Home | History | Annotate | Download | only in NEON
      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 (at) 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_NEON_H
     11 #define EIGEN_COMPLEX_NEON_H
     12 
     13 namespace Eigen {
     14 
     15 namespace internal {
     16 
     17 static uint32x4_t p4ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET4(0x00000000, 0x80000000, 0x00000000, 0x80000000);
     18 static uint32x2_t p2ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET2(0x00000000, 0x80000000);
     19 
     20 //---------- float ----------
     21 struct Packet2cf
     22 {
     23   EIGEN_STRONG_INLINE Packet2cf() {}
     24   EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
     25   Packet4f  v;
     26 };
     27 
     28 template<> struct packet_traits<std::complex<float> >  : default_packet_traits
     29 {
     30   typedef Packet2cf type;
     31   enum {
     32     Vectorizable = 1,
     33     AlignedOnScalar = 1,
     34     size = 2,
     35 
     36     HasAdd    = 1,
     37     HasSub    = 1,
     38     HasMul    = 1,
     39     HasDiv    = 1,
     40     HasNegate = 1,
     41     HasAbs    = 0,
     42     HasAbs2   = 0,
     43     HasMin    = 0,
     44     HasMax    = 0,
     45     HasSetLinear = 0
     46   };
     47 };
     48 
     49 template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; };
     50 
     51 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
     52 {
     53   float32x2_t r64;
     54   r64 = vld1_f32((float *)&from);
     55 
     56   return Packet2cf(vcombine_f32(r64, r64));
     57 }
     58 
     59 template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v,b.v)); }
     60 template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v,b.v)); }
     61 template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate<Packet4f>(a.v)); }
     62 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
     63 {
     64   Packet4ui b = vreinterpretq_u32_f32(a.v);
     65   return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR)));
     66 }
     67 
     68 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
     69 {
     70   Packet4f v1, v2;
     71   float32x2_t a_lo, a_hi;
     72 
     73   // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
     74   v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0));
     75   // Get the real values of a | a1_im | a1_im | a2_im | a2_im |
     76   v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1));
     77   // Multiply the real a with b
     78   v1 = vmulq_f32(v1, b.v);
     79   // Multiply the imag a with b
     80   v2 = vmulq_f32(v2, b.v);
     81   // Conjugate v2
     82   v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR));
     83   // Swap real/imag elements in v2.
     84   a_lo = vrev64_f32(vget_low_f32(v2));
     85   a_hi = vrev64_f32(vget_high_f32(v2));
     86   v2 = vcombine_f32(a_lo, a_hi);
     87   // Add and return the result
     88   return Packet2cf(vaddq_f32(v1, v2));
     89 }
     90 
     91 template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
     92 {
     93   return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
     94 }
     95 template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
     96 {
     97   return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
     98 }
     99 template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
    100 {
    101   return Packet2cf(vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
    102 }
    103 template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
    104 {
    105   return Packet2cf(vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
    106 }
    107 
    108 template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
    109 template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
    110 
    111 template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
    112 
    113 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
    114 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
    115 
    116 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { __pld((float *)addr); }
    117 
    118 template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
    119 {
    120   std::complex<float> EIGEN_ALIGN16 x[2];
    121   vst1q_f32((float *)x, a.v);
    122   return x[0];
    123 }
    124 
    125 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
    126 {
    127   float32x2_t a_lo, a_hi;
    128   Packet4f a_r128;
    129 
    130   a_lo = vget_low_f32(a.v);
    131   a_hi = vget_high_f32(a.v);
    132   a_r128 = vcombine_f32(a_hi, a_lo);
    133 
    134   return Packet2cf(a_r128);
    135 }
    136 
    137 template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a)
    138 {
    139   return Packet2cf(vrev64q_f32(a.v));
    140 }
    141 
    142 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
    143 {
    144   float32x2_t a1, a2;
    145   std::complex<float> s;
    146 
    147   a1 = vget_low_f32(a.v);
    148   a2 = vget_high_f32(a.v);
    149   a2 = vadd_f32(a1, a2);
    150   vst1_f32((float *)&s, a2);
    151 
    152   return s;
    153 }
    154 
    155 template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
    156 {
    157   Packet4f sum1, sum2, sum;
    158 
    159   // Add the first two 64-bit float32x2_t of vecs[0]
    160   sum1 = vcombine_f32(vget_low_f32(vecs[0].v), vget_low_f32(vecs[1].v));
    161   sum2 = vcombine_f32(vget_high_f32(vecs[0].v), vget_high_f32(vecs[1].v));
    162   sum = vaddq_f32(sum1, sum2);
    163 
    164   return Packet2cf(sum);
    165 }
    166 
    167 template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
    168 {
    169   float32x2_t a1, a2, v1, v2, prod;
    170   std::complex<float> s;
    171 
    172   a1 = vget_low_f32(a.v);
    173   a2 = vget_high_f32(a.v);
    174    // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
    175   v1 = vdup_lane_f32(a1, 0);
    176   // Get the real values of a | a1_im | a1_im | a2_im | a2_im |
    177   v2 = vdup_lane_f32(a1, 1);
    178   // Multiply the real a with b
    179   v1 = vmul_f32(v1, a2);
    180   // Multiply the imag a with b
    181   v2 = vmul_f32(v2, a2);
    182   // Conjugate v2
    183   v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR));
    184   // Swap real/imag elements in v2.
    185   v2 = vrev64_f32(v2);
    186   // Add v1, v2
    187   prod = vadd_f32(v1, v2);
    188 
    189   vst1_f32((float *)&s, prod);
    190 
    191   return s;
    192 }
    193 
    194 template<int Offset>
    195 struct palign_impl<Offset,Packet2cf>
    196 {
    197   EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second)
    198   {
    199     if (Offset==1)
    200     {
    201       first.v = vextq_f32(first.v, second.v, 2);
    202     }
    203   }
    204 };
    205 
    206 template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
    207 {
    208   EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
    209   { return padd(pmul(x,y),c); }
    210 
    211   EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
    212   {
    213     return internal::pmul(a, pconj(b));
    214   }
    215 };
    216 
    217 template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
    218 {
    219   EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
    220   { return padd(pmul(x,y),c); }
    221 
    222   EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
    223   {
    224     return internal::pmul(pconj(a), b);
    225   }
    226 };
    227 
    228 template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
    229 {
    230   EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
    231   { return padd(pmul(x,y),c); }
    232 
    233   EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
    234   {
    235     return pconj(internal::pmul(a, b));
    236   }
    237 };
    238 
    239 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
    240 {
    241   // TODO optimize it for AltiVec
    242   Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
    243   Packet4f s, rev_s;
    244   float32x2_t a_lo, a_hi;
    245 
    246   // this computes the norm
    247   s = vmulq_f32(b.v, b.v);
    248   a_lo = vrev64_f32(vget_low_f32(s));
    249   a_hi = vrev64_f32(vget_high_f32(s));
    250   rev_s = vcombine_f32(a_lo, a_hi);
    251 
    252   return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s)));
    253 }
    254 
    255 } // end namespace internal
    256 
    257 } // end namespace Eigen
    258 
    259 #endif // EIGEN_COMPLEX_NEON_H
    260