int
hypre_SMGAxpy( double alpha,
hypre_StructVector *x,
hypre_StructVector *y,
hypre_Index base_index,
hypre_Index base_stride )
{
int ierr = 0;
hypre_Box *x_data_box;
hypre_Box *y_data_box;
int xi;
int yi;
double *xp;
double *yp;
hypre_BoxArray *boxes;
hypre_Box *box;
hypre_Index loop_size;
hypre_IndexRef start;
int i;
int loopi, loopj, loopk;
box = hypre_BoxCreate();
boxes = hypre_StructGridBoxes(hypre_StructVectorGrid(y));
hypre_ForBoxI(i, boxes)
{
hypre_CopyBox(hypre_BoxArrayBox(boxes, i), box);
hypre_ProjectBox(box, base_index, base_stride);
start = hypre_BoxIMin(box);
x_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(x), i);
y_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(y), i);
xp = hypre_StructVectorBoxData(x, i);
yp = hypre_StructVectorBoxData(y, i);
hypre_BoxGetStrideSize(box, base_stride, loop_size);
hypre_BoxLoop2Begin(loop_size,
x_data_box, start, base_stride, xi,
y_data_box, start, base_stride, yi);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,xi,yi
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop2For(loopi, loopj, loopk, xi, yi)
{
yp[yi] += alpha * xp[xi];
}
int
hypre_SMGSetStructVectorConstantValues( hypre_StructVector *vector,
double values,
hypre_BoxArray *box_array,
hypre_Index stride )
{
int ierr = 0;
hypre_Box *v_data_box;
int vi;
double *vp;
hypre_Box *box;
hypre_Index loop_size;
hypre_IndexRef start;
int loopi, loopj, loopk;
int i;
/*-----------------------------------------------------------------------
* Set the vector coefficients
*-----------------------------------------------------------------------*/
hypre_ForBoxI(i, box_array)
{
box = hypre_BoxArrayBox(box_array, i);
start = hypre_BoxIMin(box);
v_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(vector), i);
vp = hypre_StructVectorBoxData(vector, i);
hypre_BoxGetStrideSize(box, stride, loop_size);
hypre_BoxLoop1Begin(loop_size,
v_data_box, start, stride, vi);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,vi
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop1For(loopi, loopj, loopk, vi)
{
vp[vi] = values;
}
hypre_BoxLoop1End(vi);
}
HYPRE_Int
hypre_SemiInterp( void *interp_vdata,
hypre_StructMatrix *P,
hypre_StructVector *xc,
hypre_StructVector *e )
{
hypre_SemiInterpData *interp_data = interp_vdata;
HYPRE_Int P_stored_as_transpose;
hypre_ComputePkg *compute_pkg;
hypre_IndexRef cindex;
hypre_IndexRef findex;
hypre_IndexRef stride;
hypre_StructGrid *fgrid;
HYPRE_Int *fgrid_ids;
hypre_StructGrid *cgrid;
hypre_BoxArray *cgrid_boxes;
HYPRE_Int *cgrid_ids;
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *P_dbox;
hypre_Box *xc_dbox;
hypre_Box *e_dbox;
HYPRE_Int Pi;
HYPRE_Int xci;
HYPRE_Int ei;
HYPRE_Int constant_coefficient;
HYPRE_Real *Pp0, *Pp1;
HYPRE_Real *xcp;
HYPRE_Real *ep, *ep0, *ep1;
hypre_Index loop_size;
hypre_Index start;
hypre_Index startc;
hypre_Index stridec;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
HYPRE_Int compute_i, fi, ci, j;
/*-----------------------------------------------------------------------
* Initialize some things
*-----------------------------------------------------------------------*/
hypre_BeginTiming(interp_data -> time_index);
P_stored_as_transpose = (interp_data -> P_stored_as_transpose);
compute_pkg = (interp_data -> compute_pkg);
cindex = (interp_data -> cindex);
findex = (interp_data -> findex);
stride = (interp_data -> stride);
stencil = hypre_StructMatrixStencil(P);
stencil_shape = hypre_StructStencilShape(stencil);
constant_coefficient = hypre_StructMatrixConstantCoefficient(P);
hypre_assert( constant_coefficient==0 || constant_coefficient==1 );
/* ... constant_coefficient==2 for P shouldn't happen, see
hypre_PFMGCreateInterpOp in pfmg_setup_interp.c */
if (constant_coefficient) hypre_StructVectorClearBoundGhostValues(e, 0);
hypre_SetIndex3(stridec, 1, 1, 1);
/*-----------------------------------------------------------------------
* Compute e at coarse points (injection)
*-----------------------------------------------------------------------*/
fgrid = hypre_StructVectorGrid(e);
fgrid_ids = hypre_StructGridIDs(fgrid);
cgrid = hypre_StructVectorGrid(xc);
cgrid_boxes = hypre_StructGridBoxes(cgrid);
cgrid_ids = hypre_StructGridIDs(cgrid);
fi = 0;
hypre_ForBoxI(ci, cgrid_boxes)
{
while (fgrid_ids[fi] != cgrid_ids[ci])
{
fi++;
}
compute_box = hypre_BoxArrayBox(cgrid_boxes, ci);
hypre_CopyIndex(hypre_BoxIMin(compute_box), startc);
hypre_StructMapCoarseToFine(startc, cindex, stride, start);
e_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(e), fi);
xc_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(xc), ci);
ep = hypre_StructVectorBoxData(e, fi);
xcp = hypre_StructVectorBoxData(xc, ci);
hypre_BoxGetSize(compute_box, loop_size);
hypre_BoxLoop2Begin(hypre_StructMatrixNDim(P), loop_size,
e_dbox, start, stride, ei,
xc_dbox, startc, stridec, xci);
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(HYPRE_BOX_PRIVATE,ei,xci) HYPRE_SMP_SCHEDULE
#endif
hypre_BoxLoop2For(ei, xci)
{
ep[ei] = xcp[xci];
}
hypre_BoxLoop2End(ei, xci);
}
/*-----------------------------------------------------------------------
* Compute e at fine points
*-----------------------------------------------------------------------*/
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
{
ep = hypre_StructVectorData(e);
hypre_InitializeIndtComputations(compute_pkg, ep, &comm_handle);
compute_box_aa = hypre_ComputePkgIndtBoxes(compute_pkg);
}
break;
case 1:
{
hypre_FinalizeIndtComputations(comm_handle);
compute_box_aa = hypre_ComputePkgDeptBoxes(compute_pkg);
}
break;
}
hypre_ForBoxArrayI(fi, compute_box_aa)
{
compute_box_a = hypre_BoxArrayArrayBoxArray(compute_box_aa, fi);
P_dbox = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(P), fi);
e_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(e), fi);
if (P_stored_as_transpose)
{
if ( constant_coefficient )
{
Pp0 = hypre_StructMatrixBoxData(P, fi, 1);
Pp1 = hypre_StructMatrixBoxData(P, fi, 0) -
hypre_CCBoxOffsetDistance(P_dbox, stencil_shape[0]);
}
else
{
Pp0 = hypre_StructMatrixBoxData(P, fi, 1);
Pp1 = hypre_StructMatrixBoxData(P, fi, 0) -
hypre_BoxOffsetDistance(P_dbox, stencil_shape[0]);
}
}
else
{
Pp0 = hypre_StructMatrixBoxData(P, fi, 0);
Pp1 = hypre_StructMatrixBoxData(P, fi, 1);
}
ep = hypre_StructVectorBoxData(e, fi);
ep0 = ep + hypre_BoxOffsetDistance(e_dbox, stencil_shape[0]);
ep1 = ep + hypre_BoxOffsetDistance(e_dbox, stencil_shape[1]);
hypre_ForBoxI(j, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, j);
hypre_CopyIndex(hypre_BoxIMin(compute_box), start);
hypre_StructMapFineToCoarse(start, findex, stride, startc);
hypre_BoxGetStrideSize(compute_box, stride, loop_size);
if ( constant_coefficient )
{
Pi = hypre_CCBoxIndexRank( P_dbox, startc );
hypre_BoxLoop1Begin(hypre_StructMatrixNDim(P), loop_size,
e_dbox, start, stride, ei);
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(HYPRE_BOX_PRIVATE,ei) HYPRE_SMP_SCHEDULE
#endif
hypre_BoxLoop1For(ei)
{
ep[ei] = (Pp0[Pi] * ep0[ei] +
Pp1[Pi] * ep1[ei]);
}
hypre_BoxLoop1End(ei);
}
else
{
hypre_BoxLoop2Begin(hypre_StructMatrixNDim(P), loop_size,
P_dbox, startc, stridec, Pi,
e_dbox, start, stride, ei);
#ifdef HYPRE_USING_OPENMP
#pragma omp parallel for private(HYPRE_BOX_PRIVATE,Pi,ei) HYPRE_SMP_SCHEDULE
#endif
hypre_BoxLoop2For(Pi, ei)
{
ep[ei] = (Pp0[Pi] * ep0[ei] +
Pp1[Pi] * ep1[ei]);
}
hypre_BoxLoop2End(Pi, ei);
}
}
HYPRE_Int
hypre_SparseMSGInterp( void *interp_vdata,
hypre_StructMatrix *P,
hypre_StructVector *xc,
hypre_StructVector *e )
{
HYPRE_Int ierr = 0;
hypre_SparseMSGInterpData *interp_data = interp_vdata;
hypre_ComputePkg *compute_pkg;
hypre_IndexRef cindex;
hypre_IndexRef findex;
hypre_IndexRef stride;
hypre_IndexRef strideP;
hypre_StructGrid *fgrid;
HYPRE_Int *fgrid_ids;
hypre_StructGrid *cgrid;
hypre_BoxArray *cgrid_boxes;
HYPRE_Int *cgrid_ids;
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *P_dbox;
hypre_Box *xc_dbox;
hypre_Box *e_dbox;
HYPRE_Int Pi;
HYPRE_Int xci;
HYPRE_Int ei;
double *Pp0, *Pp1;
double *xcp;
double *ep, *ep0, *ep1;
hypre_Index loop_size;
hypre_Index start;
hypre_Index startc;
hypre_Index startP;
hypre_Index stridec;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
HYPRE_Int compute_i, fi, ci, j;
HYPRE_Int loopi, loopj, loopk;
/*-----------------------------------------------------------------------
* Initialize some things
*-----------------------------------------------------------------------*/
hypre_BeginTiming(interp_data -> time_index);
compute_pkg = (interp_data -> compute_pkg);
cindex = (interp_data -> cindex);
findex = (interp_data -> findex);
stride = (interp_data -> stride);
strideP = (interp_data -> strideP);
stencil = hypre_StructMatrixStencil(P);
stencil_shape = hypre_StructStencilShape(stencil);
hypre_SetIndex(stridec, 1, 1, 1);
/*-----------------------------------------------------------------------
* Compute e at coarse points (injection)
*-----------------------------------------------------------------------*/
fgrid = hypre_StructVectorGrid(e);
fgrid_ids = hypre_StructGridIDs(fgrid);
cgrid = hypre_StructVectorGrid(xc);
cgrid_boxes = hypre_StructGridBoxes(cgrid);
cgrid_ids = hypre_StructGridIDs(cgrid);
fi = 0;
hypre_ForBoxI(ci, cgrid_boxes)
{
while (fgrid_ids[fi] != cgrid_ids[ci])
{
fi++;
}
compute_box = hypre_BoxArrayBox(cgrid_boxes, ci);
hypre_CopyIndex(hypre_BoxIMin(compute_box), startc);
hypre_StructMapCoarseToFine(startc, cindex, stride, start);
e_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(e), fi);
xc_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(xc), ci);
ep = hypre_StructVectorBoxData(e, fi);
xcp = hypre_StructVectorBoxData(xc, ci);
hypre_BoxGetSize(compute_box, loop_size);
hypre_BoxLoop2Begin(loop_size,
e_dbox, start, stride, ei,
xc_dbox, startc, stridec, xci);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,ei,xci
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop2For(loopi, loopj, loopk, ei, xci)
{
ep[ei] = xcp[xci];
}
hypre_BoxLoop2End(ei, xci);
}
/*-----------------------------------------------------------------------
* Compute e at fine points
*-----------------------------------------------------------------------*/
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
{
ep = hypre_StructVectorData(e);
hypre_InitializeIndtComputations(compute_pkg, ep, &comm_handle);
compute_box_aa = hypre_ComputePkgIndtBoxes(compute_pkg);
}
break;
case 1:
{
hypre_FinalizeIndtComputations(comm_handle);
compute_box_aa = hypre_ComputePkgDeptBoxes(compute_pkg);
}
break;
}
hypre_ForBoxArrayI(fi, compute_box_aa)
{
compute_box_a = hypre_BoxArrayArrayBoxArray(compute_box_aa, fi);
P_dbox = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(P), fi);
e_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(e), fi);
Pp0 = hypre_StructMatrixBoxData(P, fi, 0);
Pp1 = hypre_StructMatrixBoxData(P, fi, 1);
ep = hypre_StructVectorBoxData(e, fi);
ep0 = ep + hypre_BoxOffsetDistance(e_dbox, stencil_shape[0]);
ep1 = ep + hypre_BoxOffsetDistance(e_dbox, stencil_shape[1]);
hypre_ForBoxI(j, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, j);
hypre_CopyIndex(hypre_BoxIMin(compute_box), start);
hypre_StructMapFineToCoarse(start, findex, stride, startc);
hypre_StructMapCoarseToFine(startc, cindex, strideP, startP);
hypre_BoxGetStrideSize(compute_box, stride, loop_size);
hypre_BoxLoop2Begin(loop_size,
P_dbox, startP, strideP, Pi,
e_dbox, start, stride, ei);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,Pi,ei
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop2For(loopi, loopj, loopk, Pi, ei)
{
ep[ei] = (Pp0[Pi] * ep0[ei] +
Pp1[Pi] * ep1[ei]);
}
hypre_BoxLoop2End(Pi, ei);
}
}
int
hypre_SMGResidual( void *residual_vdata,
hypre_StructMatrix *A,
hypre_StructVector *x,
hypre_StructVector *b,
hypre_StructVector *r )
{
int ierr = 0;
hypre_SMGResidualData *residual_data = residual_vdata;
hypre_IndexRef base_stride = (residual_data -> base_stride);
hypre_BoxArray *base_points = (residual_data -> base_points);
hypre_ComputePkg *compute_pkg = (residual_data -> compute_pkg);
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *A_data_box;
hypre_Box *x_data_box;
hypre_Box *b_data_box;
hypre_Box *r_data_box;
int Ai;
int xi;
int bi;
int ri;
double *Ap;
double *xp;
double *bp;
double *rp;
hypre_Index loop_size;
hypre_IndexRef start;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
int stencil_size;
int compute_i, i, j, si;
int loopi, loopj, loopk;
hypre_BeginTiming(residual_data -> time_index);
/*-----------------------------------------------------------------------
* Compute residual r = b - Ax
*-----------------------------------------------------------------------*/
stencil = hypre_StructMatrixStencil(A);
stencil_shape = hypre_StructStencilShape(stencil);
stencil_size = hypre_StructStencilSize(stencil);
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
{
xp = hypre_StructVectorData(x);
hypre_InitializeIndtComputations(compute_pkg, xp, &comm_handle);
compute_box_aa = hypre_ComputePkgIndtBoxes(compute_pkg);
/*----------------------------------------
* Copy b into r
*----------------------------------------*/
compute_box_a = base_points;
hypre_ForBoxI(i, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, i);
start = hypre_BoxIMin(compute_box);
b_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(b), i);
r_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(r), i);
bp = hypre_StructVectorBoxData(b, i);
rp = hypre_StructVectorBoxData(r, i);
hypre_BoxGetStrideSize(compute_box, base_stride, loop_size);
hypre_BoxLoop2Begin(loop_size,
b_data_box, start, base_stride, bi,
r_data_box, start, base_stride, ri);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,bi,ri
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop2For(loopi, loopj, loopk, bi, ri)
{
rp[ri] = bp[bi];
}
hypre_BoxLoop2End(bi, ri);
}
}
break;
case 1:
{
hypre_FinalizeIndtComputations(comm_handle);
compute_box_aa = hypre_ComputePkgDeptBoxes(compute_pkg);
}
break;
}
/*--------------------------------------------------------------------
* Compute r -= A*x
*--------------------------------------------------------------------*/
hypre_ForBoxArrayI(i, compute_box_aa)
{
compute_box_a = hypre_BoxArrayArrayBoxArray(compute_box_aa, i);
A_data_box = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(A), i);
x_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(x), i);
r_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(r), i);
rp = hypre_StructVectorBoxData(r, i);
hypre_ForBoxI(j, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, j);
start = hypre_BoxIMin(compute_box);
for (si = 0; si < stencil_size; si++)
{
Ap = hypre_StructMatrixBoxData(A, i, si);
xp = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[si]);
hypre_BoxGetStrideSize(compute_box, base_stride,
loop_size);
hypre_BoxLoop3Begin(loop_size,
A_data_box, start, base_stride, Ai,
x_data_box, start, base_stride, xi,
r_data_box, start, base_stride, ri);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,Ai,xi,ri
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop3For(loopi, loopj, loopk, Ai, xi, ri)
{
rp[ri] -= Ap[Ai] * xp[xi];
}
hypre_BoxLoop3End(Ai, xi, ri);
}
int
hypre_PointRelax( void *relax_vdata,
hypre_StructMatrix *A,
hypre_StructVector *b,
hypre_StructVector *x )
{
hypre_PointRelaxData *relax_data = (hypre_PointRelaxData *)relax_vdata;
int max_iter = (relax_data -> max_iter);
int zero_guess = (relax_data -> zero_guess);
double weight = (relax_data -> weight);
int num_pointsets = (relax_data -> num_pointsets);
int *pointset_ranks = (relax_data -> pointset_ranks);
hypre_Index *pointset_strides = (relax_data -> pointset_strides);
hypre_StructVector *t = (relax_data -> t);
int diag_rank = (relax_data -> diag_rank);
hypre_ComputePkg **compute_pkgs = (relax_data -> compute_pkgs);
hypre_ComputePkg *compute_pkg;
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *A_data_box;
hypre_Box *b_data_box;
hypre_Box *x_data_box;
hypre_Box *t_data_box;
int Ai;
int bi;
int xi;
int ti;
double *Ap;
double *bp;
double *xp;
double *tp;
hypre_IndexRef stride;
hypre_IndexRef start;
hypre_Index loop_size;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
int stencil_size;
int iter, p, compute_i, i, j, si;
int loopi, loopj, loopk;
int pointset;
int ierr = 0;
/*----------------------------------------------------------
* Initialize some things and deal with special cases
*----------------------------------------------------------*/
hypre_BeginTiming(relax_data -> time_index);
hypre_StructMatrixDestroy(relax_data -> A);
hypre_StructVectorDestroy(relax_data -> b);
hypre_StructVectorDestroy(relax_data -> x);
(relax_data -> A) = hypre_StructMatrixRef(A);
(relax_data -> x) = hypre_StructVectorRef(x);
(relax_data -> b) = hypre_StructVectorRef(b);
(relax_data -> num_iterations) = 0;
/* if max_iter is zero, return */
if (max_iter == 0)
{
/* if using a zero initial guess, return zero */
if (zero_guess)
{
hypre_StructVectorSetConstantValues(x, 0.0);
}
hypre_EndTiming(relax_data -> time_index);
return ierr;
}
stencil = hypre_StructMatrixStencil(A);
stencil_shape = hypre_StructStencilShape(stencil);
stencil_size = hypre_StructStencilSize(stencil);
/*----------------------------------------------------------
* Do zero_guess iteration
*----------------------------------------------------------*/
p = 0;
iter = 0;
if (zero_guess)
{
pointset = pointset_ranks[p];
compute_pkg = compute_pkgs[pointset];
stride = pointset_strides[pointset];
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
{
compute_box_aa = hypre_ComputePkgIndtBoxes(compute_pkg);
}
break;
case 1:
{
compute_box_aa = hypre_ComputePkgDeptBoxes(compute_pkg);
}
break;
}
hypre_ForBoxArrayI(i, compute_box_aa)
{
compute_box_a = hypre_BoxArrayArrayBoxArray(compute_box_aa, i);
A_data_box =
hypre_BoxArrayBox(hypre_StructMatrixDataSpace(A), i);
b_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(b), i);
x_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(x), i);
Ap = hypre_StructMatrixBoxData(A, i, diag_rank);
bp = hypre_StructVectorBoxData(b, i);
xp = hypre_StructVectorBoxData(x, i);
hypre_ForBoxI(j, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, j);
start = hypre_BoxIMin(compute_box);
hypre_BoxGetStrideSize(compute_box, stride, loop_size);
hypre_BoxLoop3Begin(loop_size,
A_data_box, start, stride, Ai,
b_data_box, start, stride, bi,
x_data_box, start, stride, xi);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,Ai,bi,xi
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop3For(loopi, loopj, loopk, Ai, bi, xi)
{
xp[xi] = bp[bi] / Ap[Ai];
}
hypre_BoxLoop3End(Ai, bi, xi);
}
}
}
int
hypre_SMGResidual( void *residual_vdata,
hypre_StructMatrix *A,
hypre_StructVector *x,
hypre_StructVector *b,
hypre_StructVector *r )
{
int ierr = 0;
hypre_SMGResidualData *residual_data = (hypre_SMGResidualData *)residual_vdata;
hypre_IndexRef base_stride = (residual_data -> base_stride);
hypre_BoxArray *base_points = (residual_data -> base_points);
hypre_ComputePkg *compute_pkg = (residual_data -> compute_pkg);
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *A_data_box;
hypre_Box *x_data_box;
hypre_Box *b_data_box;
hypre_Box *r_data_box;
int Ai;
int xi;
int bi;
int ri;
double *Ap;
double *xp;
double *bp;
double *rp;
hypre_Index loop_size;
hypre_IndexRef start;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
int stencil_size;
int compute_i, i, j, si;
int loopi, loopj, loopk;
hypre_BeginTiming(residual_data -> time_index);
/*-----------------------------------------------------------------------
* Compute residual r = b - Ax
*-----------------------------------------------------------------------*/
stencil = hypre_StructMatrixStencil(A);
stencil_shape = hypre_StructStencilShape(stencil);
stencil_size = hypre_StructStencilSize(stencil);
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
{
xp = hypre_StructVectorData(x);
hypre_InitializeIndtComputations(compute_pkg, xp, &comm_handle);
compute_box_aa = hypre_ComputePkgIndtBoxes(compute_pkg);
/*----------------------------------------
* Copy b into r
*----------------------------------------*/
compute_box_a = base_points;
hypre_ForBoxI(i, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, i);
start = hypre_BoxIMin(compute_box);
b_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(b), i);
r_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(r), i);
bp = hypre_StructVectorBoxData(b, i);
rp = hypre_StructVectorBoxData(r, i);
hypre_BoxGetStrideSize(compute_box, base_stride, loop_size);
hypre_BoxLoop2Begin(loop_size,
b_data_box, start, base_stride, bi,
r_data_box, start, base_stride, ri);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,bi,ri
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop2For(loopi, loopj, loopk, bi, ri)
{
rp[ri] = bp[bi];
}
hypre_BoxLoop2End(bi, ri);
}
}
break;
case 1:
{
hypre_FinalizeIndtComputations(comm_handle);
compute_box_aa = hypre_ComputePkgDeptBoxes(compute_pkg);
}
break;
}
/*--------------------------------------------------------------------
* Compute r -= A*x
*--------------------------------------------------------------------*/
hypre_ForBoxArrayI(i, compute_box_aa)
{
compute_box_a = hypre_BoxArrayArrayBoxArray(compute_box_aa, i);
A_data_box = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(A), i);
x_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(x), i);
r_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(r), i);
rp = hypre_StructVectorBoxData(r, i);
hypre_ForBoxI(j, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, j);
start = hypre_BoxIMin(compute_box);
for (si = 0; si < stencil_size; si++)
{
Ap = hypre_StructMatrixBoxData(A, i, si);
xp = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[si]);
hypre_BoxGetStrideSize(compute_box, base_stride,
loop_size);
hypre_BoxLoop3Begin(loop_size,
A_data_box, start, base_stride, Ai,
x_data_box, start, base_stride, xi,
r_data_box, start, base_stride, ri);
#if 0
/* The following portion is preprocessed to be handled by ROSE outliner */
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,Ai,xi,ri
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop3For(loopi, loopj, loopk, Ai, xi, ri)
{
rp[ri] -= Ap[Ai] * xp[xi];
}
hypre_BoxLoop3End(Ai, xi, ri);
#else
for (hypre__block = 0; hypre__block < hypre__num_blocks;
hypre__block++)
{
loopi = 0;
loopj = 0;
loopk = 0;
hypre__nx = hypre__mx;
hypre__ny = hypre__my;
hypre__nz = hypre__mz;
if (hypre__num_blocks > 1)
{
if (hypre__dir == 0)
{
loopi =
hypre__block * hypre__div +
(((hypre__mod) <
(hypre__block)) ? (hypre__mod)
: (hypre__block));
hypre__nx =
hypre__div +
((hypre__mod > hypre__block) ? 1 : 0);
}
else if (hypre__dir == 1)
{
loopj =
hypre__block * hypre__div +
(((hypre__mod) <
(hypre__block)) ? (hypre__mod)
: (hypre__block));
hypre__ny =
hypre__div +
((hypre__mod > hypre__block) ? 1 : 0);
}
else if (hypre__dir == 2)
{
loopk =
hypre__block * hypre__div +
(((hypre__mod) <
(hypre__block)) ? (hypre__mod)
: (hypre__block));
hypre__nz =
hypre__div +
((hypre__mod > hypre__block) ? 1 : 0);
}
};
Ai =
hypre__i1start + loopi * hypre__sx1 +
loopj * hypre__sy1 + loopk * hypre__sz1;
xi =
hypre__i2start + loopi * hypre__sx2 +
loopj * hypre__sy2 + loopk * hypre__sz2;
ri =
hypre__i3start + loopi * hypre__sx3 +
loopj * hypre__sy3 + loopk * hypre__sz3;
//begin of the loop
#if 0 //
for (loopk = 0; loopk < hypre__nz; loopk++)
{
for (loopj = 0; loopj < hypre__ny; loopj++)
{
for (loopi = 0; loopi < hypre__nx; loopi++)
{
{
rp[ri] -= Ap[Ai] * xp[xi];
}
Ai += hypre__sx1;
xi += hypre__sx2;
ri += hypre__sx3;
}
Ai += hypre__sy1 - hypre__nx * hypre__sx1;
xi += hypre__sy2 - hypre__nx * hypre__sx2;
ri += hypre__sy3 - hypre__nx * hypre__sx3;
}
Ai += hypre__sz1 - hypre__ny * hypre__sy1;
xi += hypre__sz2 - hypre__ny * hypre__sy2;
ri += hypre__sz3 - hypre__ny * hypre__sy3;
} // end of the loop
#else
#if BLCR_CHECKPOINTING
// Only checkpoint it at the first occurrance.
if (g_checkpoint_flag == 0)
{
int err;
cr_checkpoint_args_t cr_args;
cr_checkpoint_handle_t cr_handle;
cr_initialize_checkpoint_args_t(&cr_args);
cr_args.cr_scope = CR_SCOPE_PROC;// a process
cr_args.cr_target = 0; //self
cr_args.cr_signal = SIGKILL; // kill after checkpointing
cr_args.cr_fd = open("dump.yy", O_WRONLY|O_CREAT|O_LARGEFILE, 0400);
if (cr_args.cr_fd < 0) {
printf("Error: cannot open file for checkpoiting context\n");
abort();
}
g_checkpoint_flag ++;
printf("Checkpoiting: starting here ..\n");
err = cr_request_checkpoint(&cr_args, &cr_handle);
if (err < 0) {
printf("cannot request checkpoining! err=%d\n",err);
abort();
}
// block until the request is served
cr_enter_cs(cr);
cr_leave_cs(cr);
printf("Checkpoiting: restarting here ..\n");
}
#endif
OUT__1__6755__(&Ai,&xi,&ri,&Ap,&xp,&rp,&loopi,&loopj,&loopk,&hypre__sx1,&hypre__sy1,&hypre__sz1,&hypre__sx2,&hypre__sy2,&hypre__sz2,&hypre__sx3,&hypre__sy3,&hypre__sz3,&hypre__nx,&hypre__ny,&hypre__nz);
#endif
}
};
int hypre_SMGResidual(void *residual_vdata,hypre_StructMatrix *A,hypre_StructVector *x,hypre_StructVector *b,hypre_StructVector *r)
{
int ierr = 0;
hypre_SMGResidualData *residual_data = residual_vdata;
hypre_IndexRef base_stride = (residual_data -> base_stride);
hypre_BoxArray *base_points = (residual_data -> base_points);
hypre_ComputePkg *compute_pkg = (residual_data -> compute_pkg);
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *A_data_box;
hypre_Box *x_data_box;
hypre_Box *b_data_box;
hypre_Box *r_data_box;
int Ai;
int xi;
int bi;
int ri;
double *Ap;
double *xp;
double *bp;
double *rp;
hypre_Index loop_size;
hypre_IndexRef start;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
int stencil_size;
int compute_i;
int i;
int j;
int si;
int loopi;
int loopj;
int loopk;
/* New static variables, precomputed */
hypre_BeginTiming((residual_data -> time_index));
stencil = (A -> stencil);
stencil_shape = (stencil -> shape);
stencil_size = (stencil -> size);
(stencil_size <= 15)?0 : ((__assert_fail(("stencil_size <= 15"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(203),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
for (compute_i = 0; compute_i < 2; compute_i++) {
switch(compute_i){
case 0:
{
{
xp = (x -> data);
hypre_InitializeIndtComputations(compute_pkg,xp,&comm_handle);
compute_box_aa = (compute_pkg -> indt_boxes);
compute_box_a = base_points;
for (i = 0; i < (compute_box_a -> size); i++) {
compute_box = ((compute_box_a -> boxes) + i);
start = (compute_box -> imin);
b_data_box = ((( *(b -> data_space)).boxes) + i);
r_data_box = ((( *(r -> data_space)).boxes) + i);
bp = ((b -> data) + ((b -> data_indices)[i]));
rp = ((r -> data) + ((r -> data_indices)[i]));
hypre_BoxGetStrideSize(compute_box,base_stride,loop_size);
{
int hypre__i1start = (((start[0]) - ((b_data_box -> imin)[0])) + ((((start[1]) - ((b_data_box -> imin)[1])) + (((start[2]) - ((b_data_box -> imin)[2])) * (((0 < ((((b_data_box -> imax)[1]) - ((b_data_box -> imin)[1])) + 1))?((((b_data_box -> imax)[1]) - ((b_data_box -> imin)[1])) + 1) : 0)))) * (((0 < ((((b_data_box -> imax)[0]) - ((b_data_box -> imin)[0])) + 1))?((((b_data_box -> imax)[0]) - ((b_data_box -> imin)[0])) + 1) : 0))));
int hypre__i2start = (((start[0]) - ((r_data_box -> imin)[0])) + ((((start[1]) - ((r_data_box -> imin)[1])) + (((start[2]) - ((r_data_box -> imin)[2])) * (((0 < ((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1))?((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1) : 0)))) * (((0 < ((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1))?((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1) : 0))));
int hypre__sx1 = (base_stride[0]);
int hypre__sy1 = ((base_stride[1]) * (((0 < ((((b_data_box -> imax)[0]) - ((b_data_box -> imin)[0])) + 1))?((((b_data_box -> imax)[0]) - ((b_data_box -> imin)[0])) + 1) : 0)));
int hypre__sz1 = (((base_stride[2]) * (((0 < ((((b_data_box -> imax)[0]) - ((b_data_box -> imin)[0])) + 1))?((((b_data_box -> imax)[0]) - ((b_data_box -> imin)[0])) + 1) : 0))) * (((0 < ((((b_data_box -> imax)[1]) - ((b_data_box -> imin)[1])) + 1))?((((b_data_box -> imax)[1]) - ((b_data_box -> imin)[1])) + 1) : 0)));
int hypre__sx2 = (base_stride[0]);
int hypre__sy2 = ((base_stride[1]) * (((0 < ((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1))?((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1) : 0)));
int hypre__sz2 = (((base_stride[2]) * (((0 < ((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1))?((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1) : 0))) * (((0 < ((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1))?((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1) : 0)));
int hypre__nx = (loop_size[0]);
int hypre__ny = (loop_size[1]);
int hypre__nz = (loop_size[2]);
int hypre__mx = hypre__nx;
int hypre__my = hypre__ny;
int hypre__mz = hypre__nz;
int hypre__dir;
int hypre__max;
int hypre__div;
int hypre__mod;
int hypre__block;
int hypre__num_blocks;
hypre__dir = 0;
hypre__max = hypre__nx;
if (hypre__ny > hypre__max) {
hypre__dir = 1;
hypre__max = hypre__ny;
}
if (hypre__nz > hypre__max) {
hypre__dir = 2;
hypre__max = hypre__nz;
}
hypre__num_blocks = 1;
if (hypre__max < hypre__num_blocks) {
hypre__num_blocks = hypre__max;
}
if (hypre__num_blocks > 0) {
hypre__div = (hypre__max / hypre__num_blocks);
hypre__mod = (hypre__max % hypre__num_blocks);
}
/* # 236 "smg_residual.c" */
(hypre__sx1 == 1)?0 : ((__assert_fail(("hypre__sx1 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(357),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
(hypre__sx2 == 1)?0 : ((__assert_fail(("hypre__sx2 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(358),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
if (hypre__num_blocks == 1) {
int ii;
int jj;
int kk;
const double *bp_0 = (bp + hypre__i1start);
double *rp_0 = (rp + hypre__i2start);
for (kk = 0; kk < hypre__mz; kk++) {
for (jj = 0; jj < hypre__my; jj++) {
const double *bpp = ((bp_0 + (jj * hypre__sy1)) + (kk * hypre__sz1));
double *rpp = ((rp_0 + (jj * hypre__sy2)) + (kk * hypre__sz2));
for (ii = 0; ii < hypre__mx; ii++) {
rpp[ii] = (bpp[ii]);
}
}
}
/* hypre__num_blocks > 1 */
}
else {
for (hypre__block = 0; hypre__block < hypre__num_blocks; hypre__block++) {
loopi = 0;
loopj = 0;
loopk = 0;
hypre__nx = hypre__mx;
hypre__ny = hypre__my;
hypre__nz = hypre__mz;
if (hypre__dir == 0) {
loopi = ((hypre__block * hypre__div) + (((hypre__mod < hypre__block)?hypre__mod : hypre__block)));
hypre__nx = (hypre__div + (((hypre__mod > hypre__block)?1 : 0)));
}
else if (hypre__dir == 1) {
loopj = ((hypre__block * hypre__div) + (((hypre__mod < hypre__block)?hypre__mod : hypre__block)));
hypre__ny = (hypre__div + (((hypre__mod > hypre__block)?1 : 0)));
}
else if (hypre__dir == 2) {
loopk = ((hypre__block * hypre__div) + (((hypre__mod < hypre__block)?hypre__mod : hypre__block)));
hypre__nz = (hypre__div + (((hypre__mod > hypre__block)?1 : 0)));
}
bi = (((hypre__i1start + loopi) + (loopj * hypre__sy1)) + (loopk * hypre__sz1));
ri = (((hypre__i2start + loopi) + (loopj * hypre__sy2)) + (loopk * hypre__sz2));
/* AAA */
{
int ii;
int jj;
int kk;
const double *bp_0 = (bp + bi);
double *rp_0 = (rp + ri);
for (kk = 0; kk < hypre__nz; kk++) {
for (jj = 0; jj < hypre__ny; jj++) {
const double *bpp = ((bp_0 + (jj * hypre__sy1)) + (kk * hypre__sz1));
double *rpp = ((rp_0 + (jj * hypre__sy2)) + (kk * hypre__sz2));
for (ii = 0; ii < hypre__nx; ii++) {
rpp[ii] = (bpp[ii]);
}
}
}
/* AAA */
}
}
/* hypre__num_blocks > 1 */
}
}
}
}
break;
}
case 1:
{
{
hypre_FinalizeIndtComputations(comm_handle);
compute_box_aa = (compute_pkg -> dept_boxes);
}
break;
}
/* switch */
}
/*--------------------------------------------------------------------
* Compute r -= A*x
*--------------------------------------------------------------------*/
for (i = 0; i < (compute_box_aa -> size); i++) {
int dxp_s[15UL];
compute_box_a = ((compute_box_aa -> box_arrays)[i]);
A_data_box = ((( *(A -> data_space)).boxes) + i);
x_data_box = ((( *(x -> data_space)).boxes) + i);
r_data_box = ((( *(r -> data_space)).boxes) + i);
rp = ((r -> data) + ((r -> data_indices)[i]));
for (si = 0; si < stencil_size; si++) {
dxp_s[si] = (((stencil_shape[si])[0]) + ((((stencil_shape[si])[1]) + (((stencil_shape[si])[2]) * (((0 < ((((x_data_box -> imax)[1]) - ((x_data_box -> imin)[1])) + 1))?((((x_data_box -> imax)[1]) - ((x_data_box -> imin)[1])) + 1) : 0)))) * (((0 < ((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1))?((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1) : 0))));
}
for (j = 0; j < (compute_box_a -> size); j++) {{
int hypre__i1start;
int hypre__i2start;
int hypre__i3start;
int hypre__sx1;
int hypre__sy1;
int hypre__sz1;
int hypre__sx2;
int hypre__sy2;
int hypre__sz2;
int hypre__sx3;
int hypre__sy3;
int hypre__sz3;
int hypre__nx;
int hypre__ny;
int hypre__nz;
int hypre__mx;
int hypre__my;
int hypre__mz;
int hypre__dir;
int hypre__max;
int hypre__div;
int hypre__mod;
int hypre__block;
int hypre__num_blocks;
compute_box = ((compute_box_a -> boxes) + j);
start = (compute_box -> imin);
hypre__i1start = (((start[0]) - ((A_data_box -> imin)[0])) + ((((start[1]) - ((A_data_box -> imin)[1])) + (((start[2]) - ((A_data_box -> imin)[2])) * (((0 < ((((A_data_box -> imax)[1]) - ((A_data_box -> imin)[1])) + 1))?((((A_data_box -> imax)[1]) - ((A_data_box -> imin)[1])) + 1) : 0)))) * (((0 < ((((A_data_box -> imax)[0]) - ((A_data_box -> imin)[0])) + 1))?((((A_data_box -> imax)[0]) - ((A_data_box -> imin)[0])) + 1) : 0))));
hypre__i2start = (((start[0]) - ((x_data_box -> imin)[0])) + ((((start[1]) - ((x_data_box -> imin)[1])) + (((start[2]) - ((x_data_box -> imin)[2])) * (((0 < ((((x_data_box -> imax)[1]) - ((x_data_box -> imin)[1])) + 1))?((((x_data_box -> imax)[1]) - ((x_data_box -> imin)[1])) + 1) : 0)))) * (((0 < ((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1))?((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1) : 0))));
hypre__i3start = (((start[0]) - ((r_data_box -> imin)[0])) + ((((start[1]) - ((r_data_box -> imin)[1])) + (((start[2]) - ((r_data_box -> imin)[2])) * (((0 < ((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1))?((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1) : 0)))) * (((0 < ((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1))?((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1) : 0))));
hypre_BoxGetStrideSize(compute_box,base_stride,loop_size);
hypre__sx1 = (base_stride[0]);
hypre__sy1 = ((base_stride[1]) * (((0 < ((((A_data_box -> imax)[0]) - ((A_data_box -> imin)[0])) + 1))?((((A_data_box -> imax)[0]) - ((A_data_box -> imin)[0])) + 1) : 0)));
hypre__sz1 = (((base_stride[2]) * (((0 < ((((A_data_box -> imax)[0]) - ((A_data_box -> imin)[0])) + 1))?((((A_data_box -> imax)[0]) - ((A_data_box -> imin)[0])) + 1) : 0))) * (((0 < ((((A_data_box -> imax)[1]) - ((A_data_box -> imin)[1])) + 1))?((((A_data_box -> imax)[1]) - ((A_data_box -> imin)[1])) + 1) : 0)));
hypre__sx2 = (base_stride[0]);
hypre__sy2 = ((base_stride[1]) * (((0 < ((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1))?((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1) : 0)));
hypre__sz2 = (((base_stride[2]) * (((0 < ((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1))?((((x_data_box -> imax)[0]) - ((x_data_box -> imin)[0])) + 1) : 0))) * (((0 < ((((x_data_box -> imax)[1]) - ((x_data_box -> imin)[1])) + 1))?((((x_data_box -> imax)[1]) - ((x_data_box -> imin)[1])) + 1) : 0)));
hypre__sx3 = (base_stride[0]);
hypre__sy3 = ((base_stride[1]) * (((0 < ((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1))?((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1) : 0)));
hypre__sz3 = (((base_stride[2]) * (((0 < ((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1))?((((r_data_box -> imax)[0]) - ((r_data_box -> imin)[0])) + 1) : 0))) * (((0 < ((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1))?((((r_data_box -> imax)[1]) - ((r_data_box -> imin)[1])) + 1) : 0)));
/* Based on BG/L Milestone #46 */
(hypre__sx1 == 1)?0 : ((__assert_fail(("hypre__sx1 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(602),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
(hypre__sx2 == 1)?0 : ((__assert_fail(("hypre__sx2 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(603),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
(hypre__sx3 == 1)?0 : ((__assert_fail(("hypre__sx3 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(604),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
hypre__mx = (loop_size[0]);
hypre__my = (loop_size[1]);
hypre__mz = (loop_size[2]);
hypre__dir = 0;
hypre__max = hypre__mx;
if (hypre__my > hypre__max) {
hypre__dir = 1;
hypre__max = hypre__my;
}
if (hypre__mz > hypre__max) {
hypre__dir = 2;
hypre__max = hypre__mz;
}
hypre__num_blocks = 1;
if (hypre__max < hypre__num_blocks) {
hypre__num_blocks = hypre__max;
}
if (hypre__num_blocks > 0) {
hypre__div = (hypre__max / hypre__num_blocks);
hypre__mod = (hypre__max % hypre__num_blocks);
}
else
continue;
if (hypre__num_blocks == 1) {
int si;
int ii;
int jj;
int kk;
const double *Ap_0 = ((A -> data) + hypre__i1start);
const double *xp_0 = (((x -> data) + hypre__i2start) + ((x -> data_indices)[i]));
ri = hypre__i3start;
void *__out_argv1__1527__[21];
*(__out_argv1__1527__ + 0) = ((void *)(&xp_0));
*(__out_argv1__1527__ + 1) = ((void *)(&Ap_0));
*(__out_argv1__1527__ + 2) = ((void *)(&kk));
*(__out_argv1__1527__ + 3) = ((void *)(&jj));
*(__out_argv1__1527__ + 4) = ((void *)(&ii));
*(__out_argv1__1527__ + 5) = ((void *)(&si));
*(__out_argv1__1527__ + 6) = ((void *)(&hypre__mz));
*(__out_argv1__1527__ + 7) = ((void *)(&hypre__my));
*(__out_argv1__1527__ + 8) = ((void *)(&hypre__mx));
*(__out_argv1__1527__ + 9) = ((void *)(&hypre__sz3));
*(__out_argv1__1527__ + 10) = ((void *)(&hypre__sy3));
*(__out_argv1__1527__ + 11) = ((void *)(&hypre__sz2));
*(__out_argv1__1527__ + 12) = ((void *)(&hypre__sy2));
*(__out_argv1__1527__ + 13) = ((void *)(&hypre__sz1));
*(__out_argv1__1527__ + 14) = ((void *)(&hypre__sy1));
*(__out_argv1__1527__ + 15) = ((void *)(&dxp_s));
*(__out_argv1__1527__ + 16) = ((void *)(&i));
*(__out_argv1__1527__ + 17) = ((void *)(&stencil_size));
*(__out_argv1__1527__ + 18) = ((void *)(&rp));
*(__out_argv1__1527__ + 19) = ((void *)(&ri));
*(__out_argv1__1527__ + 20) = ((void *)(&A));
#ifdef USE_DLOPEN
if (g_execution_flag == 0){
printf("Opening the .so file ...\n");
FunctionLib = dlopen("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/OUT__1__6119__.so",RTLD_LAZY);
dlError = dlerror();
if( dlError ) {
printf("cannot open .so file!\n");
exit(1);
}
/* Find the first loaded function */
OUT__1__6119__ = dlsym( FunctionLib, "OUT__1__6119__");
dlError = dlerror();
if( dlError )
{
printf("cannot find OUT__1__6755__() !\n");
exit(1);
}
//remove("/tmp/peri.result");
//time1=time_stamp();
} // end if (flag ==0)
g_execution_flag ++;
(*OUT__1__6119__)(__out_argv1__1527__);
#else
OUT__1__6119__(__out_argv1__1527__);
#endif
/* hypre__num_blocks > 1 */
}
else {
for (si = 0; si < stencil_size; si++) {
Ap = ((A -> data) + (((A -> data_indices)[i])[si]));
xp = (((x -> data) + ((x -> data_indices)[i])) + (dxp_s[si]));
for (hypre__block = 0; hypre__block < hypre__num_blocks; hypre__block++) {
loopi = 0;
loopj = 0;
loopk = 0;
hypre__nx = hypre__mx;
hypre__ny = hypre__my;
hypre__nz = hypre__mz;
if (hypre__dir == 0) {
loopi = ((hypre__block * hypre__div) + (((hypre__mod < hypre__block)?hypre__mod : hypre__block)));
hypre__nx = (hypre__div + (((hypre__mod > hypre__block)?1 : 0)));
}
else if (hypre__dir == 1) {
loopj = ((hypre__block * hypre__div) + (((hypre__mod < hypre__block)?hypre__mod : hypre__block)));
hypre__ny = (hypre__div + (((hypre__mod > hypre__block)?1 : 0)));
}
else if (hypre__dir == 2) {
loopk = ((hypre__block * hypre__div) + (((hypre__mod < hypre__block)?hypre__mod : hypre__block)));
hypre__nz = (hypre__div + (((hypre__mod > hypre__block)?1 : 0)));
}
Ai = (((hypre__i1start + (loopi * hypre__sx1)) + (loopj * hypre__sy1)) + (loopk * hypre__sz1));
xi = (((hypre__i2start + (loopi * hypre__sx2)) + (loopj * hypre__sy2)) + (loopk * hypre__sz2));
ri = (((hypre__i3start + (loopi * hypre__sx3)) + (loopj * hypre__sy3)) + (loopk * hypre__sz3));
/* CORE LOOP BEGIN */
(hypre__sx1 == 1)?0 : ((__assert_fail(("hypre__sx1 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(689),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
(hypre__sx2 == 1)?0 : ((__assert_fail(("hypre__sx2 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(690),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
(hypre__sx3 == 1)?0 : ((__assert_fail(("hypre__sx3 == 1"),("/home/liao6/svnrepos/benchmarks/smg2000/struct_ls/smg_residual.c"),(691),("int hypre_SMGResidual(void *, struct hypre_StructMatrix_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *, struct hypre_StructVector_struct *)")) , 0));
{
/* In essence, this loop computes:
*
FOR_ALL i, j, k DO
rp[ri + i + j*DJ_R + k*DK_R]
-= Ap[Ai + i + j*DJ_A + k*DK_A]
* xp[xi + i + j*DJ_X + k*DK_X];
*/
// 1. promoting loop invariant expressions
// j loop increment for Ai, xi, and ri
int DJA0 = (hypre__sy1 - (hypre__nx * hypre__sx1));
int DJX0 = (hypre__sy2 - (hypre__nx * hypre__sx2));
int DJR0 = (hypre__sy3 - (hypre__nx * hypre__sx3));
// k loop increment for Ai, xi, and ri
int DKA0 = (hypre__sz1 - (hypre__ny * hypre__sy1));
int DKX0 = (hypre__sz2 - (hypre__ny * hypre__sy2));
int DKR0 = (hypre__sz3 - (hypre__ny * hypre__sy3));
// pre-compute array index offset changes for one iteration within each level of loop
// one iteration of j loop on ri
int DJR1 = (DJR0 + (hypre__nx * hypre__sx3));
// one iteration of k loop on ri
int DKR1 = (DKR0 + (hypre__ny * DJR1));
// one iteration of j loop on Ai
int DJA1 = (DJA0 + (hypre__nx * hypre__sx1));
// one iteration of k loop on Ai
int DKA1 = (DKA0 + (hypre__ny * DJA1));
// one iteration of j loop on xi
int DJX1 = (DJX0 + (hypre__nx * hypre__sx2));
// one iteration of k loop on xi
int DKX1 = (DKX0 + (hypre__ny * DJX1));
for (loopk = 0; loopk < hypre__nz; loopk++) {
for (loopj = 0; loopj < hypre__ny; loopj++) {
for (loopi = 0; loopi < hypre__nx; loopi++) {{
rp[((ri + (loopi * hypre__sx1)) + (loopj * DJR1)) + (loopk * DKR1)] -= ((Ap[((Ai + (loopi * hypre__sx1)) + (loopj * DJA1)) + (loopk * DKA1)]) * (xp[((xi + (loopi * hypre__sx2)) + (loopj * DJX1)) + (loopk * DKX1)]));
//rp[ri] -= Ap[Ai] * xp[xi];
}
//Ai += hypre__sx1; // 2. merging loop index changes
//xi += hypre__sx2;
//ri += hypre__sx3;
}
//Ai += DJA0;//(hypre__sy1 - (hypre__nx * hypre__sx1));
//xi += DJX0;//(hypre__sy2 - (hypre__nx * hypre__sx2));
//ri += DJR0;//(hypre__sy3 - (hypre__nx * hypre__sx3));
}
//Ai += DKA0;//(hypre__sz1 - (hypre__ny * hypre__sy1));
//xi += DKX0; //(hypre__sz2 - (hypre__ny * hypre__sy2));
//ri += DKR0;//(hypre__sz3 - (hypre__ny * hypre__sy3));
}
}
/* CORE LOOP END */
/* hypre__block */
}
/* si */
}
/* else hypre__num_blocks > 1 */
}
/* j */
}
}
/* i */
}
/* compute_i */
}
hypre_IncFLOPCount((residual_data -> flops));
hypre_EndTiming((residual_data -> time_index));
return ierr;
}
HYPRE_Int
hypre_SMGResidual( void *residual_vdata,
hypre_StructMatrix *A,
hypre_StructVector *x,
hypre_StructVector *b,
hypre_StructVector *r )
{
HYPRE_Int ierr;
hypre_SMGResidualData *residual_data = residual_vdata;
hypre_IndexRef base_stride = (residual_data -> base_stride);
hypre_BoxArray *base_points = (residual_data -> base_points);
hypre_ComputePkg *compute_pkg = (residual_data -> compute_pkg);
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *A_data_box;
hypre_Box *x_data_box;
hypre_Box *b_data_box;
hypre_Box *r_data_box;
HYPRE_Int Ai;
HYPRE_Int xi;
HYPRE_Int bi;
HYPRE_Int ri;
double *Ap0;
double *xp0;
double *bp;
double *rp;
hypre_Index loop_size;
hypre_IndexRef start;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
HYPRE_Int stencil_size;
HYPRE_Int compute_i, i, j, si;
HYPRE_Int loopi, loopj, loopk;
double *Ap1, *Ap2;
double *Ap3, *Ap4;
double *Ap5, *Ap6;
double *Ap7, *Ap8, *Ap9;
double *Ap10, *Ap11, *Ap12, *Ap13, *Ap14;
double *Ap15, *Ap16, *Ap17, *Ap18;
double *Ap19, *Ap20, *Ap21, *Ap22, *Ap23, *Ap24, *Ap25, *Ap26;
double *xp1, *xp2;
double *xp3, *xp4;
double *xp5, *xp6;
double *xp7, *xp8, *xp9;
double *xp10, *xp11, *xp12, *xp13, *xp14;
double *xp15, *xp16, *xp17, *xp18;
double *xp19, *xp20, *xp21, *xp22, *xp23, *xp24, *xp25, *xp26;
hypre_BeginTiming(residual_data -> time_index);
/*-----------------------------------------------------------------------
* Compute residual r = b - Ax
*-----------------------------------------------------------------------*/
stencil = hypre_StructMatrixStencil(A);
stencil_shape = hypre_StructStencilShape(stencil);
stencil_size = hypre_StructStencilSize(stencil);
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
{
xp0 = hypre_StructVectorData(x);
hypre_InitializeIndtComputations(compute_pkg, xp0, &comm_handle);
compute_box_aa = hypre_ComputePkgIndtBoxes(compute_pkg);
/*----------------------------------------
* Copy b into r
*----------------------------------------*/
compute_box_a = base_points;
hypre_ForBoxI(i, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, i);
start = hypre_BoxIMin(compute_box);
b_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(b), i);
r_data_box =
hypre_BoxArrayBox(hypre_StructVectorDataSpace(r), i);
bp = hypre_StructVectorBoxData(b, i);
rp = hypre_StructVectorBoxData(r, i);
hypre_BoxGetStrideSize(compute_box, base_stride, loop_size);
hypre_BoxLoop2Begin(loop_size,
b_data_box, start, base_stride, bi,
r_data_box, start, base_stride, ri);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,bi,ri
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop2For(loopi, loopj, loopk, bi, ri)
{
rp[ri] = bp[bi];
}
hypre_BoxLoop2End(bi, ri);
}
}
break;
case 1:
{
hypre_FinalizeIndtComputations(comm_handle);
compute_box_aa = hypre_ComputePkgDeptBoxes(compute_pkg);
}
break;
}
/*--------------------------------------------------------------------
* Compute r -= A*x
*--------------------------------------------------------------------*/
hypre_ForBoxArrayI(i, compute_box_aa)
{
compute_box_a = hypre_BoxArrayArrayBoxArray(compute_box_aa, i);
A_data_box = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(A), i);
x_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(x), i);
r_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(r), i);
rp = hypre_StructVectorBoxData(r, i);
/*--------------------------------------------------------------
* Switch statement to direct control (based on stencil size) to
* code to get pointers and offsets fo A and x.
*--------------------------------------------------------------*/
switch (stencil_size)
{
case 1:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
break;
case 3:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
Ap1 = hypre_StructMatrixBoxData(A, i, 1);
Ap2 = hypre_StructMatrixBoxData(A, i, 2);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
xp1 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[1]);
xp2 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[2]);
break;
case 5:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
Ap1 = hypre_StructMatrixBoxData(A, i, 1);
Ap2 = hypre_StructMatrixBoxData(A, i, 2);
Ap3 = hypre_StructMatrixBoxData(A, i, 3);
Ap4 = hypre_StructMatrixBoxData(A, i, 4);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
xp1 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[1]);
xp2 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[2]);
xp3 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[3]);
xp4 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[4]);
break;
case 7:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
Ap1 = hypre_StructMatrixBoxData(A, i, 1);
Ap2 = hypre_StructMatrixBoxData(A, i, 2);
Ap3 = hypre_StructMatrixBoxData(A, i, 3);
Ap4 = hypre_StructMatrixBoxData(A, i, 4);
Ap5 = hypre_StructMatrixBoxData(A, i, 5);
Ap6 = hypre_StructMatrixBoxData(A, i, 6);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
xp1 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[1]);
xp2 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[2]);
xp3 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[3]);
xp4 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[4]);
xp5 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[5]);
xp6 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[6]);
break;
case 9:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
Ap1 = hypre_StructMatrixBoxData(A, i, 1);
Ap2 = hypre_StructMatrixBoxData(A, i, 2);
Ap3 = hypre_StructMatrixBoxData(A, i, 3);
Ap4 = hypre_StructMatrixBoxData(A, i, 4);
Ap5 = hypre_StructMatrixBoxData(A, i, 5);
Ap6 = hypre_StructMatrixBoxData(A, i, 6);
Ap7 = hypre_StructMatrixBoxData(A, i, 7);
Ap8 = hypre_StructMatrixBoxData(A, i, 8);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
xp1 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[1]);
xp2 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[2]);
xp3 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[3]);
xp4 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[4]);
xp5 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[5]);
xp6 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[6]);
xp7 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[7]);
xp8 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[8]);
break;
case 15:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
Ap1 = hypre_StructMatrixBoxData(A, i, 1);
Ap2 = hypre_StructMatrixBoxData(A, i, 2);
Ap3 = hypre_StructMatrixBoxData(A, i, 3);
Ap4 = hypre_StructMatrixBoxData(A, i, 4);
Ap5 = hypre_StructMatrixBoxData(A, i, 5);
Ap6 = hypre_StructMatrixBoxData(A, i, 6);
Ap7 = hypre_StructMatrixBoxData(A, i, 7);
Ap8 = hypre_StructMatrixBoxData(A, i, 8);
Ap9 = hypre_StructMatrixBoxData(A, i, 9);
Ap10 = hypre_StructMatrixBoxData(A, i, 10);
Ap11 = hypre_StructMatrixBoxData(A, i, 11);
Ap12 = hypre_StructMatrixBoxData(A, i, 12);
Ap13 = hypre_StructMatrixBoxData(A, i, 13);
Ap14 = hypre_StructMatrixBoxData(A, i, 14);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
xp1 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[1]);
xp2 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[2]);
xp3 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[3]);
xp4 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[4]);
xp5 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[5]);
xp6 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[6]);
xp7 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[7]);
xp8 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[8]);
xp9 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[9]);
xp10 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[10]);
xp11 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[11]);
xp12 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[12]);
xp13 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[13]);
xp14 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[14]);
break;
case 19:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
Ap1 = hypre_StructMatrixBoxData(A, i, 1);
Ap2 = hypre_StructMatrixBoxData(A, i, 2);
Ap3 = hypre_StructMatrixBoxData(A, i, 3);
Ap4 = hypre_StructMatrixBoxData(A, i, 4);
Ap5 = hypre_StructMatrixBoxData(A, i, 5);
Ap6 = hypre_StructMatrixBoxData(A, i, 6);
Ap7 = hypre_StructMatrixBoxData(A, i, 7);
Ap8 = hypre_StructMatrixBoxData(A, i, 8);
Ap9 = hypre_StructMatrixBoxData(A, i, 9);
Ap10 = hypre_StructMatrixBoxData(A, i, 10);
Ap11 = hypre_StructMatrixBoxData(A, i, 11);
Ap12 = hypre_StructMatrixBoxData(A, i, 12);
Ap13 = hypre_StructMatrixBoxData(A, i, 13);
Ap14 = hypre_StructMatrixBoxData(A, i, 14);
Ap15 = hypre_StructMatrixBoxData(A, i, 15);
Ap16 = hypre_StructMatrixBoxData(A, i, 16);
Ap17 = hypre_StructMatrixBoxData(A, i, 17);
Ap18 = hypre_StructMatrixBoxData(A, i, 18);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
xp1 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[1]);
xp2 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[2]);
xp3 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[3]);
xp4 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[4]);
xp5 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[5]);
xp6 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[6]);
xp7 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[7]);
xp8 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[8]);
xp9 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[9]);
xp10 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[10]);
xp11 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[11]);
xp12 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[12]);
xp13 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[13]);
xp14 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[14]);
xp15 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[15]);
xp16 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[16]);
xp17 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[17]);
xp18 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[18]);
break;
case 27:
Ap0 = hypre_StructMatrixBoxData(A, i, 0);
Ap1 = hypre_StructMatrixBoxData(A, i, 1);
Ap2 = hypre_StructMatrixBoxData(A, i, 2);
Ap3 = hypre_StructMatrixBoxData(A, i, 3);
Ap4 = hypre_StructMatrixBoxData(A, i, 4);
Ap5 = hypre_StructMatrixBoxData(A, i, 5);
Ap6 = hypre_StructMatrixBoxData(A, i, 6);
Ap7 = hypre_StructMatrixBoxData(A, i, 7);
Ap8 = hypre_StructMatrixBoxData(A, i, 8);
Ap9 = hypre_StructMatrixBoxData(A, i, 9);
Ap10 = hypre_StructMatrixBoxData(A, i, 10);
Ap11 = hypre_StructMatrixBoxData(A, i, 11);
Ap12 = hypre_StructMatrixBoxData(A, i, 12);
Ap13 = hypre_StructMatrixBoxData(A, i, 13);
Ap14 = hypre_StructMatrixBoxData(A, i, 14);
Ap15 = hypre_StructMatrixBoxData(A, i, 15);
Ap16 = hypre_StructMatrixBoxData(A, i, 16);
Ap17 = hypre_StructMatrixBoxData(A, i, 17);
Ap18 = hypre_StructMatrixBoxData(A, i, 18);
Ap19 = hypre_StructMatrixBoxData(A, i, 19);
Ap20 = hypre_StructMatrixBoxData(A, i, 20);
Ap21 = hypre_StructMatrixBoxData(A, i, 21);
Ap22 = hypre_StructMatrixBoxData(A, i, 22);
Ap23 = hypre_StructMatrixBoxData(A, i, 23);
Ap24 = hypre_StructMatrixBoxData(A, i, 24);
Ap25 = hypre_StructMatrixBoxData(A, i, 25);
Ap26 = hypre_StructMatrixBoxData(A, i, 26);
xp0 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[0]);
xp1 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[1]);
xp2 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[2]);
xp3 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[3]);
xp4 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[4]);
xp5 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[5]);
xp6 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[6]);
xp7 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[7]);
xp8 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[8]);
xp9 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[9]);
xp10 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[10]);
xp11 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[11]);
xp12 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[12]);
xp13 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[13]);
xp14 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[14]);
xp15 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[15]);
xp16 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[16]);
xp17 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[17]);
xp18 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[18]);
xp19 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[19]);
xp20 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[20]);
xp21 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[21]);
xp22 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[22]);
xp23 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[23]);
xp24 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[24]);
xp25 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[25]);
xp26 = hypre_StructVectorBoxData(x, i) +
hypre_BoxOffsetDistance(x_data_box, stencil_shape[26]);
break;
default:
;
}
hypre_ForBoxI(j, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, j);
start = hypre_BoxIMin(compute_box);
/*------------------------------------------------------
* Switch statement to direct control to appropriate
* box loop depending on stencil size
*------------------------------------------------------*/
switch (stencil_size)
{
case 1:
hypre_BoxGetStrideSize(compute_box, base_stride, loop_size);
hypre_BoxLoop3Begin(loop_size,
A_data_box, start, base_stride, Ai,
x_data_box, start, base_stride, xi,
r_data_box, start, base_stride, ri);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,Ai,xi,ri
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop3For(loopi, loopj, loopk, Ai, xi, ri)
{
rp[ri] = rp[ri]
- Ap0[Ai] * xp0[xi];
}
hypre_BoxLoop3End(Ai, xi, ri);
break;
case 3:
hypre_BoxGetStrideSize(compute_box, base_stride, loop_size);
hypre_BoxLoop3Begin(loop_size,
A_data_box, start, base_stride, Ai,
x_data_box, start, base_stride, xi,
r_data_box, start, base_stride, ri);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,Ai,xi,ri
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop3For(loopi, loopj, loopk, Ai, xi, ri)
{
rp[ri] = rp[ri]
- Ap0[Ai] * xp0[xi]
- Ap1[Ai] * xp1[xi]
- Ap2[Ai] * xp2[xi];
}
int
hypre_SemiRestrict( void *restrict_vdata,
hypre_StructMatrix *R,
hypre_StructVector *r,
hypre_StructVector *rc )
{
int ierr = 0;
hypre_SemiRestrictData *restrict_data = (hypre_SemiRestrictData *)restrict_vdata;
int R_stored_as_transpose;
hypre_ComputePkg *compute_pkg;
hypre_IndexRef cindex;
hypre_IndexRef stride;
hypre_StructGrid *fgrid;
int *fgrid_ids;
hypre_StructGrid *cgrid;
hypre_BoxArray *cgrid_boxes;
int *cgrid_ids;
hypre_CommHandle *comm_handle;
hypre_BoxArrayArray *compute_box_aa;
hypre_BoxArray *compute_box_a;
hypre_Box *compute_box;
hypre_Box *R_dbox;
hypre_Box *r_dbox;
hypre_Box *rc_dbox;
int Ri;
int ri;
int rci;
double *Rp0, *Rp1;
double *rp, *rp0, *rp1;
double *rcp;
hypre_Index loop_size;
hypre_IndexRef start;
hypre_Index startc;
hypre_Index stridec;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
int compute_i, fi, ci, j;
int loopi, loopj, loopk;
/*-----------------------------------------------------------------------
* Initialize some things.
*-----------------------------------------------------------------------*/
hypre_BeginTiming(restrict_data -> time_index);
R_stored_as_transpose = (restrict_data -> R_stored_as_transpose);
compute_pkg = (restrict_data -> compute_pkg);
cindex = (restrict_data -> cindex);
stride = (restrict_data -> stride);
stencil = hypre_StructMatrixStencil(R);
stencil_shape = hypre_StructStencilShape(stencil);
hypre_SetIndex(stridec, 1, 1, 1);
/*--------------------------------------------------------------------
* Restrict the residual.
*--------------------------------------------------------------------*/
fgrid = hypre_StructVectorGrid(r);
fgrid_ids = hypre_StructGridIDs(fgrid);
cgrid = hypre_StructVectorGrid(rc);
cgrid_boxes = hypre_StructGridBoxes(cgrid);
cgrid_ids = hypre_StructGridIDs(cgrid);
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
{
rp = hypre_StructVectorData(r);
hypre_InitializeIndtComputations(compute_pkg, rp, &comm_handle);
compute_box_aa = hypre_ComputePkgIndtBoxes(compute_pkg);
}
break;
case 1:
{
hypre_FinalizeIndtComputations(comm_handle);
compute_box_aa = hypre_ComputePkgDeptBoxes(compute_pkg);
}
break;
}
fi = 0;
hypre_ForBoxArrayI(ci, cgrid_boxes)
{
while (fgrid_ids[fi] != cgrid_ids[ci])
{
fi++;
}
compute_box_a = hypre_BoxArrayArrayBoxArray(compute_box_aa, fi);
R_dbox = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(R), fi);
r_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(r), fi);
rc_dbox = hypre_BoxArrayBox(hypre_StructVectorDataSpace(rc), ci);
if (R_stored_as_transpose)
{
Rp0 = hypre_StructMatrixBoxData(R, fi, 1) -
hypre_BoxOffsetDistance(R_dbox, stencil_shape[1]);
Rp1 = hypre_StructMatrixBoxData(R, fi, 0);
}
else
{
Rp0 = hypre_StructMatrixBoxData(R, fi, 0);
Rp1 = hypre_StructMatrixBoxData(R, fi, 1);
}
rp = hypre_StructVectorBoxData(r, fi);
rp0 = rp + hypre_BoxOffsetDistance(r_dbox, stencil_shape[0]);
rp1 = rp + hypre_BoxOffsetDistance(r_dbox, stencil_shape[1]);
rcp = hypre_StructVectorBoxData(rc, ci);
hypre_ForBoxI(j, compute_box_a)
{
compute_box = hypre_BoxArrayBox(compute_box_a, j);
start = hypre_BoxIMin(compute_box);
hypre_StructMapFineToCoarse(start, cindex, stride, startc);
hypre_BoxGetStrideSize(compute_box, stride, loop_size);
hypre_BoxLoop3Begin(loop_size,
R_dbox, startc, stridec, Ri,
r_dbox, start, stride, ri,
rc_dbox, startc, stridec, rci);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,Ri,ri,rci
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop3For(loopi, loopj, loopk, Ri, ri, rci)
{
rcp[rci] = rp[ri] + (Rp0[Ri] * rp0[ri] +
Rp1[Ri] * rp1[ri]);
}
hypre_BoxLoop3End(Ri, ri, rci);
}
}