1 2 //g++ -O3 -g0 -DNDEBUG sparse_product.cpp -I.. -I/home/gael/Coding/LinearAlgebra/mtl4/ -DDENSITY=0.005 -DSIZE=10000 && ./a.out 3 //g++ -O3 -g0 -DNDEBUG sparse_product.cpp -I.. -I/home/gael/Coding/LinearAlgebra/mtl4/ -DDENSITY=0.05 -DSIZE=2000 && ./a.out 4 // -DNOGMM -DNOMTL -DCSPARSE 5 // -I /home/gael/Coding/LinearAlgebra/CSparse/Include/ /home/gael/Coding/LinearAlgebra/CSparse/Lib/libcsparse.a 6 #ifndef SIZE 7 #define SIZE 100000 8 #endif 9 10 #ifndef NBPERROW 11 #define NBPERROW 24 12 #endif 13 14 #ifndef REPEAT 15 #define REPEAT 2 16 #endif 17 18 #ifndef NBTRIES 19 #define NBTRIES 2 20 #endif 21 22 #ifndef KK 23 #define KK 10 24 #endif 25 26 #ifndef NOGOOGLE 27 #define EIGEN_GOOGLEHASH_SUPPORT 28 #include <google/sparse_hash_map> 29 #endif 30 31 #include "BenchSparseUtil.h" 32 33 #define CHECK_MEM 34 // #define CHECK_MEM std/**/::cout << "check mem\n"; getchar(); 35 36 #define BENCH(X) \ 37 timer.reset(); \ 38 for (int _j=0; _j<NBTRIES; ++_j) { \ 39 timer.start(); \ 40 for (int _k=0; _k<REPEAT; ++_k) { \ 41 X \ 42 } timer.stop(); } 43 44 typedef std::vector<Vector2i> Coordinates; 45 typedef std::vector<float> Values; 46 47 EIGEN_DONT_INLINE Scalar* setinnerrand_eigen(const Coordinates& coords, const Values& vals); 48 EIGEN_DONT_INLINE Scalar* setrand_eigen_dynamic(const Coordinates& coords, const Values& vals); 49 EIGEN_DONT_INLINE Scalar* setrand_eigen_compact(const Coordinates& coords, const Values& vals); 50 EIGEN_DONT_INLINE Scalar* setrand_eigen_sumeq(const Coordinates& coords, const Values& vals); 51 EIGEN_DONT_INLINE Scalar* setrand_eigen_gnu_hash(const Coordinates& coords, const Values& vals); 52 EIGEN_DONT_INLINE Scalar* setrand_eigen_google_dense(const Coordinates& coords, const Values& vals); 53 EIGEN_DONT_INLINE Scalar* setrand_eigen_google_sparse(const Coordinates& coords, const Values& vals); 54 EIGEN_DONT_INLINE Scalar* setrand_scipy(const Coordinates& coords, const Values& vals); 55 EIGEN_DONT_INLINE Scalar* setrand_ublas_mapped(const Coordinates& coords, const Values& vals); 56 EIGEN_DONT_INLINE Scalar* setrand_ublas_coord(const Coordinates& coords, const Values& vals); 57 EIGEN_DONT_INLINE Scalar* setrand_ublas_compressed(const Coordinates& coords, const Values& vals); 58 EIGEN_DONT_INLINE Scalar* setrand_ublas_genvec(const Coordinates& coords, const Values& vals); 59 EIGEN_DONT_INLINE Scalar* setrand_mtl(const Coordinates& coords, const Values& vals); 60 61 int main(int argc, char *argv[]) 62 { 63 int rows = SIZE; 64 int cols = SIZE; 65 bool fullyrand = true; 66 67 BenchTimer timer; 68 Coordinates coords; 69 Values values; 70 if(fullyrand) 71 { 72 Coordinates pool; 73 pool.reserve(cols*NBPERROW); 74 std::cerr << "fill pool" << "\n"; 75 for (int i=0; i<cols*NBPERROW; ) 76 { 77 // DynamicSparseMatrix<int> stencil(SIZE,SIZE); 78 Vector2i ij(internal::random<int>(0,rows-1),internal::random<int>(0,cols-1)); 79 // if(stencil.coeffRef(ij.x(), ij.y())==0) 80 { 81 // stencil.coeffRef(ij.x(), ij.y()) = 1; 82 pool.push_back(ij); 83 84 } 85 ++i; 86 } 87 std::cerr << "pool ok" << "\n"; 88 int n = cols*NBPERROW*KK; 89 coords.reserve(n); 90 values.reserve(n); 91 for (int i=0; i<n; ++i) 92 { 93 int i = internal::random<int>(0,pool.size()); 94 coords.push_back(pool[i]); 95 values.push_back(internal::random<Scalar>()); 96 } 97 } 98 else 99 { 100 for (int j=0; j<cols; ++j) 101 for (int i=0; i<NBPERROW; ++i) 102 { 103 coords.push_back(Vector2i(internal::random<int>(0,rows-1),j)); 104 values.push_back(internal::random<Scalar>()); 105 } 106 } 107 std::cout << "nnz = " << coords.size() << "\n"; 108 CHECK_MEM 109 110 // dense matrices 111 #ifdef DENSEMATRIX 112 { 113 BENCH(setrand_eigen_dense(coords,values);) 114 std::cout << "Eigen Dense\t" << timer.value() << "\n"; 115 } 116 #endif 117 118 // eigen sparse matrices 119 // if (!fullyrand) 120 // { 121 // BENCH(setinnerrand_eigen(coords,values);) 122 // std::cout << "Eigen fillrand\t" << timer.value() << "\n"; 123 // } 124 { 125 BENCH(setrand_eigen_dynamic(coords,values);) 126 std::cout << "Eigen dynamic\t" << timer.value() << "\n"; 127 } 128 // { 129 // BENCH(setrand_eigen_compact(coords,values);) 130 // std::cout << "Eigen compact\t" << timer.value() << "\n"; 131 // } 132 { 133 BENCH(setrand_eigen_sumeq(coords,values);) 134 std::cout << "Eigen sumeq\t" << timer.value() << "\n"; 135 } 136 { 137 // BENCH(setrand_eigen_gnu_hash(coords,values);) 138 // std::cout << "Eigen std::map\t" << timer.value() << "\n"; 139 } 140 { 141 BENCH(setrand_scipy(coords,values);) 142 std::cout << "scipy\t" << timer.value() << "\n"; 143 } 144 #ifndef NOGOOGLE 145 { 146 BENCH(setrand_eigen_google_dense(coords,values);) 147 std::cout << "Eigen google dense\t" << timer.value() << "\n"; 148 } 149 { 150 BENCH(setrand_eigen_google_sparse(coords,values);) 151 std::cout << "Eigen google sparse\t" << timer.value() << "\n"; 152 } 153 #endif 154 155 #ifndef NOUBLAS 156 { 157 // BENCH(setrand_ublas_mapped(coords,values);) 158 // std::cout << "ublas mapped\t" << timer.value() << "\n"; 159 } 160 { 161 BENCH(setrand_ublas_genvec(coords,values);) 162 std::cout << "ublas vecofvec\t" << timer.value() << "\n"; 163 } 164 /*{ 165 timer.reset(); 166 timer.start(); 167 for (int k=0; k<REPEAT; ++k) 168 setrand_ublas_compressed(coords,values); 169 timer.stop(); 170 std::cout << "ublas comp\t" << timer.value() << "\n"; 171 } 172 { 173 timer.reset(); 174 timer.start(); 175 for (int k=0; k<REPEAT; ++k) 176 setrand_ublas_coord(coords,values); 177 timer.stop(); 178 std::cout << "ublas coord\t" << timer.value() << "\n"; 179 }*/ 180 #endif 181 182 183 // MTL4 184 #ifndef NOMTL 185 { 186 BENCH(setrand_mtl(coords,values)); 187 std::cout << "MTL\t" << timer.value() << "\n"; 188 } 189 #endif 190 191 return 0; 192 } 193 194 EIGEN_DONT_INLINE Scalar* setinnerrand_eigen(const Coordinates& coords, const Values& vals) 195 { 196 using namespace Eigen; 197 SparseMatrix<Scalar> mat(SIZE,SIZE); 198 //mat.startFill(2000000/*coords.size()*/); 199 for (int i=0; i<coords.size(); ++i) 200 { 201 mat.insert(coords[i].x(), coords[i].y()) = vals[i]; 202 } 203 mat.finalize(); 204 CHECK_MEM; 205 return 0; 206 } 207 208 EIGEN_DONT_INLINE Scalar* setrand_eigen_dynamic(const Coordinates& coords, const Values& vals) 209 { 210 using namespace Eigen; 211 DynamicSparseMatrix<Scalar> mat(SIZE,SIZE); 212 mat.reserve(coords.size()/10); 213 for (int i=0; i<coords.size(); ++i) 214 { 215 mat.coeffRef(coords[i].x(), coords[i].y()) += vals[i]; 216 } 217 mat.finalize(); 218 CHECK_MEM; 219 return &mat.coeffRef(coords[0].x(), coords[0].y()); 220 } 221 222 EIGEN_DONT_INLINE Scalar* setrand_eigen_sumeq(const Coordinates& coords, const Values& vals) 223 { 224 using namespace Eigen; 225 int n = coords.size()/KK; 226 DynamicSparseMatrix<Scalar> mat(SIZE,SIZE); 227 for (int j=0; j<KK; ++j) 228 { 229 DynamicSparseMatrix<Scalar> aux(SIZE,SIZE); 230 mat.reserve(n); 231 for (int i=j*n; i<(j+1)*n; ++i) 232 { 233 aux.insert(coords[i].x(), coords[i].y()) += vals[i]; 234 } 235 aux.finalize(); 236 mat += aux; 237 } 238 return &mat.coeffRef(coords[0].x(), coords[0].y()); 239 } 240 241 EIGEN_DONT_INLINE Scalar* setrand_eigen_compact(const Coordinates& coords, const Values& vals) 242 { 243 using namespace Eigen; 244 DynamicSparseMatrix<Scalar> setter(SIZE,SIZE); 245 setter.reserve(coords.size()/10); 246 for (int i=0; i<coords.size(); ++i) 247 { 248 setter.coeffRef(coords[i].x(), coords[i].y()) += vals[i]; 249 } 250 SparseMatrix<Scalar> mat = setter; 251 CHECK_MEM; 252 return &mat.coeffRef(coords[0].x(), coords[0].y()); 253 } 254 255 EIGEN_DONT_INLINE Scalar* setrand_eigen_gnu_hash(const Coordinates& coords, const Values& vals) 256 { 257 using namespace Eigen; 258 SparseMatrix<Scalar> mat(SIZE,SIZE); 259 { 260 RandomSetter<SparseMatrix<Scalar>, StdMapTraits > setter(mat); 261 for (int i=0; i<coords.size(); ++i) 262 { 263 setter(coords[i].x(), coords[i].y()) += vals[i]; 264 } 265 CHECK_MEM; 266 } 267 return &mat.coeffRef(coords[0].x(), coords[0].y()); 268 } 269 270 #ifndef NOGOOGLE 271 EIGEN_DONT_INLINE Scalar* setrand_eigen_google_dense(const Coordinates& coords, const Values& vals) 272 { 273 using namespace Eigen; 274 SparseMatrix<Scalar> mat(SIZE,SIZE); 275 { 276 RandomSetter<SparseMatrix<Scalar>, GoogleDenseHashMapTraits> setter(mat); 277 for (int i=0; i<coords.size(); ++i) 278 setter(coords[i].x(), coords[i].y()) += vals[i]; 279 CHECK_MEM; 280 } 281 return &mat.coeffRef(coords[0].x(), coords[0].y()); 282 } 283 284 EIGEN_DONT_INLINE Scalar* setrand_eigen_google_sparse(const Coordinates& coords, const Values& vals) 285 { 286 using namespace Eigen; 287 SparseMatrix<Scalar> mat(SIZE,SIZE); 288 { 289 RandomSetter<SparseMatrix<Scalar>, GoogleSparseHashMapTraits> setter(mat); 290 for (int i=0; i<coords.size(); ++i) 291 setter(coords[i].x(), coords[i].y()) += vals[i]; 292 CHECK_MEM; 293 } 294 return &mat.coeffRef(coords[0].x(), coords[0].y()); 295 } 296 #endif 297 298 299 template <class T> 300 void coo_tocsr(const int n_row, 301 const int n_col, 302 const int nnz, 303 const Coordinates Aij, 304 const Values Ax, 305 int Bp[], 306 int Bj[], 307 T Bx[]) 308 { 309 //compute number of non-zero entries per row of A coo_tocsr 310 std::fill(Bp, Bp + n_row, 0); 311 312 for (int n = 0; n < nnz; n++){ 313 Bp[Aij[n].x()]++; 314 } 315 316 //cumsum the nnz per row to get Bp[] 317 for(int i = 0, cumsum = 0; i < n_row; i++){ 318 int temp = Bp[i]; 319 Bp[i] = cumsum; 320 cumsum += temp; 321 } 322 Bp[n_row] = nnz; 323 324 //write Aj,Ax into Bj,Bx 325 for(int n = 0; n < nnz; n++){ 326 int row = Aij[n].x(); 327 int dest = Bp[row]; 328 329 Bj[dest] = Aij[n].y(); 330 Bx[dest] = Ax[n]; 331 332 Bp[row]++; 333 } 334 335 for(int i = 0, last = 0; i <= n_row; i++){ 336 int temp = Bp[i]; 337 Bp[i] = last; 338 last = temp; 339 } 340 341 //now Bp,Bj,Bx form a CSR representation (with possible duplicates) 342 } 343 344 template< class T1, class T2 > 345 bool kv_pair_less(const std::pair<T1,T2>& x, const std::pair<T1,T2>& y){ 346 return x.first < y.first; 347 } 348 349 350 template<class I, class T> 351 void csr_sort_indices(const I n_row, 352 const I Ap[], 353 I Aj[], 354 T Ax[]) 355 { 356 std::vector< std::pair<I,T> > temp; 357 358 for(I i = 0; i < n_row; i++){ 359 I row_start = Ap[i]; 360 I row_end = Ap[i+1]; 361 362 temp.clear(); 363 364 for(I jj = row_start; jj < row_end; jj++){ 365 temp.push_back(std::make_pair(Aj[jj],Ax[jj])); 366 } 367 368 std::sort(temp.begin(),temp.end(),kv_pair_less<I,T>); 369 370 for(I jj = row_start, n = 0; jj < row_end; jj++, n++){ 371 Aj[jj] = temp[n].first; 372 Ax[jj] = temp[n].second; 373 } 374 } 375 } 376 377 template <class I, class T> 378 void csr_sum_duplicates(const I n_row, 379 const I n_col, 380 I Ap[], 381 I Aj[], 382 T Ax[]) 383 { 384 I nnz = 0; 385 I row_end = 0; 386 for(I i = 0; i < n_row; i++){ 387 I jj = row_end; 388 row_end = Ap[i+1]; 389 while( jj < row_end ){ 390 I j = Aj[jj]; 391 T x = Ax[jj]; 392 jj++; 393 while( jj < row_end && Aj[jj] == j ){ 394 x += Ax[jj]; 395 jj++; 396 } 397 Aj[nnz] = j; 398 Ax[nnz] = x; 399 nnz++; 400 } 401 Ap[i+1] = nnz; 402 } 403 } 404 405 EIGEN_DONT_INLINE Scalar* setrand_scipy(const Coordinates& coords, const Values& vals) 406 { 407 using namespace Eigen; 408 SparseMatrix<Scalar> mat(SIZE,SIZE); 409 mat.resizeNonZeros(coords.size()); 410 // std::cerr << "setrand_scipy...\n"; 411 coo_tocsr<Scalar>(SIZE,SIZE, coords.size(), coords, vals, mat._outerIndexPtr(), mat._innerIndexPtr(), mat._valuePtr()); 412 // std::cerr << "coo_tocsr ok\n"; 413 414 csr_sort_indices(SIZE, mat._outerIndexPtr(), mat._innerIndexPtr(), mat._valuePtr()); 415 416 csr_sum_duplicates(SIZE, SIZE, mat._outerIndexPtr(), mat._innerIndexPtr(), mat._valuePtr()); 417 418 mat.resizeNonZeros(mat._outerIndexPtr()[SIZE]); 419 420 return &mat.coeffRef(coords[0].x(), coords[0].y()); 421 } 422 423 424 #ifndef NOUBLAS 425 EIGEN_DONT_INLINE Scalar* setrand_ublas_mapped(const Coordinates& coords, const Values& vals) 426 { 427 using namespace boost; 428 using namespace boost::numeric; 429 using namespace boost::numeric::ublas; 430 mapped_matrix<Scalar> aux(SIZE,SIZE); 431 for (int i=0; i<coords.size(); ++i) 432 { 433 aux(coords[i].x(), coords[i].y()) += vals[i]; 434 } 435 CHECK_MEM; 436 compressed_matrix<Scalar> mat(aux); 437 return 0;// &mat(coords[0].x(), coords[0].y()); 438 } 439 /*EIGEN_DONT_INLINE Scalar* setrand_ublas_coord(const Coordinates& coords, const Values& vals) 440 { 441 using namespace boost; 442 using namespace boost::numeric; 443 using namespace boost::numeric::ublas; 444 coordinate_matrix<Scalar> aux(SIZE,SIZE); 445 for (int i=0; i<coords.size(); ++i) 446 { 447 aux(coords[i].x(), coords[i].y()) = vals[i]; 448 } 449 compressed_matrix<Scalar> mat(aux); 450 return 0;//&mat(coords[0].x(), coords[0].y()); 451 } 452 EIGEN_DONT_INLINE Scalar* setrand_ublas_compressed(const Coordinates& coords, const Values& vals) 453 { 454 using namespace boost; 455 using namespace boost::numeric; 456 using namespace boost::numeric::ublas; 457 compressed_matrix<Scalar> mat(SIZE,SIZE); 458 for (int i=0; i<coords.size(); ++i) 459 { 460 mat(coords[i].x(), coords[i].y()) = vals[i]; 461 } 462 return 0;//&mat(coords[0].x(), coords[0].y()); 463 }*/ 464 EIGEN_DONT_INLINE Scalar* setrand_ublas_genvec(const Coordinates& coords, const Values& vals) 465 { 466 using namespace boost; 467 using namespace boost::numeric; 468 using namespace boost::numeric::ublas; 469 470 // ublas::vector<coordinate_vector<Scalar> > foo; 471 generalized_vector_of_vector<Scalar, row_major, ublas::vector<coordinate_vector<Scalar> > > aux(SIZE,SIZE); 472 for (int i=0; i<coords.size(); ++i) 473 { 474 aux(coords[i].x(), coords[i].y()) += vals[i]; 475 } 476 CHECK_MEM; 477 compressed_matrix<Scalar,row_major> mat(aux); 478 return 0;//&mat(coords[0].x(), coords[0].y()); 479 } 480 #endif 481 482 #ifndef NOMTL 483 EIGEN_DONT_INLINE void setrand_mtl(const Coordinates& coords, const Values& vals); 484 #endif 485 486
