Lines Matching refs:bs2

315   const PetscInt *ajtmp, *bjtmp, *bdiag = b->diag, *pj, bs2 = a->bs2;  in MatLUFactorNumeric_SeqBAIJ_5()  local
327 PetscCall(PetscMalloc2(bs2 * n, &rtmp, bs2, &mwork)); in MatLUFactorNumeric_SeqBAIJ_5()
328 PetscCall(PetscArrayzero(rtmp, bs2 * n)); in MatLUFactorNumeric_SeqBAIJ_5()
335 for (j = 0; j < nz; j++) PetscCall(PetscArrayzero(rtmp + bs2 * bjtmp[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5()
340 for (j = 0; j < nz; j++) PetscCall(PetscArrayzero(rtmp + bs2 * bjtmp[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5()
345 v = aa + bs2 * ai[r[i]]; in MatLUFactorNumeric_SeqBAIJ_5()
346 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(rtmp + bs2 * ic[ajtmp[j]], v + bs2 * j, bs2)); in MatLUFactorNumeric_SeqBAIJ_5()
353 pc = rtmp + bs2 * row; in MatLUFactorNumeric_SeqBAIJ_5()
354 for (flg = 0, j = 0; j < bs2; j++) { in MatLUFactorNumeric_SeqBAIJ_5()
361 pv = b->a + bs2 * bdiag[row]; in MatLUFactorNumeric_SeqBAIJ_5()
366 pv = b->a + bs2 * (bdiag[row + 1] + 1); in MatLUFactorNumeric_SeqBAIJ_5()
371 v = rtmp + bs2 * pj[j]; in MatLUFactorNumeric_SeqBAIJ_5()
373 pv += bs2; in MatLUFactorNumeric_SeqBAIJ_5()
381 pv = b->a + bs2 * bi[i]; in MatLUFactorNumeric_SeqBAIJ_5()
384 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(pv + bs2 * j, rtmp + bs2 * pj[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5()
387 pv = b->a + bs2 * bdiag[i]; in MatLUFactorNumeric_SeqBAIJ_5()
389 PetscCall(PetscArraycpy(pv, rtmp + bs2 * pj[0], bs2)); in MatLUFactorNumeric_SeqBAIJ_5()
394 pv = b->a + bs2 * (bdiag[i + 1] + 1); in MatLUFactorNumeric_SeqBAIJ_5()
397 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(pv + bs2 * j, rtmp + bs2 * pj[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5()
682 const PetscInt *ajtmp, *bjtmp, *bdiag = b->diag, *pj, bs2 = a->bs2; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering() local
692 PetscCall(PetscMalloc2(bs2 * n, &rtmp, bs2, &mwork)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
693 PetscCall(PetscArrayzero(rtmp, bs2 * n)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
700 for (j = 0; j < nz; j++) PetscCall(PetscArrayzero(rtmp + bs2 * bjtmp[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
705 for (j = 0; j < nz; j++) PetscCall(PetscArrayzero(rtmp + bs2 * bjtmp[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
710 v = aa + bs2 * ai[i]; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
711 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(rtmp + bs2 * ajtmp[j], v + bs2 * j, bs2)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
718 pc = rtmp + bs2 * row; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
719 for (flg = 0, j = 0; j < bs2; j++) { in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
726 pv = b->a + bs2 * bdiag[row]; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
731 pv = b->a + bs2 * (bdiag[row + 1] + 1); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
736 vv = rtmp + bs2 * pj[j]; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
738 pv += bs2; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
746 pv = b->a + bs2 * bi[i]; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
749 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(pv + bs2 * j, rtmp + bs2 * pj[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
752 pv = b->a + bs2 * bdiag[i]; in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
754 PetscCall(PetscArraycpy(pv, rtmp + bs2 * pj[0], bs2)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
759 pv = b->a + bs2 * (bdiag[i + 1] + 1); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
762 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(pv + bs2 * j, rtmp + bs2 * pj[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering()
785 const PetscInt *ajtmp, *bjtmp, *bdiag = b->diag, *pj, bs2 = a->bs2; in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering() local
795 PetscCall(PetscMalloc2(bs2 * n, &rtmp, bs2, &mwork)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
796 PetscCall(PetscArrayzero(rtmp, bs2 * n)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
803 for (j = 0; j < nz; j++) PetscCall(PetscArrayzero(rtmp + bs2 * bjtmp[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
808 for (j = 0; j < nz; j++) PetscCall(PetscArrayzero(rtmp + bs2 * bjtmp[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
813 v = aa + bs2 * ai[i]; in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
814 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(rtmp + bs2 * ajtmp[j], v + bs2 * j, bs2)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
821 pc = rtmp + bs2 * row; in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
822 for (flg = 0, j = 0; j < bs2; j++) { in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
829 pv = b->a + bs2 * bdiag[row]; in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
834 pv = b->a + bs2 * (bdiag[row + 1] + 1); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
839 v = rtmp + bs2 * pj[j]; in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
849 pv = b->a + bs2 * bi[i]; in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
852 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(pv + bs2 * j, rtmp + bs2 * pj[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
855 pv = b->a + bs2 * bdiag[i]; in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
857 PetscCall(PetscArraycpy(pv, rtmp + bs2 * pj[0], bs2)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
862 pv = b->a + bs2 * (bdiag[i + 1] + 1); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
865 for (j = 0; j < nz; j++) PetscCall(PetscArraycpy(pv + bs2 * j, rtmp + bs2 * pj[j], bs2)); in MatLUFactorNumeric_SeqBAIJ_9_NaturalOrdering()
882 PetscInt nz, bs = A->rmap->bs, bs2 = a->bs2; in MatSolve_SeqBAIJ_9_NaturalOrdering() local
897 v = aa + bs2 * ai[i]; in MatSolve_SeqBAIJ_9_NaturalOrdering()
980 v += bs2; in MatSolve_SeqBAIJ_9_NaturalOrdering()
990 v = aa + bs2 * (adiag[i + 1] + 1); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1079 v += bs2; in MatSolve_SeqBAIJ_9_NaturalOrdering()
1092 a0 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[0]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1094 a1 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[4]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1096 a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[8]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1101 a3 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[9]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1103 a4 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[13]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1105 a5 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[17]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1110 a0 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[18]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1112 a1 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[22]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1114 a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[26]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1119 a3 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[27]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1121 a4 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[31]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1123 a5 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[35]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1128 a0 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[36]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1130 a1 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[40]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1132 a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[44]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1137 a3 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[45]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1139 a4 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[49]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1141 a5 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[53]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1146 a0 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[54]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1148 a1 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[58]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1150 a2 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[62]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1155 a3 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[63]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1157 a4 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[67]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1159 a5 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[71]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1164 a3 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[72]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1166 a4 = _mm256_loadu_pd(&(aa + bs2 * adiag[i])[76]); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1168 … a2 = _mm256_maskload_pd(&(aa + bs2 * adiag[i])[80], _mm256_set_epi64x(0LL, 0LL, 0LL, 1LL << 63)); in MatSolve_SeqBAIJ_9_NaturalOrdering()
1180 PetscCall(PetscLogFlops(2.0 * (a->bs2) * (a->nz) - A->rmap->bs * A->cmap->n)); in MatSolve_SeqBAIJ_9_NaturalOrdering()