/*BHEADER********************************************************************** * Copyright (c) 2008, Lawrence Livermore National Security, LLC. * Produced at the Lawrence Livermore National Laboratory. * This file is part of HYPRE. See file COPYRIGHT for details. * * HYPRE is free software; you can redistribute it and/or modify it under the * terms of the GNU Lesser General Public License (as published by the Free * Software Foundation) version 2.1 dated February 1999. * * $Revision: 2.4 $ ***********************************************************************EHEADER*/ /****************************************************************************** * * Structured inner product routine for overlapped grids. Computes the * inner product of two vectors over an overlapped grid. * *****************************************************************************/ #include "headers.h" /*-------------------------------------------------------------------------- * hypre_StructOverlapInnerProd *--------------------------------------------------------------------------*/ #ifdef HYPRE_USE_PTHREADS double *local_result_ref[hypre_MAX_THREADS]; #endif double final_innerprod_result_overlap; double hypre_StructOverlapInnerProd( hypre_StructVector *x, hypre_StructVector *y ) { double local_result, overlap_result; double process_result; hypre_Box *x_data_box; hypre_Box *y_data_box; hypre_BoxArray *overlap_boxes; int xi; int yi; double *xp; double *yp; hypre_BoxArray *boxes; hypre_Box *boxi, *boxj, intersect_box; hypre_BoxNeighbors *neighbors= hypre_StructGridNeighbors(hypre_StructVectorGrid(y)); int *neighbors_procs= hypre_BoxNeighborsProcs(neighbors); hypre_BoxArray *selected_nboxes; hypre_BoxArray *tmp_box_array, *tmp2_box_array; hypre_Index loop_size; hypre_IndexRef start; hypre_Index unit_stride; int i, j; int myid; int boxarray_size; int loopi, loopj, loopk; #ifdef HYPRE_USE_PTHREADS int threadid = hypre_GetThreadID(); #endif local_result = 0.0; process_result = 0.0; hypre_SetIndex(unit_stride, 1, 1, 1); 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); hypre_BoxLoop2For(loopi, loopj, loopk, xi, yi) { local_result += xp[xi] * yp[yi]; } hypre_BoxLoop2End(xi, yi); /*-------------------------------------------------------------------- * intersect all boxes from (i+1) to boxarray_size *--------------------------------------------------------------------*/ overlap_boxes= hypre_BoxArrayCreate(0); for (j= (i+1); j< boxarray_size; j++) { boxj= hypre_BoxArrayBox(boxes, j); hypre_IntersectBoxes(boxi, boxj, &intersect_box); if (hypre_BoxVolume(&intersect_box)) { hypre_AppendBox(&intersect_box, overlap_boxes); } } if (hypre_BoxArraySize(overlap_boxes) > 1) { hypre_UnionBoxes(overlap_boxes); } /*-------------------------------------------------------------------- * compute inner product over overlap *--------------------------------------------------------------------*/ if (hypre_BoxArraySize(overlap_boxes)) { overlap_result= 0.0; hypre_ForBoxI(j, overlap_boxes) { boxj = hypre_BoxArrayBox(overlap_boxes, j); start = hypre_BoxIMin(boxj); hypre_BoxGetSize(boxj, loop_size); hypre_BoxLoop2Begin(loop_size, x_data_box, start, unit_stride, xi, y_data_box, start, unit_stride, yi); hypre_BoxLoop2For(loopi, loopj, loopk, xi, yi) { overlap_result += xp[xi] * yp[yi]; } hypre_BoxLoop2End(xi, yi); } local_result-= overlap_result; } hypre_BoxArrayDestroy(overlap_boxes); } /*----------------------------------------------------------------------- * Determine the across processor overlap. The inner product is computed * and subtracted on processors that share the overlap except the one * with the lowest processor id. Therefore, on this processor, we need * to subtract only overlaps with boxes on processors with id < myid. *-----------------------------------------------------------------------*/ boxes= hypre_BoxNeighborsBoxes(neighbors); selected_nboxes= hypre_BoxArrayCreate(0); /* extract only boxes on processors with ids < myid. */ hypre_ForBoxI(i, boxes) { if (neighbors_procs[i] < myid) { hypre_AppendBox(hypre_BoxArrayBox(boxes, i), selected_nboxes); } } boxes= hypre_StructGridBoxes(hypre_StructVectorGrid(y)); overlap_boxes= hypre_BoxArrayCreate(0); hypre_ForBoxI(i, boxes) { boxi= hypre_BoxArrayBox(boxes, i); hypre_ForBoxI(j, selected_nboxes) { boxj= hypre_BoxArrayBox(selected_nboxes, j); hypre_IntersectBoxes(boxi, boxj, &intersect_box); if (hypre_BoxVolume(&intersect_box)) { hypre_AppendBox(&intersect_box, overlap_boxes); } } } hypre_BoxArrayDestroy(selected_nboxes); /*----------------------------------------------------------------------- * Union the overlap_boxes and then begin to compute and subtract chunks * and norms. *-----------------------------------------------------------------------*/ if (hypre_BoxArraySize(overlap_boxes) > 1) { hypre_UnionBoxes(overlap_boxes); } if (hypre_BoxArraySize(overlap_boxes)) { overlap_result= 0.0; hypre_ForBoxI(i, boxes) { boxi = hypre_BoxArrayBox(boxes, i); 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); tmp_box_array= hypre_BoxArrayCreate(0); hypre_ForBoxI(j, overlap_boxes) { boxj= hypre_BoxArrayBox(overlap_boxes, j); hypre_IntersectBoxes(boxi, boxj, &intersect_box); if (hypre_BoxVolume(&intersect_box)) { start= hypre_BoxIMin(&intersect_box); hypre_BoxGetSize(&intersect_box, loop_size); hypre_BoxLoop2Begin(loop_size, x_data_box, start, unit_stride, xi, y_data_box, start, unit_stride, yi); hypre_BoxLoop2For(loopi, loopj, loopk, xi, yi) { overlap_result += xp[xi] * yp[yi]; } hypre_BoxLoop2End(xi, yi); hypre_AppendBox(&intersect_box, tmp_box_array); } /* if (hypre_BoxVolume(&intersect_box)) */ } /* hypre_ForBoxI(j, overlap_boxes) */ /*------------------------------------------------------------------------- * Subtract the intersection boxes so that norm on higher degree overlaps * on this processor are computed only once. *-------------------------------------------------------------------------*/ tmp2_box_array= hypre_BoxArrayCreate(0); hypre_SubtractBoxArrays(overlap_boxes, tmp_box_array, tmp2_box_array); hypre_BoxArrayDestroy(tmp_box_array); hypre_BoxArrayDestroy(tmp2_box_array); } /* hypre_ForBoxI(i, boxes) */ local_result-= overlap_result; } /* if (hypre_BoxArraySize(overlap_boxes)) */ hypre_BoxArrayDestroy(overlap_boxes); #ifdef HYPRE_USE_PTHREADS if (threadid != hypre_NumThreads) { for (i = 0; i < hypre_NumThreads; i++) process_result += *local_result_ref[i]; } else process_result = *local_result_ref[threadid]; #else process_result = local_result; #endif MPI_Allreduce(&process_result, &final_innerprod_result_overlap, 1, MPI_DOUBLE, MPI_SUM, hypre_StructVectorComm(x)); #ifdef HYPRE_USE_PTHREADS if (threadid == 0 || threadid == hypre_NumThreads) #endif hypre_IncFLOPCount(2*hypre_StructVectorGlobalSize(x)); return final_innerprod_result_overlap; }