int lis_quad_sqrt(LIS_QUAD *a, const LIS_QUAD *b)
{
LIS_QUAD_DECLAR;
#ifndef USE_SSE2
LIS_QUAD_SQRT(a->hi,a->lo,b->hi,b->lo);
#else
LIS_QUAD_SQRT_SSE2(a->hi,a->lo,b->hi,b->lo);
#endif
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;
}
