int
hypre_SMGRelaxSetupTempVec( void *relax_vdata,
hypre_StructMatrix *A,
hypre_StructVector *b,
hypre_StructVector *x )
{
hypre_SMGRelaxData *relax_data = relax_vdata;
hypre_StructVector *temp_vec = (relax_data -> temp_vec);
int ierr = 0;
/*----------------------------------------------------------
* Set up data
*----------------------------------------------------------*/
if ((relax_data -> temp_vec) == NULL)
{
temp_vec = hypre_StructVectorCreate(hypre_StructVectorComm(b),
hypre_StructVectorGrid(b));
hypre_StructVectorSetNumGhost(temp_vec, hypre_StructVectorNumGhost(b));
hypre_StructVectorInitialize(temp_vec);
hypre_StructVectorAssemble(temp_vec);
(relax_data -> temp_vec) = temp_vec;
}
(relax_data -> setup_temp_vec) = 0;
return ierr;
}
void *
hypre_StructKrylovCreateVector( void *vvector )
{
hypre_StructVector *vector = vvector;
hypre_StructVector *new_vector;
new_vector = hypre_StructVectorCreate( hypre_StructVectorComm(vector),
hypre_StructVectorGrid(vector) );
hypre_StructVectorInitialize(new_vector);
hypre_StructVectorAssemble(new_vector);
return ( (void *) new_vector );
}
void *
hypre_StructKrylovCreateVector( void *vvector )
{
hypre_StructVector *vector = vvector;
hypre_StructVector *new_vector;
HYPRE_Int *num_ghost= hypre_StructVectorNumGhost(vector);
new_vector = hypre_StructVectorCreate( hypre_StructVectorComm(vector),
hypre_StructVectorGrid(vector) );
hypre_StructVectorSetNumGhost(new_vector, num_ghost);
hypre_StructVectorInitialize(new_vector);
hypre_StructVectorAssemble(new_vector);
return ( (void *) new_vector );
}
void *
hypre_StructKrylovCreateVectorArray(int n, void *vvector )
{
hypre_StructVector *vector = vvector;
hypre_StructVector **new_vector;
int i;
new_vector = hypre_CTAlloc(hypre_StructVector*,n);
for (i=0; i < n; i++)
{
HYPRE_StructVectorCreate(hypre_StructVectorComm(vector),
hypre_StructVectorGrid(vector),
(HYPRE_StructVector *) &new_vector[i] );
HYPRE_StructVectorInitialize((HYPRE_StructVector) new_vector[i]);
HYPRE_StructVectorAssemble((HYPRE_StructVector) new_vector[i]);
}
return ( (void *) new_vector );
}
hypre_StructVector *
hypre_StructVectorCreate( MPI_Comm comm,
hypre_StructGrid *grid )
{
hypre_StructVector *vector;
int i;
vector = hypre_CTAlloc(hypre_StructVector, 1);
hypre_StructVectorComm(vector) = comm;
hypre_StructGridRef(grid, &hypre_StructVectorGrid(vector));
hypre_StructVectorDataAlloced(vector) = 1;
hypre_StructVectorRefCount(vector) = 1;
/* set defaults */
for (i = 0; i < 6; i++)
hypre_StructVectorNumGhost(vector)[i] = 1;
return vector;
}
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;
}
double
hypre_StructOverlapInnerProd( hypre_StructVector *x,
hypre_StructVector *y )
{
double final_innerprod_result;
double local_result, overlap_result;
double process_result;
hypre_Box *x_data_box;
hypre_Box *y_data_box;
hypre_BoxArray *overlap_boxes;
HYPRE_Int xi;
HYPRE_Int yi;
double *xp;
double *yp;
hypre_BoxArray *boxes;
hypre_Box *boxi, *boxj, intersect_box;
hypre_StructGrid *grid= hypre_StructVectorGrid(y);
hypre_BoxManager *boxman = hypre_StructGridBoxMan(grid);
hypre_BoxArray *neighbor_boxes;
HYPRE_Int *neighbors_procs= NULL;
hypre_BoxArray *selected_nboxes;
hypre_BoxArray *tmp_box_array, *tmp2_box_array;
hypre_Index loop_size;
hypre_IndexRef start;
hypre_Index unit_stride;
HYPRE_Int i, j;
HYPRE_Int myid;
HYPRE_Int boxarray_size;
HYPRE_Int loopi, loopj, loopk;
#ifdef HYPRE_USE_PTHREADS
HYPRE_Int threadid = hypre_GetThreadID();
#endif
local_result = 0.0;
process_result = 0.0;
hypre_SetIndex(unit_stride, 1, 1, 1);
hypre_MPI_Comm_rank(hypre_StructVectorComm(y), &myid);
/*-----------------------------------------------------------------------
* Determine the overlapped boxes on this local processor.
*-----------------------------------------------------------------------*/
boxes = hypre_StructGridBoxes(hypre_StructVectorGrid(y));
boxarray_size= hypre_BoxArraySize(boxes);
/*-----------------------------------------------------------------------
* To compute the inner product over this local processor, given a box,
* the inner product between x & y is computed over the whole box and
* over any overlapping between this box and overlap_boxes. The latter
* result is subtracted from the former. Overlapping between more than
* two boxes are handled.
*-----------------------------------------------------------------------*/
hypre_ForBoxI(i, boxes)
{
boxi = hypre_BoxArrayBox(boxes, i);
start = hypre_BoxIMin(boxi);
x_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(x), i);
y_data_box = hypre_BoxArrayBox(hypre_StructVectorDataSpace(y), i);
xp = hypre_StructVectorBoxData(x, i);
yp = hypre_StructVectorBoxData(y, i);
hypre_BoxGetSize(boxi, loop_size);
#ifdef HYPRE_USE_PTHREADS
local_result_ref[threadid] = &local_result;
#endif
hypre_BoxLoop2Begin(loop_size,
x_data_box, start, unit_stride, xi,
y_data_box, start, unit_stride, yi);
#define HYPRE_BOX_SMP_PRIVATE loopk,loopi,loopj,xi,yi
#define HYPRE_SMP_REDUCTION_OP +
#define HYPRE_SMP_REDUCTION_VARS local_result
#include "hypre_box_smp_forloop.h"
hypre_BoxLoop2For(loopi, loopj, loopk, xi, yi)
{
local_result += xp[xi] * yp[yi];
}