hypre_StructMatrix *
hypre_SparseMSGCreateRAPOp( hypre_StructMatrix *R,
hypre_StructMatrix *A,
hypre_StructMatrix *P,
hypre_StructGrid *coarse_grid,
HYPRE_Int cdir )
{
hypre_StructMatrix *RAP;
hypre_StructStencil *stencil;
stencil = hypre_StructMatrixStencil(A);
switch (hypre_StructStencilNDim(stencil))
{
case 2:
RAP = hypre_SparseMSG2CreateRAPOp(R ,A, P, coarse_grid, cdir);
break;
case 3:
RAP = hypre_SparseMSG3CreateRAPOp(R ,A, P, coarse_grid, cdir);
break;
}
return RAP;
}
int
hypre_StructMatrixDestroy( hypre_StructMatrix *matrix )
{
int i;
int ierr = 0;
if (matrix)
{
hypre_StructMatrixRefCount(matrix) --;
if (hypre_StructMatrixRefCount(matrix) == 0)
{
if (hypre_StructMatrixDataAlloced(matrix))
{
hypre_SharedTFree(hypre_StructMatrixData(matrix));
}
hypre_CommPkgDestroy(hypre_StructMatrixCommPkg(matrix));
hypre_ForBoxI(i, hypre_StructMatrixDataSpace(matrix))
hypre_TFree(hypre_StructMatrixDataIndices(matrix)[i]);
hypre_TFree(hypre_StructMatrixDataIndices(matrix));
hypre_BoxArrayDestroy(hypre_StructMatrixDataSpace(matrix));
hypre_TFree(hypre_StructMatrixSymmElements(matrix));
hypre_StructStencilDestroy(hypre_StructMatrixUserStencil(matrix));
hypre_StructStencilDestroy(hypre_StructMatrixStencil(matrix));
hypre_StructGridDestroy(hypre_StructMatrixGrid(matrix));
hypre_TFree(matrix);
}
}
return ierr;
}
HYPRE_Int
hypre_SMGResidualSetup( void *residual_vdata,
hypre_StructMatrix *A,
hypre_StructVector *x,
hypre_StructVector *b,
hypre_StructVector *r )
{
HYPRE_Int ierr = 0;
hypre_SMGResidualData *residual_data = residual_vdata;
hypre_IndexRef base_index = (residual_data -> base_index);
hypre_IndexRef base_stride = (residual_data -> base_stride);
hypre_StructGrid *grid;
hypre_StructStencil *stencil;
hypre_BoxArray *base_points;
hypre_ComputeInfo *compute_info;
hypre_ComputePkg *compute_pkg;
/*----------------------------------------------------------
* Set up base points and the compute package
*----------------------------------------------------------*/
grid = hypre_StructMatrixGrid(A);
stencil = hypre_StructMatrixStencil(A);
base_points = hypre_BoxArrayDuplicate(hypre_StructGridBoxes(grid));
hypre_ProjectBoxArray(base_points, base_index, base_stride);
hypre_CreateComputeInfo(grid, stencil, &compute_info);
hypre_ComputeInfoProjectComp(compute_info, base_index, base_stride);
hypre_ComputePkgCreate(compute_info, hypre_StructVectorDataSpace(x), 1,
grid, &compute_pkg);
/*----------------------------------------------------------
* Set up the residual data structure
*----------------------------------------------------------*/
(residual_data -> A) = hypre_StructMatrixRef(A);
(residual_data -> x) = hypre_StructVectorRef(x);
(residual_data -> b) = hypre_StructVectorRef(b);
(residual_data -> r) = hypre_StructVectorRef(r);
(residual_data -> base_points) = base_points;
(residual_data -> compute_pkg) = compute_pkg;
/*-----------------------------------------------------
* Compute flops
*-----------------------------------------------------*/
(residual_data -> flops) =
(hypre_StructMatrixGlobalSize(A) + hypre_StructVectorGlobalSize(x)) /
(hypre_IndexX(base_stride) *
hypre_IndexY(base_stride) *
hypre_IndexZ(base_stride) );
return ierr;
}
HYPRE_Int
hypre_SparseMSGSetupRAPOp( hypre_StructMatrix *R,
hypre_StructMatrix *A,
hypre_StructMatrix *P,
HYPRE_Int cdir,
hypre_Index cindex,
hypre_Index cstride,
hypre_Index stridePR,
hypre_StructMatrix *Ac )
{
HYPRE_Int ierr = 0;
hypre_StructStencil *stencil;
stencil = hypre_StructMatrixStencil(A);
switch (hypre_StructStencilNDim(stencil))
{
case 2:
/*--------------------------------------------------------------------
* Set lower triangular (+ diagonal) coefficients
*--------------------------------------------------------------------*/
ierr = hypre_SparseMSG2BuildRAPSym(A, P, R, cdir,
cindex, cstride, stridePR, Ac);
/*--------------------------------------------------------------------
* For non-symmetric A, set upper triangular coefficients as well
*--------------------------------------------------------------------*/
if(!hypre_StructMatrixSymmetric(A))
ierr += hypre_SparseMSG2BuildRAPNoSym(A, P, R, cdir,
cindex, cstride, stridePR, Ac);
break;
case 3:
/*--------------------------------------------------------------------
* Set lower triangular (+ diagonal) coefficients
*--------------------------------------------------------------------*/
ierr = hypre_SparseMSG3BuildRAPSym(A, P, R, cdir,
cindex, cstride, stridePR, Ac);
/*--------------------------------------------------------------------
* For non-symmetric A, set upper triangular coefficients as well
*--------------------------------------------------------------------*/
if(!hypre_StructMatrixSymmetric(A))
ierr += hypre_SparseMSG3BuildRAPNoSym(A, P, R, cdir,
cindex, cstride, stridePR, Ac);
break;
}
hypre_StructMatrixAssemble(Ac);
return ierr;
}
int
hypre_StructMatvecSetup( void *matvec_vdata,
hypre_StructMatrix *A,
hypre_StructVector *x )
{
int ierr = 0;
hypre_StructMatvecData *matvec_data = matvec_vdata;
hypre_StructGrid *grid;
hypre_StructStencil *stencil;
hypre_BoxArrayArray *send_boxes;
hypre_BoxArrayArray *recv_boxes;
int **send_processes;
int **recv_processes;
hypre_BoxArrayArray *indt_boxes;
hypre_BoxArrayArray *dept_boxes;
hypre_Index unit_stride;
hypre_ComputePkg *compute_pkg;
/*----------------------------------------------------------
* Set up the compute package
*----------------------------------------------------------*/
grid = hypre_StructMatrixGrid(A);
stencil = hypre_StructMatrixStencil(A);
hypre_CreateComputeInfo(grid, stencil,
&send_boxes, &recv_boxes,
&send_processes, &recv_processes,
&indt_boxes, &dept_boxes);
hypre_SetIndex(unit_stride, 1, 1, 1);
hypre_ComputePkgCreate(send_boxes, recv_boxes,
unit_stride, unit_stride,
send_processes, recv_processes,
indt_boxes, dept_boxes,
unit_stride,
grid, hypre_StructVectorDataSpace(x), 1,
&compute_pkg);
/*----------------------------------------------------------
* Set up the matvec data structure
*----------------------------------------------------------*/
(matvec_data -> A) = hypre_StructMatrixRef(A);
(matvec_data -> x) = hypre_StructVectorRef(x);
(matvec_data -> compute_pkg) = compute_pkg;
return ierr;
}
hypre_StructMatrix *
hypre_SMGCreateRAPOp( hypre_StructMatrix *R,
hypre_StructMatrix *A,
hypre_StructMatrix *PT,
hypre_StructGrid *coarse_grid )
{
hypre_StructMatrix *RAP;
hypre_StructStencil *stencil;
#if NEWRAP
HYPRE_Int cdir;
HYPRE_Int P_stored_as_transpose = 1;
#endif
stencil = hypre_StructMatrixStencil(A);
#if OLDRAP
switch (hypre_StructStencilDim(stencil))
{
case 2:
RAP = hypre_SMG2CreateRAPOp(R ,A, PT, coarse_grid);
break;
case 3:
RAP = hypre_SMG3CreateRAPOp(R ,A, PT, coarse_grid);
break;
}
#endif
#if NEWRAP
switch (hypre_StructStencilDim(stencil))
{
case 2:
cdir = 1;
RAP = hypre_SemiCreateRAPOp(R ,A, PT, coarse_grid, cdir,
P_stored_as_transpose);
break;
case 3:
cdir = 2;
RAP = hypre_SemiCreateRAPOp(R ,A, PT, coarse_grid, cdir,
P_stored_as_transpose);
break;
}
#endif
return RAP;
}
double *
hypre_StructMatrixExtractPointerByIndex( hypre_StructMatrix *matrix,
int b,
hypre_Index index )
{
hypre_StructStencil *stencil;
int rank;
stencil = hypre_StructMatrixStencil(matrix);
rank = hypre_StructStencilElementRank( stencil, index );
if ( rank >= 0 )
return hypre_StructMatrixBoxData(matrix, b, rank);
else
return NULL; /* error - invalid index */
}
HYPRE_Int
hypre_SparseMSGInterpSetup( void *interp_vdata,
hypre_StructMatrix *P,
hypre_StructVector *xc,
hypre_StructVector *e,
hypre_Index cindex,
hypre_Index findex,
hypre_Index stride,
hypre_Index strideP )
{
hypre_SparseMSGInterpData *interp_data = interp_vdata;
hypre_StructGrid *grid;
hypre_StructStencil *stencil;
hypre_ComputeInfo *compute_info;
hypre_ComputePkg *compute_pkg;
HYPRE_Int ierr = 0;
/*----------------------------------------------------------
* Set up the compute package
*----------------------------------------------------------*/
grid = hypre_StructVectorGrid(e);
stencil = hypre_StructMatrixStencil(P);
hypre_CreateComputeInfo(grid, stencil, &compute_info);
hypre_ComputeInfoProjectSend(compute_info, cindex, stride);
hypre_ComputeInfoProjectRecv(compute_info, cindex, stride);
hypre_ComputeInfoProjectComp(compute_info, findex, stride);
hypre_ComputePkgCreate(compute_info, hypre_StructVectorDataSpace(e), 1,
grid, &compute_pkg);
/*----------------------------------------------------------
* Set up the interp data structure
*----------------------------------------------------------*/
(interp_data -> P) = hypre_StructMatrixRef(P);
(interp_data -> compute_pkg) = compute_pkg;
hypre_CopyIndex(cindex, (interp_data -> cindex));
hypre_CopyIndex(findex, (interp_data -> findex));
hypre_CopyIndex(stride, (interp_data -> stride));
hypre_CopyIndex(strideP, (interp_data -> strideP));
return ierr;
}
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);
}
}
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_SMGResidualSetup( 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_index = (residual_data -> base_index);
hypre_IndexRef base_stride = (residual_data -> base_stride);
hypre_Index unit_stride;
hypre_StructGrid *grid;
hypre_StructStencil *stencil;
hypre_BoxArrayArray *send_boxes;
hypre_BoxArrayArray *recv_boxes;
int **send_processes;
int **recv_processes;
hypre_BoxArrayArray *indt_boxes;
hypre_BoxArrayArray *dept_boxes;
hypre_BoxArray *base_points;
hypre_ComputePkg *compute_pkg;
/*----------------------------------------------------------
* Set up base points and the compute package
*----------------------------------------------------------*/
grid = hypre_StructMatrixGrid(A);
stencil = hypre_StructMatrixStencil(A);
hypre_SetIndex(unit_stride, 1, 1, 1);
base_points = hypre_BoxArrayDuplicate(hypre_StructGridBoxes(grid));
hypre_ProjectBoxArray(base_points, base_index, base_stride);
hypre_CreateComputeInfo(grid, stencil,
&send_boxes, &recv_boxes,
&send_processes, &recv_processes,
&indt_boxes, &dept_boxes);
hypre_ProjectBoxArrayArray(indt_boxes, base_index, base_stride);
hypre_ProjectBoxArrayArray(dept_boxes, base_index, base_stride);
hypre_ComputePkgCreate(send_boxes, recv_boxes,
unit_stride, unit_stride,
send_processes, recv_processes,
indt_boxes, dept_boxes,
base_stride, grid,
hypre_StructVectorDataSpace(x), 1,
&compute_pkg);
/*----------------------------------------------------------
* Set up the residual data structure
*----------------------------------------------------------*/
(residual_data -> A) = hypre_StructMatrixRef(A);
(residual_data -> x) = hypre_StructVectorRef(x);
(residual_data -> b) = hypre_StructVectorRef(b);
(residual_data -> r) = hypre_StructVectorRef(r);
(residual_data -> base_points) = base_points;
(residual_data -> compute_pkg) = compute_pkg;
/*-----------------------------------------------------
* Compute flops
*-----------------------------------------------------*/
(residual_data -> flops) =
(hypre_StructMatrixGlobalSize(A) + hypre_StructVectorGlobalSize(x)) /
(hypre_IndexX(base_stride) *
hypre_IndexY(base_stride) *
hypre_IndexZ(base_stride) );
return ierr;
}
HYPRE_Int
hypre_SMGSetupRAPOp( hypre_StructMatrix *R,
hypre_StructMatrix *A,
hypre_StructMatrix *PT,
hypre_StructMatrix *Ac,
hypre_Index cindex,
hypre_Index cstride )
{
HYPRE_Int ierr = 0;
#if NEWRAP
HYPRE_Int cdir;
HYPRE_Int P_stored_as_transpose = 1;
#endif
hypre_StructStencil *stencil;
stencil = hypre_StructMatrixStencil(A);
#if OLDRAP
switch (hypre_StructStencilDim(stencil))
{
case 2:
/*--------------------------------------------------------------------
* Set lower triangular (+ diagonal) coefficients
*--------------------------------------------------------------------*/
ierr = hypre_SMG2BuildRAPSym(A, PT, R, Ac, cindex, cstride);
/*--------------------------------------------------------------------
* For non-symmetric A, set upper triangular coefficients as well
*--------------------------------------------------------------------*/
if(!hypre_StructMatrixSymmetric(A))
{
ierr += hypre_SMG2BuildRAPNoSym(A, PT, R, Ac, cindex, cstride);
/*-----------------------------------------------------------------
* Collapse stencil for periodic probems on coarsest grid.
*-----------------------------------------------------------------*/
ierr = hypre_SMG2RAPPeriodicNoSym(Ac, cindex, cstride);
}
else
{
/*-----------------------------------------------------------------
* Collapse stencil for periodic problems on coarsest grid.
*-----------------------------------------------------------------*/
ierr = hypre_SMG2RAPPeriodicSym(Ac, cindex, cstride);
}
break;
case 3:
/*--------------------------------------------------------------------
* Set lower triangular (+ diagonal) coefficients
*--------------------------------------------------------------------*/
ierr = hypre_SMG3BuildRAPSym(A, PT, R, Ac, cindex, cstride);
/*--------------------------------------------------------------------
* For non-symmetric A, set upper triangular coefficients as well
*--------------------------------------------------------------------*/
if(!hypre_StructMatrixSymmetric(A))
{
ierr += hypre_SMG3BuildRAPNoSym(A, PT, R, Ac, cindex, cstride);
/*-----------------------------------------------------------------
* Collapse stencil for periodic probems on coarsest grid.
*-----------------------------------------------------------------*/
ierr = hypre_SMG3RAPPeriodicNoSym(Ac, cindex, cstride);
}
else
{
/*-----------------------------------------------------------------
* Collapse stencil for periodic problems on coarsest grid.
*-----------------------------------------------------------------*/
ierr = hypre_SMG3RAPPeriodicSym(Ac, cindex, cstride);
}
break;
}
#endif
#if NEWRAP
switch (hypre_StructStencilDim(stencil))
{
case 2:
cdir = 1;
ierr = hypre_SemiBuildRAP(A, PT, R, cdir, cindex, cstride,
P_stored_as_transpose, Ac);
break;
case 3:
cdir = 2;
ierr = hypre_SemiBuildRAP(A, PT, R, cdir, cindex, cstride,
P_stored_as_transpose, Ac);
break;
}
#endif
hypre_StructMatrixAssemble(Ac);
return ierr;
}
int
hypre_SemiRestrictSetup( void *restrict_vdata,
hypre_StructMatrix *R,
int R_stored_as_transpose,
hypre_StructVector *r,
hypre_StructVector *rc,
hypre_Index cindex,
hypre_Index findex,
hypre_Index stride )
{
hypre_SemiRestrictData *restrict_data = (hypre_SemiRestrictData *)restrict_vdata;
hypre_StructGrid *grid;
hypre_StructStencil *stencil;
hypre_BoxArrayArray *send_boxes;
hypre_BoxArrayArray *recv_boxes;
int **send_processes;
int **recv_processes;
hypre_BoxArrayArray *indt_boxes;
hypre_BoxArrayArray *dept_boxes;
hypre_ComputePkg *compute_pkg;
int ierr = 0;
/*----------------------------------------------------------
* Set up the compute package
*----------------------------------------------------------*/
grid = hypre_StructVectorGrid(r);
stencil = hypre_StructMatrixStencil(R);
hypre_CreateComputeInfo(grid, stencil,
&send_boxes, &recv_boxes,
&send_processes, &recv_processes,
&indt_boxes, &dept_boxes);
hypre_ProjectBoxArrayArray(send_boxes, findex, stride);
hypre_ProjectBoxArrayArray(recv_boxes, findex, stride);
hypre_ProjectBoxArrayArray(indt_boxes, cindex, stride);
hypre_ProjectBoxArrayArray(dept_boxes, cindex, stride);
hypre_ComputePkgCreate(send_boxes, recv_boxes,
stride, stride,
send_processes, recv_processes,
indt_boxes, dept_boxes,
stride, grid,
hypre_StructVectorDataSpace(r), 1,
&compute_pkg);
/*----------------------------------------------------------
* Set up the restrict data structure
*----------------------------------------------------------*/
(restrict_data -> R) = hypre_StructMatrixRef(R);
(restrict_data -> R_stored_as_transpose) = R_stored_as_transpose;
(restrict_data -> compute_pkg) = compute_pkg;
hypre_CopyIndex(cindex ,(restrict_data -> cindex));
hypre_CopyIndex(stride ,(restrict_data -> stride));
return ierr;
}
int
hypre_SMGRelaxSetupARem( void *relax_vdata,
hypre_StructMatrix *A,
hypre_StructVector *b,
hypre_StructVector *x )
{
hypre_SMGRelaxData *relax_data = relax_vdata;
int num_spaces = (relax_data -> num_spaces);
int *space_indices = (relax_data -> space_indices);
int *space_strides = (relax_data -> space_strides);
hypre_StructVector *temp_vec = (relax_data -> temp_vec);
hypre_StructStencil *stencil = hypre_StructMatrixStencil(A);
hypre_Index *stencil_shape = hypre_StructStencilShape(stencil);
int stencil_size = hypre_StructStencilSize(stencil);
int stencil_dim = hypre_StructStencilDim(stencil);
hypre_StructMatrix *A_rem;
void **residual_data;
hypre_Index base_index;
hypre_Index base_stride;
int num_stencil_indices;
int *stencil_indices;
int i;
int ierr = 0;
/*----------------------------------------------------------
* Free up old data before putting new data into structure
*----------------------------------------------------------*/
hypre_SMGRelaxDestroyARem(relax_vdata);
/*----------------------------------------------------------
* Set up data
*----------------------------------------------------------*/
hypre_CopyIndex((relax_data -> base_index), base_index);
hypre_CopyIndex((relax_data -> base_stride), base_stride);
stencil_indices = hypre_TAlloc(int, stencil_size);
num_stencil_indices = 0;
for (i = 0; i < stencil_size; i++)
{
if (hypre_IndexD(stencil_shape[i], (stencil_dim - 1)) != 0)
{
stencil_indices[num_stencil_indices] = i;
num_stencil_indices++;
}
}
A_rem = hypre_StructMatrixCreateMask(A, num_stencil_indices, stencil_indices);
hypre_TFree(stencil_indices);
/* Set up residual_data */
residual_data = hypre_TAlloc(void *, num_spaces);
for (i = 0; i < num_spaces; i++)
{
hypre_IndexD(base_index, (stencil_dim - 1)) = space_indices[i];
hypre_IndexD(base_stride, (stencil_dim - 1)) = space_strides[i];
residual_data[i] = hypre_SMGResidualCreate();
hypre_SMGResidualSetBase(residual_data[i], base_index, base_stride);
hypre_SMGResidualSetup(residual_data[i], A_rem, x, b, temp_vec);
}
(relax_data -> A_rem) = A_rem;
(relax_data -> residual_data) = residual_data;
(relax_data -> setup_a_rem) = 0;
return ierr;
}
int
hypre_SMG2BuildRAPSym( hypre_StructMatrix *A,
hypre_StructMatrix *PT,
hypre_StructMatrix *R,
hypre_StructMatrix *RAP,
hypre_Index cindex,
hypre_Index cstride )
{
hypre_Index index;
hypre_StructStencil *fine_stencil;
int fine_stencil_size;
hypre_StructGrid *fgrid;
int *fgrid_ids;
hypre_StructGrid *cgrid;
hypre_BoxArray *cgrid_boxes;
int *cgrid_ids;
hypre_Box *cgrid_box;
hypre_IndexRef cstart;
hypre_Index stridec;
hypre_Index fstart;
hypre_IndexRef stridef;
hypre_Index loop_size;
int fi, ci;
int loopi, loopj, loopk;
hypre_Box *A_dbox;
hypre_Box *PT_dbox;
hypre_Box *R_dbox;
hypre_Box *RAP_dbox;
double *pa, *pb;
double *ra, *rb;
double *a_cc, *a_cw, *a_ce, *a_cs, *a_cn;
double *a_csw, *a_cse, *a_cnw;
double *rap_cc, *rap_cw, *rap_cs;
double *rap_csw, *rap_cse;
int iA, iAm1, iAp1;
int iAc;
int iP, iP1;
int iR;
int yOffsetA;
int xOffsetP;
int yOffsetP;
int ierr = 0;
fine_stencil = hypre_StructMatrixStencil(A);
fine_stencil_size = hypre_StructStencilSize(fine_stencil);
stridef = cstride;
hypre_SetIndex(stridec, 1, 1, 1);
fgrid = hypre_StructMatrixGrid(A);
fgrid_ids = hypre_StructGridIDs(fgrid);
cgrid = hypre_StructMatrixGrid(RAP);
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++;
}
cgrid_box = hypre_BoxArrayBox(cgrid_boxes, ci);
cstart = hypre_BoxIMin(cgrid_box);
hypre_StructMapCoarseToFine(cstart, cindex, cstride, fstart);
A_dbox = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(A), fi);
PT_dbox = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(PT), fi);
R_dbox = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(R), fi);
RAP_dbox = hypre_BoxArrayBox(hypre_StructMatrixDataSpace(RAP), ci);
/*-----------------------------------------------------------------
* Extract pointers for interpolation operator:
* pa is pointer for weight for f-point above c-point
* pb is pointer for weight for f-point below c-point
*-----------------------------------------------------------------*/
hypre_SetIndex(index,0,1,0);
pa = hypre_StructMatrixExtractPointerByIndex(PT, fi, index);
hypre_SetIndex(index,0,-1,0);
pb = hypre_StructMatrixExtractPointerByIndex(PT, fi, index);
/*-----------------------------------------------------------------
* Extract pointers for restriction operator:
* ra is pointer for weight for f-point above c-point
* rb is pointer for weight for f-point below c-point
*-----------------------------------------------------------------*/
hypre_SetIndex(index,0,1,0);
ra = hypre_StructMatrixExtractPointerByIndex(R, fi, index);
hypre_SetIndex(index,0,-1,0);
rb = hypre_StructMatrixExtractPointerByIndex(R, fi, index);
/*-----------------------------------------------------------------
* Extract pointers for 5-point fine grid operator:
*
* a_cc is pointer for center coefficient
* a_cw is pointer for west coefficient
* a_ce is pointer for east coefficient
* a_cs is pointer for south coefficient
* a_cn is pointer for north coefficient
*-----------------------------------------------------------------*/
hypre_SetIndex(index,0,0,0);
a_cc = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
hypre_SetIndex(index,-1,0,0);
a_cw = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
hypre_SetIndex(index,1,0,0);
a_ce = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
hypre_SetIndex(index,0,-1,0);
a_cs = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
hypre_SetIndex(index,0,1,0);
a_cn = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
/*-----------------------------------------------------------------
* Extract additional pointers for 9-point fine grid operator:
*
* a_csw is pointer for southwest coefficient
* a_cse is pointer for southeast coefficient
* a_cnw is pointer for northwest coefficient
* a_cne is pointer for northeast coefficient
*-----------------------------------------------------------------*/
if(fine_stencil_size > 5)
{
hypre_SetIndex(index,-1,-1,0);
a_csw = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
hypre_SetIndex(index,1,-1,0);
a_cse = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
hypre_SetIndex(index,-1,1,0);
a_cnw = hypre_StructMatrixExtractPointerByIndex(A, fi, index);
}
/*-----------------------------------------------------------------
* Extract pointers for coarse grid operator - always 9-point:
*
* We build only the lower triangular part (plus diagonal).
*
* rap_cc is pointer for center coefficient (etc.)
*-----------------------------------------------------------------*/
hypre_SetIndex(index,0,0,0);
rap_cc = hypre_StructMatrixExtractPointerByIndex(RAP, ci, index);
hypre_SetIndex(index,-1,0,0);
rap_cw = hypre_StructMatrixExtractPointerByIndex(RAP, ci, index);
hypre_SetIndex(index,0,-1,0);
rap_cs = hypre_StructMatrixExtractPointerByIndex(RAP, ci, index);
hypre_SetIndex(index,-1,-1,0);
rap_csw = hypre_StructMatrixExtractPointerByIndex(RAP, ci, index);
hypre_SetIndex(index,1,-1,0);
rap_cse = hypre_StructMatrixExtractPointerByIndex(RAP, ci, index);
/*-----------------------------------------------------------------
* Define offsets for fine grid stencil and interpolation
*
* In the BoxLoop below I assume iA and iP refer to data associated
* with the point which we are building the stencil for. The below
* Offsets are used in refering to data associated with other points.
*-----------------------------------------------------------------*/
hypre_SetIndex(index,0,1,0);
yOffsetA = hypre_BoxOffsetDistance(A_dbox,index);
yOffsetP = hypre_BoxOffsetDistance(PT_dbox,index);
hypre_SetIndex(index,1,0,0);
xOffsetP = hypre_BoxOffsetDistance(PT_dbox,index);
/*-----------------------------------------------------------------
* Switch statement to direct control to apropriate BoxLoop depending
* on stencil size. Default is full 9-point.
*-----------------------------------------------------------------*/
switch (fine_stencil_size)
{
/*--------------------------------------------------------------
* Loop for symmetric 5-point fine grid operator; produces a
* symmetric 9-point coarse grid operator. We calculate only the
* lower triangular stencil entries: (southwest, south, southeast,
* west, and center).
*--------------------------------------------------------------*/
case 5:
hypre_BoxGetSize(cgrid_box, loop_size);
hypre_BoxLoop4Begin(loop_size,
PT_dbox, cstart, stridec, iP,
R_dbox, cstart, stridec, iR,
A_dbox, fstart, stridef, iA,
RAP_dbox, cstart, stridec, iAc);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,iP,iR,iA,iAc,iAm1,iAp1,iP1
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop4For(loopi, loopj, loopk, iP, iR, iA, iAc)
{
iAm1 = iA - yOffsetA;
iAp1 = iA + yOffsetA;
iP1 = iP - yOffsetP - xOffsetP;
rap_csw[iAc] = rb[iR] * a_cw[iAm1] * pa[iP1];
iP1 = iP - yOffsetP;
rap_cs[iAc] = rb[iR] * a_cc[iAm1] * pa[iP1]
+ rb[iR] * a_cs[iAm1]
+ a_cs[iA] * pa[iP1];
iP1 = iP - yOffsetP + xOffsetP;
rap_cse[iAc] = rb[iR] * a_ce[iAm1] * pa[iP1];
iP1 = iP - xOffsetP;
rap_cw[iAc] = a_cw[iA]
+ rb[iR] * a_cw[iAm1] * pb[iP1]
+ ra[iR] * a_cw[iAp1] * pa[iP1];
rap_cc[iAc] = a_cc[iA]
+ rb[iR] * a_cc[iAm1] * pb[iP]
+ ra[iR] * a_cc[iAp1] * pa[iP]
+ rb[iR] * a_cn[iAm1]
+ ra[iR] * a_cs[iAp1]
+ a_cs[iA] * pb[iP]
+ a_cn[iA] * pa[iP];
}
hypre_BoxLoop4End(iP, iR, iA, iAc);
break;
/*--------------------------------------------------------------
* Loop for symmetric 9-point fine grid operator; produces a
* symmetric 9-point coarse grid operator. We calculate only the
* lower triangular stencil entries: (southwest, south, southeast,
* west, and center).
*--------------------------------------------------------------*/
default:
hypre_BoxGetSize(cgrid_box, loop_size);
hypre_BoxLoop4Begin(loop_size,
PT_dbox, cstart, stridec, iP,
R_dbox, cstart, stridec, iR,
A_dbox, fstart, stridef, iA,
RAP_dbox, cstart, stridec, iAc);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,iP,iR,iA,iAc,iAm1,iAp1,iP1
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop4For(loopi, loopj, loopk, iP, iR, iA, iAc)
{
iAm1 = iA - yOffsetA;
iAp1 = iA + yOffsetA;
iP1 = iP - yOffsetP - xOffsetP;
rap_csw[iAc] = rb[iR] * a_cw[iAm1] * pa[iP1]
+ rb[iR] * a_csw[iAm1]
+ a_csw[iA] * pa[iP1];
iP1 = iP - yOffsetP;
rap_cs[iAc] = rb[iR] * a_cc[iAm1] * pa[iP1]
+ rb[iR] * a_cs[iAm1]
+ a_cs[iA] * pa[iP1];
iP1 = iP - yOffsetP + xOffsetP;
rap_cse[iAc] = rb[iR] * a_ce[iAm1] * pa[iP1]
+ rb[iR] * a_cse[iAm1]
+ a_cse[iA] * pa[iP1];
iP1 = iP - xOffsetP;
rap_cw[iAc] = a_cw[iA]
+ rb[iR] * a_cw[iAm1] * pb[iP1]
+ ra[iR] * a_cw[iAp1] * pa[iP1]
+ rb[iR] * a_cnw[iAm1]
+ ra[iR] * a_csw[iAp1]
+ a_csw[iA] * pb[iP1]
+ a_cnw[iA] * pa[iP1];
rap_cc[iAc] = a_cc[iA]
+ rb[iR] * a_cc[iAm1] * pb[iP]
+ ra[iR] * a_cc[iAp1] * pa[iP]
+ rb[iR] * a_cn[iAm1]
+ ra[iR] * a_cs[iAp1]
+ a_cs[iA] * pb[iP]
+ a_cn[iA] * pa[iP];
}
hypre_BoxLoop4End(iP, iR, iA, iAc);
break;
} /* end switch statement */
HYPRE_Int
hypre_PFMGSetupRAPOp( hypre_StructMatrix *R,
hypre_StructMatrix *A,
hypre_StructMatrix *P,
HYPRE_Int cdir,
hypre_Index cindex,
hypre_Index cstride,
HYPRE_Int rap_type,
hypre_StructMatrix *Ac )
{
HYPRE_Int P_stored_as_transpose = 0;
hypre_StructStencil *stencil;
stencil = hypre_StructMatrixStencil(A);
if (rap_type == 0)
{
switch (hypre_StructStencilNDim(stencil))
{
case 2:
/*--------------------------------------------------------------------
* Set lower triangular (+ diagonal) coefficients
*--------------------------------------------------------------------*/
hypre_PFMG2BuildRAPSym(A, P, R, cdir, cindex, cstride, Ac);
/*--------------------------------------------------------------------
* For non-symmetric A, set upper triangular coefficients as well
*--------------------------------------------------------------------*/
if(!hypre_StructMatrixSymmetric(A))
hypre_PFMG2BuildRAPNoSym(A, P, R, cdir, cindex, cstride, Ac);
break;
case 3:
/*--------------------------------------------------------------------
* Set lower triangular (+ diagonal) coefficients
*--------------------------------------------------------------------*/
hypre_PFMG3BuildRAPSym(A, P, R, cdir, cindex, cstride, Ac);
/*--------------------------------------------------------------------
* For non-symmetric A, set upper triangular coefficients as well
*--------------------------------------------------------------------*/
if(!hypre_StructMatrixSymmetric(A))
hypre_PFMG3BuildRAPNoSym(A, P, R, cdir, cindex, cstride, Ac);
break;
}
}
else if (rap_type == 1)
{
switch (hypre_StructStencilNDim(stencil))
{
case 2:
hypre_PFMGBuildCoarseOp5(A, P, R, cdir, cindex, cstride, Ac);
break;
case 3:
hypre_PFMGBuildCoarseOp7(A, P, R, cdir, cindex, cstride, Ac);
break;
}
}
else if (rap_type == 2)
{
hypre_SemiBuildRAP(A, P, R, cdir, cindex, cstride,
P_stored_as_transpose, Ac);
}
hypre_StructMatrixAssemble(Ac);
return hypre_error_flag;
}
int
hypre_PointRelaxSetup( void *relax_vdata,
hypre_StructMatrix *A,
hypre_StructVector *b,
hypre_StructVector *x )
{
hypre_PointRelaxData *relax_data = (hypre_PointRelaxData *)relax_vdata;
int num_pointsets = (relax_data -> num_pointsets);
int *pointset_sizes = (relax_data -> pointset_sizes);
hypre_Index *pointset_strides = (relax_data -> pointset_strides);
hypre_Index **pointset_indices = (relax_data -> pointset_indices);
hypre_StructVector *t;
int diag_rank;
hypre_ComputePkg **compute_pkgs;
hypre_Index unit_stride;
hypre_Index diag_index;
hypre_IndexRef stride;
hypre_IndexRef index;
hypre_StructGrid *grid;
hypre_StructStencil *stencil;
hypre_BoxArrayArray *send_boxes;
hypre_BoxArrayArray *recv_boxes;
int **send_processes;
int **recv_processes;
hypre_BoxArrayArray *indt_boxes;
hypre_BoxArrayArray *dept_boxes;
hypre_BoxArrayArray *orig_indt_boxes;
hypre_BoxArrayArray *orig_dept_boxes;
hypre_BoxArrayArray *box_aa;
hypre_BoxArray *box_a;
hypre_Box *box;
int box_aa_size;
int box_a_size;
hypre_BoxArrayArray *new_box_aa;
hypre_BoxArray *new_box_a;
hypre_Box *new_box;
double scale;
int frac;
int i, j, k, p, m, compute_i;
int ierr = 0;
/*----------------------------------------------------------
* Set up the temp vector
*----------------------------------------------------------*/
if ((relax_data -> t) == NULL)
{
t = hypre_StructVectorCreate(hypre_StructVectorComm(b),
hypre_StructVectorGrid(b));
hypre_StructVectorSetNumGhost(t, hypre_StructVectorNumGhost(b));
hypre_StructVectorInitialize(t);
hypre_StructVectorAssemble(t);
(relax_data -> t) = t;
}
/*----------------------------------------------------------
* Find the matrix diagonal
*----------------------------------------------------------*/
grid = hypre_StructMatrixGrid(A);
stencil = hypre_StructMatrixStencil(A);
hypre_SetIndex(diag_index, 0, 0, 0);
diag_rank = hypre_StructStencilElementRank(stencil, diag_index);
/*----------------------------------------------------------
* Set up the compute packages
*----------------------------------------------------------*/
hypre_SetIndex(unit_stride, 1, 1, 1);
compute_pkgs = hypre_CTAlloc(hypre_ComputePkg *, num_pointsets);
for (p = 0; p < num_pointsets; p++)
{
hypre_CreateComputeInfo(grid, stencil,
&send_boxes, &recv_boxes,
&send_processes, &recv_processes,
&orig_indt_boxes, &orig_dept_boxes);
stride = pointset_strides[p];
for (compute_i = 0; compute_i < 2; compute_i++)
{
switch(compute_i)
{
case 0:
box_aa = orig_indt_boxes;
break;
case 1:
box_aa = orig_dept_boxes;
break;
}
box_aa_size = hypre_BoxArrayArraySize(box_aa);
new_box_aa = hypre_BoxArrayArrayCreate(box_aa_size);
for (i = 0; i < box_aa_size; i++)
{
box_a = hypre_BoxArrayArrayBoxArray(box_aa, i);
box_a_size = hypre_BoxArraySize(box_a);
new_box_a = hypre_BoxArrayArrayBoxArray(new_box_aa, i);
hypre_BoxArraySetSize(new_box_a, box_a_size * pointset_sizes[p]);
k = 0;
for (m = 0; m < pointset_sizes[p]; m++)
{
index = pointset_indices[p][m];
for (j = 0; j < box_a_size; j++)
{
box = hypre_BoxArrayBox(box_a, j);
new_box = hypre_BoxArrayBox(new_box_a, k);
hypre_CopyBox(box, new_box);
hypre_ProjectBox(new_box, index, stride);
k++;
}
}
}
switch(compute_i)
{
case 0:
indt_boxes = new_box_aa;
break;
case 1:
dept_boxes = new_box_aa;
break;
}
}
hypre_ComputePkgCreate(send_boxes, recv_boxes,
unit_stride, unit_stride,
send_processes, recv_processes,
indt_boxes, dept_boxes,
stride, grid,
hypre_StructVectorDataSpace(x), 1,
&compute_pkgs[p]);
hypre_BoxArrayArrayDestroy(orig_indt_boxes);
hypre_BoxArrayArrayDestroy(orig_dept_boxes);
}
/*----------------------------------------------------------
* Set up the relax data structure
*----------------------------------------------------------*/
(relax_data -> A) = hypre_StructMatrixRef(A);
(relax_data -> x) = hypre_StructVectorRef(x);
(relax_data -> b) = hypre_StructVectorRef(b);
(relax_data -> diag_rank) = diag_rank;
(relax_data -> compute_pkgs) = compute_pkgs;
/*-----------------------------------------------------
* Compute flops
*-----------------------------------------------------*/
scale = 0.0;
for (p = 0; p < num_pointsets; p++)
{
stride = pointset_strides[p];
frac = hypre_IndexX(stride);
frac *= hypre_IndexY(stride);
frac *= hypre_IndexZ(stride);
scale += (pointset_sizes[p] / frac);
}
(relax_data -> flops) = scale * (hypre_StructMatrixGlobalSize(A) +
hypre_StructVectorGlobalSize(x));
return ierr;
}
int
hypre_StructMatrixInitializeShell( hypre_StructMatrix *matrix )
{
int ierr = 0;
hypre_StructGrid *grid;
hypre_StructStencil *user_stencil;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
int stencil_size;
int num_values;
int *symm_elements;
int *num_ghost;
int extra_ghost[] = {0, 0, 0, 0, 0, 0};
hypre_BoxArray *data_space;
hypre_BoxArray *boxes;
hypre_Box *box;
hypre_Box *data_box;
int **data_indices;
int data_size;
int data_box_volume;
int i, j, d;
grid = hypre_StructMatrixGrid(matrix);
/*-----------------------------------------------------------------------
* Set up stencil and num_values:
* The stencil is a "symmetrized" version of the user's stencil
* as computed by hypre_StructStencilSymmetrize.
*
* The `symm_elements' array is used to determine what data is
* explicitely stored (symm_elements[i] < 0) and what data does is
* not explicitely stored (symm_elements[i] >= 0), but is instead
* stored as the transpose coefficient at a neighboring grid point.
*-----------------------------------------------------------------------*/
if (hypre_StructMatrixStencil(matrix) == NULL)
{
user_stencil = hypre_StructMatrixUserStencil(matrix);
hypre_StructStencilSymmetrize(user_stencil, &stencil, &symm_elements);
stencil_shape = hypre_StructStencilShape(stencil);
stencil_size = hypre_StructStencilSize(stencil);
if (!hypre_StructMatrixSymmetric(matrix))
{
/* store all element data */
for (i = 0; i < stencil_size; i++)
symm_elements[i] = -1;
num_values = stencil_size;
}
else
{
num_values = (stencil_size + 1) / 2;
}
hypre_StructMatrixStencil(matrix) = stencil;
hypre_StructMatrixSymmElements(matrix) = symm_elements;
hypre_StructMatrixNumValues(matrix) = num_values;
}
/*-----------------------------------------------------------------------
* Set ghost-layer size for symmetric storage
* - All stencil coeffs are to be available at each point in the
* grid, as well as in the user-specified ghost layer.
*-----------------------------------------------------------------------*/
num_ghost = hypre_StructMatrixNumGhost(matrix);
stencil = hypre_StructMatrixStencil(matrix);
stencil_shape = hypre_StructStencilShape(stencil);
stencil_size = hypre_StructStencilSize(stencil);
symm_elements = hypre_StructMatrixSymmElements(matrix);
for (i = 0; i < stencil_size; i++)
{
if (symm_elements[i] >= 0)
{
for (d = 0; d < 3; d++)
{
extra_ghost[2*d] =
hypre_max(extra_ghost[2*d], -hypre_IndexD(stencil_shape[i], d));
extra_ghost[2*d + 1] =
hypre_max(extra_ghost[2*d + 1], hypre_IndexD(stencil_shape[i], d));
}
}
}
for (d = 0; d < 3; d++)
{
num_ghost[2*d] += extra_ghost[2*d];
num_ghost[2*d + 1] += extra_ghost[2*d + 1];
}
/*-----------------------------------------------------------------------
* Set up data_space
*-----------------------------------------------------------------------*/
if (hypre_StructMatrixDataSpace(matrix) == NULL)
{
boxes = hypre_StructGridBoxes(grid);
data_space = hypre_BoxArrayCreate(hypre_BoxArraySize(boxes));
hypre_ForBoxI(i, boxes)
{
box = hypre_BoxArrayBox(boxes, i);
data_box = hypre_BoxArrayBox(data_space, i);
hypre_CopyBox(box, data_box);
for (d = 0; d < 3; d++)
{
hypre_BoxIMinD(data_box, d) -= num_ghost[2*d];
hypre_BoxIMaxD(data_box, d) += num_ghost[2*d + 1];
}
}
hypre_StructMatrix *
hypre_PFMGCreateRAPOp( hypre_StructMatrix *R,
hypre_StructMatrix *A,
hypre_StructMatrix *P,
hypre_StructGrid *coarse_grid,
HYPRE_Int cdir,
HYPRE_Int rap_type )
{
hypre_StructMatrix *RAP=NULL;
hypre_StructStencil *stencil;
HYPRE_Int P_stored_as_transpose = 0;
HYPRE_Int constant_coefficient;
stencil = hypre_StructMatrixStencil(A);
if (rap_type == 0)
{
switch (hypre_StructStencilNDim(stencil))
{
case 2:
RAP = hypre_PFMG2CreateRAPOp(R ,A, P, coarse_grid, cdir);
break;
case 3:
RAP = hypre_PFMG3CreateRAPOp(R ,A, P, coarse_grid, cdir);
break;
}
}
else if (rap_type == 1)
{
switch (hypre_StructStencilNDim(stencil))
{
case 2:
RAP = hypre_PFMGCreateCoarseOp5(R ,A, P, coarse_grid, cdir);
break;
case 3:
RAP = hypre_PFMGCreateCoarseOp7(R ,A, P, coarse_grid, cdir);
break;
}
}
else if (rap_type == 2)
{
RAP = hypre_SemiCreateRAPOp(R ,A, P, coarse_grid, cdir,
P_stored_as_transpose);
}
constant_coefficient = hypre_StructMatrixConstantCoefficient(A);
if ( constant_coefficient==2 && rap_type==0 )
{
/* A has variable diagonal, so, in the Galerkin case, P (and R) is
entirely variable coefficient. Thus RAP will be variable coefficient */
hypre_StructMatrixSetConstantCoefficient( RAP, 0 );
}
else
{
hypre_StructMatrixSetConstantCoefficient( RAP, constant_coefficient );
}
return RAP;
}
HYPRE_Int
hypre_SMGSolve( void *smg_vdata,
hypre_StructMatrix *A,
hypre_StructVector *b,
hypre_StructVector *x )
{
hypre_SMGData *smg_data = smg_vdata;
double tol = (smg_data -> tol);
HYPRE_Int max_iter = (smg_data -> max_iter);
HYPRE_Int rel_change = (smg_data -> rel_change);
HYPRE_Int zero_guess = (smg_data -> zero_guess);
HYPRE_Int num_levels = (smg_data -> num_levels);
HYPRE_Int num_pre_relax = (smg_data -> num_pre_relax);
HYPRE_Int num_post_relax = (smg_data -> num_post_relax);
hypre_IndexRef base_index = (smg_data -> base_index);
hypre_IndexRef base_stride = (smg_data -> base_stride);
hypre_StructMatrix **A_l = (smg_data -> A_l);
hypre_StructMatrix **PT_l = (smg_data -> PT_l);
hypre_StructMatrix **R_l = (smg_data -> R_l);
hypre_StructVector **b_l = (smg_data -> b_l);
hypre_StructVector **x_l = (smg_data -> x_l);
hypre_StructVector **r_l = (smg_data -> r_l);
hypre_StructVector **e_l = (smg_data -> e_l);
void **relax_data_l = (smg_data -> relax_data_l);
void **residual_data_l = (smg_data -> residual_data_l);
void **restrict_data_l = (smg_data -> restrict_data_l);
void **interp_data_l = (smg_data -> interp_data_l);
HYPRE_Int logging = (smg_data -> logging);
double *norms = (smg_data -> norms);
double *rel_norms = (smg_data -> rel_norms);
double b_dot_b = 0, r_dot_r, eps = 0;
double e_dot_e = 0, x_dot_x = 1;
HYPRE_Int i, l;
#if DEBUG
char filename[255];
#endif
/*-----------------------------------------------------
* Initialize some things and deal with special cases
*-----------------------------------------------------*/
hypre_BeginTiming(smg_data -> time_index);
hypre_StructMatrixDestroy(A_l[0]);
hypre_StructVectorDestroy(b_l[0]);
hypre_StructVectorDestroy(x_l[0]);
A_l[0] = hypre_StructMatrixRef(A);
b_l[0] = hypre_StructVectorRef(b);
x_l[0] = hypre_StructVectorRef(x);
(smg_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(smg_data -> time_index);
return hypre_error_flag;
}
/* part of convergence check */
if (tol > 0.0)
{
/* eps = (tol^2) */
b_dot_b = hypre_StructInnerProd(b_l[0], b_l[0]);
eps = tol*tol;
/* if rhs is zero, return a zero solution */
if (b_dot_b == 0.0)
{
hypre_StructVectorSetConstantValues(x, 0.0);
if (logging > 0)
{
norms[0] = 0.0;
rel_norms[0] = 0.0;
}
hypre_EndTiming(smg_data -> time_index);
return hypre_error_flag;
}
}
/*-----------------------------------------------------
* Do V-cycles:
* For each index l, "fine" = l, "coarse" = (l+1)
*-----------------------------------------------------*/
for (i = 0; i < max_iter; i++)
{
/*--------------------------------------------------
* Down cycle
*--------------------------------------------------*/
/* fine grid pre-relaxation */
if (num_levels > 1)
{
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[0], 0, 0);
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[0], 1, 1);
}
hypre_SMGRelaxSetMaxIter(relax_data_l[0], num_pre_relax);
hypre_SMGRelaxSetZeroGuess(relax_data_l[0], zero_guess);
hypre_SMGRelax(relax_data_l[0], A_l[0], b_l[0], x_l[0]);
zero_guess = 0;
/* compute fine grid residual (b - Ax) */
hypre_SMGResidual(residual_data_l[0], A_l[0], x_l[0], b_l[0], r_l[0]);
/* convergence check */
if (tol > 0.0)
{
r_dot_r = hypre_StructInnerProd(r_l[0], r_l[0]);
if (logging > 0)
{
norms[i] = sqrt(r_dot_r);
if (b_dot_b > 0)
rel_norms[i] = sqrt(r_dot_r/b_dot_b);
else
rel_norms[i] = 0.0;
}
/* always do at least 1 V-cycle */
if ((r_dot_r/b_dot_b < eps) && (i > 0))
{
if (rel_change)
{
if ((e_dot_e/x_dot_x) < eps)
break;
}
else
{
break;
}
}
}
if (num_levels > 1)
{
/* restrict fine grid residual */
hypre_SemiRestrict(restrict_data_l[0], R_l[0], r_l[0], b_l[1]);
#if DEBUG
if(hypre_StructStencilDim(hypre_StructMatrixStencil(A)) == 3)
{
hypre_sprintf(filename, "zout_xdown.%02d", 0);
hypre_StructVectorPrint(filename, x_l[0], 0);
hypre_sprintf(filename, "zout_rdown.%02d", 0);
hypre_StructVectorPrint(filename, r_l[0], 0);
hypre_sprintf(filename, "zout_b.%02d", 1);
hypre_StructVectorPrint(filename, b_l[1], 0);
}
#endif
for (l = 1; l <= (num_levels - 2); l++)
{
/* pre-relaxation */
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[l], 0, 0);
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[l], 1, 1);
hypre_SMGRelaxSetMaxIter(relax_data_l[l], num_pre_relax);
hypre_SMGRelaxSetZeroGuess(relax_data_l[l], 1);
hypre_SMGRelax(relax_data_l[l], A_l[l], b_l[l], x_l[l]);
/* compute residual (b - Ax) */
hypre_SMGResidual(residual_data_l[l],
A_l[l], x_l[l], b_l[l], r_l[l]);
/* restrict residual */
hypre_SemiRestrict(restrict_data_l[l], R_l[l], r_l[l], b_l[l+1]);
#if DEBUG
if(hypre_StructStencilDim(hypre_StructMatrixStencil(A)) == 3)
{
hypre_sprintf(filename, "zout_xdown.%02d", l);
hypre_StructVectorPrint(filename, x_l[l], 0);
hypre_sprintf(filename, "zout_rdown.%02d", l);
hypre_StructVectorPrint(filename, r_l[l], 0);
hypre_sprintf(filename, "zout_b.%02d", l+1);
hypre_StructVectorPrint(filename, b_l[l+1], 0);
}
#endif
}
/*--------------------------------------------------
* Bottom
*--------------------------------------------------*/
hypre_SMGRelaxSetZeroGuess(relax_data_l[l], 1);
hypre_SMGRelax(relax_data_l[l], A_l[l], b_l[l], x_l[l]);
#if DEBUG
if(hypre_StructStencilDim(hypre_StructMatrixStencil(A)) == 3)
{
hypre_sprintf(filename, "zout_xbottom.%02d", l);
hypre_StructVectorPrint(filename, x_l[l], 0);
}
#endif
/*--------------------------------------------------
* Up cycle
*--------------------------------------------------*/
for (l = (num_levels - 2); l >= 1; l--)
{
/* interpolate error and correct (x = x + Pe_c) */
hypre_SemiInterp(interp_data_l[l], PT_l[l], x_l[l+1], e_l[l]);
hypre_StructAxpy(1.0, e_l[l], x_l[l]);
#if DEBUG
if(hypre_StructStencilDim(hypre_StructMatrixStencil(A)) == 3)
{
hypre_sprintf(filename, "zout_eup.%02d", l);
hypre_StructVectorPrint(filename, e_l[l], 0);
hypre_sprintf(filename, "zout_xup.%02d", l);
hypre_StructVectorPrint(filename, x_l[l], 0);
}
#endif
/* post-relaxation */
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[l], 0, 1);
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[l], 1, 0);
hypre_SMGRelaxSetMaxIter(relax_data_l[l], num_post_relax);
hypre_SMGRelaxSetZeroGuess(relax_data_l[l], 0);
hypre_SMGRelax(relax_data_l[l], A_l[l], b_l[l], x_l[l]);
}
/* interpolate error and correct on fine grid (x = x + Pe_c) */
hypre_SemiInterp(interp_data_l[0], PT_l[0], x_l[1], e_l[0]);
hypre_SMGAxpy(1.0, e_l[0], x_l[0], base_index, base_stride);
#if DEBUG
if(hypre_StructStencilDim(hypre_StructMatrixStencil(A)) == 3)
{
hypre_sprintf(filename, "zout_eup.%02d", 0);
hypre_StructVectorPrint(filename, e_l[0], 0);
hypre_sprintf(filename, "zout_xup.%02d", 0);
hypre_StructVectorPrint(filename, x_l[0], 0);
}
#endif
}
/* part of convergence check */
if ((tol > 0.0) && (rel_change))
{
if (num_levels > 1)
{
e_dot_e = hypre_StructInnerProd(e_l[0], e_l[0]);
x_dot_x = hypre_StructInnerProd(x_l[0], x_l[0]);
}
else
{
e_dot_e = 0.0;
x_dot_x = 1.0;
}
}
/* fine grid post-relaxation */
if (num_levels > 1)
{
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[0], 0, 1);
hypre_SMGRelaxSetRegSpaceRank(relax_data_l[0], 1, 0);
}
hypre_SMGRelaxSetMaxIter(relax_data_l[0], num_post_relax);
hypre_SMGRelaxSetZeroGuess(relax_data_l[0], 0);
hypre_SMGRelax(relax_data_l[0], A_l[0], b_l[0], x_l[0]);
(smg_data -> num_iterations) = (i + 1);
}
hypre_EndTiming(smg_data -> time_index);
return hypre_error_flag;
}
int
hypre_SemiInterpSetup( void *interp_vdata,
hypre_StructMatrix *P,
int P_stored_as_transpose,
hypre_StructVector *xc,
hypre_StructVector *e,
hypre_Index cindex,
hypre_Index findex,
hypre_Index stride )
{
hypre_SemiInterpData *interp_data = interp_vdata;
hypre_StructGrid *grid;
hypre_StructStencil *stencil;
hypre_BoxArrayArray *send_boxes;
hypre_BoxArrayArray *recv_boxes;
int **send_processes;
int **recv_processes;
hypre_BoxArrayArray *indt_boxes;
hypre_BoxArrayArray *dept_boxes;
hypre_ComputePkg *compute_pkg;
int ierr = 0;
/*----------------------------------------------------------
* Set up the compute package
*----------------------------------------------------------*/
grid = hypre_StructVectorGrid(e);
stencil = hypre_StructMatrixStencil(P);
hypre_CreateComputeInfo(grid, stencil,
&send_boxes, &recv_boxes,
&send_processes, &recv_processes,
&indt_boxes, &dept_boxes);
hypre_ProjectBoxArrayArray(send_boxes, cindex, stride);
hypre_ProjectBoxArrayArray(recv_boxes, cindex, stride);
hypre_ProjectBoxArrayArray(indt_boxes, findex, stride);
hypre_ProjectBoxArrayArray(dept_boxes, findex, stride);
hypre_ComputePkgCreate(send_boxes, recv_boxes,
stride, stride,
send_processes, recv_processes,
indt_boxes, dept_boxes,
stride, grid,
hypre_StructVectorDataSpace(e), 1,
&compute_pkg);
/*----------------------------------------------------------
* Set up the interp data structure
*----------------------------------------------------------*/
(interp_data -> P) = hypre_StructMatrixRef(P);
(interp_data -> P_stored_as_transpose) = P_stored_as_transpose;
(interp_data -> compute_pkg) = compute_pkg;
hypre_CopyIndex(cindex, (interp_data -> cindex));
hypre_CopyIndex(findex, (interp_data -> findex));
hypre_CopyIndex(stride, (interp_data -> stride));
return ierr;
}
int
hypre_SMGRelaxSetupASol( void *relax_vdata,
hypre_StructMatrix *A,
hypre_StructVector *b,
hypre_StructVector *x )
{
hypre_SMGRelaxData *relax_data = relax_vdata;
int num_spaces = (relax_data -> num_spaces);
int *space_indices = (relax_data -> space_indices);
int *space_strides = (relax_data -> space_strides);
hypre_StructVector *temp_vec = (relax_data -> temp_vec);
int num_pre_relax = (relax_data -> num_pre_relax);
int num_post_relax = (relax_data -> num_post_relax);
hypre_StructStencil *stencil = hypre_StructMatrixStencil(A);
hypre_Index *stencil_shape = hypre_StructStencilShape(stencil);
int stencil_size = hypre_StructStencilSize(stencil);
int stencil_dim = hypre_StructStencilDim(stencil);
hypre_StructMatrix *A_sol;
void **solve_data;
hypre_Index base_index;
hypre_Index base_stride;
int num_stencil_indices;
int *stencil_indices;
int i;
int ierr = 0;
/*----------------------------------------------------------
* Free up old data before putting new data into structure
*----------------------------------------------------------*/
hypre_SMGRelaxDestroyASol(relax_vdata);
/*----------------------------------------------------------
* Set up data
*----------------------------------------------------------*/
hypre_CopyIndex((relax_data -> base_index), base_index);
hypre_CopyIndex((relax_data -> base_stride), base_stride);
stencil_indices = hypre_TAlloc(int, stencil_size);
num_stencil_indices = 0;
for (i = 0; i < stencil_size; i++)
{
if (hypre_IndexD(stencil_shape[i], (stencil_dim - 1)) == 0)
{
stencil_indices[num_stencil_indices] = i;
num_stencil_indices++;
}
}
A_sol = hypre_StructMatrixCreateMask(A, num_stencil_indices, stencil_indices);
hypre_StructStencilDim(hypre_StructMatrixStencil(A_sol)) = stencil_dim - 1;
hypre_TFree(stencil_indices);
/* Set up solve_data */
solve_data = hypre_TAlloc(void *, num_spaces);
for (i = 0; i < num_spaces; i++)
{
hypre_IndexD(base_index, (stencil_dim - 1)) = space_indices[i];
hypre_IndexD(base_stride, (stencil_dim - 1)) = space_strides[i];
if (stencil_dim > 2)
{
solve_data[i] = hypre_SMGCreate(relax_data -> comm);
hypre_SMGSetNumPreRelax( solve_data[i], num_pre_relax);
hypre_SMGSetNumPostRelax( solve_data[i], num_post_relax);
hypre_SMGSetBase(solve_data[i], base_index, base_stride);
hypre_SMGSetMemoryUse(solve_data[i], (relax_data -> memory_use));
hypre_SMGSetTol(solve_data[i], 0.0);
hypre_SMGSetMaxIter(solve_data[i], 1);
hypre_SMGSetup(solve_data[i], A_sol, temp_vec, x);
}
else
{
solve_data[i] = hypre_CyclicReductionCreate(relax_data -> comm);
hypre_CyclicReductionSetBase(solve_data[i], base_index, base_stride);
hypre_CyclicReductionSetup(solve_data[i], A_sol, temp_vec, x);
}
}
(relax_data -> A_sol) = A_sol;
(relax_data -> solve_data) = solve_data;
(relax_data -> setup_a_sol) = 0;
return ierr;
}
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_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);
}
}
}
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_SMGRelaxSetup( void *relax_vdata,
hypre_StructMatrix *A,
hypre_StructVector *b,
hypre_StructVector *x )
{
hypre_SMGRelaxData *relax_data = relax_vdata;
int stencil_dim;
int a_sol_test;
int ierr = 0;
stencil_dim = hypre_StructStencilDim(hypre_StructMatrixStencil(A));
(relax_data -> stencil_dim) = stencil_dim;
hypre_StructMatrixDestroy(relax_data -> A);
hypre_StructVectorDestroy(relax_data -> b);
hypre_StructVectorDestroy(relax_data -> x);
(relax_data -> A) = hypre_StructMatrixRef(A);
(relax_data -> b) = hypre_StructVectorRef(b);
(relax_data -> x) = hypre_StructVectorRef(x);
/*----------------------------------------------------------
* Set up memory according to memory_use parameter.
*
* If a subset of the solver memory is not to be set up
* until the solve is actually done, it's "setup" tag
* should have a value greater than 1.
*----------------------------------------------------------*/
if ((stencil_dim - 1) <= (relax_data -> memory_use))
{
a_sol_test = 1;
}
else
{
a_sol_test = 0;
}
/*----------------------------------------------------------
* Set up the solver
*----------------------------------------------------------*/
if ((relax_data -> setup_temp_vec) > 0)
{
ierr = hypre_SMGRelaxSetupTempVec(relax_vdata, A, b, x);
}
if ((relax_data -> setup_a_rem) > 0)
{
ierr = hypre_SMGRelaxSetupARem(relax_vdata, A, b, x);
}
if ((relax_data -> setup_a_sol) > a_sol_test)
{
ierr = hypre_SMGRelaxSetupASol(relax_vdata, A, b, x);
}
if ((relax_data -> base_box_array) == NULL)
{
ierr = hypre_SMGRelaxSetupBaseBoxArray(relax_vdata, A, b, x);
}
return ierr;
}
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
}
};
hypre_SStructPMatrix *
hypre_SysPFMGCreateRAPOp( hypre_SStructPMatrix *R,
hypre_SStructPMatrix *A,
hypre_SStructPMatrix *P,
hypre_SStructPGrid *coarse_grid,
HYPRE_Int cdir )
{
hypre_SStructPMatrix *RAP;
HYPRE_Int ndim;
HYPRE_Int nvars;
hypre_SStructVariable vartype;
hypre_SStructStencil **RAP_stencils;
hypre_StructMatrix *RAP_s;
hypre_StructMatrix *R_s;
hypre_StructMatrix *A_s;
hypre_StructMatrix *P_s;
hypre_Index **RAP_shapes;
hypre_StructStencil *sstencil;
hypre_Index *shape;
HYPRE_Int s;
HYPRE_Int *sstencil_sizes;
HYPRE_Int stencil_size;
hypre_StructGrid *cgrid;
HYPRE_Int vi,vj;
HYPRE_Int sten_cntr;
HYPRE_Int P_stored_as_transpose = 0;
ndim = hypre_StructStencilDim(hypre_SStructPMatrixSStencil(A, 0, 0));
nvars = hypre_SStructPMatrixNVars(A);
vartype = hypre_SStructPGridVarType(coarse_grid, 0);
cgrid = hypre_SStructPGridVTSGrid(coarse_grid, vartype);
RAP_stencils = hypre_CTAlloc(hypre_SStructStencil *, nvars);
RAP_shapes = hypre_CTAlloc(hypre_Index *, nvars);
sstencil_sizes = hypre_CTAlloc(HYPRE_Int, nvars);
/*--------------------------------------------------------------------------
* Symmetry within a block is exploited, but not symmetry of the form
* A_{vi,vj} = A_{vj,vi}^T.
*--------------------------------------------------------------------------*/
for (vi = 0; vi < nvars; vi++)
{
R_s = hypre_SStructPMatrixSMatrix(R, vi, vi);
stencil_size = 0;
for (vj = 0; vj < nvars; vj++)
{
A_s = hypre_SStructPMatrixSMatrix(A, vi, vj);
P_s = hypre_SStructPMatrixSMatrix(P, vj, vj);
sstencil_sizes[vj] = 0;
if (A_s != NULL)
{
RAP_s = hypre_SemiCreateRAPOp(R_s, A_s, P_s,
cgrid, cdir,
P_stored_as_transpose);
/* Just want stencil for RAP */
hypre_StructMatrixInitializeShell(RAP_s);
sstencil = hypre_StructMatrixStencil(RAP_s);
shape = hypre_StructStencilShape(sstencil);
sstencil_sizes[vj] = hypre_StructStencilSize(sstencil);
stencil_size += sstencil_sizes[vj];
RAP_shapes[vj] = hypre_CTAlloc(hypre_Index,
sstencil_sizes[vj]);
for (s = 0; s < sstencil_sizes[vj]; s++)
{
hypre_CopyIndex(shape[s],RAP_shapes[vj][s]);
}
hypre_StructMatrixDestroy(RAP_s);
}
}
HYPRE_SStructStencilCreate(ndim, stencil_size, &RAP_stencils[vi]);
sten_cntr = 0;
for (vj = 0; vj < nvars; vj++)
{
if (sstencil_sizes[vj] > 0)
{
for (s = 0; s < sstencil_sizes[vj]; s++)
{
HYPRE_SStructStencilSetEntry(RAP_stencils[vi],
sten_cntr, RAP_shapes[vj][s], vj);
sten_cntr++;
}
hypre_TFree(RAP_shapes[vj]);
}
}
}
/* create RAP Pmatrix */
hypre_SStructPMatrixCreate(hypre_SStructPMatrixComm(A),
coarse_grid, RAP_stencils, &RAP);
hypre_TFree(RAP_shapes);
hypre_TFree(sstencil_sizes);
return RAP;
}
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_StructMatrix *
hypre_StructMatrixCreateMask( hypre_StructMatrix *matrix,
int num_stencil_indices,
int *stencil_indices )
{
hypre_StructMatrix *mask;
hypre_StructStencil *stencil;
hypre_Index *stencil_shape;
int stencil_size;
hypre_Index *mask_stencil_shape;
int mask_stencil_size;
hypre_BoxArray *data_space;
int **data_indices;
int **mask_data_indices;
int i, j;
stencil = hypre_StructMatrixStencil(matrix);
stencil_shape = hypre_StructStencilShape(stencil);
stencil_size = hypre_StructStencilSize(stencil);
mask = hypre_CTAlloc(hypre_StructMatrix, 1);
hypre_StructMatrixComm(mask) = hypre_StructMatrixComm(matrix);
hypre_StructGridRef(hypre_StructMatrixGrid(matrix),
&hypre_StructMatrixGrid(mask));
hypre_StructMatrixUserStencil(mask) =
hypre_StructStencilRef(hypre_StructMatrixUserStencil(matrix));
mask_stencil_size = num_stencil_indices;
mask_stencil_shape = hypre_CTAlloc(hypre_Index, num_stencil_indices);
for (i = 0; i < num_stencil_indices; i++)
{
hypre_CopyIndex(stencil_shape[stencil_indices[i]],
mask_stencil_shape[i]);
}
hypre_StructMatrixStencil(mask) =
hypre_StructStencilCreate(hypre_StructStencilDim(stencil),
mask_stencil_size,
mask_stencil_shape);
hypre_StructMatrixNumValues(mask) = hypre_StructMatrixNumValues(matrix);
hypre_StructMatrixDataSpace(mask) =
hypre_BoxArrayDuplicate(hypre_StructMatrixDataSpace(matrix));
hypre_StructMatrixData(mask) = hypre_StructMatrixData(matrix);
hypre_StructMatrixDataAlloced(mask) = 0;
hypre_StructMatrixDataSize(mask) = hypre_StructMatrixDataSize(matrix);
data_space = hypre_StructMatrixDataSpace(matrix);
data_indices = hypre_StructMatrixDataIndices(matrix);
mask_data_indices = hypre_CTAlloc(int *, hypre_BoxArraySize(data_space));
hypre_ForBoxI(i, data_space)
{
mask_data_indices[i] = hypre_TAlloc(int, num_stencil_indices);
for (j = 0; j < num_stencil_indices; j++)
{
mask_data_indices[i][j] = data_indices[i][stencil_indices[j]];
}
}
