1*81278733SSatish Balay #include "sbaij.h" 2*81278733SSatish Balay #include "src/inline/ilu.h" 3*81278733SSatish Balay 4*81278733SSatish Balay /* 5*81278733SSatish Balay Version for when blocks are 4 by 4 Using natural ordering 6*81278733SSatish Balay */ 7*81278733SSatish Balay #undef __FUNC__ 8*81278733SSatish Balay #define __FUNC__ "MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering" 9*81278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_4_NaturalOrdering(Mat A,Mat *B) 10*81278733SSatish Balay { 11*81278733SSatish Balay Mat C = *B; 12*81278733SSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data; 13*81278733SSatish Balay int ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j; 14*81278733SSatish Balay int *ai,*aj,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili; 15*81278733SSatish Balay MatScalar *ba = b->a,*aa,*ap,*dk,*uik; 16*81278733SSatish Balay MatScalar *u,*diag,*rtmp,*rtmp_ptr; 17*81278733SSatish Balay 18*81278733SSatish Balay PetscFunctionBegin; 19*81278733SSatish Balay 20*81278733SSatish Balay /* initialization */ 21*81278733SSatish Balay ierr = PetscMalloc(16*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 22*81278733SSatish Balay ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr); 23*81278733SSatish Balay ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr); 24*81278733SSatish Balay jl = il + mbs; 25*81278733SSatish Balay for (i=0; i<mbs; i++) { 26*81278733SSatish Balay jl[i] = mbs; il[0] = 0; 27*81278733SSatish Balay } 28*81278733SSatish Balay ierr = PetscMalloc(32*sizeof(MatScalar),&dk);CHKERRQ(ierr); 29*81278733SSatish Balay uik = dk + 16; 30*81278733SSatish Balay 31*81278733SSatish Balay ai = a->i; aj = a->j; aa = a->a; 32*81278733SSatish Balay 33*81278733SSatish Balay /* for each row k */ 34*81278733SSatish Balay for (k = 0; k<mbs; k++){ 35*81278733SSatish Balay 36*81278733SSatish Balay /*initialize k-th row with elements nonzero in row k of A */ 37*81278733SSatish Balay jmin = ai[k]; jmax = ai[k+1]; 38*81278733SSatish Balay if (jmin < jmax) { 39*81278733SSatish Balay ap = aa + jmin*16; 40*81278733SSatish Balay for (j = jmin; j < jmax; j++){ 41*81278733SSatish Balay vj = aj[j]; /* block col. index */ 42*81278733SSatish Balay rtmp_ptr = rtmp + vj*16; 43*81278733SSatish Balay for (i=0; i<16; i++) *rtmp_ptr++ = *ap++; 44*81278733SSatish Balay } 45*81278733SSatish Balay } 46*81278733SSatish Balay 47*81278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 48*81278733SSatish Balay ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr); 49*81278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */ 50*81278733SSatish Balay 51*81278733SSatish Balay while (i < mbs){ 52*81278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 53*81278733SSatish Balay 54*81278733SSatish Balay /* compute multiplier */ 55*81278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 56*81278733SSatish Balay 57*81278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */ 58*81278733SSatish Balay diag = ba + i*16; 59*81278733SSatish Balay u = ba + ili*16; 60*81278733SSatish Balay 61*81278733SSatish Balay uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]); 62*81278733SSatish Balay uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]); 63*81278733SSatish Balay uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]); 64*81278733SSatish Balay uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]); 65*81278733SSatish Balay 66*81278733SSatish Balay uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]); 67*81278733SSatish Balay uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]); 68*81278733SSatish Balay uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]); 69*81278733SSatish Balay uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]); 70*81278733SSatish Balay 71*81278733SSatish Balay uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]); 72*81278733SSatish Balay uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]); 73*81278733SSatish Balay uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]); 74*81278733SSatish Balay uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]); 75*81278733SSatish Balay 76*81278733SSatish Balay uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]); 77*81278733SSatish Balay uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]); 78*81278733SSatish Balay uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]); 79*81278733SSatish Balay uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]); 80*81278733SSatish Balay 81*81278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 82*81278733SSatish Balay dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 83*81278733SSatish Balay dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 84*81278733SSatish Balay dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 85*81278733SSatish Balay dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 86*81278733SSatish Balay 87*81278733SSatish Balay dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 88*81278733SSatish Balay dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 89*81278733SSatish Balay dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 90*81278733SSatish Balay dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 91*81278733SSatish Balay 92*81278733SSatish Balay dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 93*81278733SSatish Balay dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 94*81278733SSatish Balay dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 95*81278733SSatish Balay dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 96*81278733SSatish Balay 97*81278733SSatish Balay dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 98*81278733SSatish Balay dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 99*81278733SSatish Balay dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 100*81278733SSatish Balay dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 101*81278733SSatish Balay 102*81278733SSatish Balay /* update -U(i,k) */ 103*81278733SSatish Balay ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr); 104*81278733SSatish Balay 105*81278733SSatish Balay /* add multiple of row i to k-th row ... */ 106*81278733SSatish Balay jmin = ili + 1; jmax = bi[i+1]; 107*81278733SSatish Balay if (jmin < jmax){ 108*81278733SSatish Balay for (j=jmin; j<jmax; j++) { 109*81278733SSatish Balay /* rtmp += -U(i,k)^T * U_bar(i,j) */ 110*81278733SSatish Balay rtmp_ptr = rtmp + bj[j]*16; 111*81278733SSatish Balay u = ba + j*16; 112*81278733SSatish Balay rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3]; 113*81278733SSatish Balay rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3]; 114*81278733SSatish Balay rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3]; 115*81278733SSatish Balay rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3]; 116*81278733SSatish Balay 117*81278733SSatish Balay rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7]; 118*81278733SSatish Balay rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7]; 119*81278733SSatish Balay rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7]; 120*81278733SSatish Balay rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7]; 121*81278733SSatish Balay 122*81278733SSatish Balay rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11]; 123*81278733SSatish Balay rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11]; 124*81278733SSatish Balay rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11]; 125*81278733SSatish Balay rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11]; 126*81278733SSatish Balay 127*81278733SSatish Balay rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15]; 128*81278733SSatish Balay rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15]; 129*81278733SSatish Balay rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15]; 130*81278733SSatish Balay rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15]; 131*81278733SSatish Balay } 132*81278733SSatish Balay 133*81278733SSatish Balay /* ... add i to row list for next nonzero entry */ 134*81278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 135*81278733SSatish Balay j = bj[jmin]; 136*81278733SSatish Balay jl[i] = jl[j]; jl[j] = i; /* update jl */ 137*81278733SSatish Balay } 138*81278733SSatish Balay i = nexti; 139*81278733SSatish Balay } 140*81278733SSatish Balay 141*81278733SSatish Balay /* save nonzero entries in k-th row of U ... */ 142*81278733SSatish Balay 143*81278733SSatish Balay /* invert diagonal block */ 144*81278733SSatish Balay diag = ba+k*16; 145*81278733SSatish Balay ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr); 146*81278733SSatish Balay ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr); 147*81278733SSatish Balay 148*81278733SSatish Balay jmin = bi[k]; jmax = bi[k+1]; 149*81278733SSatish Balay if (jmin < jmax) { 150*81278733SSatish Balay for (j=jmin; j<jmax; j++){ 151*81278733SSatish Balay vj = bj[j]; /* block col. index of U */ 152*81278733SSatish Balay u = ba + j*16; 153*81278733SSatish Balay rtmp_ptr = rtmp + vj*16; 154*81278733SSatish Balay for (k1=0; k1<16; k1++){ 155*81278733SSatish Balay *u++ = *rtmp_ptr; 156*81278733SSatish Balay *rtmp_ptr++ = 0.0; 157*81278733SSatish Balay } 158*81278733SSatish Balay } 159*81278733SSatish Balay 160*81278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */ 161*81278733SSatish Balay il[k] = jmin; 162*81278733SSatish Balay i = bj[jmin]; 163*81278733SSatish Balay jl[k] = jl[i]; jl[i] = k; 164*81278733SSatish Balay } 165*81278733SSatish Balay } 166*81278733SSatish Balay 167*81278733SSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 168*81278733SSatish Balay ierr = PetscFree(il);CHKERRQ(ierr); 169*81278733SSatish Balay ierr = PetscFree(dk);CHKERRQ(ierr); 170*81278733SSatish Balay 171*81278733SSatish Balay C->factor = FACTOR_CHOLESKY; 172*81278733SSatish Balay C->assembled = PETSC_TRUE; 173*81278733SSatish Balay C->preallocated = PETSC_TRUE; 174*81278733SSatish Balay PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */ 175*81278733SSatish Balay PetscFunctionReturn(0); 176*81278733SSatish Balay } 177