif
|
|---|
| 296 | }
|
|---|
| 297 |
|
|---|
| 298 | template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); }
|
|---|
| 299 | template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); }
|
|---|
| 300 | template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); }
|
|---|
| 301 | template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); }
|
|---|
| 302 |
|
|---|
| 303 | // FIXME force unaligned load, this is a temporary fix
|
|---|
| 304 | template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from)
|
|---|
| 305 | { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
|
|---|
| 306 | template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from)
|
|---|
| 307 | { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
|
|---|
| 308 | template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from)
|
|---|
| 309 | { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
|
|---|
| 310 |
|
|---|
| 311 | template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
|
|---|
| 312 |
|
|---|
| 313 | // FIXME force unaligned store, this is a temporary fix
|
|---|
| 314 | template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
|
|---|
| 315 | template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
|
|---|
| 316 |
|
|---|
| 317 | template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
|
|---|
| 318 |
|
|---|
| 319 | template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a)
|
|---|
| 320 | {
|
|---|
| 321 | EIGEN_ALIGN16 double res[2];
|
|---|
| 322 | _mm_store_pd(res, a.v);
|
|---|
| 323 | return std::complex<double>(res[0],res[1]);
|
|---|
| 324 | }
|
|---|
| 325 |
|
|---|
| 326 | template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
|
|---|
| 327 |
|
|---|
| 328 | template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
|
|---|
| 329 | {
|
|---|
| 330 | return pfirst(a);
|
|---|
| 331 | }
|
|---|
| 332 |
|
|---|
| 333 | template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
|
|---|
| 334 | {
|
|---|
| 335 | return vecs[0];
|
|---|
| 336 | }
|
|---|
| 337 |
|
|---|
| 338 | template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
|
|---|
| 339 | {
|
|---|
| 340 | return pfirst(a);
|
|---|
| 341 | }
|
|---|
| 342 |
|
|---|
| 343 | template<int Offset>
|
|---|
| 344 | struct palign_impl<Offset,Packet1cd>
|
|---|
| 345 | {
|
|---|
| 346 | static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
|
|---|
| 347 | {
|
|---|
| 348 | // FIXME is it sure we never have to align a Packet1cd?
|
|---|
| 349 | // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
|
|---|
| 350 | }
|
|---|
| 351 | };
|
|---|
| 352 |
|
|---|
| 353 | template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
|
|---|
| 354 | {
|
|---|
| 355 | EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
|
|---|
| 356 | { return padd(pmul(x,y),c); }
|
|---|
| 357 |
|
|---|
| 358 | EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
|
|---|
| 359 | {
|
|---|
| 360 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 361 | return internal::pmul(a, pconj(b));
|
|---|
| 362 | #else
|
|---|
| 363 | const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
|
|---|
| 364 | return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
|
|---|
| 365 | _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
|
|---|
| 366 | vec2d_swizzle1(b.v, 1, 0))));
|
|---|
| 367 | #endif
|
|---|
| 368 | }
|
|---|
| 369 | };
|
|---|
| 370 |
|
|---|
| 371 | template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
|
|---|
| 372 | {
|
|---|
| 373 | EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
|
|---|
| 374 | { return padd(pmul(x,y),c); }
|
|---|
| 375 |
|
|---|
| 376 | EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
|
|---|
| 377 | {
|
|---|
| 378 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 379 | return internal::pmul(pconj(a), b);
|
|---|
| 380 | #else
|
|---|
| 381 | const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
|
|---|
| 382 | return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
|
|---|
| 383 | _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
|
|---|
| 384 | vec2d_swizzle1(b.v, 1, 0)), mask)));
|
|---|
| 385 | #endif
|
|---|
| 386 | }
|
|---|
| 387 | };
|
|---|
| 388 |
|
|---|
| 389 | template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
|
|---|
| 390 | {
|
|---|
| 391 | EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
|
|---|
| 392 | { return padd(pmul(x,y),c); }
|
|---|
| 393 |
|
|---|
| 394 | EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
|
|---|
| 395 | {
|
|---|
| 396 | #ifdef EIGEN_VECTORIZE_SSE3
|
|---|
| 397 | return pconj(internal::pmul(a, b));
|
|---|
| 398 | #else
|
|---|
| 399 | const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
|
|---|
| 400 | return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
|
|---|
| 401 | _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
|
|---|
| 402 | vec2d_swizzle1(b.v, 1, 0))));
|
|---|
| 403 | #endif
|
|---|
| 404 | }
|
|---|
| 405 | };
|
|---|
| 406 |
|
|---|
| 407 | template<> struct conj_helper<Packet2d, Packet1cd, false,false>
|
|---|
| 408 | {
|
|---|
| 409 | EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
|
|---|
| 410 | { return padd(c, pmul(x,y)); }
|
|---|
| 411 |
|
|---|
| 412 | EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
|
|---|
| 413 | { return Packet1cd(Eigen::internal::pmul(x, y.v)); }
|
|---|
| 414 | };
|
|---|
| 415 |
|
|---|
| 416 | template<> struct conj_helper<Packet1cd, Packet2d, false,false>
|
|---|
| 417 | {
|
|---|
| 418 | EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
|
|---|
| 419 | { return padd(c, pmul(x,y)); }
|
|---|
| 420 |
|
|---|
| 421 | EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
|
|---|
| 422 | { return Packet1cd(Eigen::internal::pmul(x.v, y)); }
|
|---|
| 423 | };
|
|---|
| 424 |
|
|---|
| 425 | template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
|
|---|
| 426 | {
|
|---|
| 427 | // TODO optimize it for SSE3 and 4
|
|---|
| 428 | Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
|
|---|
| 429 | __m128d s = _mm_mul_pd(b.v,b.v);
|
|---|
| 430 | return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1))));
|
|---|
| 431 | }
|
|---|
| 432 |
|
|---|
| 433 | EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
|
|---|
| 434 | {
|
|---|
| 435 | return Packet1cd(preverse(x.v));
|
|---|
| 436 | }
|
|---|
| 437 |
|
|---|
| 438 | } // end namespace internal
|
|---|
| 439 |
|
|---|
| 440 | } // end namespace Eigen
|
|---|
| 441 |
|
|---|
| 442 | #endif // EIGEN_COMPLEX_SSE_H
|
|---|
