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