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_MATRIX_MATRIX_H 11 #define EIGEN_TRIANGULAR_MATRIX_MATRIX_H 12 13 namespace Eigen { 14 15 namespace internal { 16 17 // template<typename Scalar, int mr, int StorageOrder, bool Conjugate, int Mode> 18 // struct gemm_pack_lhs_triangular 19 // { 20 // Matrix<Scalar,mr,mr, 21 // void operator()(Scalar* blockA, const EIGEN_RESTRICT Scalar* _lhs, int lhsStride, int depth, int rows) 22 // { 23 // conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; 24 // const_blas_data_mapper<Scalar, StorageOrder> lhs(_lhs,lhsStride); 25 // int count = 0; 26 // const int peeled_mc = (rows/mr)*mr; 27 // for(int i=0; i<peeled_mc; i+=mr) 28 // { 29 // for(int k=0; k<depth; k++) 30 // for(int w=0; w<mr; w++) 31 // blockA[count++] = cj(lhs(i+w, k)); 32 // } 33 // for(int i=peeled_mc; i<rows; i++) 34 // { 35 // for(int k=0; k<depth; k++) 36 // blockA[count++] = cj(lhs(i, k)); 37 // } 38 // } 39 // }; 40 41 /* Optimized triangular matrix * matrix (_TRMM++) product built on top of 42 * the general matrix matrix product. 43 */ 44 template <typename Scalar, typename Index, 45 int Mode, bool LhsIsTriangular, 46 int LhsStorageOrder, bool ConjugateLhs, 47 int RhsStorageOrder, bool ConjugateRhs, 48 int ResStorageOrder, int Version = Specialized> 49 struct product_triangular_matrix_matrix; 50 51 template <typename Scalar, typename Index, 52 int Mode, bool LhsIsTriangular, 53 int LhsStorageOrder, bool ConjugateLhs, 54 int RhsStorageOrder, bool ConjugateRhs, int Version> 55 struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular, 56 LhsStorageOrder,ConjugateLhs, 57 RhsStorageOrder,ConjugateRhs,RowMajor,Version> 58 { 59 static EIGEN_STRONG_INLINE void run( 60 Index rows, Index cols, Index depth, 61 const Scalar* lhs, Index lhsStride, 62 const Scalar* rhs, Index rhsStride, 63 Scalar* res, Index resStride, 64 Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) 65 { 66 product_triangular_matrix_matrix<Scalar, Index, 67 (Mode&(UnitDiag|ZeroDiag)) | ((Mode&Upper) ? Lower : Upper), 68 (!LhsIsTriangular), 69 RhsStorageOrder==RowMajor ? ColMajor : RowMajor, 70 ConjugateRhs, 71 LhsStorageOrder==RowMajor ? ColMajor : RowMajor, 72 ConjugateLhs, 73 ColMajor> 74 ::run(cols, rows, depth, rhs, rhsStride, lhs, lhsStride, res, resStride, alpha, blocking); 75 } 76 }; 77 78 // implements col-major += alpha * op(triangular) * op(general) 79 template <typename Scalar, typename Index, int Mode, 80 int LhsStorageOrder, bool ConjugateLhs, 81 int RhsStorageOrder, bool ConjugateRhs, int Version> 82 struct product_triangular_matrix_matrix<Scalar,Index,Mode,true, 83 LhsStorageOrder,ConjugateLhs, 84 RhsStorageOrder,ConjugateRhs,ColMajor,Version> 85 { 86 87 typedef gebp_traits<Scalar,Scalar> Traits; 88 enum { 89 SmallPanelWidth = 2 * EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr), 90 IsLower = (Mode&Lower) == Lower, 91 SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1 92 }; 93 94 static EIGEN_DONT_INLINE void run( 95 Index _rows, Index _cols, Index _depth, 96 const Scalar* _lhs, Index lhsStride, 97 const Scalar* _rhs, Index rhsStride, 98 Scalar* res, Index resStride, 99 Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) 100 { 101 // strip zeros 102 Index diagSize = (std::min)(_rows,_depth); 103 Index rows = IsLower ? _rows : diagSize; 104 Index depth = IsLower ? diagSize : _depth; 105 Index cols = _cols; 106 107 const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride); 108 const_blas_data_mapper<Scalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride); 109 110 Index kc = blocking.kc(); // cache block size along the K direction 111 Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction 112 113 std::size_t sizeA = kc*mc; 114 std::size_t sizeB = kc*cols; 115 std::size_t sizeW = kc*Traits::WorkSpaceFactor; 116 117 ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); 118 ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); 119 ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW()); 120 121 Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer; 122 triangularBuffer.setZero(); 123 if((Mode&ZeroDiag)==ZeroDiag) 124 triangularBuffer.diagonal().setZero(); 125 else 126 triangularBuffer.diagonal().setOnes(); 127 128 gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel; 129 gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; 130 gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs; 131 132 for(Index k2=IsLower ? depth : 0; 133 IsLower ? k2>0 : k2<depth; 134 IsLower ? k2-=kc : k2+=kc) 135 { 136 Index actual_kc = (std::min)(IsLower ? k2 : depth-k2, kc); 137 Index actual_k2 = IsLower ? k2-actual_kc : k2; 138 139 // align blocks with the end of the triangular part for trapezoidal lhs 140 if((!IsLower)&&(k2<rows)&&(k2+actual_kc>rows)) 141 { 142 actual_kc = rows-k2; 143 k2 = k2+actual_kc-kc; 144 } 145 146 pack_rhs(blockB, &rhs(actual_k2,0), rhsStride, actual_kc, cols); 147 148 // the selected lhs's panel has to be split in three different parts: 149 // 1 - the part which is zero => skip it 150 // 2 - the diagonal block => special kernel 151 // 3 - the dense panel below (lower case) or above (upper case) the diagonal block => GEPP 152 153 // the block diagonal, if any: 154 if(IsLower || actual_k2<rows) 155 { 156 // for each small vertical panels of lhs 157 for (Index k1=0; k1<actual_kc; k1+=SmallPanelWidth) 158 { 159 Index actualPanelWidth = std::min<Index>(actual_kc-k1, SmallPanelWidth); 160 Index lengthTarget = IsLower ? actual_kc-k1-actualPanelWidth : k1; 161 Index startBlock = actual_k2+k1; 162 Index blockBOffset = k1; 163 164 // => GEBP with the micro triangular block 165 // The trick is to pack this micro block while filling the opposite triangular part with zeros. 166 // To this end we do an extra triangular copy to a small temporary buffer 167 for (Index k=0;k<actualPanelWidth;++k) 168 { 169 if (SetDiag) 170 triangularBuffer.coeffRef(k,k) = lhs(startBlock+k,startBlock+k); 171 for (Index i=IsLower ? k+1 : 0; IsLower ? i<actualPanelWidth : i<k; ++i) 172 triangularBuffer.coeffRef(i,k) = lhs(startBlock+i,startBlock+k); 173 } 174 pack_lhs(blockA, triangularBuffer.data(), triangularBuffer.outerStride(), actualPanelWidth, actualPanelWidth); 175 176 gebp_kernel(res+startBlock, resStride, blockA, blockB, actualPanelWidth, actualPanelWidth, cols, alpha, 177 actualPanelWidth, actual_kc, 0, blockBOffset, blockW); 178 179 // GEBP with remaining micro panel 180 if (lengthTarget>0) 181 { 182 Index startTarget = IsLower ? actual_k2+k1+actualPanelWidth : actual_k2; 183 184 pack_lhs(blockA, &lhs(startTarget,startBlock), lhsStride, actualPanelWidth, lengthTarget); 185 186 gebp_kernel(res+startTarget, resStride, blockA, blockB, lengthTarget, actualPanelWidth, cols, alpha, 187 actualPanelWidth, actual_kc, 0, blockBOffset, blockW); 188 } 189 } 190 } 191 // the part below (lower case) or above (upper case) the diagonal => GEPP 192 { 193 Index start = IsLower ? k2 : 0; 194 Index end = IsLower ? rows : (std::min)(actual_k2,rows); 195 for(Index i2=start; i2<end; i2+=mc) 196 { 197 const Index actual_mc = (std::min)(i2+mc,end)-i2; 198 gemm_pack_lhs<Scalar, Index, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>() 199 (blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc); 200 201 gebp_kernel(res+i2, resStride, blockA, blockB, actual_mc, actual_kc, cols, alpha, -1, -1, 0, 0, blockW); 202 } 203 } 204 } 205 } 206 }; 207 208 // implements col-major += alpha * op(general) * op(triangular) 209 template <typename Scalar, typename Index, int Mode, 210 int LhsStorageOrder, bool ConjugateLhs, 211 int RhsStorageOrder, bool ConjugateRhs, int Version> 212 struct product_triangular_matrix_matrix<Scalar,Index,Mode,false, 213 LhsStorageOrder,ConjugateLhs, 214 RhsStorageOrder,ConjugateRhs,ColMajor,Version> 215 { 216 typedef gebp_traits<Scalar,Scalar> Traits; 217 enum { 218 SmallPanelWidth = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr), 219 IsLower = (Mode&Lower) == Lower, 220 SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1 221 }; 222 223 static EIGEN_DONT_INLINE void run( 224 Index _rows, Index _cols, Index _depth, 225 const Scalar* _lhs, Index lhsStride, 226 const Scalar* _rhs, Index rhsStride, 227 Scalar* res, Index resStride, 228 Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) 229 { 230 // strip zeros 231 Index diagSize = (std::min)(_cols,_depth); 232 Index rows = _rows; 233 Index depth = IsLower ? _depth : diagSize; 234 Index cols = IsLower ? diagSize : _cols; 235 236 const_blas_data_mapper<Scalar, Index, LhsStorageOrder> lhs(_lhs,lhsStride); 237 const_blas_data_mapper<Scalar, Index, RhsStorageOrder> rhs(_rhs,rhsStride); 238 239 Index kc = blocking.kc(); // cache block size along the K direction 240 Index mc = (std::min)(rows,blocking.mc()); // cache block size along the M direction 241 242 std::size_t sizeA = kc*mc; 243 std::size_t sizeB = kc*cols; 244 std::size_t sizeW = kc*Traits::WorkSpaceFactor; 245 246 ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA()); 247 ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB()); 248 ei_declare_aligned_stack_constructed_variable(Scalar, blockW, sizeW, blocking.blockW()); 249 250 Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer; 251 triangularBuffer.setZero(); 252 if((Mode&ZeroDiag)==ZeroDiag) 253 triangularBuffer.diagonal().setZero(); 254 else 255 triangularBuffer.diagonal().setOnes(); 256 257 gebp_kernel<Scalar, Scalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel; 258 gemm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs; 259 gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs; 260 gemm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel; 261 262 for(Index k2=IsLower ? 0 : depth; 263 IsLower ? k2<depth : k2>0; 264 IsLower ? k2+=kc : k2-=kc) 265 { 266 Index actual_kc = (std::min)(IsLower ? depth-k2 : k2, kc); 267 Index actual_k2 = IsLower ? k2 : k2-actual_kc; 268 269 // align blocks with the end of the triangular part for trapezoidal rhs 270 if(IsLower && (k2<cols) && (actual_k2+actual_kc>cols)) 271 { 272 actual_kc = cols-k2; 273 k2 = actual_k2 + actual_kc - kc; 274 } 275 276 // remaining size 277 Index rs = IsLower ? (std::min)(cols,actual_k2) : cols - k2; 278 // size of the triangular part 279 Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc; 280 281 Scalar* geb = blockB+ts*ts; 282 283 pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, actual_kc, rs); 284 285 // pack the triangular part of the rhs padding the unrolled blocks with zeros 286 if(ts>0) 287 { 288 for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth) 289 { 290 Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth); 291 Index actual_j2 = actual_k2 + j2; 292 Index panelOffset = IsLower ? j2+actualPanelWidth : 0; 293 Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2; 294 // general part 295 pack_rhs_panel(blockB+j2*actual_kc, 296 &rhs(actual_k2+panelOffset, actual_j2), rhsStride, 297 panelLength, actualPanelWidth, 298 actual_kc, panelOffset); 299 300 // append the triangular part via a temporary buffer 301 for (Index j=0;j<actualPanelWidth;++j) 302 { 303 if (SetDiag) 304 triangularBuffer.coeffRef(j,j) = rhs(actual_j2+j,actual_j2+j); 305 for (Index k=IsLower ? j+1 : 0; IsLower ? k<actualPanelWidth : k<j; ++k) 306 triangularBuffer.coeffRef(k,j) = rhs(actual_j2+k,actual_j2+j); 307 } 308 309 pack_rhs_panel(blockB+j2*actual_kc, 310 triangularBuffer.data(), triangularBuffer.outerStride(), 311 actualPanelWidth, actualPanelWidth, 312 actual_kc, j2); 313 } 314 } 315 316 for (Index i2=0; i2<rows; i2+=mc) 317 { 318 const Index actual_mc = (std::min)(mc,rows-i2); 319 pack_lhs(blockA, &lhs(i2, actual_k2), lhsStride, actual_kc, actual_mc); 320 321 // triangular kernel 322 if(ts>0) 323 { 324 for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth) 325 { 326 Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth); 327 Index panelLength = IsLower ? actual_kc-j2 : j2+actualPanelWidth; 328 Index blockOffset = IsLower ? j2 : 0; 329 330 gebp_kernel(res+i2+(actual_k2+j2)*resStride, resStride, 331 blockA, blockB+j2*actual_kc, 332 actual_mc, panelLength, actualPanelWidth, 333 alpha, 334 actual_kc, actual_kc, // strides 335 blockOffset, blockOffset,// offsets 336 blockW); // workspace 337 } 338 } 339 gebp_kernel(res+i2+(IsLower ? 0 : k2)*resStride, resStride, 340 blockA, geb, actual_mc, actual_kc, rs, 341 alpha, 342 -1, -1, 0, 0, blockW); 343 } 344 } 345 } 346 }; 347 348 /*************************************************************************** 349 * Wrapper to product_triangular_matrix_matrix 350 ***************************************************************************/ 351 352 template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs> 353 struct traits<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false> > 354 : traits<ProductBase<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false>, Lhs, Rhs> > 355 {}; 356 357 } // end namespace internal 358 359 template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs> 360 struct TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false> 361 : public ProductBase<TriangularProduct<Mode,LhsIsTriangular,Lhs,false,Rhs,false>, Lhs, Rhs > 362 { 363 EIGEN_PRODUCT_PUBLIC_INTERFACE(TriangularProduct) 364 365 TriangularProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs) {} 366 367 template<typename Dest> void scaleAndAddTo(Dest& dst, Scalar alpha) const 368 { 369 typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs); 370 typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs); 371 372 Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(m_lhs) 373 * RhsBlasTraits::extractScalarFactor(m_rhs); 374 375 typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar, 376 Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType; 377 378 enum { IsLower = (Mode&Lower) == Lower }; 379 Index stripedRows = ((!LhsIsTriangular) || (IsLower)) ? lhs.rows() : (std::min)(lhs.rows(),lhs.cols()); 380 Index stripedCols = ((LhsIsTriangular) || (!IsLower)) ? rhs.cols() : (std::min)(rhs.cols(),rhs.rows()); 381 Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows())) 382 : ((IsLower) ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols())); 383 384 BlockingType blocking(stripedRows, stripedCols, stripedDepth); 385 386 internal::product_triangular_matrix_matrix<Scalar, Index, 387 Mode, LhsIsTriangular, 388 (internal::traits<_ActualLhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate, 389 (internal::traits<_ActualRhsType>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate, 390 (internal::traits<Dest >::Flags&RowMajorBit) ? RowMajor : ColMajor> 391 ::run( 392 stripedRows, stripedCols, stripedDepth, // sizes 393 &lhs.coeffRef(0,0), lhs.outerStride(), // lhs info 394 &rhs.coeffRef(0,0), rhs.outerStride(), // rhs info 395 &dst.coeffRef(0,0), dst.outerStride(), // result info 396 actualAlpha, blocking 397 ); 398 } 399 }; 400 401 } // end namespace Eigen 402 403 #endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_H 404