• 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) 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_TRIANGULAR_SOLVER_MATRIX_H
11 #define EIGEN_TRIANGULAR_SOLVER_MATRIX_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 // if the rhs is row major, let's transpose the product
18 template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
19 struct triangular_solve_matrix<Scalar,Index,Side,Mode,Conjugate,TriStorageOrder,RowMajor,OtherInnerStride>
20 {
21   static void run(
22     Index size, Index cols,
23     const Scalar*  tri, Index triStride,
24     Scalar* _other, Index otherIncr, Index otherStride,
25     level3_blocking<Scalar,Scalar>& blocking)
26   {
27     triangular_solve_matrix<
28       Scalar, Index, Side==OnTheLeft?OnTheRight:OnTheLeft,
29       (Mode&UnitDiag) | ((Mode&Upper) ? Lower : Upper),
30       NumTraits<Scalar>::IsComplex && Conjugate,
31       TriStorageOrder==RowMajor ? ColMajor : RowMajor, ColMajor, OtherInnerStride>
32       ::run(size, cols, tri, triStride, _other, otherIncr, otherStride, blocking);
33   }
34 };
35 
36 /* Optimized triangular solver with multiple right hand side and the triangular matrix on the left
37  */
38 template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder,int OtherInnerStride>
39 struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>
40 {
41   static EIGEN_DONT_INLINE void run(
42     Index size, Index otherSize,
43     const Scalar* _tri, Index triStride,
44     Scalar* _other, Index otherIncr, Index otherStride,
45     level3_blocking<Scalar,Scalar>& blocking);
46 };
47 template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
48 EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>::run(
49     Index size, Index otherSize,
50     const Scalar* _tri, Index triStride,
51     Scalar* _other, Index otherIncr, Index otherStride,
52     level3_blocking<Scalar,Scalar>& blocking)
53   {
54     Index cols = otherSize;
55 
56     typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> TriMapper;
57     typedef blas_data_mapper<Scalar, Index, ColMajor, Unaligned, OtherInnerStride> OtherMapper;
58     TriMapper tri(_tri, triStride);
59     OtherMapper other(_other, otherStride, otherIncr);
60 
61     typedef gebp_traits<Scalar,Scalar> Traits;
62 
63     enum {
64       SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
65       IsLower = (Mode&Lower) == Lower
66     };
67 
68     Index kc = blocking.kc();                   // cache block size along the K direction
69     Index mc = (std::min)(size,blocking.mc());  // cache block size along the M direction
70 
71     std::size_t sizeA = kc*mc;
72     std::size_t sizeB = kc*cols;
73 
74     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
75     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
76 
77     conj_if<Conjugate> conj;
78     gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
79     gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, TriStorageOrder> pack_lhs;
80     gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs;
81 
82     // the goal here is to subdivise the Rhs panels such that we keep some cache
83     // coherence when accessing the rhs elements
84     std::ptrdiff_t l1, l2, l3;
85     manage_caching_sizes(GetAction, &l1, &l2, &l3);
86     Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) * std::max<Index>(otherStride,size)) : 0;
87     subcols = std::max<Index>((subcols/Traits::nr)*Traits::nr, Traits::nr);
88 
89     for(Index k2=IsLower ? 0 : size;
90         IsLower ? k2<size : k2>0;
91         IsLower ? k2+=kc : k2-=kc)
92     {
93       const Index actual_kc = (std::min)(IsLower ? size-k2 : k2, kc);
94 
95       // We have selected and packed a big horizontal panel R1 of rhs. Let B be the packed copy of this panel,
96       // and R2 the remaining part of rhs. The corresponding vertical panel of lhs is split into
97       // A11 (the triangular part) and A21 the remaining rectangular part.
98       // Then the high level algorithm is:
99       //  - B = R1                    => general block copy (done during the next step)
100       //  - R1 = A11^-1 B             => tricky part
101       //  - update B from the new R1  => actually this has to be performed continuously during the above step
102       //  - R2 -= A21 * B             => GEPP
103 
104       // The tricky part: compute R1 = A11^-1 B while updating B from R1
105       // The idea is to split A11 into multiple small vertical panels.
106       // Each panel can be split into a small triangular part T1k which is processed without optimization,
107       // and the remaining small part T2k which is processed using gebp with appropriate block strides
108       for(Index j2=0; j2<cols; j2+=subcols)
109       {
110         Index actual_cols = (std::min)(cols-j2,subcols);
111         // for each small vertical panels [T1k^T, T2k^T]^T of lhs
112         for (Index k1=0; k1<actual_kc; k1+=SmallPanelWidth)
113         {
114           Index actualPanelWidth = std::min<Index>(actual_kc-k1, SmallPanelWidth);
115           // tr solve
116           for (Index k=0; k<actualPanelWidth; ++k)
117           {
118             // TODO write a small kernel handling this (can be shared with trsv)
119             Index i  = IsLower ? k2+k1+k : k2-k1-k-1;
120             Index rs = actualPanelWidth - k - 1; // remaining size
121             Index s  = TriStorageOrder==RowMajor ? (IsLower ? k2+k1 : i+1)
122                                                  :  IsLower ? i+1 : i-rs;
123 
124             Scalar a = (Mode & UnitDiag) ? Scalar(1) : Scalar(1)/conj(tri(i,i));
125             for (Index j=j2; j<j2+actual_cols; ++j)
126             {
127               if (TriStorageOrder==RowMajor)
128               {
129                 Scalar b(0);
130                 const Scalar* l = &tri(i,s);
131                 typename OtherMapper::LinearMapper r = other.getLinearMapper(s,j);
132                 for (Index i3=0; i3<k; ++i3)
133                   b += conj(l[i3]) * r(i3);
134 
135                 other(i,j) = (other(i,j) - b)*a;
136               }
137               else
138               {
139                 Scalar& otherij = other(i,j);
140                 otherij *= a;
141                 Scalar b = otherij;
142                 typename OtherMapper::LinearMapper r = other.getLinearMapper(s,j);
143                 typename TriMapper::LinearMapper l = tri.getLinearMapper(s,i);
144                 for (Index i3=0;i3<rs;++i3)
145                   r(i3) -= b * conj(l(i3));
146               }
147             }
148           }
149 
150           Index lengthTarget = actual_kc-k1-actualPanelWidth;
151           Index startBlock   = IsLower ? k2+k1 : k2-k1-actualPanelWidth;
152           Index blockBOffset = IsLower ? k1 : lengthTarget;
153 
154           // update the respective rows of B from other
155           pack_rhs(blockB+actual_kc*j2, other.getSubMapper(startBlock,j2), actualPanelWidth, actual_cols, actual_kc, blockBOffset);
156 
157           // GEBP
158           if (lengthTarget>0)
159           {
160             Index startTarget  = IsLower ? k2+k1+actualPanelWidth : k2-actual_kc;
161 
162             pack_lhs(blockA, tri.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget);
163 
164             gebp_kernel(other.getSubMapper(startTarget,j2), blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1),
165                         actualPanelWidth, actual_kc, 0, blockBOffset);
166           }
167         }
168       }
169 
170       // R2 -= A21 * B => GEPP
171       {
172         Index start = IsLower ? k2+kc : 0;
173         Index end   = IsLower ? size : k2-kc;
174         for(Index i2=start; i2<end; i2+=mc)
175         {
176           const Index actual_mc = (std::min)(mc,end-i2);
177           if (actual_mc>0)
178           {
179             pack_lhs(blockA, tri.getSubMapper(i2, IsLower ? k2 : k2-kc), actual_kc, actual_mc);
180 
181             gebp_kernel(other.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, Scalar(-1), -1, -1, 0, 0);
182           }
183         }
184       }
185     }
186   }
187 
188 /* Optimized triangular solver with multiple left hand sides and the triangular matrix on the right
189  */
190 template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
191 struct triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>
192 {
193   static EIGEN_DONT_INLINE void run(
194     Index size, Index otherSize,
195     const Scalar* _tri, Index triStride,
196     Scalar* _other, Index otherIncr, Index otherStride,
197     level3_blocking<Scalar,Scalar>& blocking);
198 };
199 template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder, int OtherInnerStride>
200 EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor,OtherInnerStride>::run(
201     Index size, Index otherSize,
202     const Scalar* _tri, Index triStride,
203     Scalar* _other, Index otherIncr, Index otherStride,
204     level3_blocking<Scalar,Scalar>& blocking)
205   {
206     Index rows = otherSize;
207     typedef typename NumTraits<Scalar>::Real RealScalar;
208 
209     typedef blas_data_mapper<Scalar, Index, ColMajor, Unaligned, OtherInnerStride> LhsMapper;
210     typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> RhsMapper;
211     LhsMapper lhs(_other, otherStride, otherIncr);
212     RhsMapper rhs(_tri, triStride);
213 
214     typedef gebp_traits<Scalar,Scalar> Traits;
215     enum {
216       RhsStorageOrder   = TriStorageOrder,
217       SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
218       IsLower = (Mode&Lower) == Lower
219     };
220 
221     Index kc = blocking.kc();                   // cache block size along the K direction
222     Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
223 
224     std::size_t sizeA = kc*mc;
225     std::size_t sizeB = kc*size;
226 
227     ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
228     ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
229 
230     conj_if<Conjugate> conj;
231     gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
232     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
233     gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel;
234     gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, typename Traits::LhsPacket4Packing, ColMajor, false, true> pack_lhs_panel;
235 
236     for(Index k2=IsLower ? size : 0;
237         IsLower ? k2>0 : k2<size;
238         IsLower ? k2-=kc : k2+=kc)
239     {
240       const Index actual_kc = (std::min)(IsLower ? k2 : size-k2, kc);
241       Index actual_k2 = IsLower ? k2-actual_kc : k2 ;
242 
243       Index startPanel = IsLower ? 0 : k2+actual_kc;
244       Index rs = IsLower ? actual_k2 : size - actual_k2 - actual_kc;
245       Scalar* geb = blockB+actual_kc*actual_kc;
246 
247       if (rs>0) pack_rhs(geb, rhs.getSubMapper(actual_k2,startPanel), actual_kc, rs);
248 
249       // triangular packing (we only pack the panels off the diagonal,
250       // neglecting the blocks overlapping the diagonal
251       {
252         for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth)
253         {
254           Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
255           Index actual_j2 = actual_k2 + j2;
256           Index panelOffset = IsLower ? j2+actualPanelWidth : 0;
257           Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2;
258 
259           if (panelLength>0)
260           pack_rhs_panel(blockB+j2*actual_kc,
261                          rhs.getSubMapper(actual_k2+panelOffset, actual_j2),
262                          panelLength, actualPanelWidth,
263                          actual_kc, panelOffset);
264         }
265       }
266 
267       for(Index i2=0; i2<rows; i2+=mc)
268       {
269         const Index actual_mc = (std::min)(mc,rows-i2);
270 
271         // triangular solver kernel
272         {
273           // for each small block of the diagonal (=> vertical panels of rhs)
274           for (Index j2 = IsLower
275                       ? (actual_kc - ((actual_kc%SmallPanelWidth) ? Index(actual_kc%SmallPanelWidth)
276                                                                   : Index(SmallPanelWidth)))
277                       : 0;
278                IsLower ? j2>=0 : j2<actual_kc;
279                IsLower ? j2-=SmallPanelWidth : j2+=SmallPanelWidth)
280           {
281             Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
282             Index absolute_j2 = actual_k2 + j2;
283             Index panelOffset = IsLower ? j2+actualPanelWidth : 0;
284             Index panelLength = IsLower ? actual_kc - j2 - actualPanelWidth : j2;
285 
286             // GEBP
287             if(panelLength>0)
288             {
289               gebp_kernel(lhs.getSubMapper(i2,absolute_j2),
290                           blockA, blockB+j2*actual_kc,
291                           actual_mc, panelLength, actualPanelWidth,
292                           Scalar(-1),
293                           actual_kc, actual_kc, // strides
294                           panelOffset, panelOffset); // offsets
295             }
296 
297             // unblocked triangular solve
298             for (Index k=0; k<actualPanelWidth; ++k)
299             {
300               Index j = IsLower ? absolute_j2+actualPanelWidth-k-1 : absolute_j2+k;
301 
302               typename LhsMapper::LinearMapper r = lhs.getLinearMapper(i2,j);
303               for (Index k3=0; k3<k; ++k3)
304               {
305                 Scalar b = conj(rhs(IsLower ? j+1+k3 : absolute_j2+k3,j));
306                 typename LhsMapper::LinearMapper a = lhs.getLinearMapper(i2,IsLower ? j+1+k3 : absolute_j2+k3);
307                 for (Index i=0; i<actual_mc; ++i)
308                   r(i) -= a(i) * b;
309               }
310               if((Mode & UnitDiag)==0)
311               {
312                 Scalar inv_rjj = RealScalar(1)/conj(rhs(j,j));
313                 for (Index i=0; i<actual_mc; ++i)
314                   r(i) *= inv_rjj;
315               }
316             }
317 
318             // pack the just computed part of lhs to A
319             pack_lhs_panel(blockA, lhs.getSubMapper(i2,absolute_j2),
320                            actualPanelWidth, actual_mc,
321                            actual_kc, j2);
322           }
323         }
324 
325         if (rs>0)
326           gebp_kernel(lhs.getSubMapper(i2, startPanel), blockA, geb,
327                       actual_mc, actual_kc, rs, Scalar(-1),
328                       -1, -1, 0, 0);
329       }
330     }
331   }
332 
333 } // end namespace internal
334 
335 } // end namespace Eigen
336 
337 #endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_H
338