1 2 /* 3 This is included by sbaij.c to generate unsigned short and regular versions of these two functions 4 */ 5 #undef __FUNCT__ 6 #if defined(USESHORT) 7 #define __FUNCT__ "MatMult_SeqSBAIJ_1_Hermitian_ushort" 8 PetscErrorCode MatMult_SeqSBAIJ_1_Hermitian_ushort(Mat A,Vec xx,Vec zz) 9 #else 10 #define __FUNCT__ "MatMult_SeqSBAIJ_1_Hermitian" 11 PetscErrorCode MatMult_SeqSBAIJ_1_Hermitian(Mat A,Vec xx,Vec zz) 12 #endif 13 { 14 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data; 15 const PetscScalar *x; 16 PetscScalar *z,x1,sum; 17 const MatScalar *v; 18 MatScalar vj; 19 PetscErrorCode ierr; 20 PetscInt mbs=a->mbs,i,j,nz; 21 const PetscInt *ai=a->i; 22 #if defined(USESHORT) 23 const unsigned short *ib=a->jshort; 24 unsigned short ibt; 25 #else 26 const PetscInt *ib=a->j; 27 PetscInt ibt; 28 #endif 29 PetscInt nonzerorow = 0; 30 31 PetscFunctionBegin; 32 ierr = VecSet(zz,0.0);CHKERRQ(ierr); 33 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 34 ierr = VecGetArray(zz,&z);CHKERRQ(ierr); 35 36 v = a->a; 37 for (i=0; i<mbs; i++) { 38 nz = ai[i+1] - ai[i]; /* length of i_th row of A */ 39 if (!nz) continue; /* Move to the next row if the current row is empty */ 40 nonzerorow++; 41 x1 = x[i]; 42 sum = v[0]*x1; /* diagonal term */ 43 for (j=1; j<nz; j++) { 44 ibt = ib[j]; 45 vj = v[j]; 46 sum += vj * x[ibt]; /* (strict upper triangular part of A)*x */ 47 z[ibt] += PetscConj(v[j]) * x1; /* (strict lower triangular part of A)*x */ 48 } 49 z[i] += sum; 50 v += nz; 51 ib += nz; 52 } 53 54 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 55 ierr = VecRestoreArray(zz,&z);CHKERRQ(ierr); 56 ierr = PetscLogFlops(2.0*(2.0*a->nz - nonzerorow) - nonzerorow);CHKERRQ(ierr); 57 PetscFunctionReturn(0); 58 } 59 60 #undef __FUNCT__ 61 #if defined(USESHORT) 62 #define __FUNCT__ "MatMult_SeqSBAIJ_1_ushort" 63 PetscErrorCode MatMult_SeqSBAIJ_1_ushort(Mat A,Vec xx,Vec zz) 64 #else 65 #define __FUNCT__ "MatMult_SeqSBAIJ_1" 66 PetscErrorCode MatMult_SeqSBAIJ_1(Mat A,Vec xx,Vec zz) 67 #endif 68 { 69 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data; 70 const PetscScalar *x; 71 PetscScalar *z,x1,sum; 72 const MatScalar *v; 73 MatScalar vj; 74 PetscErrorCode ierr; 75 PetscInt mbs=a->mbs,i,j,nz; 76 const PetscInt *ai=a->i; 77 #if defined(USESHORT) 78 const unsigned short *ib=a->jshort; 79 unsigned short ibt; 80 #else 81 const PetscInt *ib=a->j; 82 PetscInt ibt; 83 #endif 84 PetscInt nonzerorow=0; 85 86 PetscFunctionBegin; 87 ierr = VecSet(zz,0.0);CHKERRQ(ierr); 88 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 89 ierr = VecGetArray(zz,&z);CHKERRQ(ierr); 90 91 v = a->a; 92 for (i=0; i<mbs; i++) { 93 nz = ai[i+1] - ai[i]; /* length of i_th row of A */ 94 if (!nz) continue; /* Move to the next row if the current row is empty */ 95 nonzerorow++; 96 x1 = x[i]; 97 sum = v[0]*x1; /* diagonal term */ 98 PetscPrefetchBlock(ib+nz,nz,0,PETSC_PREFETCH_HINT_NTA); /* Indices for the next row (assumes same size as this one) */ 99 PetscPrefetchBlock(v+nz,nz,0,PETSC_PREFETCH_HINT_NTA); /* Entries for the next row */ 100 for (j=1; j<nz; j++) { 101 ibt = ib[j]; 102 vj = v[j]; 103 z[ibt] += vj * x1; /* (strict lower triangular part of A)*x */ 104 sum += vj * x[ibt]; /* (strict upper triangular part of A)*x */ 105 } 106 z[i] += sum; 107 v += nz; 108 ib += nz; 109 } 110 111 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 112 ierr = VecRestoreArray(zz,&z);CHKERRQ(ierr); 113 ierr = PetscLogFlops(2.0*(2.0*a->nz - nonzerorow) - nonzerorow);CHKERRQ(ierr); 114 PetscFunctionReturn(0); 115 } 116 117 #undef __FUNCT__ 118 #if defined(USESHORT) 119 #define __FUNCT__ "MatSOR_SeqSBAIJ_ushort" 120 PetscErrorCode MatSOR_SeqSBAIJ_ushort(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 121 #else 122 #define __FUNCT__ "MatSOR_SeqSBAIJ" 123 PetscErrorCode MatSOR_SeqSBAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 124 #endif 125 { 126 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data; 127 const MatScalar *aa=a->a,*v,*v1,*aidiag; 128 PetscScalar *x,*t,sum; 129 const PetscScalar *b; 130 MatScalar tmp; 131 PetscErrorCode ierr; 132 PetscInt m=a->mbs,bs=A->rmap->bs,j; 133 const PetscInt *ai=a->i; 134 #if defined(USESHORT) 135 const unsigned short *aj=a->jshort,*vj,*vj1; 136 #else 137 const PetscInt *aj=a->j,*vj,*vj1; 138 #endif 139 PetscInt nz,nz1,i; 140 141 PetscFunctionBegin; 142 if (flag & SOR_EISENSTAT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for Eisenstat"); 143 144 its = its*lits; 145 if (its <= 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits); 146 147 if (bs > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented"); 148 149 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 150 ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr); 151 152 if (!a->idiagvalid) { 153 if (!a->idiag) { 154 ierr = PetscMalloc(m*sizeof(PetscScalar),&a->idiag);CHKERRQ(ierr); 155 } 156 for (i=0; i<a->mbs; i++) a->idiag[i] = 1.0/a->a[a->i[i]]; 157 a->idiagvalid = PETSC_TRUE; 158 } 159 160 if (!a->sor_work) { 161 ierr = PetscMalloc(m*sizeof(PetscScalar),&a->sor_work);CHKERRQ(ierr); 162 } 163 t = a->sor_work; 164 165 aidiag = a->idiag; 166 167 if (flag == SOR_APPLY_UPPER) { 168 /* apply (U + D/omega) to the vector */ 169 PetscScalar d; 170 for (i=0; i<m; i++) { 171 d = fshift + aa[ai[i]]; 172 nz = ai[i+1] - ai[i] - 1; 173 vj = aj + ai[i] + 1; 174 v = aa + ai[i] + 1; 175 sum = b[i]*d/omega; 176 PetscSparseDensePlusDot(sum,b,v,vj,nz); 177 x[i] = sum; 178 } 179 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); 180 } 181 182 if (flag & SOR_ZERO_INITIAL_GUESS) { 183 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP){ 184 ierr = PetscMemcpy(t,b,m*sizeof(PetscScalar));CHKERRQ(ierr); 185 186 v = aa + 1; 187 vj = aj + 1; 188 for (i=0; i<m; i++){ 189 nz = ai[i+1] - ai[i] - 1; 190 tmp = - (x[i] = omega*t[i]*aidiag[i]); 191 for (j=0; j<nz; j++) { 192 t[vj[j]] += tmp*v[j]; 193 } 194 v += nz + 1; 195 vj += nz + 1; 196 } 197 ierr = PetscLogFlops(2*a->nz);CHKERRQ(ierr); 198 } 199 200 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP){ 201 int nz2; 202 if (!(flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP)){ 203 #if defined(PETSC_USE_BACKWARD_LOOP) 204 v = aa + ai[m] - 1; 205 vj = aj + ai[m] - 1; 206 for (i=m-1; i>=0; i--){ 207 sum = b[i]; 208 nz = ai[i+1] - ai[i] - 1; 209 {PetscInt __i;for(__i=0;__i<nz;__i++) sum -= v[-__i] * x[vj[-__i]];} 210 #else 211 v = aa + ai[m-1] + 1; 212 vj = aj + ai[m-1] + 1; 213 nz = 0; 214 for (i=m-1; i>=0; i--){ 215 sum = b[i]; 216 nz2 = ai[i] - ai[i-1] - 1; 217 PETSC_Prefetch(v-nz2-1,0,PETSC_PREFETCH_HINT_NTA); 218 PETSC_Prefetch(vj-nz2-1,0,PETSC_PREFETCH_HINT_NTA); 219 PetscSparseDenseMinusDot(sum,x,v,vj,nz); 220 nz = nz2; 221 #endif 222 x[i] = omega*sum*aidiag[i]; 223 v -= nz + 1; 224 vj -= nz + 1; 225 } 226 ierr = PetscLogFlops(2*a->nz);CHKERRQ(ierr); 227 } else { 228 v = aa + ai[m-1] + 1; 229 vj = aj + ai[m-1] + 1; 230 nz = 0; 231 for (i=m-1; i>=0; i--){ 232 sum = t[i]; 233 nz2 = ai[i] - ai[i-1] - 1; 234 PETSC_Prefetch(v-nz2-1,0,PETSC_PREFETCH_HINT_NTA); 235 PETSC_Prefetch(vj-nz2-1,0,PETSC_PREFETCH_HINT_NTA); 236 PetscSparseDenseMinusDot(sum,x,v,vj,nz); 237 x[i] = (1-omega)*x[i] + omega*sum*aidiag[i]; 238 nz = nz2; 239 v -= nz + 1; 240 vj -= nz + 1; 241 } 242 ierr = PetscLogFlops(2*a->nz);CHKERRQ(ierr); 243 } 244 } 245 its--; 246 } 247 248 while (its--) { 249 /* 250 forward sweep: 251 for i=0,...,m-1: 252 sum[i] = (b[i] - U(i,:)x )/d[i]; 253 x[i] = (1-omega)x[i] + omega*sum[i]; 254 b = b - x[i]*U^T(i,:); 255 256 */ 257 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP){ 258 ierr = PetscMemcpy(t,b,m*sizeof(PetscScalar));CHKERRQ(ierr); 259 260 for (i=0; i<m; i++){ 261 v = aa + ai[i] + 1; v1=v; 262 vj = aj + ai[i] + 1; vj1=vj; 263 nz = ai[i+1] - ai[i] - 1; nz1=nz; 264 sum = t[i]; 265 ierr = PetscLogFlops(4.0*nz-2);CHKERRQ(ierr); 266 while (nz1--) sum -= (*v1++)*x[*vj1++]; 267 x[i] = (1-omega)*x[i] + omega*sum*aidiag[i]; 268 while (nz--) t[*vj++] -= x[i]*(*v++); 269 } 270 } 271 272 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP){ 273 /* 274 backward sweep: 275 b = b - x[i]*U^T(i,:), i=0,...,n-2 276 for i=m-1,...,0: 277 sum[i] = (b[i] - U(i,:)x )/d[i]; 278 x[i] = (1-omega)x[i] + omega*sum[i]; 279 */ 280 /* if there was a forward sweep done above then I thing the next two for loops are not needed */ 281 ierr = PetscMemcpy(t,b,m*sizeof(PetscScalar));CHKERRQ(ierr); 282 283 for (i=0; i<m-1; i++){ /* update rhs */ 284 v = aa + ai[i] + 1; 285 vj = aj + ai[i] + 1; 286 nz = ai[i+1] - ai[i] - 1; 287 ierr = PetscLogFlops(2.0*nz-1);CHKERRQ(ierr); 288 while (nz--) t[*vj++] -= x[i]*(*v++); 289 } 290 for (i=m-1; i>=0; i--){ 291 v = aa + ai[i] + 1; 292 vj = aj + ai[i] + 1; 293 nz = ai[i+1] - ai[i] - 1; 294 ierr = PetscLogFlops(2.0*nz-1);CHKERRQ(ierr); 295 sum = t[i]; 296 while (nz--) sum -= x[*vj++]*(*v++); 297 x[i] = (1-omega)*x[i] + omega*sum*aidiag[i]; 298 } 299 } 300 } 301 302 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 303 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 304 PetscFunctionReturn(0); 305 } 306