void lis_quad_sub(LIS_QUAD *a, const LIS_QUAD *b, const LIS_QUAD *c)
{
LIS_QUAD_DECLAR;
#ifndef USE_SSE2
LIS_QUAD_ADD(a->hi,a->lo,b->hi,b->lo,-c->hi,-c->lo);
#else
LIS_QUAD_ADD_SSE2(a->hi,a->lo,b->hi,b->lo,-c->hi,-c->lo);
#endif
}
LIS_INT lis_reduce_mp(LIS_COMMTABLE commtable, LIS_VECTOR X)
{
LIS_INT neib,i,is,inum,neibpetot,pad;
LIS_SCALAR *x,*xl;
LIS_SCALAR *ws,*wr;
LIS_QUAD_DECLAR;
LIS_DEBUG_FUNC_IN;
neibpetot = commtable->neibpetot;
ws = commtable->ws;
wr = commtable->wr;
pad = commtable->pad;
x = X->value;
xl = X->value_lo;
for(neib=0;neib<neibpetot;neib++)
{
is = commtable->import_ptr[neib];
inum = commtable->import_ptr[neib+1] - is;
for(i=0;i<inum;i++)
{
wr[is*2+i] = x[commtable->import_index[is+i]+pad];
wr[is*2+inum+i] = xl[commtable->import_index[is+i]+pad];
}
MPI_Isend(&wr[is*2],inum*2,MPI_DOUBLE,commtable->neibpe[neib],0,commtable->comm,&commtable->req1[neib]);
}
for(neib=0;neib<neibpetot;neib++)
{
is = commtable->export_ptr[neib];
inum = commtable->export_ptr[neib+1] - is;
MPI_Irecv(&ws[is*2],inum*2,MPI_DOUBLE,commtable->neibpe[neib],0,commtable->comm,&commtable->req2[neib]);
}
MPI_Waitall(neibpetot, commtable->req2, commtable->sta2);
for(neib=0;neib<neibpetot;neib++)
{
is = commtable->export_ptr[neib];
inum = commtable->export_ptr[neib+1] - is;
for(i=0;i<inum;i++)
{
/*x[commtable->export_index[i]] += ws[i];*/
#ifndef USE_SSE2
LIS_QUAD_ADD(x[commtable->export_index[is+i]],xl[commtable->export_index[is+i]],x[commtable->export_index[is+i]],xl[commtable->export_index[is+i]],ws[is*2+i],ws[is*2+inum+i]);
#else
LIS_QUAD_ADD_SSE2(x[commtable->export_index[is+i]],xl[commtable->export_index[is+i]],x[commtable->export_index[is+i]],xl[commtable->export_index[is+i]],ws[is*2+i],ws[is*2+inum+i]);
#endif
}
}
MPI_Waitall(neibpetot, commtable->req1, commtable->sta1);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
void lis_mpi_msum(LIS_QUAD *invec, LIS_QUAD *inoutvec, LIS_INT *len, MPI_Datatype *datatype)
{
LIS_INT i;
LIS_QUAD_DECLAR;
LIS_DEBUG_FUNC_IN;
for(i=0;i<*len;i++)
{
#ifndef USE_SSE2
LIS_QUAD_ADD(inoutvec[i].hi,inoutvec[i].lo,inoutvec[i].hi,inoutvec[i].lo,invec[i].hi,invec[i].lo);
#else
LIS_QUAD_ADD_SSE2(inoutvec[i].hi,inoutvec[i].lo,inoutvec[i].hi,inoutvec[i].lo,invec[i].hi,invec[i].lo);
#endif
}
LIS_DEBUG_FUNC_OUT;
}
LIS_INT lis_matvec_ilu(LIS_MATRIX A, LIS_MATRIX_ILU LU, LIS_VECTOR X, LIS_VECTOR Y)
{
LIS_INT i,j,jj,n,np;
LIS_SCALAR *x,*y;
#ifdef _OPENMP
LIS_INT nprocs,k;
LIS_SCALAR t,*w;
#endif
#ifdef USE_QUAD_PRECISION
LIS_INT j0,j1;
#ifdef _OPENMP
LIS_SCALAR *ww,*wwl;
#endif
#endif
LIS_QUAD_DECLAR;
LIS_DEBUG_FUNC_IN;
np = A->np;
n = LU->n;
x = X->value;
y = Y->value;
#ifdef USE_QUAD_PRECISION
if( X->precision==LIS_PRECISION_DEFAULT )
#endif
{
#ifdef USE_MPI
LIS_MATVEC_SENDRECV;
#endif
#ifdef _OPENMP
nprocs = omp_get_max_threads();
w = (LIS_SCALAR *)lis_malloc( nprocs*np*sizeof(LIS_SCALAR),"lis_matvect_crs::w" );
#pragma omp parallel private(i,j,k,jj,t)
{
k = omp_get_thread_num();
#pragma omp for
for(j=0;j<nprocs;j++)
{
memset( &w[j*np], 0, np*sizeof(LIS_SCALAR) );
}
#pragma omp for
for(i=0;i<n;i++)
{
for(j=0;j<LU->nnz[i];j++)
{
jj = k*np + LU->index[i][j];
w[jj] += LU->value[i][j] * X->value[i];
}
}
#pragma omp for
for(i=0;i<np;i++)
{
t = 0.0;
for(j=0;j<nprocs;j++)
{
t += w[j*np+i];
}
Y->value[i] = t;
}
}
lis_free(w);
#else
for(i=0;i<np;i++)
{
Y->value[i] = 0.0;
}
for(i=0;i<n;i++)
{
for(j=0;j<LU->nnz[i];j++)
{
jj = LU->index[i][j];
Y->value[jj] += LU->value[i][j] * X->value[i];
}
}
#endif
}
#ifdef USE_QUAD_PRECISION
else
{
#ifdef USE_MPI
lis_send_recv_mp(A->commtable,X);
#endif
#ifdef _OPENMP
#ifndef USE_FMA2_SSE2
nprocs = omp_get_max_threads();
ww = (LIS_SCALAR *)lis_malloc( 2*nprocs*np*sizeof(LIS_SCALAR),"lis_matvect_crs_mp::ww" );
wwl = &ww[nprocs*np];
#ifndef USE_SSE2
#pragma omp parallel private(i,j,jj,k,p1,p2,tq,bhi,blo,chi,clo,sh,sl,th,tl,eh,el)
#else
#pragma omp parallel private(i,j,jj,k,bh,ch,sh,wh,th,bl,cl,sl,wl,tl,p1,p2,t0,t1,t2,eh)
#endif
{
k = omp_get_thread_num();
#pragma omp for
for(j=0;j<nprocs;j++)
{
memset( &ww[j*np], 0, np*sizeof(LIS_SCALAR) );
memset( &wwl[j*np], 0, np*sizeof(LIS_SCALAR) );
}
#pragma omp for
for(i=0;i<n;i++)
{
for(j=0;j<LU->nnz[i];j++)
{
jj = k*np + LU->index[i][j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(ww[jj],wwl[jj],ww[jj],wwl[jj],X->value[i],X->value_lo[i],LU->value[i][j]);
#else
LIS_QUAD_FMAD_SSE2(ww[jj],wwl[jj],ww[jj],wwl[jj],X->value[i],X->value_lo[i],LU->value[i][j]);
#endif
}
}
#pragma omp for
for(i=0;i<np;i++)
{
Y->value[i] = Y->value_lo[i] = 0.0;
for(j=0;j<nprocs;j++)
{
#ifndef USE_SSE2
LIS_QUAD_ADD(Y->value[i],Y->value_lo[i],Y->value[i],Y->value_lo[i],ww[j*np+i],wwl[j*np+i]);
#else
LIS_QUAD_ADD_SSE2(Y->value[i],Y->value_lo[i],Y->value[i],Y->value_lo[i],ww[j*np+i],wwl[j*np+i]);
#endif
}
}
}
lis_free(ww);
#else
nprocs = omp_get_max_threads();
ww = (LIS_SCALAR *)lis_malloc( 2*nprocs*np*sizeof(LIS_SCALAR), "lis_matvect_crs_mp2::ww" );
wwl = &ww[nprocs*np];
#pragma omp parallel private(i,j,j0,j1,k,bh,ch,sh,wh,th,bl,cl,sl,wl,tl,p1,p2,t0,t1,t2,eh)
{
k = omp_get_thread_num();
#pragma omp for
for(j=0;j<nprocs;j++)
{
memset( &ww[j*np], 0, np*sizeof(LIS_SCALAR) );
memset( &wwl[j*np], 0, np*sizeof(LIS_SCALAR) );
}
#pragma omp for
for(i=0; i<n; i++)
{
for(j=0;j<LU->nnz[i]-1;j+=2)
{
j0 = k*np + LU->index[i][j];
j1 = k*np + LU->index[i][j+1];
#ifdef USE_SSE2
LIS_QUAD_FMAD2_SSE2_STSD(ww[j0],wwl[j0],ww[j1],wwl[j1],ww[j0],wwl[j0],ww[j1],wwl[j1],X->value[i],X->value_lo[i],X->value[i],X->value_lo[i],LU->value[i][j]);
#endif
}
for(;j<LU->nnz[i];j++)
{
j0 = LU->index[i][j];
#ifdef USE_SSE2
LIS_QUAD_FMAD_SSE2(ww[j0],wwl[j0],ww[j0],wwl[j0],X->value[i],X->value_lo[i],LU->value[i][j]);
#endif
}
}
#pragma omp for
for(i=0;i<np;i++)
{
Y->value[i] = Y->value_lo[i] = 0.0;
for(j=0;j<nprocs;j++)
{
#ifdef USE_SSE2
LIS_QUAD_ADD_SSE2(Y->value[i],Y->value_lo[i],Y->value[i],Y->value_lo[i],ww[j*np+i],wwl[j*np+i]);
#endif
}
}
}
lis_free(ww);
#endif
#else
#ifndef USE_FMA2_SSE2
for(i=0;i<np;i++)
{
Y->value[i] = 0.0;
Y->value_lo[i] = 0.0;
}
for(i=0;i<n;i++)
{
for(j=0;j<LU->nnz[i];j++)
{
jj = LU->index[i][j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(Y->value[jj],Y->value_lo[jj],Y->value[jj],Y->value_lo[jj],X->value[i],X->value_lo[i],LU->value[i][j]);
#else
LIS_QUAD_FMAD_SSE2(Y->value[jj],Y->value_lo[jj],Y->value[jj],Y->value_lo[jj],X->value[i],X->value_lo[i],LU->value[i][j]);
#endif
}
}
#else
for(i=0; i<np; i++)
{
Y->value[i] = 0.0;
Y->value_lo[i] = 0.0;
}
for(i=0; i<n; i++)
{
for(j=0;j<LU->nnz[i]-1;j+=2)
{
j0 = LU->index[i][j];
j1 = LU->index[i][j+1];
#ifdef USE_SSE2
LIS_QUAD_FMAD2_SSE2_STSD(Y->value[j0],Y->value_lo[j0],Y->value[j1],Y->value_lo[j1],Y->value[j0],Y->value_lo[j0],Y->value[j1],Y->value_lo[j1],X->value[i],X->value_lo[i],X->value[i],X->value_lo[i],LU->value[i][j]);
#endif
}
for(;j<LU->nnz[i];j++)
{
j0 = LU->index[i][j];
#ifdef USE_SSE2
LIS_QUAD_FMAD_SSE2(Y->value[j0],Y->value_lo[j0],Y->value[j0],Y->value_lo[j0],X->value[i],X->value_lo[i],LU->value[i][j]);
#endif
}
}
#endif
#endif
}
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_psolve_adds(LIS_SOLVER solver, LIS_VECTOR B, LIS_VECTOR X)
{
LIS_INT i,k,n,np,iter,ptype;
LIS_SCALAR *b,*x,*w,*r,*rl;
LIS_VECTOR W,R;
LIS_PRECON precon;
LIS_QUAD_DECLAR;
LIS_DEBUG_FUNC_IN;
precon = solver->precon;
n = precon->A->n;
np = precon->A->np;
W = precon->work[0];
R = precon->work[1];
b = B->value;
x = X->value;
w = W->value;
r = R->value;
rl = R->value_lo;
iter = solver->options[LIS_OPTIONS_ADDS_ITER];
ptype = solver->options[LIS_OPTIONS_PRECON];
#ifdef USE_QUAD_PRECISION
if( solver->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_set_all(0.0,X);
lis_vector_copy(B,R);
for(k=0;k<iter+1;k++)
{
for(i=n;i<np;i++)
{
r[i] = 0.0;
}
lis_psolve_xxx[ptype](solver,R,W);
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for(i=0;i<n;i++)
{
x[i] += w[i];
}
if(k!=iter)
{
lis_matvec(precon->A,X,R);
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for(i=0;i<n;i++)
{
r[i] = b[i] - r[i];
}
}
}
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_set_allex_nm(0.0,X);
lis_vector_copyex_mm(B,R);
for(k=0;k<iter+1;k++)
{
for(i=n;i<np;i++)
{
r[i] = 0.0;
rl[i] = 0.0;
}
lis_psolve_xxx[ptype](solver,R,W);
for(i=0;i<n;i++)
{
#ifndef USE_SSE2
LIS_QUAD_ADD(X->value[i],X->value_lo[i],X->value[i],X->value_lo[i],W->value[i],W->value_lo[i]);
#else
LIS_QUAD_ADD_SSE2(X->value[i],X->value_lo[i],X->value[i],X->value_lo[i],W->value[i],W->value_lo[i]);
#endif
/* x[i] += w[i];*/
}
if(k==iter) break;
lis_matvec(precon->A,X,R);
for(i=0;i<n;i++)
{
#ifndef USE_SSE2
LIS_QUAD_ADD(R->value[i],R->value_lo[i],B->value[i],B->value_lo[i],-R->value[i],-R->value_lo[i]);
#else
LIS_QUAD_ADD_SSE2(R->value[i],R->value_lo[i],B->value[i],B->value_lo[i],-R->value[i],-R->value_lo[i]);
#endif
/* r[i] = b[i] - r[i];*/
}
}
}
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_vector_nrm2ex_mm(LIS_VECTOR vx, LIS_QUAD_PTR *val)
{
LIS_INT i,n;
LIS_SCALAR *x,*xl;
LIS_QUAD_PTR dotm2,dotm,tmpm;
#ifdef _OPENMP
LIS_INT is,ie,nprocs,my_rank;
LIS_SCALAR *gt;
#endif
#ifdef USE_MPI
MPI_Comm comm;
#endif
LIS_QUAD_DECLAR;
LIS_DEBUG_FUNC_IN;
n = vx->n;
x = vx->value;
xl = vx->value_lo;
dotm2.hi = &vx->work[0];
dotm2.lo = &vx->work[2];
dotm.hi = &vx->work[8];
dotm.lo = &vx->work[9];
tmpm.hi = &vx->work[10];
tmpm.lo = &vx->work[11];
#ifdef USE_MPI
comm = vx->comm;
#endif
#ifdef _OPENMP
gt = lis_vec_tmp;
nprocs = omp_get_max_threads();
#ifndef USE_SSE2
#pragma omp parallel private(i,is,ie,my_rank,p1,p2,tq,bhi,blo,chi,clo,sh,eh,sl,el,th,tl)
#else
#pragma omp parallel private(i,bh,ch,sh,wh,th,bl,cl,sl,wl,tl,p1,p2,t0,t1,t2,eh,is,ie,my_rank)
#endif
{
my_rank = omp_get_thread_num();
LIS_GET_ISIE(my_rank,nprocs,n,is,ie);
#ifndef USE_FMA2_SSE2
gt[my_rank*LIS_VEC_TMP_PADD] = gt[my_rank*LIS_VEC_TMP_PADD+1] = 0.0;
#pragma cdir nodep
for(i=is;i<ie;i++)
{
LIS_QUAD_FSA(gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+1],gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+1],x[i],xl[i]);
}
#else
gt[my_rank*LIS_VEC_TMP_PADD ] = gt[my_rank*LIS_VEC_TMP_PADD+1] = 0.0;
gt[my_rank*LIS_VEC_TMP_PADD+2] = gt[my_rank*LIS_VEC_TMP_PADD+3] = 0.0;
#ifdef USE_VEC_COMP
#pragma cdir nodep
#endif
for(i=is;i<ie-1;i+=2)
{
LIS_QUAD_FSA2_SSE2(gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+2],gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+2],x[i],xl[i]);
}
LIS_QUAD_ADD_SSE2(gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+1],gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+2],gt[my_rank*LIS_VEC_TMP_PADD+1],gt[my_rank*LIS_VEC_TMP_PADD+3]);
for(;i<ie;i++)
{
LIS_QUAD_FSA_SSE2(gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+1],gt[my_rank*LIS_VEC_TMP_PADD],gt[my_rank*LIS_VEC_TMP_PADD+1],x[i],xl[i]);
}
#endif
}
dotm.hi[0] = dotm.lo[0] = 0.0;
for(i=0;i<nprocs;i++)
{
#ifndef USE_SSE2
LIS_QUAD_ADD(dotm.hi[0],dotm.lo[0],dotm.hi[0],dotm.lo[0],gt[i*LIS_VEC_TMP_PADD],gt[i*LIS_VEC_TMP_PADD+1]);
#else
LIS_QUAD_ADD_SSE2(dotm.hi[0],dotm.lo[0],dotm.hi[0],dotm.lo[0],gt[i*LIS_VEC_TMP_PADD],gt[i*LIS_VEC_TMP_PADD+1]);
#endif
}
#else
#ifndef USE_FMA2_SSE2
dotm.hi[0] = dotm.lo[0] = 0.0;
#pragma cdir nodep
for(i=0;i<n;i++)
{
LIS_QUAD_FSA(dotm.hi[0],dotm.lo[0],dotm.hi[0],dotm.lo[0],x[i],xl[i]);
}
#else
dotm2.hi[0] = dotm2.hi[1] = 0.0;
dotm2.lo[0] = dotm2.lo[1] = 0.0;
for(i=0;i<n-1;i+=2)
{
LIS_QUAD_FSA2_SSE2(dotm2.hi[0],dotm2.lo[0],dotm2.hi[0],dotm2.lo[0],x[i],xl[i]);
}
LIS_QUAD_ADD_SSE2(dotm.hi[0],dotm.lo[0],dotm2.hi[0],dotm2.lo[0],dotm2.hi[1],dotm2.lo[1]);
for(;i<n;i++)
{
LIS_QUAD_FSA_SSE2(dotm.hi[0],dotm.lo[0],dotm.hi[0],dotm.lo[0],x[i],xl[i]);
}
#endif
#endif
#ifdef USE_MPI
MPI_Allreduce(dotm.hi,tmpm.hi,1,LIS_MPI_MSCALAR,LIS_MPI_MSUM,comm);
#ifndef USE_SSE2
LIS_QUAD_SQRT(val->hi[0],val->lo[0],tmpm.hi[0],tmpm.lo[0]);
#else
LIS_QUAD_SQRT_SSE2(val->hi[0],val->lo[0],tmpm.hi[0],tmpm.lo[0]);
#endif
#else
#ifndef USE_SSE2
LIS_QUAD_SQRT(val->hi[0],val->lo[0],dotm.hi[0],dotm.lo[0]);
#else
LIS_QUAD_SQRT_SSE2(val->hi[0],val->lo[0],dotm.hi[0],dotm.lo[0]);
#endif
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}