1*81278733SSatish Balay #include "sbaij.h" 2*81278733SSatish Balay #include "src/inline/ilu.h" 3*81278733SSatish Balay 4*81278733SSatish Balay /* Version for when blocks are 4 by 4 */ 5*81278733SSatish Balay #undef __FUNC__ 6*81278733SSatish Balay #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4" 7*81278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_4(Mat A,Mat *B) 8*81278733SSatish Balay { 9*81278733SSatish Balay Mat C = *B; 10*81278733SSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 11*81278733SSatish Balay IS perm = b->row; 12*81278733SSatish Balay int *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 13*81278733SSatish Balay int *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 14*81278733SSatish Balay MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 15*81278733SSatish Balay MatScalar *u,*diag,*rtmp,*rtmp_ptr; 16*81278733SSatish Balay 17*81278733SSatish Balay PetscFunctionBegin; 18*81278733SSatish Balay /* initialization */ 19*81278733SSatish Balay ierr = PetscMalloc(16*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 20*81278733SSatish Balay ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 21*81278733SSatish Balay ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr); 22*81278733SSatish Balay jl = il + mbs; 23*81278733SSatish Balay for (i=0; i<mbs; i++) { 24*81278733SSatish Balay jl[i] = mbs; il[0] = 0; 25*81278733SSatish Balay } 26*81278733SSatish Balay ierr = PetscMalloc(32*sizeof(MatScalar),&dk);CHKERRQ(ierr); 27*81278733SSatish Balay uik = dk + 16; 28*81278733SSatish Balay ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr); 29*81278733SSatish Balay 30*81278733SSatish Balay /* check permutation */ 31*81278733SSatish Balay if (!a->permute){ 32*81278733SSatish Balay ai = a->i; aj = a->j; aa = a->a; 33*81278733SSatish Balay } else { 34*81278733SSatish Balay ai = a->inew; aj = a->jnew; 35*81278733SSatish Balay ierr = PetscMalloc(16*ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr); 36*81278733SSatish Balay ierr = PetscMemcpy(aa,a->a,16*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr); 37*81278733SSatish Balay ierr = PetscMalloc(ai[mbs]*sizeof(int),&a2anew);CHKERRQ(ierr); 38*81278733SSatish Balay ierr = PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr); 39*81278733SSatish Balay 40*81278733SSatish Balay for (i=0; i<mbs; i++){ 41*81278733SSatish Balay jmin = ai[i]; jmax = ai[i+1]; 42*81278733SSatish Balay for (j=jmin; j<jmax; j++){ 43*81278733SSatish Balay while (a2anew[j] != j){ 44*81278733SSatish Balay k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k; 45*81278733SSatish Balay for (k1=0; k1<16; k1++){ 46*81278733SSatish Balay dk[k1] = aa[k*16+k1]; 47*81278733SSatish Balay aa[k*16+k1] = aa[j*16+k1]; 48*81278733SSatish Balay aa[j*16+k1] = dk[k1]; 49*81278733SSatish Balay } 50*81278733SSatish Balay } 51*81278733SSatish Balay /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */ 52*81278733SSatish Balay if (i > aj[j]){ 53*81278733SSatish Balay /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */ 54*81278733SSatish Balay ap = aa + j*16; /* ptr to the beginning of j-th block of aa */ 55*81278733SSatish Balay for (k=0; k<16; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */ 56*81278733SSatish Balay for (k=0; k<4; k++){ /* j-th block of aa <- dk^T */ 57*81278733SSatish Balay for (k1=0; k1<4; k1++) *ap++ = dk[k + 4*k1]; 58*81278733SSatish Balay } 59*81278733SSatish Balay } 60*81278733SSatish Balay } 61*81278733SSatish Balay } 62*81278733SSatish Balay ierr = PetscFree(a2anew);CHKERRQ(ierr); 63*81278733SSatish Balay } 64*81278733SSatish Balay 65*81278733SSatish Balay /* for each row k */ 66*81278733SSatish Balay for (k = 0; k<mbs; k++){ 67*81278733SSatish Balay 68*81278733SSatish Balay /*initialize k-th row with elements nonzero in row perm(k) of A */ 69*81278733SSatish Balay jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1]; 70*81278733SSatish Balay if (jmin < jmax) { 71*81278733SSatish Balay ap = aa + jmin*16; 72*81278733SSatish Balay for (j = jmin; j < jmax; j++){ 73*81278733SSatish Balay vj = perm_ptr[aj[j]]; /* block col. index */ 74*81278733SSatish Balay rtmp_ptr = rtmp + vj*16; 75*81278733SSatish Balay for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 76*81278733SSatish Balay } 77*81278733SSatish Balay } 78*81278733SSatish Balay 79*81278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 80*81278733SSatish Balay ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 81*81278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */ 82*81278733SSatish Balay 83*81278733SSatish Balay while (i < mbs){ 84*81278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 85*81278733SSatish Balay 86*81278733SSatish Balay /* compute multiplier */ 87*81278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 88*81278733SSatish Balay 89*81278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */ 90*81278733SSatish Balay diag = ba + i*16; 91*81278733SSatish Balay u = ba + ili*16; 92*81278733SSatish Balay 93*81278733SSatish Balay uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 94*81278733SSatish Balay uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 95*81278733SSatish Balay uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 96*81278733SSatish Balay uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 97*81278733SSatish Balay 98*81278733SSatish Balay uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 99*81278733SSatish Balay uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 100*81278733SSatish Balay uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 101*81278733SSatish Balay uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 102*81278733SSatish Balay 103*81278733SSatish Balay uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 104*81278733SSatish Balay uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 105*81278733SSatish Balay uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 106*81278733SSatish Balay uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 107*81278733SSatish Balay 108*81278733SSatish Balay uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 109*81278733SSatish Balay uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 110*81278733SSatish Balay uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 111*81278733SSatish Balay uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 112*81278733SSatish Balay 113*81278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 114*81278733SSatish Balay dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 115*81278733SSatish Balay dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 116*81278733SSatish Balay dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 117*81278733SSatish Balay dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 118*81278733SSatish Balay 119*81278733SSatish Balay dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 120*81278733SSatish Balay dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 121*81278733SSatish Balay dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 122*81278733SSatish Balay dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 123*81278733SSatish Balay 124*81278733SSatish Balay dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 125*81278733SSatish Balay dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 126*81278733SSatish Balay dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 127*81278733SSatish Balay dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 128*81278733SSatish Balay 129*81278733SSatish Balay dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 130*81278733SSatish Balay dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 131*81278733SSatish Balay dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 132*81278733SSatish Balay dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 133*81278733SSatish Balay 134*81278733SSatish Balay /* update -U(i,k) */ 135*81278733SSatish Balay ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 136*81278733SSatish Balay 137*81278733SSatish Balay /* add multiple of row i to k-th row ... */ 138*81278733SSatish Balay jmin = ili + 1; jmax = bi[i+1]; 139*81278733SSatish Balay if (jmin < jmax){ 140*81278733SSatish Balay for (j=jmin; j<jmax; j++) { 141*81278733SSatish Balay /* rtmp += -U(i,k)^T * U_bar(i,j) */ 142*81278733SSatish Balay rtmp_ptr = rtmp + bj[j]*16; 143*81278733SSatish Balay u = ba + j*16; 144*81278733SSatish Balay rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 145*81278733SSatish Balay rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 146*81278733SSatish Balay rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 147*81278733SSatish Balay rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 148*81278733SSatish Balay 149*81278733SSatish Balay rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 150*81278733SSatish Balay rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 151*81278733SSatish Balay rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 152*81278733SSatish Balay rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 153*81278733SSatish Balay 154*81278733SSatish Balay rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 155*81278733SSatish Balay rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 156*81278733SSatish Balay rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 157*81278733SSatish Balay rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 158*81278733SSatish Balay 159*81278733SSatish Balay rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 160*81278733SSatish Balay rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 161*81278733SSatish Balay rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 162*81278733SSatish Balay rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 163*81278733SSatish Balay } 164*81278733SSatish Balay 165*81278733SSatish Balay /* ... add i to row list for next nonzero entry */ 166*81278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 167*81278733SSatish Balay j = bj[jmin]; 168*81278733SSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 169*81278733SSatish Balay } 170*81278733SSatish Balay i = nexti; 171*81278733SSatish Balay } 172*81278733SSatish Balay 173*81278733SSatish Balay /* save nonzero entries in k-th row of U ... */ 174*81278733SSatish Balay 175*81278733SSatish Balay /* invert diagonal block */ 176*81278733SSatish Balay diag = ba+k*16; 177*81278733SSatish Balay ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 178*81278733SSatish Balay ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 179*81278733SSatish Balay 180*81278733SSatish Balay jmin = bi[k]; jmax = bi[k+1]; 181*81278733SSatish Balay if (jmin < jmax) { 182*81278733SSatish Balay for (j=jmin; j<jmax; j++){ 183*81278733SSatish Balay vj = bj[j]; /* block col. index of U */ 184*81278733SSatish Balay u = ba + j*16; 185*81278733SSatish Balay rtmp_ptr = rtmp + vj*16; 186*81278733SSatish Balay for (k1=0; k1<16; k1++){ 187*81278733SSatish Balay *u++ = *rtmp_ptr; 188*81278733SSatish Balay *rtmp_ptr++ = 0.0; 189*81278733SSatish Balay } 190*81278733SSatish Balay } 191*81278733SSatish Balay 192*81278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */ 193*81278733SSatish Balay il[k] = jmin; 194*81278733SSatish Balay i = bj[jmin]; 195*81278733SSatish Balay jl[k] = jl[i]; jl[i] = k; 196*81278733SSatish Balay } 197*81278733SSatish Balay } 198*81278733SSatish Balay 199*81278733SSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 200*81278733SSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 201*81278733SSatish Balay ierr = PetscFree(dk);CHKERRQ(ierr); 202*81278733SSatish Balay if (a->permute){ 203*81278733SSatish Balay ierr = PetscFree(aa);CHKERRQ(ierr); 204*81278733SSatish Balay } 205*81278733SSatish Balay 206*81278733SSatish Balay ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr); 207*81278733SSatish Balay C->factor = FACTOR_CHOLESKY; 208*81278733SSatish Balay C->assembled = PETSC_TRUE; 209*81278733SSatish Balay C->preallocated = PETSC_TRUE; 210*81278733SSatish Balay PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 211*81278733SSatish Balay PetscFunctionReturn(0); 212*81278733SSatish Balay } 213