1 // Ceres Solver - A fast non-linear least squares minimizer
2 // Copyright 2013 Google Inc. All rights reserved.
3 // http://code.google.com/p/ceres-solver/
4 //
5 // Redistribution and use in source and binary forms, with or without
6 // modification, are permitted provided that the following conditions are met:
7 //
8 // * Redistributions of source code must retain the above copyright notice,
9 //   this list of conditions and the following disclaimer.
10 // * Redistributions in binary form must reproduce the above copyright notice,
11 //   this list of conditions and the following disclaimer in the documentation
12 //   and/or other materials provided with the distribution.
13 // * Neither the name of Google Inc. nor the names of its contributors may be
14 //   used to endorse or promote products derived from this software without
15 //   specific prior written permission.
16 //
17 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18 // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 // ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
21 // LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22 // CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23 // SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24 // INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25 // CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26 // ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27 // POSSIBILITY OF SUCH DAMAGE.
28 //
29 // Author: sameeragarwal@google.com (Sameer Agarwal)
30 //
31 // Simple blas functions for use in the Schur Eliminator. These are
32 // fairly basic implementations which already yield a significant
33 // speedup in the eliminator performance.
34 
35 #ifndef CERES_INTERNAL_SMALL_BLAS_H_
36 #define CERES_INTERNAL_SMALL_BLAS_H_
37 
38 #include "ceres/internal/port.h"
39 #include "ceres/internal/eigen.h"
40 #include "glog/logging.h"
41 
42 namespace ceres {
43 namespace internal {
44 
45 // The following three macros are used to share code and reduce
46 // template junk across the various GEMM variants.
47 #define CERES_GEMM_BEGIN(name)                                          \
48   template<int kRowA, int kColA, int kRowB, int kColB, int kOperation>  \
49   inline void name(const double* A,                                     \
50                    const int num_row_a,                                 \
51                    const int num_col_a,                                 \
52                    const double* B,                                     \
53                    const int num_row_b,                                 \
54                    const int num_col_b,                                 \
55                    double* C,                                           \
56                    const int start_row_c,                               \
57                    const int start_col_c,                               \
58                    const int row_stride_c,                              \
59                    const int col_stride_c)
60 
61 #define CERES_GEMM_NAIVE_HEADER                                         \
62   DCHECK_GT(num_row_a, 0);                                              \
63   DCHECK_GT(num_col_a, 0);                                              \
64   DCHECK_GT(num_row_b, 0);                                              \
65   DCHECK_GT(num_col_b, 0);                                              \
66   DCHECK_GE(start_row_c, 0);                                            \
67   DCHECK_GE(start_col_c, 0);                                            \
68   DCHECK_GT(row_stride_c, 0);                                           \
69   DCHECK_GT(col_stride_c, 0);                                           \
70   DCHECK((kRowA == Eigen::Dynamic) || (kRowA == num_row_a));            \
71   DCHECK((kColA == Eigen::Dynamic) || (kColA == num_col_a));            \
72   DCHECK((kRowB == Eigen::Dynamic) || (kRowB == num_row_b));            \
73   DCHECK((kColB == Eigen::Dynamic) || (kColB == num_col_b));            \
74   const int NUM_ROW_A = (kRowA != Eigen::Dynamic ? kRowA : num_row_a);  \
75   const int NUM_COL_A = (kColA != Eigen::Dynamic ? kColA : num_col_a);  \
76   const int NUM_ROW_B = (kColB != Eigen::Dynamic ? kRowB : num_row_b);  \
77   const int NUM_COL_B = (kColB != Eigen::Dynamic ? kColB : num_col_b);
78 
79 #define CERES_GEMM_EIGEN_HEADER                                         \
80   const typename EigenTypes<kRowA, kColA>::ConstMatrixRef               \
81   Aref(A, num_row_a, num_col_a);                                        \
82   const typename EigenTypes<kRowB, kColB>::ConstMatrixRef               \
83   Bref(B, num_row_b, num_col_b);                                        \
84   MatrixRef Cref(C, row_stride_c, col_stride_c);                        \
85 
86 #define CERES_CALL_GEMM(name)                                           \
87   name<kRowA, kColA, kRowB, kColB, kOperation>(                         \
88       A, num_row_a, num_col_a,                                          \
89       B, num_row_b, num_col_b,                                          \
90       C, start_row_c, start_col_c, row_stride_c, col_stride_c);
91 
92 
93 // For the matrix-matrix functions below, there are three variants for
94 // each functionality. Foo, FooNaive and FooEigen. Foo is the one to
95 // be called by the user. FooNaive is a basic loop based
96 // implementation and FooEigen uses Eigen's implementation. Foo
97 // chooses between FooNaive and FooEigen depending on how many of the
98 // template arguments are fixed at compile time. Currently, FooEigen
99 // is called if all matrix dimensions are compile time
100 // constants. FooNaive is called otherwise. This leads to the best
101 // performance currently.
102 //
103 // The MatrixMatrixMultiply variants compute:
104 //
105 //   C op A * B;
106 //
107 // The MatrixTransposeMatrixMultiply variants compute:
108 //
109 //   C op A' * B
110 //
111 // where op can be +=, -=, or =.
112 //
113 // The template parameters (kRowA, kColA, kRowB, kColB) allow
114 // specialization of the loop at compile time. If this information is
115 // not available, then Eigen::Dynamic should be used as the template
116 // argument.
117 //
118 //   kOperation =  1  -> C += A * B
119 //   kOperation = -1  -> C -= A * B
120 //   kOperation =  0  -> C  = A * B
121 //
122 // The functions can write into matrices C which are larger than the
123 // matrix A * B. This is done by specifying the true size of C via
124 // row_stride_c and col_stride_c, and then indicating where A * B
125 // should be written into by start_row_c and start_col_c.
126 //
127 // Graphically if row_stride_c = 10, col_stride_c = 12, start_row_c =
128 // 4 and start_col_c = 5, then if A = 3x2 and B = 2x4, we get
129 //
130 //   ------------
131 //   ------------
132 //   ------------
133 //   ------------
134 //   -----xxxx---
135 //   -----xxxx---
136 //   -----xxxx---
137 //   ------------
138 //   ------------
139 //   ------------
140 //
CERES_GEMM_BEGIN(MatrixMatrixMultiplyEigen)141 CERES_GEMM_BEGIN(MatrixMatrixMultiplyEigen) {
142   CERES_GEMM_EIGEN_HEADER
143   Eigen::Block<MatrixRef, kRowA, kColB>
144     block(Cref, start_row_c, start_col_c, num_row_a, num_col_b);
145 
146   if (kOperation > 0) {
147     block.noalias() += Aref * Bref;
148   } else if (kOperation < 0) {
149     block.noalias() -= Aref * Bref;
150   } else {
151     block.noalias() = Aref * Bref;
152   }
153 }
154 
CERES_GEMM_BEGIN(MatrixMatrixMultiplyNaive)155 CERES_GEMM_BEGIN(MatrixMatrixMultiplyNaive) {
156   CERES_GEMM_NAIVE_HEADER
157   DCHECK_EQ(NUM_COL_A, NUM_ROW_B);
158 
159   const int NUM_ROW_C = NUM_ROW_A;
160   const int NUM_COL_C = NUM_COL_B;
161   DCHECK_LE(start_row_c + NUM_ROW_C, row_stride_c);
162   DCHECK_LE(start_col_c + NUM_COL_C, col_stride_c);
163 
164   for (int row = 0; row < NUM_ROW_C; ++row) {
165     for (int col = 0; col < NUM_COL_C; ++col) {
166       double tmp = 0.0;
167       for (int k = 0; k < NUM_COL_A; ++k) {
168         tmp += A[row * NUM_COL_A + k] * B[k * NUM_COL_B + col];
169       }
170 
171       const int index = (row + start_row_c) * col_stride_c + start_col_c + col;
172       if (kOperation > 0) {
173         C[index] += tmp;
174       } else if (kOperation < 0) {
175         C[index] -= tmp;
176       } else {
177         C[index] = tmp;
178       }
179     }
180   }
181 }
182 
CERES_GEMM_BEGIN(MatrixMatrixMultiply)183 CERES_GEMM_BEGIN(MatrixMatrixMultiply) {
184 #ifdef CERES_NO_CUSTOM_BLAS
185 
186   CERES_CALL_GEMM(MatrixMatrixMultiplyEigen)
187   return;
188 
189 #else
190 
191   if (kRowA != Eigen::Dynamic && kColA != Eigen::Dynamic &&
192       kRowB != Eigen::Dynamic && kColB != Eigen::Dynamic) {
193     CERES_CALL_GEMM(MatrixMatrixMultiplyEigen)
194   } else {
195     CERES_CALL_GEMM(MatrixMatrixMultiplyNaive)
196   }
197 
198 #endif
199 }
200 
CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiplyEigen)201 CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiplyEigen) {
202   CERES_GEMM_EIGEN_HEADER
203   Eigen::Block<MatrixRef, kColA, kColB> block(Cref,
204                                               start_row_c, start_col_c,
205                                               num_col_a, num_col_b);
206   if (kOperation > 0) {
207     block.noalias() += Aref.transpose() * Bref;
208   } else if (kOperation < 0) {
209     block.noalias() -= Aref.transpose() * Bref;
210   } else {
211     block.noalias() = Aref.transpose() * Bref;
212   }
213 }
214 
CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiplyNaive)215 CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiplyNaive) {
216   CERES_GEMM_NAIVE_HEADER
217   DCHECK_EQ(NUM_ROW_A, NUM_ROW_B);
218 
219   const int NUM_ROW_C = NUM_COL_A;
220   const int NUM_COL_C = NUM_COL_B;
221   DCHECK_LE(start_row_c + NUM_ROW_C, row_stride_c);
222   DCHECK_LE(start_col_c + NUM_COL_C, col_stride_c);
223 
224   for (int row = 0; row < NUM_ROW_C; ++row) {
225     for (int col = 0; col < NUM_COL_C; ++col) {
226       double tmp = 0.0;
227       for (int k = 0; k < NUM_ROW_A; ++k) {
228         tmp += A[k * NUM_COL_A + row] * B[k * NUM_COL_B + col];
229       }
230 
231       const int index = (row + start_row_c) * col_stride_c + start_col_c + col;
232       if (kOperation > 0) {
233         C[index]+= tmp;
234       } else if (kOperation < 0) {
235         C[index]-= tmp;
236       } else {
237         C[index]= tmp;
238       }
239     }
240   }
241 }
242 
CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiply)243 CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiply) {
244 #ifdef CERES_NO_CUSTOM_BLAS
245 
246   CERES_CALL_GEMM(MatrixTransposeMatrixMultiplyEigen)
247   return;
248 
249 #else
250 
251   if (kRowA != Eigen::Dynamic && kColA != Eigen::Dynamic &&
252       kRowB != Eigen::Dynamic && kColB != Eigen::Dynamic) {
253     CERES_CALL_GEMM(MatrixTransposeMatrixMultiplyEigen)
254   } else {
255     CERES_CALL_GEMM(MatrixTransposeMatrixMultiplyNaive)
256   }
257 
258 #endif
259 }
260 
261 // Matrix-Vector multiplication
262 //
263 // c op A * b;
264 //
265 // where op can be +=, -=, or =.
266 //
267 // The template parameters (kRowA, kColA) allow specialization of the
268 // loop at compile time. If this information is not available, then
269 // Eigen::Dynamic should be used as the template argument.
270 //
271 // kOperation =  1  -> c += A' * b
272 // kOperation = -1  -> c -= A' * b
273 // kOperation =  0  -> c  = A' * b
274 template<int kRowA, int kColA, int kOperation>
MatrixVectorMultiply(const double * A,const int num_row_a,const int num_col_a,const double * b,double * c)275 inline void MatrixVectorMultiply(const double* A,
276                                  const int num_row_a,
277                                  const int num_col_a,
278                                  const double* b,
279                                  double* c) {
280 #ifdef CERES_NO_CUSTOM_BLAS
281   const typename EigenTypes<kRowA, kColA>::ConstMatrixRef
282       Aref(A, num_row_a, num_col_a);
283   const typename EigenTypes<kColA>::ConstVectorRef bref(b, num_col_a);
284   typename EigenTypes<kRowA>::VectorRef cref(c, num_row_a);
285 
286   // lazyProduct works better than .noalias() for matrix-vector
287   // products.
288   if (kOperation > 0) {
289     cref += Aref.lazyProduct(bref);
290   } else if (kOperation < 0) {
291     cref -= Aref.lazyProduct(bref);
292   } else {
293     cref = Aref.lazyProduct(bref);
294   }
295 #else
296 
297   DCHECK_GT(num_row_a, 0);
298   DCHECK_GT(num_col_a, 0);
299   DCHECK((kRowA == Eigen::Dynamic) || (kRowA == num_row_a));
300   DCHECK((kColA == Eigen::Dynamic) || (kColA == num_col_a));
301 
302   const int NUM_ROW_A = (kRowA != Eigen::Dynamic ? kRowA : num_row_a);
303   const int NUM_COL_A = (kColA != Eigen::Dynamic ? kColA : num_col_a);
304 
305   for (int row = 0; row < NUM_ROW_A; ++row) {
306     double tmp = 0.0;
307     for (int col = 0; col < NUM_COL_A; ++col) {
308       tmp += A[row * NUM_COL_A + col] * b[col];
309     }
310 
311     if (kOperation > 0) {
312       c[row] += tmp;
313     } else if (kOperation < 0) {
314       c[row] -= tmp;
315     } else {
316       c[row] = tmp;
317     }
318   }
319 #endif  // CERES_NO_CUSTOM_BLAS
320 }
321 
322 // Similar to MatrixVectorMultiply, except that A is transposed, i.e.,
323 //
324 // c op A' * b;
325 template<int kRowA, int kColA, int kOperation>
MatrixTransposeVectorMultiply(const double * A,const int num_row_a,const int num_col_a,const double * b,double * c)326 inline void MatrixTransposeVectorMultiply(const double* A,
327                                           const int num_row_a,
328                                           const int num_col_a,
329                                           const double* b,
330                                           double* c) {
331 #ifdef CERES_NO_CUSTOM_BLAS
332   const typename EigenTypes<kRowA, kColA>::ConstMatrixRef
333       Aref(A, num_row_a, num_col_a);
334   const typename EigenTypes<kRowA>::ConstVectorRef bref(b, num_row_a);
335   typename EigenTypes<kColA>::VectorRef cref(c, num_col_a);
336 
337   // lazyProduct works better than .noalias() for matrix-vector
338   // products.
339   if (kOperation > 0) {
340     cref += Aref.transpose().lazyProduct(bref);
341   } else if (kOperation < 0) {
342     cref -= Aref.transpose().lazyProduct(bref);
343   } else {
344     cref = Aref.transpose().lazyProduct(bref);
345   }
346 #else
347 
348   DCHECK_GT(num_row_a, 0);
349   DCHECK_GT(num_col_a, 0);
350   DCHECK((kRowA == Eigen::Dynamic) || (kRowA == num_row_a));
351   DCHECK((kColA == Eigen::Dynamic) || (kColA == num_col_a));
352 
353   const int NUM_ROW_A = (kRowA != Eigen::Dynamic ? kRowA : num_row_a);
354   const int NUM_COL_A = (kColA != Eigen::Dynamic ? kColA : num_col_a);
355 
356   for (int row = 0; row < NUM_COL_A; ++row) {
357     double tmp = 0.0;
358     for (int col = 0; col < NUM_ROW_A; ++col) {
359       tmp += A[col * NUM_COL_A + row] * b[col];
360     }
361 
362     if (kOperation > 0) {
363       c[row] += tmp;
364     } else if (kOperation < 0) {
365       c[row] -= tmp;
366     } else {
367       c[row] = tmp;
368     }
369   }
370 #endif  // CERES_NO_CUSTOM_BLAS
371 }
372 
373 #undef CERES_GEMM_BEGIN
374 #undef CERES_GEMM_EIGEN_HEADER
375 #undef CERES_GEMM_NAIVE_HEADER
376 #undef CERES_CALL_GEMM
377 
378 }  // namespace internal
379 }  // namespace ceres
380 
381 #endif  // CERES_INTERNAL_SMALL_BLAS_H_
382