• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2009 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_SELFADJOINT_MATRIX_VECTOR_H
11 #define EIGEN_SELFADJOINT_MATRIX_VECTOR_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 /* Optimized selfadjoint matrix * vector product:
18  * This algorithm processes 2 columns at onces that allows to both reduce
19  * the number of load/stores of the result by a factor 2 and to reduce
20  * the instruction dependency.
21  */
22 
23 template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version=Specialized>
24 struct selfadjoint_matrix_vector_product;
25 
26 template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version>
27 struct selfadjoint_matrix_vector_product
28 
29 {
30 static EIGEN_DONT_INLINE void run(
31   Index size,
32   const Scalar*  lhs, Index lhsStride,
33   const Scalar*  rhs,
34   Scalar* res,
35   Scalar alpha);
36 };
37 
38 template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version>
run(Index size,const Scalar * lhs,Index lhsStride,const Scalar * rhs,Scalar * res,Scalar alpha)39 EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run(
40   Index size,
41   const Scalar*  lhs, Index lhsStride,
42   const Scalar*  rhs,
43   Scalar* res,
44   Scalar alpha)
45 {
46   typedef typename packet_traits<Scalar>::type Packet;
47   typedef typename NumTraits<Scalar>::Real RealScalar;
48   const Index PacketSize = sizeof(Packet)/sizeof(Scalar);
49 
50   enum {
51     IsRowMajor = StorageOrder==RowMajor ? 1 : 0,
52     IsLower = UpLo == Lower ? 1 : 0,
53     FirstTriangular = IsRowMajor == IsLower
54   };
55 
56   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> cj0;
57   conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> cj1;
58   conj_helper<RealScalar,Scalar,false, ConjugateRhs> cjd;
59 
60   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> pcj0;
61   conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> pcj1;
62 
63   Scalar cjAlpha = ConjugateRhs ? numext::conj(alpha) : alpha;
64 
65 
66   Index bound = (std::max)(Index(0),size-8) & 0xfffffffe;
67   if (FirstTriangular)
68     bound = size - bound;
69 
70   for (Index j=FirstTriangular ? bound : 0;
71        j<(FirstTriangular ? size : bound);j+=2)
72   {
73     const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride;
74     const Scalar* EIGEN_RESTRICT A1 = lhs + (j+1)*lhsStride;
75 
76     Scalar t0 = cjAlpha * rhs[j];
77     Packet ptmp0 = pset1<Packet>(t0);
78     Scalar t1 = cjAlpha * rhs[j+1];
79     Packet ptmp1 = pset1<Packet>(t1);
80 
81     Scalar t2(0);
82     Packet ptmp2 = pset1<Packet>(t2);
83     Scalar t3(0);
84     Packet ptmp3 = pset1<Packet>(t3);
85 
86     Index starti = FirstTriangular ? 0 : j+2;
87     Index endi   = FirstTriangular ? j : size;
88     Index alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti);
89     Index alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
90 
91     res[j]   += cjd.pmul(numext::real(A0[j]), t0);
92     res[j+1] += cjd.pmul(numext::real(A1[j+1]), t1);
93     if(FirstTriangular)
94     {
95       res[j]   += cj0.pmul(A1[j],   t1);
96       t3       += cj1.pmul(A1[j],   rhs[j]);
97     }
98     else
99     {
100       res[j+1] += cj0.pmul(A0[j+1],t0);
101       t2 += cj1.pmul(A0[j+1], rhs[j+1]);
102     }
103 
104     for (Index i=starti; i<alignedStart; ++i)
105     {
106       res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
107       t2 += cj1.pmul(A0[i], rhs[i]);
108       t3 += cj1.pmul(A1[i], rhs[i]);
109     }
110     // Yes this an optimization for gcc 4.3 and 4.4 (=> huge speed up)
111     // gcc 4.2 does this optimization automatically.
112     const Scalar* EIGEN_RESTRICT a0It  = A0  + alignedStart;
113     const Scalar* EIGEN_RESTRICT a1It  = A1  + alignedStart;
114     const Scalar* EIGEN_RESTRICT rhsIt = rhs + alignedStart;
115           Scalar* EIGEN_RESTRICT resIt = res + alignedStart;
116     for (Index i=alignedStart; i<alignedEnd; i+=PacketSize)
117     {
118       Packet A0i = ploadu<Packet>(a0It);  a0It  += PacketSize;
119       Packet A1i = ploadu<Packet>(a1It);  a1It  += PacketSize;
120       Packet Bi  = ploadu<Packet>(rhsIt); rhsIt += PacketSize; // FIXME should be aligned in most cases
121       Packet Xi  = pload <Packet>(resIt);
122 
123       Xi    = pcj0.pmadd(A0i,ptmp0, pcj0.pmadd(A1i,ptmp1,Xi));
124       ptmp2 = pcj1.pmadd(A0i,  Bi, ptmp2);
125       ptmp3 = pcj1.pmadd(A1i,  Bi, ptmp3);
126       pstore(resIt,Xi); resIt += PacketSize;
127     }
128     for (Index i=alignedEnd; i<endi; i++)
129     {
130       res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
131       t2 += cj1.pmul(A0[i], rhs[i]);
132       t3 += cj1.pmul(A1[i], rhs[i]);
133     }
134 
135     res[j]   += alpha * (t2 + predux(ptmp2));
136     res[j+1] += alpha * (t3 + predux(ptmp3));
137   }
138   for (Index j=FirstTriangular ? 0 : bound;j<(FirstTriangular ? bound : size);j++)
139   {
140     const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride;
141 
142     Scalar t1 = cjAlpha * rhs[j];
143     Scalar t2(0);
144     res[j] += cjd.pmul(numext::real(A0[j]), t1);
145     for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++)
146     {
147       res[i] += cj0.pmul(A0[i], t1);
148       t2 += cj1.pmul(A0[i], rhs[i]);
149     }
150     res[j] += alpha * t2;
151   }
152 }
153 
154 } // end namespace internal
155 
156 /***************************************************************************
157 * Wrapper to product_selfadjoint_vector
158 ***************************************************************************/
159 
160 namespace internal {
161 
162 template<typename Lhs, int LhsMode, typename Rhs>
163 struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true>
164 {
165   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
166 
167   typedef internal::blas_traits<Lhs> LhsBlasTraits;
168   typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
169   typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
170 
171   typedef internal::blas_traits<Rhs> RhsBlasTraits;
172   typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
173   typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
174 
175   enum { LhsUpLo = LhsMode&(Upper|Lower) };
176 
177   template<typename Dest>
178   static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
179   {
180     typedef typename Dest::Scalar ResScalar;
181     typedef typename Rhs::Scalar RhsScalar;
182     typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
183 
184     eigen_assert(dest.rows()==a_lhs.rows() && dest.cols()==a_rhs.cols());
185 
186     typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
187     typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
188 
189     Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
190                                * RhsBlasTraits::extractScalarFactor(a_rhs);
191 
192     enum {
193       EvalToDest = (Dest::InnerStrideAtCompileTime==1),
194       UseRhs = (ActualRhsTypeCleaned::InnerStrideAtCompileTime==1)
195     };
196 
197     internal::gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,!EvalToDest> static_dest;
198     internal::gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!UseRhs> static_rhs;
199 
200     ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
201                                                   EvalToDest ? dest.data() : static_dest.data());
202 
203     ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,rhs.size(),
204         UseRhs ? const_cast<RhsScalar*>(rhs.data()) : static_rhs.data());
205 
206     if(!EvalToDest)
207     {
208       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
209       Index size = dest.size();
210       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
211       #endif
212       MappedDest(actualDestPtr, dest.size()) = dest;
213     }
214 
215     if(!UseRhs)
216     {
217       #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
218       Index size = rhs.size();
219       EIGEN_DENSE_STORAGE_CTOR_PLUGIN
220       #endif
221       Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, rhs.size()) = rhs;
222     }
223 
224 
225     internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor,
226                                                 int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run
227       (
228         lhs.rows(),                             // size
229         &lhs.coeffRef(0,0),  lhs.outerStride(), // lhs info
230         actualRhsPtr,                           // rhs info
231         actualDestPtr,                          // result info
232         actualAlpha                             // scale factor
233       );
234 
235     if(!EvalToDest)
236       dest = MappedDest(actualDestPtr, dest.size());
237   }
238 };
239 
240 template<typename Lhs, typename Rhs, int RhsMode>
241 struct selfadjoint_product_impl<Lhs,0,true,Rhs,RhsMode,false>
242 {
243   typedef typename Product<Lhs,Rhs>::Scalar Scalar;
244   enum { RhsUpLo = RhsMode&(Upper|Lower)  };
245 
246   template<typename Dest>
247   static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
248   {
249     // let's simply transpose the product
250     Transpose<Dest> destT(dest);
251     selfadjoint_product_impl<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false,
252                              Transpose<const Lhs>, 0, true>::run(destT, a_rhs.transpose(), a_lhs.transpose(), alpha);
253   }
254 };
255 
256 } // end namespace internal
257 
258 } // end namespace Eigen
259 
260 #endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H
261