private static void inertial1d(vtx_data **graph, /* graph data structure */
int nvtxs, /* number of vtxs in graph */
bool cube_or_mesh, /* 0 => hypercube, d => d-dimensional mesh */
int nsets, /* number of sets to divide into */
float *x, /* x coordinates of vertices */
int *sets, /* set each vertex gets assigned to */
double[] goal, /* desired set sizes */
bool using_vwgts /* are vertex weights being used? */
)
{
double *value; /* values passed to median routine */
double time; /* timing variables */
int * space; /* space required by median routine */
int i; /* loop counter */
value = (double *)Marshal.AllocHGlobal((nvtxs + 1) * sizeof(double));
/* Copy values into double precision array. */
for (i = 1; i <= nvtxs; i++)
{
value[i] = x[i];
}
/* Now find the median value and partition based upon it. */
space = (int *)Marshal.AllocHGlobal(nvtxs * sizeof(int));
time = seconds();
rec_median_1(graph, value, nvtxs, space, cube_or_mesh, nsets, goal, using_vwgts, sets, true);
median_time += seconds() - time;
Marshal.FreeHGlobal((IntPtr)space);
Marshal.FreeHGlobal((IntPtr)value);
}
private static bool SameStructure(int node1, int node2, /* two vertices which might have same nonzeros */
vtx_data **graph, /* data structure for storing graph */
int *scatter /* array for checking vertex labelling */
)
{
bool same; /* are two vertices indistinguisable? */
int i; /* loop counter */
scatter[node1] = node1;
for (i = 1; i < graph[node1]->nedges; i++)
{
scatter[graph[node1]->edges[i]] = node1;
}
for (i = 1; i < graph[node2]->nedges; i++)
{
if (scatter[graph[node2]->edges[i]] != node1)
{
break;
}
}
same = (i == graph[node2]->nedges && scatter[node2] == node1);
return(same);
}
/* Check an extended eigenpair of A by direct multiplication. Uses
* the Ay = extval*y + Dg form of the problem for convenience. */
public static double checkeig_ext(double *err, double *work, /* work vector of length n */
vtx_data **A, double *y, int n, double extval, double *vwsqrt,
double *gvec, double eigtol,
bool warnings /* don't want to see warning messages in one of the
* contexts this is called */
)
{
double resid; /* the extended eigen residual */
splarax(err, A, n, y, vwsqrt, work);
scadd(err, 1, n, -extval, y);
cpvec(work, 1, n, gvec); /* only need if going to re-use gvec */
scale_diag(work, 1, n, vwsqrt);
scadd(err, 1, n, -1.0, work);
resid = ch_norm(err, 1, n);
if (DEBUG_EVECS > 0)
{
{
Trace.WriteLine($" extended residual: {resid:g}");
}
}
if (warnings && WARNING_EVECS > 0 && resid > eigtol)
{
Trace.WriteLine($"WARNING: Extended residual ({resid:g}) greater than tolerance ({eigtol:g}).");
}
return(resid);
}
private static double compute_cube_edata(refine_edata *edata, /* desire data for current edge */
refine_vdata *vdata, /* data for all vertices */
int nsets_tot, /* total number of processors */
vtx_data **comm_graph, /* communication graph */
int *node2vtx /* maps mesh nodes to graph vertices */
)
{
double desire; /* edge's interest in flipping */
float ewgt; /* edge weight */
int offset; /* offset into vdata array */
int vtx1, vtx2; /* vertices on either side of wire */
vtx1 = node2vtx[edata->node1];
vtx2 = node2vtx[edata->node2];
offset = nsets_tot * edata->dim;
desire = (vdata[offset + vtx1].above - vdata[offset + vtx1].same) +
(vdata[offset + vtx2].above - vdata[offset + vtx2].same);
/* Subtract off potential doubly counted edge. */
if (is_an_edge(comm_graph[vtx1], vtx2, &ewgt))
{
desire -= 2 * ewgt;
}
return(desire);
}
public static void count_weights(vtx_data **graph, /* data structure for graph */
int nvtxs, /* number of vtxs in graph */
int *sets, /* set each vertex is assigned to */
int nsets, /* number of sets in this division */
double [] weights, /* vertex weights in each set */
bool using_vwgts /* are vertex weights being used? */
)
{
int i; /* loop counters */
/* Compute the sum of vertex weights for each set. */
for (i = 0; i < nsets; i++)
{
weights[i] = 0;
}
if (using_vwgts)
{
for (i = 1; i <= nvtxs; i++)
{
weights[sets[i]] += graph[i]->vwgt;
}
}
else
{
for (i = 1; i <= nvtxs; i++)
{
weights[sets[i]]++;
}
}
}
public static double compute_mesh_edata(refine_edata *edata, /* desire data for current edge */
refine_vdata *vdata, /* data for all vertices */
int[] mesh_dims /*[3]*/, /* dimensions of processor mesh */
vtx_data **comm_graph, /* communication graph */
int *node2vtx /* maps mesh nodes to graph vertices */
)
{
double desire; /* edge's interest in flipping */
float ewgt; /* edge weight */
int vtx1, vtx2; /* vertices on either side of wire */
int off; /* index into vdata */
vtx1 = node2vtx[edata->node1];
vtx2 = node2vtx[edata->node2];
off = edata->dim * mesh_dims[0] * mesh_dims[1] * mesh_dims[2];
desire = (vdata[off + vtx1].above - vdata[off + vtx1].below - vdata[off + vtx1].same) +
(vdata[off + vtx2].below - vdata[off + vtx2].above - vdata[off + vtx2].same);
/* Subtract off potential doubly counted edge. */
if (is_an_edge(comm_graph[vtx1], vtx2, &ewgt))
{
desire -= 2 * ewgt;
}
return(desire);
}
/* Set up data structures for refine_part. */
private static void make_maps_ref(vtx_data **graph, /* graph data structure */
bilist *set_list, /* lists of vertices in each set */
bilist *vtx_elems, /* start of storage for vertices */
int *assignment, /* set assignments for graph */
int *sub_assign, /* assignment file for subgraph */
int set1, int set2, /* set value denoting subgraph */
int *glob2loc, /* graph -> subgraph numbering map */
int *loc2glob, /* subgraph -> graph numbering map */
int *psub_nvtxs, /* number of vtxs in subgraph */
int *pvwgt_max, /* returned largest vwgt */
int *pvwgt_sum1, int *pvwgt_sum2 /* returned set sizes */
)
{
bilist *ptr; /* loops through set lists */
int vwgt_max; /* largest vertex weight in subgraph */
int vwgt_sum1, vwgt_sum2; /* sum of vertex weights in sets */
int vtx; /* vertex in subgraph */
int size; /* array spacing */
int j; /* loop counter */
size = (int)(&(vtx_elems[1]) - &(vtx_elems[0]));
j = 1;
vwgt_max = vwgt_sum1 = vwgt_sum2 = 0;
for (ptr = set_list[set1].next; ptr != null; ptr = ptr->next)
{
vtx = ((int)(ptr - vtx_elems)) / size;
sub_assign[j] = 0;
glob2loc[vtx] = j;
loc2glob[j] = vtx;
if (graph[vtx]->vwgt > vwgt_max)
{
vwgt_max = graph[vtx]->vwgt;
}
vwgt_sum1 += graph[vtx]->vwgt;
j++;
}
for (ptr = set_list[set2].next; ptr != null; ptr = ptr->next)
{
vtx = ((int)(ptr - vtx_elems)) / size;
sub_assign[j] = 1;
glob2loc[vtx] = j;
loc2glob[j] = vtx;
if (graph[vtx]->vwgt > vwgt_max)
{
vwgt_max = graph[vtx]->vwgt;
}
vwgt_sum2 += graph[vtx]->vwgt;
assignment[vtx] = set1;
j++;
}
*pvwgt_sum1 = vwgt_sum1;
*pvwgt_sum2 = vwgt_sum2;
*pvwgt_max = vwgt_max;
*psub_nvtxs = j - 1;
}
/// <summary>
///
/// </summary>
/// <param name="graph">data structure with vtx weights</param>
/// <param name="yvecs">ptr to list of y-vectors (lengths nvtxs+1)</param>
/// <param name="nvtxs">number of vertices in graph</param>
/// <param name="ndims">number of vectors for dividing</param>
/// <param name="cubeOrMesh">0 => hypercube, d => d-dimensional mesh</param>
/// <param name="nsets">number of sets to divide into</param>
/// <param name="wsqrt">sqrt of vertex weights</param>
/// <param name="sets">processor assignment for my vtxs</param>
/// <param name="active">space for nvtxs integers</param>
/// <param name="mediantype">which partitioning strategy to use</param>
/// <param name="goal">desired set sizes</param>
/// <param name="vertexWeightMax">largest vertex weight</param>
public static void Assign(vtx_data **graph, double *[] yvecs, int nvtxs, int ndims, bool cubeOrMesh, int nsets, double *wsqrt, int *sets, int *active, MappingType mediantype, double[] goal, int vertexWeightMax)
{
if (yvecs == null)
{
throw new ArgumentNullException(nameof(yvecs));
}
if (DEBUG_TRACE)
{
Trace.WriteLine($"<Entering {nameof(Assign)}, {nameof(nvtxs)} = {nvtxs:D}, {nameof(ndims)} = {ndims:D}>");
}
var useVertexWeights = vertexWeightMax != 1;
switch (ndims)
{
case 1:
{
break;
}
case 2:
{
var theta = opt2d(graph, yvecs, nvtxs, nvtxs);
Rotate2d(yvecs, nvtxs, theta);
break;
}
case 3:
{
if (DEBUG_ASSIGN)
{
var temp = tri_prod(yvecs[1], yvecs[2], yvecs[3], wsqrt, nvtxs);
Trace.WriteLine($"Before rotation, 3-way orthogonality = {temp:e}");
}
double phi, gamma, theta; /* angles for optimal rotation */
opt3d(graph, yvecs, nvtxs, nvtxs, wsqrt, &theta, &phi, &gamma, useVertexWeights);
Rotate3d(yvecs, nvtxs, theta, phi, gamma);
if (DEBUG_ASSIGN)
{
var temp = tri_prod(yvecs[1], yvecs[2], yvecs[3], wsqrt, nvtxs);
Trace.WriteLine($"After rotation ({theta:f},{phi:f},{gamma:f}), 3-way orthogonality = {temp:e}");
}
break;
}
default:
{
throw new ArgumentOutOfRangeException(nameof(ndims));
}
}
/* Unscale yvecs to get xvecs. */
y2x(yvecs, ndims, nvtxs, wsqrt);
mapper(graph, yvecs, nvtxs, active, sets, ndims, cubeOrMesh, nsets, mediantype, goal, vertexWeightMax);
}
public static void map2d(vtx_data **graph, /* data structure with vertex weights */
double *[] xvecs, /* vectors to partition */
int vertexCount, /* number of vertices */
int *sets, /* set each vertex gets assigned to */
double[] goal, /* desired set sizes */
int maxVertexWeight /* largest vertex weight */
)
{
double *[][] vals = new double *[4][]; /* values in sorted lists */
for (var i = 0; i < vals.Length; i++)
{
vals[i] = new double *[MAXSETS];
}
double[] dist = new double[4]; /* trial separation point */
double[] size = new double[4]; /* sizes of each set being modified */
int *[][] indices = new int *[4][]; /* indices sorting lists */
for (var i = 0; i < indices.Length; i++)
{
indices[i] = new int *[MAXSETS];
}
int[][] startvtx = new int[4][]; /* indices defining separation */
for (var i = 0; i < startvtx.Length; i++)
{
startvtx[i] = new int[MAXSETS];
}
N_VTX_CHECKS = N_VTX_MOVES = 0;
/* Generate all the lists of values that need to be sorted. */
genvals2d(xvecs, vals, vertexCount);
/* Sort the lists of values. */
sorts2d(vals, indices, vertexCount);
/* Now initialize dists and assign to sets. */
inits2d(graph, xvecs, vals, indices, vertexCount, dist, startvtx, size, sets);
/* Determine the largest and smallest allowed set sizes. */
/* (For now, assume all sets must be same size, but can easily change.) */
if (DEBUG_BPMATCH == DebugFlagBP.ErrorChecking)
{
Trace.WriteLine(" Calling check before movevtxs");
checkbp(graph, xvecs, sets, dist, vertexCount, nsection);
}
movevtxs(graph, vertexCount, nsets, dist, indices, vals, startvtx, sets, size, goal, maxVertexWeight);
if (DEBUG_BPMATCH != DebugFlagBP.NoDebugging)
{
Trace.WriteLine($" {nameof(N_VTX_CHECKS)} = {N_VTX_CHECKS:d}, {nameof(N_VTX_MOVES)} = {N_VTX_MOVES:d}");
checkbp(graph, xvecs, sets, dist, vertexCount, nsection);
}
free2d(vals, indices);
}
static float *old_ewgts; /* space for old edge weights */
public static void compress_ewgts(vtx_data **graph, /* list of graph info for each vertex */
int nvtxs, /* number of vertices in graph */
int nedges, /* number of edges in graph */
double ewgt_max, /* largest edge weight */
bool useEdgeWeights /* are edge weights being used? */
)
{
float *old_ewptr; /* loops through old edge weights */
float *new_ewptr; /* loops through old edge weights */
float *self_ptr; /* points to self edge in new_ewgts */
float *new_ewgts; /* space for new edge weights */
int ewgt; /* new edge weight value */
double sum; /* sum of all the new edge weights */
double ratio; /* amount to reduce edge weights */
int i, j; /* loop counter */
/* Check easy cases first. */
if (!useEdgeWeights)
{
old_ewgts = null;
}
else if (ewgt_max < EWGT_RATIO_MAX * nvtxs)
{
/* If not too heavy, leave it alone. */
old_ewgts = null;
Trace.WriteLine($"In compress_ewgts, but not too heavy, ewgt_max = {ewgt_max:g}, nvtxs = {nvtxs:d}");
}
else /* Otherwise, compress edge weights. */
/* Allocate space for new edge weights */
{
old_ewgts = graph[1]->ewgts;
new_ewgts = (float *)Marshal.AllocHGlobal((2 * nedges + nvtxs) * sizeof(float));
ratio = (EWGT_RATIO_MAX * nvtxs) / ewgt_max;
Trace.WriteLine("In compress_ewgts, ewgt_max = {ewgt_max:g}, nvtxs = {nvtxs:d}, ratio = {ratio:e}");
old_ewptr = old_ewgts;
new_ewptr = new_ewgts;
for (i = 1; i <= nvtxs; i++)
{
self_ptr = new_ewptr++;
old_ewptr++;
sum = 0;
for (j = graph[i]->nedges - 1; j != 0; j--)
{
ewgt = (int)(*(old_ewptr++) * ratio + 1.0);
*(new_ewptr++) = (float)ewgt;
sum += (float)ewgt;
}
*self_ptr = (float)-sum;
graph[i]->ewgts = self_ptr;
}
}
}
/// <summary>
/// Find a maximal matching in a graph using one of several algorithms.
/// </summary>
public static int maxmatch(vtx_data **graph, /* array of vtx data for graph */
int nvtxs, /* number of vertices in graph */
int nedges, /* number of edges in graph */
int *mflag, /* flag indicating vtx selected or not */
bool useEdgeWeights, /* are edge weights being used? */
int igeom, /* geometric dimensionality */
float **coords /* coordinates for each vertex */
)
{
int nmerged = 0; /* number of matching edges found */
if (MATCH_TYPE == MatchingRoutine.maxmatch1 || (MATCH_TYPE == MatchingRoutine.maxmatch5_geometric && coords == null))
{
/* Dumb, fast routine. */
nmerged = maxmatch1(graph, nvtxs, mflag, useEdgeWeights);
}
else if (MATCH_TYPE == MatchingRoutine.maxmatch2)
{
/* More random but somewhat slower. */
nmerged = maxmatch2(graph, nvtxs, mflag, useEdgeWeights);
}
else if (MATCH_TYPE == MatchingRoutine.maxmatch3)
{
/* Much more random but slower still. */
nmerged = maxmatch3(graph, nvtxs, mflag, useEdgeWeights);
}
else if (MATCH_TYPE == MatchingRoutine.maxmatch4_Luby)
{
/* Truly random but very slow. */
nmerged = maxmatch4(graph, nvtxs, nedges, mflag, useEdgeWeights);
}
else if (MATCH_TYPE == MatchingRoutine.maxmatch5_geometric && coords != null)
{
/* Geometric nearness. */
nmerged = maxmatch5(graph, nvtxs, mflag, igeom, coords);
}
else if (MATCH_TYPE == MatchingRoutine.maxmatch9_minimumVertexDegree)
{
/* Minimum degree of merged vertex */
nmerged = maxmatch9(graph, nvtxs, mflag, useEdgeWeights);
}
if (DEBUG_COARSEN)
{
Trace.WriteLine($"Number of matching edges = {nmerged:D}");
}
return(nmerged);
}
public static void coarsen1(vtx_data **graph, /* array of vtx data for graph */
int nvtxs, /* number of vertices in graph */
int nedges, /* number of edges in graph */
vtx_data ***pcgraph, /* coarsened version of graph */
int *pcnvtxs, /* number of vtxs in coarsened graph */
int *pcnedges, /* number of edges in coarsened graph */
int **pv2cv, /* pointer to v2cv */
int igeom, /* dimension for geometric information */
float **coords, /* coordinates for vertices */
float **ccoords, /* coordinates for coarsened vertices */
bool useEdgeWeights /* are edge weights being used? */
)
{
double time; /* time routine is entered */
int * v2cv; /* maps from vtxs to cvtxs */
int * mflag; /* flag indicating vtx matched or not */
int cnvtxs; /* number of vtxs in coarse graph */
int nmerged; /* number of edges contracted */
time = seconds();
/* Allocate and initialize space. */
v2cv = (int *)Marshal.AllocHGlobal((nvtxs + 1) * sizeof(int));
mflag = (int *)Marshal.AllocHGlobal((nvtxs + 1) * sizeof(int));
/* Find a maximal matching in the graph. */
nmerged = maxmatch(graph, nvtxs, nedges, mflag, useEdgeWeights, igeom, coords);
match_time += seconds() - time;
/* Now construct coarser graph by contracting along matching edges. */
/* Pairs of values in mflag array indicate matched vertices. */
/* A zero value indicates that vertex is unmatched. */
/*
* makecgraph(graph, nvtxs, pcgraph, pcnvtxs, pcnedges, mflag, *pv2cv, nmerged, useEdgeWeights, igeom, coords, ccoords);
* makecgraph2(graph, nvtxs, nedges, pcgraph, pcnvtxs, pcnedges, mflag, *pv2cv, nmerged, useEdgeWeights, igeom, coords, ccoords);
*/
makev2cv(mflag, nvtxs, v2cv);
Marshal.FreeHGlobal((IntPtr)mflag);
cnvtxs = nvtxs - nmerged;
makefgraph(graph, nvtxs, nedges, pcgraph, cnvtxs, pcnedges, v2cv, useEdgeWeights, igeom, coords,
ccoords);
*pcnvtxs = cnvtxs;
*pv2cv = v2cv;
coarsen_time += seconds() - time;
}
public static void median_assign(vtx_data **graph, /* data structure with vertex weights */
double *vals, /* values of which to find median */
int nvtxs, /* number of values I own */
double[] goal, /* desired sizes for sets */
bool using_vwgts, /* are vertex weights being used? */
int *sets, /* assigned set for each vertex */
double wlow, /* sum of weights below guess */
double whigh, /* sum of weights above guess */
double guess /* median value */
)
{
for (var i = 1; i <= nvtxs; i++)
{
if (vals[i] < guess)
{
sets[i] = 0;
}
else if (vals[i] > guess)
{
sets[i] = 1;
}
else
{
if (goal[0] - wlow > goal[1] - whigh)
{
sets[i] = 0;
if (using_vwgts)
{
wlow += graph[i]->vwgt;
}
else
{
wlow++;
}
}
else
{
sets[i] = 1;
if (using_vwgts)
{
whigh += graph[i]->vwgt;
}
else
{
whigh++;
}
}
}
}
}
public static void print_sep_size(int *list, vtx_data **graph /* array of vtx data for graph */
)
{
int sep_size, sep_weight;
int i;
sep_size = sep_weight = 0;
for (i = 0; list[i] != 0; i++)
{
sep_size++;
sep_weight += graph[list[i]]->vwgt;
}
Trace.WriteLine($" {nameof(sep_size)} = {sep_size:d}, {nameof(sep_weight)} = {sep_weight:d}");
}
/// <summary>
/// BFS to find connected component.
/// </summary>
/// <param name="graph">graph data structure</param>
/// <param name="root">start vertex for DFS</param>
/// <param name="count">number of vertices in component</param>
/// <param name="mark">has vtx been seen?</param>
/// <param name="vtxlist">space for storing vtxs to search</param>
/// <param name="currentComponentNumber">current component number</param>
/// <returns></returns>
public static int bfsearch(vtx_data **graph, int root, int *count, int *mark, int *vtxlist, int currentComponentNumber)
{
var firstVertex = 1;
var lastVertexInList = 1;
mark[root] = currentComponentNumber;
vtxlist[0] = root;
// Copy root's neighbors to vtxlist, incrementing count
var edgePointer = graph[root]->edges;
if (graph[root]->nedges - 1 < 0)
{
throw new InvalidOperationException("A node should have at least one edge");
}
for (var i = graph[root]->nedges - 1; i != 0; i--)
{
// neighbor of vertex
var neighbor = *(++edgePointer);
vtxlist[lastVertexInList++] = neighbor;
mark[neighbor] = currentComponentNumber;
}
while (firstVertex < lastVertexInList)
{
// vertex being processed
var vtx = vtxlist[firstVertex++];
// Loop through neighbors, copying to vtxlist if unmarked.
var iptr = graph[vtx]->edges;
if (graph[vtx]->nedges - 1 < 0)
{
throw new InvalidOperationException("A node should have at least one edge");
}
for (var i = graph[vtx]->nedges - 1; i != 0; i--)
{
// neighbor of vertex
var neighbor = *(++iptr);
if (mark[neighbor] != currentComponentNumber)
{
mark[neighbor] = currentComponentNumber;
vtxlist[lastVertexInList++] = neighbor;
}
}
}
*count += lastVertexInList;
return(vtxlist[lastVertexInList - 1]);
}
public static void countup_vtx_sep(vtx_data **graph, /* list of graph info for each vertex */
int nvtxs, /* number of vertices in graph */
int *sets /* local partitioning of vtxs */
)
{
int vtx, set; /* vertex and set in graph */
int sep_size; /* size of the separator */
int i, j, k; /* loop counters */
sep_size = 0;
j = k = 0;
for (i = 1; i <= nvtxs; i++)
{
if (sets[i] == 0)
{
j += graph[i]->vwgt;
}
if (sets[i] == 1)
{
k += graph[i]->vwgt;
}
if (sets[i] == 2)
{
sep_size += graph[i]->vwgt;
}
}
Trace.WriteLine($"Set sizes = {j:d}/{k:d}, Separator size = {sep_size:d}\n");
/* Now check that it really is a separator. */
for (i = 1; i <= nvtxs; i++)
{
set = sets[i];
if (set != 2)
{
for (j = 1; j < graph[i]->nedges; j++)
{
vtx = graph[i]->edges[j];
if (sets[vtx] != 2 && sets[vtx] != set)
{
Trace.WriteLine($"Error: {i:d} (set {set:d}) adjacent to {vtx:d} (set {sets[vtx]:d})");
}
}
}
}
}
public static void make_connected(
/* Add edges to make graph connected. */
vtx_data **graph, /* graph data structure */
int nvtxs, /* number of vertices in graph */
int *nedges, /* number of edges in graph */
int *mark, /* space for nvtxs+1 ints */
int *vtxlist, /* space for nvtxs ints */
connect_data **cdata, /* space for connectivity data */
bool useEdgeWeights /* are edges of graph weighted? */
)
{
edgeslist *new_edges; /* list of edges connecting graph */
edgeslist *prev_edge; /* pointer for manipulating edge list */
edgeslist *curr_edge; /* pointer for manipulating edge list */
edgeslist *next_edge; /* pointer for manipulating edge list */
int nadded; /* number of edges being added */
/* First find edges needed to make graph connected. */
nadded = find_edges(graph, nvtxs, mark, vtxlist, &new_edges);
/* Now add these needed edges to graph data structure if needed. */
if (nadded == 0)
{
*cdata = null;
}
else
{
*cdata = (connect_data *)Marshal.AllocHGlobal(sizeof(connect_data));
(*cdata)->old_edges = null;
(*cdata)->old_ewgts = null;
add_edges(graph, new_edges, &(*cdata)->old_edges, &(*cdata)->old_ewgts, useEdgeWeights);
*nedges += nadded;
/* Now, reverse the order of the new_edges list for consistency with */
/* the removal order. */
curr_edge = new_edges->next;
new_edges->next = null;
prev_edge = new_edges;
while (curr_edge != null)
{
next_edge = curr_edge->next;
curr_edge->next = prev_edge;
prev_edge = curr_edge;
curr_edge = next_edge;
}
(*cdata)->new_edges = prev_edge;
}
}
/// <summary>
/// Breadth first search algorithm to find & mark connected components.
/// Returns list of edges to connect them together.
/// </summary>
/// <param name="graph">graph data structure</param>
/// <param name="nvtxs">number of vertices in graph</param>
/// <param name="mark">space for nvtxs+1 ints</param>
/// <param name="vtxlist">space for nvtxs ints</param>
/// <param name="edges">list of edges connecting graph</param>
/// <returns></returns>
public static int find_edges(vtx_data **graph, int nvtxs, int *mark, int *vtxlist, edgeslist **edges)
{
Trace.WriteLine($"<Entering {nameof(find_edges)}>");
for (var i = 1; i <= nvtxs; i++)
{
mark[i] = -1;
}
var visitedVertexCount = 0;
var edgesAdded = 0;
*edges = null;
// vertex to start the dfs
var root = (int)(nvtxs * drandom()) + 1;
// last vertex seen in BFS
var last = bfsearch(graph, root, &visitedVertexCount, mark, vtxlist, edgesAdded);
while (visitedVertexCount != nvtxs)
{
// Are there any remaining vertices?
// Find starting vtx for next BFS.
root = (int)(nvtxs * drandom()) + 1;
//Trace.WriteLine($"looking for next root from: {root}");
while (mark[root] >= 0)
{
root++;
if (root > nvtxs)
{
root = 1;
}
}
// Add new edge to list needed for connectivity.
var newedge = (edgeslist *)Marshal.AllocHGlobal(sizeof(edgeslist));
newedge->next = *edges;
newedge->vtx1 = last;
newedge->vtx2 = root;
*edges = newedge;
edgesAdded++;
last = bfsearch(graph, root, &visitedVertexCount, mark, vtxlist, edgesAdded);
}
return(edgesAdded);
}
/* Find vertices on boundary of partition, and change their assignments. */
public static int find_bndy(vtx_data **graph, /* array of vtx data for graph */
int nvtxs, /* number of vertices in graph */
int *assignment, /* processor each vertex gets assigned to */
int new_val, /* assignment value for boundary vtxs */
int **pbndy_list /* returned list, end with zero */
)
{
int *bndy_list; /* returned list, end with zero */
int *edges; /* loops through edge list */
int list_length; /* returned number of vtxs on boundary */
int set, set2; /* set a vertex is in */
int i, j; /* loop counters */
bndy_list = (int *)Marshal.AllocHGlobal((nvtxs + 1) * sizeof(int));
list_length = 0;
for (i = 1; i <= nvtxs; i++)
{
set = assignment[i];
edges = graph[i]->edges;
for (j = graph[i]->nedges - 1; j != 0; j--)
{
set2 = assignment[*(++edges)];
if (set2 != set)
{
bndy_list[list_length++] = i;
break;
}
}
}
bndy_list[list_length] = 0;
for (i = 0; i < list_length; i++)
{
assignment[bndy_list[i]] = new_val;
}
/* Shrink out unnecessary space */
*pbndy_list = (int *)Marshal.ReAllocHGlobal((IntPtr)bndy_list, new IntPtr((list_length + 1) * sizeof(int)));
return(list_length);
}
private static void clear_dvals(vtx_data **graph, /* data structure for graph */
int nvtxs, /* number of vtxs in graph */
int *ldvals, /* d-values for each transition */
int *rdvals, /* d-values for each transition */
int *bspace, /* list of activated vertices */
int list_length /* number of activated vertices */
)
{
int *edges; /* loops through edge list */
int vtx; /* vertex in bspace */
int neighbor; /* neighbor of vtx */
int i, j; /* loop counters */
if (list_length > .05 * nvtxs)
{
/* Do it directly. */
for (i = 1; i <= nvtxs; i++)
{
ldvals[i] = rdvals[i] = 0;
}
}
else
{
/* Do it more carefully */
for (i = 0; i < list_length; i++)
{
vtx = bspace[i];
if (vtx < 0)
{
vtx = -vtx;
}
ldvals[vtx] = rdvals[vtx] = 0;
edges = graph[vtx]->edges;
for (j = graph[vtx]->nedges - 1; j != 0; j--)
{
neighbor = *(++edges);
ldvals[neighbor] = rdvals[neighbor] = 0;
}
}
}
}
/// <summary>
/// Breadth first search algorithm to find & mark connected components.
/// </summary>
/// <param name="graph">graph data structure</param>
/// <param name="nvtxs">number of vertices in graph</param>
/// <param name="mark">space for nvtxs+1 ints</param>
/// <param name="vtxlist">space for nvtxs ints</param>
/// <returns>The number of connected components.</returns>
public static int find_comps(vtx_data **graph, int nvtxs, int *mark, int *vtxlist)
{
Trace.WriteLine($"<Entering {nameof(find_comps)}>");
for (var i = 1; i <= nvtxs; i++)
{
mark[i] = -1;
}
var visitedVertexCount = 0;
var componentCount = 0;
// vertex to start the dfs
var root = (int)(nvtxs * drandom()) + 1;
bfsearch(graph, root, &visitedVertexCount, mark, vtxlist, componentCount);
// Are there any remaining vertices?
while (visitedVertexCount != nvtxs)
{
// Find starting vtx for next BFS.
root = (int)(nvtxs * drandom()) + 1;
while (mark[root] >= 0)
{
root++;
if (root > nvtxs)
{
root = 1;
}
}
// Add new edge to list needed for connectivity.
componentCount++;
bfsearch(graph, root, &visitedVertexCount, mark, vtxlist, componentCount);
Trace.WriteLine($"Visited {visitedVertexCount} out of {nvtxs} vertices");
}
Trace.WriteLine($"<Exiting {nameof(find_comps)}>");
return(componentCount + 1);
}
public static void makevwsqrt(double *vwsqrt, vtx_data **graph, int nvtxs)
/* Make vector of square roots of vertex weights. */
/* vector returned */
/* graph data structure */
/* number of vertices in graph */
{
int vwgt; /* vertex weight */
int i; /* loop counter */
for (i = 1; i <= nvtxs; i++)
{
vwgt = graph[i]->vwgt;
if (vwgt <= NSQRTS)
{
vwsqrt[i] = SQRTS[vwgt];
}
else
{
vwsqrt[i] = Math.Sqrt(vwgt);
}
}
}
public static void restore_ewgts(vtx_data **graph, /* list of graph info for each vertex */
int nvtxs /* number of vertices in graph */
)
{
int i; /* loop counter */
/* Check easy case first. */
if (old_ewgts == null)
{
return;
}
else /* otherwise, compress edge weights. */
{
Marshal.FreeHGlobal((IntPtr)(graph[1]->ewgts));
for (i = 1; i <= nvtxs; i++)
{
graph[i]->ewgts = old_ewgts;
old_ewgts += graph[i]->nedges;
}
old_ewgts = null;
}
}
private static void compute_cube_vdata(refine_vdata *vdata, /* preference data for a vertex */
vtx_data **comm_graph, /* communication graph data structure */
int vtx, /* current vertex */
int mask, /* bit set in current hypercube dimension */
int *vtx2node /* maps graph vtxs to mesh nodes */
)
{
float same; /* my preference to stay where I am */
float change; /* my preference to change this bit */
float ewgt; /* weight of an edge */
int neighbor; /* neighboring vtx in comm_graph */
int my_side; /* which side of hypercube I'm on */
int neighb_side; /* which side of hypercube neighbor's on */
int j; /* loop counter */
my_side = (vtx2node[vtx] & mask);
change = 0;
same = 0;
for (j = 1; j < comm_graph[vtx]->nedges; j++)
{
neighbor = comm_graph[vtx]->edges[j];
ewgt = comm_graph[vtx]->ewgts[j];
neighb_side = (vtx2node[neighbor] & mask);
if (neighb_side != my_side)
{
change += ewgt;
}
else
{
same += ewgt;
}
}
vdata->same = same;
vdata->above = change;
}
/// <summary>
/// Undo the construction of the subgraph.
/// </summary>
/// <param name="subgraph">subgraph data structure</param>
/// <param name="subnvtxs">number of vtxs in subgraph</param>
/// <param name="loc2glob">mapping from subgraph to graph numbering</param>
/// <param name="degree">degrees of vertices in graph</param>
/// <param name="useEdgeWeights">are edge weights being used?</param>
public static void remake_graph(vtx_data **subgraph, int subnvtxs, int *loc2glob, int *degree, bool useEdgeWeights)
{
Trace.WriteLine($"<Entering {nameof(remake_graph)}>");
vtx_data *subgptr; /* loops through subgraph */
double ewgtsum; /* sum of weights of subgraph edges */
int nedges; /* vertex degree in subgraph */
/* For each vertex in subgraph, expand the edge set back out. */
for (var i = 1; i <= subnvtxs; i++)
{
subgptr = subgraph[i];
subgptr->edges[0] = loc2glob[i];
nedges = subgptr->nedges;
/* Change edges back to global numbering system. */
var iptr = subgptr->edges; /* loops through adjacency list */
for (var j = nedges - 1; j != 0; j--)
{
iptr++;
*iptr = loc2glob[*iptr];
}
subgptr->nedges = degree[i];
/* Now get the diagonal value right. */
if (useEdgeWeights)
{
ewgtsum = 0;
var fptr = subgptr->ewgts; /* loops through edge weights */
for (var j = degree[i] - 1; j != 0; j--)
{
ewgtsum += *(++fptr);
}
subgptr->ewgts[0] = (float)-ewgtsum;
}
}
}
public static bool flatten(vtx_data **graph, /* array of vtx data for graph */
int nvtxs, /* number of vertices in graph */
int nedges, /* number of edges in graph */
vtx_data ***pcgraph, /* coarsened version of graph */
int *pcnvtxs, /* number of vtxs in coarsened graph */
int *pcnedges, /* number of edges in coarsened graph */
int **pv2cv, /* pointer to v2cv */
bool useEdgeWeights, /* are edge weights being used? */
int igeom, /* dimensions of geometric data */
float **coords, /* coordinates for vertices */
float **ccoords /* coordinates for coarsened vertices */
)
{
double Thresh; /* minimal acceptable size reduction */
int * v2cv; /* map from vtxs to coarse vtxs */
int cnvtxs; /* number of vertices in flattened graph */
Thresh = .9;
v2cv = (int *)Marshal.AllocHGlobal((nvtxs + 1) * sizeof(int));
find_flat(graph, nvtxs, &cnvtxs, v2cv);
if (cnvtxs <= Thresh * nvtxs) /* Sufficient shrinkage? */
{
makefgraph(graph, nvtxs, nedges, pcgraph, cnvtxs, pcnedges, v2cv, useEdgeWeights, igeom, coords,
ccoords);
*pcnvtxs = cnvtxs;
*pv2cv = v2cv;
return(true);
}
/* Not worth bothering */
Marshal.FreeHGlobal((IntPtr)v2cv);
return(false);
}
/// <summary>
/// Free a graph data structure.
/// </summary>
public static void free_graph(vtx_data **graph)
{
if (graph == null)
{
return;
}
if (graph[1] != null)
{
if (graph[1]->ewgts != null)
{
Marshal.FreeHGlobal((IntPtr)graph[1]->ewgts);
}
if (graph[1]->edges != null)
{
Marshal.FreeHGlobal((IntPtr)graph[1]->edges);
}
Marshal.FreeHGlobal((IntPtr)graph[1]);
}
Marshal.FreeHGlobal((IntPtr)graph);
}
/// <summary>
/// Benchmark certain kernel operations
/// </summary>
/// <param name="A">matrix/graph being analyzed</param>
/// <param name="n">number of rows/columns in matrix</param>
/// <param name="vwsqrt">square roots of vertex weights</param>
public static void time_kernels(vtx_data **A, int n, double *vwsqrt)
{
const double minTime = 0.5;
const double targetTime = 1.0;
int i;
float *vwsqrtFloat;
double time;
if (DEBUG_TRACE)
{
Trace.WriteLine("<Entering time_kernels>");
}
const int beg = 1;
var end = n;
var dvec1 = mkvec(beg, end);
var dvec2 = mkvec(beg, end);
var dvec3 = mkvec(beg - 1, end);
var svec1 = mkvec_float(beg, end);
var svec2 = mkvec_float(beg, end);
var svec3 = mkvec_float(beg - 1, end);
if (vwsqrt == null)
{
vwsqrtFloat = null;
}
else
{
vwsqrtFloat = mkvec_float(beg - 1, end);
for (i = beg - 1; i <= end; i++)
{
vwsqrtFloat[i] = (float)vwsqrt[i];
}
}
vecran(dvec1, beg, end);
vecran(dvec2, beg, end);
vecran(dvec3, beg, end);
for (i = beg; i <= end; i++)
{
svec1[i] = (float)dvec1[i];
svec2[i] = (float)dvec2[i];
svec3[i] = (float)dvec3[i];
}
// Set number of loops so that ch_norm() takes about one second.
// This should insulate against inaccurate timings on faster machines.
var normDvec = 0.0d;
var loops = 1;
double timeDvec = 0;
while (timeDvec < minTime)
{
time = seconds();
for (i = loops; i != 0; i--)
{
normDvec = ch_norm(dvec1, beg, end);
}
timeDvec = seconds() - time;
if (timeDvec < minTime)
{
loops = 10 * loops;
}
}
loops = (int)((targetTime / timeDvec) * loops);
if (loops < 1)
{
loops = 1;
}
Trace.WriteLine(" Kernel benchmarking");
Trace.WriteLine($"Time (in seconds) for {loops:d} loops of each operation:\n");
Trace.WriteLine("Routine Double Float Discrepancy Description");
Trace.WriteLine("------- ------ ----- ----------- -----------");
/* Norm operation */
time = seconds();
for (i = loops; i != 0; i--)
{
normDvec = ch_norm(dvec1, beg, end);
}
timeDvec = seconds() - time;
time = seconds();
var normSvec = 0.0d;
for (i = loops; i != 0; i--)
{
normSvec = norm_float(svec1, beg, end);
}
var timeSvec = seconds() - time;
var diff = normDvec - normSvec;
Trace.Write($"norm {timeDvec:f} {timeSvec:f} {diff:e}");
Trace.WriteLine(" 2 norm");
/* Dot operation */
time = seconds();
var dotDvec = 0.0d;
for (i = loops; i != 0; i--)
{
dotDvec = dot(dvec1, beg, end, dvec2);
}
timeDvec = seconds() - time;
time = seconds();
var dotSvec = 0.0d;
for (i = loops; i != 0; i--)
{
dotSvec = dot_float(svec1, beg, end, svec2);
}
timeSvec = seconds() - time;
diff = dotDvec - dotSvec;
Trace.Write($"dot {timeDvec:f} {timeSvec:f} {diff:e}");
Trace.WriteLine(" scalar product");
/* Scadd operation */
const double factor = 1.01d;
const float factorFloat = (float)factor;
var fac = factor;
time = seconds();
for (i = loops; i != 0; i--)
{
scadd(dvec1, beg, end, fac, dvec2);
fac = -fac; /* to keep things in scale */
}
timeDvec = seconds() - time;
var facFloat = factorFloat;
time = seconds();
for (i = loops; i != 0; i--)
{
scadd_float(svec1, beg, end, facFloat, svec2);
facFloat = -facFloat; /* to keep things in scale */
}
timeSvec = seconds() - time;
diff = checkvec(dvec1, beg, end, svec1);
Trace.Write($"scadd {timeDvec:f} {timeSvec:f} {diff:e}");
Trace.WriteLine(" vec1 <- vec1 + alpha*vec2");
/* Update operation */
time = seconds();
for (i = loops; i != 0; i--)
{
update(dvec1, beg, end, dvec2, factor, dvec3);
}
timeDvec = seconds() - time;
time = seconds();
for (i = loops; i != 0; i--)
{
update_float(svec1, beg, end, svec2, factorFloat, svec3);
}
timeSvec = seconds() - time;
diff = checkvec(dvec1, beg, end, svec1);
Trace.Write($"update {timeDvec:f} {timeSvec:f} {diff:g}");
Trace.WriteLine(" vec1 <- vec2 + alpha*vec3");
/* splarax operation */
if (PERTURB)
{
if (NPERTURB > 0 && PERTURB_MAX > 0.0)
{
perturb_init(n);
if (DEBUG_PERTURB)
{
Trace.WriteLine($"Matrix being perturbed with scale {PERTURB_MAX:e}");
}
}
else if (DEBUG_PERTURB)
{
Trace.WriteLine("Matrix not being perturbed");
}
}
time = seconds();
for (i = loops; i != 0; i--)
{
splarax(dvec1, A, n, dvec2, vwsqrt, dvec3);
}
timeDvec = seconds() - time;
time = seconds();
for (i = loops; i != 0; i--)
{
splarax_float(svec1, A, n, svec2, vwsqrtFloat, svec3);
}
timeSvec = seconds() - time;
diff = checkvec(dvec1, beg, end, svec1);
Trace.Write($"splarax {timeDvec:f} {timeSvec:f} {diff:e}");
Trace.WriteLine(" sparse matrix vector multiply");
if (PERTURB && NPERTURB > 0 && PERTURB_MAX > 0.0)
{
perturb_clear();
}
Trace.WriteLine("");
/* Free memory */
frvec(dvec1, 1);
frvec(dvec2, 1);
frvec(dvec3, 0);
frvec_float(svec1, 1);
frvec_float(svec2, 1);
frvec_float(svec3, 0);
if (vwsqrtFloat != null)
{
frvec_float(vwsqrtFloat, beg - 1);
}
}
/* See comments in lanczos_SO() */
public static void lanczos_SO_float(vtx_data **A, /* sparse matrix in row linked list format */
int n, /* problem size */
int d, /* problem dimension = number of eigvecs to find */
double *[] y, /* columns of y are eigenvectors of A */
double[] lambda, /* ritz approximation to eigenvals of A */
double[] bound, /* on ritz pair approximations to eig pairs of A */
double eigtol, /* tolerance on eigenvectors */
double *vwsqrt, /* square roots of vertex weights */
double maxdeg, /* maximum degree of graph */
int version, /* flags which version of sel. orth. to use */
bool cube_or_mesh, /* 0 => hypercube, d => d-dimensional mesh */
int nsets, /* number of sets to divide into */
int *assignment, /* set number of each vtx (length n+1) */
int *active, /* space for nvtxs integers */
MappingType mediantype, /* which partitioning strategy to use */
double[] goal, /* desired set sizes */
int vwgt_max /* largest vertex weight */
)
{
double bis_safety; /* real safety factor for T bisection */
int i, j, k; /* indices */
int maxj; /* maximum number of Lanczos iterations */
float * u;
float * r; /* Lanczos vectors */
double * u_double; /* double version of u */
double * alpha;
double * beta; /* the Lanczos scalars from each step */
double * ritz; /* copy of alpha for ql */
double * workj; /* work vector, e.g. copy of beta for ql */
float * workn; /* work vector, e.g. product Av for checkeig */
double * workn_double; /* work vector, e.g. product Av for checkeig */
double * s; /* eigenvector of T */
float ** q; /* columns of q are Lanczos basis vectors */
double * bj; /* beta(j)*(last el. of corr. eigvec s of T) */
double Sres; /* how well Tevec calculated eigvec s */
double Sres_max; /* Max value of Sres */
bool inc_bis_safety; /* need to increase bisection safety */
double * Ares; /* how well Lanczos calc. eigpair lambda,y */
int * index; /* the Ritz index of an eigenpair */
orthlink_float **solist; /* vec. of structs with vecs. to orthog. against */
scanlink * scanlist; /* linked list of fields to do with min ritz vals */
scanlink * curlnk; /* for traversing the scanlist */
double bji_tol; /* tol on bji est. of eigen residual of A */
bool converged; /* has the iteration converged? */
double goodtol; /* error tolerance for a good Ritz vector */
int ngood; /* total number of good Ritz pairs at current step */
int maxngood; /* biggest val of ngood through current step */
int left_ngood; /* number of good Ritz pairs on left end */
int right_ngood; /* number of good Ritz pairs on right end */
int lastpause; /* Most recent step with good ritz vecs */
bool firstpause; /* Is this the first pause? */
bool nopauses; /* Have there been any pauses? */
int interval; /* number of steps between pauses */
double time; /* Current clock time */
int left_goodlim; /* number of ritz pairs checked on left end */
int right_goodlim; /* number of ritz pairs checked on right end */
double Anorm; /* Norm estimate of the Laplacian matrix */
int pausemode; /* which Lanczos pausing criterion to use */
bool pause; /* whether to pause */
int temp; /* used to prevent redundant index computations */
int * old_assignment = null; /* set # of each vtx on previous pause, length n+1 */
int * assgn_pntr; /* pntr to assignment vector */
int * old_assgn_pntr; /* pntr to previous assignment vector */
int assigndiff; /* # of differences between old and new assignment */
int assigntol; /* tolerance on convergence of assignment vector */
bool ritzval_flag; /* status flag for get_ritzvals() */
bool memory_ok; /* True until lanczos runs out of memory */
float * vwsqrt_float; /* float version of vwsqrt */
if (DEBUG_TRACE)
{
Trace.WriteLine("<Entering lanczos_so_float>");
}
if (DEBUG_EVECS > 0)
{
Trace.WriteLine($"Selective orthogonalization Lanczos (v. {version:d}), matrix size = {n:d}.");
}
/* Initialize time. */
time = lanc_seconds();
if (n < d + 1)
{
throw new ArgumentOutOfRangeException(nameof(n), "ERROR: System too small for number of eigenvalues requested.");
/* d+1 since don't use zero eigenvalue pair */
}
/* Allocate space. */
maxj = LANCZOS_MAXITNS;
u = mkvec_float(1, n);
u_double = mkvec(1, n);
r = mkvec_float(1, n);
workn = mkvec_float(1, n);
workn_double = mkvec(1, n);
Ares = mkvec(0, d);
index = (int *)Marshal.AllocHGlobal((d + 1) * sizeof(int));
alpha = mkvec(1, maxj);
beta = mkvec(0, maxj);
ritz = mkvec(1, maxj);
s = mkvec(1, maxj);
bj = mkvec(1, maxj);
workj = mkvec(0, maxj);
q = (float **)Marshal.AllocHGlobal((maxj + 1) * sizeof(float *));
solist = (orthlink_float **)Marshal.AllocHGlobal((maxj + 1) * sizeof(orthlink_float *));
scanlist = mkscanlist(d);
if (LANCZOS_CONVERGENCE_MODE == 1)
{
old_assignment = (int *)Marshal.AllocHGlobal((n + 1) * sizeof(int));
}
/* Set some constants governing the orthogonalization heuristic. */
ngood = 0;
maxngood = 0;
bji_tol = eigtol;
assigntol = (int)(eigtol * n);
Anorm = 2 * maxdeg; /* Gershgorin estimate for ||A|| */
goodtol = Anorm * Math.Sqrt(DOUBLE_EPSILON); /* Parlett & Scott's bound, p.224 */
interval = 2 + Math.Min(LANCZOS_SO_INTERVAL - 2, n / (2 * LANCZOS_SO_INTERVAL));
bis_safety = BISECTION_SAFETY;
if (DEBUG_EVECS > 0)
{
Trace.WriteLine($" maxdeg {maxdeg:g}");
Trace.WriteLine($" goodtol {goodtol:g}");
Trace.WriteLine($" interval {interval:d}");
Trace.WriteLine($" maxj {maxj:d}");
if (LANCZOS_CONVERGENCE_MODE == 1)
{
Trace.WriteLine($" assigntol {assigntol:d}");
}
}
/* Make a float copy of vwsqrt */
if (vwsqrt == null)
{
vwsqrt_float = null;
}
else
{
vwsqrt_float = mkvec_float(0, n);
double_to_float(vwsqrt_float, 1, n, vwsqrt);
}
/* Initialize space. */
vecran_float(r, 1, n);
if (vwsqrt_float == null)
{
orthog1_float(r, 1, n);
}
else
{
orthogvec_float(r, 1, n, vwsqrt_float);
}
beta[0] = norm_float(r, 1, n);
q[0] = mkvec_float(1, n);
setvec_float(q[0], 1, n, 0.0f);
setvec(bj, 1, maxj, double.MaxValue);
/* Main Lanczos loop. */
j = 1;
lastpause = 0;
pausemode = 1;
left_ngood = 0;
right_ngood = 0;
left_goodlim = 0;
right_goodlim = 0;
converged = false;
Sres_max = 0.0;
inc_bis_safety = false;
ritzval_flag = false;
memory_ok = true;
firstpause = false;
nopauses = true;
init_time += lanc_seconds() - time;
while ((j <= maxj) && (!converged) && (!ritzval_flag) && memory_ok)
{
time = lanc_seconds();
/* Allocate next Lanczos vector. If fail, back up to last pause. */
q[j] = mkvec_ret_float(1, n);
if (q[j] == null)
{
memory_ok = false;
if (DEBUG_EVECS > 0 || WARNING_EVECS > 0)
{
Trace.WriteLine("WARNING: Lanczos out of memory; computing best approximation available.");
}
if (nopauses)
{
throw new InvalidOperationException("ERROR: Sorry, can't salvage Lanczos.");
/* ... save yourselves, men. */
}
for (i = lastpause + 1; i <= j - 1; i++)
{
frvec_float(q[i], 1);
}
j = lastpause;
}
vecscale_float(q[j], 1, n, (float)(1.0 / beta[j - 1]), r);
blas_time += lanc_seconds() - time;
time = lanc_seconds();
splarax_float(u, A, n, q[j], vwsqrt_float, workn);
splarax_time += lanc_seconds() - time;
time = lanc_seconds();
update_float(r, 1, n, u, (float)(-beta[j - 1]), q[j - 1]);
alpha[j] = dot_float(r, 1, n, q[j]);
update_float(r, 1, n, r, (float)(-alpha[j]), q[j]);
blas_time += lanc_seconds() - time;
time = lanc_seconds();
if (vwsqrt_float == null)
{
orthog1_float(r, 1, n);
}
else
{
orthogvec_float(r, 1, n, vwsqrt_float);
}
if ((j == (lastpause + 1)) || (j == (lastpause + 2)))
{
sorthog_float(r, n, solist, ngood);
}
orthog_time += lanc_seconds() - time;
beta[j] = norm_float(r, 1, n);
time = lanc_seconds();
pause = lanpause_float(j, lastpause, interval, q, n, &pausemode, version, beta[j]);
pause_time += lanc_seconds() - time;
if (pause)
{
nopauses = false;
if (lastpause == 0)
{
firstpause = true;
}
else
{
firstpause = false;
}
lastpause = j;
/* Compute limits for checking Ritz pair convergence. */
if (version == 1)
{
if (left_ngood + 2 > left_goodlim)
{
left_goodlim = left_ngood + 2;
}
if (right_ngood + 3 > right_goodlim)
{
right_goodlim = right_ngood + 3;
}
}
if (version == 2)
{
if (left_ngood + 2 > left_goodlim)
{
left_goodlim = left_ngood + 2;
}
right_goodlim = 0;
}
/* Special case: need at least d Ritz vals on left. */
left_goodlim = Math.Max(left_goodlim, d);
/* Special case: can't find more than j total Ritz vals. */
if (left_goodlim + right_goodlim > j)
{
left_goodlim = Math.Min(left_goodlim, j);
right_goodlim = j - left_goodlim;
}
/* Find Ritz vals using faster of Sturm bisection or QL. */
time = lanc_seconds();
blas_time += lanc_seconds() - time;
time = lanc_seconds();
if (inc_bis_safety)
{
bis_safety *= 10;
inc_bis_safety = false;
}
ritzval_flag = get_ritzvals(alpha, beta, j, Anorm, workj, ritz, d, left_goodlim,
right_goodlim, eigtol, bis_safety);
ql_time += lanc_seconds() - time;
/* If get_ritzvals() fails, back up to last pause point and exit main loop. */
if (ritzval_flag)
{
if (DEBUG_EVECS > 0 || WARNING_EVECS > 0)
{
Trace.WriteLine("ERROR: Lanczos failed in computing eigenvalues of T; computing");
Trace.WriteLine(" best readily available approximation to eigenvector.\n");
}
if (firstpause)
{
throw new InvalidOperationException("ERROR: Sorry, can't salvage Lanczos.");
/* ... save yourselves, men. */
}
for (i = lastpause + 1; i <= j; i++)
{
frvec_float(q[i], 1);
}
j = lastpause;
get_ritzvals(alpha, beta, j, Anorm, workj, ritz, d, left_goodlim, right_goodlim, eigtol,
bis_safety);
}
/* Scan for minimum evals of tridiagonal. */
time = lanc_seconds();
scanmin(ritz, 1, j, &scanlist);
scan_time += lanc_seconds() - time;
/* Compute Ritz pair bounds at left end. */
time = lanc_seconds();
setvec(bj, 1, j, 0.0);
for (i = 1; i <= left_goodlim; i++)
{
Sres = Tevec(alpha, beta - 1, j, ritz[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
if (Sres > SRESTOL)
{
inc_bis_safety = true;
}
bj[i] = s[j] * beta[j];
}
/* Compute Ritz pair bounds at right end. */
for (i = j; i > j - right_goodlim; i--)
{
Sres = Tevec(alpha, beta - 1, j, ritz[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
if (Sres > SRESTOL)
{
inc_bis_safety = true;
}
bj[i] = s[j] * beta[j];
}
ritz_time += lanc_seconds() - time;
/* Show the portion of the spectrum checked for convergence. */
if (DEBUG_EVECS > 2)
{
time = lanc_seconds();
Trace.WriteLine(string.Format("index Ritz vals bji bounds (j = %d)\n", j));
for (i = 1; i <= left_goodlim; i++)
{
Trace.Write($" {i:d}");
doubleout(ritz[i], 1);
doubleout(bj[i], 1);
Trace.WriteLine("");
}
Trace.WriteLine("");
curlnk = scanlist;
while (curlnk != null)
{
temp = curlnk->indx;
if ((temp > left_goodlim) && (temp < j - right_goodlim))
{
Trace.Write($" {temp:d}");
doubleout(ritz[temp], 1);
doubleout(bj[temp], 1);
Trace.WriteLine("");
}
curlnk = curlnk->pntr;
}
Trace.WriteLine("");
for (i = j - right_goodlim + 1; i <= j; i++)
{
Trace.Write($" {i:d}");
doubleout(ritz[i], 1);
doubleout(bj[i], 1);
Trace.WriteLine("");
}
Trace.WriteLine(" -------------------");
Trace.WriteLine(string.Format(" goodtol: %19.16f\n", goodtol));
debug_time += lanc_seconds() - time;
}
/* Check for convergence. */
time = lanc_seconds();
if (LANCZOS_CONVERGENCE_MODE != 1 || d > 1)
{
/* check convergence of residual bound */
converged = true;
if (j < d)
{
converged = false;
}
else
{
curlnk = scanlist;
while (curlnk != null)
{
if (bj[curlnk->indx] > bji_tol)
{
converged = false;
}
curlnk = curlnk->pntr;
}
}
}
if (LANCZOS_CONVERGENCE_MODE == 1 && d == 1)
{
/* check change in partition */
if (firstpause)
{
converged = true;
if (j < d)
{
converged = false;
}
else
{
curlnk = scanlist;
while (curlnk != null)
{
if (bj[curlnk->indx] > bji_tol)
{
converged = false;
}
curlnk = curlnk->pntr;
}
}
if (!converged)
{
/* compute current approx. to eigenvectors */
i = d;
curlnk = scanlist;
while (curlnk != null)
{
lambda[i] = curlnk->val;
bound[i] = bj[curlnk->indx];
index[i] = curlnk->indx;
curlnk = curlnk->pntr;
i--;
}
for (i = 1; i <= d; i++)
{
Sres = Tevec(alpha, beta - 1, j, lambda[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
if (Sres > SRESTOL)
{
inc_bis_safety = true;
}
setvec(y[i], 1, n, 0.0);
for (k = 1; k <= j; k++)
{
scadd_mixed(y[i], 1, n, s[k], q[k]);
}
}
Assign(A, y, n, d, cube_or_mesh, nsets, vwsqrt, assignment, active, mediantype, goal, vwgt_max);
}
}
else
{
/* copy assignment to old_assignment */
assgn_pntr = assignment;
old_assgn_pntr = old_assignment;
for (i = n + 1; i != 0; i--)
{
*old_assgn_pntr++ = *assgn_pntr++;
}
/* compute current approx. to eigenvectors */
i = d;
curlnk = scanlist;
while (curlnk != null)
{
lambda[i] = curlnk->val;
bound[i] = bj[curlnk->indx];
index[i] = curlnk->indx;
curlnk = curlnk->pntr;
i--;
}
for (i = 1; i <= d; i++)
{
Sres = Tevec(alpha, beta - 1, j, lambda[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
if (Sres > SRESTOL)
{
inc_bis_safety = true;
}
setvec(y[i], 1, n, 0.0);
for (k = 1; k <= j; k++)
{
scadd_mixed(y[i], 1, n, s[k], q[k]);
}
}
/* write new assignment */
Assign(A, y, n, d, cube_or_mesh, nsets, vwsqrt, assignment, active, mediantype, goal, vwgt_max);
assigndiff = 0;
assgn_pntr = assignment;
old_assgn_pntr = old_assignment;
for (i = n + 1; i != 0; i--)
{
if (*old_assgn_pntr++ != *assgn_pntr++)
{
assigndiff++;
}
}
assigndiff = Math.Min(assigndiff, n - assigndiff);
if (DEBUG_EVECS > 1)
{
Trace.WriteLine($" j {j:d}, change from last assignment {assigndiff:d}\n");
}
if (assigndiff <= assigntol)
{
converged = true;
}
else
{
converged = false;
}
}
}
scan_time += lanc_seconds() - time;
/* Show current estimates of evals and bounds (for help in tuning) */
if (DEBUG_EVECS > 2 && !converged)
{
time = lanc_seconds();
/* Collect eigenvalue and bound information for display, return. */
i = d;
curlnk = scanlist;
while (curlnk != null)
{
lambda[i] = curlnk->val;
bound[i] = bj[curlnk->indx];
index[i] = curlnk->indx;
curlnk = curlnk->pntr;
i--;
}
/* Compute eigenvectors and display associated info. */
Trace.WriteLine($"j {j:d;} lambda Ares est. Ares index");
for (i = 1; i <= d; i++)
{
Sres = Tevec(alpha, beta - 1, j, lambda[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
if (Sres > SRESTOL)
{
inc_bis_safety = true;
}
setvec(y[i], 1, n, 0.0);
for (k = 1; k <= j; k++)
{
scadd_mixed(y[i], 1, n, s[k], q[k]);
}
float_to_double(u_double, 1, n, u);
Ares[i] = checkeig(workn_double, A, y[i], n, lambda[i], vwsqrt, u_double);
Trace.Write($"{i:d}.");
doubleout(lambda[i], 1);
doubleout(bound[i], 1);
doubleout(Ares[i], 1);
Trace.WriteLine($" {index[i]:d}");
}
Trace.WriteLine("");
debug_time += lanc_seconds() - time;
}
if (!converged)
{
ngood = 0;
left_ngood = 0; /* for setting left_goodlim on next loop */
right_ngood = 0; /* for setting right_goodlim on next loop */
/* Compute converged Ritz pairs on left end */
time = lanc_seconds();
for (i = 1; i <= left_goodlim; i++)
{
if (bj[i] <= goodtol)
{
ngood += 1;
left_ngood += 1;
if (ngood > maxngood)
{
maxngood = ngood;
solist[ngood] = makeorthlnk_float();
(solist[ngood])->vec = mkvec_float(1, n);
}
(solist[ngood])->index = i;
Sres = Tevec(alpha, beta - 1, j, ritz[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
if (Sres > SRESTOL)
{
inc_bis_safety = true;
}
setvec_float((solist[ngood])->vec, 1, n, 0.0f);
for (k = 1; k <= j; k++)
{
scadd_float((solist[ngood])->vec, 1, n, (float)(s[k]), q[k]);
}
}
}
/* Compute converged Ritz pairs on right end */
for (i = j; i > j - right_goodlim; i--)
{
if (bj[i] <= goodtol)
{
ngood += 1;
right_ngood += 1;
if (ngood > maxngood)
{
maxngood = ngood;
solist[ngood] = makeorthlnk_float();
(solist[ngood])->vec = mkvec_float(1, n);
}
(solist[ngood])->index = i;
Sres = Tevec(alpha, beta - 1, j, ritz[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
if (Sres > SRESTOL)
{
inc_bis_safety = true;
}
setvec_float((solist[ngood])->vec, 1, n, 0.0f);
for (k = 1; k <= j; k++)
{
scadd_float((solist[ngood])->vec, 1, n, (float)(s[k]), q[k]);
}
}
}
ritz_time += lanc_seconds() - time;
if (DEBUG_EVECS > 2)
{
time = lanc_seconds();
Trace.Write($" j {j:d}; goodlim lft {left_goodlim:d}, rgt {right_goodlim:d}; list ");
solistout_float(solist, ngood, j);
Trace.WriteLine("---------------------end of iteration---------------------\n");
debug_time += lanc_seconds() - time;
}
}
}
j++;
}
j--;
/* Collect eigenvalue and bound information. Only compute and display info for
* the eigpairs actually used in the partitioning since don't want to spend the
* time or space to compute the null-space of the Laplacian. */
time = lanc_seconds();
i = d;
curlnk = scanlist;
while (curlnk != null)
{
lambda[i] = curlnk->val;
bound[i] = bj[curlnk->indx];
index[i] = curlnk->indx;
curlnk = curlnk->pntr;
i--;
}
scan_time += lanc_seconds() - time;
/* Compute eigenvectors. */
time = lanc_seconds();
for (i = 1; i <= d; i++)
{
Sres = Tevec(alpha, beta - 1, j, lambda[i], s);
if (Sres > Sres_max)
{
Sres_max = Sres;
}
setvec(y[i], 1, n, 0.0);
for (k = 1; k <= j; k++)
{
scadd_mixed(y[i], 1, n, s[k], q[k]);
}
}
evec_time += lanc_seconds() - time;
time = lanc_seconds();
float_to_double(u_double, 1, n, u);
warnings(workn_double, A, y, n, lambda, vwsqrt, Ares, bound, index, d, j, maxj, Sres_max, eigtol,
u_double, Anorm);
debug_time += lanc_seconds() - time;
/* free up memory */
time = lanc_seconds();
frvec_float(u, 1);
frvec(u_double, 1);
frvec_float(r, 1);
frvec_float(workn, 1);
frvec(workn_double, 1);
frvec(Ares, 0);
Marshal.FreeHGlobal((IntPtr)index);
frvec(alpha, 1);
frvec(beta, 0);
frvec(ritz, 1);
frvec(s, 1);
frvec(bj, 1);
frvec(workj, 0);
for (i = 0; i <= j; i++)
{
frvec_float(q[i], 1);
}
Marshal.FreeHGlobal((IntPtr)q);
if (vwsqrt_float != null)
{
frvec_float(vwsqrt_float, 0);
}
while (scanlist != null)
{
curlnk = scanlist->pntr;
Marshal.FreeHGlobal((IntPtr)scanlist);
scanlist = curlnk;
}
for (i = 1; i <= maxngood; i++)
{
frvec_float((solist[i])->vec, 1);
Marshal.FreeHGlobal((IntPtr)solist[i]);
}
Marshal.FreeHGlobal((IntPtr)solist);
if (LANCZOS_CONVERGENCE_MODE == 1)
{
Marshal.FreeHGlobal((IntPtr)old_assignment);
}
init_time += lanc_seconds() - time;
}
/* Confirm that the bipartite match algorithm did the right thing. */
public static void checkbp(
vtx_data **graph, /* graph data structure for vertex weights */
double *[] xvecs, /* values to partition */
int *sets, /* set assignments returned */
double[] dists, /* distances that separate sets */
int nvtxs, /* number of vertices */
int ndims /* number of dimensions for division */
)
{
int[] signs = new int[MAXDIMS]; /* signs for coordinates of target points */
int[] sizes = new int[MAXSETS]; /* size of each set */
int[] weights = new int[MAXSETS]; /* size of each set */
double setval = 0.0; /* value from assigned set */
double val, bestval = 0.0; /* value to decide set assignment */
const double tolerance = 1.0e-8; /* numerical tolerance */
bool error = false; /* are errors encountered? */
int bestset = -1; /* set vtx should be assigned to */
var nsets = 1 << ndims; /* number of sets */
for (var i = 0; i < nsets; i++)
{
sizes[i] = 0;
weights[i] = 0;
}
for (var i = 1; i <= nvtxs; i++)
{
/* Is vertex closest to the set it is assigned to? */
for (var j = 0; j < MAXDIMS; j++)
{
signs[j] = -1;
}
bestval = 0;
for (var j = 0; j < nsets; j++)
{
val = -dists[j];
for (var k = 1; k <= ndims; k++)
{
val += 2 * signs[k - 1] * xvecs[k][i];
}
if (j == sets[i])
{
setval = val;
}
if (j == 0 || val < bestval)
{
bestval = val;
bestset = j;
}
if (signs[0] == 1 && signs[1] == 1)
{
signs[2] *= -1;
}
if (signs[0] == 1)
{
signs[1] *= -1;
}
signs[0] *= -1;
}
if (Math.Abs(setval - bestval) >= tolerance * (Math.Abs(setval) + Math.Abs(bestval)))
{
Trace.WriteLine($" Vtx {i:d} in set {sets[i]:d} ({setval:e}), but should be in {bestset:d} ({bestval:e})");
error = true;
}
++sizes[sets[i]];
weights[sets[i]] += graph[i]->vwgt;
}
Console.Write(" Sizes:");
for (var i = 0; i < nsets; i++)
{
Console.Write(" {0:d}({1:d})", sizes[i], weights[i]);
}
Trace.WriteLine("");
if (error)
{
throw new InvalidOperationException("ERROR in checkbp");
}
}
