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 (at) 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/eigen.h" 39 #include "glog/logging.h" 40 41 namespace ceres { 42 namespace internal { 43 44 // Remove the ".noalias()" annotation from the matrix matrix 45 // mutliplies to produce a correct build with the Android NDK, 46 // including versions 6, 7, 8, and 8b, when built with STLPort and the 47 // non-standalone toolchain (i.e. ndk-build). This appears to be a 48 // compiler bug; if the workaround is not in place, the line 49 // 50 // block.noalias() -= A * B; 51 // 52 // gets compiled to 53 // 54 // block.noalias() += A * B; 55 // 56 // which breaks schur elimination. Introducing a temporary by removing the 57 // .noalias() annotation causes the issue to disappear. Tracking this 58 // issue down was tricky, since the test suite doesn't run when built with 59 // the non-standalone toolchain. 60 // 61 // TODO(keir): Make a reproduction case for this and send it upstream. 62 #ifdef CERES_WORK_AROUND_ANDROID_NDK_COMPILER_BUG 63 #define CERES_MAYBE_NOALIAS 64 #else 65 #define CERES_MAYBE_NOALIAS .noalias() 66 #endif 67 68 // The following three macros are used to share code and reduce 69 // template junk across the various GEMM variants. 70 #define CERES_GEMM_BEGIN(name) \ 71 template<int kRowA, int kColA, int kRowB, int kColB, int kOperation> \ 72 inline void name(const double* A, \ 73 const int num_row_a, \ 74 const int num_col_a, \ 75 const double* B, \ 76 const int num_row_b, \ 77 const int num_col_b, \ 78 double* C, \ 79 const int start_row_c, \ 80 const int start_col_c, \ 81 const int row_stride_c, \ 82 const int col_stride_c) 83 84 #define CERES_GEMM_NAIVE_HEADER \ 85 DCHECK_GT(num_row_a, 0); \ 86 DCHECK_GT(num_col_a, 0); \ 87 DCHECK_GT(num_row_b, 0); \ 88 DCHECK_GT(num_col_b, 0); \ 89 DCHECK_GE(start_row_c, 0); \ 90 DCHECK_GE(start_col_c, 0); \ 91 DCHECK_GT(row_stride_c, 0); \ 92 DCHECK_GT(col_stride_c, 0); \ 93 DCHECK((kRowA == Eigen::Dynamic) || (kRowA == num_row_a)); \ 94 DCHECK((kColA == Eigen::Dynamic) || (kColA == num_col_a)); \ 95 DCHECK((kRowB == Eigen::Dynamic) || (kRowB == num_row_b)); \ 96 DCHECK((kColB == Eigen::Dynamic) || (kColB == num_col_b)); \ 97 const int NUM_ROW_A = (kRowA != Eigen::Dynamic ? kRowA : num_row_a); \ 98 const int NUM_COL_A = (kColA != Eigen::Dynamic ? kColA : num_col_a); \ 99 const int NUM_ROW_B = (kColB != Eigen::Dynamic ? kRowB : num_row_b); \ 100 const int NUM_COL_B = (kColB != Eigen::Dynamic ? kColB : num_col_b); 101 102 #define CERES_GEMM_EIGEN_HEADER \ 103 const typename EigenTypes<kRowA, kColA>::ConstMatrixRef \ 104 Aref(A, num_row_a, num_col_a); \ 105 const typename EigenTypes<kRowB, kColB>::ConstMatrixRef \ 106 Bref(B, num_row_b, num_col_b); \ 107 MatrixRef Cref(C, row_stride_c, col_stride_c); \ 108 109 #define CERES_CALL_GEMM(name) \ 110 name<kRowA, kColA, kRowB, kColB, kOperation>( \ 111 A, num_row_a, num_col_a, \ 112 B, num_row_b, num_col_b, \ 113 C, start_row_c, start_col_c, row_stride_c, col_stride_c); 114 115 116 // For the matrix-matrix functions below, there are three variants for 117 // each functionality. Foo, FooNaive and FooEigen. Foo is the one to 118 // be called by the user. FooNaive is a basic loop based 119 // implementation and FooEigen uses Eigen's implementation. Foo 120 // chooses between FooNaive and FooEigen depending on how many of the 121 // template arguments are fixed at compile time. Currently, FooEigen 122 // is called if all matrix dimensions are compile time 123 // constants. FooNaive is called otherwise. This leads to the best 124 // performance currently. 125 // 126 // The MatrixMatrixMultiply variants compute: 127 // 128 // C op A * B; 129 // 130 // The MatrixTransposeMatrixMultiply variants compute: 131 // 132 // C op A' * B 133 // 134 // where op can be +=, -=, or =. 135 // 136 // The template parameters (kRowA, kColA, kRowB, kColB) allow 137 // specialization of the loop at compile time. If this information is 138 // not available, then Eigen::Dynamic should be used as the template 139 // argument. 140 // 141 // kOperation = 1 -> C += A * B 142 // kOperation = -1 -> C -= A * B 143 // kOperation = 0 -> C = A * B 144 // 145 // The functions can write into matrices C which are larger than the 146 // matrix A * B. This is done by specifying the true size of C via 147 // row_stride_c and col_stride_c, and then indicating where A * B 148 // should be written into by start_row_c and start_col_c. 149 // 150 // Graphically if row_stride_c = 10, col_stride_c = 12, start_row_c = 151 // 4 and start_col_c = 5, then if A = 3x2 and B = 2x4, we get 152 // 153 // ------------ 154 // ------------ 155 // ------------ 156 // ------------ 157 // -----xxxx--- 158 // -----xxxx--- 159 // -----xxxx--- 160 // ------------ 161 // ------------ 162 // ------------ 163 // 164 CERES_GEMM_BEGIN(MatrixMatrixMultiplyEigen) { 165 CERES_GEMM_EIGEN_HEADER 166 Eigen::Block<MatrixRef, kRowA, kColB> 167 block(Cref, start_row_c, start_col_c, num_row_a, num_col_b); 168 169 if (kOperation > 0) { 170 block CERES_MAYBE_NOALIAS += Aref * Bref; 171 } else if (kOperation < 0) { 172 block CERES_MAYBE_NOALIAS -= Aref * Bref; 173 } else { 174 block CERES_MAYBE_NOALIAS = Aref * Bref; 175 } 176 } 177 178 CERES_GEMM_BEGIN(MatrixMatrixMultiplyNaive) { 179 CERES_GEMM_NAIVE_HEADER 180 DCHECK_EQ(NUM_COL_A, NUM_ROW_B); 181 182 const int NUM_ROW_C = NUM_ROW_A; 183 const int NUM_COL_C = NUM_COL_B; 184 DCHECK_LE(start_row_c + NUM_ROW_C, row_stride_c); 185 DCHECK_LE(start_col_c + NUM_COL_C, col_stride_c); 186 187 for (int row = 0; row < NUM_ROW_C; ++row) { 188 for (int col = 0; col < NUM_COL_C; ++col) { 189 double tmp = 0.0; 190 for (int k = 0; k < NUM_COL_A; ++k) { 191 tmp += A[row * NUM_COL_A + k] * B[k * NUM_COL_B + col]; 192 } 193 194 const int index = (row + start_row_c) * col_stride_c + start_col_c + col; 195 if (kOperation > 0) { 196 C[index] += tmp; 197 } else if (kOperation < 0) { 198 C[index] -= tmp; 199 } else { 200 C[index] = tmp; 201 } 202 } 203 } 204 } 205 206 CERES_GEMM_BEGIN(MatrixMatrixMultiply) { 207 #ifdef CERES_NO_CUSTOM_BLAS 208 209 CERES_CALL_GEMM(MatrixMatrixMultiplyEigen) 210 return; 211 212 #else 213 214 if (kRowA != Eigen::Dynamic && kColA != Eigen::Dynamic && 215 kRowB != Eigen::Dynamic && kColB != Eigen::Dynamic) { 216 CERES_CALL_GEMM(MatrixMatrixMultiplyEigen) 217 } else { 218 CERES_CALL_GEMM(MatrixMatrixMultiplyNaive) 219 } 220 221 #endif 222 } 223 224 CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiplyEigen) { 225 CERES_GEMM_EIGEN_HEADER 226 Eigen::Block<MatrixRef, kColA, kColB> block(Cref, 227 start_row_c, start_col_c, 228 num_col_a, num_col_b); 229 if (kOperation > 0) { 230 block CERES_MAYBE_NOALIAS += Aref.transpose() * Bref; 231 } else if (kOperation < 0) { 232 block CERES_MAYBE_NOALIAS -= Aref.transpose() * Bref; 233 } else { 234 block CERES_MAYBE_NOALIAS = Aref.transpose() * Bref; 235 } 236 } 237 238 CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiplyNaive) { 239 CERES_GEMM_NAIVE_HEADER 240 DCHECK_EQ(NUM_ROW_A, NUM_ROW_B); 241 242 const int NUM_ROW_C = NUM_COL_A; 243 const int NUM_COL_C = NUM_COL_B; 244 DCHECK_LE(start_row_c + NUM_ROW_C, row_stride_c); 245 DCHECK_LE(start_col_c + NUM_COL_C, col_stride_c); 246 247 for (int row = 0; row < NUM_ROW_C; ++row) { 248 for (int col = 0; col < NUM_COL_C; ++col) { 249 double tmp = 0.0; 250 for (int k = 0; k < NUM_ROW_A; ++k) { 251 tmp += A[k * NUM_COL_A + row] * B[k * NUM_COL_B + col]; 252 } 253 254 const int index = (row + start_row_c) * col_stride_c + start_col_c + col; 255 if (kOperation > 0) { 256 C[index]+= tmp; 257 } else if (kOperation < 0) { 258 C[index]-= tmp; 259 } else { 260 C[index]= tmp; 261 } 262 } 263 } 264 } 265 266 CERES_GEMM_BEGIN(MatrixTransposeMatrixMultiply) { 267 #ifdef CERES_NO_CUSTOM_BLAS 268 269 CERES_CALL_GEMM(MatrixTransposeMatrixMultiplyEigen) 270 return; 271 272 #else 273 274 if (kRowA != Eigen::Dynamic && kColA != Eigen::Dynamic && 275 kRowB != Eigen::Dynamic && kColB != Eigen::Dynamic) { 276 CERES_CALL_GEMM(MatrixTransposeMatrixMultiplyEigen) 277 } else { 278 CERES_CALL_GEMM(MatrixTransposeMatrixMultiplyNaive) 279 } 280 281 #endif 282 } 283 284 // Matrix-Vector multiplication 285 // 286 // c op A * b; 287 // 288 // where op can be +=, -=, or =. 289 // 290 // The template parameters (kRowA, kColA) allow specialization of the 291 // loop at compile time. If this information is not available, then 292 // Eigen::Dynamic should be used as the template argument. 293 // 294 // kOperation = 1 -> c += A' * b 295 // kOperation = -1 -> c -= A' * b 296 // kOperation = 0 -> c = A' * b 297 template<int kRowA, int kColA, int kOperation> 298 inline void MatrixVectorMultiply(const double* A, 299 const int num_row_a, 300 const int num_col_a, 301 const double* b, 302 double* c) { 303 #ifdef CERES_NO_CUSTOM_BLAS 304 const typename EigenTypes<kRowA, kColA>::ConstMatrixRef 305 Aref(A, num_row_a, num_col_a); 306 const typename EigenTypes<kColA>::ConstVectorRef bref(b, num_col_a); 307 typename EigenTypes<kRowA>::VectorRef cref(c, num_row_a); 308 309 // lazyProduct works better than .noalias() for matrix-vector 310 // products. 311 if (kOperation > 0) { 312 cref += Aref.lazyProduct(bref); 313 } else if (kOperation < 0) { 314 cref -= Aref.lazyProduct(bref); 315 } else { 316 cref = Aref.lazyProduct(bref); 317 } 318 #else 319 320 DCHECK_GT(num_row_a, 0); 321 DCHECK_GT(num_col_a, 0); 322 DCHECK((kRowA == Eigen::Dynamic) || (kRowA == num_row_a)); 323 DCHECK((kColA == Eigen::Dynamic) || (kColA == num_col_a)); 324 325 const int NUM_ROW_A = (kRowA != Eigen::Dynamic ? kRowA : num_row_a); 326 const int NUM_COL_A = (kColA != Eigen::Dynamic ? kColA : num_col_a); 327 328 for (int row = 0; row < NUM_ROW_A; ++row) { 329 double tmp = 0.0; 330 for (int col = 0; col < NUM_COL_A; ++col) { 331 tmp += A[row * NUM_COL_A + col] * b[col]; 332 } 333 334 if (kOperation > 0) { 335 c[row] += tmp; 336 } else if (kOperation < 0) { 337 c[row] -= tmp; 338 } else { 339 c[row] = tmp; 340 } 341 } 342 #endif // CERES_NO_CUSTOM_BLAS 343 } 344 345 // Similar to MatrixVectorMultiply, except that A is transposed, i.e., 346 // 347 // c op A' * b; 348 template<int kRowA, int kColA, int kOperation> 349 inline void MatrixTransposeVectorMultiply(const double* A, 350 const int num_row_a, 351 const int num_col_a, 352 const double* b, 353 double* c) { 354 #ifdef CERES_NO_CUSTOM_BLAS 355 const typename EigenTypes<kRowA, kColA>::ConstMatrixRef 356 Aref(A, num_row_a, num_col_a); 357 const typename EigenTypes<kRowA>::ConstVectorRef bref(b, num_row_a); 358 typename EigenTypes<kColA>::VectorRef cref(c, num_col_a); 359 360 // lazyProduct works better than .noalias() for matrix-vector 361 // products. 362 if (kOperation > 0) { 363 cref += Aref.transpose().lazyProduct(bref); 364 } else if (kOperation < 0) { 365 cref -= Aref.transpose().lazyProduct(bref); 366 } else { 367 cref = Aref.transpose().lazyProduct(bref); 368 } 369 #else 370 371 DCHECK_GT(num_row_a, 0); 372 DCHECK_GT(num_col_a, 0); 373 DCHECK((kRowA == Eigen::Dynamic) || (kRowA == num_row_a)); 374 DCHECK((kColA == Eigen::Dynamic) || (kColA == num_col_a)); 375 376 const int NUM_ROW_A = (kRowA != Eigen::Dynamic ? kRowA : num_row_a); 377 const int NUM_COL_A = (kColA != Eigen::Dynamic ? kColA : num_col_a); 378 379 for (int row = 0; row < NUM_COL_A; ++row) { 380 double tmp = 0.0; 381 for (int col = 0; col < NUM_ROW_A; ++col) { 382 tmp += A[col * NUM_COL_A + row] * b[col]; 383 } 384 385 if (kOperation > 0) { 386 c[row] += tmp; 387 } else if (kOperation < 0) { 388 c[row] -= tmp; 389 } else { 390 c[row] = tmp; 391 } 392 } 393 #endif // CERES_NO_CUSTOM_BLAS 394 } 395 396 397 #undef CERES_MAYBE_NOALIAS 398 #undef CERES_GEMM_BEGIN 399 #undef CERES_GEMM_EIGEN_HEADER 400 #undef CERES_GEMM_NAIVE_HEADER 401 #undef CERES_CALL_GEMM 402 403 } // namespace internal 404 } // namespace ceres 405 406 #endif // CERES_INTERNAL_SMALL_BLAS_H_ 407