1 #define PETSCMAT_DLL 2 3 /* 4 Factorization code for BAIJ format. 5 */ 6 #include "../src/mat/impls/baij/seq/baij.h" 7 #include "../src/mat/blockinvert.h" 8 9 /* ------------------------------------------------------------*/ 10 /* 11 Version for when blocks are 5 by 5 12 */ 13 #undef __FUNCT__ 14 #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_5" 15 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5(Mat C,Mat A,const MatFactorInfo *info) 16 { 17 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data; 18 IS isrow = b->row,isicol = b->icol; 19 PetscErrorCode ierr; 20 const PetscInt *r,*ic,*bi = b->i,*bj = b->j,*ajtmpold,*ajtmp; 21 PetscInt i,j,n = a->mbs,nz,row,idx; 22 const PetscInt *diag_offset = b->diag,*ai=a->i,*aj=a->j,*pj; 23 MatScalar *w,*pv,*rtmp,*x,*pc; 24 const MatScalar *v,*aa = a->a; 25 MatScalar p1,p2,p3,p4,m1,m2,m3,m4,m5,m6,m7,m8,m9,x1,x2,x3,x4; 26 MatScalar p5,p6,p7,p8,p9,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15,x16; 27 MatScalar x17,x18,x19,x20,x21,x22,x23,x24,x25,p10,p11,p12,p13,p14; 28 MatScalar p15,p16,p17,p18,p19,p20,p21,p22,p23,p24,p25,m10,m11,m12; 29 MatScalar m13,m14,m15,m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 30 MatScalar *ba = b->a; 31 PetscReal shift = info->shiftinblocks; 32 33 PetscFunctionBegin; 34 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 35 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 36 ierr = PetscMalloc(25*(n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 37 38 #define PETSC_USE_MEMZERO 1 39 #define PETSC_USE_MEMCPY 1 40 41 for (i=0; i<n; i++) { 42 nz = bi[i+1] - bi[i]; 43 ajtmp = bj + bi[i]; 44 for (j=0; j<nz; j++) { 45 #if defined(PETSC_USE_MEMZERO) 46 ierr = PetscMemzero(rtmp+25*ajtmp[j],25*sizeof(PetscScalar));CHKERRQ(ierr); 47 #else 48 x = rtmp+25*ajtmp[j]; 49 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 50 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = x[16] = x[17] = 0.0; 51 x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 52 #endif 53 } 54 /* load in initial (unfactored row) */ 55 idx = r[i]; 56 nz = ai[idx+1] - ai[idx]; 57 ajtmpold = aj + ai[idx]; 58 v = aa + 25*ai[idx]; 59 for (j=0; j<nz; j++) { 60 #if defined(PETSC_USE_MEMCPY) 61 ierr = PetscMemcpy(rtmp+25*ic[ajtmpold[j]],v,25*sizeof(PetscScalar));CHKERRQ(ierr); 62 #else 63 x = rtmp+25*ic[ajtmpold[j]]; 64 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 65 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 66 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 67 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; 68 x[18] = v[18]; x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; 69 x[22] = v[22]; x[23] = v[23]; x[24] = v[24]; 70 #endif 71 v += 25; 72 } 73 row = *ajtmp++; 74 while (row < i) { 75 pc = rtmp + 25*row; 76 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 77 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 78 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 79 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; p19 = pc[18]; 80 p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; p24 = pc[23]; 81 p25 = pc[24]; 82 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 83 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 84 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 85 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || 86 p20 != 0.0 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || 87 p24 != 0.0 || p25 != 0.0) { 88 pv = ba + 25*diag_offset[row]; 89 pj = bj + diag_offset[row] + 1; 90 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 91 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 92 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 93 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 94 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; 95 x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 96 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 97 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 98 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 99 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 100 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 101 102 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 103 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 104 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 105 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 106 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 107 108 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 109 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 110 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 111 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 112 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 113 114 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 115 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 116 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 117 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 118 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 119 120 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 121 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 122 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 123 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 124 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 125 126 nz = bi[row+1] - diag_offset[row] - 1; 127 pv += 25; 128 for (j=0; j<nz; j++) { 129 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 130 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 131 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 132 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; 133 x18 = pv[17]; x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; 134 x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; x25 = pv[24]; 135 x = rtmp + 25*pj[j]; 136 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 137 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 138 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 139 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 140 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 141 142 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 143 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 144 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 145 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 146 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 147 148 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 149 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 150 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 151 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 152 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 153 154 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 155 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 156 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 157 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 158 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 159 160 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 161 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 162 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 163 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 164 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 165 166 pv += 25; 167 } 168 ierr = PetscLogFlops(250.0*nz+225.0);CHKERRQ(ierr); 169 } 170 row = *ajtmp++; 171 } 172 /* finished row so stick it into b->a */ 173 pv = ba + 25*bi[i]; 174 pj = bj + bi[i]; 175 nz = bi[i+1] - bi[i]; 176 for (j=0; j<nz; j++) { 177 #if defined(PETSC_USE_MEMCPY) 178 ierr = PetscMemcpy(pv,rtmp+25*pj[j],25*sizeof(PetscScalar));CHKERRQ(ierr); 179 #else 180 x = rtmp+25*pj[j]; 181 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 182 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 183 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 184 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; 185 pv[17] = x[17]; pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; 186 pv[21] = x[21]; pv[22] = x[22]; pv[23] = x[23]; pv[24] = x[24]; 187 #endif 188 pv += 25; 189 } 190 /* invert diagonal block */ 191 w = ba + 25*diag_offset[i]; 192 ierr = Kernel_A_gets_inverse_A_5(w,shift);CHKERRQ(ierr); 193 } 194 195 ierr = PetscFree(rtmp);CHKERRQ(ierr); 196 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 197 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 198 C->ops->solve = MatSolve_SeqBAIJ_5; 199 C->ops->solvetranspose = MatSolveTranspose_SeqBAIJ_5; 200 C->assembled = PETSC_TRUE; 201 ierr = PetscLogFlops(1.3333*125*b->mbs);CHKERRQ(ierr); /* from inverting diagonal blocks */ 202 PetscFunctionReturn(0); 203 } 204 205 /* MatLUFactorNumeric_SeqBAIJ_5_newdatastruct - 206 copied from MatLUFactorNumeric_SeqBAIJ_N_newdatastruct() and manually re-implemented 207 Kernel_A_gets_A_times_B() 208 Kernel_A_gets_A_minus_B_times_C() 209 Kernel_A_gets_inverse_A() 210 */ 211 #undef __FUNCT__ 212 #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_5_newdatastruct" 213 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 214 { 215 Mat C=B; 216 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ *)C->data; 217 IS isrow = b->row,isicol = b->icol; 218 PetscErrorCode ierr; 219 const PetscInt *r,*ic,*ics; 220 PetscInt i,j,k,n=a->mbs,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 221 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 222 MatScalar *rtmp,*pc,*mwork,*v,*pv,*aa=a->a; 223 PetscInt bs2 = a->bs2,flg; 224 PetscReal shift = info->shiftinblocks; 225 226 PetscFunctionBegin; 227 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 228 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 229 230 /* generate work space needed by the factorization */ 231 ierr = PetscMalloc((bs2*n+bs2+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 232 mwork = rtmp + bs2*n; 233 ierr = PetscMemzero(rtmp,bs2*n*sizeof(MatScalar));CHKERRQ(ierr); 234 ics = ic; 235 236 for (i=0; i<n; i++){ 237 /* zero rtmp */ 238 /* L part */ 239 nz = bi[i+1] - bi[i]; 240 bjtmp = bj + bi[i]; 241 for (j=0; j<nz; j++){ 242 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 243 } 244 245 /* U part */ 246 nz = bi[2*n-i+1] - bi[2*n-i]; 247 bjtmp = bj + bi[2*n-i]; 248 for (j=0; j<nz; j++){ 249 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 250 } 251 252 /* load in initial (unfactored row) */ 253 nz = ai[r[i]+1] - ai[r[i]]; 254 ajtmp = aj + ai[r[i]]; 255 v = aa + bs2*ai[r[i]]; 256 for (j=0; j<nz; j++) { 257 ierr = PetscMemcpy(rtmp+bs2*ic[ajtmp[j]],v+bs2*j,bs2*sizeof(MatScalar));CHKERRQ(ierr); 258 } 259 260 /* elimination */ 261 bjtmp = bj + bi[i]; 262 nzL = bi[i+1] - bi[i]; 263 for(k=0;k < nzL;k++) { 264 row = bjtmp[k]; 265 pc = rtmp + bs2*row; 266 for (flg=0,j=0; j<bs2; j++) { if (pc[j]!=0.0) { flg = 1; break; }} 267 if (flg) { 268 pv = b->a + bs2*bdiag[row]; 269 /* Kernel_A_gets_A_times_B(bs,pc,pv,mwork); *pc = *pc * (*pv); */ 270 ierr = Kernel_A_gets_A_times_B_5(pc,pv,mwork);CHKERRQ(ierr); 271 272 pj = b->j + bi[2*n-row]; /* begining of U(row,:) */ 273 pv = b->a + bs2*bi[2*n-row]; 274 nz = bi[2*n-row+1] - bi[2*n-row] - 1; /* num of entries inU(row,:), excluding diag */ 275 for (j=0; j<nz; j++) { 276 /* Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); */ 277 /* rtmp+bs2*pj[j] = rtmp+bs2*pj[j] - (*pc)*(pv+bs2*j) */ 278 v = rtmp + bs2*pj[j]; 279 ierr = Kernel_A_gets_A_minus_B_times_C_5(v,pc,pv);CHKERRQ(ierr); 280 pv += bs2; 281 } 282 ierr = PetscLogFlops(250*nz+225);CHKERRQ(ierr); /* flops = 2*bs^3*nz + 2*bs^3 - bs2) */ 283 } 284 } 285 286 /* finished row so stick it into b->a */ 287 /* L part */ 288 pv = b->a + bs2*bi[i] ; 289 pj = b->j + bi[i] ; 290 nz = bi[i+1] - bi[i]; 291 for (j=0; j<nz; j++) { 292 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 293 } 294 295 /* Mark diagonal and invert diagonal for simplier triangular solves */ 296 pv = b->a + bs2*bdiag[i]; 297 pj = b->j + bdiag[i]; 298 ierr = PetscMemcpy(pv,rtmp+bs2*pj[0],bs2*sizeof(MatScalar));CHKERRQ(ierr); 299 /* ierr = Kernel_A_gets_inverse_A(bs,pv,v_pivots,v_work);CHKERRQ(ierr); */ 300 ierr = Kernel_A_gets_inverse_A_5(pv,shift);CHKERRQ(ierr); 301 302 /* U part */ 303 pv = b->a + bs2*bi[2*n-i]; 304 pj = b->j + bi[2*n-i]; 305 nz = bi[2*n-i+1] - bi[2*n-i] - 1; 306 for (j=0; j<nz; j++){ 307 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 308 } 309 } 310 311 ierr = PetscFree(rtmp);CHKERRQ(ierr); 312 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 313 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 314 315 C->assembled = PETSC_TRUE; 316 ierr = PetscLogFlops(1.3333*bs2*n);CHKERRQ(ierr); /* from inverting diagonal blocks */ 317 PetscFunctionReturn(0); 318 } 319 320 #undef __FUNCT__ 321 #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_5_newdatastruct_v2" 322 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_newdatastruct_v2(Mat B,Mat A,const MatFactorInfo *info) 323 { 324 Mat C=B; 325 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ *)C->data; 326 IS isrow = b->row,isicol = b->icol; 327 PetscErrorCode ierr; 328 const PetscInt *r,*ic,*ics; 329 PetscInt i,j,k,n=a->mbs,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 330 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 331 MatScalar *rtmp,*pc,*mwork,*v,*pv,*aa=a->a; 332 PetscInt bs2 = a->bs2,flg; 333 PetscReal shift = info->shiftinblocks; 334 335 PetscFunctionBegin; 336 ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 337 ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 338 339 /* generate work space needed by the factorization */ 340 ierr = PetscMalloc((bs2*n+bs2+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 341 mwork = rtmp + bs2*n; 342 ierr = PetscMemzero(rtmp,bs2*n*sizeof(MatScalar));CHKERRQ(ierr); 343 ics = ic; 344 345 for (i=0; i<n; i++){ 346 /* zero rtmp */ 347 /* L part */ 348 nz = bi[i+1] - bi[i]; 349 bjtmp = bj + bi[i]; 350 for (j=0; j<nz; j++){ 351 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 352 } 353 354 /* U part */ 355 nz = bdiag[i] - bdiag[i+1]; 356 bjtmp = bj + bdiag[i+1]+1; 357 for (j=0; j<nz; j++){ 358 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 359 } 360 361 /* load in initial (unfactored row) */ 362 nz = ai[r[i]+1] - ai[r[i]]; 363 ajtmp = aj + ai[r[i]]; 364 v = aa + bs2*ai[r[i]]; 365 for (j=0; j<nz; j++) { 366 ierr = PetscMemcpy(rtmp+bs2*ic[ajtmp[j]],v+bs2*j,bs2*sizeof(MatScalar));CHKERRQ(ierr); 367 } 368 369 /* elimination */ 370 bjtmp = bj + bi[i]; 371 nzL = bi[i+1] - bi[i]; 372 for(k=0;k < nzL;k++) { 373 row = bjtmp[k]; 374 pc = rtmp + bs2*row; 375 for (flg=0,j=0; j<bs2; j++) { if (pc[j]!=0.0) { flg = 1; break; }} 376 if (flg) { 377 pv = b->a + bs2*bdiag[row]; 378 /* Kernel_A_gets_A_times_B(bs,pc,pv,mwork); *pc = *pc * (*pv); */ 379 ierr = Kernel_A_gets_A_times_B_5(pc,pv,mwork);CHKERRQ(ierr); 380 381 pj = b->j + bdiag[row+1]+1; /* begining of U(row,:) */ 382 pv = b->a + bs2*(bdiag[row+1]+1); 383 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries inU(row,:), excluding diag */ 384 for (j=0; j<nz; j++) { 385 /* Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); */ 386 /* rtmp+bs2*pj[j] = rtmp+bs2*pj[j] - (*pc)*(pv+bs2*j) */ 387 v = rtmp + bs2*pj[j]; 388 ierr = Kernel_A_gets_A_minus_B_times_C_5(v,pc,pv);CHKERRQ(ierr); 389 pv += bs2; 390 } 391 ierr = PetscLogFlops(250*nz+225);CHKERRQ(ierr); /* flops = 2*bs^3*nz + 2*bs^3 - bs2) */ 392 } 393 } 394 395 /* finished row so stick it into b->a */ 396 /* L part */ 397 pv = b->a + bs2*bi[i] ; 398 pj = b->j + bi[i] ; 399 nz = bi[i+1] - bi[i]; 400 for (j=0; j<nz; j++) { 401 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 402 } 403 404 /* Mark diagonal and invert diagonal for simplier triangular solves */ 405 pv = b->a + bs2*bdiag[i]; 406 pj = b->j + bdiag[i]; 407 ierr = PetscMemcpy(pv,rtmp+bs2*pj[0],bs2*sizeof(MatScalar));CHKERRQ(ierr); 408 /* ierr = Kernel_A_gets_inverse_A(bs,pv,v_pivots,v_work);CHKERRQ(ierr); */ 409 ierr = Kernel_A_gets_inverse_A_5(pv,shift);CHKERRQ(ierr); 410 411 /* U part */ 412 pv = b->a + bs2*(bdiag[i+1]+1); 413 pj = b->j + bdiag[i+1]+1; 414 nz = bdiag[i] - bdiag[i+1] - 1; 415 for (j=0; j<nz; j++){ 416 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 417 } 418 } 419 420 ierr = PetscFree(rtmp);CHKERRQ(ierr); 421 ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 422 ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 423 424 C->assembled = PETSC_TRUE; 425 ierr = PetscLogFlops(1.3333*bs2*n);CHKERRQ(ierr); /* from inverting diagonal blocks */ 426 PetscFunctionReturn(0); 427 } 428 429 /* 430 Version for when blocks are 5 by 5 Using natural ordering 431 */ 432 #undef __FUNCT__ 433 #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering" 434 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering(Mat C,Mat A,const MatFactorInfo *info) 435 { 436 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data,*b = (Mat_SeqBAIJ *)C->data; 437 PetscErrorCode ierr; 438 PetscInt i,j,n = a->mbs,*bi = b->i,*bj = b->j; 439 PetscInt *ajtmpold,*ajtmp,nz,row; 440 PetscInt *diag_offset = b->diag,*ai=a->i,*aj=a->j,*pj; 441 MatScalar *pv,*v,*rtmp,*pc,*w,*x; 442 MatScalar x1,x2,x3,x4,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15; 443 MatScalar x16,x17,x18,x19,x20,x21,x22,x23,x24,x25; 444 MatScalar p1,p2,p3,p4,p5,p6,p7,p8,p9,p10,p11,p12,p13,p14,p15; 445 MatScalar p16,p17,p18,p19,p20,p21,p22,p23,p24,p25; 446 MatScalar m1,m2,m3,m4,m5,m6,m7,m8,m9,m10,m11,m12,m13,m14,m15; 447 MatScalar m16,m17,m18,m19,m20,m21,m22,m23,m24,m25; 448 MatScalar *ba = b->a,*aa = a->a; 449 PetscReal shift = info->shiftinblocks; 450 451 PetscFunctionBegin; 452 ierr = PetscMalloc(25*(n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 453 for (i=0; i<n; i++) { 454 nz = bi[i+1] - bi[i]; 455 ajtmp = bj + bi[i]; 456 for (j=0; j<nz; j++) { 457 x = rtmp+25*ajtmp[j]; 458 x[0] = x[1] = x[2] = x[3] = x[4] = x[5] = x[6] = x[7] = x[8] = x[9] = 0.0; 459 x[10] = x[11] = x[12] = x[13] = x[14] = x[15] = 0.0; 460 x[16] = x[17] = x[18] = x[19] = x[20] = x[21] = x[22] = x[23] = x[24] = 0.0; 461 } 462 /* load in initial (unfactored row) */ 463 nz = ai[i+1] - ai[i]; 464 ajtmpold = aj + ai[i]; 465 v = aa + 25*ai[i]; 466 for (j=0; j<nz; j++) { 467 x = rtmp+25*ajtmpold[j]; 468 x[0] = v[0]; x[1] = v[1]; x[2] = v[2]; x[3] = v[3]; 469 x[4] = v[4]; x[5] = v[5]; x[6] = v[6]; x[7] = v[7]; x[8] = v[8]; 470 x[9] = v[9]; x[10] = v[10]; x[11] = v[11]; x[12] = v[12]; x[13] = v[13]; 471 x[14] = v[14]; x[15] = v[15]; x[16] = v[16]; x[17] = v[17]; x[18] = v[18]; 472 x[19] = v[19]; x[20] = v[20]; x[21] = v[21]; x[22] = v[22]; x[23] = v[23]; 473 x[24] = v[24]; 474 v += 25; 475 } 476 row = *ajtmp++; 477 while (row < i) { 478 pc = rtmp + 25*row; 479 p1 = pc[0]; p2 = pc[1]; p3 = pc[2]; p4 = pc[3]; 480 p5 = pc[4]; p6 = pc[5]; p7 = pc[6]; p8 = pc[7]; p9 = pc[8]; 481 p10 = pc[9]; p11 = pc[10]; p12 = pc[11]; p13 = pc[12]; p14 = pc[13]; 482 p15 = pc[14]; p16 = pc[15]; p17 = pc[16]; p18 = pc[17]; 483 p19 = pc[18]; p20 = pc[19]; p21 = pc[20]; p22 = pc[21]; p23 = pc[22]; 484 p24 = pc[23]; p25 = pc[24]; 485 if (p1 != 0.0 || p2 != 0.0 || p3 != 0.0 || p4 != 0.0 || p5 != 0.0 || 486 p6 != 0.0 || p7 != 0.0 || p8 != 0.0 || p9 != 0.0 || p10 != 0.0 || 487 p11 != 0.0 || p12 != 0.0 || p13 != 0.0 || p14 != 0.0 || p15 != 0.0 488 || p16 != 0.0 || p17 != 0.0 || p18 != 0.0 || p19 != 0.0 || p20 != 0.0 489 || p21 != 0.0 || p22 != 0.0 || p23 != 0.0 || p24 != 0.0 || p25 != 0.0) { 490 pv = ba + 25*diag_offset[row]; 491 pj = bj + diag_offset[row] + 1; 492 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 493 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 494 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; x14 = pv[13]; 495 x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; x19 = pv[18]; 496 x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; x23 = pv[22]; x24 = pv[23]; 497 x25 = pv[24]; 498 pc[0] = m1 = p1*x1 + p6*x2 + p11*x3 + p16*x4 + p21*x5; 499 pc[1] = m2 = p2*x1 + p7*x2 + p12*x3 + p17*x4 + p22*x5; 500 pc[2] = m3 = p3*x1 + p8*x2 + p13*x3 + p18*x4 + p23*x5; 501 pc[3] = m4 = p4*x1 + p9*x2 + p14*x3 + p19*x4 + p24*x5; 502 pc[4] = m5 = p5*x1 + p10*x2 + p15*x3 + p20*x4 + p25*x5; 503 504 pc[5] = m6 = p1*x6 + p6*x7 + p11*x8 + p16*x9 + p21*x10; 505 pc[6] = m7 = p2*x6 + p7*x7 + p12*x8 + p17*x9 + p22*x10; 506 pc[7] = m8 = p3*x6 + p8*x7 + p13*x8 + p18*x9 + p23*x10; 507 pc[8] = m9 = p4*x6 + p9*x7 + p14*x8 + p19*x9 + p24*x10; 508 pc[9] = m10 = p5*x6 + p10*x7 + p15*x8 + p20*x9 + p25*x10; 509 510 pc[10] = m11 = p1*x11 + p6*x12 + p11*x13 + p16*x14 + p21*x15; 511 pc[11] = m12 = p2*x11 + p7*x12 + p12*x13 + p17*x14 + p22*x15; 512 pc[12] = m13 = p3*x11 + p8*x12 + p13*x13 + p18*x14 + p23*x15; 513 pc[13] = m14 = p4*x11 + p9*x12 + p14*x13 + p19*x14 + p24*x15; 514 pc[14] = m15 = p5*x11 + p10*x12 + p15*x13 + p20*x14 + p25*x15; 515 516 pc[15] = m16 = p1*x16 + p6*x17 + p11*x18 + p16*x19 + p21*x20; 517 pc[16] = m17 = p2*x16 + p7*x17 + p12*x18 + p17*x19 + p22*x20; 518 pc[17] = m18 = p3*x16 + p8*x17 + p13*x18 + p18*x19 + p23*x20; 519 pc[18] = m19 = p4*x16 + p9*x17 + p14*x18 + p19*x19 + p24*x20; 520 pc[19] = m20 = p5*x16 + p10*x17 + p15*x18 + p20*x19 + p25*x20; 521 522 pc[20] = m21 = p1*x21 + p6*x22 + p11*x23 + p16*x24 + p21*x25; 523 pc[21] = m22 = p2*x21 + p7*x22 + p12*x23 + p17*x24 + p22*x25; 524 pc[22] = m23 = p3*x21 + p8*x22 + p13*x23 + p18*x24 + p23*x25; 525 pc[23] = m24 = p4*x21 + p9*x22 + p14*x23 + p19*x24 + p24*x25; 526 pc[24] = m25 = p5*x21 + p10*x22 + p15*x23 + p20*x24 + p25*x25; 527 528 nz = bi[row+1] - diag_offset[row] - 1; 529 pv += 25; 530 for (j=0; j<nz; j++) { 531 x1 = pv[0]; x2 = pv[1]; x3 = pv[2]; x4 = pv[3]; 532 x5 = pv[4]; x6 = pv[5]; x7 = pv[6]; x8 = pv[7]; x9 = pv[8]; 533 x10 = pv[9]; x11 = pv[10]; x12 = pv[11]; x13 = pv[12]; 534 x14 = pv[13]; x15 = pv[14]; x16 = pv[15]; x17 = pv[16]; x18 = pv[17]; 535 x19 = pv[18]; x20 = pv[19]; x21 = pv[20]; x22 = pv[21]; x23 = pv[22]; 536 x24 = pv[23]; x25 = pv[24]; 537 x = rtmp + 25*pj[j]; 538 x[0] -= m1*x1 + m6*x2 + m11*x3 + m16*x4 + m21*x5; 539 x[1] -= m2*x1 + m7*x2 + m12*x3 + m17*x4 + m22*x5; 540 x[2] -= m3*x1 + m8*x2 + m13*x3 + m18*x4 + m23*x5; 541 x[3] -= m4*x1 + m9*x2 + m14*x3 + m19*x4 + m24*x5; 542 x[4] -= m5*x1 + m10*x2 + m15*x3 + m20*x4 + m25*x5; 543 544 x[5] -= m1*x6 + m6*x7 + m11*x8 + m16*x9 + m21*x10; 545 x[6] -= m2*x6 + m7*x7 + m12*x8 + m17*x9 + m22*x10; 546 x[7] -= m3*x6 + m8*x7 + m13*x8 + m18*x9 + m23*x10; 547 x[8] -= m4*x6 + m9*x7 + m14*x8 + m19*x9 + m24*x10; 548 x[9] -= m5*x6 + m10*x7 + m15*x8 + m20*x9 + m25*x10; 549 550 x[10] -= m1*x11 + m6*x12 + m11*x13 + m16*x14 + m21*x15; 551 x[11] -= m2*x11 + m7*x12 + m12*x13 + m17*x14 + m22*x15; 552 x[12] -= m3*x11 + m8*x12 + m13*x13 + m18*x14 + m23*x15; 553 x[13] -= m4*x11 + m9*x12 + m14*x13 + m19*x14 + m24*x15; 554 x[14] -= m5*x11 + m10*x12 + m15*x13 + m20*x14 + m25*x15; 555 556 x[15] -= m1*x16 + m6*x17 + m11*x18 + m16*x19 + m21*x20; 557 x[16] -= m2*x16 + m7*x17 + m12*x18 + m17*x19 + m22*x20; 558 x[17] -= m3*x16 + m8*x17 + m13*x18 + m18*x19 + m23*x20; 559 x[18] -= m4*x16 + m9*x17 + m14*x18 + m19*x19 + m24*x20; 560 x[19] -= m5*x16 + m10*x17 + m15*x18 + m20*x19 + m25*x20; 561 562 x[20] -= m1*x21 + m6*x22 + m11*x23 + m16*x24 + m21*x25; 563 x[21] -= m2*x21 + m7*x22 + m12*x23 + m17*x24 + m22*x25; 564 x[22] -= m3*x21 + m8*x22 + m13*x23 + m18*x24 + m23*x25; 565 x[23] -= m4*x21 + m9*x22 + m14*x23 + m19*x24 + m24*x25; 566 x[24] -= m5*x21 + m10*x22 + m15*x23 + m20*x24 + m25*x25; 567 pv += 25; 568 } 569 ierr = PetscLogFlops(250.0*nz+225.0);CHKERRQ(ierr); 570 } 571 row = *ajtmp++; 572 } 573 /* finished row so stick it into b->a */ 574 pv = ba + 25*bi[i]; 575 pj = bj + bi[i]; 576 nz = bi[i+1] - bi[i]; 577 for (j=0; j<nz; j++) { 578 x = rtmp+25*pj[j]; 579 pv[0] = x[0]; pv[1] = x[1]; pv[2] = x[2]; pv[3] = x[3]; 580 pv[4] = x[4]; pv[5] = x[5]; pv[6] = x[6]; pv[7] = x[7]; pv[8] = x[8]; 581 pv[9] = x[9]; pv[10] = x[10]; pv[11] = x[11]; pv[12] = x[12]; 582 pv[13] = x[13]; pv[14] = x[14]; pv[15] = x[15]; pv[16] = x[16]; pv[17] = x[17]; 583 pv[18] = x[18]; pv[19] = x[19]; pv[20] = x[20]; pv[21] = x[21]; pv[22] = x[22]; 584 pv[23] = x[23]; pv[24] = x[24]; 585 pv += 25; 586 } 587 /* invert diagonal block */ 588 w = ba + 25*diag_offset[i]; 589 ierr = Kernel_A_gets_inverse_A_5(w,shift);CHKERRQ(ierr); 590 } 591 592 ierr = PetscFree(rtmp);CHKERRQ(ierr); 593 C->ops->solve = MatSolve_SeqBAIJ_5_NaturalOrdering; 594 C->ops->solvetranspose = MatSolveTranspose_SeqBAIJ_5_NaturalOrdering; 595 C->assembled = PETSC_TRUE; 596 ierr = PetscLogFlops(1.3333*125*b->mbs);CHKERRQ(ierr); /* from inverting diagonal blocks */ 597 PetscFunctionReturn(0); 598 } 599 600 #undef __FUNCT__ 601 #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering_newdatastruct" 602 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering_newdatastruct(Mat B,Mat A,const MatFactorInfo *info) 603 { 604 Mat C=B; 605 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ *)C->data; 606 PetscErrorCode ierr; 607 PetscInt i,j,k,n=a->mbs,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 608 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 609 MatScalar *rtmp,*pc,*mwork,*v,*pv,*aa=a->a; 610 PetscInt bs2 = a->bs2,flg; 611 PetscReal shift = info->shiftinblocks; 612 613 PetscFunctionBegin; 614 /* generate work space needed by the factorization */ 615 ierr = PetscMalloc((bs2*n+bs2+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 616 mwork = rtmp + bs2*n; 617 ierr = PetscMemzero(rtmp,bs2*n*sizeof(MatScalar));CHKERRQ(ierr); 618 619 for (i=0; i<n; i++){ 620 /* zero rtmp */ 621 /* L part */ 622 nz = bi[i+1] - bi[i]; 623 bjtmp = bj + bi[i]; 624 for (j=0; j<nz; j++){ 625 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 626 } 627 628 /* U part */ 629 nz = bi[2*n-i+1] - bi[2*n-i]; 630 bjtmp = bj + bi[2*n-i]; 631 for (j=0; j<nz; j++){ 632 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 633 } 634 635 /* load in initial (unfactored row) */ 636 nz = ai[i+1] - ai[i]; 637 ajtmp = aj + ai[i]; 638 v = aa + bs2*ai[i]; 639 for (j=0; j<nz; j++) { 640 ierr = PetscMemcpy(rtmp+bs2*ajtmp[j],v+bs2*j,bs2*sizeof(MatScalar));CHKERRQ(ierr); 641 } 642 643 /* elimination */ 644 bjtmp = bj + bi[i]; 645 nzL = bi[i+1] - bi[i]; 646 for(k=0;k < nzL;k++) { 647 row = bjtmp[k]; 648 pc = rtmp + bs2*row; 649 for (flg=0,j=0; j<bs2; j++) { if (pc[j]!=0.0) { flg = 1; break; }} 650 if (flg) { 651 pv = b->a + bs2*bdiag[row]; 652 /* Kernel_A_gets_A_times_B(bs,pc,pv,mwork); *pc = *pc * (*pv); */ 653 ierr = Kernel_A_gets_A_times_B_5(pc,pv,mwork);CHKERRQ(ierr); 654 655 pj = b->j + bi[2*n-row]; /* begining of U(row,:) */ 656 pv = b->a + bs2*bi[2*n-row]; 657 nz = bi[2*n-row+1] - bi[2*n-row] - 1; /* num of entries inU(row,:), excluding diag */ 658 for (j=0; j<nz; j++) { 659 /* Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); */ 660 /* rtmp+bs2*pj[j] = rtmp+bs2*pj[j] - (*pc)*(pv+bs2*j) */ 661 v = rtmp + bs2*pj[j]; 662 ierr = Kernel_A_gets_A_minus_B_times_C_5(v,pc,pv);CHKERRQ(ierr); 663 pv += bs2; 664 } 665 ierr = PetscLogFlops(250*nz+225);CHKERRQ(ierr); /* flops = 2*bs^3*nz + 2*bs^3 - bs2) */ 666 } 667 } 668 669 /* finished row so stick it into b->a */ 670 /* L part */ 671 pv = b->a + bs2*bi[i] ; 672 pj = b->j + bi[i] ; 673 nz = bi[i+1] - bi[i]; 674 for (j=0; j<nz; j++) { 675 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 676 } 677 678 /* Mark diagonal and invert diagonal for simplier triangular solves */ 679 pv = b->a + bs2*bdiag[i]; 680 pj = b->j + bdiag[i]; 681 ierr = PetscMemcpy(pv,rtmp+bs2*pj[0],bs2*sizeof(MatScalar));CHKERRQ(ierr); 682 /* ierr = Kernel_A_gets_inverse_A(bs,pv,v_pivots,v_work);CHKERRQ(ierr); */ 683 ierr = Kernel_A_gets_inverse_A_5(pv,shift);CHKERRQ(ierr); 684 685 /* U part */ 686 pv = b->a + bs2*bi[2*n-i]; 687 pj = b->j + bi[2*n-i]; 688 nz = bi[2*n-i+1] - bi[2*n-i] - 1; 689 for (j=0; j<nz; j++){ 690 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 691 } 692 } 693 694 ierr = PetscFree(rtmp);CHKERRQ(ierr); 695 C->assembled = PETSC_TRUE; 696 ierr = PetscLogFlops(1.3333*bs2*n);CHKERRQ(ierr); /* from inverting diagonal blocks */ 697 PetscFunctionReturn(0); 698 } 699 700 #undef __FUNCT__ 701 #define __FUNCT__ "MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering_newdatastruct_v2" 702 PetscErrorCode MatLUFactorNumeric_SeqBAIJ_5_NaturalOrdering_newdatastruct_v2(Mat B,Mat A,const MatFactorInfo *info) 703 { 704 Mat C=B; 705 Mat_SeqBAIJ *a=(Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ *)C->data; 706 PetscErrorCode ierr; 707 PetscInt i,j,k,n=a->mbs,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j; 708 PetscInt *ajtmp,*bjtmp,nz,nzL,row,*bdiag=b->diag,*pj; 709 MatScalar *rtmp,*pc,*mwork,*v,*pv,*aa=a->a; 710 PetscInt bs2 = a->bs2,flg; 711 PetscReal shift = info->shiftinblocks; 712 713 PetscFunctionBegin; 714 /* generate work space needed by the factorization */ 715 ierr = PetscMalloc((bs2*n+bs2+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr); 716 mwork = rtmp + bs2*n; 717 ierr = PetscMemzero(rtmp,bs2*n*sizeof(MatScalar));CHKERRQ(ierr); 718 719 for (i=0; i<n; i++){ 720 /* zero rtmp */ 721 /* L part */ 722 nz = bi[i+1] - bi[i]; 723 bjtmp = bj + bi[i]; 724 for (j=0; j<nz; j++){ 725 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 726 } 727 728 /* U part */ 729 nz = bdiag[i] - bdiag[i+1]; 730 bjtmp = bj + bdiag[i+1]+1; 731 for (j=0; j<nz; j++){ 732 ierr = PetscMemzero(rtmp+bs2*bjtmp[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 733 } 734 735 /* load in initial (unfactored row) */ 736 nz = ai[i+1] - ai[i]; 737 ajtmp = aj + ai[i]; 738 v = aa + bs2*ai[i]; 739 for (j=0; j<nz; j++) { 740 ierr = PetscMemcpy(rtmp+bs2*ajtmp[j],v+bs2*j,bs2*sizeof(MatScalar));CHKERRQ(ierr); 741 } 742 743 /* elimination */ 744 bjtmp = bj + bi[i]; 745 nzL = bi[i+1] - bi[i]; 746 for(k=0;k < nzL;k++) { 747 row = bjtmp[k]; 748 pc = rtmp + bs2*row; 749 for (flg=0,j=0; j<bs2; j++) { if (pc[j]!=0.0) { flg = 1; break; }} 750 if (flg) { 751 pv = b->a + bs2*bdiag[row]; 752 /* Kernel_A_gets_A_times_B(bs,pc,pv,mwork); *pc = *pc * (*pv); */ 753 ierr = Kernel_A_gets_A_times_B_5(pc,pv,mwork);CHKERRQ(ierr); 754 755 pj = b->j + bdiag[row+1]+1; /* begining of U(row,:) */ 756 pv = b->a + bs2*(bdiag[row+1]+1); 757 nz = bdiag[row] - bdiag[row+1] - 1; /* num of entries inU(row,:), excluding diag */ 758 for (j=0; j<nz; j++) { 759 /* Kernel_A_gets_A_minus_B_times_C(bs,rtmp+bs2*pj[j],pc,pv+bs2*j); */ 760 /* rtmp+bs2*pj[j] = rtmp+bs2*pj[j] - (*pc)*(pv+bs2*j) */ 761 v = rtmp + bs2*pj[j]; 762 ierr = Kernel_A_gets_A_minus_B_times_C_5(v,pc,pv);CHKERRQ(ierr); 763 pv += bs2; 764 } 765 ierr = PetscLogFlops(250*nz+225);CHKERRQ(ierr); /* flops = 2*bs^3*nz + 2*bs^3 - bs2) */ 766 } 767 } 768 769 /* finished row so stick it into b->a */ 770 /* L part */ 771 pv = b->a + bs2*bi[i] ; 772 pj = b->j + bi[i] ; 773 nz = bi[i+1] - bi[i]; 774 for (j=0; j<nz; j++) { 775 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 776 } 777 778 /* Mark diagonal and invert diagonal for simplier triangular solves */ 779 pv = b->a + bs2*bdiag[i]; 780 pj = b->j + bdiag[i]; 781 ierr = PetscMemcpy(pv,rtmp+bs2*pj[0],bs2*sizeof(MatScalar));CHKERRQ(ierr); 782 /* ierr = Kernel_A_gets_inverse_A(bs,pv,v_pivots,v_work);CHKERRQ(ierr); */ 783 ierr = Kernel_A_gets_inverse_A_5(pv,shift);CHKERRQ(ierr); 784 785 /* U part */ 786 pv = b->a + bs2*(bdiag[i+1]+1); 787 pj = b->j + bdiag[i+1]+1; 788 nz = bdiag[i] - bdiag[i+1] - 1; 789 for (j=0; j<nz; j++){ 790 ierr = PetscMemcpy(pv+bs2*j,rtmp+bs2*pj[j],bs2*sizeof(MatScalar));CHKERRQ(ierr); 791 } 792 } 793 794 ierr = PetscFree(rtmp);CHKERRQ(ierr); 795 C->assembled = PETSC_TRUE; 796 ierr = PetscLogFlops(1.3333*bs2*n);CHKERRQ(ierr); /* from inverting diagonal blocks */ 797 PetscFunctionReturn(0); 798 } 799