1 2 #if !defined(__BAIJ_H) 3 #define __BAIJ_H 4 #include "private/matimpl.h" 5 #include "../src/mat/impls/aij/seq/aij.h" 6 #include "../src/mat/impls/baij/seq/ftn-kernels/fsolvebaij.h" 7 8 /* 9 MATSEQBAIJ format - Block compressed row storage. The i[] and j[] 10 arrays start at 0. 11 */ 12 13 /* This header is shared by the SeqSBAIJ matrix */ 14 #define SEQBAIJHEADER \ 15 PetscInt bs2; /* square of block size */ \ 16 PetscInt mbs,nbs; /* rows/bs, columns/bs */ \ 17 PetscScalar *mult_work; /* work array for matrix vector product*/ \ 18 MatScalar *saved_values; \ 19 \ 20 Mat sbaijMat; /* mat in sbaij format */ \ 21 \ 22 \ 23 MatScalar *idiag; /* inverse of block diagonal */ \ 24 PetscTruth idiagvalid /* if above has correct/current values */ 25 26 27 typedef struct { 28 SEQAIJHEADER(MatScalar); 29 SEQBAIJHEADER; 30 } Mat_SeqBAIJ; 31 32 EXTERN_C_BEGIN 33 EXTERN PetscErrorCode MatSeqBAIJSetPreallocation_SeqBAIJ(Mat,PetscInt,PetscInt,PetscInt*); 34 EXTERN_C_END 35 EXTERN PetscErrorCode MatILUFactorSymbolic_SeqBAIJ(Mat,Mat,IS,IS,const MatFactorInfo*); 36 EXTERN PetscErrorCode MatICCFactorSymbolic_SeqBAIJ(Mat,Mat,IS,const MatFactorInfo*); 37 EXTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqBAIJ(Mat,Mat,IS,const MatFactorInfo*); 38 EXTERN PetscErrorCode MatCholeskyFactorNumeric_SeqBAIJ_N(Mat,Mat,const MatFactorInfo*); 39 EXTERN PetscErrorCode MatCholeskyFactorNumeric_SeqBAIJ_N_NaturalOrdering(Mat,Mat,const MatFactorInfo*); 40 EXTERN PetscErrorCode MatDuplicate_SeqBAIJ(Mat,MatDuplicateOption,Mat*); 41 EXTERN PetscErrorCode MatMissingDiagonal_SeqBAIJ(Mat,PetscTruth*,PetscInt*); 42 EXTERN PetscErrorCode MatMarkDiagonal_SeqBAIJ(Mat); 43 EXTERN PetscErrorCode MatILUDTFactor_SeqBAIJ(Mat,IS,IS,const MatFactorInfo*,Mat*); 44 45 EXTERN PetscErrorCode MatLUFactorSymbolic_SeqBAIJ(Mat,Mat,IS,IS,const MatFactorInfo*); 46 EXTERN PetscErrorCode MatLUFactor_SeqBAIJ(Mat,IS,IS,const MatFactorInfo*); 47 EXTERN PetscErrorCode MatIncreaseOverlap_SeqBAIJ(Mat,PetscInt,IS*,PetscInt); 48 EXTERN PetscErrorCode MatGetSubMatrix_SeqBAIJ(Mat,IS,IS,MatReuse,Mat*); 49 EXTERN PetscErrorCode MatGetSubMatrices_SeqBAIJ(Mat,PetscInt,const IS[],const IS[],MatReuse,Mat*[]); 50 EXTERN PetscErrorCode MatMultTranspose_SeqBAIJ(Mat,Vec,Vec); 51 EXTERN PetscErrorCode MatMultTransposeAdd_SeqBAIJ(Mat,Vec,Vec,Vec); 52 EXTERN PetscErrorCode MatScale_SeqBAIJ(Mat,PetscScalar); 53 EXTERN PetscErrorCode MatNorm_SeqBAIJ(Mat,NormType,PetscReal *); 54 EXTERN PetscErrorCode MatEqual_SeqBAIJ(Mat,Mat,PetscTruth*); 55 EXTERN PetscErrorCode MatGetDiagonal_SeqBAIJ(Mat,Vec); 56 EXTERN PetscErrorCode MatDiagonalScale_SeqBAIJ(Mat,Vec,Vec); 57 EXTERN PetscErrorCode MatGetInfo_SeqBAIJ(Mat,MatInfoType,MatInfo *); 58 EXTERN PetscErrorCode MatZeroEntries_SeqBAIJ(Mat); 59 60 EXTERN PetscErrorCode MatSeqBAIJ_UpdateFactorNumeric_NaturalOrdering(Mat); 61 62 EXTERN PetscErrorCode MatSolve_SeqBAIJ_1(Mat,Vec,Vec); 63 EXTERN PetscErrorCode MatSolve_SeqBAIJ_1_NaturalOrdering(Mat,Vec,Vec); 64 EXTERN PetscErrorCode MatSolve_SeqBAIJ_2(Mat,Vec,Vec); 65 EXTERN PetscErrorCode MatSolve_SeqBAIJ_2_newdatastruct(Mat,Vec,Vec); 66 EXTERN PetscErrorCode MatSolve_SeqBAIJ_2_NaturalOrdering(Mat,Vec,Vec); 67 EXTERN PetscErrorCode MatSolve_SeqBAIJ_2_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 68 EXTERN PetscErrorCode MatSolve_SeqBAIJ_2_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 69 EXTERN PetscErrorCode MatSolve_SeqBAIJ_3_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 70 EXTERN PetscErrorCode MatSolve_SeqBAIJ_3_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 71 EXTERN PetscErrorCode MatSolve_SeqBAIJ_3(Mat,Vec,Vec); 72 EXTERN PetscErrorCode MatSolve_SeqBAIJ_3_newdatastruct(Mat,Vec,Vec); 73 EXTERN PetscErrorCode MatSolve_SeqBAIJ_3_NaturalOrdering(Mat,Vec,Vec); 74 EXTERN PetscErrorCode MatSolve_SeqBAIJ_3_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 75 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4(Mat,Vec,Vec); 76 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4_newdatastruct(Mat,Vec,Vec); 77 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4_NaturalOrdering(Mat,Vec,Vec); 78 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 79 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 80 #if defined(PETSC_HAVE_SSE) 81 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4_SSE_Demotion(Mat,Vec,Vec); 82 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4_NaturalOrdering_SSE_Demotion(Mat,Vec,Vec); 83 EXTERN PetscErrorCode MatSolve_SeqBAIJ_4_NaturalOrdering_SSE_Demotion_usj(Mat,Vec,Vec); 84 #endif 85 EXTERN PetscErrorCode MatSolve_SeqBAIJ_5(Mat,Vec,Vec); 86 EXTERN PetscErrorCode MatSolve_SeqBAIJ_5_newdatastruct(Mat,Vec,Vec); 87 EXTERN PetscErrorCode MatSolve_SeqBAIJ_5_NaturalOrdering(Mat,Vec,Vec); 88 EXTERN PetscErrorCode MatSolve_SeqBAIJ_5_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 89 EXTERN PetscErrorCode MatSolve_SeqBAIJ_5_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 90 EXTERN PetscErrorCode MatSolve_SeqBAIJ_6(Mat,Vec,Vec); 91 EXTERN PetscErrorCode MatSolve_SeqBAIJ_6_newdatastruct(Mat,Vec,Vec); 92 EXTERN PetscErrorCode MatSolve_SeqBAIJ_6_NaturalOrdering(Mat,Vec,Vec); 93 EXTERN PetscErrorCode MatSolve_SeqBAIJ_6_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 94 EXTERN PetscErrorCode MatSolve_SeqBAIJ_6_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 95 EXTERN PetscErrorCode MatSolve_SeqBAIJ_7(Mat,Vec,Vec); 96 EXTERN PetscErrorCode MatSolve_SeqBAIJ_7_newdatastruct(Mat,Vec,Vec); 97 EXTERN PetscErrorCode MatSolve_SeqBAIJ_7_NaturalOrdering(Mat,Vec,Vec); 98 EXTERN PetscErrorCode MatSolve_SeqBAIJ_7_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 99 EXTERN PetscErrorCode MatSolve_SeqBAIJ_7_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 100 EXTERN PetscErrorCode MatSolve_SeqBAIJ_N(Mat,Vec,Vec); 101 EXTERN PetscErrorCode MatSolve_SeqBAIJ_N_newdatastruct(Mat,Vec,Vec); 102 EXTERN PetscErrorCode MatSolve_SeqBAIJ_N_NaturalOrdering_newdatastruct(Mat,Vec,Vec); 103 EXTERN PetscErrorCode MatSolve_SeqBAIJ_N_NaturalOrdering_newdatastruct_v2(Mat,Vec,Vec); 104 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_Update(Mat,Vec,Vec); 105 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_1(Mat,Vec,Vec); 106 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_1_NaturalOrdering(Mat,Vec,Vec); 107 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_2(Mat,Vec,Vec); 108 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_2_NaturalOrdering(Mat,Vec,Vec); 109 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_3(Mat,Vec,Vec); 110 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_3_NaturalOrdering(Mat,Vec,Vec); 111 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_4(Mat,Vec,Vec); 112 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_4_NaturalOrdering(Mat,Vec,Vec); 113 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_5(Mat,Vec,Vec); 114 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_5_NaturalOrdering(Mat,Vec,Vec); 115 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_6(Mat,Vec,Vec); 116 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_6_NaturalOrdering(Mat,Vec,Vec); 117 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_7(Mat,Vec,Vec); 118 EXTERN PetscErrorCode MatSolveTranspose_SeqBAIJ_7_NaturalOrdering(Mat,Vec,Vec); 119 120 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_N_newdatastruct(Mat,Mat,const MatFactorInfo*); 121 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_N_newdatastruct_v2(Mat,Mat,const MatFactorInfo*); 122 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_1(Mat,Mat,const MatFactorInfo*); 123 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_2(Mat,Mat,const MatFactorInfo*); 124 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_2_newdatastruct(Mat,Mat,const MatFactorInfo*); 125 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_2_NaturalOrdering(Mat,Mat,const MatFactorInfo*); 126 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_2_NaturalOrdering_newdatastruct(Mat,Mat,const MatFactorInfo*); 127 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_2_NaturalOrdering_newdatastruct_v2(Mat,Mat,const MatFactorInfo*); 128 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3(Mat,Mat,const MatFactorInfo*); 129 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3_newdatastruct(Mat,Mat,const MatFactorInfo*); 130 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3_NaturalOrdering(Mat,Mat,const MatFactorInfo*); 131 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3_NaturalOrdering_newdatastruct(Mat,Mat,const MatFactorInfo*); 132 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_3_NaturalOrdering_newdatastruct_v2(Mat,Mat,const MatFactorInfo*); 133 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4(Mat,Mat,const MatFactorInfo*); 134 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4_newdatastruct(Mat,Mat,const MatFactorInfo*); 135 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering(Mat,Mat,const MatFactorInfo*); 136 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering_newdatastruct(Mat,Mat,const MatFactorInfo*); 137 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering_newdatastruct_v2(Mat,Mat,const MatFactorInfo*); 138 #if defined(PETSC_HAVE_SSE) 139 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering_SSE(Mat,Mat,const MatFactorInfo*); 140 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_4_NaturalOrdering_SSE_usj(Mat,Mat,const MatFactorInfo*); 141 #else 142 #endif 143 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5(Mat,Mat,const MatFactorInfo*); 144 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_newdatastruct(Mat,Mat,const MatFactorInfo*); 145 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering(Mat,Mat,const MatFactorInfo*); 146 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering_newdatastruct(Mat,Mat,const MatFactorInfo*); 147 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering_newdatastruct_v2(Mat,Mat,const MatFactorInfo*); 148 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_6(Mat,Mat,const MatFactorInfo*); 149 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_6_newdatastruct(Mat,Mat,const MatFactorInfo*); 150 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_6_NaturalOrdering(Mat,Mat,const MatFactorInfo*); 151 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_6_NaturalOrdering_newdatastruct(Mat,Mat,const MatFactorInfo*); 152 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_6_NaturalOrdering_newdatastruct_v2(Mat,Mat,const MatFactorInfo*); 153 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_7(Mat,Mat,const MatFactorInfo*); 154 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_7_newdatastruct(Mat,Mat,const MatFactorInfo*); 155 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_7_NaturalOrdering(Mat,Mat,const MatFactorInfo*); 156 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_7_NaturalOrdering_newdatastruct(Mat,Mat,const MatFactorInfo*); 157 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_7_NaturalOrdering_newdatastruct_v2(Mat,Mat,const MatFactorInfo*); 158 EXTERN PetscErrorCode MatLUFactorNumeric_SeqBAIJ_N(Mat,Mat,const MatFactorInfo*); 159 160 EXTERN PetscErrorCode MatMult_SeqBAIJ_1(Mat,Vec,Vec); 161 EXTERN PetscErrorCode MatMult_SeqBAIJ_2(Mat,Vec,Vec); 162 EXTERN PetscErrorCode MatMult_SeqBAIJ_3(Mat,Vec,Vec); 163 EXTERN PetscErrorCode MatMult_SeqBAIJ_4(Mat,Vec,Vec); 164 EXTERN PetscErrorCode MatMult_SeqBAIJ_5(Mat,Vec,Vec); 165 EXTERN PetscErrorCode MatMult_SeqBAIJ_6(Mat,Vec,Vec); 166 EXTERN PetscErrorCode MatMult_SeqBAIJ_7(Mat,Vec,Vec); 167 EXTERN PetscErrorCode MatMult_SeqBAIJ_N(Mat,Vec,Vec); 168 169 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_1(Mat,Vec,Vec,Vec); 170 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_2(Mat,Vec,Vec,Vec); 171 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_3(Mat,Vec,Vec,Vec); 172 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_4(Mat,Vec,Vec,Vec); 173 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_5(Mat,Vec,Vec,Vec); 174 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_6(Mat,Vec,Vec,Vec); 175 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_7(Mat,Vec,Vec,Vec); 176 EXTERN PetscErrorCode MatMultAdd_SeqBAIJ_N(Mat,Vec,Vec,Vec); 177 EXTERN PetscErrorCode MatLoad_SeqBAIJ(PetscViewer, const MatType,Mat*); 178 179 /* 180 Kernel_A_gets_A_times_B_2: A = A * B with size bs=2 181 182 Input Parameters: 183 + A,B - square bs by bs arrays stored in column major order 184 - W - bs*bs work arrary 185 186 Output Parameter: 187 . A = A * B 188 */ 189 190 #define Kernel_A_gets_A_times_B_2(A,B,W) 0;\ 191 {\ 192 PetscMemcpy(W,A,4*sizeof(MatScalar));\ 193 A[0] = W[0]*B[0] + W[2]*B[1];\ 194 A[1] = W[1]*B[0] + W[3]*B[1];\ 195 A[2] = W[0]*B[2] + W[2]*B[3];\ 196 A[3] = W[1]*B[2] + W[3]*B[3];\ 197 } 198 199 /* 200 Kernel_A_gets_A_minus_B_times_C_2: A = A - B * C with size bs=2 201 202 Input Parameters: 203 + A,B,C - square bs by bs arrays stored in column major order 204 205 Output Parameter: 206 . A = A - B*C 207 */ 208 209 #define Kernel_A_gets_A_minus_B_times_C_2(A,B,C) 0;\ 210 {\ 211 A[0] -= B[0]*C[0] + B[2]*C[1];\ 212 A[1] -= B[1]*C[0] + B[3]*C[1];\ 213 A[2] -= B[0]*C[2] + B[2]*C[3];\ 214 A[3] -= B[1]*C[2] + B[3]*C[3];\ 215 } 216 217 /* 218 Kernel_A_gets_A_times_B_3: A = A * B with size bs=3 219 220 Input Parameters: 221 + A,B - square bs by bs arrays stored in column major order 222 - W - bs*bs work arrary 223 224 Output Parameter: 225 . A = A * B 226 */ 227 228 #define Kernel_A_gets_A_times_B_3(A,B,W) 0;\ 229 {\ 230 PetscMemcpy(W,A,9*sizeof(MatScalar));\ 231 A[0] = W[0]*B[0] + W[3]*B[1] + W[6]*B[2];\ 232 A[1] = W[1]*B[0] + W[4]*B[1] + W[7]*B[2];\ 233 A[2] = W[2]*B[0] + W[5]*B[1] + W[8]*B[2];\ 234 A[3] = W[0]*B[3] + W[3]*B[4] + W[6]*B[5];\ 235 A[4] = W[1]*B[3] + W[4]*B[4] + W[7]*B[5];\ 236 A[5] = W[2]*B[3] + W[5]*B[4] + W[8]*B[5];\ 237 A[6] = W[0]*B[6] + W[3]*B[7] + W[6]*B[8];\ 238 A[7] = W[1]*B[6] + W[4]*B[7] + W[7]*B[8];\ 239 A[8] = W[2]*B[6] + W[5]*B[7] + W[8]*B[8];\ 240 } 241 242 /* 243 Kernel_A_gets_A_minus_B_times_C_3: A = A - B * C with size bs=3 244 245 Input Parameters: 246 + A,B,C - square bs by bs arrays stored in column major order 247 248 Output Parameter: 249 . A = A - B*C 250 */ 251 252 #define Kernel_A_gets_A_minus_B_times_C_3(A,B,C) 0;\ 253 {\ 254 A[0] -= B[0]*C[0] + B[3]*C[1] + B[6]*C[2];\ 255 A[1] -= B[1]*C[0] + B[4]*C[1] + B[7]*C[2];\ 256 A[2] -= B[2]*C[0] + B[5]*C[1] + B[8]*C[2];\ 257 A[3] -= B[0]*C[3] + B[3]*C[4] + B[6]*C[5];\ 258 A[4] -= B[1]*C[3] + B[4]*C[4] + B[7]*C[5];\ 259 A[5] -= B[2]*C[3] + B[5]*C[4] + B[8]*C[5];\ 260 A[6] -= B[0]*C[6] + B[3]*C[7] + B[6]*C[8];\ 261 A[7] -= B[1]*C[6] + B[4]*C[7] + B[7]*C[8];\ 262 A[8] -= B[2]*C[6] + B[5]*C[7] + B[8]*C[8];\ 263 } 264 265 /* 266 Kernel_A_gets_A_times_B_4: A = A * B with size bs=4 267 268 Input Parameters: 269 + A,B - square bs by bs arrays stored in column major order 270 - W - bs*bs work arrary 271 272 Output Parameter: 273 . A = A * B 274 */ 275 276 #define Kernel_A_gets_A_times_B_4(A,B,W) 0;\ 277 {\ 278 PetscMemcpy(W,A,16*sizeof(MatScalar));\ 279 A[0] = W[0]*B[0] + W[4]*B[1] + W[8]*B[2] + W[12]*B[3];\ 280 A[1] = W[1]*B[0] + W[5]*B[1] + W[9]*B[2] + W[13]*B[3];\ 281 A[2] = W[2]*B[0] + W[6]*B[1] + W[10]*B[2] + W[14]*B[3];\ 282 A[3] = W[3]*B[0] + W[7]*B[1] + W[11]*B[2] + W[15]*B[3];\ 283 A[4] = W[0]*B[4] + W[4]*B[5] + W[8]*B[6] + W[12]*B[7];\ 284 A[5] = W[1]*B[4] + W[5]*B[5] + W[9]*B[6] + W[13]*B[7];\ 285 A[6] = W[2]*B[4] + W[6]*B[5] + W[10]*B[6] + W[14]*B[7];\ 286 A[7] = W[3]*B[4] + W[7]*B[5] + W[11]*B[6] + W[15]*B[7];\ 287 A[8] = W[0]*B[8] + W[4]*B[9] + W[8]*B[10] + W[12]*B[11];\ 288 A[9] = W[1]*B[8] + W[5]*B[9] + W[9]*B[10] + W[13]*B[11];\ 289 A[10] = W[2]*B[8] + W[6]*B[9] + W[10]*B[10] + W[14]*B[11];\ 290 A[11] = W[3]*B[8] + W[7]*B[9] + W[11]*B[10] + W[15]*B[11];\ 291 A[12] = W[0]*B[12] + W[4]*B[13] + W[8]*B[14] + W[12]*B[15];\ 292 A[13] = W[1]*B[12] + W[5]*B[13] + W[9]*B[14] + W[13]*B[15];\ 293 A[14] = W[2]*B[12] + W[6]*B[13] + W[10]*B[14] + W[14]*B[15];\ 294 A[15] = W[3]*B[12] + W[7]*B[13] + W[11]*B[14] + W[15]*B[15];\ 295 } 296 297 /* 298 Kernel_A_gets_A_minus_B_times_C_4: A = A - B * C with size bs=4 299 300 Input Parameters: 301 + A,B,C - square bs by bs arrays stored in column major order 302 303 Output Parameter: 304 . A = A - B*C 305 */ 306 307 #define Kernel_A_gets_A_minus_B_times_C_4(A,B,C) 0;\ 308 {\ 309 A[0] -= B[0]*C[0] + B[4]*C[1] + B[8]*C[2] + B[12]*C[3];\ 310 A[1] -= B[1]*C[0] + B[5]*C[1] + B[9]*C[2] + B[13]*C[3];\ 311 A[2] -= B[2]*C[0] + B[6]*C[1] + B[10]*C[2] + B[14]*C[3];\ 312 A[3] -= B[3]*C[0] + B[7]*C[1] + B[11]*C[2] + B[15]*C[3];\ 313 A[4] -= B[0]*C[4] + B[4]*C[5] + B[8]*C[6] + B[12]*C[7];\ 314 A[5] -= B[1]*C[4] + B[5]*C[5] + B[9]*C[6] + B[13]*C[7];\ 315 A[6] -= B[2]*C[4] + B[6]*C[5] + B[10]*C[6] + B[14]*C[7];\ 316 A[7] -= B[3]*C[4] + B[7]*C[5] + B[11]*C[6] + B[15]*C[7];\ 317 A[8] -= B[0]*C[8] + B[4]*C[9] + B[8]*C[10] + B[12]*C[11];\ 318 A[9] -= B[1]*C[8] + B[5]*C[9] + B[9]*C[10] + B[13]*C[11];\ 319 A[10] -= B[2]*C[8] + B[6]*C[9] + B[10]*C[10] + B[14]*C[11];\ 320 A[11] -= B[3]*C[8] + B[7]*C[9] + B[11]*C[10] + B[15]*C[11];\ 321 A[12] -= B[0]*C[12] + B[4]*C[13] + B[8]*C[14] + B[12]*C[15];\ 322 A[13] -= B[1]*C[12] + B[5]*C[13] + B[9]*C[14] + B[13]*C[15];\ 323 A[14] -= B[2]*C[12] + B[6]*C[13] + B[10]*C[14] + B[14]*C[15];\ 324 A[15] -= B[3]*C[12] + B[7]*C[13] + B[11]*C[14] + B[15]*C[15];\ 325 } 326 327 #define Kernel_A_gets_A_times_B_5(A,B,W) 0;\ 328 {\ 329 PetscMemcpy(W,A,25*sizeof(MatScalar));\ 330 A[0] = W[0]*B[0] + W[5]*B[1] + W[10]*B[2] + W[15]*B[3] + W[20]*B[4];\ 331 A[1] = W[1]*B[0] + W[6]*B[1] + W[11]*B[2] + W[16]*B[3] + W[21]*B[4];\ 332 A[2] = W[2]*B[0] + W[7]*B[1] + W[12]*B[2] + W[17]*B[3] + W[22]*B[4];\ 333 A[3] = W[3]*B[0] + W[8]*B[1] + W[13]*B[2] + W[18]*B[3] + W[23]*B[4];\ 334 A[4] = W[4]*B[0] + W[9]*B[1] + W[14]*B[2] + W[19]*B[3] + W[24]*B[4];\ 335 A[5] = W[0]*B[5] + W[5]*B[6] + W[10]*B[7] + W[15]*B[8] + W[20]*B[9];\ 336 A[6] = W[1]*B[5] + W[6]*B[6] + W[11]*B[7] + W[16]*B[8] + W[21]*B[9];\ 337 A[7] = W[2]*B[5] + W[7]*B[6] + W[12]*B[7] + W[17]*B[8] + W[22]*B[9];\ 338 A[8] = W[3]*B[5] + W[8]*B[6] + W[13]*B[7] + W[18]*B[8] + W[23]*B[9];\ 339 A[9] = W[4]*B[5] + W[9]*B[6] + W[14]*B[7] + W[19]*B[8] + W[24]*B[9];\ 340 A[10] = W[0]*B[10] + W[5]*B[11] + W[10]*B[12] + W[15]*B[13] + W[20]*B[14];\ 341 A[11] = W[1]*B[10] + W[6]*B[11] + W[11]*B[12] + W[16]*B[13] + W[21]*B[14];\ 342 A[12] = W[2]*B[10] + W[7]*B[11] + W[12]*B[12] + W[17]*B[13] + W[22]*B[14];\ 343 A[13] = W[3]*B[10] + W[8]*B[11] + W[13]*B[12] + W[18]*B[13] + W[23]*B[14];\ 344 A[14] = W[4]*B[10] + W[9]*B[11] + W[14]*B[12] + W[19]*B[13] + W[24]*B[14];\ 345 A[15] = W[0]*B[15] + W[5]*B[16] + W[10]*B[17] + W[15]*B[18] + W[20]*B[19];\ 346 A[16] = W[1]*B[15] + W[6]*B[16] + W[11]*B[17] + W[16]*B[18] + W[21]*B[19];\ 347 A[17] = W[2]*B[15] + W[7]*B[16] + W[12]*B[17] + W[17]*B[18] + W[22]*B[19];\ 348 A[18] = W[3]*B[15] + W[8]*B[16] + W[13]*B[17] + W[18]*B[18] + W[23]*B[19];\ 349 A[19] = W[4]*B[15] + W[9]*B[16] + W[14]*B[17] + W[19]*B[18] + W[24]*B[19];\ 350 A[20] = W[0]*B[20] + W[5]*B[21] + W[10]*B[22] + W[15]*B[23] + W[20]*B[24];\ 351 A[21] = W[1]*B[20] + W[6]*B[21] + W[11]*B[22] + W[16]*B[23] + W[21]*B[24];\ 352 A[22] = W[2]*B[20] + W[7]*B[21] + W[12]*B[22] + W[17]*B[23] + W[22]*B[24];\ 353 A[23] = W[3]*B[20] + W[8]*B[21] + W[13]*B[22] + W[18]*B[23] + W[23]*B[24];\ 354 A[24] = W[4]*B[20] + W[9]*B[21] + W[14]*B[22] + W[19]*B[23] + W[24]*B[24];\ 355 } 356 357 /* 358 Kernel_A_gets_A_minus_B_times_C_5: A = A - B * C with size bs=5 359 360 Input Parameters: 361 + A,B,C - square bs by bs arrays stored in column major order 362 363 Output Parameter: 364 . A = A - B*C 365 */ 366 367 #define Kernel_A_gets_A_minus_B_times_C_5(A,B,C) 0;\ 368 {\ 369 A[0] -= B[0]*C[0] + B[5]*C[1] + B[10]*C[2] + B[15]*C[3] + B[20]*C[4];\ 370 A[1] -= B[1]*C[0] + B[6]*C[1] + B[11]*C[2] + B[16]*C[3] + B[21]*C[4];\ 371 A[2] -= B[2]*C[0] + B[7]*C[1] + B[12]*C[2] + B[17]*C[3] + B[22]*C[4];\ 372 A[3] -= B[3]*C[0] + B[8]*C[1] + B[13]*C[2] + B[18]*C[3] + B[23]*C[4];\ 373 A[4] -= B[4]*C[0] + B[9]*C[1] + B[14]*C[2] + B[19]*C[3] + B[24]*C[4];\ 374 A[5] -= B[0]*C[5] + B[5]*C[6] + B[10]*C[7] + B[15]*C[8] + B[20]*C[9];\ 375 A[6] -= B[1]*C[5] + B[6]*C[6] + B[11]*C[7] + B[16]*C[8] + B[21]*C[9];\ 376 A[7] -= B[2]*C[5] + B[7]*C[6] + B[12]*C[7] + B[17]*C[8] + B[22]*C[9];\ 377 A[8] -= B[3]*C[5] + B[8]*C[6] + B[13]*C[7] + B[18]*C[8] + B[23]*C[9];\ 378 A[9] -= B[4]*C[5] + B[9]*C[6] + B[14]*C[7] + B[19]*C[8] + B[24]*C[9];\ 379 A[10] -= B[0]*C[10] + B[5]*C[11] + B[10]*C[12] + B[15]*C[13] + B[20]*C[14];\ 380 A[11] -= B[1]*C[10] + B[6]*C[11] + B[11]*C[12] + B[16]*C[13] + B[21]*C[14];\ 381 A[12] -= B[2]*C[10] + B[7]*C[11] + B[12]*C[12] + B[17]*C[13] + B[22]*C[14];\ 382 A[13] -= B[3]*C[10] + B[8]*C[11] + B[13]*C[12] + B[18]*C[13] + B[23]*C[14];\ 383 A[14] -= B[4]*C[10] + B[9]*C[11] + B[14]*C[12] + B[19]*C[13] + B[24]*C[14];\ 384 A[15] -= B[0]*C[15] + B[5]*C[16] + B[10]*C[17] + B[15]*C[18] + B[20]*C[19];\ 385 A[16] -= B[1]*C[15] + B[6]*C[16] + B[11]*C[17] + B[16]*C[18] + B[21]*C[19];\ 386 A[17] -= B[2]*C[15] + B[7]*C[16] + B[12]*C[17] + B[17]*C[18] + B[22]*C[19];\ 387 A[18] -= B[3]*C[15] + B[8]*C[16] + B[13]*C[17] + B[18]*C[18] + B[23]*C[19];\ 388 A[19] -= B[4]*C[15] + B[9]*C[16] + B[14]*C[17] + B[19]*C[18] + B[24]*C[19];\ 389 A[20] -= B[0]*C[20] + B[5]*C[21] + B[10]*C[22] + B[15]*C[23] + B[20]*C[24];\ 390 A[21] -= B[1]*C[20] + B[6]*C[21] + B[11]*C[22] + B[16]*C[23] + B[21]*C[24];\ 391 A[22] -= B[2]*C[20] + B[7]*C[21] + B[12]*C[22] + B[17]*C[23] + B[22]*C[24];\ 392 A[23] -= B[3]*C[20] + B[8]*C[21] + B[13]*C[22] + B[18]*C[23] + B[23]*C[24];\ 393 A[24] -= B[4]*C[20] + B[9]*C[21] + B[14]*C[22] + B[19]*C[23] + B[24]*C[24];\ 394 } 395 396 #define Kernel_A_gets_A_times_B_6(A,B,W) 0;\ 397 {\ 398 PetscMemcpy(W,A,36*sizeof(MatScalar));\ 399 A[0] = W[0]*B[0] + W[6]*B[1] + W[12]*B[2] + W[18]*B[3] + W[24]*B[4] + W[30]*B[5];\ 400 A[1] = W[1]*B[0] + W[7]*B[1] + W[13]*B[2] + W[19]*B[3] + W[25]*B[4] + W[31]*B[5];\ 401 A[2] = W[2]*B[0] + W[8]*B[1] + W[14]*B[2] + W[20]*B[3] + W[26]*B[4] + W[32]*B[5];\ 402 A[3] = W[3]*B[0] + W[9]*B[1] + W[15]*B[2] + W[21]*B[3] + W[27]*B[4] + W[33]*B[5];\ 403 A[4] = W[4]*B[0] + W[10]*B[1] + W[16]*B[2] + W[22]*B[3] + W[28]*B[4] + W[34]*B[5];\ 404 A[5] = W[5]*B[0] + W[11]*B[1] + W[17]*B[2] + W[23]*B[3] + W[29]*B[4] + W[35]*B[5];\ 405 A[6] = W[0]*B[6] + W[6]*B[7] + W[12]*B[8] + W[18]*B[9] + W[24]*B[10] + W[30]*B[11];\ 406 A[7] = W[1]*B[6] + W[7]*B[7] + W[13]*B[8] + W[19]*B[9] + W[25]*B[10] + W[31]*B[11];\ 407 A[8] = W[2]*B[6] + W[8]*B[7] + W[14]*B[8] + W[20]*B[9] + W[26]*B[10] + W[32]*B[11];\ 408 A[9] = W[3]*B[6] + W[9]*B[7] + W[15]*B[8] + W[21]*B[9] + W[27]*B[10] + W[33]*B[11];\ 409 A[10] = W[4]*B[6] + W[10]*B[7] + W[16]*B[8] + W[22]*B[9] + W[28]*B[10] + W[34]*B[11];\ 410 A[11] = W[5]*B[6] + W[11]*B[7] + W[17]*B[8] + W[23]*B[9] + W[29]*B[10] + W[35]*B[11];\ 411 A[12] = W[0]*B[12] + W[6]*B[13] + W[12]*B[14] + W[18]*B[15] + W[24]*B[16] + W[30]*B[17];\ 412 A[13] = W[1]*B[12] + W[7]*B[13] + W[13]*B[14] + W[19]*B[15] + W[25]*B[16] + W[31]*B[17];\ 413 A[14] = W[2]*B[12] + W[8]*B[13] + W[14]*B[14] + W[20]*B[15] + W[26]*B[16] + W[32]*B[17];\ 414 A[15] = W[3]*B[12] + W[9]*B[13] + W[15]*B[14] + W[21]*B[15] + W[27]*B[16] + W[33]*B[17];\ 415 A[16] = W[4]*B[12] + W[10]*B[13] + W[16]*B[14] + W[22]*B[15] + W[28]*B[16] + W[34]*B[17];\ 416 A[17] = W[5]*B[12] + W[11]*B[13] + W[17]*B[14] + W[23]*B[15] + W[29]*B[16] + W[35]*B[17];\ 417 A[18] = W[0]*B[18] + W[6]*B[19] + W[12]*B[20] + W[18]*B[21] + W[24]*B[22] + W[30]*B[23];\ 418 A[19] = W[1]*B[18] + W[7]*B[19] + W[13]*B[20] + W[19]*B[21] + W[25]*B[22] + W[31]*B[23];\ 419 A[20] = W[2]*B[18] + W[8]*B[19] + W[14]*B[20] + W[20]*B[21] + W[26]*B[22] + W[32]*B[23];\ 420 A[21] = W[3]*B[18] + W[9]*B[19] + W[15]*B[20] + W[21]*B[21] + W[27]*B[22] + W[33]*B[23];\ 421 A[22] = W[4]*B[18] + W[10]*B[19] + W[16]*B[20] + W[22]*B[21] + W[28]*B[22] + W[34]*B[23];\ 422 A[23] = W[5]*B[18] + W[11]*B[19] + W[17]*B[20] + W[23]*B[21] + W[29]*B[22] + W[35]*B[23];\ 423 A[24] = W[0]*B[24] + W[6]*B[25] + W[12]*B[26] + W[18]*B[27] + W[24]*B[28] + W[30]*B[29];\ 424 A[25] = W[1]*B[24] + W[7]*B[25] + W[13]*B[26] + W[19]*B[27] + W[25]*B[28] + W[31]*B[29];\ 425 A[26] = W[2]*B[24] + W[8]*B[25] + W[14]*B[26] + W[20]*B[27] + W[26]*B[28] + W[32]*B[29];\ 426 A[27] = W[3]*B[24] + W[9]*B[25] + W[15]*B[26] + W[21]*B[27] + W[27]*B[28] + W[33]*B[29];\ 427 A[28] = W[4]*B[24] + W[10]*B[25] + W[16]*B[26] + W[22]*B[27] + W[28]*B[28] + W[34]*B[29];\ 428 A[29] = W[5]*B[24] + W[11]*B[25] + W[17]*B[26] + W[23]*B[27] + W[29]*B[28] + W[35]*B[29];\ 429 A[30] = W[0]*B[30] + W[6]*B[31] + W[12]*B[32] + W[18]*B[33] + W[24]*B[34] + W[30]*B[35];\ 430 A[31] = W[1]*B[30] + W[7]*B[31] + W[13]*B[32] + W[19]*B[33] + W[25]*B[34] + W[31]*B[35];\ 431 A[32] = W[2]*B[30] + W[8]*B[31] + W[14]*B[32] + W[20]*B[33] + W[26]*B[34] + W[32]*B[35];\ 432 A[33] = W[3]*B[30] + W[9]*B[31] + W[15]*B[32] + W[21]*B[33] + W[27]*B[34] + W[33]*B[35];\ 433 A[34] = W[4]*B[30] + W[10]*B[31] + W[16]*B[32] + W[22]*B[33] + W[28]*B[34] + W[34]*B[35];\ 434 A[35] = W[5]*B[30] + W[11]*B[31] + W[17]*B[32] + W[23]*B[33] + W[29]*B[34] + W[35]*B[35];\ 435 } 436 437 /* 438 Kernel_A_gets_A_minus_B_times_C_6: A = A - B * C with size bs=6 439 440 Input Parameters: 441 + A,B,C - square bs by bs arrays stored in column major order 442 443 Output Parameter: 444 . A = A - B*C 445 */ 446 447 #define Kernel_A_gets_A_minus_B_times_C_6(A,B,C) 0;\ 448 {\ 449 A[0] -= B[0]*C[0] + B[6]*C[1] + B[12]*C[2] + B[18]*C[3] + B[24]*C[4] + B[30]*C[5];\ 450 A[1] -= B[1]*C[0] + B[7]*C[1] + B[13]*C[2] + B[19]*C[3] + B[25]*C[4] + B[31]*C[5];\ 451 A[2] -= B[2]*C[0] + B[8]*C[1] + B[14]*C[2] + B[20]*C[3] + B[26]*C[4] + B[32]*C[5];\ 452 A[3] -= B[3]*C[0] + B[9]*C[1] + B[15]*C[2] + B[21]*C[3] + B[27]*C[4] + B[33]*C[5];\ 453 A[4] -= B[4]*C[0] + B[10]*C[1] + B[16]*C[2] + B[22]*C[3] + B[28]*C[4] + B[34]*C[5];\ 454 A[5] -= B[5]*C[0] + B[11]*C[1] + B[17]*C[2] + B[23]*C[3] + B[29]*C[4] + B[35]*C[5];\ 455 A[6] -= B[0]*C[6] + B[6]*C[7] + B[12]*C[8] + B[18]*C[9] + B[24]*C[10] + B[30]*C[11];\ 456 A[7] -= B[1]*C[6] + B[7]*C[7] + B[13]*C[8] + B[19]*C[9] + B[25]*C[10] + B[31]*C[11];\ 457 A[8] -= B[2]*C[6] + B[8]*C[7] + B[14]*C[8] + B[20]*C[9] + B[26]*C[10] + B[32]*C[11];\ 458 A[9] -= B[3]*C[6] + B[9]*C[7] + B[15]*C[8] + B[21]*C[9] + B[27]*C[10] + B[33]*C[11];\ 459 A[10] -= B[4]*C[6] + B[10]*C[7] + B[16]*C[8] + B[22]*C[9] + B[28]*C[10] + B[34]*C[11];\ 460 A[11] -= B[5]*C[6] + B[11]*C[7] + B[17]*C[8] + B[23]*C[9] + B[29]*C[10] + B[35]*C[11];\ 461 A[12] -= B[0]*C[12] + B[6]*C[13] + B[12]*C[14] + B[18]*C[15] + B[24]*C[16] + B[30]*C[17];\ 462 A[13] -= B[1]*C[12] + B[7]*C[13] + B[13]*C[14] + B[19]*C[15] + B[25]*C[16] + B[31]*C[17];\ 463 A[14] -= B[2]*C[12] + B[8]*C[13] + B[14]*C[14] + B[20]*C[15] + B[26]*C[16] + B[32]*C[17];\ 464 A[15] -= B[3]*C[12] + B[9]*C[13] + B[15]*C[14] + B[21]*C[15] + B[27]*C[16] + B[33]*C[17];\ 465 A[16] -= B[4]*C[12] + B[10]*C[13] + B[16]*C[14] + B[22]*C[15] + B[28]*C[16] + B[34]*C[17];\ 466 A[17] -= B[5]*C[12] + B[11]*C[13] + B[17]*C[14] + B[23]*C[15] + B[29]*C[16] + B[35]*C[17];\ 467 A[18] -= B[0]*C[18] + B[6]*C[19] + B[12]*C[20] + B[18]*C[21] + B[24]*C[22] + B[30]*C[23];\ 468 A[19] -= B[1]*C[18] + B[7]*C[19] + B[13]*C[20] + B[19]*C[21] + B[25]*C[22] + B[31]*C[23];\ 469 A[20] -= B[2]*C[18] + B[8]*C[19] + B[14]*C[20] + B[20]*C[21] + B[26]*C[22] + B[32]*C[23];\ 470 A[21] -= B[3]*C[18] + B[9]*C[19] + B[15]*C[20] + B[21]*C[21] + B[27]*C[22] + B[33]*C[23];\ 471 A[22] -= B[4]*C[18] + B[10]*C[19] + B[16]*C[20] + B[22]*C[21] + B[28]*C[22] + B[34]*C[23];\ 472 A[23] -= B[5]*C[18] + B[11]*C[19] + B[17]*C[20] + B[23]*C[21] + B[29]*C[22] + B[35]*C[23];\ 473 A[24] -= B[0]*C[24] + B[6]*C[25] + B[12]*C[26] + B[18]*C[27] + B[24]*C[28] + B[30]*C[29];\ 474 A[25] -= B[1]*C[24] + B[7]*C[25] + B[13]*C[26] + B[19]*C[27] + B[25]*C[28] + B[31]*C[29];\ 475 A[26] -= B[2]*C[24] + B[8]*C[25] + B[14]*C[26] + B[20]*C[27] + B[26]*C[28] + B[32]*C[29];\ 476 A[27] -= B[3]*C[24] + B[9]*C[25] + B[15]*C[26] + B[21]*C[27] + B[27]*C[28] + B[33]*C[29];\ 477 A[28] -= B[4]*C[24] + B[10]*C[25] + B[16]*C[26] + B[22]*C[27] + B[28]*C[28] + B[34]*C[29];\ 478 A[29] -= B[5]*C[24] + B[11]*C[25] + B[17]*C[26] + B[23]*C[27] + B[29]*C[28] + B[35]*C[29];\ 479 A[30] -= B[0]*C[30] + B[6]*C[31] + B[12]*C[32] + B[18]*C[33] + B[24]*C[34] + B[30]*C[35];\ 480 A[31] -= B[1]*C[30] + B[7]*C[31] + B[13]*C[32] + B[19]*C[33] + B[25]*C[34] + B[31]*C[35];\ 481 A[32] -= B[2]*C[30] + B[8]*C[31] + B[14]*C[32] + B[20]*C[33] + B[26]*C[34] + B[32]*C[35];\ 482 A[33] -= B[3]*C[30] + B[9]*C[31] + B[15]*C[32] + B[21]*C[33] + B[27]*C[34] + B[33]*C[35];\ 483 A[34] -= B[4]*C[30] + B[10]*C[31] + B[16]*C[32] + B[22]*C[33] + B[28]*C[34] + B[34]*C[35];\ 484 A[35] -= B[5]*C[30] + B[11]*C[31] + B[17]*C[32] + B[23]*C[33] + B[29]*C[34] + B[35]*C[35];\ 485 } 486 487 #define Kernel_A_gets_A_times_B_7(A,B,W) 0;\ 488 {\ 489 PetscMemcpy(W,A,49*sizeof(MatScalar));\ 490 A[0] = W[0]*B[0] + W[7]*B[1] + W[14]*B[2] + W[21]*B[3] + W[28]*B[4] + W[35]*B[5] + W[42]*B[6];\ 491 A[1] = W[1]*B[0] + W[8]*B[1] + W[15]*B[2] + W[22]*B[3] + W[29]*B[4] + W[36]*B[5] + W[43]*B[6];\ 492 A[2] = W[2]*B[0] + W[9]*B[1] + W[16]*B[2] + W[23]*B[3] + W[30]*B[4] + W[37]*B[5] + W[44]*B[6];\ 493 A[3] = W[3]*B[0] + W[10]*B[1] + W[17]*B[2] + W[24]*B[3] + W[31]*B[4] + W[38]*B[5] + W[45]*B[6];\ 494 A[4] = W[4]*B[0] + W[11]*B[1] + W[18]*B[2] + W[25]*B[3] + W[32]*B[4] + W[39]*B[5] + W[46]*B[6];\ 495 A[5] = W[5]*B[0] + W[12]*B[1] + W[19]*B[2] + W[26]*B[3] + W[33]*B[4] + W[40]*B[5] + W[47]*B[6];\ 496 A[6] = W[6]*B[0] + W[13]*B[1] + W[20]*B[2] + W[27]*B[3] + W[34]*B[4] + W[41]*B[5] + W[48]*B[6];\ 497 A[7] = W[0]*B[7] + W[7]*B[8] + W[14]*B[9] + W[21]*B[10] + W[28]*B[11] + W[35]*B[12] + W[42]*B[13];\ 498 A[8] = W[1]*B[7] + W[8]*B[8] + W[15]*B[9] + W[22]*B[10] + W[29]*B[11] + W[36]*B[12] + W[43]*B[13];\ 499 A[9] = W[2]*B[7] + W[9]*B[8] + W[16]*B[9] + W[23]*B[10] + W[30]*B[11] + W[37]*B[12] + W[44]*B[13];\ 500 A[10] = W[3]*B[7] + W[10]*B[8] + W[17]*B[9] + W[24]*B[10] + W[31]*B[11] + W[38]*B[12] + W[45]*B[13];\ 501 A[11] = W[4]*B[7] + W[11]*B[8] + W[18]*B[9] + W[25]*B[10] + W[32]*B[11] + W[39]*B[12] + W[46]*B[13];\ 502 A[12] = W[5]*B[7] + W[12]*B[8] + W[19]*B[9] + W[26]*B[10] + W[33]*B[11] + W[40]*B[12] + W[47]*B[13];\ 503 A[13] = W[6]*B[7] + W[13]*B[8] + W[20]*B[9] + W[27]*B[10] + W[34]*B[11] + W[41]*B[12] + W[48]*B[13];\ 504 A[14] = W[0]*B[14] + W[7]*B[15] + W[14]*B[16] + W[21]*B[17] + W[28]*B[18] + W[35]*B[19] + W[42]*B[20];\ 505 A[15] = W[1]*B[14] + W[8]*B[15] + W[15]*B[16] + W[22]*B[17] + W[29]*B[18] + W[36]*B[19] + W[43]*B[20];\ 506 A[16] = W[2]*B[14] + W[9]*B[15] + W[16]*B[16] + W[23]*B[17] + W[30]*B[18] + W[37]*B[19] + W[44]*B[20];\ 507 A[17] = W[3]*B[14] + W[10]*B[15] + W[17]*B[16] + W[24]*B[17] + W[31]*B[18] + W[38]*B[19] + W[45]*B[20];\ 508 A[18] = W[4]*B[14] + W[11]*B[15] + W[18]*B[16] + W[25]*B[17] + W[32]*B[18] + W[39]*B[19] + W[46]*B[20];\ 509 A[19] = W[5]*B[14] + W[12]*B[15] + W[19]*B[16] + W[26]*B[17] + W[33]*B[18] + W[40]*B[19] + W[47]*B[20];\ 510 A[20] = W[6]*B[14] + W[13]*B[15] + W[20]*B[16] + W[27]*B[17] + W[34]*B[18] + W[41]*B[19] + W[48]*B[20];\ 511 A[21] = W[0]*B[21] + W[7]*B[22] + W[14]*B[23] + W[21]*B[24] + W[28]*B[25] + W[35]*B[26] + W[42]*B[27];\ 512 A[22] = W[1]*B[21] + W[8]*B[22] + W[15]*B[23] + W[22]*B[24] + W[29]*B[25] + W[36]*B[26] + W[43]*B[27];\ 513 A[23] = W[2]*B[21] + W[9]*B[22] + W[16]*B[23] + W[23]*B[24] + W[30]*B[25] + W[37]*B[26] + W[44]*B[27];\ 514 A[24] = W[3]*B[21] + W[10]*B[22] + W[17]*B[23] + W[24]*B[24] + W[31]*B[25] + W[38]*B[26] + W[45]*B[27];\ 515 A[25] = W[4]*B[21] + W[11]*B[22] + W[18]*B[23] + W[25]*B[24] + W[32]*B[25] + W[39]*B[26] + W[46]*B[27];\ 516 A[26] = W[5]*B[21] + W[12]*B[22] + W[19]*B[23] + W[26]*B[24] + W[33]*B[25] + W[40]*B[26] + W[47]*B[27];\ 517 A[27] = W[6]*B[21] + W[13]*B[22] + W[20]*B[23] + W[27]*B[24] + W[34]*B[25] + W[41]*B[26] + W[48]*B[27];\ 518 A[28] = W[0]*B[28] + W[7]*B[29] + W[14]*B[30] + W[21]*B[31] + W[28]*B[32] + W[35]*B[33] + W[42]*B[34];\ 519 A[29] = W[1]*B[28] + W[8]*B[29] + W[15]*B[30] + W[22]*B[31] + W[29]*B[32] + W[36]*B[33] + W[43]*B[34];\ 520 A[30] = W[2]*B[28] + W[9]*B[29] + W[16]*B[30] + W[23]*B[31] + W[30]*B[32] + W[37]*B[33] + W[44]*B[34];\ 521 A[31] = W[3]*B[28] + W[10]*B[29] + W[17]*B[30] + W[24]*B[31] + W[31]*B[32] + W[38]*B[33] + W[45]*B[34];\ 522 A[32] = W[4]*B[28] + W[11]*B[29] + W[18]*B[30] + W[25]*B[31] + W[32]*B[32] + W[39]*B[33] + W[46]*B[34];\ 523 A[33] = W[5]*B[28] + W[12]*B[29] + W[19]*B[30] + W[26]*B[31] + W[33]*B[32] + W[40]*B[33] + W[47]*B[34];\ 524 A[34] = W[6]*B[28] + W[13]*B[29] + W[20]*B[30] + W[27]*B[31] + W[34]*B[32] + W[41]*B[33] + W[48]*B[34];\ 525 A[35] = W[0]*B[35] + W[7]*B[36] + W[14]*B[37] + W[21]*B[38] + W[28]*B[39] + W[35]*B[40] + W[42]*B[41];\ 526 A[36] = W[1]*B[35] + W[8]*B[36] + W[15]*B[37] + W[22]*B[38] + W[29]*B[39] + W[36]*B[40] + W[43]*B[41];\ 527 A[37] = W[2]*B[35] + W[9]*B[36] + W[16]*B[37] + W[23]*B[38] + W[30]*B[39] + W[37]*B[40] + W[44]*B[41];\ 528 A[38] = W[3]*B[35] + W[10]*B[36] + W[17]*B[37] + W[24]*B[38] + W[31]*B[39] + W[38]*B[40] + W[45]*B[41];\ 529 A[39] = W[4]*B[35] + W[11]*B[36] + W[18]*B[37] + W[25]*B[38] + W[32]*B[39] + W[39]*B[40] + W[46]*B[41];\ 530 A[40] = W[5]*B[35] + W[12]*B[36] + W[19]*B[37] + W[26]*B[38] + W[33]*B[39] + W[40]*B[40] + W[47]*B[41];\ 531 A[41] = W[6]*B[35] + W[13]*B[36] + W[20]*B[37] + W[27]*B[38] + W[34]*B[39] + W[41]*B[40] + W[48]*B[41];\ 532 A[42] = W[0]*B[42] + W[7]*B[43] + W[14]*B[44] + W[21]*B[45] + W[28]*B[46] + W[35]*B[47] + W[42]*B[48];\ 533 A[43] = W[1]*B[42] + W[8]*B[43] + W[15]*B[44] + W[22]*B[45] + W[29]*B[46] + W[36]*B[47] + W[43]*B[48];\ 534 A[44] = W[2]*B[42] + W[9]*B[43] + W[16]*B[44] + W[23]*B[45] + W[30]*B[46] + W[37]*B[47] + W[44]*B[48];\ 535 A[45] = W[3]*B[42] + W[10]*B[43] + W[17]*B[44] + W[24]*B[45] + W[31]*B[46] + W[38]*B[47] + W[45]*B[48];\ 536 A[46] = W[4]*B[42] + W[11]*B[43] + W[18]*B[44] + W[25]*B[45] + W[32]*B[46] + W[39]*B[47] + W[46]*B[48];\ 537 A[47] = W[5]*B[42] + W[12]*B[43] + W[19]*B[44] + W[26]*B[45] + W[33]*B[46] + W[40]*B[47] + W[47]*B[48];\ 538 A[48] = W[6]*B[42] + W[13]*B[43] + W[20]*B[44] + W[27]*B[45] + W[34]*B[46] + W[41]*B[47] + W[48]*B[48];\ 539 } 540 541 /* 542 Kernel_A_gets_A_minus_B_times_C_7: A = A - B * C with size bs=7 543 544 Input Parameters: 545 + A,B,C - square bs by bs arrays stored in column major order 546 547 Output Parameter: 548 . A = A - B*C 549 */ 550 551 #define Kernel_A_gets_A_minus_B_times_C_7(A,B,C) 0;\ 552 {\ 553 A[0] -= B[0]*C[0] + B[7]*C[1] + B[14]*C[2] + B[21]*C[3] + B[28]*C[4] + B[35]*C[5] + B[42]*C[6];\ 554 A[1] -= B[1]*C[0] + B[8]*C[1] + B[15]*C[2] + B[22]*C[3] + B[29]*C[4] + B[36]*C[5] + B[43]*C[6];\ 555 A[2] -= B[2]*C[0] + B[9]*C[1] + B[16]*C[2] + B[23]*C[3] + B[30]*C[4] + B[37]*C[5] + B[44]*C[6];\ 556 A[3] -= B[3]*C[0] + B[10]*C[1] + B[17]*C[2] + B[24]*C[3] + B[31]*C[4] + B[38]*C[5] + B[45]*C[6];\ 557 A[4] -= B[4]*C[0] + B[11]*C[1] + B[18]*C[2] + B[25]*C[3] + B[32]*C[4] + B[39]*C[5] + B[46]*C[6];\ 558 A[5] -= B[5]*C[0] + B[12]*C[1] + B[19]*C[2] + B[26]*C[3] + B[33]*C[4] + B[40]*C[5] + B[47]*C[6];\ 559 A[6] -= B[6]*C[0] + B[13]*C[1] + B[20]*C[2] + B[27]*C[3] + B[34]*C[4] + B[41]*C[5] + B[48]*C[6];\ 560 A[7] -= B[0]*C[7] + B[7]*C[8] + B[14]*C[9] + B[21]*C[10] + B[28]*C[11] + B[35]*C[12] + B[42]*C[13];\ 561 A[8] -= B[1]*C[7] + B[8]*C[8] + B[15]*C[9] + B[22]*C[10] + B[29]*C[11] + B[36]*C[12] + B[43]*C[13];\ 562 A[9] -= B[2]*C[7] + B[9]*C[8] + B[16]*C[9] + B[23]*C[10] + B[30]*C[11] + B[37]*C[12] + B[44]*C[13];\ 563 A[10] -= B[3]*C[7] + B[10]*C[8] + B[17]*C[9] + B[24]*C[10] + B[31]*C[11] + B[38]*C[12] + B[45]*C[13];\ 564 A[11] -= B[4]*C[7] + B[11]*C[8] + B[18]*C[9] + B[25]*C[10] + B[32]*C[11] + B[39]*C[12] + B[46]*C[13];\ 565 A[12] -= B[5]*C[7] + B[12]*C[8] + B[19]*C[9] + B[26]*C[10] + B[33]*C[11] + B[40]*C[12] + B[47]*C[13];\ 566 A[13] -= B[6]*C[7] + B[13]*C[8] + B[20]*C[9] + B[27]*C[10] + B[34]*C[11] + B[41]*C[12] + B[48]*C[13];\ 567 A[14] -= B[0]*C[14] + B[7]*C[15] + B[14]*C[16] + B[21]*C[17] + B[28]*C[18] + B[35]*C[19] + B[42]*C[20];\ 568 A[15] -= B[1]*C[14] + B[8]*C[15] + B[15]*C[16] + B[22]*C[17] + B[29]*C[18] + B[36]*C[19] + B[43]*C[20];\ 569 A[16] -= B[2]*C[14] + B[9]*C[15] + B[16]*C[16] + B[23]*C[17] + B[30]*C[18] + B[37]*C[19] + B[44]*C[20];\ 570 A[17] -= B[3]*C[14] + B[10]*C[15] + B[17]*C[16] + B[24]*C[17] + B[31]*C[18] + B[38]*C[19] + B[45]*C[20];\ 571 A[18] -= B[4]*C[14] + B[11]*C[15] + B[18]*C[16] + B[25]*C[17] + B[32]*C[18] + B[39]*C[19] + B[46]*C[20];\ 572 A[19] -= B[5]*C[14] + B[12]*C[15] + B[19]*C[16] + B[26]*C[17] + B[33]*C[18] + B[40]*C[19] + B[47]*C[20];\ 573 A[20] -= B[6]*C[14] + B[13]*C[15] + B[20]*C[16] + B[27]*C[17] + B[34]*C[18] + B[41]*C[19] + B[48]*C[20];\ 574 A[21] -= B[0]*C[21] + B[7]*C[22] + B[14]*C[23] + B[21]*C[24] + B[28]*C[25] + B[35]*C[26] + B[42]*C[27];\ 575 A[22] -= B[1]*C[21] + B[8]*C[22] + B[15]*C[23] + B[22]*C[24] + B[29]*C[25] + B[36]*C[26] + B[43]*C[27];\ 576 A[23] -= B[2]*C[21] + B[9]*C[22] + B[16]*C[23] + B[23]*C[24] + B[30]*C[25] + B[37]*C[26] + B[44]*C[27];\ 577 A[24] -= B[3]*C[21] + B[10]*C[22] + B[17]*C[23] + B[24]*C[24] + B[31]*C[25] + B[38]*C[26] + B[45]*C[27];\ 578 A[25] -= B[4]*C[21] + B[11]*C[22] + B[18]*C[23] + B[25]*C[24] + B[32]*C[25] + B[39]*C[26] + B[46]*C[27];\ 579 A[26] -= B[5]*C[21] + B[12]*C[22] + B[19]*C[23] + B[26]*C[24] + B[33]*C[25] + B[40]*C[26] + B[47]*C[27];\ 580 A[27] -= B[6]*C[21] + B[13]*C[22] + B[20]*C[23] + B[27]*C[24] + B[34]*C[25] + B[41]*C[26] + B[48]*C[27];\ 581 A[28] -= B[0]*C[28] + B[7]*C[29] + B[14]*C[30] + B[21]*C[31] + B[28]*C[32] + B[35]*C[33] + B[42]*C[34];\ 582 A[29] -= B[1]*C[28] + B[8]*C[29] + B[15]*C[30] + B[22]*C[31] + B[29]*C[32] + B[36]*C[33] + B[43]*C[34];\ 583 A[30] -= B[2]*C[28] + B[9]*C[29] + B[16]*C[30] + B[23]*C[31] + B[30]*C[32] + B[37]*C[33] + B[44]*C[34];\ 584 A[31] -= B[3]*C[28] + B[10]*C[29] + B[17]*C[30] + B[24]*C[31] + B[31]*C[32] + B[38]*C[33] + B[45]*C[34];\ 585 A[32] -= B[4]*C[28] + B[11]*C[29] + B[18]*C[30] + B[25]*C[31] + B[32]*C[32] + B[39]*C[33] + B[46]*C[34];\ 586 A[33] -= B[5]*C[28] + B[12]*C[29] + B[19]*C[30] + B[26]*C[31] + B[33]*C[32] + B[40]*C[33] + B[47]*C[34];\ 587 A[34] -= B[6]*C[28] + B[13]*C[29] + B[20]*C[30] + B[27]*C[31] + B[34]*C[32] + B[41]*C[33] + B[48]*C[34];\ 588 A[35] -= B[0]*C[35] + B[7]*C[36] + B[14]*C[37] + B[21]*C[38] + B[28]*C[39] + B[35]*C[40] + B[42]*C[41];\ 589 A[36] -= B[1]*C[35] + B[8]*C[36] + B[15]*C[37] + B[22]*C[38] + B[29]*C[39] + B[36]*C[40] + B[43]*C[41];\ 590 A[37] -= B[2]*C[35] + B[9]*C[36] + B[16]*C[37] + B[23]*C[38] + B[30]*C[39] + B[37]*C[40] + B[44]*C[41];\ 591 A[38] -= B[3]*C[35] + B[10]*C[36] + B[17]*C[37] + B[24]*C[38] + B[31]*C[39] + B[38]*C[40] + B[45]*C[41];\ 592 A[39] -= B[4]*C[35] + B[11]*C[36] + B[18]*C[37] + B[25]*C[38] + B[32]*C[39] + B[39]*C[40] + B[46]*C[41];\ 593 A[40] -= B[5]*C[35] + B[12]*C[36] + B[19]*C[37] + B[26]*C[38] + B[33]*C[39] + B[40]*C[40] + B[47]*C[41];\ 594 A[41] -= B[6]*C[35] + B[13]*C[36] + B[20]*C[37] + B[27]*C[38] + B[34]*C[39] + B[41]*C[40] + B[48]*C[41];\ 595 A[42] -= B[0]*C[42] + B[7]*C[43] + B[14]*C[44] + B[21]*C[45] + B[28]*C[46] + B[35]*C[47] + B[42]*C[48];\ 596 A[43] -= B[1]*C[42] + B[8]*C[43] + B[15]*C[44] + B[22]*C[45] + B[29]*C[46] + B[36]*C[47] + B[43]*C[48];\ 597 A[44] -= B[2]*C[42] + B[9]*C[43] + B[16]*C[44] + B[23]*C[45] + B[30]*C[46] + B[37]*C[47] + B[44]*C[48];\ 598 A[45] -= B[3]*C[42] + B[10]*C[43] + B[17]*C[44] + B[24]*C[45] + B[31]*C[46] + B[38]*C[47] + B[45]*C[48];\ 599 A[46] -= B[4]*C[42] + B[11]*C[43] + B[18]*C[44] + B[25]*C[45] + B[32]*C[46] + B[39]*C[47] + B[46]*C[48];\ 600 A[47] -= B[5]*C[42] + B[12]*C[43] + B[19]*C[44] + B[26]*C[45] + B[33]*C[46] + B[40]*C[47] + B[47]*C[48];\ 601 A[48] -= B[6]*C[42] + B[13]*C[43] + B[20]*C[44] + B[27]*C[45] + B[34]*C[46] + B[41]*C[47] + B[48]*C[48];\ 602 } 603 604 605 #endif 606