1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
|
From 1e8803e0efe8e469e6a306481e6359c968a7f8b1 Mon Sep 17 00:00:00 2001
From: Nuno Nobre <nuno.nobre@stfc.ac.uk>
Date: Thu, 13 Nov 2025 14:36:14 -0700
Subject: [PATCH] Add sHWPM_CombBLAS.hpp
---
SRC/single/sHWPM_CombBLAS.hpp | 147 ++++++++++++++++++++++++++++++++++
1 file changed, 147 insertions(+)
create mode 100644 SRC/single/sHWPM_CombBLAS.hpp
diff --git a/SRC/single/sHWPM_CombBLAS.hpp b/SRC/single/sHWPM_CombBLAS.hpp
new file mode 100644
index 00000000..20020715
--- /dev/null
+++ b/SRC/single/sHWPM_CombBLAS.hpp
@@ -0,0 +1,147 @@
+/*! \file
+Copyright (c) 2003, The Regents of the University of California, through
+Lawrence Berkeley National Laboratory (subject to receipt of any required
+approvals from U.S. Dept. of Energy)
+
+All rights reserved.
+
+The source code is distributed under BSD license, see the file License.txt
+at the top-level directory.
+*/
+
+/*! @file
+ * \brief Get HWPM, heavy-weight perfect matching.
+ *
+ * <pre>
+ * -- Distributed SuperLU routine (version 9.0) --
+ * Lawrence Berkeley National Lab, Univ. of California Berkeley.
+ * April 2, 2020
+ * </pre>
+ */
+
+#pragma once
+#ifndef sHWPM_CombBLAS_hpp
+#define sHWPM_CombBLAS_hpp
+
+#include "CombBLAS/CombBLAS.h"
+#include "ApproxWeightPerfectMatching.h"
+#include "superlu_sdefs.h"
+
+
+/*! \brief
+ *
+ * <pre>
+ * Purpose
+ * =======
+ * Get perm_r from HWPM, heavy-weight perfect matching, as a
+ * numerical pivoting permutation.
+ *
+ * Arguments
+ * =========
+ *
+ * A (input) SuperMatrix*
+ * The distributed input matrix A of dimension (A->nrow, A->ncol).
+ * A may be overwritten by diag(R)*A*diag(C)*Pc^T.
+ * The type of A can be: Stype = SLU_NR_loc; Dtype = SLU_D; Mtype = SLU_GE.
+ *
+ * perm (input) int_t*
+ * Permutation vector describing the transformation performed to
+ * the original matrix A.
+ *
+ * grid (input) gridinfo_t*
+ * SuperLU's 2D process mesh.
+ *
+ * ScalePermstruct (output) sScalePermstruct_t*
+ * ScalePermstruct->perm_r stores the permutation obtained from HWPM.
+ *
+ * </pre>
+ */
+void
+sGetHWPM(SuperMatrix *A, gridinfo_t *grid, sScalePermstruct_t *ScalePermstruct)
+{
+ NRformat_loc *Astore;
+ int_t i, irow, fst_row, j, jcol, m, n, m_loc;
+ int_t lirow, ljcol;
+ int_t nnz_loc; /* number of local nonzeros */
+ float *nzval_a;
+ int iam, p, procs;
+ int_t *perm=nullptr; // placeholder for load balancing permutation for CombBLAS
+ procs = grid->nprow * grid->npcol;
+
+ if(grid->nprow != grid->npcol)
+ {
+ printf("HWPM only supports square process grid. Retuning without a permutation.\n");
+ }
+ combblas::SpParMat < int_t, float, combblas::SpDCCols<int_t,float> > Adcsc(grid->comm);
+ std::vector< std::vector < std::tuple<int_t,int_t,float> > > data(procs);
+
+ /* ------------------------------------------------------------
+ INITIALIZATION.
+ ------------------------------------------------------------*/
+ iam = grid->iam;
+#if ( DEBUGlevel>=1 )
+ CHECK_MALLOC(iam, "Enter pdCSR_loc_to_2DBlock()");
+#endif
+ Astore = (NRformat_loc *) A->Store;
+ n = A->ncol;
+ m = A->nrow;
+ m_loc = Astore->m_loc;
+ fst_row = Astore->fst_row;
+
+ /* ------------------------------------------------------------
+ FIRST PASS OF A:
+ COUNT THE NUMBER OF NONZEROS TO BE SENT TO EACH PROCESS,
+ THEN ALLOCATE SPACE.
+ Re-distribute A from distributed CSR storage to 2D block storage
+ conforming CombBLAS API.
+ ------------------------------------------------------------*/
+ nzval_a = (float *) Astore->nzval;
+ nnz_loc = 0;
+ for (i = 0; i < m_loc; ++i) {
+ for (j = Astore->rowptr[i]; j < Astore->rowptr[i+1]; ++j) {
+ if(perm != NULL)
+ {
+ irow = perm[i+fst_row]; /* Row number in P*A*P^T */
+ jcol = perm[Astore->colind[j]]; /* Column number in P*A*P^T */
+ }
+ else
+ {
+ irow = i+fst_row;
+ jcol = Astore->colind[j];
+ }
+ p = Adcsc.Owner(m, n , irow, jcol, lirow, ljcol);
+ ++ nnz_loc;
+ data[p].push_back(std::make_tuple(lirow,ljcol,nzval_a[j]));
+
+ }
+ }
+
+ Adcsc.SparseCommon(data, nnz_loc, m, n, std::plus<float>());
+ combblas::FullyDistVec<int_t, int_t> mateRow2Col ( Adcsc.getcommgrid(), m, (int_t) -1);
+ combblas::FullyDistVec<int_t, int_t> mateCol2Row ( Adcsc.getcommgrid(), n, (int_t) -1);
+ combblas::AWPM(Adcsc, mateRow2Col, mateCol2Row,true);
+
+ // now gather the matching vector
+ MPI_Comm World = mateRow2Col.getcommgrid()->GetWorld();
+ int * rdispls = new int[procs];
+ int sendcnt = mateRow2Col.LocArrSize();
+ int * recvcnt = new int[procs];
+ MPI_Allgather(&sendcnt, 1, MPI_INT, recvcnt, 1, MPI_INT, World);
+ rdispls[0] = 0;
+ for(int i=0; i<procs-1; ++i)
+ {
+ rdispls[i+1] = rdispls[i] + recvcnt[i];
+ }
+ int_t *senddata = (int_t *)mateRow2Col.GetLocArr();
+
+ MPI_Allgatherv(senddata, sendcnt, combblas::MPIType<int_t>(), ScalePermstruct->perm_r, recvcnt, rdispls, combblas::MPIType<int_t>(), World);
+
+ delete[] rdispls;
+ delete[] recvcnt;
+
+#if ( DEBUGlevel>=1 )
+ CHECK_MALLOC(iam, "Exit sHWPM_CombBLAS()");
+#endif
+}
+
+#endif /* sHWPM_CombBLAS_hpp */
|
