fdurastante/amg4psblas
1
0
Fork 0
mirror of https://github.com/sfilippone/amg4psblas.git synced 2026-10-06 22:55:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/PolySmooth' into l1aggregation

Browse Source
This commit is contained in:
Cirdans-Home
2024-06-04 12:06:33 +00:00
parent 0f3c3380cb 67594f8b07
commit 5b17e1bbf1
26 changed files with 80 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Make.inc.in
+1 -1
View File
@@ -75,7 +75,7 @@ CDEFINES=$(AMGCDEFINES)
AMGFDEFINES=@AMGFDEFINES@ $(PSBFDEFINES)
FDEFINES=$(AMGFDEFINES)
CXXDEFINES=@AMGCXXDEFINES@
CXXDEFINES=@AMGCXXDEFINES@ $(PSBCXXDEFINES)
@COMPILERULES@
amgprec/impl/aggregator/MatchBoxPC.cpp
+2 -2
View File
@@ -73,8 +73,8 @@ void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
#endif
// Rimosso per tornare al vecchio matching #define OMP
#ifdef OMP
fprintf(stderr,"Warning: using buggy OpenMP matching!\n");
#ifdef OPENMP
//fprintf(stderr,"Warning: using buggy OpenMP matching!\n");
dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(NLVer, NLEdge,
verLocPtr, verLocInd, edgeLocWeight,
verDistance, Mate,
amgprec/impl/aggregator/MatchBoxPC.h
+3 -3
View File
@@ -59,7 +59,7 @@
#include <assert.h>
#include <map>
#include <vector>
#ifdef OMP
#ifdef OPENMP
// OpenMP is included and used if and only if the OpenMP version of the matching
// is required
#include "omp.h"
@@ -178,7 +178,7 @@ extern "C"
#define MilanRealMin MINUS_INFINITY
#endif
#ifdef OMP
#ifdef OPENMP
/* These functions are only used in the experimental OMP implementation, if that
is disabled there is no reason to actually compile or reference them. */
@@ -431,7 +431,7 @@ is disabled there is no reason to actually compile or reference them. */
#endif
#ifndef OMP
#ifndef OPENMP
//Function of find the owner of a ghost vertex using binary search:
inline MilanInt findOwnerOfGhost(MilanLongInt vtxIndex, MilanLongInt *mVerDistance,
MilanInt myRank, MilanInt numProcs);
amgprec/impl/aggregator/algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.cpp
+15 -8
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
// ***********************************************************************
//
// MatchboxP: A C++ library for approximate weighted matching
@@ -126,8 +126,10 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
fflush(stdout);
#endif
MilanLongInt StartIndex = verDistance[myRank]; // The starting vertex owned by the current rank
MilanLongInt EndIndex = verDistance[myRank + 1] - 1; // The ending vertex owned by the current rank
// The starting vertex owned by the current rank
MilanLongInt StartIndex = verDistance[myRank];
// The ending vertex owned by the current rank
MilanLongInt EndIndex = verDistance[myRank + 1] - 1;
MPI_Status computeStatus;
@@ -145,7 +147,8 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
// only one message will be sent in the initialization phase -
// one of: REQUEST/FAILURE/SUCCESS
vector<MilanLongInt> QLocalVtx, QGhostVtx, QMsgType;
vector<MilanInt> QOwner; // Changed by Fabio to be an integer, addresses needs to be integers!
// Changed by Fabio to be an integer, addresses needs to be integers!
vector<MilanInt> QOwner;
MilanLongInt *PCounter = new MilanLongInt[numProcs];
for (int i = 0; i < numProcs; i++)
@@ -153,7 +156,8 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
MilanLongInt NumMessagesBundled = 0;
// TODO when the last computational section will be refactored this could be eliminated
MilanInt ghostOwner = 0; // Changed by Fabio to be an integer, addresses needs to be integers!
// Changed by Fabio to be an integer, addresses needs to be integers!
MilanInt ghostOwner = 0;
MilanLongInt *candidateMate = nullptr;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")NV: " << NLVer << " Edges: " << NLEdge;
@@ -168,9 +172,12 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
MilanLongInt myCard = 0;
// Build the Ghost Vertex Set: Vg
map<MilanLongInt, MilanLongInt> Ghost2LocalMap; // Map each ghost vertex to a local vertex
vector<MilanLongInt> Counter; // Store the edge count for each ghost vertex
MilanLongInt numGhostVertices = 0, numGhostEdges = 0; // Number of Ghost vertices
// Map each ghost vertex to a local vertex
map<MilanLongInt, MilanLongInt> Ghost2LocalMap;
// Store the edge count for each ghost vertex
vector<MilanLongInt> Counter;
// Number of Ghost vertices
MilanLongInt numGhostVertices = 0, numGhostEdges = 0;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")About to compute Ghost Vertices...";
amgprec/impl/aggregator/amg_c_soc1_map_bld.F90
+2 -1
View File
@@ -275,7 +275,8 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk))*nths+kk
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
amgprec/impl/aggregator/amg_c_soc2_map_bld.F90
+2 -1
View File
@@ -309,7 +309,8 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk))*nths+kk
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
amgprec/impl/aggregator/amg_d_soc1_map_bld.F90
+2 -1
View File
@@ -275,7 +275,8 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1_psb_lpk_+locnaggr(kk))*nths+kk
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
amgprec/impl/aggregator/amg_d_soc2_map_bld.F90
+3 -2
View File
@@ -309,8 +309,9 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk))*nths+kk
if (itmp < (bnds(kk)-1_psb_lpk_+locnaggr(kk))) then
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
!$omp end atomic
amgprec/impl/aggregator/amg_s_soc1_map_bld.F90
+2 -1
View File
@@ -275,7 +275,8 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk))*nths+kk
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
amgprec/impl/aggregator/amg_s_soc2_map_bld.F90
+2 -1
View File
@@ -309,7 +309,8 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk))*nths+kk
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
amgprec/impl/aggregator/amg_z_soc1_map_bld.F90
+2 -1
View File
@@ -275,7 +275,8 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk))*nths+kk
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
amgprec/impl/aggregator/amg_z_soc2_map_bld.F90
+2 -1
View File
@@ -309,7 +309,8 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk))*nths+kk
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
amgprec/impl/aggregator/clean.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
// TODO comment
void clean(MilanLongInt NLVer,
amgprec/impl/aggregator/computeCandidateMate.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
/**
* Execute the research fr the Candidate Mate without controlling if the vertices are already matched.
* Returns the vertices with the highest weight
amgprec/impl/aggregator/extractUChunk.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void extractUChunk(
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
amgprec/impl/aggregator/findOwnerOfGhost.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
/// Find the owner of a ghost node:
MilanInt findOwnerOfGhost(MilanLongInt vtxIndex, MilanLongInt *mVerDistance,
MilanInt myRank, MilanInt numProcs)
amgprec/impl/aggregator/initialize.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt StartIndex, MilanLongInt EndIndex,
MilanLongInt *numGhostEdges,
amgprec/impl/aggregator/isAlreadyMatched.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
/**
* //TODO documentation
* @param k
amgprec/impl/aggregator/parallelComputeCandidateMateB.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void PARALLEL_COMPUTE_CANDIDATE_MATE_B(MilanLongInt NLVer,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
amgprec/impl/aggregator/processCrossEdge.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void PROCESS_CROSS_EDGE(MilanLongInt *edge,
MilanLongInt *S)
{
amgprec/impl/aggregator/processExposedVertex.cpp
+28 -25
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
MilanLongInt *candidateMate,
MilanLongInt *verLocInd,
@@ -113,32 +113,35 @@ void PARALLEL_PROCESS_EXPOSED_VERTEX_B(MilanLongInt NLVer,
} // End of if(w >=0)
else {
// This piece of code is executed a really small amount of times
adj11 = verLocPtr[v];
adj12 = verLocPtr[v + 1];
for (k1 = adj11; k1 < adj12; k1++) {
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#pragma omp critical(adjuse)
{
// This piece of code is executed a really small number of times
adj11 = verLocPtr[v];
adj12 = verLocPtr[v + 1];
for (k1 = adj11; k1 < adj12; k1++) {
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
fflush(stdout);
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
fflush(stdout);
#endif
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
privateQLocalVtx.push_back(v + StartIndex);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
PCounter[ghostOwner]++;
privateQLocalVtx.push_back(v + StartIndex);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
}
}
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
amgprec/impl/aggregator/processMatchedVertices.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void processMatchedVertices(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
amgprec/impl/aggregator/processMatchedVerticesAndSendMessages.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
//#define DEBUG_HANG_
void processMatchedVerticesAndSendMessages(
MilanLongInt NLVer,
amgprec/impl/aggregator/processMessages.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
//#define DEBUG_HANG_
void processMessages(
amgprec/impl/aggregator/queueTransfer.cpp
+2 -3
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void queuesTransfer(vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &QLocalVtx,
@@ -17,8 +17,6 @@ void queuesTransfer(vector<MilanLongInt> &U,
U.insert(U.end(), privateU.begin(), privateU.end());
}
privateU.clear();
#pragma omp critical(sendMessageTransfer)
{
@@ -28,6 +26,7 @@ void queuesTransfer(vector<MilanLongInt> &U,
QOwner.insert(QOwner.end(), privateQOwner.begin(), privateQOwner.end());
}
privateU.clear();
privateQLocalVtx.clear();
privateQGhostVtx.clear();
privateQMsgType.clear();
amgprec/impl/aggregator/sendBundledMessages.cpp
+1 -1
View File
@@ -1,5 +1,5 @@
#include "MatchBoxPC.h"
#ifdef OMP
#ifdef OPENMP
void sendBundledMessages(MilanLongInt *numGhostEdges,
MilanInt *BufferSize,
MilanLongInt *Buffer,