11c2a3de1SBarry Smith 2c6db04a5SJed Brown #include <../src/mat/impls/sbaij/seq/sbaij.h> 3af0996ceSBarry Smith #include <petsc/private/kernels/blockinvert.h> 481278733SSatish Balay 581278733SSatish Balay /* 681278733SSatish Balay Version for when blocks are 5 by 5 Using natural ordering 781278733SSatish Balay */ 8*d71ae5a4SJacob Faibussowitsch PetscErrorCode MatCholeskyFactorNumeric_SeqSBAIJ_5_NaturalOrdering(Mat C, Mat A, const MatFactorInfo *info) 9*d71ae5a4SJacob Faibussowitsch { 1081278733SSatish Balay Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ *)A->data, *b = (Mat_SeqSBAIJ *)C->data; 1113f74950SBarry Smith PetscInt i, j, mbs = a->mbs, *bi = b->i, *bj = b->j; 12c0c53c30SBarry Smith PetscInt *ai, *aj, k, k1, jmin, jmax, *jl, *il, vj, nexti, ili, ipvt[5]; 1381278733SSatish Balay MatScalar *ba = b->a, *aa, *ap, *dk, *uik; 14c0c53c30SBarry Smith MatScalar *u, *d, *rtmp, *rtmp_ptr, work[25]; 15182b8fbaSHong Zhang PetscReal shift = info->shiftamount; 16a455e926SHong Zhang PetscBool allowzeropivot, zeropivotdetected; 1781278733SSatish Balay 1881278733SSatish Balay PetscFunctionBegin; 1981278733SSatish Balay /* initialization */ 200164db54SHong Zhang allowzeropivot = PetscNot(A->erroriffailure); 219566063dSJacob Faibussowitsch PetscCall(PetscCalloc1(25 * mbs, &rtmp)); 229566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(mbs, &il, mbs, &jl)); 236df5ee2eSHong Zhang il[0] = 0; 246df5ee2eSHong Zhang for (i = 0; i < mbs; i++) jl[i] = mbs; 256df5ee2eSHong Zhang 269566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(25, &dk, 25, &uik)); 279371c9d4SSatish Balay ai = a->i; 289371c9d4SSatish Balay aj = a->j; 299371c9d4SSatish Balay aa = a->a; 3081278733SSatish Balay 3181278733SSatish Balay /* for each row k */ 3281278733SSatish Balay for (k = 0; k < mbs; k++) { 3381278733SSatish Balay /*initialize k-th row with elements nonzero in row k of A */ 349371c9d4SSatish Balay jmin = ai[k]; 359371c9d4SSatish Balay jmax = ai[k + 1]; 3681278733SSatish Balay if (jmin < jmax) { 3781278733SSatish Balay ap = aa + jmin * 25; 3881278733SSatish Balay for (j = jmin; j < jmax; j++) { 3981278733SSatish Balay vj = aj[j]; /* block col. index */ 4081278733SSatish Balay rtmp_ptr = rtmp + vj * 25; 4181278733SSatish Balay for (i = 0; i < 25; i++) *rtmp_ptr++ = *ap++; 4281278733SSatish Balay } 4381278733SSatish Balay } 4481278733SSatish Balay 4581278733SSatish Balay /* modify k-th row by adding in those rows i with U(i,k) != 0 */ 469566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(dk, rtmp + k * 25, 25)); 4781278733SSatish Balay i = jl[k]; /* first row to be added to k_th row */ 4881278733SSatish Balay 4981278733SSatish Balay while (i < mbs) { 5081278733SSatish Balay nexti = jl[i]; /* next row to be added to k_th row */ 5181278733SSatish Balay 5281278733SSatish Balay /* compute multiplier */ 5381278733SSatish Balay ili = il[i]; /* index of first nonzero element in U(i,k:bms-1) */ 5481278733SSatish Balay 5581278733SSatish Balay /* uik = -inv(Di)*U_bar(i,k) */ 5681278733SSatish Balay d = ba + i * 25; 5781278733SSatish Balay u = ba + ili * 25; 5881278733SSatish Balay 5981278733SSatish Balay uik[0] = -(d[0] * u[0] + d[5] * u[1] + d[10] * u[2] + d[15] * u[3] + d[20] * u[4]); 6081278733SSatish Balay uik[1] = -(d[1] * u[0] + d[6] * u[1] + d[11] * u[2] + d[16] * u[3] + d[21] * u[4]); 6181278733SSatish Balay uik[2] = -(d[2] * u[0] + d[7] * u[1] + d[12] * u[2] + d[17] * u[3] + d[22] * u[4]); 6281278733SSatish Balay uik[3] = -(d[3] * u[0] + d[8] * u[1] + d[13] * u[2] + d[18] * u[3] + d[23] * u[4]); 6381278733SSatish Balay uik[4] = -(d[4] * u[0] + d[9] * u[1] + d[14] * u[2] + d[19] * u[3] + d[24] * u[4]); 6481278733SSatish Balay 6581278733SSatish Balay uik[5] = -(d[0] * u[5] + d[5] * u[6] + d[10] * u[7] + d[15] * u[8] + d[20] * u[9]); 6681278733SSatish Balay uik[6] = -(d[1] * u[5] + d[6] * u[6] + d[11] * u[7] + d[16] * u[8] + d[21] * u[9]); 6781278733SSatish Balay uik[7] = -(d[2] * u[5] + d[7] * u[6] + d[12] * u[7] + d[17] * u[8] + d[22] * u[9]); 6881278733SSatish Balay uik[8] = -(d[3] * u[5] + d[8] * u[6] + d[13] * u[7] + d[18] * u[8] + d[23] * u[9]); 6981278733SSatish Balay uik[9] = -(d[4] * u[5] + d[9] * u[6] + d[14] * u[7] + d[19] * u[8] + d[24] * u[9]); 7081278733SSatish Balay 7181278733SSatish Balay uik[10] = -(d[0] * u[10] + d[5] * u[11] + d[10] * u[12] + d[15] * u[13] + d[20] * u[14]); 7281278733SSatish Balay uik[11] = -(d[1] * u[10] + d[6] * u[11] + d[11] * u[12] + d[16] * u[13] + d[21] * u[14]); 7381278733SSatish Balay uik[12] = -(d[2] * u[10] + d[7] * u[11] + d[12] * u[12] + d[17] * u[13] + d[22] * u[14]); 7481278733SSatish Balay uik[13] = -(d[3] * u[10] + d[8] * u[11] + d[13] * u[12] + d[18] * u[13] + d[23] * u[14]); 7581278733SSatish Balay uik[14] = -(d[4] * u[10] + d[9] * u[11] + d[14] * u[12] + d[19] * u[13] + d[24] * u[14]); 7681278733SSatish Balay 7781278733SSatish Balay uik[15] = -(d[0] * u[15] + d[5] * u[16] + d[10] * u[17] + d[15] * u[18] + d[20] * u[19]); 7881278733SSatish Balay uik[16] = -(d[1] * u[15] + d[6] * u[16] + d[11] * u[17] + d[16] * u[18] + d[21] * u[19]); 7981278733SSatish Balay uik[17] = -(d[2] * u[15] + d[7] * u[16] + d[12] * u[17] + d[17] * u[18] + d[22] * u[19]); 8081278733SSatish Balay uik[18] = -(d[3] * u[15] + d[8] * u[16] + d[13] * u[17] + d[18] * u[18] + d[23] * u[19]); 8181278733SSatish Balay uik[19] = -(d[4] * u[15] + d[9] * u[16] + d[14] * u[17] + d[19] * u[18] + d[24] * u[19]); 8281278733SSatish Balay 8381278733SSatish Balay uik[20] = -(d[0] * u[20] + d[5] * u[21] + d[10] * u[22] + d[15] * u[23] + d[20] * u[24]); 8481278733SSatish Balay uik[21] = -(d[1] * u[20] + d[6] * u[21] + d[11] * u[22] + d[16] * u[23] + d[21] * u[24]); 8581278733SSatish Balay uik[22] = -(d[2] * u[20] + d[7] * u[21] + d[12] * u[22] + d[17] * u[23] + d[22] * u[24]); 8681278733SSatish Balay uik[23] = -(d[3] * u[20] + d[8] * u[21] + d[13] * u[22] + d[18] * u[23] + d[23] * u[24]); 8781278733SSatish Balay uik[24] = -(d[4] * u[20] + d[9] * u[21] + d[14] * u[22] + d[19] * u[23] + d[24] * u[24]); 8881278733SSatish Balay 8981278733SSatish Balay /* update D(k) += -U(i,k)^T * U_bar(i,k) */ 9081278733SSatish Balay dk[0] += uik[0] * u[0] + uik[1] * u[1] + uik[2] * u[2] + uik[3] * u[3] + uik[4] * u[4]; 9181278733SSatish Balay dk[1] += uik[5] * u[0] + uik[6] * u[1] + uik[7] * u[2] + uik[8] * u[3] + uik[9] * u[4]; 9281278733SSatish Balay dk[2] += uik[10] * u[0] + uik[11] * u[1] + uik[12] * u[2] + uik[13] * u[3] + uik[14] * u[4]; 9381278733SSatish Balay dk[3] += uik[15] * u[0] + uik[16] * u[1] + uik[17] * u[2] + uik[18] * u[3] + uik[19] * u[4]; 9481278733SSatish Balay dk[4] += uik[20] * u[0] + uik[21] * u[1] + uik[22] * u[2] + uik[23] * u[3] + uik[24] * u[4]; 9581278733SSatish Balay 9681278733SSatish Balay dk[5] += uik[0] * u[5] + uik[1] * u[6] + uik[2] * u[7] + uik[3] * u[8] + uik[4] * u[9]; 9781278733SSatish Balay dk[6] += uik[5] * u[5] + uik[6] * u[6] + uik[7] * u[7] + uik[8] * u[8] + uik[9] * u[9]; 9881278733SSatish Balay dk[7] += uik[10] * u[5] + uik[11] * u[6] + uik[12] * u[7] + uik[13] * u[8] + uik[14] * u[9]; 9981278733SSatish Balay dk[8] += uik[15] * u[5] + uik[16] * u[6] + uik[17] * u[7] + uik[18] * u[8] + uik[19] * u[9]; 10081278733SSatish Balay dk[9] += uik[20] * u[5] + uik[21] * u[6] + uik[22] * u[7] + uik[23] * u[8] + uik[24] * u[9]; 10181278733SSatish Balay 10281278733SSatish Balay dk[10] += uik[0] * u[10] + uik[1] * u[11] + uik[2] * u[12] + uik[3] * u[13] + uik[4] * u[14]; 10381278733SSatish Balay dk[11] += uik[5] * u[10] + uik[6] * u[11] + uik[7] * u[12] + uik[8] * u[13] + uik[9] * u[14]; 10481278733SSatish Balay dk[12] += uik[10] * u[10] + uik[11] * u[11] + uik[12] * u[12] + uik[13] * u[13] + uik[14] * u[14]; 10581278733SSatish Balay dk[13] += uik[15] * u[10] + uik[16] * u[11] + uik[17] * u[12] + uik[18] * u[13] + uik[19] * u[14]; 10681278733SSatish Balay dk[14] += uik[20] * u[10] + uik[21] * u[11] + uik[22] * u[12] + uik[23] * u[13] + uik[24] * u[14]; 10781278733SSatish Balay 10881278733SSatish Balay dk[15] += uik[0] * u[15] + uik[1] * u[16] + uik[2] * u[17] + uik[3] * u[18] + uik[4] * u[19]; 10981278733SSatish Balay dk[16] += uik[5] * u[15] + uik[6] * u[16] + uik[7] * u[17] + uik[8] * u[18] + uik[9] * u[19]; 11081278733SSatish Balay dk[17] += uik[10] * u[15] + uik[11] * u[16] + uik[12] * u[17] + uik[13] * u[18] + uik[14] * u[19]; 11181278733SSatish Balay dk[18] += uik[15] * u[15] + uik[16] * u[16] + uik[17] * u[17] + uik[18] * u[18] + uik[19] * u[19]; 11281278733SSatish Balay dk[19] += uik[20] * u[15] + uik[21] * u[16] + uik[22] * u[17] + uik[23] * u[18] + uik[24] * u[19]; 11381278733SSatish Balay 11481278733SSatish Balay dk[20] += uik[0] * u[20] + uik[1] * u[21] + uik[2] * u[22] + uik[3] * u[23] + uik[4] * u[24]; 11581278733SSatish Balay dk[21] += uik[5] * u[20] + uik[6] * u[21] + uik[7] * u[22] + uik[8] * u[23] + uik[9] * u[24]; 11681278733SSatish Balay dk[22] += uik[10] * u[20] + uik[11] * u[21] + uik[12] * u[22] + uik[13] * u[23] + uik[14] * u[24]; 11781278733SSatish Balay dk[23] += uik[15] * u[20] + uik[16] * u[21] + uik[17] * u[22] + uik[18] * u[23] + uik[19] * u[24]; 11881278733SSatish Balay dk[24] += uik[20] * u[20] + uik[21] * u[21] + uik[22] * u[22] + uik[23] * u[23] + uik[24] * u[24]; 11981278733SSatish Balay 1209566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(125.0 * 4.0)); 121187a9f4bSHong Zhang 12281278733SSatish Balay /* update -U(i,k) */ 1239566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(ba + ili * 25, uik, 25)); 12481278733SSatish Balay 12581278733SSatish Balay /* add multiple of row i to k-th row ... */ 1269371c9d4SSatish Balay jmin = ili + 1; 1279371c9d4SSatish Balay jmax = bi[i + 1]; 12881278733SSatish Balay if (jmin < jmax) { 12981278733SSatish Balay for (j = jmin; j < jmax; j++) { 13081278733SSatish Balay /* rtmp += -U(i,k)^T * U_bar(i,j) */ 13181278733SSatish Balay rtmp_ptr = rtmp + bj[j] * 25; 13281278733SSatish Balay u = ba + j * 25; 13381278733SSatish Balay rtmp_ptr[0] += uik[0] * u[0] + uik[1] * u[1] + uik[2] * u[2] + uik[3] * u[3] + uik[4] * u[4]; 13481278733SSatish Balay rtmp_ptr[1] += uik[5] * u[0] + uik[6] * u[1] + uik[7] * u[2] + uik[8] * u[3] + uik[9] * u[4]; 13581278733SSatish Balay rtmp_ptr[2] += uik[10] * u[0] + uik[11] * u[1] + uik[12] * u[2] + uik[13] * u[3] + uik[14] * u[4]; 13681278733SSatish Balay rtmp_ptr[3] += uik[15] * u[0] + uik[16] * u[1] + uik[17] * u[2] + uik[18] * u[3] + uik[19] * u[4]; 13781278733SSatish Balay rtmp_ptr[4] += uik[20] * u[0] + uik[21] * u[1] + uik[22] * u[2] + uik[23] * u[3] + uik[24] * u[4]; 13881278733SSatish Balay 13981278733SSatish Balay rtmp_ptr[5] += uik[0] * u[5] + uik[1] * u[6] + uik[2] * u[7] + uik[3] * u[8] + uik[4] * u[9]; 14081278733SSatish Balay rtmp_ptr[6] += uik[5] * u[5] + uik[6] * u[6] + uik[7] * u[7] + uik[8] * u[8] + uik[9] * u[9]; 14181278733SSatish Balay rtmp_ptr[7] += uik[10] * u[5] + uik[11] * u[6] + uik[12] * u[7] + uik[13] * u[8] + uik[14] * u[9]; 14281278733SSatish Balay rtmp_ptr[8] += uik[15] * u[5] + uik[16] * u[6] + uik[17] * u[7] + uik[18] * u[8] + uik[19] * u[9]; 14381278733SSatish Balay rtmp_ptr[9] += uik[20] * u[5] + uik[21] * u[6] + uik[22] * u[7] + uik[23] * u[8] + uik[24] * u[9]; 14481278733SSatish Balay 14581278733SSatish Balay rtmp_ptr[10] += uik[0] * u[10] + uik[1] * u[11] + uik[2] * u[12] + uik[3] * u[13] + uik[4] * u[14]; 14681278733SSatish Balay rtmp_ptr[11] += uik[5] * u[10] + uik[6] * u[11] + uik[7] * u[12] + uik[8] * u[13] + uik[9] * u[14]; 14781278733SSatish Balay rtmp_ptr[12] += uik[10] * u[10] + uik[11] * u[11] + uik[12] * u[12] + uik[13] * u[13] + uik[14] * u[14]; 14881278733SSatish Balay rtmp_ptr[13] += uik[15] * u[10] + uik[16] * u[11] + uik[17] * u[12] + uik[18] * u[13] + uik[19] * u[14]; 14981278733SSatish Balay rtmp_ptr[14] += uik[20] * u[10] + uik[21] * u[11] + uik[22] * u[12] + uik[23] * u[13] + uik[24] * u[14]; 15081278733SSatish Balay 15181278733SSatish Balay rtmp_ptr[15] += uik[0] * u[15] + uik[1] * u[16] + uik[2] * u[17] + uik[3] * u[18] + uik[4] * u[19]; 15281278733SSatish Balay rtmp_ptr[16] += uik[5] * u[15] + uik[6] * u[16] + uik[7] * u[17] + uik[8] * u[18] + uik[9] * u[19]; 15381278733SSatish Balay rtmp_ptr[17] += uik[10] * u[15] + uik[11] * u[16] + uik[12] * u[17] + uik[13] * u[18] + uik[14] * u[19]; 15481278733SSatish Balay rtmp_ptr[18] += uik[15] * u[15] + uik[16] * u[16] + uik[17] * u[17] + uik[18] * u[18] + uik[19] * u[19]; 15581278733SSatish Balay rtmp_ptr[19] += uik[20] * u[15] + uik[21] * u[16] + uik[22] * u[17] + uik[23] * u[18] + uik[24] * u[19]; 15681278733SSatish Balay 15781278733SSatish Balay rtmp_ptr[20] += uik[0] * u[20] + uik[1] * u[21] + uik[2] * u[22] + uik[3] * u[23] + uik[4] * u[24]; 15881278733SSatish Balay rtmp_ptr[21] += uik[5] * u[20] + uik[6] * u[21] + uik[7] * u[22] + uik[8] * u[23] + uik[9] * u[24]; 15981278733SSatish Balay rtmp_ptr[22] += uik[10] * u[20] + uik[11] * u[21] + uik[12] * u[22] + uik[13] * u[23] + uik[14] * u[24]; 16081278733SSatish Balay rtmp_ptr[23] += uik[15] * u[20] + uik[16] * u[21] + uik[17] * u[22] + uik[18] * u[23] + uik[19] * u[24]; 16181278733SSatish Balay rtmp_ptr[24] += uik[20] * u[20] + uik[21] * u[21] + uik[22] * u[22] + uik[23] * u[23] + uik[24] * u[24]; 16281278733SSatish Balay } 1639566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(2.0 * 125.0 * (jmax - jmin))); 16481278733SSatish Balay 16581278733SSatish Balay /* ... add i to row list for next nonzero entry */ 16681278733SSatish Balay il[i] = jmin; /* update il(i) in column k+1, ... mbs-1 */ 16781278733SSatish Balay j = bj[jmin]; 1689371c9d4SSatish Balay jl[i] = jl[j]; 1699371c9d4SSatish Balay jl[j] = i; /* update jl */ 17081278733SSatish Balay } 17181278733SSatish Balay i = nexti; 17281278733SSatish Balay } 17381278733SSatish Balay 17481278733SSatish Balay /* save nonzero entries in k-th row of U ... */ 17581278733SSatish Balay 17681278733SSatish Balay /* invert diagonal block */ 17781278733SSatish Balay d = ba + k * 25; 1789566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(d, dk, 25)); 1799566063dSJacob Faibussowitsch PetscCall(PetscKernel_A_gets_inverse_A_5(d, ipvt, work, shift, allowzeropivot, &zeropivotdetected)); 1807b6c816cSBarry Smith if (zeropivotdetected) C->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 18181278733SSatish Balay 1829371c9d4SSatish Balay jmin = bi[k]; 1839371c9d4SSatish Balay jmax = bi[k + 1]; 18481278733SSatish Balay if (jmin < jmax) { 18581278733SSatish Balay for (j = jmin; j < jmax; j++) { 18681278733SSatish Balay vj = bj[j]; /* block col. index of U */ 18781278733SSatish Balay u = ba + j * 25; 18881278733SSatish Balay rtmp_ptr = rtmp + vj * 25; 18981278733SSatish Balay for (k1 = 0; k1 < 25; k1++) { 19081278733SSatish Balay *u++ = *rtmp_ptr; 19181278733SSatish Balay *rtmp_ptr++ = 0.0; 19281278733SSatish Balay } 19381278733SSatish Balay } 19481278733SSatish Balay 19581278733SSatish Balay /* ... add k to row list for first nonzero entry in k-th row */ 19681278733SSatish Balay il[k] = jmin; 19781278733SSatish Balay i = bj[jmin]; 1989371c9d4SSatish Balay jl[k] = jl[i]; 1999371c9d4SSatish Balay jl[i] = k; 20081278733SSatish Balay } 20181278733SSatish Balay } 20281278733SSatish Balay 2039566063dSJacob Faibussowitsch PetscCall(PetscFree(rtmp)); 2049566063dSJacob Faibussowitsch PetscCall(PetscFree2(il, jl)); 2059566063dSJacob Faibussowitsch PetscCall(PetscFree2(dk, uik)); 20681278733SSatish Balay 2074f79d315SHong Zhang C->ops->solve = MatSolve_SeqSBAIJ_5_NaturalOrdering_inplace; 2084f79d315SHong Zhang C->ops->solvetranspose = MatSolve_SeqSBAIJ_5_NaturalOrdering_inplace; 2094f79d315SHong Zhang C->ops->forwardsolve = MatForwardSolve_SeqSBAIJ_5_NaturalOrdering_inplace; 2104f79d315SHong Zhang C->ops->backwardsolve = MatBackwardSolve_SeqSBAIJ_5_NaturalOrdering_inplace; 21181278733SSatish Balay C->assembled = PETSC_TRUE; 21281278733SSatish Balay C->preallocated = PETSC_TRUE; 21326fbe8dcSKarl Rupp 2149566063dSJacob Faibussowitsch PetscCall(PetscLogFlops(1.3333 * 125 * b->mbs)); /* from inverting diagonal blocks */ 21581278733SSatish Balay PetscFunctionReturn(0); 21681278733SSatish Balay } 217