Compare commits

...
169 Commits
Author SHA1 Message Date
sfilippone c1e8bc0c57 Do not use OpenMP code for serial version 2024-07-05 14:23:44 +02:00
sfilippone 2f5072166d Switch off OpenMP in certain sections of MatchBOXP 2024-07-05 13:56:52 +02:00
sfilippone 89e2d53e8b Silence debug print. 2024-07-05 13:56:38 +02:00
sfilippone bfe0a32e09 Merge branch 'mboxomp' into PolySmooth 2024-07-05 13:08:56 +02:00
sfilippone e88d176fed Switch off OpenMP in processExposedVertex 2024-07-05 12:10:26 +02:00
sfilippone c96727a97c Merge branch 'PolySmooth' into mboxomp 2024-07-03 11:01:08 +02:00
sfilippone 6362db0cc5 Try improve OpenMP version of matchbox 2024-07-03 11:00:13 +02:00
sfilippone 9239b16175 Merge branch 'PolySmooth' of github.com:sfilippone/amg4psblas into PolySmooth 2024-07-03 10:58:33 +02:00
sfilippone 96a700cb9d Workaround for save_smoothers bug with gcc-13.3.0 2024-07-03 10:58:00 +02:00
sfilippone 41d91120d4 Merge branch 'PolySmooth' into mboxomp 2024-06-11 09:57:53 +02:00
sfilippone 5d20407b15 Merge branch 'PolySmooth' of github.com:sfilippone/amg4psblas into PolySmooth 2024-06-11 09:56:39 +02:00
sfilippone 322e3f65d1 Merge branch 'PolySmooth' of github.com:sfilippone/amg4psblas into PolySmooth 2024-06-11 09:55:07 +02:00
sfilippone 3ff1ad9372 Fix noise in %memory_use() 2024-06-11 09:54:55 +02:00
sfilippone 818ead5878 Try changes for matching 2024-06-11 09:52:11 +02:00
sfilippone 803d311d1c S versions. Take out parallel in a few places 2024-06-06 17:12:31 +02:00
sfilippone 677e4fe6bc Modify MatchBox names with D in preparation for S version 2024-06-05 15:13:18 +02:00
sfilippone 02a83575a2 Reorganize MatchBox (prepare for S OpenMP) 2024-06-05 13:11:41 +02:00
sfilippone cfbec1f6ea Cosmetic changes to MatchBoxPC 2024-06-05 10:38:30 +02:00
sfilippone e11a134a1f Additional tinkering with OpenMP matchbox 2024-06-04 17:39:31 +02:00
sfilippone 6d05120930 Merge branch 'PolySmooth' of github.com:sfilippone/amg4psblas into PolySmooth 2024-06-04 16:56:07 +02:00
sfilippone bd2d1e3b26 Additional OpenMP tinkering in matchboxp 2024-06-04 16:53:52 +02:00
sfilippone 67594f8b07 Fixes for OpenMP 2024-05-30 17:35:15 +02:00
sfilippone 301fb57bb1 Merge branch 'PolySmooth' of github.com:sfilippone/amg4psblas into PolySmooth 2024-05-30 17:26:49 +02:00
sfilippone 13eee99ea3 Use ifdef OPENMP 2024-05-30 17:26:34 +02:00
sfilippone fb802c62cd Merge PSBCXXDEFINES into AMGCXXDEFINES 2024-05-30 17:25:25 +02:00
Salvatore Filippone 897c5229a6 Improve behaviour of OpenMP matching 2024-05-30 08:13:08 -04:00
Salvatore Filippone ab5eaac5ed Cosmetic changes 2024-05-30 08:12:48 -04:00
sfilippone 234071869d Merge branch 'PolySmooth' of github.com:sfilippone/amg4psblas into PolySmooth 2024-05-10 15:03:38 +02:00
sfilippone 3e3b343131 Fix potential overflow issue in SOC_MAP_BLD 2024-05-10 15:02:36 +02:00
Cirdans-Home 5790aa0cbd Revert "First hardcoded implementation of l1 smooth aggregation"
This reverts commit a17f503486.
2024-04-26 10:55:47 +00:00
Cirdans-Home a17f503486 First hardcoded implementation of l1 smooth aggregation 2024-04-26 10:54:43 +00:00
Cirdans-Home 74dccb6c44 Added timers and removed unuseful spmm 2024-04-22 07:52:14 +00:00
sfilippone e83bde6896 New timings 2024-04-17 14:45:20 +02:00
Salvatore Filippone 83d435b49e Default GLOBAL=.true. for MEMORY_USE 2024-03-16 15:22:24 +01:00
Salvatore Filippone af3fda9690 Additional output fixes for memory_use 2024-03-16 14:26:19 +01:00
Salvatore Filippone 678237cf29 Fixed implementation of GLOBAL vs VERBOSITY 2024-03-16 13:31:47 +01:00
Salvatore Filippone 3671285c7a Modified memory_use impl with GLOBAL and VERBOSITY 2024-03-16 13:15:35 +01:00
sfilippone a747cc6abb Defined memory_use method 2024-03-15 15:47:56 +01:00
Cirdans-Home d385d99e71 Fixed Cheby1 Implementation 2024-03-05 18:43:34 +01:00
Salvatore Filippone 4e6e3d5f09 Fix merge conflict 2024-02-21 16:50:55 +01:00
Salvatore Filippone 7c48b96936 Work version of polynomial smoother 2024-02-21 16:45:59 +01:00
Salvatore Filippone 12478a2fff Define COARSE_INVFILL 2024-02-21 16:45:53 +01:00
Cirdans-Home ea8974f88c Fixed build and apply to actually use degree 2024-02-02 19:25:21 +01:00
Cirdans-Home 54d608d2dd Isolated under ifdef buggy matching 2024-02-01 14:50:49 +01:00
sfilippone 47bafd7fe7 Add missing file 2023-11-28 14:00:24 +01:00
sfilippone ccef858192 Cleanup dead code 2023-11-22 14:45:16 +01:00
sfilippone 30a5c7be03 Added POLY smoothers, also in SAMPLES/ADVANCED 2023-11-22 13:33:20 +01:00
sfilippone 737ebb9a96 Test program working 2023-11-21 12:01:07 +01:00
sfilippone dc15b931a0 New test program. 2023-11-21 11:36:13 +01:00
sfilippone 23aabd794d Defined new variant of polynomial smoother. 2023-11-20 11:32:51 +01:00
sfilippone a67454ef5c Prepare for new variant. 2023-11-17 16:01:52 +01:00
sfilippone 79317cb392 Additional fields for rho(BA) estimate. 2023-11-17 15:35:15 +01:00
sfilippone 847ed6ae60 Estimate rho(BA) 2023-11-17 14:56:51 +01:00
sfilippone 6ad82037c5 Add comments in smoother fields 2023-11-17 14:49:20 +01:00
sfilippone bee9d63e9c Take out debug statement 2023-11-17 14:46:22 +01:00
sfilippone bb262275a1 Temporary checkpoint, working version, to be investigated further. 2023-11-17 14:45:18 +01:00
sfilippone 14cd4cde76 Fix inpout file. 2023-11-17 13:16:41 +01:00
sfilippone ec9fcb1bcc Adjustments for POLYNOMIAL smoothers. 2023-11-17 13:14:06 +01:00
sfilippone 2dd1cbd3dc Fix coeff generation. Change name of polynomial coefficients module. 2023-11-17 13:13:30 +01:00
sfilippone ea2f75776c Implement structure for polynomial smoother 2023-11-13 16:58:30 +01:00
sfilippone 1dcb542e4a Include solver_sweeps in test program 2023-10-30 16:10:12 +01:00
sfilippone 84ea60c94c Defined Jacobi and L1-JACOBI solvers. 2023-10-27 15:49:13 +02:00
sfilippone e8b50152fa Fix comments on LOCAL_SOLVER GLOBAL_SOLVER 2023-10-25 15:28:44 +02:00
sfilippone e6894501dd Add setting of MUMPS_LOC_GLOB when using MUMPS at lower levels 2023-10-25 15:28:36 +02:00
sfilippone 975fc6265f Fix spurious free 2023-10-24 18:23:44 +02:00
Salvatore Filippone a97f56d673 Add L1-JACOBI as subsolver. 2023-10-24 18:01:47 +02:00
sfilippone b1f05482a6 Add 'DECOUPLED' as possible choice 2023-10-04 10:59:23 +02:00
sfilippone fb490cee7e Merge branch 'dev-openmp' into development 2023-09-07 10:15:01 +02:00
Salvatore Filippone 24c85c7114 Merged hierarchy_ and smoothers_ free methods from dev 2023-08-31 08:35:27 -04:00
sfilippone 53998a1da9 Fixed out of bound accesses. 2023-08-28 13:25:53 +02:00
sfilippone 0bcc9d7b55 Merge branch 'dev-openmp' into development 2023-08-22 10:38:25 +02:00
sfilippone 11421f53a2 Minor updates on sample output 2023-08-22 10:36:44 +02:00
sfilippone d33bcfe107 Completed SOC2 OpenMP. 2023-08-13 09:52:25 +02:00
sfilippone 5bcd36f394 Fixed SOC1 and SOC2 OpenMP 2023-08-08 09:25:15 +02:00
sfilippone 73495edf09 Finish SOC1 OpenMP 2023-08-07 08:59:32 +02:00
sfilippone 9e82d2e311 Final OMP version of SOC1. 2023-08-04 09:30:32 +02:00
sfilippone c1ecb4ebec Fixed SOC1 and begin work on SOC2 2023-08-03 13:26:24 +02:00
sfilippone e78449d0f5 Prepare for SOC2 OpenMP 2023-07-31 13:26:05 +02:00
sfilippone e3de565b6d Updated commeents in SOC1 2023-07-26 14:47:05 +02:00
sfilippone 7b9c722a1a Fixed OpenMP version of SOC1. 2023-07-26 14:32:13 +02:00
sfilippone 2fd718be6f Updates and measurements for OpenMP build 2023-07-19 15:59:11 +02:00
sfilippone 3a5e73e4c8 adjust NTH in samples/pdegen 2023-06-14 14:05:42 +02:00
sfilippone 494b8b925f OpenMP loop in samples data generation 2023-06-05 11:46:39 +02:00
sfilippone 73e5d49913 Added timers to build phases 2023-06-02 11:37:58 +02:00
Salvatore Filippone a612cea167 Debug for matchboxp 2023-02-10 07:53:04 -05:00
Salvatore Filippone ebe9b45177 Modify MATCHBOXP to fix OpenMP. Performance to be reviewed 2023-02-10 07:50:58 -05:00
Salvatore Filippone 32994c7ce8 Better parameters in matchboxp_mod 2022-12-13 06:44:12 -05:00
Salvatore Filippone d59c9e6c0a Updates towards OpenMP version. 2022-11-22 03:02:51 -05:00
Salvatore Filippone 0d624df346 Merge branch 'petrilli-m' into openmp-match 2022-11-14 04:49:50 -05:00
StefanoPetrilli 6414d3aef3 U and privateU are now vectors 2022-07-23 12:47:43 -05:00
StefanoPetrilli a259e8ab53 extractUChunch optimization 2022-07-23 11:34:43 -05:00
StefanoPetrilli 500403dbda Replaced some staticQueues with vectors for performance reasons 2022-07-23 11:13:21 -05:00
StefanoPetrilli 066c1a5e62 optimization processMatchedVerticesAndSendMessages.cpp 2022-07-23 09:27:35 -05:00
StefanoPetrilli 1ab166b38b Improved performance of processMatchedVerticesAndSendMessages.cpp 2022-07-23 08:24:50 -05:00
StefanoPetrilli 5efee20041 Optimization, replaced all useless atomic with reduction 2022-07-23 05:52:27 -05:00
StefanoPetrilli aa45e2fe93 processMatchedVerticesAndSendMessages.cpp unoptimized 2022-07-23 05:14:26 -05:00
StefanoPetrilli e328f3969c queueTransfer optimization in processMatchedVertices 2022-07-22 07:25:09 -05:00
StefanoPetrilli 9d1a416f99 add rm to exec.sh 2022-07-21 15:45:31 -05:00
StefanoPetrilli 9b065602a8 Fixed race condition in processExposedVertices 2022-07-20 16:24:37 -05:00
StefanoPetrilli abf258e2e8 isAlreadyMatched is now atomic 2022-07-20 15:45:29 -05:00
StefanoPetrilli cdf92ea2b2 processMatchedVerticess add send messages with error 2022-07-20 15:37:29 -05:00
StefanoPetrilli 22d9baf296 isAlreadyMatched substituted with atomic read in one place 2022-07-18 14:11:53 -05:00
StefanoPetrilli 44f174a571 findOwnerOfGhost optimization and refactor 2022-07-17 13:44:58 -05:00
StefanoPetrilli 3e945c75b4 Refactoring, removed all useless Pointer passed in functions 2022-07-17 13:20:49 -05:00
StefanoPetrilli a71fe82752 PROCESS_CROSS_EDGE refactoring 2022-07-17 12:03:48 -05:00
StefanoPetrilli 4f07a70ed1 initialize refactoring 2022-07-17 11:48:52 -05:00
StefanoPetrilli cb660e044d Remoe MateLock 2022-07-17 11:27:17 -05:00
StefanoPetrilli d24c8c2d46 processCrossEdges is now atomic 2022-07-17 09:43:48 -05:00
StefanoPetrilli 9ab54adf3f processMatchedVertices parallelized 2022-07-17 08:59:23 -05:00
StefanoPetrilli 71d4cdc319 processMatchedVertices rollback to critical regions 2022-07-17 06:11:11 -05:00
StefanoPetrilli 1374f21ba8 refactor increment on variables passed by reference in processMatchedVertices.cpp 2022-07-16 13:54:40 -05:00
StefanoPetrilli a9bb6b26fa processMatchedVertices partially working mixed critical and lock version 2022-07-16 11:20:39 -05:00
StefanoPetrilli 561cadee0f parallelQueues working 2022-07-15 07:27:30 -05:00
StefanoPetrilli 5ca78fb871 Refactoring isAlreadyMatched and processCrossEdge 2022-07-14 17:10:36 -05:00
StefanoPetrilli f17082b337 Refactoring: eliminatino of SPtr inside processMessages 2022-07-14 15:27:53 -05:00
StefanoPetrilli 1ea1be33ba Refactoring, eliminated useless passed variables 2022-07-14 15:23:32 -05:00
StefanoPetrilli 47c6f4f2f8 comments 2022-07-13 16:19:52 -05:00
StefanoPetrilli dc1675766f processMessages.cpp further refactoring 2022-07-13 16:19:38 -05:00
StefanoPetrilli ccac816f52 processCrossEdge small refactoring 2022-07-12 13:24:12 -05:00
StefanoPetrilli c7e8193514 omp task in clean.cpp, lock destroy 2022-07-12 12:12:15 -05:00
StefanoPetrilli 36bd3a51a2 Makefile fix 2022-07-11 16:31:58 -05:00
StefanoPetrilli 32777cc15c clean partial refactoring 2022-07-10 11:09:10 -05:00
StefanoPetrilli 64c23f93f8 processMessags partial refactoring, message const refactoring 2022-07-10 10:01:50 -05:00
StefanoPetrilli d19443052d Insert private queue error in processMatchedVertices.cpp 2022-07-10 05:24:31 -05:00
StefanoPetrilli df1e4a4616 PROCESS_CROSS_EDGE refactoring 2022-07-10 04:31:51 -05:00
StefanoPetrilli 3de1e607eb sendBundledMessages refactoring 2022-07-10 03:39:58 -05:00
StefanoPetrilli 9b13aef1ce processMathedVertices refactoring 2022-07-08 13:32:24 -05:00
StefanoPetrilli 6dcae6d0c1 fix private queues in PARALLEL_PROCESS_EXPOSED_VERTEX_B 2022-07-06 15:33:29 -05:00
StefanoPetrilli 63b7602d3a refactoring queueTransfer 2022-07-06 13:12:31 -05:00
StefanoPetrilli b66de7f25c Refactoring PARALLEL_PROCESS_EXPOSED_VERTEX_B 2022-07-06 12:58:00 -05:00
StefanoPetrilli 46047b2202 refactoring parallelComputeCandidateMateB 2022-06-30 16:48:18 -05:00
StefanoPetrilli 7cfe198d0f Format 2022-06-26 10:45:06 -05:00
StefanoPetrilli 1aca17cd44 initialize fix 2022-06-26 10:02:11 -05:00
StefanoPetrilli ea040ae5ee Reformat initialize, refactoring of initialize completed 2022-06-26 04:48:49 -05:00
StefanoPetrilli 7741abd45d Initialize parallelized with task 2022-06-26 04:40:13 -05:00
StefanoPetrilli b5e52d31f5 Refactoring private queues, still not working 2022-06-25 15:25:13 -05:00
StefanoPetrilli deab695294 Refactoring Initialization 2022-06-25 12:10:14 -05:00
StefanoPetrilli a54f084ffb refactoring, initialization 2022-06-25 10:16:30 -05:00
StefanoPetrilli bf0532867d Functions in different files 2022-06-25 08:48:49 -05:00
StefanoPetrilli 2044c5c8eb Merge fix, lock error 2022-06-14 14:47:45 -05:00
StefanoPetrilli f38f3cf09a Merge branch 'tmp' into ompmpi_aggregator_stefano_petrilli 2022-06-14 14:44:24 -05:00
StefanoPetrilli 6fd571ecb2 Lock error 2022-06-14 14:33:31 -05:00
StefanoPetrilli bf35c1659b Further improved critical region U 2022-06-13 16:53:12 -05:00
StefanoPetrilli b2230a6d6d Improved critical region U 2022-06-13 16:09:00 -05:00
StefanoPetrilli 6c20cd7819 PROCESS MATCHED VERTICES draft of parallelization 2022-06-10 15:34:29 -05:00
StefanoPetrilli f921aa47c4 Master region for tempCounter.clear()
(Might have solved stucked runs)
2022-06-08 15:19:57 -05:00
StefanoPetrilli 532701031e Extendend parallel region after SEND PACKET BUNDLE
Nothing parallelizable founded
2022-06-02 09:15:31 -05:00
StefanoPetrilli b079d71f30 Further optimizations PARALLEL_PROCESS_EXPOSED_VERTEX_B 2022-06-02 07:29:21 -05:00
StefanoPetrilli e2ca97ca47 Removed one critical region from PARALLEL_PROCESS_EXPOSED_VERTEX_B 2022-05-31 16:04:56 -05:00
StefanoPetrilli 5bc4f2a080 PROCESS MATCHED VERTICES parallelization improvement 2022-05-30 14:27:26 -05:00
StefanoPetrilli 2c8dc2ffdd PROCESS MATCHED VERTICES parallelization draft 2022-05-30 13:49:34 -05:00
StefanoPetrilli f3d7b3ab5e False sharing fix 2022-05-29 12:01:28 -05:00
StefanoPetrilli 766ef320c2 Refactoring + critical(Mate) 2022-05-29 12:01:24 -05:00
StefanoPetrilli 002239f5b6 False sharing fix 2022-05-22 17:35:08 -05:00
StefanoPetrilli 70b7c4db55 PARALLEL_PROCESS_EXPOSED_VERTEX_B named critical sections 2022-05-22 16:50:07 -05:00
StefanoPetrilli 2cac21b345 fix and reformatting 2022-05-21 11:46:40 -05:00
StefanoPetrilli 6180f29f39 PARALLEL_COMPUTE_CANDIDATE_MATE_B is now paralle and correct 2022-05-21 11:23:39 -05:00
StefanoPetrilli b4bfdd83e5 computeCandidateMate and isAlreadyMatched 2022-05-21 10:22:58 -05:00
StefanoPetrilli 1140669ea7 firstComputeCandidateMate 2022-05-21 07:01:42 -05:00
StefanoPetrilli 919e2a2918 PARALLEL_PROCESS_EXPOSED_VERTEX_B is actually not parallelizable. Atleast not as I was doing. 2022-05-21 05:56:05 -05:00
StefanoPetrilli baffff3d93 Instable PARALLEL_PROCESS_EXPOSED_VERTEX_B 2022-05-09 16:52:03 -05:00
StefanoPetrilli 25a603debe PARALLEL_COMPUTE_CANDIDATE_MATE_B 2022-05-08 15:11:56 -05:00
StefanoPetrilli a20f0d47e7 Solved the static queue out of scope problem 2022-05-08 12:15:46 -05:00
StefanoPetrilli 76e04ee997 The OMP and MPI version is now separated in two different files 2022-05-05 15:57:58 -05:00
StefanoPetrilli 0a8debe43a Single parallel regions with multiple for cycles
Added OMP for testing
2022-05-01 15:26:47 -05:00
StefanoPetrilli 8f6dc5fac2 verGhostPtrInitialization is now parallelized 2022-05-01 06:05:16 -05:00
StefanoPetrilli 7d40fde21d verGhostIndInitialization and Ghost2LocalInitialization cycles parallelization 2022-05-01 05:42:42 -05:00
StefanoPetrilli 1760afbe97 Time tracking in algoDistEdge 2022-05-01 04:47:03 -05:00
StefanoPetrilli 60f90804d5 Time tracking in MatchBox 2022-05-01 04:42:33 -05:00
226 changed files with 21526 additions and 1518 deletions
+1 -1
View File
@@ -75,7 +75,7 @@ CDEFINES=$(AMGCDEFINES)
AMGFDEFINES=@AMGFDEFINES@ $(PSBFDEFINES)
FDEFINES=$(AMGFDEFINES)
CXXDEFINES=@AMGCXXDEFINES@
CXXDEFINES=@AMGCXXDEFINES@ $(PSBCXXDEFINES)
@COMPILERULES@
-1
View File
@@ -1,4 +1,3 @@
AMG4PSBLAS
Algebraic Multigrid Package based on PSBLAS (Parallel Sparse BLAS version 3.8)
+12 -9
View File
@@ -9,9 +9,10 @@ FINCLUDES=$(FMFLAG)$(HERE) $(FMFLAG)$(INCDIR) $(PSBLAS_INCLUDES)
DMODOBJS=amg_d_prec_type.o \
amg_d_inner_mod.o amg_d_ilu_solver.o amg_d_diag_solver.o amg_d_jac_smoother.o amg_d_as_smoother.o \
amg_d_poly_smoother.o amg_d_poly_coeff_mod.o\
amg_d_umf_solver.o amg_d_slu_solver.o amg_d_sludist_solver.o amg_d_id_solver.o\
amg_d_base_solver_mod.o amg_d_base_smoother_mod.o amg_d_onelev_mod.o \
amg_d_gs_solver.o amg_d_mumps_solver.o \
amg_d_gs_solver.o amg_d_mumps_solver.o amg_d_jac_solver.o \
amg_d_base_aggregator_mod.o \
amg_d_dec_aggregator_mod.o amg_d_symdec_aggregator_mod.o \
amg_d_ainv_solver.o amg_d_base_ainv_mod.o \
@@ -20,9 +21,9 @@ DMODOBJS=amg_d_prec_type.o \
SMODOBJS=amg_s_prec_type.o amg_s_ilu_fact_mod.o \
amg_s_inner_mod.o amg_s_ilu_solver.o amg_s_diag_solver.o amg_s_jac_smoother.o amg_s_as_smoother.o \
amg_s_slu_solver.o amg_s_id_solver.o\
amg_s_poly_smoother.o amg_s_slu_solver.o amg_s_id_solver.o\
amg_s_base_solver_mod.o amg_s_base_smoother_mod.o amg_s_onelev_mod.o \
amg_s_gs_solver.o amg_s_mumps_solver.o \
amg_s_gs_solver.o amg_s_mumps_solver.o amg_s_jac_solver.o \
amg_s_base_aggregator_mod.o \
amg_s_dec_aggregator_mod.o amg_s_symdec_aggregator_mod.o \
amg_s_ainv_solver.o amg_s_base_ainv_mod.o \
@@ -33,7 +34,7 @@ ZMODOBJS=amg_z_prec_type.o amg_z_ilu_fact_mod.o \
amg_z_inner_mod.o amg_z_ilu_solver.o amg_z_diag_solver.o amg_z_jac_smoother.o amg_z_as_smoother.o \
amg_z_umf_solver.o amg_z_slu_solver.o amg_z_sludist_solver.o amg_z_id_solver.o\
amg_z_base_solver_mod.o amg_z_base_smoother_mod.o amg_z_onelev_mod.o \
amg_z_gs_solver.o amg_z_mumps_solver.o \
amg_z_gs_solver.o amg_z_mumps_solver.o amg_z_jac_solver.o \
amg_z_base_aggregator_mod.o \
amg_z_dec_aggregator_mod.o amg_z_symdec_aggregator_mod.o \
amg_z_ainv_solver.o amg_z_base_ainv_mod.o \
@@ -43,7 +44,7 @@ CMODOBJS=amg_c_prec_type.o amg_c_ilu_fact_mod.o \
amg_c_inner_mod.o amg_c_ilu_solver.o amg_c_diag_solver.o amg_c_jac_smoother.o amg_c_as_smoother.o \
amg_c_slu_solver.o amg_c_id_solver.o\
amg_c_base_solver_mod.o amg_c_base_smoother_mod.o amg_c_onelev_mod.o \
amg_c_gs_solver.o amg_c_mumps_solver.o \
amg_c_gs_solver.o amg_c_mumps_solver.o amg_c_jac_solver.o \
amg_c_base_aggregator_mod.o \
amg_c_dec_aggregator_mod.o amg_c_symdec_aggregator_mod.o \
amg_c_ainv_solver.o amg_c_base_ainv_mod.o \
@@ -155,7 +156,7 @@ amg_z_base_ainv_mod.o: amg_z_base_solver_mod.o amg_base_ainv_mod.o
amg_s_base_solver_mod.o amg_d_base_solver_mod.o amg_c_base_solver_mod.o amg_z_base_solver_mod.o: amg_base_prec_type.o
amg_d_mumps_solver.o amg_d_gs_solver.o amg_d_id_solver.o amg_d_sludist_solver.o amg_d_slu_solver.o \
amg_d_umf_solver.o amg_d_diag_solver.o amg_d_ilu_solver.o: amg_d_base_solver_mod.o amg_d_prec_type.o
amg_d_umf_solver.o amg_d_diag_solver.o amg_d_ilu_solver.o amg_d_jac_solver.o: amg_d_base_solver_mod.o amg_d_prec_type.o
#amg_d_ilu_fact_mod.o: amg_base_prec_type.o amg_d_base_solver_mod.o
#amg_d_ilu_solver.o amg_d_iluk_fact.o: amg_d_ilu_fact_mod.o
@@ -164,9 +165,11 @@ amg_d_jac_smoother.o: amg_d_diag_solver.o
amg_dprecinit.o amg_dprecset.o: amg_d_diag_solver.o amg_d_ilu_solver.o \
amg_d_umf_solver.o amg_d_as_smoother.o amg_d_jac_smoother.o \
amg_d_id_solver.o amg_d_slu_solver.o amg_d_sludist_solver.o
amg_d_poly_smoother.o: amg_d_base_smoother_mod.o amg_d_poly_coeff_mod.o
amg_s_poly_smoother.o: amg_s_base_smoother_mod.o amg_d_poly_coeff_mod.o
amg_s_mumps_solver.o amg_s_gs_solver.o amg_s_id_solver.o amg_s_slu_solver.o \
amg_s_diag_solver.o amg_s_ilu_solver.o: amg_s_base_solver_mod.o amg_s_prec_type.o
amg_s_diag_solver.o amg_s_ilu_solver.o amg_s_jac_solver.o: amg_s_base_solver_mod.o amg_s_prec_type.o
amg_s_ilu_fact_mod.o: amg_base_prec_type.o amg_s_base_solver_mod.o
amg_s_ilu_solver.o amg_s_iluk_fact.o: amg_s_ilu_fact_mod.o
amg_s_as_smoother.o amg_s_jac_smoother.o: amg_s_base_smoother_mod.o
@@ -176,7 +179,7 @@ amg_sprecinit.o amg_sprecset.o: amg_s_diag_solver.o amg_s_ilu_solver.o \
amg_s_id_solver.o amg_s_slu_solver.o
amg_z_mumps_solver.o amg_z_gs_solver.o amg_z_id_solver.o amg_z_sludist_solver.o amg_z_slu_solver.o \
amg_z_umf_solver.o amg_z_diag_solver.o amg_z_ilu_solver.o: amg_z_base_solver_mod.o amg_z_prec_type.o
amg_z_umf_solver.o amg_z_diag_solver.o amg_z_ilu_solver.o amg_z_jac_solver.o: amg_z_base_solver_mod.o amg_z_prec_type.o
amg_z_ilu_fact_mod.o: amg_base_prec_type.o amg_z_base_solver_mod.o
amg_z_ilu_solver.o amg_z_iluk_fact.o: amg_z_ilu_fact_mod.o
amg_z_as_smoother.o amg_z_jac_smoother.o: amg_z_base_smoother_mod.o
@@ -186,7 +189,7 @@ amg_zprecinit.o amg_zprecset.o: amg_z_diag_solver.o amg_z_ilu_solver.o \
amg_z_id_solver.o amg_z_slu_solver.o amg_z_sludist_solver.o
amg_c_mumps_solver.o amg_c_gs_solver.o amg_c_id_solver.o amg_c_sludist_solver.o amg_c_slu_solver.o \
amg_c_diag_solver.o amg_c_ilu_solver.o: amg_c_base_solver_mod.o amg_c_prec_type.o
amg_c_diag_solver.o amg_c_ilu_solver.o amg_c_jac_solver.o: amg_c_base_solver_mod.o amg_c_prec_type.o
amg_c_ilu_fact_mod.o: amg_base_prec_type.o amg_c_base_solver_mod.o
amg_c_ilu_solver.o amg_c_iluk_fact.o: amg_c_ilu_fact_mod.o
amg_c_as_smoother.o amg_c_jac_smoother.o: amg_c_base_smoother_mod.o
+39 -16
View File
@@ -215,7 +215,8 @@ module amg_base_prec_type
integer(psb_ipk_), parameter :: amg_fbgs_ = 6
integer(psb_ipk_), parameter :: amg_l1_gs_ = 7
integer(psb_ipk_), parameter :: amg_l1_fbgs_ = 8
integer(psb_ipk_), parameter :: amg_max_prec_ = 8
integer(psb_ipk_), parameter :: amg_poly_ = 9
integer(psb_ipk_), parameter :: amg_max_prec_ = 9
!
! Constants for pre/post signaling. Now only used internally
!
@@ -233,9 +234,9 @@ module amg_base_prec_type
integer(psb_ipk_), parameter :: amg_diag_scale_ = amg_slv_delta_+1
integer(psb_ipk_), parameter :: amg_l1_diag_scale_ = amg_slv_delta_+2
integer(psb_ipk_), parameter :: amg_gs_ = amg_slv_delta_+3
! !$ integer(psb_ipk_), parameter :: amg_ilu_n_ = amg_slv_delta_+4
! !$ integer(psb_ipk_), parameter :: amg_milu_n_ = amg_slv_delta_+5
! !$ integer(psb_ipk_), parameter :: amg_ilu_t_ = amg_slv_delta_+6
integer(psb_ipk_), parameter :: amg_ilu_n_ = amg_slv_delta_+4
integer(psb_ipk_), parameter :: amg_milu_n_ = amg_slv_delta_+5
integer(psb_ipk_), parameter :: amg_ilu_t_ = amg_slv_delta_+6
integer(psb_ipk_), parameter :: amg_slu_ = amg_slv_delta_+7
integer(psb_ipk_), parameter :: amg_umf_ = amg_slv_delta_+8
integer(psb_ipk_), parameter :: amg_sludist_ = amg_slv_delta_+9
@@ -319,6 +320,16 @@ module amg_base_prec_type
integer(psb_ipk_), parameter :: amg_distr_mat_ = 0
integer(psb_ipk_), parameter :: amg_repl_mat_ = 1
integer(psb_ipk_), parameter :: amg_max_coarse_mat_ = amg_repl_mat_
!
! Legal values for entry: amg_poly_variant_
!
integer(psb_ipk_), parameter :: amg_poly_lottes_ = 0
integer(psb_ipk_), parameter :: amg_poly_lottes_beta_ = 1
integer(psb_ipk_), parameter :: amg_poly_new_ = 2
integer(psb_ipk_), parameter :: amg_poly_dbg_ = 8
integer(psb_ipk_), parameter :: amg_poly_rho_est_power_ = 0
!
! Legal values for entry: amg_prec_status_
!
@@ -390,12 +401,12 @@ module amg_base_prec_type
& ml_names(0:7)=(/'none ','additive ',&
& 'multiplicative', 'VCycle ','WCycle ',&
& 'KCycle ','KCycleSym ','new ML '/)
character(len=15), parameter :: &
character(len=16), parameter :: &
& amg_fact_names(0:amg_max_sub_solve_)=(/&
& 'none ','Jacobi ',&
& 'L1-Jacobi ','none ','none ',&
& 'none ','none ','L1-GS ',&
& 'L1-FBGS ','none ','Point Jacobi ',&
& 'L1-FBGS ','Polynomial ','none ','Point Jacobi ',&
& 'L1-Jacobi ','Gauss-Seidel ','ILU(n) ',&
& 'MILU(n) ','ILU(t,n) ',&
& 'SuperLU ','UMFPACK LU ',&
@@ -457,12 +468,12 @@ contains
character(len=*), parameter :: name='amg_stringval'
! Local variable
integer :: index_tab
character(len=15) ::string2
character(len=128) ::string2
index_tab=index(string,char(9))
if (index_tab.NE.0) then
string2=string(1:index_tab-1)
string2=string(1:index_tab-1)
else
string2=string
string2=string
endif
select case(psb_toupper(trim(string2)))
case('NONE')
@@ -482,11 +493,11 @@ contains
case('BGS','BWGS')
val = amg_bwgs_
case('ILU')
val = psb_ilu_n_
val = amg_ilu_n_
case('MILU')
val = psb_milu_n_
val = amg_milu_n_
case('ILUT')
val = psb_ilu_t_
val = amg_ilu_t_
case('MUMPS')
val = amg_mumps_
case('UMF')
@@ -557,6 +568,18 @@ contains
val = amg_krm_
case('AS')
val = amg_as_
case('POLY')
val = amg_poly_
case('POLY_LOTTES')
val = amg_poly_lottes_
case('POLY_LOTTES_BETA')
val = amg_poly_lottes_beta_
case('POLY_NEW')
val = amg_poly_new_
case('POLY_DBG')
val = amg_poly_dbg_
case('POLY_RHO_EST_POWER')
val = amg_poly_rho_est_power_
case('A_NORMI')
val = amg_max_norm_
case('USER_CHOICE')
@@ -667,10 +690,10 @@ contains
& ml_names(pm%ml_cycle)
select case (pm%ml_cycle)
case (amg_add_ml_)
write(iout,*) ' Number of smoother sweeps : ',&
write(iout,*) ' Number of smoother sweeps/degree : ',&
& pm%sweeps_pre
case (amg_mult_ml_,amg_vcycle_ml_, amg_wcycle_ml_, amg_kcycle_ml_, amg_kcyclesym_ml_)
write(iout,*) ' Number of smoother sweeps : pre: ',&
write(iout,*) ' Number of smoother sweeps/degree : pre: ',&
& pm%sweeps_pre ,' post: ', pm%sweeps_post
end select
@@ -1036,8 +1059,8 @@ contains
integer(psb_ipk_), intent(in) :: ip
logical :: is_legal_ilu_fact
is_legal_ilu_fact = ((ip==psb_ilu_n_).or.&
& (ip==psb_milu_n_).or.(ip==psb_ilu_t_))
is_legal_ilu_fact = ((ip==amg_ilu_n_).or.&
& (ip==amg_milu_n_).or.(ip==amg_ilu_t_))
return
end function is_legal_ilu_fact
function is_legal_d_omega(ip)
+7 -7
View File
@@ -234,7 +234,7 @@ contains
! Arguments
class(amg_c_ilu_solver_type), intent(inout) :: sv
sv%fact_type = psb_ilu_n_
sv%fact_type = amg_ilu_n_
sv%fill_in = 0
sv%thresh = szero
@@ -255,13 +255,13 @@ contains
info = psb_success_
call amg_check_def(sv%fact_type,&
& 'Factorization',psb_ilu_n_,is_legal_ilu_fact)
& 'Factorization',amg_ilu_n_,is_legal_ilu_fact)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
call amg_check_def(sv%fill_in,&
& 'Level',izero,is_int_non_negative)
case(psb_ilu_t_)
case(amg_ilu_t_)
call amg_check_def(sv%thresh,&
& 'Eps',szero,is_legal_s_fact_thrs)
end select
@@ -439,9 +439,9 @@ contains
write(iout_,*) trim(prefix_), ' Incomplete factorization solver: ',&
& amg_fact_names(sv%fact_type)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
case(psb_ilu_t_)
case(amg_ilu_t_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
write(iout_,*) trim(prefix_), ' Fill threshold :',sv%thresh
end select
@@ -496,7 +496,7 @@ contains
implicit none
integer(psb_ipk_) :: val
val = psb_ilu_n_
val = amg_ilu_n_
end function c_ilu_solver_get_id
function c_ilu_solver_get_wrksize() result(val)
+585
View File
@@ -0,0 +1,585 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
!
! File: amg_c_jac_solver_mod.f90
!
! Module: amg_c_jac_solver_mod
!
! This module defines:
! - the amg_c_jac_solver_type data structure containing the ingredients
! for a local Jacobi iteration. The iterations are local to a process
! (they operate on the block diagonal).
!
!
module amg_c_jac_solver
use amg_c_base_solver_mod
type, extends(amg_c_base_solver_type) :: amg_c_jac_solver_type
type(psb_cspmat_type) :: a
type(psb_c_vect_type), allocatable :: dv
complex(psb_spk_), allocatable :: d(:)
integer(psb_ipk_) :: sweeps
real(psb_spk_) :: eps
contains
procedure, pass(sv) :: dump => amg_c_jac_solver_dmp
procedure, pass(sv) :: check => c_jac_solver_check
procedure, pass(sv) :: clone => amg_c_jac_solver_clone
procedure, pass(sv) :: clone_settings => amg_c_jac_solver_clone_settings
procedure, pass(sv) :: clear_data => amg_c_jac_solver_clear_data
procedure, pass(sv) :: build => amg_c_jac_solver_bld
procedure, pass(sv) :: cnv => amg_c_jac_solver_cnv
procedure, pass(sv) :: apply_v => amg_c_jac_solver_apply_vect
procedure, pass(sv) :: apply_a => amg_c_jac_solver_apply
procedure, pass(sv) :: free => c_jac_solver_free
procedure, pass(sv) :: cseti => c_jac_solver_cseti
procedure, pass(sv) :: csetc => c_jac_solver_csetc
procedure, pass(sv) :: csetr => c_jac_solver_csetr
procedure, pass(sv) :: descr => c_jac_solver_descr
procedure, pass(sv) :: default => c_jac_solver_default
procedure, pass(sv) :: sizeof => c_jac_solver_sizeof
procedure, pass(sv) :: get_nzeros => c_jac_solver_get_nzeros
procedure, nopass :: get_wrksz => c_jac_solver_get_wrksize
procedure, nopass :: get_fmt => c_jac_solver_get_fmt
procedure, nopass :: get_id => c_jac_solver_get_id
procedure, nopass :: is_iterative => c_jac_solver_is_iterative
end type amg_c_jac_solver_type
type, extends(amg_c_jac_solver_type) :: amg_c_l1_jac_solver_type
contains
procedure, pass(sv) :: build => amg_c_l1_jac_solver_bld
procedure, pass(sv) :: descr => c_l1_jac_solver_descr
procedure, nopass :: get_fmt => c_l1_jac_solver_get_fmt
procedure, nopass :: get_id => c_l1_jac_solver_get_id
end type amg_c_l1_jac_solver_type
private :: c_jac_solver_bld, c_jac_solver_apply, &
& c_jac_solver_free, &
& c_jac_solver_descr, c_jac_solver_sizeof, &
& c_jac_solver_default, c_jac_solver_dmp, &
& c_jac_solver_apply_vect, c_jac_solver_get_nzeros, &
& c_jac_solver_get_fmt, c_jac_solver_check,&
& c_jac_solver_is_iterative, &
& c_jac_solver_get_id, c_jac_solver_get_wrksize
interface
subroutine amg_c_jac_solver_apply_vect(alpha,sv,x,beta,y,desc_data,&
& trans,work,wv,info,init,initu)
import :: psb_desc_type, amg_c_jac_solver_type, psb_c_vect_type, psb_spk_, &
& psb_cspmat_type, psb_c_base_sparse_mat, psb_c_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_c_jac_solver_type), intent(inout) :: sv
type(psb_c_vect_type),intent(inout) :: x
type(psb_c_vect_type),intent(inout) :: y
complex(psb_spk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
complex(psb_spk_),target, intent(inout) :: work(:)
type(psb_c_vect_type),intent(inout) :: wv(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
type(psb_c_vect_type),intent(inout), optional :: initu
end subroutine amg_c_jac_solver_apply_vect
end interface
interface
subroutine amg_c_jac_solver_apply(alpha,sv,x,beta,y,desc_data,trans,work,info,init,initu)
import :: psb_desc_type, amg_c_jac_solver_type, psb_c_vect_type, psb_spk_, &
& psb_cspmat_type, psb_c_base_sparse_mat, psb_c_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_c_jac_solver_type), intent(inout) :: sv
complex(psb_spk_),intent(inout) :: x(:)
complex(psb_spk_),intent(inout) :: y(:)
complex(psb_spk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
complex(psb_spk_),target, intent(inout) :: work(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
complex(psb_spk_),intent(inout), optional :: initu(:)
end subroutine amg_c_jac_solver_apply
end interface
interface
subroutine amg_c_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_c_jac_solver_type, psb_c_vect_type, psb_spk_, &
& psb_cspmat_type, psb_c_base_sparse_mat, psb_c_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_cspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_c_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_cspmat_type), intent(in), target, optional :: b
class(psb_c_base_sparse_mat), intent(in), optional :: amold
class(psb_c_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_c_jac_solver_bld
end interface
interface
subroutine amg_c_l1_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_c_l1_jac_solver_type, psb_c_vect_type, psb_spk_, &
& psb_cspmat_type, psb_c_base_sparse_mat, psb_c_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_cspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_c_l1_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_cspmat_type), intent(in), target, optional :: b
class(psb_c_base_sparse_mat), intent(in), optional :: amold
class(psb_c_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_c_l1_jac_solver_bld
end interface
interface
subroutine amg_c_jac_solver_cnv(sv,info,amold,vmold,imold)
import :: amg_c_jac_solver_type, psb_spk_, &
& psb_c_base_sparse_mat, psb_c_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
class(amg_c_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
class(psb_c_base_sparse_mat), intent(in), optional :: amold
class(psb_c_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_c_jac_solver_cnv
end interface
interface
subroutine amg_c_jac_solver_dmp(sv,desc,level,info,prefix,head,solver,global_num)
import :: psb_desc_type, amg_c_jac_solver_type, psb_c_vect_type, psb_spk_, &
& psb_cspmat_type, psb_c_base_sparse_mat, psb_c_base_vect_type, &
& psb_ipk_
implicit none
class(amg_c_jac_solver_type), intent(in) :: sv
type(psb_desc_type), intent(in) :: desc
integer(psb_ipk_), intent(in) :: level
integer(psb_ipk_), intent(out) :: info
character(len=*), intent(in), optional :: prefix, head
logical, optional, intent(in) :: solver, global_num
end subroutine amg_c_jac_solver_dmp
end interface
interface
subroutine amg_c_jac_solver_clone(sv,svout,info)
import :: psb_desc_type, psb_cspmat_type, psb_c_base_sparse_mat, &
& psb_c_vect_type, psb_c_base_vect_type, psb_spk_, &
& amg_c_base_solver_type, amg_c_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
class(amg_c_base_solver_type), allocatable, intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_c_jac_solver_clone
end interface
!!$ interface
!!$ subroutine amg_c_l1_jac_solver_clone(sv,svout,info)
!!$ import :: psb_desc_type, psb_cspmat_type, psb_c_base_sparse_mat, &
!!$ & psb_c_vect_type, psb_c_base_vect_type, psb_spk_, &
!!$ & amg_c_base_solver_type, amg_c_l1_jac_solver_type, psb_ipk_
!!$ Implicit None
!!$
!!$ ! Arguments
!!$ class(amg_c_l1_jac_solver_type), intent(inout) :: sv
!!$ class(amg_c_base_solver_type), allocatable, intent(inout) :: svout
!!$ integer(psb_ipk_), intent(out) :: info
!!$ end subroutine amg_c_l1_jac_solver_clone
!!$ end interface
interface
subroutine amg_c_jac_solver_clone_settings(sv,svout,info)
import :: psb_desc_type, psb_cspmat_type, psb_c_base_sparse_mat, &
& psb_c_vect_type, psb_c_base_vect_type, psb_spk_, &
& amg_c_base_solver_type, amg_c_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
class(amg_c_base_solver_type), intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_c_jac_solver_clone_settings
end interface
interface
subroutine amg_c_jac_solver_clear_data(sv,info)
import :: psb_desc_type, psb_cspmat_type, psb_c_base_sparse_mat, &
& psb_c_vect_type, psb_c_base_vect_type, psb_spk_, &
& amg_c_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
end subroutine amg_c_jac_solver_clear_data
end interface
contains
subroutine c_jac_solver_default(sv)
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
sv%sweeps = ione
sv%eps = dzero
return
end subroutine c_jac_solver_default
subroutine c_jac_solver_check(sv,info)
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='c_jac_solver_check'
call psb_erractionsave(err_act)
info = psb_success_
call amg_check_def(sv%sweeps,&
& 'Jacobi sweeps',ione,is_int_positive)
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine c_jac_solver_check
subroutine c_jac_solver_cseti(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
integer(psb_ipk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='c_jac_solver_cseti'
info = psb_success_
call psb_erractionsave(err_act)
select case(psb_toupper(what))
case('SOLVER_SWEEPS')
sv%sweeps = val
case default
call sv%amg_c_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine c_jac_solver_cseti
subroutine c_jac_solver_csetc(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
character(len=*), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act, ival
character(len=20) :: name='c_jac_solver_csetc'
info = psb_success_
call psb_erractionsave(err_act)
call sv%amg_c_base_solver_type%set(what,val,info,idx=idx)
if (info /= psb_success_) then
info = psb_err_from_subroutine_
call psb_errpush(info, name)
goto 9999
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine c_jac_solver_csetc
subroutine c_jac_solver_csetr(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
real(psb_spk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='c_jac_solver_csetr'
call psb_erractionsave(err_act)
info = psb_success_
select case(psb_toupper(what))
case('SOLVER_EPS')
sv%eps = val
case default
call sv%amg_c_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine c_jac_solver_csetr
subroutine c_jac_solver_free(sv,info)
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='c_jac_solver_free'
call psb_erractionsave(err_act)
info = psb_success_
call sv%a%free()
if (allocated(sv%dv)) then
call sv%dv%free(info)
deallocate(sv%dv)
end if
if (allocated(sv%d)) deallocate(sv%d)
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine c_jac_solver_free
subroutine c_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_c_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_c_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine c_jac_solver_descr
function c_jac_solver_get_nzeros(sv) result(val)
implicit none
! Arguments
class(amg_c_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = 0
val = val + sv%a%get_nzeros()
val = val + sv%dv%get_nrows()
return
end function c_jac_solver_get_nzeros
function c_jac_solver_sizeof(sv) result(val)
implicit none
! Arguments
class(amg_c_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = psb_sizeof_ip
val = val + sv%a%sizeof()
val = val + sv%dv%sizeof()
return
end function c_jac_solver_sizeof
function c_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "Jacobi solver"
end function c_jac_solver_get_fmt
function c_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_jac_
end function c_jac_solver_get_id
!
! If this is true, then the solver needs a starting
! guess. Currently only handled in JAC smoother.
!
function c_jac_solver_is_iterative() result(val)
implicit none
logical :: val
val = .true.
end function c_jac_solver_is_iterative
function c_jac_solver_get_wrksize() result(val)
implicit none
integer(psb_ipk_) :: val
val = 2
end function c_jac_solver_get_wrksize
subroutine c_l1_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_c_l1_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_c_l1_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine c_l1_jac_solver_descr
function c_l1_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "L1-Jacobi solver"
end function c_l1_jac_solver_get_fmt
function c_l1_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_l1_jac_
end function c_l1_jac_solver_get_id
end module amg_c_jac_solver
+2 -1
View File
@@ -78,7 +78,8 @@ module amg_c_mumps_solver
!
! Controls to be set before MUMPS instantiation:
!
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: LOCAL 1==amg_global_solver_: GLOBAL
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: 'LOCAL_SOLVER
! 1==amg_global_solver_: 'GLOBAL_SOLVER'
! IPAR(2) : MUMPS_PRINT_ERR print verbosity (see MUMPS)
! IPAR(3) : MUMPS_SYM 0: non-symmetric 2: symmetric
integer(psb_ipk_), dimension(3) :: ipar
+18
View File
@@ -187,6 +187,7 @@ module amg_c_onelev_mod
procedure, pass(lv) :: clone => c_base_onelev_clone
procedure, pass(lv) :: cnv => amg_c_base_onelev_cnv
procedure, pass(lv) :: descr => amg_c_base_onelev_descr
procedure, pass(lv) :: memory_use => amg_c_base_onelev_memory_use
procedure, pass(lv) :: default => c_base_onelev_default
procedure, pass(lv) :: free => amg_c_base_onelev_free
procedure, pass(lv) :: free_smoothers => amg_c_base_onelev_free_smoothers
@@ -273,6 +274,23 @@ module amg_c_onelev_mod
end subroutine amg_c_base_onelev_descr
end interface
interface
subroutine amg_c_base_onelev_memory_use(lv,il,nl,ilmin,info,iout,verbosity, prefix,global)
import :: psb_cspmat_type, psb_c_vect_type, psb_c_base_vect_type, &
& psb_clinmap_type, psb_spk_, amg_c_onelev_type, &
& psb_ipk_, psb_epk_, psb_desc_type
Implicit None
! Arguments
class(amg_c_onelev_type), intent(in) :: lv
integer(psb_ipk_), intent(in) :: il,nl,ilmin
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_c_base_onelev_memory_use
end interface
interface
subroutine amg_c_base_onelev_cnv(lv,info,amold,vmold,imold)
import :: amg_c_onelev_type, psb_c_base_vect_type, psb_spk_, &
+17
View File
@@ -139,6 +139,7 @@ module amg_c_prec_type
procedure, pass(prec) :: smoothers_build => amg_c_smoothers_bld
procedure, pass(prec) :: smoothers_free => amg_c_smoothers_free
procedure, pass(prec) :: descr => amg_cfile_prec_descr
procedure, pass(prec) :: memory_use => amg_cfile_prec_memory_use
end type amg_cprec_type
private :: amg_c_dump, amg_c_get_compl, amg_c_cmp_compl,&
@@ -170,6 +171,22 @@ module amg_c_prec_type
end subroutine amg_cfile_prec_descr
end interface
interface amg_memory_use
subroutine amg_cfile_prec_memory_use(prec,info,iout,root,verbosity,prefix,global)
import :: amg_cprec_type, psb_ipk_
implicit none
! Arguments
class(amg_cprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_cfile_prec_memory_use
end interface
interface amg_sizeof
module procedure amg_cprec_sizeof
end interface
+7 -7
View File
@@ -234,7 +234,7 @@ contains
! Arguments
class(amg_d_ilu_solver_type), intent(inout) :: sv
sv%fact_type = psb_ilu_n_
sv%fact_type = amg_ilu_n_
sv%fill_in = 0
sv%thresh = dzero
@@ -255,13 +255,13 @@ contains
info = psb_success_
call amg_check_def(sv%fact_type,&
& 'Factorization',psb_ilu_n_,is_legal_ilu_fact)
& 'Factorization',amg_ilu_n_,is_legal_ilu_fact)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
call amg_check_def(sv%fill_in,&
& 'Level',izero,is_int_non_negative)
case(psb_ilu_t_)
case(amg_ilu_t_)
call amg_check_def(sv%thresh,&
& 'Eps',dzero,is_legal_d_fact_thrs)
end select
@@ -439,9 +439,9 @@ contains
write(iout_,*) trim(prefix_), ' Incomplete factorization solver: ',&
& amg_fact_names(sv%fact_type)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
case(psb_ilu_t_)
case(amg_ilu_t_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
write(iout_,*) trim(prefix_), ' Fill threshold :',sv%thresh
end select
@@ -496,7 +496,7 @@ contains
implicit none
integer(psb_ipk_) :: val
val = psb_ilu_n_
val = amg_ilu_n_
end function d_ilu_solver_get_id
function d_ilu_solver_get_wrksize() result(val)
+585
View File
@@ -0,0 +1,585 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
!
! File: amg_d_jac_solver_mod.f90
!
! Module: amg_d_jac_solver_mod
!
! This module defines:
! - the amg_d_jac_solver_type data structure containing the ingredients
! for a local Jacobi iteration. The iterations are local to a process
! (they operate on the block diagonal).
!
!
module amg_d_jac_solver
use amg_d_base_solver_mod
type, extends(amg_d_base_solver_type) :: amg_d_jac_solver_type
type(psb_dspmat_type) :: a
type(psb_d_vect_type), allocatable :: dv
real(psb_dpk_), allocatable :: d(:)
integer(psb_ipk_) :: sweeps
real(psb_dpk_) :: eps
contains
procedure, pass(sv) :: dump => amg_d_jac_solver_dmp
procedure, pass(sv) :: check => d_jac_solver_check
procedure, pass(sv) :: clone => amg_d_jac_solver_clone
procedure, pass(sv) :: clone_settings => amg_d_jac_solver_clone_settings
procedure, pass(sv) :: clear_data => amg_d_jac_solver_clear_data
procedure, pass(sv) :: build => amg_d_jac_solver_bld
procedure, pass(sv) :: cnv => amg_d_jac_solver_cnv
procedure, pass(sv) :: apply_v => amg_d_jac_solver_apply_vect
procedure, pass(sv) :: apply_a => amg_d_jac_solver_apply
procedure, pass(sv) :: free => d_jac_solver_free
procedure, pass(sv) :: cseti => d_jac_solver_cseti
procedure, pass(sv) :: csetc => d_jac_solver_csetc
procedure, pass(sv) :: csetr => d_jac_solver_csetr
procedure, pass(sv) :: descr => d_jac_solver_descr
procedure, pass(sv) :: default => d_jac_solver_default
procedure, pass(sv) :: sizeof => d_jac_solver_sizeof
procedure, pass(sv) :: get_nzeros => d_jac_solver_get_nzeros
procedure, nopass :: get_wrksz => d_jac_solver_get_wrksize
procedure, nopass :: get_fmt => d_jac_solver_get_fmt
procedure, nopass :: get_id => d_jac_solver_get_id
procedure, nopass :: is_iterative => d_jac_solver_is_iterative
end type amg_d_jac_solver_type
type, extends(amg_d_jac_solver_type) :: amg_d_l1_jac_solver_type
contains
procedure, pass(sv) :: build => amg_d_l1_jac_solver_bld
procedure, pass(sv) :: descr => d_l1_jac_solver_descr
procedure, nopass :: get_fmt => d_l1_jac_solver_get_fmt
procedure, nopass :: get_id => d_l1_jac_solver_get_id
end type amg_d_l1_jac_solver_type
private :: d_jac_solver_bld, d_jac_solver_apply, &
& d_jac_solver_free, &
& d_jac_solver_descr, d_jac_solver_sizeof, &
& d_jac_solver_default, d_jac_solver_dmp, &
& d_jac_solver_apply_vect, d_jac_solver_get_nzeros, &
& d_jac_solver_get_fmt, d_jac_solver_check,&
& d_jac_solver_is_iterative, &
& d_jac_solver_get_id, d_jac_solver_get_wrksize
interface
subroutine amg_d_jac_solver_apply_vect(alpha,sv,x,beta,y,desc_data,&
& trans,work,wv,info,init,initu)
import :: psb_desc_type, amg_d_jac_solver_type, psb_d_vect_type, psb_dpk_, &
& psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_d_jac_solver_type), intent(inout) :: sv
type(psb_d_vect_type),intent(inout) :: x
type(psb_d_vect_type),intent(inout) :: y
real(psb_dpk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
real(psb_dpk_),target, intent(inout) :: work(:)
type(psb_d_vect_type),intent(inout) :: wv(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
type(psb_d_vect_type),intent(inout), optional :: initu
end subroutine amg_d_jac_solver_apply_vect
end interface
interface
subroutine amg_d_jac_solver_apply(alpha,sv,x,beta,y,desc_data,trans,work,info,init,initu)
import :: psb_desc_type, amg_d_jac_solver_type, psb_d_vect_type, psb_dpk_, &
& psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_d_jac_solver_type), intent(inout) :: sv
real(psb_dpk_),intent(inout) :: x(:)
real(psb_dpk_),intent(inout) :: y(:)
real(psb_dpk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
real(psb_dpk_),target, intent(inout) :: work(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
real(psb_dpk_),intent(inout), optional :: initu(:)
end subroutine amg_d_jac_solver_apply
end interface
interface
subroutine amg_d_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_d_jac_solver_type, psb_d_vect_type, psb_dpk_, &
& psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_dspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_d_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_dspmat_type), intent(in), target, optional :: b
class(psb_d_base_sparse_mat), intent(in), optional :: amold
class(psb_d_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_d_jac_solver_bld
end interface
interface
subroutine amg_d_l1_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_d_l1_jac_solver_type, psb_d_vect_type, psb_dpk_, &
& psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_dspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_d_l1_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_dspmat_type), intent(in), target, optional :: b
class(psb_d_base_sparse_mat), intent(in), optional :: amold
class(psb_d_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_d_l1_jac_solver_bld
end interface
interface
subroutine amg_d_jac_solver_cnv(sv,info,amold,vmold,imold)
import :: amg_d_jac_solver_type, psb_dpk_, &
& psb_d_base_sparse_mat, psb_d_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
class(amg_d_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
class(psb_d_base_sparse_mat), intent(in), optional :: amold
class(psb_d_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_d_jac_solver_cnv
end interface
interface
subroutine amg_d_jac_solver_dmp(sv,desc,level,info,prefix,head,solver,global_num)
import :: psb_desc_type, amg_d_jac_solver_type, psb_d_vect_type, psb_dpk_, &
& psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type, &
& psb_ipk_
implicit none
class(amg_d_jac_solver_type), intent(in) :: sv
type(psb_desc_type), intent(in) :: desc
integer(psb_ipk_), intent(in) :: level
integer(psb_ipk_), intent(out) :: info
character(len=*), intent(in), optional :: prefix, head
logical, optional, intent(in) :: solver, global_num
end subroutine amg_d_jac_solver_dmp
end interface
interface
subroutine amg_d_jac_solver_clone(sv,svout,info)
import :: psb_desc_type, psb_dspmat_type, psb_d_base_sparse_mat, &
& psb_d_vect_type, psb_d_base_vect_type, psb_dpk_, &
& amg_d_base_solver_type, amg_d_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
class(amg_d_base_solver_type), allocatable, intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_d_jac_solver_clone
end interface
!!$ interface
!!$ subroutine amg_d_l1_jac_solver_clone(sv,svout,info)
!!$ import :: psb_desc_type, psb_dspmat_type, psb_d_base_sparse_mat, &
!!$ & psb_d_vect_type, psb_d_base_vect_type, psb_dpk_, &
!!$ & amg_d_base_solver_type, amg_d_l1_jac_solver_type, psb_ipk_
!!$ Implicit None
!!$
!!$ ! Arguments
!!$ class(amg_d_l1_jac_solver_type), intent(inout) :: sv
!!$ class(amg_d_base_solver_type), allocatable, intent(inout) :: svout
!!$ integer(psb_ipk_), intent(out) :: info
!!$ end subroutine amg_d_l1_jac_solver_clone
!!$ end interface
interface
subroutine amg_d_jac_solver_clone_settings(sv,svout,info)
import :: psb_desc_type, psb_dspmat_type, psb_d_base_sparse_mat, &
& psb_d_vect_type, psb_d_base_vect_type, psb_dpk_, &
& amg_d_base_solver_type, amg_d_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
class(amg_d_base_solver_type), intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_d_jac_solver_clone_settings
end interface
interface
subroutine amg_d_jac_solver_clear_data(sv,info)
import :: psb_desc_type, psb_dspmat_type, psb_d_base_sparse_mat, &
& psb_d_vect_type, psb_d_base_vect_type, psb_dpk_, &
& amg_d_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
end subroutine amg_d_jac_solver_clear_data
end interface
contains
subroutine d_jac_solver_default(sv)
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
sv%sweeps = ione
sv%eps = dzero
return
end subroutine d_jac_solver_default
subroutine d_jac_solver_check(sv,info)
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='d_jac_solver_check'
call psb_erractionsave(err_act)
info = psb_success_
call amg_check_def(sv%sweeps,&
& 'Jacobi sweeps',ione,is_int_positive)
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_jac_solver_check
subroutine d_jac_solver_cseti(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
integer(psb_ipk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='d_jac_solver_cseti'
info = psb_success_
call psb_erractionsave(err_act)
select case(psb_toupper(what))
case('SOLVER_SWEEPS')
sv%sweeps = val
case default
call sv%amg_d_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_jac_solver_cseti
subroutine d_jac_solver_csetc(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
character(len=*), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act, ival
character(len=20) :: name='d_jac_solver_csetc'
info = psb_success_
call psb_erractionsave(err_act)
call sv%amg_d_base_solver_type%set(what,val,info,idx=idx)
if (info /= psb_success_) then
info = psb_err_from_subroutine_
call psb_errpush(info, name)
goto 9999
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_jac_solver_csetc
subroutine d_jac_solver_csetr(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
real(psb_dpk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='d_jac_solver_csetr'
call psb_erractionsave(err_act)
info = psb_success_
select case(psb_toupper(what))
case('SOLVER_EPS')
sv%eps = val
case default
call sv%amg_d_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_jac_solver_csetr
subroutine d_jac_solver_free(sv,info)
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='d_jac_solver_free'
call psb_erractionsave(err_act)
info = psb_success_
call sv%a%free()
if (allocated(sv%dv)) then
call sv%dv%free(info)
deallocate(sv%dv)
end if
if (allocated(sv%d)) deallocate(sv%d)
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_jac_solver_free
subroutine d_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_d_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_d_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_jac_solver_descr
function d_jac_solver_get_nzeros(sv) result(val)
implicit none
! Arguments
class(amg_d_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = 0
val = val + sv%a%get_nzeros()
val = val + sv%dv%get_nrows()
return
end function d_jac_solver_get_nzeros
function d_jac_solver_sizeof(sv) result(val)
implicit none
! Arguments
class(amg_d_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = psb_sizeof_ip
val = val + sv%a%sizeof()
val = val + sv%dv%sizeof()
return
end function d_jac_solver_sizeof
function d_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "Jacobi solver"
end function d_jac_solver_get_fmt
function d_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_jac_
end function d_jac_solver_get_id
!
! If this is true, then the solver needs a starting
! guess. Currently only handled in JAC smoother.
!
function d_jac_solver_is_iterative() result(val)
implicit none
logical :: val
val = .true.
end function d_jac_solver_is_iterative
function d_jac_solver_get_wrksize() result(val)
implicit none
integer(psb_ipk_) :: val
val = 2
end function d_jac_solver_get_wrksize
subroutine d_l1_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_d_l1_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_d_l1_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_l1_jac_solver_descr
function d_l1_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "L1-Jacobi solver"
end function d_l1_jac_solver_get_fmt
function d_l1_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_l1_jac_
end function d_l1_jac_solver_get_id
end module amg_d_jac_solver
+41 -13
View File
@@ -143,9 +143,10 @@ contains
type(psb_ld_coo_sparse_mat) :: tmpcoo
logical :: display_out_, print_out_, reproducible_
logical, parameter :: dump=.false., debug=.false., dump_mate=.false., &
& debug_ilaggr=.false., debug_sync=.false.
& debug_ilaggr=.false., debug_sync=.false., debug_mate=.false.
integer(psb_ipk_), save :: idx_bldmtc=-1, idx_phase1=-1, idx_phase2=-1, idx_phase3=-1
logical, parameter :: do_timings=.true.
integer, parameter :: ilaggr_neginit=-1, ilaggr_nonlocal=-2
ictxt = desc_a%get_ctxt()
call psb_info(ictxt,iam,np)
@@ -187,7 +188,7 @@ contains
call desc_a%l2gip(ilv,info,owned=.false.)
call psb_geall(ilaggr,desc_a,info)
ilaggr = -1
ilaggr = ilaggr_neginit
call psb_geasb(ilaggr,desc_a,info)
nr = a%get_nrows()
nc = a%get_ncols()
@@ -213,7 +214,20 @@ contains
call psb_barrier(ictxt)
if (iam == 0) write(0,*)' out from buildmatching:', info
end if
if (debug_mate) then
block
integer(psb_lpk_), allocatable :: ckmate(:)
allocate(ckmate(nr))
ckmate(1:nr) = mate(1:nr)
call psb_msort(ckmate(1:nr))
do i=1,nr-1
if ((ckmate(i)>0) .and. (ckmate(i) == ckmate(i+1))) then
write(0,*) iam,' Duplicate mate entry at',i,' :',ckmate(i)
end if
end do
end block
end if
if (info == 0) then
if (do_timings) call psb_tic(idx_phase2)
if (debug_sync) then
@@ -259,7 +273,7 @@ contains
cycle
else
if (ilaggr(k) == -1) then
if (ilaggr(k) == ilaggr_neginit) then
wk = w(k)
widx = w(idx)
@@ -267,7 +281,7 @@ contains
nrmagg = wmax*sqrt((wk/wmax)**2+(widx/wmax)**2)
if (nrmagg > epsilon(nrmagg)) then
if (idx <= nr) then
if (ilaggr(idx) == -1) then
if (ilaggr(idx) == ilaggr_neginit) then
! Now, if both vertices are local, the aggregate is local
! (kinda obvious).
nlaggr(iam) = nlaggr(iam) + 1
@@ -275,6 +289,9 @@ contains
ilaggr(idx) = nlaggr(iam)
wtemp(k) = w(k)/nrmagg
wtemp(idx) = w(idx)/nrmagg
else
write(0,*) iam,' Inconsistent mate? ',k,mate(k),idx,&
&mate(idx),ilaggr(idx)
end if
nlpairs = nlpairs+1
else if (idx <= nc) then
@@ -294,7 +311,7 @@ contains
ilaggr(k) = nlaggr(iam)
nlpairs = nlpairs+1
else
ilaggr(k) = -2
ilaggr(k) = ilaggr_nonlocal
end if
else
! Use a statistically unbiased tie-breaking rule,
@@ -309,7 +326,7 @@ contains
ilaggr(k) = nlaggr(iam)
nlpairs = nlpairs+1
else
ilaggr(k) = -2
ilaggr(k) = ilaggr_nonlocal
end if
end if
end if
@@ -325,6 +342,12 @@ contains
nlsingl = nlsingl + 1
end if
end if
if (ilaggr(k) == ilaggr_neginit) then
write(0,*) iam,' Error: no update to ',k,mate(k),&
& abs(w(k)),nrmagg,epsilon(nrmagg),wtemp(k)
end if
else
if (ilaggr(k)<0) write(0,*) 'Strange? ',k,ilaggr(k)
end if
end if
end do
@@ -332,7 +355,7 @@ contains
if (do_timings) call psb_tic(idx_phase3)
! Ok, now compute offsets, gather halo and fix non-local
! aggregates (those where ilaggr == -2)
! aggregates (those where ilaggr == ilaggr_nonlocal)
call psb_sum(ictxt,nlaggr)
ntaggr = sum(nlaggr(0:np-1))
naggrm1 = sum(nlaggr(0:iam-1))
@@ -347,7 +370,7 @@ contains
call psb_halo(wtemp,desc_a,info)
! Cleanup as yet unmarked entries
do k=1,nr
if (ilaggr(k) == -2) then
if (ilaggr(k) == ilaggr_nonlocal) then
idx = mate(k)
if (idx > nr) then
i = ilaggr(idx)
@@ -359,9 +382,14 @@ contains
else
write(0,*) 'Error : unresolved (paired) index ',k,idx,i,nr,nc, ilv(k),ilv(idx)
end if
end if
if (ilaggr(k) <0) then
write(0,*) 'Matchboxp: Funny number: ',k,ilv(k),ilaggr(k),wtemp(k)
else if (ilaggr(k) <0) then
write(0,*) iam,'Matchboxp: Funny number: ',k,ilv(k),ilaggr(k),wtemp(k)
write(0,*) iam,' : : ',nr,nc,mate(k)
if (mate(k) <= nr) then
write(0,*) iam,' : : ',ilaggr(mate(k)),mate(mate(k)),&
& ilv(k),ilv(mate(k)), ilv(mate(mate(k))),ilaggr(mate(mate(k)))
end if
flush(0)
end if
end do
if (debug_sync) then
@@ -414,7 +442,7 @@ contains
end block
if (iam == 0) then
write(0,*) 'Matching statistics: Unmatched nodes ',&
write(0,*) iam,'Matching statistics: Unmatched nodes ',&
& nunmatched,' Singletons:',nlsingl,' Pairs:',nlpairs
end if
+2 -1
View File
@@ -78,7 +78,8 @@ module amg_d_mumps_solver
!
! Controls to be set before MUMPS instantiation:
!
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: LOCAL 1==amg_global_solver_: GLOBAL
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: 'LOCAL_SOLVER
! 1==amg_global_solver_: 'GLOBAL_SOLVER'
! IPAR(2) : MUMPS_PRINT_ERR print verbosity (see MUMPS)
! IPAR(3) : MUMPS_SYM 0: non-symmetric 2: symmetric
integer(psb_ipk_), dimension(3) :: ipar
+18
View File
@@ -188,6 +188,7 @@ module amg_d_onelev_mod
procedure, pass(lv) :: clone => d_base_onelev_clone
procedure, pass(lv) :: cnv => amg_d_base_onelev_cnv
procedure, pass(lv) :: descr => amg_d_base_onelev_descr
procedure, pass(lv) :: memory_use => amg_d_base_onelev_memory_use
procedure, pass(lv) :: default => d_base_onelev_default
procedure, pass(lv) :: free => amg_d_base_onelev_free
procedure, pass(lv) :: free_smoothers => amg_d_base_onelev_free_smoothers
@@ -274,6 +275,23 @@ module amg_d_onelev_mod
end subroutine amg_d_base_onelev_descr
end interface
interface
subroutine amg_d_base_onelev_memory_use(lv,il,nl,ilmin,info,iout,verbosity, prefix,global)
import :: psb_dspmat_type, psb_d_vect_type, psb_d_base_vect_type, &
& psb_dlinmap_type, psb_dpk_, amg_d_onelev_type, &
& psb_ipk_, psb_epk_, psb_desc_type
Implicit None
! Arguments
class(amg_d_onelev_type), intent(in) :: lv
integer(psb_ipk_), intent(in) :: il,nl,ilmin
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_d_base_onelev_memory_use
end interface
interface
subroutine amg_d_base_onelev_cnv(lv,info,amold,vmold,imold)
import :: amg_d_onelev_type, psb_d_base_vect_type, psb_dpk_, &
+548
View File
@@ -0,0 +1,548 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Daniela di Serafino
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
! File: amg_d_poly_smoother_mod.f90
!
! Module: amg_d_poly_smoother_mod
!
! This module defines:
! the amg_d_poly_smoother_type data structure containing the
! smoother for a Jacobi/block Jacobi smoother.
! The smoother stores in ND the block off-diagonal matrix.
! One special case is treated separately, when the solver is DIAG or L1-DIAG
! then the ND is the entire off-diagonal part of the matrix (including the
! main diagonal block), so that it becomes possible to implement
! a pure Jacobi or L1-Jacobi global solver.
!
module amg_d_poly_coeff_mod
use psb_base_mod
real(psb_dpk_), parameter :: amg_d_poly_a_vect(30) = [ &
& 0.3333333333333333_psb_dpk_, &
& 0.1805359927403007_psb_dpk_, &
& 0.1159278464862213_psb_dpk_, &
& 0.0820780659590383_psb_dpk_, &
& 0.0618496002413377_psb_dpk_, &
& 0.0486605823426062_psb_dpk_, &
& 0.0395132986024057_psb_dpk_, &
& 0.0328701017544880_psb_dpk_, &
& 0.0278702862721800_psb_dpk_, &
& 0.0239987409600620_psb_dpk_, &
& 0.0209304400432259_psb_dpk_, &
& 0.0184513099045066_psb_dpk_, &
& 0.0164152586042591_psb_dpk_, &
& 0.0147195638076874_psb_dpk_, &
& 0.0132901324757843_psb_dpk_, &
& 0.0120723317737698_psb_dpk_, &
& 0.0110250964606384_psb_dpk_, &
& 0.0101170330064859_psb_dpk_, &
& 0.0093237789039835_psb_dpk_, &
& 0.0086261728849515_psb_dpk_, &
& 0.0080089618703679_psb_dpk_, &
& 0.0074598709610601_psb_dpk_, &
& 0.0069689238144320_psb_dpk_, &
& 0.0065279387776372_psb_dpk_, &
& 0.0061301503808627_psb_dpk_, &
& 0.0057699215598864_psb_dpk_, &
& 0.0054425224281914_psb_dpk_, &
& 0.0051439584672521_psb_dpk_, &
& 0.0048708358327268_psb_dpk_, &
& 0.0046202548314912_psb_dpk_ ];
real(psb_dpk_), parameter :: amg_d_poly_beta_vect(900) = [ &
& 1.1250000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0238728757031315_psb_dpk_, 1.2640890537108553_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0084254478202830_psb_dpk_, 1.0886783920873087_psb_dpk_, &
& 1.3375312590961856_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0039131042728535_psb_dpk_, 1.0403581118859304_psb_dpk_, &
& 1.1486349854625493_psb_dpk_, 1.3826886924100055_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0021293014616472_psb_dpk_, 1.0217371154926094_psb_dpk_, &
& 1.0787243319260302_psb_dpk_, 1.1981006529266300_psb_dpk_, &
& 1.4132254279168215_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0012851725594023_psb_dpk_, 1.0130429303523338_psb_dpk_, &
& 1.0467821512411335_psb_dpk_, 1.1161648941967548_psb_dpk_, &
& 1.2382902021844453_psb_dpk_, 1.4352429710674484_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0008346439791242_psb_dpk_, 1.0084394943012289_psb_dpk_, &
& 1.0300870776871385_psb_dpk_, 1.0740838409200377_psb_dpk_, &
& 1.1503618670736642_psb_dpk_, 1.2711647404613990_psb_dpk_, &
& 1.4518665864936395_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0005724663119766_psb_dpk_, 1.0057742766241562_psb_dpk_, &
& 1.0205018792294143_psb_dpk_, 1.0501980344456543_psb_dpk_, &
& 1.1011557298494106_psb_dpk_, 1.1808604280685657_psb_dpk_, &
& 1.2983858538257604_psb_dpk_, 1.4648607315109978_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0004096007283281_psb_dpk_, 1.0041243950610661_psb_dpk_, &
& 1.0146021214826659_psb_dpk_, 1.0356111362667175_psb_dpk_, &
& 1.0713997252919425_psb_dpk_, 1.1268827371096291_psb_dpk_, &
& 1.2078521914072933_psb_dpk_, 1.3212193071674674_psb_dpk_, &
& 1.4752964282069962_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0003031222965291_psb_dpk_, 1.0030484066079688_psb_dpk_, &
& 1.0107702271538761_psb_dpk_, 1.0261901159764004_psb_dpk_, &
& 1.0523172493375519_psb_dpk_, 1.0925574320754976_psb_dpk_, &
& 1.1508337666397197_psb_dpk_, 1.2317225087089441_psb_dpk_, &
& 1.3406080202445980_psb_dpk_, 1.4838612440701109_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0002305859520939_psb_dpk_, 1.0023167502402850_psb_dpk_, &
& 1.0081724539630488_psb_dpk_, 1.0198298656634219_psb_dpk_, &
& 1.0395021023532465_psb_dpk_, 1.0696504270054137_psb_dpk_, &
& 1.1130575429574259_psb_dpk_, 1.1729087627556418_psb_dpk_, &
& 1.2528830057679230_psb_dpk_, 1.3572557991951903_psb_dpk_, &
& 1.4910167256413891_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0001794720082837_psb_dpk_, 1.0018018913961957_psb_dpk_, &
& 1.0063486190730762_psb_dpk_, 1.0153786456630600_psb_dpk_, &
& 1.0305694283076039_psb_dpk_, 1.0537601969394355_psb_dpk_, &
& 1.0869986259207296_psb_dpk_, 1.1325918309791341_psb_dpk_, &
& 1.1931627335817252_psb_dpk_, 1.2717129367511055_psb_dpk_, &
& 1.3716933796979953_psb_dpk_, 1.4970841857556243_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0001424192155957_psb_dpk_, 1.0014290693262966_psb_dpk_, &
& 1.0050302898629815_psb_dpk_, 1.0121691051849540_psb_dpk_, &
& 1.0241487434279255_psb_dpk_, 1.0423815888082042_psb_dpk_, &
& 1.0684200812870084_psb_dpk_, 1.1039901093675994_psb_dpk_, &
& 1.1510274824264566_psb_dpk_, 1.2117181191012512_psb_dpk_, &
& 1.2885426486512805_psb_dpk_, 1.3843261938099158_psb_dpk_, &
& 1.5022941875736890_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0001149053826193_psb_dpk_, 1.0011524637691460_psb_dpk_, &
& 1.0040535733326481_psb_dpk_, 1.0097959057315313_psb_dpk_, &
& 1.0194130047299461_psb_dpk_, 1.0340142503543679_psb_dpk_, &
& 1.0548059960662932_psb_dpk_, 1.0831142030181304_psb_dpk_, &
& 1.1204089166089239_psb_dpk_, 1.1683309565544606_psb_dpk_, &
& 1.2287212228823874_psb_dpk_, 1.3036530570781755_psb_dpk_, &
& 1.3954681405367855_psb_dpk_, 1.5068164620958386_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000940475075257_psb_dpk_, 1.0009429169634352_psb_dpk_, &
& 1.0033144905644482_psb_dpk_, 1.0080029483381612_psb_dpk_, &
& 1.0158423625914039_psb_dpk_, 1.0277208331770495_psb_dpk_, &
& 1.0445953542283146_psb_dpk_, 1.0675076120612534_psb_dpk_, &
& 1.0976009254588965_psb_dpk_, 1.1361385536615733_psb_dpk_, &
& 1.1845236142623621_psb_dpk_, 1.2443208730447588_psb_dpk_, &
& 1.3172806908339272_psb_dpk_, 1.4053654389356023_psb_dpk_, &
& 1.5107787250184523_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000779482817921_psb_dpk_, 1.0007812684725339_psb_dpk_, &
& 1.0027448797440124_psb_dpk_, 1.0066229101701514_psb_dpk_, &
& 1.0130985883697137_psb_dpk_, 1.0228944832933697_psb_dpk_, &
& 1.0367832140998394_psb_dpk_, 1.0555987571989653_psb_dpk_, &
& 1.0802484840556024_psb_dpk_, 1.1117260713149764_psb_dpk_, &
& 1.1511254343107276_psb_dpk_, 1.1996558461497355_psb_dpk_, &
& 1.2586584174494597_psb_dpk_, 1.3296241265666493_psb_dpk_, &
& 1.4142136069557629_psb_dpk_, 1.5142789173034623_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000653242183546_psb_dpk_, 1.0006545722939437_psb_dpk_, &
& 1.0022987777448662_psb_dpk_, 1.0055432691173583_psb_dpk_, &
& 1.0109550075016893_psb_dpk_, 1.0191301541168694_psb_dpk_, &
& 1.0307019481191382_psb_dpk_, 1.0463489778000818_psb_dpk_, &
& 1.0668039321569163_psb_dpk_, 1.0928629244731740_psb_dpk_, &
& 1.1253954850882542_psb_dpk_, 1.1653553270075827_psb_dpk_, &
& 1.2137919954743157_psb_dpk_, 1.2718635211544003_psb_dpk_, &
& 1.3408502062615073_psb_dpk_, 1.4221696838526183_psb_dpk_, &
& 1.5173934027630227_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000552858792859_psb_dpk_, 1.0005538659610900_psb_dpk_, &
& 1.0019444166743086_psb_dpk_, 1.0046864301776393_psb_dpk_, &
& 1.0092557508630260_psb_dpk_, 1.0161502674772371_psb_dpk_, &
& 1.0258958148322650_psb_dpk_, 1.0390523408953256_psb_dpk_, &
& 1.0562203973533295_psb_dpk_, 1.0780480145522537_psb_dpk_, &
& 1.1052380250439366_psb_dpk_, 1.1385559038570177_psb_dpk_, &
& 1.1788381980793483_psb_dpk_, 1.2270016234308427_psb_dpk_, &
& 1.2840529112630572_psb_dpk_, 1.3510994958895055_psb_dpk_, &
& 1.4293611393851839_psb_dpk_, 1.5201825990516680_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000472036358790_psb_dpk_, 1.0004728102642675_psb_dpk_, &
& 1.0016593577469159_psb_dpk_, 1.0039976891368516_psb_dpk_, &
& 1.0078911941833455_psb_dpk_, 1.0137601583069535_psb_dpk_, &
& 1.0220462561721002_psb_dpk_, 1.0332172281153209_psb_dpk_, &
& 1.0477717791157513_psb_dpk_, 1.0662447417325256_psb_dpk_, &
& 1.0892125464929936_psb_dpk_, 1.1172990456131733_psb_dpk_, &
& 1.1511817386833911_psb_dpk_, 1.1915984520803475_psb_dpk_, &
& 1.2393545273929878_psb_dpk_, 1.2953305781018039_psb_dpk_, &
& 1.3604908781568688_psb_dpk_, 1.4358924509939206_psb_dpk_, &
& 1.5226949329440265_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000406232569254_psb_dpk_, 1.0004068351374691_psb_dpk_, &
& 1.0014274431564170_psb_dpk_, 1.0034377175807407_psb_dpk_, &
& 1.0067826854070978_psb_dpk_, 1.0118204999571436_psb_dpk_, &
& 1.0189259121271075_psb_dpk_, 1.0284938700470616_psb_dpk_, &
& 1.0409432748132981_psb_dpk_, 1.0567209210598594_psb_dpk_, &
& 1.0763056524407055_psb_dpk_, 1.1002127636100871_psb_dpk_, &
& 1.1289986820268283_psb_dpk_, 1.1632659648787138_psb_dpk_, &
& 1.2036686486408621_psb_dpk_, 1.2509179912601627_psb_dpk_, &
& 1.3057886497146727_psb_dpk_, 1.3691253387497200_psb_dpk_, &
& 1.4418500199624611_psb_dpk_, 1.5249696741164267_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000352114440929_psb_dpk_, 1.0003525892395289_psb_dpk_, &
& 1.0012368357172980_psb_dpk_, 1.0029777430511673_psb_dpk_, &
& 1.0058727830027672_psb_dpk_, 1.0102297507781717_psb_dpk_, &
& 1.0163694815733537_psb_dpk_, 1.0246286588536329_psb_dpk_, &
& 1.0353627340015590_psb_dpk_, 1.0489489776835172_psb_dpk_, &
& 1.0657896841306789_psb_dpk_, 1.0863155505114006_psb_dpk_, &
& 1.1109892546943501_psb_dpk_, 1.1403092559728156_psb_dpk_, &
& 1.1748138447471401_psb_dpk_, 1.2150854687543668_psb_dpk_, &
& 1.2617553651999671_psb_dpk_, 1.3155085300984379_psb_dpk_, &
& 1.3770890582780710_psb_dpk_, 1.4473058898645985_psb_dpk_, &
& 1.5270390016420912_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000307198714835_psb_dpk_, 1.0003075769178242_psb_dpk_, &
& 1.0010787281022711_psb_dpk_, 1.0025963829693492_psb_dpk_, &
& 1.0051188625231162_psb_dpk_, 1.0089126974249720_psb_dpk_, &
& 1.0142547789760521_psb_dpk_, 1.0214345766593154_psb_dpk_, &
& 1.0307564364069204_psb_dpk_, 1.0425419742322541_psb_dpk_, &
& 1.0571325804249445_psb_dpk_, 1.0748920501551993_psb_dpk_, &
& 1.0962093570737961_psb_dpk_, 1.1215015873309027_psb_dpk_, &
& 1.1512170523743910_psb_dpk_, 1.1858385999327761_psb_dpk_, &
& 1.2258871437439198_psb_dpk_, 1.2719254338660289_psb_dpk_, &
& 1.3245620908078453_psb_dpk_, 1.3844559282498121_psb_dpk_, &
& 1.4523205908039656_psb_dpk_, 1.5289295350887884_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000269609460124_psb_dpk_, 1.0002699137181752_psb_dpk_, &
& 1.0009464748475532_psb_dpk_, 1.0022775198638552_psb_dpk_, &
& 1.0044888368184179_psb_dpk_, 1.0078128087804721_psb_dpk_, &
& 1.0124901352066715_psb_dpk_, 1.0187716022931539_psb_dpk_, &
& 1.0269199126829005_psb_dpk_, 1.0372115852204526_psb_dpk_, &
& 1.0499389358225151_psb_dpk_, 1.0654121509688057_psb_dpk_, &
& 1.0839614658147161_psb_dpk_, 1.1059394594887115_psb_dpk_, &
& 1.1317234807654135_psb_dpk_, 1.1617182180038959_psb_dpk_, &
& 1.1963584280123116_psb_dpk_, 1.2361118393501820_psb_dpk_, &
& 1.2814822465106404_psb_dpk_, 1.3330128124440397_psb_dpk_, &
& 1.3912895979940381_psb_dpk_, 1.4569453380258381_psb_dpk_, &
& 1.5306634853375161_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000237911597230_psb_dpk_, 1.0002381585998457_psb_dpk_, &
& 1.0008349974382460_psb_dpk_, 1.0020088476285827_psb_dpk_, &
& 1.0039582343156432_psb_dpk_, 1.0068870298152559_psb_dpk_, &
& 1.0110058445931565_psb_dpk_, 1.0165334547611182_psb_dpk_, &
& 1.0236982737890488_psb_dpk_, 1.0327398763510158_psb_dpk_, &
& 1.0439105824804926_psb_dpk_, 1.0574771105088172_psb_dpk_, &
& 1.0737223076000839_psb_dpk_, 1.0929469670793606_psb_dpk_, &
& 1.1154717421787756_psb_dpk_, 1.1416391663018148_psb_dpk_, &
& 1.1718157904303341_psb_dpk_, 1.2063944488757254_psb_dpk_, &
& 1.2457966652063013_psb_dpk_, 1.2904752108716941_psb_dpk_, &
& 1.3409168297942540_psb_dpk_, 1.3976451430108305_psb_dpk_, &
& 1.4612237483301715_psb_dpk_, 1.5322595309246121_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000210994601235_psb_dpk_, 1.0002111968041199_psb_dpk_, &
& 1.0007403694573151_psb_dpk_, 1.0017808593384865_psb_dpk_, &
& 1.0035081686576977_psb_dpk_, 1.0061021720448531_psb_dpk_, &
& 1.0097482505685551_psb_dpk_, 1.0146384533048582_psb_dpk_, &
& 1.0209726922414943_psb_dpk_, 1.0289599764553270_psb_dpk_, &
& 1.0388196916802268_psb_dpk_, 1.0507829315895938_psb_dpk_, &
& 1.0650938873538003_psb_dpk_, 1.0820113022982043_psb_dpk_, &
& 1.1018099987843295_psb_dpk_, 1.1247824847650900_psb_dpk_, &
& 1.1512406478277994_psb_dpk_, 1.1815175449359154_psb_dpk_, &
& 1.2159692965153148_psb_dpk_, 1.2549770940040335_psb_dpk_, &
& 1.2989493304988182_psb_dpk_, 1.3483238646890843_psb_dpk_, &
& 1.4035704288718982_psb_dpk_, 1.4651931924923849_psb_dpk_, &
& 1.5337334933563860_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000187989989242_psb_dpk_, 1.0001881567984481_psb_dpk_, &
& 1.0006595227084085_psb_dpk_, 1.0015861311895899_psb_dpk_, &
& 1.0031239047778964_psb_dpk_, 1.0054323694760092_psb_dpk_, &
& 1.0086755868504005_psb_dpk_, 1.0130231071421940_psb_dpk_, &
& 1.0186509477893992_psb_dpk_, 1.0257426018654052_psb_dpk_, &
& 1.0344900810652515_psb_dpk_, 1.0450949980170887_psb_dpk_, &
& 1.0577696928624343_psb_dpk_, 1.0727384092356933_psb_dpk_, &
& 1.0902385249817814_psb_dpk_, 1.1105218431816117_psb_dpk_, &
& 1.1338559493090710_psb_dpk_, 1.1605256406217599_psb_dpk_, &
& 1.1908344341913664_psb_dpk_, 1.2251061603103259_psb_dpk_, &
& 1.2636866483695495_psb_dpk_, 1.3069455126904677_psb_dpk_, &
& 1.3552780462128098_psb_dpk_, 1.4091072303921326_psb_dpk_, &
& 1.4688858701459975_psb_dpk_, 1.5350988632115488_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000168211938973_psb_dpk_, 1.0001683505351420_psb_dpk_, &
& 1.0005900360142315_psb_dpk_, 1.0014188084960041_psb_dpk_, &
& 1.0027938311393803_psb_dpk_, 1.0048572584314193_psb_dpk_, &
& 1.0077550080990554_psb_dpk_, 1.0116375492127350_psb_dpk_, &
& 1.0166607098595459_psb_dpk_, 1.0229865078405374_psb_dpk_, &
& 1.0307840079371537_psb_dpk_, 1.0402302093961155_psb_dpk_, &
& 1.0515109674005423_psb_dpk_, 1.0648219524284319_psb_dpk_, &
& 1.0803696515480321_psb_dpk_, 1.0983724158638981_psb_dpk_, &
& 1.1190615585080472_psb_dpk_, 1.1426825077681895_psb_dpk_, &
& 1.1694960201606786_psb_dpk_, 1.1997794584895700_psb_dpk_, &
& 1.2338281401870808_psb_dpk_, 1.2719567615042522_psb_dpk_, &
& 1.3145009034164739_psb_dpk_, 1.3618186254259919_psb_dpk_, &
& 1.4142921537855777_psb_dpk_, 1.4723296710339275_psb_dpk_, &
& 1.5363672141264497_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000151113991291_psb_dpk_, 1.0001512299115287_psb_dpk_, &
& 1.0005299814085029_psb_dpk_, 1.0012742317597600_psb_dpk_, &
& 1.0025087130476142_psb_dpk_, 1.0043606572645858_psb_dpk_, &
& 1.0069604400315522_psb_dpk_, 1.0104422369100252_psb_dpk_, &
& 1.0149446949285030_psb_dpk_, 1.0206116219981500_psb_dpk_, &
& 1.0275926969588451_psb_dpk_, 1.0360442030716124_psb_dpk_, &
& 1.0461297878595799_psb_dpk_, 1.0580212522952626_psb_dpk_, &
& 1.0718993724396861_psb_dpk_, 1.0879547567564958_psb_dpk_, &
& 1.1063887424550545_psb_dpk_, 1.1274143343577541_psb_dpk_, &
& 1.1512571899424711_psb_dpk_, 1.1781566543781672_psb_dpk_, &
& 1.2083668495540898_psb_dpk_, 1.2421578212983135_psb_dpk_, &
& 1.2798167491932815_psb_dpk_, 1.3216492236219661_psb_dpk_, &
& 1.3679805949228399_psb_dpk_, 1.4191573997915068_psb_dpk_, &
& 1.4755488703473389_psb_dpk_, 1.5375485315807513_psb_dpk_, &
& 0.0000000000000000_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000136257096588_psb_dpk_, 1.0001363546506836_psb_dpk_, &
& 1.0004778107488095_psb_dpk_, 1.0011486612681773_psb_dpk_, &
& 1.0022611433613271_psb_dpk_, 1.0039295964948667_psb_dpk_, &
& 1.0062710027404669_psb_dpk_, 1.0094055369479136_psb_dpk_, &
& 1.0134571288503909_psb_dpk_, 1.0185540391932908_psb_dpk_, &
& 1.0248294520252528_psb_dpk_, 1.0324220853457433_psb_dpk_, &
& 1.0414768223656390_psb_dpk_, 1.0521453657079123_psb_dpk_, &
& 1.0645869169533493_psb_dpk_, 1.0789688840227822_psb_dpk_, &
& 1.0954676189818162_psb_dpk_, 1.1142691889576817_psb_dpk_, &
& 1.1355701829701565_psb_dpk_, 1.1595785576006521_psb_dpk_, &
& 1.1865145245551894_psb_dpk_, 1.2166114833191515_psb_dpk_, &
& 1.2501170022543431_psb_dpk_, 1.2872938516530203_psb_dpk_, &
& 1.3284210924391027_psb_dpk_, 1.3737952243949607_psb_dpk_, &
& 1.4237313979931023_psb_dpk_, 1.4785646941265451_psb_dpk_, &
& 1.5386514762605854_psb_dpk_, 0.0000000000000000_psb_dpk_, &
& 1.0000123285767939_psb_dpk_, 1.0001233683396147_psb_dpk_, &
& 1.0004322711781202_psb_dpk_, 1.0010390719329101_psb_dpk_, &
& 1.0020451337350940_psb_dpk_, 1.0035535979966428_psb_dpk_, &
& 1.0056698406248343_psb_dpk_, 1.0085019360540697_psb_dpk_, &
& 1.0121611307132341_psb_dpk_, 1.0167623275769953_psb_dpk_, &
& 1.0224245834847208_psb_dpk_, 1.0292716209515502_psb_dpk_, &
& 1.0374323562422998_psb_dpk_, 1.0470414455308106_psb_dpk_, &
& 1.0582398510249318_psb_dpk_, 1.0711754290010183_psb_dpk_, &
& 1.0860035417614331_psb_dpk_, 1.1028876956049132_psb_dpk_, &
& 1.1220002069820316_psb_dpk_, 1.1435228990979547_psb_dpk_, &
& 1.1676478313209715_psb_dpk_, 1.1945780638597872_psb_dpk_, &
& 1.2245284602839432_psb_dpk_, 1.2577265305821996_psb_dpk_, &
& 1.2944133175813315_psb_dpk_, 1.3348443296857557_psb_dpk_, &
& 1.3792905230439911_psb_dpk_, 1.4280393364047606_psb_dpk_, &
& 1.4813957820911738_psb_dpk_, 1.5396835966986973_psb_dpk_ ]
!!$ [1.1250000000000000_psb_dpk_, 0.0_psb_dpk_, 0.0_psb_dpk__psb_dpk_,,&
!!$ & 1.0238728757031315_psb_dpk_, 1.2640890537108553_psb_dpk_, 0.0_psb_dpk_,&
!!$ & 1.0084254478202830_psb_dpk_, 1.0886783920873087_psb_dpk_, 1.3375312590961856_psb_dpk_]
real(psb_dpk_), parameter :: amg_d_poly_beta_mat(30,30)=reshape(amg_d_poly_beta_vect,[30,30])
end module amg_d_poly_coeff_mod
+374
View File
@@ -0,0 +1,374 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Daniela di Serafino
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
! File: amg_d_poly_smoother_mod.f90
!
! Module: amg_d_poly_smoother_mod
!
! This module defines:
! the amg_d_poly_smoother_type data structure containing the
! smoother for a Jacobi/block Jacobi smoother.
! The smoother stores in ND the block off-diagonal matrix.
! One special case is treated separately, when the solver is DIAG or L1-DIAG
! then the ND is the entire off-diagonal part of the matrix (including the
! main diagonal block), so that it becomes possible to implement
! a pure Jacobi or L1-Jacobi global solver.
!
module amg_d_poly_smoother
use amg_d_base_smoother_mod
use amg_d_poly_coeff_mod
type, extends(amg_d_base_smoother_type) :: amg_d_poly_smoother_type
! The local solver component is inherited from the
! parent type.
! class(amg_d_base_solver_type), allocatable :: sv
!
integer(psb_ipk_) :: pdegree, variant
integer(psb_ipk_) :: rho_estimate=amg_poly_rho_est_power_
integer(psb_ipk_) :: rho_estimate_iterations=10
type(psb_dspmat_type), pointer :: pa => null()
real(psb_dpk_), allocatable :: poly_beta(:)
real(psb_dpk_) :: cf_a = dzero
real(psb_dpk_) :: rho_ba = -done
contains
procedure, pass(sm) :: apply_v => amg_d_poly_smoother_apply_vect
!!$ procedure, pass(sm) :: apply_a => amg_d_poly_smoother_apply
procedure, pass(sm) :: dump => amg_d_poly_smoother_dmp
procedure, pass(sm) :: build => amg_d_poly_smoother_bld
procedure, pass(sm) :: cnv => amg_d_poly_smoother_cnv
procedure, pass(sm) :: clone => amg_d_poly_smoother_clone
procedure, pass(sm) :: clone_settings => amg_d_poly_smoother_clone_settings
procedure, pass(sm) :: clear_data => amg_d_poly_smoother_clear_data
procedure, pass(sm) :: free => d_poly_smoother_free
procedure, pass(sm) :: cseti => amg_d_poly_smoother_cseti
procedure, pass(sm) :: csetc => amg_d_poly_smoother_csetc
procedure, pass(sm) :: csetr => amg_d_poly_smoother_csetr
procedure, pass(sm) :: descr => amg_d_poly_smoother_descr
procedure, pass(sm) :: sizeof => d_poly_smoother_sizeof
procedure, pass(sm) :: default => d_poly_smoother_default
procedure, pass(sm) :: get_nzeros => d_poly_smoother_get_nzeros
procedure, pass(sm) :: get_wrksz => d_poly_smoother_get_wrksize
procedure, nopass :: get_fmt => d_poly_smoother_get_fmt
procedure, nopass :: get_id => d_poly_smoother_get_id
end type amg_d_poly_smoother_type
private :: d_poly_smoother_free, &
& d_poly_smoother_sizeof, d_poly_smoother_get_nzeros, &
& d_poly_smoother_get_fmt, d_poly_smoother_get_id, &
& d_poly_smoother_get_wrksize
interface
subroutine amg_d_poly_smoother_apply_vect(alpha,sm,x,beta,y,desc_data,trans,&
& sweeps,work,wv,info,init,initu)
import :: psb_desc_type, amg_d_poly_smoother_type, psb_d_vect_type, psb_dpk_, &
& psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type,&
& psb_ipk_
type(psb_desc_type), intent(in) :: desc_data
class(amg_d_poly_smoother_type), intent(inout) :: sm
type(psb_d_vect_type),intent(inout) :: x
type(psb_d_vect_type),intent(inout) :: y
real(psb_dpk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
integer(psb_ipk_), intent(in) :: sweeps
real(psb_dpk_),target, intent(inout) :: work(:)
type(psb_d_vect_type),intent(inout) :: wv(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
type(psb_d_vect_type),intent(inout), optional :: initu
end subroutine amg_d_poly_smoother_apply_vect
end interface
!!$ interface
!!$ subroutine amg_d_poly_smoother_apply(alpha,sm,x,beta,y,desc_data,trans,&
!!$ & sweeps,work,info,init,initu)
!!$ import :: psb_desc_type, amg_d_poly_smoother_type, psb_d_vect_type, psb_dpk_, &
!!$ & psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type, &
!!$ & psb_ipk_
!!$ type(psb_desc_type), intent(in) :: desc_data
!!$ class(amg_d_poly_smoother_type), intent(inout) :: sm
!!$ real(psb_dpk_),intent(inout) :: x(:)
!!$ real(psb_dpk_),intent(inout) :: y(:)
!!$ real(psb_dpk_),intent(in) :: alpha,beta
!!$ character(len=1),intent(in) :: trans
!!$ integer(psb_ipk_), intent(in) :: sweeps
!!$ real(psb_dpk_),target, intent(inout) :: work(:)
!!$ integer(psb_ipk_), intent(out) :: info
!!$ character, intent(in), optional :: init
!!$ real(psb_dpk_),intent(inout), optional :: initu(:)
!!$ end subroutine amg_d_poly_smoother_apply
!!$ end interface
!!$
interface
subroutine amg_d_poly_smoother_bld(a,desc_a,sm,info,amold,vmold,imold)
import :: psb_desc_type, amg_d_poly_smoother_type, psb_d_vect_type, psb_dpk_, &
& psb_dspmat_type, psb_d_base_sparse_mat, psb_d_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
type(psb_dspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_d_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
class(psb_d_base_sparse_mat), intent(in), optional :: amold
class(psb_d_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_d_poly_smoother_bld
end interface
interface
subroutine amg_d_poly_smoother_cnv(sm,info,amold,vmold,imold)
import :: amg_d_poly_smoother_type, psb_dpk_, &
& psb_d_base_sparse_mat, psb_d_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
class(amg_d_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
class(psb_d_base_sparse_mat), intent(in), optional :: amold
class(psb_d_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_d_poly_smoother_cnv
end interface
interface
subroutine amg_d_poly_smoother_dmp(sm,desc,level,info,prefix,head,smoother,solver,global_num)
import :: psb_dspmat_type, psb_d_vect_type, psb_d_base_vect_type, &
& psb_dpk_, amg_d_poly_smoother_type, psb_epk_, psb_desc_type, &
& psb_ipk_
implicit none
class(amg_d_poly_smoother_type), intent(in) :: sm
type(psb_desc_type), intent(in) :: desc
integer(psb_ipk_), intent(in) :: level
integer(psb_ipk_), intent(out) :: info
character(len=*), intent(in), optional :: prefix, head
logical, optional, intent(in) :: smoother, solver, global_num
end subroutine amg_d_poly_smoother_dmp
end interface
interface
subroutine amg_d_poly_smoother_clone(sm,smout,info)
import :: amg_d_poly_smoother_type, psb_dpk_, &
& amg_d_base_smoother_type, psb_ipk_
class(amg_d_poly_smoother_type), intent(inout) :: sm
class(amg_d_base_smoother_type), allocatable, intent(inout) :: smout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_d_poly_smoother_clone
end interface
interface
subroutine amg_d_poly_smoother_clone_settings(sm,smout,info)
import :: amg_d_poly_smoother_type, psb_dpk_, &
& amg_d_base_smoother_type, psb_ipk_
class(amg_d_poly_smoother_type), intent(inout) :: sm
class(amg_d_base_smoother_type), allocatable, intent(inout) :: smout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_d_poly_smoother_clone_settings
end interface
interface
subroutine amg_d_poly_smoother_clear_data(sm,info)
import :: amg_d_poly_smoother_type, psb_dpk_, &
& amg_d_base_smoother_type, psb_ipk_
class(amg_d_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
end subroutine amg_d_poly_smoother_clear_data
end interface
interface
subroutine amg_d_poly_smoother_descr(sm,info,iout,coarse,prefix)
import :: amg_d_poly_smoother_type, psb_ipk_
class(amg_d_poly_smoother_type), intent(in) :: sm
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
end subroutine amg_d_poly_smoother_descr
end interface
interface
subroutine amg_d_poly_smoother_cseti(sm,what,val,info,idx)
import :: psb_dspmat_type, psb_d_vect_type, psb_d_base_vect_type, &
& psb_dpk_, amg_d_poly_smoother_type, psb_epk_, psb_desc_type, psb_ipk_
implicit none
class(amg_d_poly_smoother_type), intent(inout) :: sm
character(len=*), intent(in) :: what
integer(psb_ipk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
end subroutine amg_d_poly_smoother_cseti
end interface
interface
subroutine amg_d_poly_smoother_csetc(sm,what,val,info,idx)
import :: psb_dspmat_type, psb_d_vect_type, psb_d_base_vect_type, &
& psb_dpk_, amg_d_poly_smoother_type, psb_epk_, psb_desc_type, psb_ipk_
implicit none
class(amg_d_poly_smoother_type), intent(inout) :: sm
character(len=*), intent(in) :: what
character(len=*), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
end subroutine amg_d_poly_smoother_csetc
end interface
interface
subroutine amg_d_poly_smoother_csetr(sm,what,val,info,idx)
import :: psb_dspmat_type, psb_d_vect_type, psb_d_base_vect_type, &
& psb_dpk_, amg_d_poly_smoother_type, psb_epk_, psb_desc_type, psb_ipk_
implicit none
class(amg_d_poly_smoother_type), intent(inout) :: sm
character(len=*), intent(in) :: what
real(psb_dpk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
end subroutine amg_d_poly_smoother_csetr
end interface
contains
subroutine d_poly_smoother_free(sm,info)
Implicit None
! Arguments
class(amg_d_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='d_poly_smoother_free'
call psb_erractionsave(err_act)
info = psb_success_
if (allocated(sm%sv)) then
call sm%sv%free(info)
if (info == psb_success_) deallocate(sm%sv,stat=info)
if (info /= psb_success_) then
info = psb_err_alloc_dealloc_
call psb_errpush(info,name)
goto 9999
end if
end if
if (allocated(sm%poly_beta)) deallocate(sm%poly_beta)
sm%pa => null()
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine d_poly_smoother_free
function d_poly_smoother_sizeof(sm) result(val)
implicit none
! Arguments
class(amg_d_poly_smoother_type), intent(in) :: sm
integer(psb_epk_) :: val
val = psb_sizeof_dp
if (allocated(sm%sv)) val = val + sm%sv%sizeof()
if (allocated(sm%poly_beta)) val = val + psb_sizeof_dp * size(sm%poly_beta)
return
end function d_poly_smoother_sizeof
subroutine d_poly_smoother_default(sm)
Implicit None
! Arguments
class(amg_d_poly_smoother_type), intent(inout) :: sm
!
! Default: BJAC with no residual check
!
sm%pdegree = 1
sm%rho_ba = -done
sm%variant = amg_poly_lottes_
sm%rho_estimate = amg_poly_rho_est_power_
sm%rho_estimate_iterations = 20
if (allocated(sm%sv)) then
call sm%sv%default()
end if
return
end subroutine d_poly_smoother_default
function d_poly_smoother_get_nzeros(sm) result(val)
implicit none
! Arguments
class(amg_d_poly_smoother_type), intent(in) :: sm
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = 0
if (allocated(sm%sv)) val = val + sm%sv%get_nzeros()
return
end function d_poly_smoother_get_nzeros
function d_poly_smoother_get_wrksize(sm) result(val)
implicit none
class(amg_d_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_) :: val
val = 4
if (allocated(sm%sv)) val = val + sm%sv%get_wrksz()
end function d_poly_smoother_get_wrksize
function d_poly_smoother_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "Polynomial smoother"
end function d_poly_smoother_get_fmt
function d_poly_smoother_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_poly_
end function d_poly_smoother_get_id
end module amg_d_poly_smoother
+17
View File
@@ -139,6 +139,7 @@ module amg_d_prec_type
procedure, pass(prec) :: smoothers_build => amg_d_smoothers_bld
procedure, pass(prec) :: smoothers_free => amg_d_smoothers_free
procedure, pass(prec) :: descr => amg_dfile_prec_descr
procedure, pass(prec) :: memory_use => amg_dfile_prec_memory_use
end type amg_dprec_type
private :: amg_d_dump, amg_d_get_compl, amg_d_cmp_compl,&
@@ -170,6 +171,22 @@ module amg_d_prec_type
end subroutine amg_dfile_prec_descr
end interface
interface amg_memory_use
subroutine amg_dfile_prec_memory_use(prec,info,iout,root,verbosity,prefix,global)
import :: amg_dprec_type, psb_ipk_
implicit none
! Arguments
class(amg_dprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_dfile_prec_memory_use
end interface
interface amg_sizeof
module procedure amg_dprec_sizeof
end interface
+7 -7
View File
@@ -234,7 +234,7 @@ contains
! Arguments
class(amg_s_ilu_solver_type), intent(inout) :: sv
sv%fact_type = psb_ilu_n_
sv%fact_type = amg_ilu_n_
sv%fill_in = 0
sv%thresh = szero
@@ -255,13 +255,13 @@ contains
info = psb_success_
call amg_check_def(sv%fact_type,&
& 'Factorization',psb_ilu_n_,is_legal_ilu_fact)
& 'Factorization',amg_ilu_n_,is_legal_ilu_fact)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
call amg_check_def(sv%fill_in,&
& 'Level',izero,is_int_non_negative)
case(psb_ilu_t_)
case(amg_ilu_t_)
call amg_check_def(sv%thresh,&
& 'Eps',szero,is_legal_s_fact_thrs)
end select
@@ -439,9 +439,9 @@ contains
write(iout_,*) trim(prefix_), ' Incomplete factorization solver: ',&
& amg_fact_names(sv%fact_type)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
case(psb_ilu_t_)
case(amg_ilu_t_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
write(iout_,*) trim(prefix_), ' Fill threshold :',sv%thresh
end select
@@ -496,7 +496,7 @@ contains
implicit none
integer(psb_ipk_) :: val
val = psb_ilu_n_
val = amg_ilu_n_
end function s_ilu_solver_get_id
function s_ilu_solver_get_wrksize() result(val)
+585
View File
@@ -0,0 +1,585 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
!
! File: amg_s_jac_solver_mod.f90
!
! Module: amg_s_jac_solver_mod
!
! This module defines:
! - the amg_s_jac_solver_type data structure containing the ingredients
! for a local Jacobi iteration. The iterations are local to a process
! (they operate on the block diagonal).
!
!
module amg_s_jac_solver
use amg_s_base_solver_mod
type, extends(amg_s_base_solver_type) :: amg_s_jac_solver_type
type(psb_sspmat_type) :: a
type(psb_s_vect_type), allocatable :: dv
real(psb_spk_), allocatable :: d(:)
integer(psb_ipk_) :: sweeps
real(psb_spk_) :: eps
contains
procedure, pass(sv) :: dump => amg_s_jac_solver_dmp
procedure, pass(sv) :: check => s_jac_solver_check
procedure, pass(sv) :: clone => amg_s_jac_solver_clone
procedure, pass(sv) :: clone_settings => amg_s_jac_solver_clone_settings
procedure, pass(sv) :: clear_data => amg_s_jac_solver_clear_data
procedure, pass(sv) :: build => amg_s_jac_solver_bld
procedure, pass(sv) :: cnv => amg_s_jac_solver_cnv
procedure, pass(sv) :: apply_v => amg_s_jac_solver_apply_vect
procedure, pass(sv) :: apply_a => amg_s_jac_solver_apply
procedure, pass(sv) :: free => s_jac_solver_free
procedure, pass(sv) :: cseti => s_jac_solver_cseti
procedure, pass(sv) :: csetc => s_jac_solver_csetc
procedure, pass(sv) :: csetr => s_jac_solver_csetr
procedure, pass(sv) :: descr => s_jac_solver_descr
procedure, pass(sv) :: default => s_jac_solver_default
procedure, pass(sv) :: sizeof => s_jac_solver_sizeof
procedure, pass(sv) :: get_nzeros => s_jac_solver_get_nzeros
procedure, nopass :: get_wrksz => s_jac_solver_get_wrksize
procedure, nopass :: get_fmt => s_jac_solver_get_fmt
procedure, nopass :: get_id => s_jac_solver_get_id
procedure, nopass :: is_iterative => s_jac_solver_is_iterative
end type amg_s_jac_solver_type
type, extends(amg_s_jac_solver_type) :: amg_s_l1_jac_solver_type
contains
procedure, pass(sv) :: build => amg_s_l1_jac_solver_bld
procedure, pass(sv) :: descr => s_l1_jac_solver_descr
procedure, nopass :: get_fmt => s_l1_jac_solver_get_fmt
procedure, nopass :: get_id => s_l1_jac_solver_get_id
end type amg_s_l1_jac_solver_type
private :: s_jac_solver_bld, s_jac_solver_apply, &
& s_jac_solver_free, &
& s_jac_solver_descr, s_jac_solver_sizeof, &
& s_jac_solver_default, s_jac_solver_dmp, &
& s_jac_solver_apply_vect, s_jac_solver_get_nzeros, &
& s_jac_solver_get_fmt, s_jac_solver_check,&
& s_jac_solver_is_iterative, &
& s_jac_solver_get_id, s_jac_solver_get_wrksize
interface
subroutine amg_s_jac_solver_apply_vect(alpha,sv,x,beta,y,desc_data,&
& trans,work,wv,info,init,initu)
import :: psb_desc_type, amg_s_jac_solver_type, psb_s_vect_type, psb_spk_, &
& psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_s_jac_solver_type), intent(inout) :: sv
type(psb_s_vect_type),intent(inout) :: x
type(psb_s_vect_type),intent(inout) :: y
real(psb_spk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
real(psb_spk_),target, intent(inout) :: work(:)
type(psb_s_vect_type),intent(inout) :: wv(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
type(psb_s_vect_type),intent(inout), optional :: initu
end subroutine amg_s_jac_solver_apply_vect
end interface
interface
subroutine amg_s_jac_solver_apply(alpha,sv,x,beta,y,desc_data,trans,work,info,init,initu)
import :: psb_desc_type, amg_s_jac_solver_type, psb_s_vect_type, psb_spk_, &
& psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_s_jac_solver_type), intent(inout) :: sv
real(psb_spk_),intent(inout) :: x(:)
real(psb_spk_),intent(inout) :: y(:)
real(psb_spk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
real(psb_spk_),target, intent(inout) :: work(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
real(psb_spk_),intent(inout), optional :: initu(:)
end subroutine amg_s_jac_solver_apply
end interface
interface
subroutine amg_s_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_s_jac_solver_type, psb_s_vect_type, psb_spk_, &
& psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_sspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_s_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_sspmat_type), intent(in), target, optional :: b
class(psb_s_base_sparse_mat), intent(in), optional :: amold
class(psb_s_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_s_jac_solver_bld
end interface
interface
subroutine amg_s_l1_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_s_l1_jac_solver_type, psb_s_vect_type, psb_spk_, &
& psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_sspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_s_l1_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_sspmat_type), intent(in), target, optional :: b
class(psb_s_base_sparse_mat), intent(in), optional :: amold
class(psb_s_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_s_l1_jac_solver_bld
end interface
interface
subroutine amg_s_jac_solver_cnv(sv,info,amold,vmold,imold)
import :: amg_s_jac_solver_type, psb_spk_, &
& psb_s_base_sparse_mat, psb_s_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
class(amg_s_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
class(psb_s_base_sparse_mat), intent(in), optional :: amold
class(psb_s_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_s_jac_solver_cnv
end interface
interface
subroutine amg_s_jac_solver_dmp(sv,desc,level,info,prefix,head,solver,global_num)
import :: psb_desc_type, amg_s_jac_solver_type, psb_s_vect_type, psb_spk_, &
& psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type, &
& psb_ipk_
implicit none
class(amg_s_jac_solver_type), intent(in) :: sv
type(psb_desc_type), intent(in) :: desc
integer(psb_ipk_), intent(in) :: level
integer(psb_ipk_), intent(out) :: info
character(len=*), intent(in), optional :: prefix, head
logical, optional, intent(in) :: solver, global_num
end subroutine amg_s_jac_solver_dmp
end interface
interface
subroutine amg_s_jac_solver_clone(sv,svout,info)
import :: psb_desc_type, psb_sspmat_type, psb_s_base_sparse_mat, &
& psb_s_vect_type, psb_s_base_vect_type, psb_spk_, &
& amg_s_base_solver_type, amg_s_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
class(amg_s_base_solver_type), allocatable, intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_s_jac_solver_clone
end interface
!!$ interface
!!$ subroutine amg_s_l1_jac_solver_clone(sv,svout,info)
!!$ import :: psb_desc_type, psb_sspmat_type, psb_s_base_sparse_mat, &
!!$ & psb_s_vect_type, psb_s_base_vect_type, psb_spk_, &
!!$ & amg_s_base_solver_type, amg_s_l1_jac_solver_type, psb_ipk_
!!$ Implicit None
!!$
!!$ ! Arguments
!!$ class(amg_s_l1_jac_solver_type), intent(inout) :: sv
!!$ class(amg_s_base_solver_type), allocatable, intent(inout) :: svout
!!$ integer(psb_ipk_), intent(out) :: info
!!$ end subroutine amg_s_l1_jac_solver_clone
!!$ end interface
interface
subroutine amg_s_jac_solver_clone_settings(sv,svout,info)
import :: psb_desc_type, psb_sspmat_type, psb_s_base_sparse_mat, &
& psb_s_vect_type, psb_s_base_vect_type, psb_spk_, &
& amg_s_base_solver_type, amg_s_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
class(amg_s_base_solver_type), intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_s_jac_solver_clone_settings
end interface
interface
subroutine amg_s_jac_solver_clear_data(sv,info)
import :: psb_desc_type, psb_sspmat_type, psb_s_base_sparse_mat, &
& psb_s_vect_type, psb_s_base_vect_type, psb_spk_, &
& amg_s_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
end subroutine amg_s_jac_solver_clear_data
end interface
contains
subroutine s_jac_solver_default(sv)
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
sv%sweeps = ione
sv%eps = dzero
return
end subroutine s_jac_solver_default
subroutine s_jac_solver_check(sv,info)
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='s_jac_solver_check'
call psb_erractionsave(err_act)
info = psb_success_
call amg_check_def(sv%sweeps,&
& 'Jacobi sweeps',ione,is_int_positive)
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_jac_solver_check
subroutine s_jac_solver_cseti(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
integer(psb_ipk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='s_jac_solver_cseti'
info = psb_success_
call psb_erractionsave(err_act)
select case(psb_toupper(what))
case('SOLVER_SWEEPS')
sv%sweeps = val
case default
call sv%amg_s_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_jac_solver_cseti
subroutine s_jac_solver_csetc(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
character(len=*), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act, ival
character(len=20) :: name='s_jac_solver_csetc'
info = psb_success_
call psb_erractionsave(err_act)
call sv%amg_s_base_solver_type%set(what,val,info,idx=idx)
if (info /= psb_success_) then
info = psb_err_from_subroutine_
call psb_errpush(info, name)
goto 9999
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_jac_solver_csetc
subroutine s_jac_solver_csetr(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
real(psb_spk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='s_jac_solver_csetr'
call psb_erractionsave(err_act)
info = psb_success_
select case(psb_toupper(what))
case('SOLVER_EPS')
sv%eps = val
case default
call sv%amg_s_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_jac_solver_csetr
subroutine s_jac_solver_free(sv,info)
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='s_jac_solver_free'
call psb_erractionsave(err_act)
info = psb_success_
call sv%a%free()
if (allocated(sv%dv)) then
call sv%dv%free(info)
deallocate(sv%dv)
end if
if (allocated(sv%d)) deallocate(sv%d)
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_jac_solver_free
subroutine s_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_s_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_s_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_jac_solver_descr
function s_jac_solver_get_nzeros(sv) result(val)
implicit none
! Arguments
class(amg_s_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = 0
val = val + sv%a%get_nzeros()
val = val + sv%dv%get_nrows()
return
end function s_jac_solver_get_nzeros
function s_jac_solver_sizeof(sv) result(val)
implicit none
! Arguments
class(amg_s_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = psb_sizeof_ip
val = val + sv%a%sizeof()
val = val + sv%dv%sizeof()
return
end function s_jac_solver_sizeof
function s_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "Jacobi solver"
end function s_jac_solver_get_fmt
function s_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_jac_
end function s_jac_solver_get_id
!
! If this is true, then the solver needs a starting
! guess. Currently only handled in JAC smoother.
!
function s_jac_solver_is_iterative() result(val)
implicit none
logical :: val
val = .true.
end function s_jac_solver_is_iterative
function s_jac_solver_get_wrksize() result(val)
implicit none
integer(psb_ipk_) :: val
val = 2
end function s_jac_solver_get_wrksize
subroutine s_l1_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_s_l1_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_s_l1_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_l1_jac_solver_descr
function s_l1_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "L1-Jacobi solver"
end function s_l1_jac_solver_get_fmt
function s_l1_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_l1_jac_
end function s_l1_jac_solver_get_id
end module amg_s_jac_solver
+41 -15
View File
@@ -143,9 +143,10 @@ contains
type(psb_ls_coo_sparse_mat) :: tmpcoo
logical :: display_out_, print_out_, reproducible_
logical, parameter :: dump=.false., debug=.false., dump_mate=.false., &
& debug_ilaggr=.false., debug_sync=.false.
& debug_ilaggr=.false., debug_sync=.false., debug_mate=.false.
integer(psb_ipk_), save :: idx_bldmtc=-1, idx_phase1=-1, idx_phase2=-1, idx_phase3=-1
logical, parameter :: do_timings=.true.
integer, parameter :: ilaggr_neginit=-1, ilaggr_nonlocal=-2
ictxt = desc_a%get_ctxt()
call psb_info(ictxt,iam,np)
@@ -187,7 +188,7 @@ contains
call desc_a%l2gip(ilv,info,owned=.false.)
call psb_geall(ilaggr,desc_a,info)
ilaggr = -1
ilaggr = ilaggr_neginit
call psb_geasb(ilaggr,desc_a,info)
nr = a%get_nrows()
nc = a%get_ncols()
@@ -213,7 +214,20 @@ contains
call psb_barrier(ictxt)
if (iam == 0) write(0,*)' out from buildmatching:', info
end if
if (debug_mate) then
block
integer(psb_lpk_), allocatable :: ckmate(:)
allocate(ckmate(nr))
ckmate(1:nr) = mate(1:nr)
call psb_msort(ckmate(1:nr))
do i=1,nr-1
if ((ckmate(i)>0) .and. (ckmate(i) == ckmate(i+1))) then
write(0,*) iam,' Duplicate mate entry at',i,' :',ckmate(i)
end if
end do
end block
end if
if (info == 0) then
if (do_timings) call psb_tic(idx_phase2)
if (debug_sync) then
@@ -258,16 +272,14 @@ contains
write(0,*) 'Impossible: mate(k) > nc'
cycle
else
if (ilaggr(k) == -1) then
if (ilaggr(k) == ilaggr_neginit) then
wk = w(k)
widx = w(idx)
wmax = max(abs(wk),abs(widx))
nrmagg = wmax*sqrt((wk/wmax)**2+(widx/wmax)**2)
if (nrmagg > epsilon(nrmagg)) then
if (idx <= nr) then
if (ilaggr(idx) == -1) then
if (ilaggr(idx) == ilaggr_neginit) then
! Now, if both vertices are local, the aggregate is local
! (kinda obvious).
nlaggr(iam) = nlaggr(iam) + 1
@@ -275,6 +287,9 @@ contains
ilaggr(idx) = nlaggr(iam)
wtemp(k) = w(k)/nrmagg
wtemp(idx) = w(idx)/nrmagg
else
write(0,*) iam,' Inconsistent mate? ',k,mate(k),idx,&
&mate(idx),ilaggr(idx)
end if
nlpairs = nlpairs+1
else if (idx <= nc) then
@@ -294,7 +309,7 @@ contains
ilaggr(k) = nlaggr(iam)
nlpairs = nlpairs+1
else
ilaggr(k) = -2
ilaggr(k) = ilaggr_nonlocal
end if
else
! Use a statistically unbiased tie-breaking rule,
@@ -309,7 +324,7 @@ contains
ilaggr(k) = nlaggr(iam)
nlpairs = nlpairs+1
else
ilaggr(k) = -2
ilaggr(k) = ilaggr_nonlocal
end if
end if
end if
@@ -325,6 +340,12 @@ contains
nlsingl = nlsingl + 1
end if
end if
if (ilaggr(k) == ilaggr_neginit) then
write(0,*) iam,' Error: no update to ',k,mate(k),&
& abs(w(k)),nrmagg,epsilon(nrmagg),wtemp(k)
end if
else
if (ilaggr(k)<0) write(0,*) 'Strange? ',k,ilaggr(k)
end if
end if
end do
@@ -332,7 +353,7 @@ contains
if (do_timings) call psb_tic(idx_phase3)
! Ok, now compute offsets, gather halo and fix non-local
! aggregates (those where ilaggr == -2)
! aggregates (those where ilaggr == ilaggr_nonlocal)
call psb_sum(ictxt,nlaggr)
ntaggr = sum(nlaggr(0:np-1))
naggrm1 = sum(nlaggr(0:iam-1))
@@ -347,7 +368,7 @@ contains
call psb_halo(wtemp,desc_a,info)
! Cleanup as yet unmarked entries
do k=1,nr
if (ilaggr(k) == -2) then
if (ilaggr(k) == ilaggr_nonlocal) then
idx = mate(k)
if (idx > nr) then
i = ilaggr(idx)
@@ -359,9 +380,14 @@ contains
else
write(0,*) 'Error : unresolved (paired) index ',k,idx,i,nr,nc, ilv(k),ilv(idx)
end if
end if
if (ilaggr(k) <0) then
write(0,*) 'Matchboxp: Funny number: ',k,ilv(k),ilaggr(k),wtemp(k)
else if (ilaggr(k) <0) then
write(0,*) iam,'Matchboxp: Funny number: ',k,ilv(k),ilaggr(k),wtemp(k)
write(0,*) iam,' : : ',nr,nc,mate(k)
if (mate(k) <= nr) then
write(0,*) iam,' : : ',ilaggr(mate(k)),mate(mate(k)),&
& ilv(k),ilv(mate(k)), ilv(mate(mate(k))),ilaggr(mate(mate(k)))
end if
flush(0)
end if
end do
if (debug_sync) then
@@ -414,7 +440,7 @@ contains
end block
if (iam == 0) then
write(0,*) 'Matching statistics: Unmatched nodes ',&
write(0,*) iam,'Matching statistics: Unmatched nodes ',&
& nunmatched,' Singletons:',nlsingl,' Pairs:',nlpairs
end if
+2 -1
View File
@@ -78,7 +78,8 @@ module amg_s_mumps_solver
!
! Controls to be set before MUMPS instantiation:
!
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: LOCAL 1==amg_global_solver_: GLOBAL
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: 'LOCAL_SOLVER
! 1==amg_global_solver_: 'GLOBAL_SOLVER'
! IPAR(2) : MUMPS_PRINT_ERR print verbosity (see MUMPS)
! IPAR(3) : MUMPS_SYM 0: non-symmetric 2: symmetric
integer(psb_ipk_), dimension(3) :: ipar
+18
View File
@@ -188,6 +188,7 @@ module amg_s_onelev_mod
procedure, pass(lv) :: clone => s_base_onelev_clone
procedure, pass(lv) :: cnv => amg_s_base_onelev_cnv
procedure, pass(lv) :: descr => amg_s_base_onelev_descr
procedure, pass(lv) :: memory_use => amg_s_base_onelev_memory_use
procedure, pass(lv) :: default => s_base_onelev_default
procedure, pass(lv) :: free => amg_s_base_onelev_free
procedure, pass(lv) :: free_smoothers => amg_s_base_onelev_free_smoothers
@@ -274,6 +275,23 @@ module amg_s_onelev_mod
end subroutine amg_s_base_onelev_descr
end interface
interface
subroutine amg_s_base_onelev_memory_use(lv,il,nl,ilmin,info,iout,verbosity, prefix,global)
import :: psb_sspmat_type, psb_s_vect_type, psb_s_base_vect_type, &
& psb_slinmap_type, psb_spk_, amg_s_onelev_type, &
& psb_ipk_, psb_epk_, psb_desc_type
Implicit None
! Arguments
class(amg_s_onelev_type), intent(in) :: lv
integer(psb_ipk_), intent(in) :: il,nl,ilmin
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_s_base_onelev_memory_use
end interface
interface
subroutine amg_s_base_onelev_cnv(lv,info,amold,vmold,imold)
import :: amg_s_onelev_type, psb_s_base_vect_type, psb_spk_, &
+374
View File
@@ -0,0 +1,374 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Daniela di Serafino
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
! File: amg_s_poly_smoother_mod.f90
!
! Module: amg_s_poly_smoother_mod
!
! This module defines:
! the amg_s_poly_smoother_type data structure containing the
! smoother for a Jacobi/block Jacobi smoother.
! The smoother stores in ND the block off-diagonal matrix.
! One special case is treated separately, when the solver is DIAG or L1-DIAG
! then the ND is the entire off-diagonal part of the matrix (including the
! main diagonal block), so that it becomes possible to implement
! a pure Jacobi or L1-Jacobi global solver.
!
module amg_s_poly_smoother
use amg_s_base_smoother_mod
use amg_d_poly_coeff_mod
type, extends(amg_s_base_smoother_type) :: amg_s_poly_smoother_type
! The local solver component is inherited from the
! parent type.
! class(amg_s_base_solver_type), allocatable :: sv
!
integer(psb_ipk_) :: pdegree, variant
integer(psb_ipk_) :: rho_estimate=amg_poly_rho_est_power_
integer(psb_ipk_) :: rho_estimate_iterations=10
type(psb_sspmat_type), pointer :: pa => null()
real(psb_spk_), allocatable :: poly_beta(:)
real(psb_spk_) :: cf_a = szero
real(psb_spk_) :: rho_ba = -sone
contains
procedure, pass(sm) :: apply_v => amg_s_poly_smoother_apply_vect
!!$ procedure, pass(sm) :: apply_a => amg_s_poly_smoother_apply
procedure, pass(sm) :: dump => amg_s_poly_smoother_dmp
procedure, pass(sm) :: build => amg_s_poly_smoother_bld
procedure, pass(sm) :: cnv => amg_s_poly_smoother_cnv
procedure, pass(sm) :: clone => amg_s_poly_smoother_clone
procedure, pass(sm) :: clone_settings => amg_s_poly_smoother_clone_settings
procedure, pass(sm) :: clear_data => amg_s_poly_smoother_clear_data
procedure, pass(sm) :: free => s_poly_smoother_free
procedure, pass(sm) :: cseti => amg_s_poly_smoother_cseti
procedure, pass(sm) :: csetc => amg_s_poly_smoother_csetc
procedure, pass(sm) :: csetr => amg_s_poly_smoother_csetr
procedure, pass(sm) :: descr => amg_s_poly_smoother_descr
procedure, pass(sm) :: sizeof => s_poly_smoother_sizeof
procedure, pass(sm) :: default => s_poly_smoother_default
procedure, pass(sm) :: get_nzeros => s_poly_smoother_get_nzeros
procedure, pass(sm) :: get_wrksz => s_poly_smoother_get_wrksize
procedure, nopass :: get_fmt => s_poly_smoother_get_fmt
procedure, nopass :: get_id => s_poly_smoother_get_id
end type amg_s_poly_smoother_type
private :: s_poly_smoother_free, &
& s_poly_smoother_sizeof, s_poly_smoother_get_nzeros, &
& s_poly_smoother_get_fmt, s_poly_smoother_get_id, &
& s_poly_smoother_get_wrksize
interface
subroutine amg_s_poly_smoother_apply_vect(alpha,sm,x,beta,y,desc_data,trans,&
& sweeps,work,wv,info,init,initu)
import :: psb_desc_type, amg_s_poly_smoother_type, psb_s_vect_type, psb_spk_, &
& psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type,&
& psb_ipk_
type(psb_desc_type), intent(in) :: desc_data
class(amg_s_poly_smoother_type), intent(inout) :: sm
type(psb_s_vect_type),intent(inout) :: x
type(psb_s_vect_type),intent(inout) :: y
real(psb_spk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
integer(psb_ipk_), intent(in) :: sweeps
real(psb_spk_),target, intent(inout) :: work(:)
type(psb_s_vect_type),intent(inout) :: wv(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
type(psb_s_vect_type),intent(inout), optional :: initu
end subroutine amg_s_poly_smoother_apply_vect
end interface
!!$ interface
!!$ subroutine amg_s_poly_smoother_apply(alpha,sm,x,beta,y,desc_data,trans,&
!!$ & sweeps,work,info,init,initu)
!!$ import :: psb_desc_type, amg_s_poly_smoother_type, psb_s_vect_type, psb_spk_, &
!!$ & psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type, &
!!$ & psb_ipk_
!!$ type(psb_desc_type), intent(in) :: desc_data
!!$ class(amg_s_poly_smoother_type), intent(inout) :: sm
!!$ real(psb_spk_),intent(inout) :: x(:)
!!$ real(psb_spk_),intent(inout) :: y(:)
!!$ real(psb_spk_),intent(in) :: alpha,beta
!!$ character(len=1),intent(in) :: trans
!!$ integer(psb_ipk_), intent(in) :: sweeps
!!$ real(psb_spk_),target, intent(inout) :: work(:)
!!$ integer(psb_ipk_), intent(out) :: info
!!$ character, intent(in), optional :: init
!!$ real(psb_spk_),intent(inout), optional :: initu(:)
!!$ end subroutine amg_s_poly_smoother_apply
!!$ end interface
!!$
interface
subroutine amg_s_poly_smoother_bld(a,desc_a,sm,info,amold,vmold,imold)
import :: psb_desc_type, amg_s_poly_smoother_type, psb_s_vect_type, psb_spk_, &
& psb_sspmat_type, psb_s_base_sparse_mat, psb_s_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
type(psb_sspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_s_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
class(psb_s_base_sparse_mat), intent(in), optional :: amold
class(psb_s_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_s_poly_smoother_bld
end interface
interface
subroutine amg_s_poly_smoother_cnv(sm,info,amold,vmold,imold)
import :: amg_s_poly_smoother_type, psb_spk_, &
& psb_s_base_sparse_mat, psb_s_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
class(amg_s_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
class(psb_s_base_sparse_mat), intent(in), optional :: amold
class(psb_s_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_s_poly_smoother_cnv
end interface
interface
subroutine amg_s_poly_smoother_dmp(sm,desc,level,info,prefix,head,smoother,solver,global_num)
import :: psb_sspmat_type, psb_s_vect_type, psb_s_base_vect_type, &
& psb_spk_, amg_s_poly_smoother_type, psb_epk_, psb_desc_type, &
& psb_ipk_
implicit none
class(amg_s_poly_smoother_type), intent(in) :: sm
type(psb_desc_type), intent(in) :: desc
integer(psb_ipk_), intent(in) :: level
integer(psb_ipk_), intent(out) :: info
character(len=*), intent(in), optional :: prefix, head
logical, optional, intent(in) :: smoother, solver, global_num
end subroutine amg_s_poly_smoother_dmp
end interface
interface
subroutine amg_s_poly_smoother_clone(sm,smout,info)
import :: amg_s_poly_smoother_type, psb_spk_, &
& amg_s_base_smoother_type, psb_ipk_
class(amg_s_poly_smoother_type), intent(inout) :: sm
class(amg_s_base_smoother_type), allocatable, intent(inout) :: smout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_s_poly_smoother_clone
end interface
interface
subroutine amg_s_poly_smoother_clone_settings(sm,smout,info)
import :: amg_s_poly_smoother_type, psb_spk_, &
& amg_s_base_smoother_type, psb_ipk_
class(amg_s_poly_smoother_type), intent(inout) :: sm
class(amg_s_base_smoother_type), allocatable, intent(inout) :: smout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_s_poly_smoother_clone_settings
end interface
interface
subroutine amg_s_poly_smoother_clear_data(sm,info)
import :: amg_s_poly_smoother_type, psb_spk_, &
& amg_s_base_smoother_type, psb_ipk_
class(amg_s_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
end subroutine amg_s_poly_smoother_clear_data
end interface
interface
subroutine amg_s_poly_smoother_descr(sm,info,iout,coarse,prefix)
import :: amg_s_poly_smoother_type, psb_ipk_
class(amg_s_poly_smoother_type), intent(in) :: sm
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
end subroutine amg_s_poly_smoother_descr
end interface
interface
subroutine amg_s_poly_smoother_cseti(sm,what,val,info,idx)
import :: psb_sspmat_type, psb_s_vect_type, psb_s_base_vect_type, &
& psb_spk_, amg_s_poly_smoother_type, psb_epk_, psb_desc_type, psb_ipk_
implicit none
class(amg_s_poly_smoother_type), intent(inout) :: sm
character(len=*), intent(in) :: what
integer(psb_ipk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
end subroutine amg_s_poly_smoother_cseti
end interface
interface
subroutine amg_s_poly_smoother_csetc(sm,what,val,info,idx)
import :: psb_sspmat_type, psb_s_vect_type, psb_s_base_vect_type, &
& psb_spk_, amg_s_poly_smoother_type, psb_epk_, psb_desc_type, psb_ipk_
implicit none
class(amg_s_poly_smoother_type), intent(inout) :: sm
character(len=*), intent(in) :: what
character(len=*), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
end subroutine amg_s_poly_smoother_csetc
end interface
interface
subroutine amg_s_poly_smoother_csetr(sm,what,val,info,idx)
import :: psb_sspmat_type, psb_s_vect_type, psb_s_base_vect_type, &
& psb_spk_, amg_s_poly_smoother_type, psb_epk_, psb_desc_type, psb_ipk_
implicit none
class(amg_s_poly_smoother_type), intent(inout) :: sm
character(len=*), intent(in) :: what
real(psb_spk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
end subroutine amg_s_poly_smoother_csetr
end interface
contains
subroutine s_poly_smoother_free(sm,info)
Implicit None
! Arguments
class(amg_s_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='s_poly_smoother_free'
call psb_erractionsave(err_act)
info = psb_success_
if (allocated(sm%sv)) then
call sm%sv%free(info)
if (info == psb_success_) deallocate(sm%sv,stat=info)
if (info /= psb_success_) then
info = psb_err_alloc_dealloc_
call psb_errpush(info,name)
goto 9999
end if
end if
if (allocated(sm%poly_beta)) deallocate(sm%poly_beta)
sm%pa => null()
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine s_poly_smoother_free
function s_poly_smoother_sizeof(sm) result(val)
implicit none
! Arguments
class(amg_s_poly_smoother_type), intent(in) :: sm
integer(psb_epk_) :: val
val = psb_sizeof_dp
if (allocated(sm%sv)) val = val + sm%sv%sizeof()
if (allocated(sm%poly_beta)) val = val + psb_sizeof_dp * size(sm%poly_beta)
return
end function s_poly_smoother_sizeof
subroutine s_poly_smoother_default(sm)
Implicit None
! Arguments
class(amg_s_poly_smoother_type), intent(inout) :: sm
!
! Default: BJAC with no residual check
!
sm%pdegree = 1
sm%rho_ba = -sone
sm%variant = amg_poly_lottes_
sm%rho_estimate = amg_poly_rho_est_power_
sm%rho_estimate_iterations = 20
if (allocated(sm%sv)) then
call sm%sv%default()
end if
return
end subroutine s_poly_smoother_default
function s_poly_smoother_get_nzeros(sm) result(val)
implicit none
! Arguments
class(amg_s_poly_smoother_type), intent(in) :: sm
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = 0
if (allocated(sm%sv)) val = val + sm%sv%get_nzeros()
return
end function s_poly_smoother_get_nzeros
function s_poly_smoother_get_wrksize(sm) result(val)
implicit none
class(amg_s_poly_smoother_type), intent(inout) :: sm
integer(psb_ipk_) :: val
val = 4
if (allocated(sm%sv)) val = val + sm%sv%get_wrksz()
end function s_poly_smoother_get_wrksize
function s_poly_smoother_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "Polynomial smoother"
end function s_poly_smoother_get_fmt
function s_poly_smoother_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_poly_
end function s_poly_smoother_get_id
end module amg_s_poly_smoother
+17
View File
@@ -139,6 +139,7 @@ module amg_s_prec_type
procedure, pass(prec) :: smoothers_build => amg_s_smoothers_bld
procedure, pass(prec) :: smoothers_free => amg_s_smoothers_free
procedure, pass(prec) :: descr => amg_sfile_prec_descr
procedure, pass(prec) :: memory_use => amg_sfile_prec_memory_use
end type amg_sprec_type
private :: amg_s_dump, amg_s_get_compl, amg_s_cmp_compl,&
@@ -170,6 +171,22 @@ module amg_s_prec_type
end subroutine amg_sfile_prec_descr
end interface
interface amg_memory_use
subroutine amg_sfile_prec_memory_use(prec,info,iout,root,verbosity,prefix,global)
import :: amg_sprec_type, psb_ipk_
implicit none
! Arguments
class(amg_sprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_sfile_prec_memory_use
end interface
interface amg_sizeof
module procedure amg_sprec_sizeof
end interface
+7 -7
View File
@@ -234,7 +234,7 @@ contains
! Arguments
class(amg_z_ilu_solver_type), intent(inout) :: sv
sv%fact_type = psb_ilu_n_
sv%fact_type = amg_ilu_n_
sv%fill_in = 0
sv%thresh = dzero
@@ -255,13 +255,13 @@ contains
info = psb_success_
call amg_check_def(sv%fact_type,&
& 'Factorization',psb_ilu_n_,is_legal_ilu_fact)
& 'Factorization',amg_ilu_n_,is_legal_ilu_fact)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
call amg_check_def(sv%fill_in,&
& 'Level',izero,is_int_non_negative)
case(psb_ilu_t_)
case(amg_ilu_t_)
call amg_check_def(sv%thresh,&
& 'Eps',dzero,is_legal_d_fact_thrs)
end select
@@ -439,9 +439,9 @@ contains
write(iout_,*) trim(prefix_), ' Incomplete factorization solver: ',&
& amg_fact_names(sv%fact_type)
select case(sv%fact_type)
case(psb_ilu_n_,psb_milu_n_)
case(amg_ilu_n_,amg_milu_n_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
case(psb_ilu_t_)
case(amg_ilu_t_)
write(iout_,*) trim(prefix_), ' Fill level:',sv%fill_in
write(iout_,*) trim(prefix_), ' Fill threshold :',sv%thresh
end select
@@ -496,7 +496,7 @@ contains
implicit none
integer(psb_ipk_) :: val
val = psb_ilu_n_
val = amg_ilu_n_
end function z_ilu_solver_get_id
function z_ilu_solver_get_wrksize() result(val)
+585
View File
@@ -0,0 +1,585 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
!
! File: amg_z_jac_solver_mod.f90
!
! Module: amg_z_jac_solver_mod
!
! This module defines:
! - the amg_z_jac_solver_type data structure containing the ingredients
! for a local Jacobi iteration. The iterations are local to a process
! (they operate on the block diagonal).
!
!
module amg_z_jac_solver
use amg_z_base_solver_mod
type, extends(amg_z_base_solver_type) :: amg_z_jac_solver_type
type(psb_zspmat_type) :: a
type(psb_z_vect_type), allocatable :: dv
complex(psb_dpk_), allocatable :: d(:)
integer(psb_ipk_) :: sweeps
real(psb_dpk_) :: eps
contains
procedure, pass(sv) :: dump => amg_z_jac_solver_dmp
procedure, pass(sv) :: check => z_jac_solver_check
procedure, pass(sv) :: clone => amg_z_jac_solver_clone
procedure, pass(sv) :: clone_settings => amg_z_jac_solver_clone_settings
procedure, pass(sv) :: clear_data => amg_z_jac_solver_clear_data
procedure, pass(sv) :: build => amg_z_jac_solver_bld
procedure, pass(sv) :: cnv => amg_z_jac_solver_cnv
procedure, pass(sv) :: apply_v => amg_z_jac_solver_apply_vect
procedure, pass(sv) :: apply_a => amg_z_jac_solver_apply
procedure, pass(sv) :: free => z_jac_solver_free
procedure, pass(sv) :: cseti => z_jac_solver_cseti
procedure, pass(sv) :: csetc => z_jac_solver_csetc
procedure, pass(sv) :: csetr => z_jac_solver_csetr
procedure, pass(sv) :: descr => z_jac_solver_descr
procedure, pass(sv) :: default => z_jac_solver_default
procedure, pass(sv) :: sizeof => z_jac_solver_sizeof
procedure, pass(sv) :: get_nzeros => z_jac_solver_get_nzeros
procedure, nopass :: get_wrksz => z_jac_solver_get_wrksize
procedure, nopass :: get_fmt => z_jac_solver_get_fmt
procedure, nopass :: get_id => z_jac_solver_get_id
procedure, nopass :: is_iterative => z_jac_solver_is_iterative
end type amg_z_jac_solver_type
type, extends(amg_z_jac_solver_type) :: amg_z_l1_jac_solver_type
contains
procedure, pass(sv) :: build => amg_z_l1_jac_solver_bld
procedure, pass(sv) :: descr => z_l1_jac_solver_descr
procedure, nopass :: get_fmt => z_l1_jac_solver_get_fmt
procedure, nopass :: get_id => z_l1_jac_solver_get_id
end type amg_z_l1_jac_solver_type
private :: z_jac_solver_bld, z_jac_solver_apply, &
& z_jac_solver_free, &
& z_jac_solver_descr, z_jac_solver_sizeof, &
& z_jac_solver_default, z_jac_solver_dmp, &
& z_jac_solver_apply_vect, z_jac_solver_get_nzeros, &
& z_jac_solver_get_fmt, z_jac_solver_check,&
& z_jac_solver_is_iterative, &
& z_jac_solver_get_id, z_jac_solver_get_wrksize
interface
subroutine amg_z_jac_solver_apply_vect(alpha,sv,x,beta,y,desc_data,&
& trans,work,wv,info,init,initu)
import :: psb_desc_type, amg_z_jac_solver_type, psb_z_vect_type, psb_dpk_, &
& psb_zspmat_type, psb_z_base_sparse_mat, psb_z_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_z_jac_solver_type), intent(inout) :: sv
type(psb_z_vect_type),intent(inout) :: x
type(psb_z_vect_type),intent(inout) :: y
complex(psb_dpk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
complex(psb_dpk_),target, intent(inout) :: work(:)
type(psb_z_vect_type),intent(inout) :: wv(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
type(psb_z_vect_type),intent(inout), optional :: initu
end subroutine amg_z_jac_solver_apply_vect
end interface
interface
subroutine amg_z_jac_solver_apply(alpha,sv,x,beta,y,desc_data,trans,work,info,init,initu)
import :: psb_desc_type, amg_z_jac_solver_type, psb_z_vect_type, psb_dpk_, &
& psb_zspmat_type, psb_z_base_sparse_mat, psb_z_base_vect_type, psb_ipk_
implicit none
type(psb_desc_type), intent(in) :: desc_data
class(amg_z_jac_solver_type), intent(inout) :: sv
complex(psb_dpk_),intent(inout) :: x(:)
complex(psb_dpk_),intent(inout) :: y(:)
complex(psb_dpk_),intent(in) :: alpha,beta
character(len=1),intent(in) :: trans
complex(psb_dpk_),target, intent(inout) :: work(:)
integer(psb_ipk_), intent(out) :: info
character, intent(in), optional :: init
complex(psb_dpk_),intent(inout), optional :: initu(:)
end subroutine amg_z_jac_solver_apply
end interface
interface
subroutine amg_z_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_z_jac_solver_type, psb_z_vect_type, psb_dpk_, &
& psb_zspmat_type, psb_z_base_sparse_mat, psb_z_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_zspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_z_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_zspmat_type), intent(in), target, optional :: b
class(psb_z_base_sparse_mat), intent(in), optional :: amold
class(psb_z_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_z_jac_solver_bld
end interface
interface
subroutine amg_z_l1_jac_solver_bld(a,desc_a,sv,info,b,amold,vmold,imold)
import :: psb_desc_type, amg_z_l1_jac_solver_type, psb_z_vect_type, psb_dpk_, &
& psb_zspmat_type, psb_z_base_sparse_mat, psb_z_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
type(psb_zspmat_type), intent(in), target :: a
Type(psb_desc_type), Intent(inout) :: desc_a
class(amg_z_l1_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
type(psb_zspmat_type), intent(in), target, optional :: b
class(psb_z_base_sparse_mat), intent(in), optional :: amold
class(psb_z_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_z_l1_jac_solver_bld
end interface
interface
subroutine amg_z_jac_solver_cnv(sv,info,amold,vmold,imold)
import :: amg_z_jac_solver_type, psb_dpk_, &
& psb_z_base_sparse_mat, psb_z_base_vect_type,&
& psb_ipk_, psb_i_base_vect_type
implicit none
class(amg_z_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
class(psb_z_base_sparse_mat), intent(in), optional :: amold
class(psb_z_base_vect_type), intent(in), optional :: vmold
class(psb_i_base_vect_type), intent(in), optional :: imold
end subroutine amg_z_jac_solver_cnv
end interface
interface
subroutine amg_z_jac_solver_dmp(sv,desc,level,info,prefix,head,solver,global_num)
import :: psb_desc_type, amg_z_jac_solver_type, psb_z_vect_type, psb_dpk_, &
& psb_zspmat_type, psb_z_base_sparse_mat, psb_z_base_vect_type, &
& psb_ipk_
implicit none
class(amg_z_jac_solver_type), intent(in) :: sv
type(psb_desc_type), intent(in) :: desc
integer(psb_ipk_), intent(in) :: level
integer(psb_ipk_), intent(out) :: info
character(len=*), intent(in), optional :: prefix, head
logical, optional, intent(in) :: solver, global_num
end subroutine amg_z_jac_solver_dmp
end interface
interface
subroutine amg_z_jac_solver_clone(sv,svout,info)
import :: psb_desc_type, psb_zspmat_type, psb_z_base_sparse_mat, &
& psb_z_vect_type, psb_z_base_vect_type, psb_dpk_, &
& amg_z_base_solver_type, amg_z_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
class(amg_z_base_solver_type), allocatable, intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_z_jac_solver_clone
end interface
!!$ interface
!!$ subroutine amg_z_l1_jac_solver_clone(sv,svout,info)
!!$ import :: psb_desc_type, psb_zspmat_type, psb_z_base_sparse_mat, &
!!$ & psb_z_vect_type, psb_z_base_vect_type, psb_dpk_, &
!!$ & amg_z_base_solver_type, amg_z_l1_jac_solver_type, psb_ipk_
!!$ Implicit None
!!$
!!$ ! Arguments
!!$ class(amg_z_l1_jac_solver_type), intent(inout) :: sv
!!$ class(amg_z_base_solver_type), allocatable, intent(inout) :: svout
!!$ integer(psb_ipk_), intent(out) :: info
!!$ end subroutine amg_z_l1_jac_solver_clone
!!$ end interface
interface
subroutine amg_z_jac_solver_clone_settings(sv,svout,info)
import :: psb_desc_type, psb_zspmat_type, psb_z_base_sparse_mat, &
& psb_z_vect_type, psb_z_base_vect_type, psb_dpk_, &
& amg_z_base_solver_type, amg_z_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
class(amg_z_base_solver_type), intent(inout) :: svout
integer(psb_ipk_), intent(out) :: info
end subroutine amg_z_jac_solver_clone_settings
end interface
interface
subroutine amg_z_jac_solver_clear_data(sv,info)
import :: psb_desc_type, psb_zspmat_type, psb_z_base_sparse_mat, &
& psb_z_vect_type, psb_z_base_vect_type, psb_dpk_, &
& amg_z_jac_solver_type, psb_ipk_
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
end subroutine amg_z_jac_solver_clear_data
end interface
contains
subroutine z_jac_solver_default(sv)
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
sv%sweeps = ione
sv%eps = dzero
return
end subroutine z_jac_solver_default
subroutine z_jac_solver_check(sv,info)
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='z_jac_solver_check'
call psb_erractionsave(err_act)
info = psb_success_
call amg_check_def(sv%sweeps,&
& 'Jacobi sweeps',ione,is_int_positive)
if (info /= psb_success_) goto 9999
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine z_jac_solver_check
subroutine z_jac_solver_cseti(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
integer(psb_ipk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='z_jac_solver_cseti'
info = psb_success_
call psb_erractionsave(err_act)
select case(psb_toupper(what))
case('SOLVER_SWEEPS')
sv%sweeps = val
case default
call sv%amg_z_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine z_jac_solver_cseti
subroutine z_jac_solver_csetc(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
character(len=*), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act, ival
character(len=20) :: name='z_jac_solver_csetc'
info = psb_success_
call psb_erractionsave(err_act)
call sv%amg_z_base_solver_type%set(what,val,info,idx=idx)
if (info /= psb_success_) then
info = psb_err_from_subroutine_
call psb_errpush(info, name)
goto 9999
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine z_jac_solver_csetc
subroutine z_jac_solver_csetr(sv,what,val,info,idx)
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
character(len=*), intent(in) :: what
real(psb_dpk_), intent(in) :: val
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: idx
integer(psb_ipk_) :: err_act
character(len=20) :: name='z_jac_solver_csetr'
call psb_erractionsave(err_act)
info = psb_success_
select case(psb_toupper(what))
case('SOLVER_EPS')
sv%eps = val
case default
call sv%amg_z_base_solver_type%set(what,val,info,idx=idx)
end select
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine z_jac_solver_csetr
subroutine z_jac_solver_free(sv,info)
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(inout) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_) :: err_act
character(len=20) :: name='z_jac_solver_free'
call psb_erractionsave(err_act)
info = psb_success_
call sv%a%free()
if (allocated(sv%dv)) then
call sv%dv%free(info)
deallocate(sv%dv)
end if
if (allocated(sv%d)) deallocate(sv%d)
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine z_jac_solver_free
subroutine z_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_z_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_z_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine z_jac_solver_descr
function z_jac_solver_get_nzeros(sv) result(val)
implicit none
! Arguments
class(amg_z_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = 0
val = val + sv%a%get_nzeros()
val = val + sv%dv%get_nrows()
return
end function z_jac_solver_get_nzeros
function z_jac_solver_sizeof(sv) result(val)
implicit none
! Arguments
class(amg_z_jac_solver_type), intent(in) :: sv
integer(psb_epk_) :: val
integer(psb_ipk_) :: i
val = psb_sizeof_ip
val = val + sv%a%sizeof()
val = val + sv%dv%sizeof()
return
end function z_jac_solver_sizeof
function z_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "Jacobi solver"
end function z_jac_solver_get_fmt
function z_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_jac_
end function z_jac_solver_get_id
!
! If this is true, then the solver needs a starting
! guess. Currently only handled in JAC smoother.
!
function z_jac_solver_is_iterative() result(val)
implicit none
logical :: val
val = .true.
end function z_jac_solver_is_iterative
function z_jac_solver_get_wrksize() result(val)
implicit none
integer(psb_ipk_) :: val
val = 2
end function z_jac_solver_get_wrksize
subroutine z_l1_jac_solver_descr(sv,info,iout,coarse,prefix)
Implicit None
! Arguments
class(amg_z_l1_jac_solver_type), intent(in) :: sv
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
logical, intent(in), optional :: coarse
character(len=*), intent(in), optional :: prefix
! Local variables
integer(psb_ipk_) :: err_act
character(len=20), parameter :: name='amg_z_l1_jac_solver_descr'
integer(psb_ipk_) :: iout_
character(1024) :: prefix_
call psb_erractionsave(err_act)
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
endif
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if (sv%eps<=dzero) then
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with ',&
& sv%sweeps,' sweeps'
else
write(iout_,*) trim(prefix_), ' L1-Jacobi iterative solver with tolerance',&
& sv%eps,' and maxit', sv%sweeps
end if
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine z_l1_jac_solver_descr
function z_l1_jac_solver_get_fmt() result(val)
implicit none
character(len=32) :: val
val = "L1-Jacobi solver"
end function z_l1_jac_solver_get_fmt
function z_l1_jac_solver_get_id() result(val)
implicit none
integer(psb_ipk_) :: val
val = amg_l1_jac_
end function z_l1_jac_solver_get_id
end module amg_z_jac_solver
+2 -1
View File
@@ -78,7 +78,8 @@ module amg_z_mumps_solver
!
! Controls to be set before MUMPS instantiation:
!
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: LOCAL 1==amg_global_solver_: GLOBAL
! IPAR(1) : MUMPS_LOC_GLOB 0==amg_local_solver_: 'LOCAL_SOLVER
! 1==amg_global_solver_: 'GLOBAL_SOLVER'
! IPAR(2) : MUMPS_PRINT_ERR print verbosity (see MUMPS)
! IPAR(3) : MUMPS_SYM 0: non-symmetric 2: symmetric
integer(psb_ipk_), dimension(3) :: ipar
+18
View File
@@ -187,6 +187,7 @@ module amg_z_onelev_mod
procedure, pass(lv) :: clone => z_base_onelev_clone
procedure, pass(lv) :: cnv => amg_z_base_onelev_cnv
procedure, pass(lv) :: descr => amg_z_base_onelev_descr
procedure, pass(lv) :: memory_use => amg_z_base_onelev_memory_use
procedure, pass(lv) :: default => z_base_onelev_default
procedure, pass(lv) :: free => amg_z_base_onelev_free
procedure, pass(lv) :: free_smoothers => amg_z_base_onelev_free_smoothers
@@ -273,6 +274,23 @@ module amg_z_onelev_mod
end subroutine amg_z_base_onelev_descr
end interface
interface
subroutine amg_z_base_onelev_memory_use(lv,il,nl,ilmin,info,iout,verbosity, prefix,global)
import :: psb_zspmat_type, psb_z_vect_type, psb_z_base_vect_type, &
& psb_zlinmap_type, psb_dpk_, amg_z_onelev_type, &
& psb_ipk_, psb_epk_, psb_desc_type
Implicit None
! Arguments
class(amg_z_onelev_type), intent(in) :: lv
integer(psb_ipk_), intent(in) :: il,nl,ilmin
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_z_base_onelev_memory_use
end interface
interface
subroutine amg_z_base_onelev_cnv(lv,info,amold,vmold,imold)
import :: amg_z_onelev_type, psb_z_base_vect_type, psb_dpk_, &
+17
View File
@@ -139,6 +139,7 @@ module amg_z_prec_type
procedure, pass(prec) :: smoothers_build => amg_z_smoothers_bld
procedure, pass(prec) :: smoothers_free => amg_z_smoothers_free
procedure, pass(prec) :: descr => amg_zfile_prec_descr
procedure, pass(prec) :: memory_use => amg_zfile_prec_memory_use
end type amg_zprec_type
private :: amg_z_dump, amg_z_get_compl, amg_z_cmp_compl,&
@@ -170,6 +171,22 @@ module amg_z_prec_type
end subroutine amg_zfile_prec_descr
end interface
interface amg_memory_use
subroutine amg_zfile_prec_memory_use(prec,info,iout,root,verbosity,prefix,global)
import :: amg_zprec_type, psb_ipk_
implicit none
! Arguments
class(amg_zprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
end subroutine amg_zfile_prec_memory_use
end interface
interface amg_sizeof
module procedure amg_zprec_sizeof
end interface
+4 -4
View File
@@ -22,22 +22,22 @@ MPFOBJS=$(SMPFOBJS) $(DMPFOBJS) $(CMPFOBJS) $(ZMPFOBJS)
MPCOBJS=amg_dslud_interface.o amg_zslud_interface.o
DINNEROBJS= amg_dmlprec_bld.o amg_dfile_prec_descr.o \
DINNEROBJS= amg_dmlprec_bld.o amg_dfile_prec_descr.o amg_dfile_prec_memory_use.o \
amg_d_smoothers_bld.o amg_d_hierarchy_bld.o amg_d_hierarchy_rebld.o \
amg_dmlprec_aply.o \
$(DMPFOBJS) amg_d_extprol_bld.o
SINNEROBJS= amg_smlprec_bld.o amg_sfile_prec_descr.o \
SINNEROBJS= amg_smlprec_bld.o amg_sfile_prec_descr.o amg_sfile_prec_memory_use.o \
amg_s_smoothers_bld.o amg_s_hierarchy_bld.o amg_s_hierarchy_rebld.o \
amg_smlprec_aply.o \
$(SMPFOBJS) amg_s_extprol_bld.o
ZINNEROBJS= amg_zmlprec_bld.o amg_zfile_prec_descr.o \
ZINNEROBJS= amg_zmlprec_bld.o amg_zfile_prec_descr.o amg_zfile_prec_memory_use.o \
amg_z_smoothers_bld.o amg_z_hierarchy_bld.o amg_z_hierarchy_rebld.o \
amg_zmlprec_aply.o \
$(ZMPFOBJS) amg_z_extprol_bld.o
CINNEROBJS= amg_cmlprec_bld.o amg_cfile_prec_descr.o \
CINNEROBJS= amg_cmlprec_bld.o amg_cfile_prec_descr.o amg_cfile_prec_memory_use.o \
amg_c_smoothers_bld.o amg_c_hierarchy_bld.o amg_c_hierarchy_rebld.o \
amg_cmlprec_aply.o \
$(CMPFOBJS) amg_c_extprol_bld.o
+16 -1
View File
@@ -62,7 +62,22 @@ amg_s_parmatch_smth_bld.o \
amg_s_parmatch_spmm_bld_inner.o
MPCOBJS=MatchBoxPC.o \
algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC.o
sendBundledMessages.o \
initialize.o \
extractUChunk.o \
isAlreadyMatched.o \
findOwnerOfGhost.o \
clean.o \
computeCandidateMate.o \
parallelComputeCandidateMateB.o \
processMatchedVertices.o \
processMatchedVerticesAndSendMessages.o \
processCrossEdge.o \
queueTransfer.o \
processMessages.o \
processExposedVertex.o \
algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC.o \
algoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP.o
OBJS = $(FOBJS) $(MPCOBJS)
+44 -7
View File
@@ -60,17 +60,43 @@ void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt* ph1_card, MilanLongInt* ph2_card ) {
#if !defined(SERIAL_MPI)
MPI_Comm C_comm=MPI_Comm_f2c(icomm);
#ifdef DEBUG
fprintf(stderr,"MatchBoxPC: rank %d nlver %ld nledge %ld [ %ld %ld ]\n",
myRank,NLVer, NLEdge,verDistance[0],verDistance[1]);
#endif
dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(NLVer, NLEdge,
#undef TIME_TRACKER
#ifdef TIME_TRACKER
double tmr = MPI_Wtime();
#endif
#if defined(OPENMP)
//fprintf(stderr,"Warning: using buggy OpenMP matching!\n");
dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(NLVer, NLEdge,
verLocPtr, verLocInd, edgeLocWeight,
verDistance, Mate,
myRank, numProcs, C_comm,
msgIndSent, msgActualSent, msgPercent,
ph0_time, ph1_time, ph2_time,
ph1_card, ph2_card );
#else
dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(NLVer, NLEdge,
verLocPtr, verLocInd, edgeLocWeight,
verDistance, Mate,
myRank, numProcs, C_comm,
msgIndSent, msgActualSent, msgPercent,
ph0_time, ph1_time, ph2_time,
ph1_card, ph2_card );
#endif
#ifdef TIME_TRACKER
tmr = MPI_Wtime() - tmr;
fprintf(stderr, "Elaboration time: %f for %ld nodes\n", tmr, NLVer);
#endif
#endif
}
@@ -88,13 +114,24 @@ void sMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
fprintf(stderr,"MatchBoxPC: rank %d nlver %ld nledge %ld [ %ld %ld ]\n",
myRank,NLVer, NLEdge,verDistance[0],verDistance[1]);
#endif
#if defined(OPENMP)
//fprintf(stderr,"Warning: using buggy OpenMP matching!\n");
salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(NLVer, NLEdge,
verLocPtr, verLocInd, edgeLocWeight,
verDistance, Mate,
myRank, numProcs, C_comm,
msgIndSent, msgActualSent, msgPercent,
ph0_time, ph1_time, ph2_time,
ph1_card, ph2_card );
#else
salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(NLVer, NLEdge,
verLocPtr, verLocInd, edgeLocWeight,
verDistance, Mate,
myRank, numProcs, C_comm,
msgIndSent, msgActualSent, msgPercent,
ph0_time, ph1_time, ph2_time,
ph1_card, ph2_card );
verLocPtr, verLocInd, edgeLocWeight,
verDistance, Mate,
myRank, numProcs, C_comm,
msgIndSent, msgActualSent, msgPercent,
ph0_time, ph1_time, ph2_time,
ph1_card, ph2_card );
#endif
#endif
}
+539 -104
View File
@@ -52,145 +52,580 @@
#ifndef _matchboxpC_H_
#define _matchboxpC_H_
//Turn on a lot of debugging information with this switch:
// Turn on a lot of debugging information with this switch:
//#define PRINT_DEBUG_INFO_
#include <stdio.h>
#include <iostream>
#include <assert.h>
#include <map>
#include <vector>
// #include "matchboxp.h"
#ifdef OPENMP
// OpenMP is included and used if and only if the OpenMP version of the matching
// is required
#include "omp.h"
#endif
#include "primitiveDataTypeDefinitions.h"
#include "dataStrStaticQueue.h"
using namespace std;
const int NUM_THREAD = 4;
const int UCHUNK = 10;
const MilanLongInt REQUEST = 1;
const MilanLongInt SUCCESS = 2;
const MilanLongInt FAILURE = 3;
const MilanLongInt SIZEINFO = 4;
const int ComputeTag = 7; // Predefined tag
const int BundleTag = 9; // Predefined tag
static vector<MilanLongInt> DEFAULT_VECTOR;
// MPI type map
template <typename T>
MPI_Datatype TypeMap();
template <>
inline MPI_Datatype TypeMap<int64_t>() { return MPI_LONG_LONG; }
template <>
inline MPI_Datatype TypeMap<int>() { return MPI_INT; }
template <>
inline MPI_Datatype TypeMap<double>() { return MPI_DOUBLE; }
template <>
inline MPI_Datatype TypeMap<float>() { return MPI_FLOAT; }
#ifdef __cplusplus
extern "C" {
extern "C"
{
#endif
#if !defined(SERIAL_MPI)
#define MilanMpiLongInt MPI_LONG_LONG
#define MilanMpiLongInt MPI_LONG_LONG
#ifndef _primitiveDataType_Definition_
#define _primitiveDataType_Definition_
//Regular integer:
#ifndef INTEGER_H
#define INTEGER_H
typedef int32_t MilanInt;
#endif
// Regular integer:
#ifndef INTEGER_H
#define INTEGER_H
typedef int32_t MilanInt;
#endif
//Regular long integer:
#ifndef LONG_INT_H
#define LONG_INT_H
#ifdef BIT64
typedef int64_t MilanLongInt;
typedef MPI_LONG MilanMpiLongInt;
#else
typedef int32_t MilanLongInt;
typedef MPI_INT MilanMpiLongInt;
#endif
#endif
// Regular long integer:
#ifndef LONG_INT_H
#define LONG_INT_H
#ifdef BIT64
typedef int64_t MilanLongInt;
typedef MPI_LONG MilanMpiLongInt;
#else
typedef int32_t MilanLongInt;
typedef MPI_INT MilanMpiLongInt;
#endif
#endif
//Regular boolean
#ifndef BOOL_H
#define BOOL_H
typedef bool MilanBool;
#endif
// Regular boolean
#ifndef BOOL_H
#define BOOL_H
typedef bool MilanBool;
#endif
//Regular double and absolute value computation:
#ifndef REAL_H
#define REAL_H
typedef double MilanReal;
typedef MPI_DOUBLE MilanMpiReal;
inline MilanReal MilanAbs(MilanReal value)
{
return fabs(value);
}
#endif
// Regular double and absolute value computation:
#ifndef REAL_H
#define REAL_H
typedef double MilanReal;
typedef MPI_DOUBLE MilanMpiReal;
inline MilanReal MilanAbs(MilanReal value)
{
return fabs(value);
}
#endif
//Regular float and absolute value computation:
#ifndef FLOAT_H
#define FLOAT_H
typedef float MilanFloat;
typedef MPI_FLOAT MilanMpiFloat;
inline MilanFloat MilanAbsFloat(MilanFloat value)
{
return fabs(value);
}
#endif
// Regular float and absolute value computation:
#ifndef FLOAT_H
#define FLOAT_H
typedef float MilanFloat;
typedef MPI_FLOAT MilanMpiFloat;
inline MilanFloat MilanAbsFloat(MilanFloat value)
{
return fabs(value);
}
#endif
//// Define the limits:
#ifndef LIMITS_H
#define LIMITS_H
//Integer Maximum and Minimum:
// #define MilanIntMax INT_MAX
// #define MilanIntMin INT_MIN
#define MilanIntMax INT32_MAX
#define MilanIntMin INT32_MIN
//// Define the limits:
#ifndef LIMITS_H
#define LIMITS_H
// Integer Maximum and Minimum:
// #define MilanIntMax INT_MAX
// #define MilanIntMin INT_MIN
#define MilanIntMax INT32_MAX
#define MilanIntMin INT32_MIN
#ifdef BIT64
#define MilanLongIntMax INT64_MAX
#define MilanLongIntMin -INT64_MAX
#else
#define MilanLongIntMax INT32_MAX
#define MilanLongIntMin -INT32_MAX
#endif
#ifdef BIT64
#define MilanLongIntMax INT64_MAX
#define MilanLongIntMin -INT64_MAX
#else
#define MilanLongIntMax INT32_MAX
#define MilanLongIntMin -INT32_MAX
#endif
#endif
#endif
// +INFINITY
const double PLUS_INFINITY = numeric_limits<int>::infinity();
const double MINUS_INFINITY = -PLUS_INFINITY;
//#define MilanRealMax LDBL_MAX
#define MilanRealMax PLUS_INFINITY
#define MilanRealMin MINUS_INFINITY
//#define MilanRealMax LDBL_MAX
#define MilanRealMax PLUS_INFINITY
#define MilanRealMin MINUS_INFINITY
#endif
//Function of find the owner of a ghost vertex using binary search:
inline MilanInt findOwnerOfGhost(MilanLongInt vtxIndex, MilanLongInt *mVerDistance,
MilanInt myRank, MilanInt numProcs);
/* These functions are only used in the experimental OMP implementation, if that
is disabled there is no reason to actually compile or reference them. */
void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC
(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt* verLocPtr, MilanLongInt* verLocInd, MilanReal* edgeLocWeight,
MilanLongInt* verDistance,
MilanLongInt* Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt* msgIndSent, MilanLongInt* msgActualSent, MilanReal* msgPercent,
MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
MilanLongInt* ph1_card, MilanLongInt* ph2_card );
// Function of find the owner of a ghost vertex using binary search:
MilanInt findOwnerOfGhost(MilanLongInt vtxIndex, MilanLongInt *mVerDistance,
MilanInt myRank, MilanInt numProcs);
MilanLongInt firstComputeCandidateMateD(MilanLongInt adj1,
MilanLongInt adj2,
MilanLongInt *verLocInd,
MilanReal *edgeLocWeight);
void salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC
(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt* verLocPtr, MilanLongInt* verLocInd, MilanFloat* edgeLocWeight,
MilanLongInt* verDistance,
MilanLongInt* Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt* msgIndSent, MilanLongInt* msgActualSent, MilanReal* msgPercent,
MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
MilanLongInt* ph1_card, MilanLongInt* ph2_card );
void queuesTransfer(vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner);
bool isAlreadyMatched(MilanLongInt node,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap);
MilanLongInt computeCandidateMateD(MilanLongInt adj1,
MilanLongInt adj2,
MilanReal *edgeLocWeight,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap);
void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt StartIndex, MilanLongInt EndIndex,
MilanLongInt *numGhostEdgesPtr,
MilanLongInt *numGhostVerticesPtr,
MilanLongInt *S,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
vector<MilanLongInt> &Counter,
vector<MilanLongInt> &verGhostPtr,
vector<MilanLongInt> &verGhostInd,
vector<MilanLongInt> &tempCounter,
vector<MilanLongInt> &GMate,
vector<MilanLongInt> &Message,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
MilanLongInt *&candidateMate,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner);
void clean(MilanLongInt NLVer,
MilanInt myRank,
MilanLongInt MessageIndex,
vector<MPI_Request> &SRequest,
vector<MPI_Status> &SStatus,
MilanInt BufferSize,
MilanLongInt *Buffer,
MilanLongInt msgActual,
MilanLongInt *msgActualSent,
MilanLongInt msgInd,
MilanLongInt *msgIndSent,
MilanLongInt NumMessagesBundled,
MilanReal *msgPercent);
void PARALLEL_COMPUTE_CANDIDATE_MATE_BD(MilanLongInt NLVer,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanInt myRank,
MilanReal *edgeLocWeight,
MilanLongInt *candidateMate);
void PARALLEL_PROCESS_EXPOSED_VERTEX_BD(MilanLongInt NLVer,
MilanLongInt *candidateMate,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *Mate,
vector<MilanLongInt> &GMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *NumMessagesBundledPtr,
MilanLongInt *SPtr,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner);
void PROCESS_CROSS_EDGE(MilanLongInt *edge,
MilanLongInt *SPtr);
void processMatchedVerticesD(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *NumMessagesBundledPtr,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner);
void processMatchedVerticesAndSendMessagesD(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *NumMessagesBundledPtr,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner,
MPI_Comm comm,
MilanLongInt *msgActual,
vector<MilanLongInt> &Message);
void sendBundledMessages(MilanLongInt *numGhostEdgesPtr,
MilanInt *BufferSizePtr,
MilanLongInt *Buffer,
vector<MilanLongInt> &PCumulative,
vector<MilanLongInt> &PMessageBundle,
vector<MilanLongInt> &PSizeInfoMessages,
MilanLongInt *PCounter,
MilanLongInt NumMessagesBundled,
MilanLongInt *msgActualPtr,
MilanLongInt *MessageIndexPtr,
MilanInt numProcs,
MilanInt myRank,
MPI_Comm comm,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MPI_Request> &SRequest,
vector<MPI_Status> &SStatus);
void processMessagesD(
MilanLongInt NLVer,
MilanLongInt *Mate,
MilanLongInt *candidateMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
vector<MilanLongInt> &GMate,
vector<MilanLongInt> &Counter,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *msgActualPtr,
MilanReal *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *verLocPtr,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanInt numProcs,
MilanInt myRank,
MPI_Comm comm,
vector<MilanLongInt> &Message,
MilanLongInt numGhostEdges,
MilanLongInt u,
MilanLongInt v,
MilanLongInt *SPtr,
vector<MilanLongInt> &U);
void extractUChunk(
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU);
void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt* verLocPtr, MilanLongInt* verLocInd, MilanReal* edgeLocWeight,
MilanLongInt* verDistance,
MilanLongInt* Mate,
MilanInt myRank, MilanInt numProcs, MilanInt icomm,
MilanLongInt* msgIndSent, MilanLongInt* msgActualSent, MilanReal* msgPercent,
MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
MilanLongInt* ph1_card, MilanLongInt* ph2_card );
MilanLongInt firstComputeCandidateMateS(MilanLongInt adj1,
MilanLongInt adj2,
MilanLongInt *verLocInd,
MilanFloat *edgeLocWeight);
void sMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt* verLocPtr, MilanLongInt* verLocInd, MilanFloat* edgeLocWeight,
MilanLongInt* verDistance,
MilanLongInt* Mate,
MilanInt myRank, MilanInt numProcs, MilanInt icomm,
MilanLongInt* msgIndSent, MilanLongInt* msgActualSent, MilanReal* msgPercent,
MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
MilanLongInt* ph1_card, MilanLongInt* ph2_card );
MilanLongInt computeCandidateMateS(MilanLongInt adj1,
MilanLongInt adj2,
MilanFloat *edgeLocWeight,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap);
void PARALLEL_COMPUTE_CANDIDATE_MATE_BS(MilanLongInt NLVer,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanInt myRank,
MilanFloat *edgeLocWeight,
MilanLongInt *candidateMate);
void PARALLEL_PROCESS_EXPOSED_VERTEX_BS(MilanLongInt NLVer,
MilanLongInt *candidateMate,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *Mate,
vector<MilanLongInt> &GMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanFloat *edgeLocWeight,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *NumMessagesBundledPtr,
MilanLongInt *SPtr,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner);
void processMatchedVerticesS(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *NumMessagesBundledPtr,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanFloat *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner);
void processMatchedVerticesAndSendMessagesS(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *NumMessagesBundledPtr,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanFloat *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner,
MPI_Comm comm,
MilanLongInt *msgActual,
vector<MilanLongInt> &Message);
void processMessagesS(
MilanLongInt NLVer,
MilanLongInt *Mate,
MilanLongInt *candidateMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
vector<MilanLongInt> &GMate,
vector<MilanLongInt> &Counter,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCardPtr,
MilanLongInt *msgIndPtr,
MilanLongInt *msgActualPtr,
MilanFloat *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *verLocPtr,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanInt numProcs,
MilanInt myRank,
MPI_Comm comm,
vector<MilanLongInt> &Message,
MilanLongInt numGhostEdges,
MilanLongInt u,
MilanLongInt v,
MilanLongInt *SPtr,
vector<MilanLongInt> &U);
void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt *verLocPtr, MilanLongInt *verLocInd, MilanReal *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt *msgIndSent, MilanLongInt *msgActualSent, MilanReal *msgPercent,
MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
MilanLongInt *ph1_card, MilanLongInt *ph2_card);
void salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateCMP(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt *verLocPtr, MilanLongInt *verLocInd, MilanFloat *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt *msgIndSent, MilanLongInt *msgActualSent, MilanReal *msgPercent,
MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
MilanLongInt *ph1_card, MilanLongInt *ph2_card);
void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt *verLocPtr, MilanLongInt *verLocInd, MilanReal *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt *msgIndSent, MilanLongInt *msgActualSent, MilanReal *msgPercent,
MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
MilanLongInt *ph1_card, MilanLongInt *ph2_card);
void salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt *verLocPtr, MilanLongInt *verLocInd, MilanFloat *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt *msgIndSent, MilanLongInt *msgActualSent, MilanReal *msgPercent,
MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
MilanLongInt *ph1_card, MilanLongInt *ph2_card);
void dMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt *verLocPtr, MilanLongInt *verLocInd, MilanReal *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *Mate,
MilanInt myRank, MilanInt numProcs, MilanInt icomm,
MilanLongInt *msgIndSent, MilanLongInt *msgActualSent, MilanReal *msgPercent,
MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
MilanLongInt *ph1_card, MilanLongInt *ph2_card);
void sMatchBoxPC(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt *verLocPtr, MilanLongInt *verLocInd, MilanFloat *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *Mate,
MilanInt myRank, MilanInt numProcs, MilanInt icomm,
MilanLongInt *msgIndSent, MilanLongInt *msgActualSent, MilanReal *msgPercent,
MilanReal *ph0_time, MilanReal *ph1_time, MilanReal *ph2_time,
MilanLongInt *ph1_card, MilanLongInt *ph2_card);
#endif
#ifdef __cplusplus
@@ -72,12 +72,6 @@
#ifdef SERIAL_MPI
#else
//MPI type map
template<typename T> MPI_Datatype TypeMap();
template<> inline MPI_Datatype TypeMap<int64_t>() { return MPI_LONG_LONG; }
template<> inline MPI_Datatype TypeMap<int>() { return MPI_INT; }
template<> inline MPI_Datatype TypeMap<double>() { return MPI_DOUBLE; }
template<> inline MPI_Datatype TypeMap<float>() { return MPI_FLOAT; }
// DOUBLE PRECISION VERSION
//WARNING: The vertex block on a given rank is contiguous
@@ -1309,16 +1303,16 @@ void dalgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
// SINGLE PRECISION VERSION
void salgoDistEdgeApproxDomEdgesLinearSearchMesgBndlSmallMateC(
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt* verLocPtr, MilanLongInt* verLocInd,
MilanFloat* edgeLocWeight,
MilanLongInt* verDistance,
MilanLongInt* Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt* msgIndSent, MilanLongInt* msgActualSent,
MilanReal* msgPercent,
MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
MilanLongInt* ph1_card, MilanLongInt* ph2_card ) {
MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt* verLocPtr, MilanLongInt* verLocInd,
MilanFloat* edgeLocWeight,
MilanLongInt* verDistance,
MilanLongInt* Mate,
MilanInt myRank, MilanInt numProcs, MPI_Comm comm,
MilanLongInt* msgIndSent, MilanLongInt* msgActualSent,
MilanReal* msgPercent,
MilanReal* ph0_time, MilanReal* ph1_time, MilanReal* ph2_time,
MilanLongInt* ph1_card, MilanLongInt* ph2_card ) {
#if !defined(SERIAL_MPI)
#ifdef PRINT_DEBUG_INFO_
cout<<"\n("<<myRank<<")Within algoEdgeApproxDominatingEdgesLinearSearchMessageBundling()"; fflush(stdout);
@@ -97,6 +97,8 @@ subroutine amg_c_dec_aggregator_build_tprol(ag,parms,ag_data,&
integer(psb_lpk_) :: ntaggr
integer(psb_ipk_) :: debug_level, debug_unit
logical :: clean_zeros
integer(psb_ipk_), save :: idx_map_bld=-1, idx_map_tprol=-1
logical, parameter :: do_timings=.false.
name='amg_c_dec_aggregator_tprol'
call psb_erractionsave(err_act)
@@ -108,6 +110,10 @@ subroutine amg_c_dec_aggregator_build_tprol(ag,parms,ag_data,&
info = psb_success_
ctxt = desc_a%get_context()
call psb_info(ctxt,me,np)
if ((do_timings).and.(idx_map_bld==-1)) &
& idx_map_bld = psb_get_timer_idx("DEC_TPROL: map_bld")
if ((do_timings).and.(idx_map_tprol==-1)) &
& idx_map_tprol = psb_get_timer_idx("DEC_TPROL: map_tprol")
call amg_check_def(parms%ml_cycle,'Multilevel cycle',&
& amg_mult_ml_,is_legal_ml_cycle)
@@ -121,10 +127,14 @@ subroutine amg_c_dec_aggregator_build_tprol(ag,parms,ag_data,&
! The decoupled aggregator based on SOC measures ignores
! ag_data except for clean_zeros; soc_map_bld is a procedure pointer.
!
if (do_timings) call psb_tic(idx_map_bld)
clean_zeros = ag%do_clean_zeros
call ag%soc_map_bld(parms%aggr_ord,parms%aggr_thresh,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
if (do_timings) call psb_toc(idx_map_bld)
if (do_timings) call psb_tic(idx_map_tprol)
if (info==psb_success_) call amg_map_to_tprol(desc_a,ilaggr,nlaggr,t_prol,info)
if (do_timings) call psb_toc(idx_map_tprol)
if (info /= psb_success_) then
info=psb_err_from_subroutine_
call psb_errpush(info,name,a_err='soc_map_bld/map_to_tprol')
+20 -14
View File
@@ -76,7 +76,7 @@ subroutine amg_c_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
integer(psb_ipk_) :: nrow, ncol, nrl, nzl, ip, nzt, i, k
integer(psb_lpk_) :: nrsave, ncsave, nzsave, nza
logical, parameter :: do_timings=.false., oldstyle=.false., debug=.false.
integer(psb_ipk_), save :: idx_spspmm=-1
integer(psb_ipk_), save :: idx_spspmm=-1, idx_cpytrans1=-1, idx_cpytrans2=-1
name='amg_ptap_bld'
if(psb_get_errstatus().ne.0) return
@@ -93,7 +93,11 @@ subroutine amg_c_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
ncol = desc_a%get_local_cols()
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("SPMM_BLD: par_spspmm")
& idx_spspmm = psb_get_timer_idx("PTAP_BLD: par_spspmm")
if ((do_timings).and.(idx_cpytrans1==-1)) &
& idx_cpytrans1 = psb_get_timer_idx("PTAP_BLD: cpy&trans1")
if ((do_timings).and.(idx_cpytrans2==-1)) &
& idx_cpytrans2 = psb_get_timer_idx("PTAP_BLD: cpy&trans2")
naggr = nlaggr(me+1)
ntaggr = sum(nlaggr)
@@ -128,6 +132,7 @@ subroutine amg_c_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
! Ok first product done.
if (present(desc_ax)) then
if (do_timings) call psb_tic(idx_cpytrans1)
block
call coo_prol%cp_to_coo(coo_restr,info)
call coo_restr%set_ncols(desc_ac%get_local_cols())
@@ -137,7 +142,7 @@ subroutine amg_c_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%set_ncols(desc_ax%get_local_cols())
end block
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans1)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -167,27 +172,28 @@ subroutine amg_c_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%transp()
nzl = coo_restr%get_nzeros()
nrl = desc_ac%get_local_rows()
i=0
nrl = desc_ac%get_local_rows()
call coo_restr%fix(info)
i=coo_restr%get_nzeros()
!
! Only keep local rows
!
do k=1, nzl
if ((1 <= coo_restr%ia(k)) .and.(coo_restr%ia(k) <= nrl)) then
i = i+1
coo_restr%val(i) = coo_restr%val(k)
coo_restr%ia(i) = coo_restr%ia(k)
coo_restr%ja(i) = coo_restr%ja(k)
search: do k=i,1,-1
if (coo_restr%ia(k) <= nrl) then
call coo_restr%set_nzeros(k)
exit search
end if
end do
call coo_restr%set_nzeros(i)
call coo_restr%fix(info)
end do search
nzl = coo_restr%get_nzeros()
call coo_restr%set_nrows(desc_ac%get_local_rows())
call coo_restr%set_ncols(desc_a%get_local_cols())
if (debug) call check_coo(me,trim(name)//' Check 2 on coo_restr:',coo_restr)
if (do_timings) call psb_tic(idx_cpytrans2)
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans2)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -72,7 +72,9 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
use psb_base_mod
use amg_base_prec_type
use amg_c_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
! Arguments
@@ -85,7 +87,7 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_), intent(out) :: info
! Local variables
integer(psb_ipk_), allocatable :: ils(:), neigh(:), irow(:), icol(:),&
integer(psb_ipk_), allocatable :: neigh(:), irow(:), icol(:),&
& ideg(:), idxs(:)
integer(psb_lpk_), allocatable :: tmpaggr(:)
complex(psb_spk_), allocatable :: val(:), diag(:)
@@ -99,6 +101,9 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: nrow, ncol, n_ne
integer(psb_lpk_) :: nrglob
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc1_p1=-1, idx_soc1_p2=-1, idx_soc1_p3=-1
integer(psb_ipk_), save :: idx_soc1_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc1_map_bld'
@@ -114,6 +119,14 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc1_p0==-1)) &
& idx_soc1_p0 = psb_get_timer_idx("SOC1_MAP: phase0")
if ((do_timings).and.(idx_soc1_p1==-1)) &
& idx_soc1_p1 = psb_get_timer_idx("SOC1_MAP: phase1")
if ((do_timings).and.(idx_soc1_p2==-1)) &
& idx_soc1_p2 = psb_get_timer_idx("SOC1_MAP: phase2")
if ((do_timings).and.(idx_soc1_p3==-1)) &
& idx_soc1_p3 = psb_get_timer_idx("SOC1_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -133,41 +146,205 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc1_p0)
call a%cp_to(acsr)
if (do_timings) call psb_toc(idx_soc1_p0)
if (clean_zeros) call acsr%clean_zeros(info)
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
idxs(i) = i
end do
else
!$omp end parallel do
else
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = acsr%irp(i+1) - acsr%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_tic(idx_soc1_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
icnt = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz, nc, i,j,m, nz, ilg, ip, rsz
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) cycle step1
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
val(1:nz) = acsr%val(acsr%irp(i):acsr%irp(i+1)-1)
!
! Build the set of all strongly coupled nodes
!
ip = 0
do k=1, nz
j = icol(k)
! If any of the neighbours is already assigned,
! we will not reset.
if (j>nr) cycle step1
if (ilaggr(j) > 0) cycle step1
if (abs(val(k)) > theta*sqrt(abs(diag(i)*diag(j)))) then
ip = ip + 1
icol(ip) = icol(k)
end if
enddo
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
!$omp end atomic
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, ip
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
step1: do ii=1, nr
if (info /= 0) cycle
i = idxs(ii)
if ((i<1).or.(i>nr)) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
@@ -176,7 +353,7 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
! Build the set of all strongly coupled nodes
!
ip = 0
ip = 0
do k=1, nz
j = icol(k)
if ((1<=j).and.(j<=nr)) then
@@ -194,8 +371,7 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! contains I even if it does not look like it from matrix)
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
icnt = icnt + 1
if (disjoint) then
naggr = naggr + 1
do k=1, ip
ilaggr(icol(k)) = naggr
@@ -204,16 +380,22 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
& ' Check 1:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_toc(idx_soc1_p1)
if (do_timings) call psb_tic(idx_soc1_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,theta)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -244,8 +426,15 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc1_p2)
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1.5:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_tic(idx_soc1_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -274,7 +463,6 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
enddo
if (ip > 0) then
icnt = icnt + 1
naggr = naggr + 1
ilaggr(i) = naggr
do k=1, ip
@@ -292,7 +480,10 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,info)&
!$omp private(ii,i,j,k,nz,icol,val,ip)
do i=1, nr
if (info /= 0) cycle
if (ilaggr(i) < 0) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if (nz == 1) then
@@ -303,15 +494,18 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc1_p3)
if (naggr > ncol) then
!write(0,*) name,'Error : naggr > ncol',naggr,ncol
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
goto 9999
@@ -336,9 +530,13 @@ subroutine amg_c_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nlaggr(:) = 0
nlaggr(me+1) = naggr
call psb_sum(ctxt,nlaggr(1:np))
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 2:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
call acsr%free()
call psb_erractionrestore(err_act)
return
@@ -68,9 +68,12 @@
!
subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
use psb_base_mod
use psb_base_mod
use amg_base_prec_type
use amg_c_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
@@ -99,6 +102,9 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: np, me
integer(psb_ipk_) :: nrow, ncol, n_ne
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc2_p1=-1, idx_soc2_p2=-1, idx_soc2_p3=-1
integer(psb_ipk_), save :: idx_soc2_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc2_map_bld'
@@ -114,6 +120,14 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc2_p0==-1)) &
& idx_soc2_p0 = psb_get_timer_idx("SOC2_MAP: phase0")
if ((do_timings).and.(idx_soc2_p1==-1)) &
& idx_soc2_p1 = psb_get_timer_idx("SOC2_MAP: phase1")
if ((do_timings).and.(idx_soc2_p2==-1)) &
& idx_soc2_p2 = psb_get_timer_idx("SOC2_MAP: phase2")
if ((do_timings).and.(idx_soc2_p3==-1)) &
& idx_soc2_p3 = psb_get_timer_idx("SOC2_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -125,6 +139,7 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc2_p0)
diag = a%get_diag(info)
if(info /= psb_success_) then
info=psb_err_from_subroutine_
@@ -137,55 +152,218 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
call a%cp_to(muij)
if (clean_zeros) call muij%clean_zeros(info)
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
if (j<= nr) muij%val(k) = abs(muij%val(k))/sqrt(abs(diag(i)*diag(j)))
end do
end do
!$omp end parallel do
!
! Compute the 1-neigbour; mark strong links with +1, weak links with -1
!
call s_neigh_coo%allocate(nr,nr,muij%get_nzeros())
ip = 0
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
s_neigh_coo%ia(k) = i
s_neigh_coo%ja(k) = j
if (j<=nr) then
ip = ip + 1
s_neigh_coo%ia(ip) = i
s_neigh_coo%ja(ip) = j
if (real(muij%val(k)) >= theta) then
s_neigh_coo%val(ip) = sone
s_neigh_coo%val(k) = sone
else
s_neigh_coo%val(ip) = -sone
s_neigh_coo%val(k) = -sone
end if
else
s_neigh_coo%val(k) = -sone
end if
end do
end do
!$omp end parallel do
!write(*,*) 'S_NEIGH: ',nr,ip
call s_neigh_coo%set_nzeros(ip)
call s_neigh_coo%set_nzeros(muij%get_nzeros())
call s_neigh%mv_from_coo(s_neigh_coo,info)
if (iorder == amg_aggr_ord_nat_) then
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) shared(ilaggr,idxs) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
end do
!$omp end parallel do
else
!$omp parallel do private(i) shared(ilaggr,idxs,muij) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = muij%irp(i+1) - muij%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_toc(idx_soc2_p0)
if (do_timings) call psb_tic(idx_soc2_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(s_neigh,bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz,nc,i,j,m,nz,ilg,ip,rsz,ip1,nzcnt
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) then
write(0,*) ' Step1:',kk,ii,i,info
cycle step1
end if
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
!
! Get the 1-neighbourhood of I
!
ip1 = s_neigh%irp(i)
nz = s_neigh%irp(i+1)-ip1
!
! If the neighbourhood only contains I, skip it
!
if (nz ==0) then
ilaggr(i) = 0
cycle step1
end if
if ((nz==1).and.(s_neigh%ja(ip1)==i)) then
ilaggr(i) = 0
cycle step1
end if
nzcnt = count(real(s_neigh%val(ip1:ip1+nz-1)) > 0)
icol(1:nzcnt) = pack(s_neigh%ja(ip1:ip1+nz-1),(real(s_neigh%val(ip1:ip1+nz-1)) > 0))
disjoint = all(ilaggr(icol(1:nzcnt)) == -(nr+1))
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
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
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, nzcnt
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!write(0,*) 'LNAG ',locnaggr(nths+1)
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
icnt = 0
step1: do ii=1, nr
i = idxs(ii)
@@ -224,16 +402,21 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
end if
if (do_timings) call psb_toc(idx_soc2_p1)
if (do_timings) call psb_tic(idx_soc2_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,muij,s_neigh)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -259,8 +442,9 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc2_p2)
if (do_timings) call psb_tic(idx_soc2_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -294,6 +478,8 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,s_neigh,info)&
!$omp private(ii,i,j,k)
do i=1, nr
if (ilaggr(i) <= 0) then
nz = (s_neigh%irp(i+1)-s_neigh%irp(i))
@@ -305,13 +491,17 @@ subroutine amg_c_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc2_p3)
if (naggr > ncol) then
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
@@ -140,6 +140,9 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
real(psb_spk_) :: anorm, omega, tmp, dg, theta
logical, parameter :: debug_new=.false.
character(len=80) :: filename
logical, parameter :: do_timings=.false.
integer(psb_ipk_), save :: idx_spspmm=-1, idx_phase1=-1, idx_gtrans=-1, idx_phase2=-1, idx_refine=-1
integer(psb_ipk_), save :: idx_phase3=-1, idx_cdasb=-1, idx_ptap=-1
name='amg_aggrmat_smth_bld'
info=psb_success_
@@ -153,6 +156,23 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
ctxt = desc_a%get_context()
call psb_info(ctxt, me, np)
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("DEC_SMTH_BLD: par_spspmm")
if ((do_timings).and.(idx_phase1==-1)) &
& idx_phase1 = psb_get_timer_idx("DEC_SMTH_BLD: phase1 ")
if ((do_timings).and.(idx_phase2==-1)) &
& idx_phase2 = psb_get_timer_idx("DEC_SMTH_BLD: phase2 ")
if ((do_timings).and.(idx_phase3==-1)) &
& idx_phase3 = psb_get_timer_idx("DEC_SMTH_BLD: phase3 ")
if ((do_timings).and.(idx_gtrans==-1)) &
& idx_gtrans = psb_get_timer_idx("DEC_SMTH_BLD: gtrans ")
if ((do_timings).and.(idx_refine==-1)) &
& idx_refine = psb_get_timer_idx("DEC_SMTH_BLD: refine ")
if ((do_timings).and.(idx_cdasb==-1)) &
& idx_cdasb = psb_get_timer_idx("DEC_SMTH_BLD: cdasb ")
if ((do_timings).and.(idx_ptap==-1)) &
& idx_ptap = psb_get_timer_idx("DEC_SMTH_BLD: ptap_bld ")
nglob = desc_a%get_global_rows()
nrow = desc_a%get_local_rows()
@@ -171,6 +191,7 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
! naggr: number of local aggregates
! nrow: local rows.
!
if (do_timings) call psb_tic(idx_phase1)
! Get the diagonal D
adiag = a%get_diag(info)
@@ -196,7 +217,7 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
!
! Build the filtered matrix Af from A
!
!$OMP parallel do private(i,j,tmp,jd) schedule(static)
do i=1, nrow
tmp = czero
jd = -1
@@ -214,11 +235,13 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
acsrf%val(jd)=acsrf%val(jd)-tmp
end if
enddo
!$OMP end parallel do
! Take out zeroed terms
call acsrf%clean_zeros(info)
end if
!$OMP parallel do private(i) schedule(static)
do i=1,size(adiag)
if (adiag(i) /= czero) then
adiag(i) = cone / adiag(i)
@@ -226,7 +249,7 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
adiag(i) = cone
end if
end do
!$OMP end parallel do
if (parms%aggr_omega_alg == amg_eig_est_) then
if (parms%aggr_eig == amg_max_norm_) then
@@ -252,8 +275,9 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(info,name,a_err='invalid amg_aggr_omega_alg_')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase1)
if (do_timings) call psb_tic(idx_phase2)
call acsrf%scal(adiag,info)
if (info /= psb_success_) goto 9999
@@ -267,6 +291,8 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_cdasb(desc_ac,info)
call psb_cd_reinit(desc_ac,info)
if (do_timings) call psb_toc(idx_phase2)
if (do_timings) call psb_tic(idx_phase3)
!
! Build the smoothed prolongator using either A or Af
! acsr1 = (I-w*D*A) Prol acsr1 = (I-w*D*Af) Prol
@@ -279,8 +305,8 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(psb_err_from_subroutine_,name,a_err='spspmm 1')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase3)
if (do_timings) call psb_tic(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done SPSPMM 1'
@@ -292,7 +318,7 @@ subroutine amg_caggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call op_prol%mv_from(coo_prol)
call op_restr%mv_from(coo_restr)
if (do_timings) call psb_toc(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done smooth_aggregate '
@@ -97,6 +97,8 @@ subroutine amg_d_dec_aggregator_build_tprol(ag,parms,ag_data,&
integer(psb_lpk_) :: ntaggr
integer(psb_ipk_) :: debug_level, debug_unit
logical :: clean_zeros
integer(psb_ipk_), save :: idx_map_bld=-1, idx_map_tprol=-1
logical, parameter :: do_timings=.false.
name='amg_d_dec_aggregator_tprol'
call psb_erractionsave(err_act)
@@ -108,6 +110,10 @@ subroutine amg_d_dec_aggregator_build_tprol(ag,parms,ag_data,&
info = psb_success_
ctxt = desc_a%get_context()
call psb_info(ctxt,me,np)
if ((do_timings).and.(idx_map_bld==-1)) &
& idx_map_bld = psb_get_timer_idx("DEC_TPROL: map_bld")
if ((do_timings).and.(idx_map_tprol==-1)) &
& idx_map_tprol = psb_get_timer_idx("DEC_TPROL: map_tprol")
call amg_check_def(parms%ml_cycle,'Multilevel cycle',&
& amg_mult_ml_,is_legal_ml_cycle)
@@ -121,10 +127,14 @@ subroutine amg_d_dec_aggregator_build_tprol(ag,parms,ag_data,&
! The decoupled aggregator based on SOC measures ignores
! ag_data except for clean_zeros; soc_map_bld is a procedure pointer.
!
if (do_timings) call psb_tic(idx_map_bld)
clean_zeros = ag%do_clean_zeros
call ag%soc_map_bld(parms%aggr_ord,parms%aggr_thresh,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
if (do_timings) call psb_toc(idx_map_bld)
if (do_timings) call psb_tic(idx_map_tprol)
if (info==psb_success_) call amg_map_to_tprol(desc_a,ilaggr,nlaggr,t_prol,info)
if (do_timings) call psb_toc(idx_map_tprol)
if (info /= psb_success_) then
info=psb_err_from_subroutine_
call psb_errpush(info,name,a_err='soc_map_bld/map_to_tprol')
+20 -14
View File
@@ -76,7 +76,7 @@ subroutine amg_d_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
integer(psb_ipk_) :: nrow, ncol, nrl, nzl, ip, nzt, i, k
integer(psb_lpk_) :: nrsave, ncsave, nzsave, nza
logical, parameter :: do_timings=.false., oldstyle=.false., debug=.false.
integer(psb_ipk_), save :: idx_spspmm=-1
integer(psb_ipk_), save :: idx_spspmm=-1, idx_cpytrans1=-1, idx_cpytrans2=-1
name='amg_ptap_bld'
if(psb_get_errstatus().ne.0) return
@@ -93,7 +93,11 @@ subroutine amg_d_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
ncol = desc_a%get_local_cols()
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("SPMM_BLD: par_spspmm")
& idx_spspmm = psb_get_timer_idx("PTAP_BLD: par_spspmm")
if ((do_timings).and.(idx_cpytrans1==-1)) &
& idx_cpytrans1 = psb_get_timer_idx("PTAP_BLD: cpy&trans1")
if ((do_timings).and.(idx_cpytrans2==-1)) &
& idx_cpytrans2 = psb_get_timer_idx("PTAP_BLD: cpy&trans2")
naggr = nlaggr(me+1)
ntaggr = sum(nlaggr)
@@ -128,6 +132,7 @@ subroutine amg_d_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
! Ok first product done.
if (present(desc_ax)) then
if (do_timings) call psb_tic(idx_cpytrans1)
block
call coo_prol%cp_to_coo(coo_restr,info)
call coo_restr%set_ncols(desc_ac%get_local_cols())
@@ -137,7 +142,7 @@ subroutine amg_d_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%set_ncols(desc_ax%get_local_cols())
end block
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans1)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -167,27 +172,28 @@ subroutine amg_d_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%transp()
nzl = coo_restr%get_nzeros()
nrl = desc_ac%get_local_rows()
i=0
nrl = desc_ac%get_local_rows()
call coo_restr%fix(info)
i=coo_restr%get_nzeros()
!
! Only keep local rows
!
do k=1, nzl
if ((1 <= coo_restr%ia(k)) .and.(coo_restr%ia(k) <= nrl)) then
i = i+1
coo_restr%val(i) = coo_restr%val(k)
coo_restr%ia(i) = coo_restr%ia(k)
coo_restr%ja(i) = coo_restr%ja(k)
search: do k=i,1,-1
if (coo_restr%ia(k) <= nrl) then
call coo_restr%set_nzeros(k)
exit search
end if
end do
call coo_restr%set_nzeros(i)
call coo_restr%fix(info)
end do search
nzl = coo_restr%get_nzeros()
call coo_restr%set_nrows(desc_ac%get_local_rows())
call coo_restr%set_ncols(desc_a%get_local_cols())
if (debug) call check_coo(me,trim(name)//' Check 2 on coo_restr:',coo_restr)
if (do_timings) call psb_tic(idx_cpytrans2)
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans2)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -72,7 +72,9 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
use psb_base_mod
use amg_base_prec_type
use amg_d_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
! Arguments
@@ -85,7 +87,7 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_), intent(out) :: info
! Local variables
integer(psb_ipk_), allocatable :: ils(:), neigh(:), irow(:), icol(:),&
integer(psb_ipk_), allocatable :: neigh(:), irow(:), icol(:),&
& ideg(:), idxs(:)
integer(psb_lpk_), allocatable :: tmpaggr(:)
real(psb_dpk_), allocatable :: val(:), diag(:)
@@ -99,6 +101,9 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: nrow, ncol, n_ne
integer(psb_lpk_) :: nrglob
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc1_p1=-1, idx_soc1_p2=-1, idx_soc1_p3=-1
integer(psb_ipk_), save :: idx_soc1_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc1_map_bld'
@@ -114,6 +119,14 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc1_p0==-1)) &
& idx_soc1_p0 = psb_get_timer_idx("SOC1_MAP: phase0")
if ((do_timings).and.(idx_soc1_p1==-1)) &
& idx_soc1_p1 = psb_get_timer_idx("SOC1_MAP: phase1")
if ((do_timings).and.(idx_soc1_p2==-1)) &
& idx_soc1_p2 = psb_get_timer_idx("SOC1_MAP: phase2")
if ((do_timings).and.(idx_soc1_p3==-1)) &
& idx_soc1_p3 = psb_get_timer_idx("SOC1_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -133,41 +146,205 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc1_p0)
call a%cp_to(acsr)
if (do_timings) call psb_toc(idx_soc1_p0)
if (clean_zeros) call acsr%clean_zeros(info)
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
idxs(i) = i
end do
else
!$omp end parallel do
else
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = acsr%irp(i+1) - acsr%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_tic(idx_soc1_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
icnt = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz, nc, i,j,m, nz, ilg, ip, rsz
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) cycle step1
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
val(1:nz) = acsr%val(acsr%irp(i):acsr%irp(i+1)-1)
!
! Build the set of all strongly coupled nodes
!
ip = 0
do k=1, nz
j = icol(k)
! If any of the neighbours is already assigned,
! we will not reset.
if (j>nr) cycle step1
if (ilaggr(j) > 0) cycle step1
if (abs(val(k)) > theta*sqrt(abs(diag(i)*diag(j)))) then
ip = ip + 1
icol(ip) = icol(k)
end if
enddo
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
!$omp end atomic
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, ip
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
step1: do ii=1, nr
if (info /= 0) cycle
i = idxs(ii)
if ((i<1).or.(i>nr)) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
@@ -176,7 +353,7 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
! Build the set of all strongly coupled nodes
!
ip = 0
ip = 0
do k=1, nz
j = icol(k)
if ((1<=j).and.(j<=nr)) then
@@ -194,8 +371,7 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! contains I even if it does not look like it from matrix)
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
icnt = icnt + 1
if (disjoint) then
naggr = naggr + 1
do k=1, ip
ilaggr(icol(k)) = naggr
@@ -204,16 +380,22 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
& ' Check 1:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_toc(idx_soc1_p1)
if (do_timings) call psb_tic(idx_soc1_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,theta)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -244,8 +426,15 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc1_p2)
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1.5:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_tic(idx_soc1_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -274,7 +463,6 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
enddo
if (ip > 0) then
icnt = icnt + 1
naggr = naggr + 1
ilaggr(i) = naggr
do k=1, ip
@@ -292,7 +480,10 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,info)&
!$omp private(ii,i,j,k,nz,icol,val,ip)
do i=1, nr
if (info /= 0) cycle
if (ilaggr(i) < 0) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if (nz == 1) then
@@ -303,15 +494,18 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc1_p3)
if (naggr > ncol) then
!write(0,*) name,'Error : naggr > ncol',naggr,ncol
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
goto 9999
@@ -336,9 +530,13 @@ subroutine amg_d_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nlaggr(:) = 0
nlaggr(me+1) = naggr
call psb_sum(ctxt,nlaggr(1:np))
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 2:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
call acsr%free()
call psb_erractionrestore(err_act)
return
@@ -68,9 +68,12 @@
!
subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
use psb_base_mod
use psb_base_mod
use amg_base_prec_type
use amg_d_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
@@ -99,6 +102,9 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: np, me
integer(psb_ipk_) :: nrow, ncol, n_ne
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc2_p1=-1, idx_soc2_p2=-1, idx_soc2_p3=-1
integer(psb_ipk_), save :: idx_soc2_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc2_map_bld'
@@ -114,6 +120,14 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc2_p0==-1)) &
& idx_soc2_p0 = psb_get_timer_idx("SOC2_MAP: phase0")
if ((do_timings).and.(idx_soc2_p1==-1)) &
& idx_soc2_p1 = psb_get_timer_idx("SOC2_MAP: phase1")
if ((do_timings).and.(idx_soc2_p2==-1)) &
& idx_soc2_p2 = psb_get_timer_idx("SOC2_MAP: phase2")
if ((do_timings).and.(idx_soc2_p3==-1)) &
& idx_soc2_p3 = psb_get_timer_idx("SOC2_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -125,6 +139,7 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc2_p0)
diag = a%get_diag(info)
if(info /= psb_success_) then
info=psb_err_from_subroutine_
@@ -137,55 +152,218 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
call a%cp_to(muij)
if (clean_zeros) call muij%clean_zeros(info)
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
if (j<= nr) muij%val(k) = abs(muij%val(k))/sqrt(abs(diag(i)*diag(j)))
end do
end do
!$omp end parallel do
!
! Compute the 1-neigbour; mark strong links with +1, weak links with -1
!
call s_neigh_coo%allocate(nr,nr,muij%get_nzeros())
ip = 0
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
s_neigh_coo%ia(k) = i
s_neigh_coo%ja(k) = j
if (j<=nr) then
ip = ip + 1
s_neigh_coo%ia(ip) = i
s_neigh_coo%ja(ip) = j
if (real(muij%val(k)) >= theta) then
s_neigh_coo%val(ip) = done
s_neigh_coo%val(k) = done
else
s_neigh_coo%val(ip) = -done
s_neigh_coo%val(k) = -done
end if
else
s_neigh_coo%val(k) = -done
end if
end do
end do
!$omp end parallel do
!write(*,*) 'S_NEIGH: ',nr,ip
call s_neigh_coo%set_nzeros(ip)
call s_neigh_coo%set_nzeros(muij%get_nzeros())
call s_neigh%mv_from_coo(s_neigh_coo,info)
if (iorder == amg_aggr_ord_nat_) then
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) shared(ilaggr,idxs) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
end do
!$omp end parallel do
else
!$omp parallel do private(i) shared(ilaggr,idxs,muij) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = muij%irp(i+1) - muij%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_toc(idx_soc2_p0)
if (do_timings) call psb_tic(idx_soc2_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(s_neigh,bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz,nc,i,j,m,nz,ilg,ip,rsz,ip1,nzcnt
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) then
write(0,*) ' Step1:',kk,ii,i,info
cycle step1
end if
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
!
! Get the 1-neighbourhood of I
!
ip1 = s_neigh%irp(i)
nz = s_neigh%irp(i+1)-ip1
!
! If the neighbourhood only contains I, skip it
!
if (nz ==0) then
ilaggr(i) = 0
cycle step1
end if
if ((nz==1).and.(s_neigh%ja(ip1)==i)) then
ilaggr(i) = 0
cycle step1
end if
nzcnt = count(real(s_neigh%val(ip1:ip1+nz-1)) > 0)
icol(1:nzcnt) = pack(s_neigh%ja(ip1:ip1+nz-1),(real(s_neigh%val(ip1:ip1+nz-1)) > 0))
disjoint = all(ilaggr(icol(1:nzcnt)) == -(nr+1))
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
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
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, nzcnt
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!write(0,*) 'LNAG ',locnaggr(nths+1)
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
icnt = 0
step1: do ii=1, nr
i = idxs(ii)
@@ -224,16 +402,21 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
end if
if (do_timings) call psb_toc(idx_soc2_p1)
if (do_timings) call psb_tic(idx_soc2_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,muij,s_neigh)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -259,8 +442,9 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc2_p2)
if (do_timings) call psb_tic(idx_soc2_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -294,6 +478,8 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,s_neigh,info)&
!$omp private(ii,i,j,k)
do i=1, nr
if (ilaggr(i) <= 0) then
nz = (s_neigh%irp(i+1)-s_neigh%irp(i))
@@ -305,13 +491,17 @@ subroutine amg_d_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc2_p3)
if (naggr > ncol) then
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
@@ -140,6 +140,9 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
real(psb_dpk_) :: anorm, omega, tmp, dg, theta
logical, parameter :: debug_new=.false.
character(len=80) :: filename
logical, parameter :: do_timings=.false.
integer(psb_ipk_), save :: idx_spspmm=-1, idx_phase1=-1, idx_gtrans=-1, idx_phase2=-1, idx_refine=-1
integer(psb_ipk_), save :: idx_phase3=-1, idx_cdasb=-1, idx_ptap=-1
name='amg_aggrmat_smth_bld'
info=psb_success_
@@ -153,6 +156,23 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
ctxt = desc_a%get_context()
call psb_info(ctxt, me, np)
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("DEC_SMTH_BLD: par_spspmm")
if ((do_timings).and.(idx_phase1==-1)) &
& idx_phase1 = psb_get_timer_idx("DEC_SMTH_BLD: phase1 ")
if ((do_timings).and.(idx_phase2==-1)) &
& idx_phase2 = psb_get_timer_idx("DEC_SMTH_BLD: phase2 ")
if ((do_timings).and.(idx_phase3==-1)) &
& idx_phase3 = psb_get_timer_idx("DEC_SMTH_BLD: phase3 ")
if ((do_timings).and.(idx_gtrans==-1)) &
& idx_gtrans = psb_get_timer_idx("DEC_SMTH_BLD: gtrans ")
if ((do_timings).and.(idx_refine==-1)) &
& idx_refine = psb_get_timer_idx("DEC_SMTH_BLD: refine ")
if ((do_timings).and.(idx_cdasb==-1)) &
& idx_cdasb = psb_get_timer_idx("DEC_SMTH_BLD: cdasb ")
if ((do_timings).and.(idx_ptap==-1)) &
& idx_ptap = psb_get_timer_idx("DEC_SMTH_BLD: ptap_bld ")
nglob = desc_a%get_global_rows()
nrow = desc_a%get_local_rows()
@@ -171,6 +191,7 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
! naggr: number of local aggregates
! nrow: local rows.
!
if (do_timings) call psb_tic(idx_phase1)
! Get the diagonal D
adiag = a%get_diag(info)
@@ -196,7 +217,7 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
!
! Build the filtered matrix Af from A
!
!$OMP parallel do private(i,j,tmp,jd) schedule(static)
do i=1, nrow
tmp = dzero
jd = -1
@@ -214,11 +235,13 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
acsrf%val(jd)=acsrf%val(jd)-tmp
end if
enddo
!$OMP end parallel do
! Take out zeroed terms
call acsrf%clean_zeros(info)
end if
!$OMP parallel do private(i) schedule(static)
do i=1,size(adiag)
if (adiag(i) /= dzero) then
adiag(i) = done / adiag(i)
@@ -226,7 +249,7 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
adiag(i) = done
end if
end do
!$OMP end parallel do
if (parms%aggr_omega_alg == amg_eig_est_) then
if (parms%aggr_eig == amg_max_norm_) then
@@ -252,8 +275,9 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(info,name,a_err='invalid amg_aggr_omega_alg_')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase1)
if (do_timings) call psb_tic(idx_phase2)
call acsrf%scal(adiag,info)
if (info /= psb_success_) goto 9999
@@ -267,6 +291,8 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_cdasb(desc_ac,info)
call psb_cd_reinit(desc_ac,info)
if (do_timings) call psb_toc(idx_phase2)
if (do_timings) call psb_tic(idx_phase3)
!
! Build the smoothed prolongator using either A or Af
! acsr1 = (I-w*D*A) Prol acsr1 = (I-w*D*Af) Prol
@@ -279,8 +305,8 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(psb_err_from_subroutine_,name,a_err='spspmm 1')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase3)
if (do_timings) call psb_tic(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done SPSPMM 1'
@@ -292,7 +318,7 @@ subroutine amg_daggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call op_prol%mv_from(coo_prol)
call op_restr%mv_from(coo_restr)
if (do_timings) call psb_toc(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done smooth_aggregate '
@@ -97,6 +97,8 @@ subroutine amg_s_dec_aggregator_build_tprol(ag,parms,ag_data,&
integer(psb_lpk_) :: ntaggr
integer(psb_ipk_) :: debug_level, debug_unit
logical :: clean_zeros
integer(psb_ipk_), save :: idx_map_bld=-1, idx_map_tprol=-1
logical, parameter :: do_timings=.false.
name='amg_s_dec_aggregator_tprol'
call psb_erractionsave(err_act)
@@ -108,6 +110,10 @@ subroutine amg_s_dec_aggregator_build_tprol(ag,parms,ag_data,&
info = psb_success_
ctxt = desc_a%get_context()
call psb_info(ctxt,me,np)
if ((do_timings).and.(idx_map_bld==-1)) &
& idx_map_bld = psb_get_timer_idx("DEC_TPROL: map_bld")
if ((do_timings).and.(idx_map_tprol==-1)) &
& idx_map_tprol = psb_get_timer_idx("DEC_TPROL: map_tprol")
call amg_check_def(parms%ml_cycle,'Multilevel cycle',&
& amg_mult_ml_,is_legal_ml_cycle)
@@ -121,10 +127,14 @@ subroutine amg_s_dec_aggregator_build_tprol(ag,parms,ag_data,&
! The decoupled aggregator based on SOC measures ignores
! ag_data except for clean_zeros; soc_map_bld is a procedure pointer.
!
if (do_timings) call psb_tic(idx_map_bld)
clean_zeros = ag%do_clean_zeros
call ag%soc_map_bld(parms%aggr_ord,parms%aggr_thresh,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
if (do_timings) call psb_toc(idx_map_bld)
if (do_timings) call psb_tic(idx_map_tprol)
if (info==psb_success_) call amg_map_to_tprol(desc_a,ilaggr,nlaggr,t_prol,info)
if (do_timings) call psb_toc(idx_map_tprol)
if (info /= psb_success_) then
info=psb_err_from_subroutine_
call psb_errpush(info,name,a_err='soc_map_bld/map_to_tprol')
+20 -14
View File
@@ -76,7 +76,7 @@ subroutine amg_s_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
integer(psb_ipk_) :: nrow, ncol, nrl, nzl, ip, nzt, i, k
integer(psb_lpk_) :: nrsave, ncsave, nzsave, nza
logical, parameter :: do_timings=.false., oldstyle=.false., debug=.false.
integer(psb_ipk_), save :: idx_spspmm=-1
integer(psb_ipk_), save :: idx_spspmm=-1, idx_cpytrans1=-1, idx_cpytrans2=-1
name='amg_ptap_bld'
if(psb_get_errstatus().ne.0) return
@@ -93,7 +93,11 @@ subroutine amg_s_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
ncol = desc_a%get_local_cols()
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("SPMM_BLD: par_spspmm")
& idx_spspmm = psb_get_timer_idx("PTAP_BLD: par_spspmm")
if ((do_timings).and.(idx_cpytrans1==-1)) &
& idx_cpytrans1 = psb_get_timer_idx("PTAP_BLD: cpy&trans1")
if ((do_timings).and.(idx_cpytrans2==-1)) &
& idx_cpytrans2 = psb_get_timer_idx("PTAP_BLD: cpy&trans2")
naggr = nlaggr(me+1)
ntaggr = sum(nlaggr)
@@ -128,6 +132,7 @@ subroutine amg_s_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
! Ok first product done.
if (present(desc_ax)) then
if (do_timings) call psb_tic(idx_cpytrans1)
block
call coo_prol%cp_to_coo(coo_restr,info)
call coo_restr%set_ncols(desc_ac%get_local_cols())
@@ -137,7 +142,7 @@ subroutine amg_s_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%set_ncols(desc_ax%get_local_cols())
end block
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans1)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -167,27 +172,28 @@ subroutine amg_s_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%transp()
nzl = coo_restr%get_nzeros()
nrl = desc_ac%get_local_rows()
i=0
nrl = desc_ac%get_local_rows()
call coo_restr%fix(info)
i=coo_restr%get_nzeros()
!
! Only keep local rows
!
do k=1, nzl
if ((1 <= coo_restr%ia(k)) .and.(coo_restr%ia(k) <= nrl)) then
i = i+1
coo_restr%val(i) = coo_restr%val(k)
coo_restr%ia(i) = coo_restr%ia(k)
coo_restr%ja(i) = coo_restr%ja(k)
search: do k=i,1,-1
if (coo_restr%ia(k) <= nrl) then
call coo_restr%set_nzeros(k)
exit search
end if
end do
call coo_restr%set_nzeros(i)
call coo_restr%fix(info)
end do search
nzl = coo_restr%get_nzeros()
call coo_restr%set_nrows(desc_ac%get_local_rows())
call coo_restr%set_ncols(desc_a%get_local_cols())
if (debug) call check_coo(me,trim(name)//' Check 2 on coo_restr:',coo_restr)
if (do_timings) call psb_tic(idx_cpytrans2)
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans2)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -72,7 +72,9 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
use psb_base_mod
use amg_base_prec_type
use amg_s_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
! Arguments
@@ -85,7 +87,7 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_), intent(out) :: info
! Local variables
integer(psb_ipk_), allocatable :: ils(:), neigh(:), irow(:), icol(:),&
integer(psb_ipk_), allocatable :: neigh(:), irow(:), icol(:),&
& ideg(:), idxs(:)
integer(psb_lpk_), allocatable :: tmpaggr(:)
real(psb_spk_), allocatable :: val(:), diag(:)
@@ -99,6 +101,9 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: nrow, ncol, n_ne
integer(psb_lpk_) :: nrglob
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc1_p1=-1, idx_soc1_p2=-1, idx_soc1_p3=-1
integer(psb_ipk_), save :: idx_soc1_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc1_map_bld'
@@ -114,6 +119,14 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc1_p0==-1)) &
& idx_soc1_p0 = psb_get_timer_idx("SOC1_MAP: phase0")
if ((do_timings).and.(idx_soc1_p1==-1)) &
& idx_soc1_p1 = psb_get_timer_idx("SOC1_MAP: phase1")
if ((do_timings).and.(idx_soc1_p2==-1)) &
& idx_soc1_p2 = psb_get_timer_idx("SOC1_MAP: phase2")
if ((do_timings).and.(idx_soc1_p3==-1)) &
& idx_soc1_p3 = psb_get_timer_idx("SOC1_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -133,41 +146,205 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc1_p0)
call a%cp_to(acsr)
if (do_timings) call psb_toc(idx_soc1_p0)
if (clean_zeros) call acsr%clean_zeros(info)
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
idxs(i) = i
end do
else
!$omp end parallel do
else
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = acsr%irp(i+1) - acsr%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_tic(idx_soc1_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
icnt = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz, nc, i,j,m, nz, ilg, ip, rsz
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) cycle step1
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
val(1:nz) = acsr%val(acsr%irp(i):acsr%irp(i+1)-1)
!
! Build the set of all strongly coupled nodes
!
ip = 0
do k=1, nz
j = icol(k)
! If any of the neighbours is already assigned,
! we will not reset.
if (j>nr) cycle step1
if (ilaggr(j) > 0) cycle step1
if (abs(val(k)) > theta*sqrt(abs(diag(i)*diag(j)))) then
ip = ip + 1
icol(ip) = icol(k)
end if
enddo
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
!$omp end atomic
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, ip
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
step1: do ii=1, nr
if (info /= 0) cycle
i = idxs(ii)
if ((i<1).or.(i>nr)) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
@@ -176,7 +353,7 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
! Build the set of all strongly coupled nodes
!
ip = 0
ip = 0
do k=1, nz
j = icol(k)
if ((1<=j).and.(j<=nr)) then
@@ -194,8 +371,7 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! contains I even if it does not look like it from matrix)
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
icnt = icnt + 1
if (disjoint) then
naggr = naggr + 1
do k=1, ip
ilaggr(icol(k)) = naggr
@@ -204,16 +380,22 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
& ' Check 1:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_toc(idx_soc1_p1)
if (do_timings) call psb_tic(idx_soc1_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,theta)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -244,8 +426,15 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc1_p2)
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1.5:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_tic(idx_soc1_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -274,7 +463,6 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
enddo
if (ip > 0) then
icnt = icnt + 1
naggr = naggr + 1
ilaggr(i) = naggr
do k=1, ip
@@ -292,7 +480,10 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,info)&
!$omp private(ii,i,j,k,nz,icol,val,ip)
do i=1, nr
if (info /= 0) cycle
if (ilaggr(i) < 0) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if (nz == 1) then
@@ -303,15 +494,18 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc1_p3)
if (naggr > ncol) then
!write(0,*) name,'Error : naggr > ncol',naggr,ncol
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
goto 9999
@@ -336,9 +530,13 @@ subroutine amg_s_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nlaggr(:) = 0
nlaggr(me+1) = naggr
call psb_sum(ctxt,nlaggr(1:np))
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 2:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
call acsr%free()
call psb_erractionrestore(err_act)
return
@@ -68,9 +68,12 @@
!
subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
use psb_base_mod
use psb_base_mod
use amg_base_prec_type
use amg_s_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
@@ -99,6 +102,9 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: np, me
integer(psb_ipk_) :: nrow, ncol, n_ne
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc2_p1=-1, idx_soc2_p2=-1, idx_soc2_p3=-1
integer(psb_ipk_), save :: idx_soc2_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc2_map_bld'
@@ -114,6 +120,14 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc2_p0==-1)) &
& idx_soc2_p0 = psb_get_timer_idx("SOC2_MAP: phase0")
if ((do_timings).and.(idx_soc2_p1==-1)) &
& idx_soc2_p1 = psb_get_timer_idx("SOC2_MAP: phase1")
if ((do_timings).and.(idx_soc2_p2==-1)) &
& idx_soc2_p2 = psb_get_timer_idx("SOC2_MAP: phase2")
if ((do_timings).and.(idx_soc2_p3==-1)) &
& idx_soc2_p3 = psb_get_timer_idx("SOC2_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -125,6 +139,7 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc2_p0)
diag = a%get_diag(info)
if(info /= psb_success_) then
info=psb_err_from_subroutine_
@@ -137,55 +152,218 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
call a%cp_to(muij)
if (clean_zeros) call muij%clean_zeros(info)
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
if (j<= nr) muij%val(k) = abs(muij%val(k))/sqrt(abs(diag(i)*diag(j)))
end do
end do
!$omp end parallel do
!
! Compute the 1-neigbour; mark strong links with +1, weak links with -1
!
call s_neigh_coo%allocate(nr,nr,muij%get_nzeros())
ip = 0
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
s_neigh_coo%ia(k) = i
s_neigh_coo%ja(k) = j
if (j<=nr) then
ip = ip + 1
s_neigh_coo%ia(ip) = i
s_neigh_coo%ja(ip) = j
if (real(muij%val(k)) >= theta) then
s_neigh_coo%val(ip) = sone
s_neigh_coo%val(k) = sone
else
s_neigh_coo%val(ip) = -sone
s_neigh_coo%val(k) = -sone
end if
else
s_neigh_coo%val(k) = -sone
end if
end do
end do
!$omp end parallel do
!write(*,*) 'S_NEIGH: ',nr,ip
call s_neigh_coo%set_nzeros(ip)
call s_neigh_coo%set_nzeros(muij%get_nzeros())
call s_neigh%mv_from_coo(s_neigh_coo,info)
if (iorder == amg_aggr_ord_nat_) then
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) shared(ilaggr,idxs) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
end do
!$omp end parallel do
else
!$omp parallel do private(i) shared(ilaggr,idxs,muij) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = muij%irp(i+1) - muij%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_toc(idx_soc2_p0)
if (do_timings) call psb_tic(idx_soc2_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(s_neigh,bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz,nc,i,j,m,nz,ilg,ip,rsz,ip1,nzcnt
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) then
write(0,*) ' Step1:',kk,ii,i,info
cycle step1
end if
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
!
! Get the 1-neighbourhood of I
!
ip1 = s_neigh%irp(i)
nz = s_neigh%irp(i+1)-ip1
!
! If the neighbourhood only contains I, skip it
!
if (nz ==0) then
ilaggr(i) = 0
cycle step1
end if
if ((nz==1).and.(s_neigh%ja(ip1)==i)) then
ilaggr(i) = 0
cycle step1
end if
nzcnt = count(real(s_neigh%val(ip1:ip1+nz-1)) > 0)
icol(1:nzcnt) = pack(s_neigh%ja(ip1:ip1+nz-1),(real(s_neigh%val(ip1:ip1+nz-1)) > 0))
disjoint = all(ilaggr(icol(1:nzcnt)) == -(nr+1))
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
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
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, nzcnt
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!write(0,*) 'LNAG ',locnaggr(nths+1)
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
icnt = 0
step1: do ii=1, nr
i = idxs(ii)
@@ -224,16 +402,21 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
end if
if (do_timings) call psb_toc(idx_soc2_p1)
if (do_timings) call psb_tic(idx_soc2_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,muij,s_neigh)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -259,8 +442,9 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc2_p2)
if (do_timings) call psb_tic(idx_soc2_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -294,6 +478,8 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,s_neigh,info)&
!$omp private(ii,i,j,k)
do i=1, nr
if (ilaggr(i) <= 0) then
nz = (s_neigh%irp(i+1)-s_neigh%irp(i))
@@ -305,13 +491,17 @@ subroutine amg_s_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc2_p3)
if (naggr > ncol) then
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
@@ -140,6 +140,9 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
real(psb_spk_) :: anorm, omega, tmp, dg, theta
logical, parameter :: debug_new=.false.
character(len=80) :: filename
logical, parameter :: do_timings=.false.
integer(psb_ipk_), save :: idx_spspmm=-1, idx_phase1=-1, idx_gtrans=-1, idx_phase2=-1, idx_refine=-1
integer(psb_ipk_), save :: idx_phase3=-1, idx_cdasb=-1, idx_ptap=-1
name='amg_aggrmat_smth_bld'
info=psb_success_
@@ -153,6 +156,23 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
ctxt = desc_a%get_context()
call psb_info(ctxt, me, np)
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("DEC_SMTH_BLD: par_spspmm")
if ((do_timings).and.(idx_phase1==-1)) &
& idx_phase1 = psb_get_timer_idx("DEC_SMTH_BLD: phase1 ")
if ((do_timings).and.(idx_phase2==-1)) &
& idx_phase2 = psb_get_timer_idx("DEC_SMTH_BLD: phase2 ")
if ((do_timings).and.(idx_phase3==-1)) &
& idx_phase3 = psb_get_timer_idx("DEC_SMTH_BLD: phase3 ")
if ((do_timings).and.(idx_gtrans==-1)) &
& idx_gtrans = psb_get_timer_idx("DEC_SMTH_BLD: gtrans ")
if ((do_timings).and.(idx_refine==-1)) &
& idx_refine = psb_get_timer_idx("DEC_SMTH_BLD: refine ")
if ((do_timings).and.(idx_cdasb==-1)) &
& idx_cdasb = psb_get_timer_idx("DEC_SMTH_BLD: cdasb ")
if ((do_timings).and.(idx_ptap==-1)) &
& idx_ptap = psb_get_timer_idx("DEC_SMTH_BLD: ptap_bld ")
nglob = desc_a%get_global_rows()
nrow = desc_a%get_local_rows()
@@ -171,6 +191,7 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
! naggr: number of local aggregates
! nrow: local rows.
!
if (do_timings) call psb_tic(idx_phase1)
! Get the diagonal D
adiag = a%get_diag(info)
@@ -196,7 +217,7 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
!
! Build the filtered matrix Af from A
!
!$OMP parallel do private(i,j,tmp,jd) schedule(static)
do i=1, nrow
tmp = szero
jd = -1
@@ -214,11 +235,13 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
acsrf%val(jd)=acsrf%val(jd)-tmp
end if
enddo
!$OMP end parallel do
! Take out zeroed terms
call acsrf%clean_zeros(info)
end if
!$OMP parallel do private(i) schedule(static)
do i=1,size(adiag)
if (adiag(i) /= szero) then
adiag(i) = sone / adiag(i)
@@ -226,7 +249,7 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
adiag(i) = sone
end if
end do
!$OMP end parallel do
if (parms%aggr_omega_alg == amg_eig_est_) then
if (parms%aggr_eig == amg_max_norm_) then
@@ -252,8 +275,9 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(info,name,a_err='invalid amg_aggr_omega_alg_')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase1)
if (do_timings) call psb_tic(idx_phase2)
call acsrf%scal(adiag,info)
if (info /= psb_success_) goto 9999
@@ -267,6 +291,8 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_cdasb(desc_ac,info)
call psb_cd_reinit(desc_ac,info)
if (do_timings) call psb_toc(idx_phase2)
if (do_timings) call psb_tic(idx_phase3)
!
! Build the smoothed prolongator using either A or Af
! acsr1 = (I-w*D*A) Prol acsr1 = (I-w*D*Af) Prol
@@ -279,8 +305,8 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(psb_err_from_subroutine_,name,a_err='spspmm 1')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase3)
if (do_timings) call psb_tic(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done SPSPMM 1'
@@ -292,7 +318,7 @@ subroutine amg_saggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call op_prol%mv_from(coo_prol)
call op_restr%mv_from(coo_restr)
if (do_timings) call psb_toc(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done smooth_aggregate '
@@ -97,6 +97,8 @@ subroutine amg_z_dec_aggregator_build_tprol(ag,parms,ag_data,&
integer(psb_lpk_) :: ntaggr
integer(psb_ipk_) :: debug_level, debug_unit
logical :: clean_zeros
integer(psb_ipk_), save :: idx_map_bld=-1, idx_map_tprol=-1
logical, parameter :: do_timings=.false.
name='amg_z_dec_aggregator_tprol'
call psb_erractionsave(err_act)
@@ -108,6 +110,10 @@ subroutine amg_z_dec_aggregator_build_tprol(ag,parms,ag_data,&
info = psb_success_
ctxt = desc_a%get_context()
call psb_info(ctxt,me,np)
if ((do_timings).and.(idx_map_bld==-1)) &
& idx_map_bld = psb_get_timer_idx("DEC_TPROL: map_bld")
if ((do_timings).and.(idx_map_tprol==-1)) &
& idx_map_tprol = psb_get_timer_idx("DEC_TPROL: map_tprol")
call amg_check_def(parms%ml_cycle,'Multilevel cycle',&
& amg_mult_ml_,is_legal_ml_cycle)
@@ -121,10 +127,14 @@ subroutine amg_z_dec_aggregator_build_tprol(ag,parms,ag_data,&
! The decoupled aggregator based on SOC measures ignores
! ag_data except for clean_zeros; soc_map_bld is a procedure pointer.
!
if (do_timings) call psb_tic(idx_map_bld)
clean_zeros = ag%do_clean_zeros
call ag%soc_map_bld(parms%aggr_ord,parms%aggr_thresh,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
if (do_timings) call psb_toc(idx_map_bld)
if (do_timings) call psb_tic(idx_map_tprol)
if (info==psb_success_) call amg_map_to_tprol(desc_a,ilaggr,nlaggr,t_prol,info)
if (do_timings) call psb_toc(idx_map_tprol)
if (info /= psb_success_) then
info=psb_err_from_subroutine_
call psb_errpush(info,name,a_err='soc_map_bld/map_to_tprol')
+20 -14
View File
@@ -76,7 +76,7 @@ subroutine amg_z_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
integer(psb_ipk_) :: nrow, ncol, nrl, nzl, ip, nzt, i, k
integer(psb_lpk_) :: nrsave, ncsave, nzsave, nza
logical, parameter :: do_timings=.false., oldstyle=.false., debug=.false.
integer(psb_ipk_), save :: idx_spspmm=-1
integer(psb_ipk_), save :: idx_spspmm=-1, idx_cpytrans1=-1, idx_cpytrans2=-1
name='amg_ptap_bld'
if(psb_get_errstatus().ne.0) return
@@ -93,7 +93,11 @@ subroutine amg_z_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
ncol = desc_a%get_local_cols()
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("SPMM_BLD: par_spspmm")
& idx_spspmm = psb_get_timer_idx("PTAP_BLD: par_spspmm")
if ((do_timings).and.(idx_cpytrans1==-1)) &
& idx_cpytrans1 = psb_get_timer_idx("PTAP_BLD: cpy&trans1")
if ((do_timings).and.(idx_cpytrans2==-1)) &
& idx_cpytrans2 = psb_get_timer_idx("PTAP_BLD: cpy&trans2")
naggr = nlaggr(me+1)
ntaggr = sum(nlaggr)
@@ -128,6 +132,7 @@ subroutine amg_z_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
! Ok first product done.
if (present(desc_ax)) then
if (do_timings) call psb_tic(idx_cpytrans1)
block
call coo_prol%cp_to_coo(coo_restr,info)
call coo_restr%set_ncols(desc_ac%get_local_cols())
@@ -137,7 +142,7 @@ subroutine amg_z_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%set_ncols(desc_ax%get_local_cols())
end block
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans1)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -167,27 +172,28 @@ subroutine amg_z_ptap_bld(a_csr,desc_a,nlaggr,parms,ac,&
call coo_restr%transp()
nzl = coo_restr%get_nzeros()
nrl = desc_ac%get_local_rows()
i=0
nrl = desc_ac%get_local_rows()
call coo_restr%fix(info)
i=coo_restr%get_nzeros()
!
! Only keep local rows
!
do k=1, nzl
if ((1 <= coo_restr%ia(k)) .and.(coo_restr%ia(k) <= nrl)) then
i = i+1
coo_restr%val(i) = coo_restr%val(k)
coo_restr%ia(i) = coo_restr%ia(k)
coo_restr%ja(i) = coo_restr%ja(k)
search: do k=i,1,-1
if (coo_restr%ia(k) <= nrl) then
call coo_restr%set_nzeros(k)
exit search
end if
end do
call coo_restr%set_nzeros(i)
call coo_restr%fix(info)
end do search
nzl = coo_restr%get_nzeros()
call coo_restr%set_nrows(desc_ac%get_local_rows())
call coo_restr%set_ncols(desc_a%get_local_cols())
if (debug) call check_coo(me,trim(name)//' Check 2 on coo_restr:',coo_restr)
if (do_timings) call psb_tic(idx_cpytrans2)
call csr_restr%cp_from_coo(coo_restr,info)
if (do_timings) call psb_toc(idx_cpytrans2)
if (info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='spcnv coo_restr')
goto 9999
@@ -72,7 +72,9 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
use psb_base_mod
use amg_base_prec_type
use amg_z_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
! Arguments
@@ -85,7 +87,7 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_), intent(out) :: info
! Local variables
integer(psb_ipk_), allocatable :: ils(:), neigh(:), irow(:), icol(:),&
integer(psb_ipk_), allocatable :: neigh(:), irow(:), icol(:),&
& ideg(:), idxs(:)
integer(psb_lpk_), allocatable :: tmpaggr(:)
complex(psb_dpk_), allocatable :: val(:), diag(:)
@@ -99,6 +101,9 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: nrow, ncol, n_ne
integer(psb_lpk_) :: nrglob
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc1_p1=-1, idx_soc1_p2=-1, idx_soc1_p3=-1
integer(psb_ipk_), save :: idx_soc1_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc1_map_bld'
@@ -114,6 +119,14 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc1_p0==-1)) &
& idx_soc1_p0 = psb_get_timer_idx("SOC1_MAP: phase0")
if ((do_timings).and.(idx_soc1_p1==-1)) &
& idx_soc1_p1 = psb_get_timer_idx("SOC1_MAP: phase1")
if ((do_timings).and.(idx_soc1_p2==-1)) &
& idx_soc1_p2 = psb_get_timer_idx("SOC1_MAP: phase2")
if ((do_timings).and.(idx_soc1_p3==-1)) &
& idx_soc1_p3 = psb_get_timer_idx("SOC1_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -133,41 +146,205 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc1_p0)
call a%cp_to(acsr)
if (do_timings) call psb_toc(idx_soc1_p0)
if (clean_zeros) call acsr%clean_zeros(info)
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
idxs(i) = i
end do
else
!$omp end parallel do
else
!$omp parallel do private(i) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = acsr%irp(i+1) - acsr%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_tic(idx_soc1_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
icnt = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz, nc, i,j,m, nz, ilg, ip, rsz
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) cycle step1
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
val(1:nz) = acsr%val(acsr%irp(i):acsr%irp(i+1)-1)
!
! Build the set of all strongly coupled nodes
!
ip = 0
do k=1, nz
j = icol(k)
! If any of the neighbours is already assigned,
! we will not reset.
if (j>nr) cycle step1
if (ilaggr(j) > 0) cycle step1
if (abs(val(k)) > theta*sqrt(abs(diag(i)*diag(j)))) then
ip = ip + 1
icol(ip) = icol(k)
end if
enddo
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
itmp = (bnds(kk)-1+locnaggr(kk)) !be careful about overflow
itmp = itmp*nths+kk
if (itmp < (bnds(kk)-1+locnaggr(kk))) then
!$omp atomic update
info = max(12345678,info)
!$omp end atomic
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, ip
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
step1: do ii=1, nr
if (info /= 0) cycle
i = idxs(ii)
if ((i<1).or.(i>nr)) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if ((nz<0).or.(nz>size(icol))) then
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
cycle step1
!goto 9999
end if
icol(1:nz) = acsr%ja(acsr%irp(i):acsr%irp(i+1)-1)
@@ -176,7 +353,7 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
! Build the set of all strongly coupled nodes
!
ip = 0
ip = 0
do k=1, nz
j = icol(k)
if ((1<=j).and.(j<=nr)) then
@@ -194,8 +371,7 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! contains I even if it does not look like it from matrix)
!
disjoint = all(ilaggr(icol(1:ip)) == -(nr+1)).or.(ip==0)
if (disjoint) then
icnt = icnt + 1
if (disjoint) then
naggr = naggr + 1
do k=1, ip
ilaggr(icol(k)) = naggr
@@ -204,16 +380,22 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
& ' Check 1:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_toc(idx_soc1_p1)
if (do_timings) call psb_tic(idx_soc1_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,theta)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -244,8 +426,15 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc1_p2)
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1.5:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
if (do_timings) call psb_tic(idx_soc1_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -274,7 +463,6 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
enddo
if (ip > 0) then
icnt = icnt + 1
naggr = naggr + 1
ilaggr(i) = naggr
do k=1, ip
@@ -292,7 +480,10 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,info)&
!$omp private(ii,i,j,k,nz,icol,val,ip)
do i=1, nr
if (info /= 0) cycle
if (ilaggr(i) < 0) then
nz = (acsr%irp(i+1)-acsr%irp(i))
if (nz == 1) then
@@ -303,15 +494,18 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc1_p3)
if (naggr > ncol) then
!write(0,*) name,'Error : naggr > ncol',naggr,ncol
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
goto 9999
@@ -336,9 +530,13 @@ subroutine amg_z_soc1_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nlaggr(:) = 0
nlaggr(me+1) = naggr
call psb_sum(ctxt,nlaggr(1:np))
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 2:',naggr,count(ilaggr(1:nr) == -(nr+1)), count(ilaggr(1:nr)>0),&
& count(ilaggr(1:nr) == -(nr+1))+count(ilaggr(1:nr)>0),nr
end if
call acsr%free()
call psb_erractionrestore(err_act)
return
@@ -68,9 +68,12 @@
!
subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,info)
use psb_base_mod
use psb_base_mod
use amg_base_prec_type
use amg_z_inner_mod
#if defined(OPENMP)
use omp_lib
#endif
implicit none
@@ -99,6 +102,9 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
integer(psb_ipk_) :: np, me
integer(psb_ipk_) :: nrow, ncol, n_ne
character(len=20) :: name, ch_err
integer(psb_ipk_), save :: idx_soc2_p1=-1, idx_soc2_p2=-1, idx_soc2_p3=-1
integer(psb_ipk_), save :: idx_soc2_p0=-1
logical, parameter :: do_timings=.true.
info=psb_success_
name = 'amg_soc2_map_bld'
@@ -114,6 +120,14 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
nrow = desc_a%get_local_rows()
ncol = desc_a%get_local_cols()
nrglob = desc_a%get_global_rows()
if ((do_timings).and.(idx_soc2_p0==-1)) &
& idx_soc2_p0 = psb_get_timer_idx("SOC2_MAP: phase0")
if ((do_timings).and.(idx_soc2_p1==-1)) &
& idx_soc2_p1 = psb_get_timer_idx("SOC2_MAP: phase1")
if ((do_timings).and.(idx_soc2_p2==-1)) &
& idx_soc2_p2 = psb_get_timer_idx("SOC2_MAP: phase2")
if ((do_timings).and.(idx_soc2_p3==-1)) &
& idx_soc2_p3 = psb_get_timer_idx("SOC2_MAP: phase3")
nr = a%get_nrows()
nc = a%get_ncols()
@@ -125,6 +139,7 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
goto 9999
end if
if (do_timings) call psb_tic(idx_soc2_p0)
diag = a%get_diag(info)
if(info /= psb_success_) then
info=psb_err_from_subroutine_
@@ -137,55 +152,218 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
!
call a%cp_to(muij)
if (clean_zeros) call muij%clean_zeros(info)
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
if (j<= nr) muij%val(k) = abs(muij%val(k))/sqrt(abs(diag(i)*diag(j)))
end do
end do
!$omp end parallel do
!
! Compute the 1-neigbour; mark strong links with +1, weak links with -1
!
call s_neigh_coo%allocate(nr,nr,muij%get_nzeros())
ip = 0
!$omp parallel do private(i,j,k) shared(nr,diag,muij) schedule(static)
do i=1, nr
do k=muij%irp(i),muij%irp(i+1)-1
j = muij%ja(k)
s_neigh_coo%ia(k) = i
s_neigh_coo%ja(k) = j
if (j<=nr) then
ip = ip + 1
s_neigh_coo%ia(ip) = i
s_neigh_coo%ja(ip) = j
if (real(muij%val(k)) >= theta) then
s_neigh_coo%val(ip) = done
s_neigh_coo%val(k) = done
else
s_neigh_coo%val(ip) = -done
s_neigh_coo%val(k) = -done
end if
else
s_neigh_coo%val(k) = -done
end if
end do
end do
!$omp end parallel do
!write(*,*) 'S_NEIGH: ',nr,ip
call s_neigh_coo%set_nzeros(ip)
call s_neigh_coo%set_nzeros(muij%get_nzeros())
call s_neigh%mv_from_coo(s_neigh_coo,info)
if (iorder == amg_aggr_ord_nat_) then
if (iorder == amg_aggr_ord_nat_) then
!$omp parallel do private(i) shared(ilaggr,idxs) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
idxs(i) = i
end do
!$omp end parallel do
else
!$omp parallel do private(i) shared(ilaggr,idxs,muij) schedule(static)
do i=1, nr
ilaggr(i) = -(nr+1)
ideg(i) = muij%irp(i+1) - muij%irp(i)
end do
!$omp end parallel do
call psb_msort(ideg,ix=idxs,dir=psb_sort_down_)
end if
if (do_timings) call psb_toc(idx_soc2_p0)
if (do_timings) call psb_tic(idx_soc2_p1)
!
! Phase one: Start with disjoint groups.
!
naggr = 0
#if defined(OPENMP)
block
integer(psb_ipk_), allocatable :: bnds(:), locnaggr(:)
integer(psb_ipk_) :: myth,nths, kk
! The parallelization makes use of a locaggr(:) array; each thread
! keeps its own version of naggr, and when the loop ends, a prefix is applied
! to locnaggr to determine:
! 1. The total number of aggregaters NAGGR;
! 2. How much should each thread shift its own aggregates
! Part 2 requires to keep track of which thread defined each entry
! of ilaggr(), so that each entry can be adjusted correctly: even
! if an entry I belongs to the range BNDS(TH)>BNDS(TH+1)-1, it may have
! been set because it is strongly connected to an entry J belonging to a
! different thread.
!$omp parallel shared(s_neigh,bnds,idxs,locnaggr,ilaggr,nr,naggr,diag,theta,nths,info) &
!$omp private(icol,val,myth,kk)
block
integer(psb_ipk_) :: ii,nlp,k,kp,n,ia,isz,nc,i,j,m,nz,ilg,ip,rsz,ip1,nzcnt
integer(psb_lpk_) :: itmp
!$omp master
nths = omp_get_num_threads()
allocate(bnds(0:nths),locnaggr(0:nths+1))
locnaggr(:) = 0
bnds(0) = 1
!$omp end master
!$omp barrier
myth = omp_get_thread_num()
rsz = nr/nths
if (myth < mod(nr,nths)) rsz = rsz + 1
bnds(myth+1) = rsz
!$omp barrier
!$omp master
do i=1,nths
bnds(i) = bnds(i) + bnds(i-1)
end do
info = 0
!$omp end master
!$omp barrier
!$omp do schedule(static) private(disjoint)
do kk=0, nths-1
step1: do ii=bnds(kk), bnds(kk+1)-1
i = idxs(ii)
if (info /= 0) then
write(0,*) ' Step1:',kk,ii,i,info
cycle step1
end if
if ((i<1).or.(i>nr)) then
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name)
cycle step1
!goto 9999
end if
if (ilaggr(i) == -(nr+1)) then
!
! Get the 1-neighbourhood of I
!
ip1 = s_neigh%irp(i)
nz = s_neigh%irp(i+1)-ip1
!
! If the neighbourhood only contains I, skip it
!
if (nz ==0) then
ilaggr(i) = 0
cycle step1
end if
if ((nz==1).and.(s_neigh%ja(ip1)==i)) then
ilaggr(i) = 0
cycle step1
end if
nzcnt = count(real(s_neigh%val(ip1:ip1+nz-1)) > 0)
icol(1:nzcnt) = pack(s_neigh%ja(ip1:ip1+nz-1),(real(s_neigh%val(ip1:ip1+nz-1)) > 0))
disjoint = all(ilaggr(icol(1:nzcnt)) == -(nr+1))
!
! If the whole strongly coupled neighborhood of I is
! as yet unconnected, turn it into the next aggregate.
! Same if ip==0 (in which case, neighborhood only
! contains I even if it does not look like it from matrix)
! The fact that DISJOINT is private and not under lock
! generates a certain un-repeatability, in that between
! computing DISJOINT and assigning, another thread might
! alter the values of ILAGGR.
! However, a certain unrepeatability is already present
! because the sequence of aggregates is computed with a
! different order than in serial mode.
! In any case, even if the enteries of ILAGGR may be
! overwritten, the important thing is that each entry is
! consistent and they generate a correct aggregation map.
!
if (disjoint) then
locnaggr(kk) = locnaggr(kk) + 1
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
cycle step1
end if
!$omp atomic write
ilaggr(i) = itmp
!$omp end atomic
do k=1, nzcnt
!$omp atomic write
ilaggr(icol(k)) = itmp
!$omp end atomic
end do
end if
end if
enddo step1
end do
!$omp end do
!$omp master
naggr = sum(locnaggr(0:nths-1))
do i=1,nths
locnaggr(i) = locnaggr(i) + locnaggr(i-1)
end do
do i=nths+1,1,-1
locnaggr(i) = locnaggr(i-1)
end do
locnaggr(0) = 0
!write(0,*) 'LNAG ',locnaggr(nths+1)
!$omp end master
!$omp barrier
!$omp do schedule(static)
do kk=0, nths-1
do ii=bnds(kk), bnds(kk+1)-1
if (ilaggr(ii) > 0) then
kp = mod(ilaggr(ii),nths)
ilaggr(ii) = (ilaggr(ii)/nths)- (bnds(kp)-1) + locnaggr(kp)
end if
end do
end do
!$omp end do
end block
!$omp end parallel
end block
if (info /= 0) then
if (info == 12345678) write(0,*) 'Overflow in encoding ILAGGR'
info=psb_err_internal_error_
call psb_errpush(info,name)
goto 9999
end if
#else
icnt = 0
step1: do ii=1, nr
i = idxs(ii)
@@ -224,16 +402,21 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
endif
enddo step1
#endif
if (debug_level >= psb_debug_outer_) then
write(debug_unit,*) me,' ',trim(name),&
& ' Check 1:',count(ilaggr == -(nr+1))
end if
if (do_timings) call psb_toc(idx_soc2_p1)
if (do_timings) call psb_tic(idx_soc2_p2)
!
! Phase two: join the neighbours
!
!$omp workshare
tmpaggr = ilaggr
!$omp end workshare
!$omp parallel do schedule(static) shared(tmpaggr,ilaggr,nr,naggr,diag,muij,s_neigh)&
!$omp private(ii,i,j,k,nz,icol,val,ip,cpling)
step2: do ii=1,nr
i = idxs(ii)
@@ -259,8 +442,9 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end if
end if
end do step2
!$omp end parallel do
if (do_timings) call psb_toc(idx_soc2_p2)
if (do_timings) call psb_tic(idx_soc2_p3)
!
! Phase three: sweep over leftovers, if any
!
@@ -294,6 +478,8 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
end do step3
! Any leftovers?
!$omp parallel do schedule(static) shared(ilaggr,s_neigh,info)&
!$omp private(ii,i,j,k)
do i=1, nr
if (ilaggr(i) <= 0) then
nz = (s_neigh%irp(i+1)-s_neigh%irp(i))
@@ -305,13 +491,17 @@ subroutine amg_z_soc2_map_bld(iorder,theta,clean_zeros,a,desc_a,nlaggr,ilaggr,in
! other processes.
ilaggr(i) = -(nrglob+nr)
else
!$omp atomic write
info=psb_err_internal_error_
!$omp end atomic
call psb_errpush(info,name,a_err='Fatal error: non-singleton leftovers')
goto 9999
cycle
endif
end if
end do
!$omp end parallel do
if (info /= 0) goto 9999
if (do_timings) call psb_toc(idx_soc2_p3)
if (naggr > ncol) then
info=psb_err_internal_error_
call psb_errpush(info,name,a_err='Fatal error: naggr>ncol')
@@ -140,6 +140,9 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
real(psb_dpk_) :: anorm, omega, tmp, dg, theta
logical, parameter :: debug_new=.false.
character(len=80) :: filename
logical, parameter :: do_timings=.false.
integer(psb_ipk_), save :: idx_spspmm=-1, idx_phase1=-1, idx_gtrans=-1, idx_phase2=-1, idx_refine=-1
integer(psb_ipk_), save :: idx_phase3=-1, idx_cdasb=-1, idx_ptap=-1
name='amg_aggrmat_smth_bld'
info=psb_success_
@@ -153,6 +156,23 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
ctxt = desc_a%get_context()
call psb_info(ctxt, me, np)
if ((do_timings).and.(idx_spspmm==-1)) &
& idx_spspmm = psb_get_timer_idx("DEC_SMTH_BLD: par_spspmm")
if ((do_timings).and.(idx_phase1==-1)) &
& idx_phase1 = psb_get_timer_idx("DEC_SMTH_BLD: phase1 ")
if ((do_timings).and.(idx_phase2==-1)) &
& idx_phase2 = psb_get_timer_idx("DEC_SMTH_BLD: phase2 ")
if ((do_timings).and.(idx_phase3==-1)) &
& idx_phase3 = psb_get_timer_idx("DEC_SMTH_BLD: phase3 ")
if ((do_timings).and.(idx_gtrans==-1)) &
& idx_gtrans = psb_get_timer_idx("DEC_SMTH_BLD: gtrans ")
if ((do_timings).and.(idx_refine==-1)) &
& idx_refine = psb_get_timer_idx("DEC_SMTH_BLD: refine ")
if ((do_timings).and.(idx_cdasb==-1)) &
& idx_cdasb = psb_get_timer_idx("DEC_SMTH_BLD: cdasb ")
if ((do_timings).and.(idx_ptap==-1)) &
& idx_ptap = psb_get_timer_idx("DEC_SMTH_BLD: ptap_bld ")
nglob = desc_a%get_global_rows()
nrow = desc_a%get_local_rows()
@@ -171,6 +191,7 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
! naggr: number of local aggregates
! nrow: local rows.
!
if (do_timings) call psb_tic(idx_phase1)
! Get the diagonal D
adiag = a%get_diag(info)
@@ -196,7 +217,7 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
!
! Build the filtered matrix Af from A
!
!$OMP parallel do private(i,j,tmp,jd) schedule(static)
do i=1, nrow
tmp = zzero
jd = -1
@@ -214,11 +235,13 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
acsrf%val(jd)=acsrf%val(jd)-tmp
end if
enddo
!$OMP end parallel do
! Take out zeroed terms
call acsrf%clean_zeros(info)
end if
!$OMP parallel do private(i) schedule(static)
do i=1,size(adiag)
if (adiag(i) /= zzero) then
adiag(i) = zone / adiag(i)
@@ -226,7 +249,7 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
adiag(i) = zone
end if
end do
!$OMP end parallel do
if (parms%aggr_omega_alg == amg_eig_est_) then
if (parms%aggr_eig == amg_max_norm_) then
@@ -252,8 +275,9 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(info,name,a_err='invalid amg_aggr_omega_alg_')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase1)
if (do_timings) call psb_tic(idx_phase2)
call acsrf%scal(adiag,info)
if (info /= psb_success_) goto 9999
@@ -267,6 +291,8 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_cdasb(desc_ac,info)
call psb_cd_reinit(desc_ac,info)
if (do_timings) call psb_toc(idx_phase2)
if (do_timings) call psb_tic(idx_phase3)
!
! Build the smoothed prolongator using either A or Af
! acsr1 = (I-w*D*A) Prol acsr1 = (I-w*D*Af) Prol
@@ -279,8 +305,8 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call psb_errpush(psb_err_from_subroutine_,name,a_err='spspmm 1')
goto 9999
end if
if (do_timings) call psb_toc(idx_phase3)
if (do_timings) call psb_tic(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done SPSPMM 1'
@@ -292,7 +318,7 @@ subroutine amg_zaggrmat_smth_bld(a,desc_a,ilaggr,nlaggr,parms,&
call op_prol%mv_from(coo_prol)
call op_restr%mv_from(coo_restr)
if (do_timings) call psb_toc(idx_ptap)
if (debug_level >= psb_debug_outer_) &
& write(debug_unit,*) me,' ',trim(name),&
& 'Done smooth_aggregate '
+90
View File
@@ -0,0 +1,90 @@
#include "MatchBoxPC.h"
// TODO comment
void clean(MilanLongInt NLVer,
MilanInt myRank,
MilanLongInt MessageIndex,
vector<MPI_Request> &SRequest,
vector<MPI_Status> &SStatus,
MilanInt BufferSize,
MilanLongInt *Buffer,
MilanLongInt msgActual,
MilanLongInt *msgActualSent,
MilanLongInt msgInd,
MilanLongInt *msgIndSent,
MilanLongInt NumMessagesBundled,
MilanReal *msgPercent)
{
// Cleanup Phase
#pragma omp parallel
{
#pragma omp master
{
#pragma omp task
{
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ") Waitall= " << endl;
fflush(stdout);
#endif
#ifdef DEBUG_HANG_
cout << "\n(" << myRank << ") Waitall " << endl;
fflush(stdout);
#endif
//return;
MPI_Waitall(MessageIndex, &SRequest[0], &SStatus[0]);
// MPI_Buffer_attach(&Buffer, BufferSize); //Attach the Buffer
if (BufferSize > 0)
{
MPI_Buffer_detach(&Buffer, &BufferSize); // Detach the Buffer
free(Buffer); // Free the memory that was allocated
}
}
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")End of function to compute matching: " << endl;
fflush(stdout);
cout << "\n(" << myRank << ")myCardinality: " << myCard << endl;
fflush(stdout);
cout << "\n(" << myRank << ")Matching took " << finishTime - startTime << "seconds" << endl;
fflush(stdout);
cout << "\n(" << myRank << ")** Getting out of the matching function **" << endl;
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ") Number of Ghost edges = " << numGhostEdges;
cout << "\n(" << myRank << ") Total number of potential message X 2 = " << numGhostEdges * 2;
cout << "\n(" << myRank << ") Number messages bundled = " << NumMessagesBundled;
cout << "\n(" << myRank << ") Total Individual Messages sent = " << msgInd;
if (msgInd > 0)
{
cout << "\n(" << myRank << ") Percentage of messages bundled = " << ((double)NumMessagesBundled / (double)(msgInd)) * 100.0 << "% \n";
}
fflush(stdout);
#endif
#pragma omp task
{
*msgActualSent = msgActual;
*msgIndSent = msgInd;
if (msgInd > 0)
{
*msgPercent = ((double)NumMessagesBundled / (double)(msgInd)) * 100.0;
}
else
{
*msgPercent = 0;
}
}
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ") Done" << endl;
fflush(stdout);
#endif
}
}
}
@@ -0,0 +1,136 @@
#include "MatchBoxPC.h"
/**
* Execute the research fr the Candidate Mate without controlling if the vertices are already matched.
* Returns the vertices with the highest weight
* @param adj1
* @param adj2
* @param verLocInd
* @param edgeLocWeight
* @return
*/
MilanLongInt firstComputeCandidateMateD(MilanLongInt adj1,
MilanLongInt adj2,
MilanLongInt *verLocInd,
MilanReal *edgeLocWeight)
{
MilanInt w = -1;
MilanReal heaviestEdgeWt = MilanRealMin; // Assign the smallest Value possible first LDBL_MIN
int finalK;
for (int k = adj1; k < adj2; k++) {
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k]))) {
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
finalK = k;
}
} // End of for loop
return finalK;
}
/**
* //TODO documentation
* @param adj1
* @param adj2
* @param edgeLocWeight
* @param k
* @param verLocInd
* @param StartIndex
* @param EndIndex
* @param GMate
* @param Mate
* @param Ghost2LocalMap
* @return
*/
MilanLongInt computeCandidateMateD(MilanLongInt adj1,
MilanLongInt adj2,
MilanReal *edgeLocWeight,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap)
{
// Start: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
MilanInt w = -1;
MilanReal heaviestEdgeWt = MilanRealMin; // Assign the smallest Value possible first LDBL_MIN
for (k = adj1; k < adj2; k++) {
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap))
continue;
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k]))) {
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
}
} // End of for loop
// End: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
return w;
}
MilanLongInt firstComputeCandidateMateS(MilanLongInt adj1,
MilanLongInt adj2,
MilanLongInt *verLocInd,
MilanFloat *edgeLocWeight)
{
MilanInt w = -1;
MilanFloat heaviestEdgeWt = 0.0f; // Assign the smallest
int finalK;
for (int k = adj1; k < adj2; k++) {
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k]))) {
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
finalK = k;
}
} // End of for loop
return finalK;
}
/**
* //TODO documentation
* @param adj1
* @param adj2
* @param edgeLocWeight
* @param k
* @param verLocInd
* @param StartIndex
* @param EndIndex
* @param GMate
* @param Mate
* @param Ghost2LocalMap
* @return
*/
MilanLongInt computeCandidateMateS(MilanLongInt adj1,
MilanLongInt adj2,
MilanFloat *edgeLocWeight,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap)
{
// Start: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
MilanInt w = -1;
MilanFloat heaviestEdgeWt = 0.0f; // Assign the smallest Value
for (k = adj1; k < adj2; k++) {
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap))
continue;
if ((edgeLocWeight[k] > heaviestEdgeWt) ||
((edgeLocWeight[k] == heaviestEdgeWt) && (w < verLocInd[k]))) {
heaviestEdgeWt = edgeLocWeight[k];
w = verLocInd[k];
}
} // End of for loop
// End: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
return w;
}
+4 -2
View File
@@ -80,9 +80,11 @@ class staticQueue
MilanLongInt squeueTail;
MilanLongInt NumNodes;
//FIXME I had to comment this piece of code in order to make everything work.
// why?
//Prevent Assignment and Pass by Value:
staticQueue(const staticQueue& src);
staticQueue& operator=(const staticQueue& rhs);
//staticQueue(const staticQueue& src);
//staticQueue& operator=(const staticQueue& rhs);
public:
//Constructors and Destructors
+30
View File
@@ -0,0 +1,30 @@
#include "MatchBoxPC.h"
void extractUChunk(
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU)
{
UChunkBeingProcessed.clear();
#pragma omp critical(U)
{
if (U.empty() && !privateU.empty()) // If U is empty but there are nodes in private U
{
while (!privateU.empty())
UChunkBeingProcessed.push_back(privateU.back());
privateU.pop_back();
}
else
{
for (int i = 0; i < UCHUNK; i++)
{ // Pop the new nodes
if (U.empty())
break;
UChunkBeingProcessed.push_back(U.back());
U.pop_back();
}
}
} // End of critical U // End of critical U
}
@@ -0,0 +1,28 @@
#include "MatchBoxPC.h"
/// Find the owner of a ghost node:
MilanInt findOwnerOfGhost(MilanLongInt vtxIndex, MilanLongInt *mVerDistance,
MilanInt myRank, MilanInt numProcs)
{
MilanLongInt mStartInd = mVerDistance[myRank];
MilanInt Start = 0;
MilanInt End = numProcs;
MilanInt Current = 0;
while (Start <= End)
{
Current = (End + Start) / 2;
// CASE-1:
if (mVerDistance[Current] == vtxIndex) return Current;
else // CASE 2:
if (mVerDistance[Current] > vtxIndex)
End = Current - 1;
else // CASE 3:
Start = Current + 1;
} // End of While()
if (mVerDistance[Current] > vtxIndex)
return (Current - 1);
return Current;
} // End of findOwnerOfGhost()
+303
View File
@@ -0,0 +1,303 @@
#include "MatchBoxPC.h"
void initialize(MilanLongInt NLVer, MilanLongInt NLEdge,
MilanLongInt StartIndex, MilanLongInt EndIndex,
MilanLongInt *numGhostEdges,
MilanLongInt *numGhostVertices,
MilanLongInt *S,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
vector<MilanLongInt> &Counter,
vector<MilanLongInt> &verGhostPtr,
vector<MilanLongInt> &verGhostInd,
vector<MilanLongInt> &tempCounter,
vector<MilanLongInt> &GMate,
vector<MilanLongInt> &Message,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
MilanLongInt *&candidateMate,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner)
{
MilanLongInt insertMe = 0;
MilanLongInt adj1, adj2;
int i, v, k, w;
// index that starts with zero to |Vg| - 1
map<MilanLongInt, MilanLongInt>::iterator storedAlready;
#pragma omp parallel private(insertMe, k, w, v, adj1, adj2) firstprivate(StartIndex, EndIndex) default(shared) num_threads(NUM_THREAD)
{
#pragma omp single
{
#ifdef TIME_TRACKER
double Ghost2LocalInitialization = MPI_Wtime();
#endif
/*
* OMP Ghost2LocalInitialization
* This loop analyzes all the edges and when finds a ghost edge
* puts it in the Ghost2LocalMap.
* A critical region is needed when inserting data in the map.
*
* Despite the critical region it is still productive to
* parallelize this cycle because the critical region is exeuted
* only when a ghost edge is found and ghost edges are a minority,
* circa 3.5% during the tests.
*/
#pragma omp task depend(out \
: *numGhostEdges, Counter, Ghost2LocalMap, insertMe, storedAlready, *numGhostVertices)
{
#pragma omp taskloop num_tasks(NUM_THREAD) reduction(+ \
: numGhostEdges[:1])
for (i = 0; i < NLEdge; i++)
{ // O(m) - Each edge stored twice
insertMe = verLocInd[i];
if ((insertMe < StartIndex) || (insertMe > EndIndex))
{ // Find a ghost
(*numGhostEdges)++;
#pragma omp critical
{
storedAlready = Ghost2LocalMap.find(insertMe);
if (storedAlready != Ghost2LocalMap.end())
{ // Has already been added
Counter[storedAlready->second]++; // Increment the counter
}
else
{ // Insert an entry for the ghost:
Ghost2LocalMap[insertMe] = *numGhostVertices; // Add a map entry
Counter.push_back(1); // Initialize the counter
(*numGhostVertices)++; // Increment the number of ghost vertices
} // End of else()
}
} // End of if ( (insertMe < StartIndex) || (insertMe > EndIndex) )
} // End of for(ghost vertices)
} // end of task depend
// *numGhostEdges = atomicNumGhostEdges;
#ifdef TIME_TRACKER
Ghost2LocalInitialization = MPI_Wtime() - Ghost2LocalInitialization;
fprintf(stderr, "Ghost2LocalInitialization time: %f\n", Ghost2LocalInitialization);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")NGhosts:" << *numGhostVertices << " GhostEdges: " << *numGhostEdges;
if (!Ghost2LocalMap.empty())
{
cout << "\n(" << myRank << ")Final Map : on process ";
cout << "\n(" << myRank << ")Key \t Value \t Counter \n";
fflush(stdout);
storedAlready = Ghost2LocalMap.begin();
do
{
cout << storedAlready->second << " - " << storedAlready->first << " : " << Counter[storedAlready->second] << endl;
fflush(stdout);
storedAlready++;
} while (storedAlready != Ghost2LocalMap.end());
}
#endif
#pragma omp task depend(out \
: verGhostPtr, tempCounter, verGhostInd, GMate) depend(in \
: *numGhostVertices, *numGhostEdges)
{
// Initialize adjacency Lists for Ghost Vertices:
try
{
verGhostPtr.reserve(*numGhostVertices + 1); // Pointer Vector
tempCounter.reserve(*numGhostVertices); // Pointer Vector
verGhostInd.reserve(*numGhostEdges); // Index Vector
GMate.reserve(*numGhostVertices); // Ghost Mate Vector
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesLinearSearch: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
// Initialize the Vectors:
verGhostPtr.resize(*numGhostVertices + 1, 0); // Pointer Vector
tempCounter.resize(*numGhostVertices, 0); // Temporary Counter
verGhostInd.resize(*numGhostEdges, -1); // Index Vector
GMate.resize(*numGhostVertices, -1); // Temporary Counter
verGhostPtr[0] = 0; // The first value
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Ghost Vertex Pointer: ";
fflush(stdout);
#endif
} // End of task
#pragma omp task depend(out \
: verGhostPtr) depend(in \
: Counter, *numGhostVertices)
{
#ifdef TIME_TRACKER
double verGhostPtrInitialization = MPI_Wtime();
#endif
for (i = 0; i < *numGhostVertices; i++)
{ // O(|Ghost Vertices|)
verGhostPtr[i + 1] = verGhostPtr[i] + Counter[i];
#ifdef PRINT_DEBUG_INFO_
cout << verGhostPtr[i] << "\t";
fflush(stdout);
#endif
}
#ifdef TIME_TRACKER
verGhostPtrInitialization = MPI_Wtime() - verGhostPtrInitialization;
fprintf(stderr, "verGhostPtrInitialization time: %f\n", verGhostPtrInitialization);
#endif
} // End of task
#ifdef PRINT_DEBUG_INFO_
if (*numGhostVertices > 0)
cout << verGhostPtr[*numGhostVertices] << "\n";
fflush(stdout);
#endif
#ifdef TIME_TRACKER
double verGhostIndInitialization = MPI_Wtime();
#endif
/*
* OMP verGhostIndInitialization
*
* In this cycle the verGhostInd is initialized
* with the datas related to ghost edges.
* The check to see if a node is a ghost node is
* executed in paralle and when a ghost node
* is found a critical region is started.
*
* Despite the critical region it's still useful to
* parallelize the for cause the ghost nodes
* are a minority hence the critical region is executed
* few times, circa 3.5% of the times in the tests.
*/
#pragma omp task depend(in \
: insertMe, Ghost2LocalMap, tempCounter, verGhostPtr) depend(out \
: verGhostInd)
{
#pragma omp taskloop num_tasks(NUM_THREAD)
for (v = 0; v < NLVer; v++)
{
adj1 = verLocPtr[v]; // Vertex Pointer
adj2 = verLocPtr[v + 1];
for (k = adj1; k < adj2; k++)
{
w = verLocInd[k]; // Get the adjacent vertex
if ((w < StartIndex) || (w > EndIndex))
{ // Find a ghost
#pragma omp critical
{
insertMe = verGhostPtr[Ghost2LocalMap[w]] + tempCounter[Ghost2LocalMap[w]]; // Where to insert
tempCounter[Ghost2LocalMap[w]]++; // Increment the counter
}
verGhostInd[insertMe] = v + StartIndex; // Add the adjacency
} // End of if((w < StartIndex) || (w > EndIndex))
} // End of for(k)
} // End of for (v)
} // end of tasklopp
#ifdef TIME_TRACKER
verGhostIndInitialization = MPI_Wtime() - verGhostIndInitialization;
fprintf(stderr, "verGhostIndInitialization time: %f\n", verGhostIndInitialization);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Ghost Vertex Index: ";
for (v = 0; v < *numGhostEdges; v++)
cout << verGhostInd[v] << "\t";
cout << endl;
fflush(stdout);
#endif
#pragma omp task depend(in \
: *numGhostEdges) depend(out \
: QLocalVtx, QGhostVtx, QMsgType, QOwner)
{
try
{
QLocalVtx.reserve(*numGhostEdges); // Local Vertex
QGhostVtx.reserve(*numGhostEdges); // Ghost Vertex
QMsgType.reserve(*numGhostEdges); // Message Type (Request/Failure)
QOwner.reserve(*numGhostEdges); // Owner of the ghost: COmpute once and use later
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesMessageBundling: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
} // end of task
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Allocating CandidateMate.. ";
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ") Setup Time :" << *ph0_time << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ") Setup Time :" << *ph0_time << endl;
fflush(stdout);
#endif
#pragma omp task depend(in \
: *numGhostVertices) depend(out \
: candidateMate, S, U, privateU, privateQLocalVtx, privateQGhostVtx, privateQMsgType, privateQOwner)
{
// Allocate Data Structures:
/*
* candidateMate was a vector and has been replaced with an array
* there is no point in using the vector (or maybe there is (???))
* so I replaced it with an array wich is slightly faster
*/
candidateMate = new MilanLongInt[NLVer + (*numGhostVertices)];
*S = (*numGhostVertices); // Initialize S with number of Ghost Vertices
/*
* Create the Queue Data Structure for the Dominating Set
*
* I had to declare the staticuQueue U before the parallel region
* to have it in the correct scope. Since we can't change the dimension
* of a staticQueue I had to destroy the previous object and instantiate
* a new one of the correct size.
*/
//new (&U) staticQueue(NLVer + (*numGhostVertices));
U.reserve(NLVer + (*numGhostVertices));
// Initialize the private vectors
privateQLocalVtx.reserve(*numGhostVertices);
privateQGhostVtx.reserve(*numGhostVertices);
privateQMsgType.reserve(*numGhostVertices);
privateQOwner.reserve(*numGhostVertices);
privateU.reserve(*numGhostVertices);
} // end of task
} // End of single region
} // End of parallel region
}
@@ -0,0 +1,45 @@
#include "MatchBoxPC.h"
/**
* //TODO documentation
* @param k
* @param verLocInd
* @param StartIndex
* @param EndIndex
* @param GMate
* @param Mate
* @param Ghost2LocalMap
* @return
*/
bool isAlreadyMatched(MilanLongInt node,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap)
{
/*
#pragma omp critical(Mate)
{
if ((node < StartIndex) || (node > EndIndex)) { //Is it a ghost vertex?
result = GMate[Ghost2LocalMap[node]] >= 0;// Already matched
} else { //A local vertex
result = (Mate[node - StartIndex] >= 0); // Already matched
}
}
*/
MilanLongInt val;
if ((node < StartIndex) || (node > EndIndex)) // if ghost vertex
{
#pragma omp atomic read
val = GMate[Ghost2LocalMap[node]];
return val >= 0; // Already matched
}
// If not ghost vertex
#pragma omp atomic read
val = Mate[node - StartIndex];
return val >= 0; // Already matched
}
@@ -0,0 +1,54 @@
#include "MatchBoxPC.h"
void PARALLEL_COMPUTE_CANDIDATE_MATE_BD(MilanLongInt NLVer,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanInt myRank,
MilanReal *edgeLocWeight,
MilanLongInt *candidateMate)
{
MilanLongInt v = -1;
#pragma omp parallel private(v) default(shared) num_threads(NUM_THREAD)
{
#pragma omp for schedule(static)
for (v = 0; v < NLVer; v++) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Processing: " << v + StartIndex << endl;
fflush(stdout);
#endif
// Start: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
candidateMate[v] = firstComputeCandidateMateD(verLocPtr[v], verLocPtr[v + 1],
verLocInd, edgeLocWeight);
// End: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
}
}
}
void PARALLEL_COMPUTE_CANDIDATE_MATE_BS(MilanLongInt NLVer,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanInt myRank,
MilanFloat *edgeLocWeight,
MilanLongInt *candidateMate)
{
MilanLongInt v = -1;
#pragma omp parallel private(v) default(shared) num_threads(NUM_THREAD)
{
#pragma omp for schedule(static)
for (v = 0; v < NLVer; v++) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Processing: " << v + StartIndex << endl;
fflush(stdout);
#endif
// Start: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
candidateMate[v] = firstComputeCandidateMateS(verLocPtr[v], verLocPtr[v + 1],
verLocInd, edgeLocWeight);
// End: PARALLEL_COMPUTE_CANDIDATE_MATE_B(v)
}
}
}
@@ -0,0 +1,23 @@
#include "MatchBoxPC.h"
void PROCESS_CROSS_EDGE(MilanLongInt *edge,
MilanLongInt *S)
{
// Start: PARALLEL_PROCESS_CROSS_EDGE_B
MilanLongInt captureCounter;
#pragma omp atomic capture
captureCounter = --(*edge); // Decrement
//assert(captureCounter >= 0);
if (captureCounter == 0)
#pragma omp atomic
(*S)--; // Decrement S
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Decrementing S: Ghost vertex " << edge << " has received all its messages";
fflush(stdout);
#endif
// End: PARALLEL_PROCESS_CROSS_EDGE_B
}
@@ -0,0 +1,368 @@
#include "MatchBoxPC.h"
void PARALLEL_PROCESS_EXPOSED_VERTEX_BD(MilanLongInt NLVer,
MilanLongInt *candidateMate,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *Mate,
vector<MilanLongInt> &GMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *S,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner)
{
MilanLongInt v = -1, k = -1, w = -1, adj11 = 0, adj12 = 0, k1 = 0;
MilanInt ghostOwner = 0, option, igw;
//#pragma omp parallel private(option, k, w, v, k1, adj11, adj12, ghostOwner) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, \
privateQGhostVtx, privateQMsgType, privateQOwner) \
default(shared) num_threads(NUM_THREAD)
{
//#pragma omp for reduction(+ \
: PCounter[:numProcs], myCard \
[:1], msgInd \
[:1], NumMessagesBundled \
[:1]) \
schedule(static)
for (v = 0; v < NLVer; v++) {
option = -1;
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
k = candidateMate[v];
candidateMate[v] = verLocInd[k];
w = candidateMate[v];
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Processing: " << v + StartIndex << endl;
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v + StartIndex << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
#pragma omp critical(Matching)
{
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap)) {
w = computeCandidateMateD(verLocPtr[v], verLocPtr[v + 1], edgeLocWeight, 0,
verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v] = w;
}
}
if (w >= 0) {
(*myCard)++;
if ((w < StartIndex) || (w > EndIndex)) { // w is a ghost vertex
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v + StartIndex) {
option = 1;
Mate[v] = w;
GMate[Ghost2LocalMap[w]] = v + StartIndex; // w is a Ghost
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == (v + StartIndex)) {
option = 3;
Mate[v] = w; // v is local
Mate[w - StartIndex] = v + StartIndex; // w is local
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v + StartIndex << "," << w << ") ";
fflush(stdout);
#endif
} // End of if ( candidateMate[w-StartIndex] == (v+StartIndex) )
} // End of Else
} // End of second if
} // End of if(w >=0)
else {
//#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);
#endif
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
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)
switch (option)
{
case -1:
break;
case 1:
privateU.push_back(v + StartIndex);
privateU.push_back(w);
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v + StartIndex << "," << w << ")";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], S);
case 2:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message (291):";
cout << "\n(" << myRank << ")Local is: " << v + StartIndex << " Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
privateQLocalVtx.push_back(v + StartIndex);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
break;
case 3:
default:
privateU.push_back(v + StartIndex);
privateU.push_back(w);
break;
}
} // End of for ( v=0; v < NLVer; v++ )
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
} // End of parallel region
}
void PARALLEL_PROCESS_EXPOSED_VERTEX_BS(MilanLongInt NLVer,
MilanLongInt *candidateMate,
MilanLongInt *verLocInd,
MilanLongInt *verLocPtr,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *Mate,
vector<MilanLongInt> &GMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanFloat *edgeLocWeight,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *S,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner)
{
MilanLongInt v = -1, k = -1, w = -1, adj11 = 0, adj12 = 0, k1 = 0;
MilanInt ghostOwner = 0, option, igw;
//#pragma omp parallel private(option, k, w, v, k1, adj11, adj12, ghostOwner) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, \
privateQGhostVtx, privateQMsgType, privateQOwner) \
default(shared) num_threads(NUM_THREAD)
{
//#pragma omp for reduction(+ \
: PCounter[:numProcs], myCard \
[:1], msgInd \
[:1], NumMessagesBundled \
[:1]) \
schedule(static)
for (v = 0; v < NLVer; v++) {
option = -1;
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
k = candidateMate[v];
candidateMate[v] = verLocInd[k];
w = candidateMate[v];
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Processing: " << v + StartIndex << endl;
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v + StartIndex << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
#pragma omp critical(Matching)
{
if (isAlreadyMatched(verLocInd[k], StartIndex, EndIndex, GMate, Mate, Ghost2LocalMap)) {
w = computeCandidateMateS(verLocPtr[v], verLocPtr[v + 1], edgeLocWeight, 0,
verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v] = w;
}
if (w >= 0) {
(*myCard)++;
if ((w < StartIndex) || (w > EndIndex)) { // w is a ghost vertex
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v + StartIndex) {
option = 1;
Mate[v] = w;
GMate[Ghost2LocalMap[w]] = v + StartIndex; // w is a Ghost
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == (v + StartIndex)) {
option = 3;
Mate[v] = w; // v is local
Mate[w - StartIndex] = v + StartIndex; // w is local
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v + StartIndex << "," << w << ") ";
fflush(stdout);
#endif
} // End of if ( candidateMate[w-StartIndex] == (v+StartIndex) )
} // End of Else
} // End of second if
}
} // End of if(w >=0)
else {
//#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);
#endif
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
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)
switch (option)
{
case -1:
break;
case 1:
privateU.push_back(v + StartIndex);
privateU.push_back(w);
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v + StartIndex << "," << w << ")";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], S);
case 2:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message (291):";
cout << "\n(" << myRank << ")Local is: " << v + StartIndex << " Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
(*msgInd)++;
(*NumMessagesBundled)++;
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
privateQLocalVtx.push_back(v + StartIndex);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
break;
case 3:
default:
privateU.push_back(v + StartIndex);
privateU.push_back(w);
break;
}
} // End of for ( v=0; v < NLVer; v++ )
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
} // End of parallel region
}
@@ -0,0 +1,584 @@
#include "MatchBoxPC.h"
void processMatchedVerticesD(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner)
{
MilanLongInt adj1, adj2, adj11, adj12, k, k1, v = -1, w = -1, ghostOwner;
int option;
MilanLongInt mateVal;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef COUNT_LOCAL_VERTEX
MilanLongInt localVertices = 0;
#endif
//#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, \
privateQMsgType, privateQOwner, UChunkBeingProcessed) \
default(shared) num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
[:1])
{
while (!U.empty()) {
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
#endif
if ((u >= StartIndex) && (u <= EndIndex)) { // Process Only the Local Vertices
#ifdef COUNT_LOCAL_VERTEX
localVertices++;
#endif
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++) {
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex)) { // If Local Vertex:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
#endif
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
#pragma omp critical
{
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMateD(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
#endif
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
}
} // End of task
} // mateval < 0
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else { // Neighbor is a ghost vertex
#pragma omp critical
{
if (candidateMate[NLVer + Ghost2LocalMap[v]] == u)
candidateMate[NLVer + Ghost2LocalMap[v]] = -1;
if (v != Mate[u - StartIndex])
option = 5; // u is local
} // End of critical
} // End of Else //A Ghost Vertex
switch (option)
{
case -1:
// No things to do
break;
case 1:
// Found a dominating edge, it is a ghost and candidateMate[NLVer + Ghost2LocalMap[w]] == v
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
break;
case 3:
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
break;
case 4:
// Could not find a dominating vertex
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 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);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
(*NumMessagesBundled)++;
PCounter[ghostOwner]++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
#pragma omp critical(U)
{
U.insert(U.end(), privateU.begin(), privateU.end());
}
#pragma omp critical(sendMessageTransfer)
{
QLocalVtx.insert(QLocalVtx.end(), privateQLocalVtx.begin(), privateQLocalVtx.end());
QGhostVtx.insert(QGhostVtx.end(), privateQGhostVtx.begin(), privateQGhostVtx.end());
QMsgType.insert(QMsgType.end(), privateQMsgType.begin(), privateQMsgType.end());
QOwner.insert(QOwner.end(), privateQOwner.begin(), privateQOwner.end());
}
privateU.clear();
privateQLocalVtx.clear();
privateQGhostVtx.clear();
privateQMsgType.clear();
privateQOwner.clear();
} // End of while ( !U.empty() )
#ifdef COUNT_LOCAL_VERTEX
printf("Count local vertexes: %ld for thread %d of processor %d\n",
localVertices,
omp_get_thread_num(),
myRank);
#endif
} // End of parallel region
}
void processMatchedVerticesS(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanFloat *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner)
{
MilanLongInt adj1, adj2, adj11, adj12, k, k1, v = -1, w = -1, ghostOwner;
int option;
MilanLongInt mateVal;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef COUNT_LOCAL_VERTEX
MilanLongInt localVertices = 0;
#endif
//#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx, \
privateQMsgType, privateQOwner, UChunkBeingProcessed) \
default(shared) num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
[:1])
{
while (!U.empty()) {
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
#endif
if ((u >= StartIndex) && (u <= EndIndex)) { // Process Only the Local Vertices
#ifdef COUNT_LOCAL_VERTEX
localVertices++;
#endif
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++) {
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex)) { // If Local Vertex:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
#endif
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
#pragma omp critical
{
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMateS(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
#endif
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
}
} // End of task
} // mateval < 0
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else { // Neighbor is a ghost vertex
#pragma omp critical
{
if (candidateMate[NLVer + Ghost2LocalMap[v]] == u)
candidateMate[NLVer + Ghost2LocalMap[v]] = -1;
if (v != Mate[u - StartIndex])
option = 5; // u is local
} // End of critical
} // End of Else //A Ghost Vertex
switch (option)
{
case -1:
// No things to do
break;
case 1:
// Found a dominating edge, it is a ghost and candidateMate[NLVer + Ghost2LocalMap[w]] == v
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
break;
case 3:
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
break;
case 4:
// Could not find a dominating vertex
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 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);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
#pragma omp atomic
PCounter[ghostOwner]++;
(*NumMessagesBundled)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// assert(ghostOwner != -1);
// assert(ghostOwner != myRank);
(*NumMessagesBundled)++;
PCounter[ghostOwner]++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
#pragma omp critical(U)
{
U.insert(U.end(), privateU.begin(), privateU.end());
}
#pragma omp critical(sendMessageTransfer)
{
QLocalVtx.insert(QLocalVtx.end(), privateQLocalVtx.begin(), privateQLocalVtx.end());
QGhostVtx.insert(QGhostVtx.end(), privateQGhostVtx.begin(), privateQGhostVtx.end());
QMsgType.insert(QMsgType.end(), privateQMsgType.begin(), privateQMsgType.end());
QOwner.insert(QOwner.end(), privateQOwner.begin(), privateQOwner.end());
}
privateU.clear();
privateQLocalVtx.clear();
privateQGhostVtx.clear();
privateQMsgType.clear();
privateQOwner.clear();
} // End of while ( !U.empty() )
#ifdef COUNT_LOCAL_VERTEX
printf("Count local vertexes: %ld for thread %d of processor %d\n",
localVertices,
omp_get_thread_num(),
myRank);
#endif
} // End of parallel region
}
@@ -0,0 +1,590 @@
#include "MatchBoxPC.h"
//#define DEBUG_HANG_
void processMatchedVerticesAndSendMessagesD(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanReal *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner,
MPI_Comm comm,
MilanLongInt *msgActual,
vector<MilanLongInt> &Message)
{
MilanLongInt initialSize = QLocalVtx.size();
MilanLongInt adj1, adj2, adj11, adj12, k, k1, v = -1, w = -1, ghostOwner;
int option;
MilanLongInt mateVal;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef COUNT_LOCAL_VERTEX
MilanLongInt localVertices = 0;
#endif
//#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(Message, privateU, StartIndex, EndIndex, privateQLocalVtx, \
privateQGhostVtx, privateQMsgType, privateQOwner, UChunkBeingProcessed) \
default(shared) \
num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
[:1], msgActual \
[:1])
{
while (!U.empty()) {
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
#endif
if ((u >= StartIndex) && (u <= EndIndex)) { // Process Only the Local Vertices
#ifdef COUNT_LOCAL_VERTEX
localVertices++;
#endif
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++) {
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex)) { // If Local Vertex:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
#endif
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
#pragma omp critical
{
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMateD(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
#endif
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
}
} // End of task
} // mateval < 0
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else { // Neighbor is a ghost vertex
#pragma omp critical
{
if (candidateMate[NLVer + Ghost2LocalMap[v]] == u)
candidateMate[NLVer + Ghost2LocalMap[v]] = -1;
if (v != Mate[u - StartIndex])
option = 5; // u is local
} // End of critical
} // End of Else //A Ghost Vertex
switch (option)
{
case -1:
// No things to do
break;
case 1:
// Found a dominating edge, it is a ghost and candidateMate[NLVer + Ghost2LocalMap[w]] == v
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
break;
case 3:
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
break;
case 4:
// Could not find a dominating vertex
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 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);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = u; // LOCAL
// Message[1] = v; // GHOST
// Message[2] = SUCCESS; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
} // End of while ( !U.empty() )
#ifdef COUNT_LOCAL_VERTEX
printf("Count local vertexes: %ld for thread %d of processor %d\n",
localVertices, mp_get_thread_num(), myRank);
#endif
} // End of parallel region
// Send the messages
#ifdef DEBUG_HANG_
cout << myRank<<" Sending: "<<QOwner.size()-initialSize<<" messages" <<endl;
#endif
for (int i = initialSize; i < QOwner.size(); i++) {
Message[0] = QLocalVtx[i];
Message[1] = QGhostVtx[i];
Message[2] = QMsgType[i];
ghostOwner = QOwner[i];
//MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
}
#ifdef DEBUG_HANG_
cout << myRank<<" Done sending messages"<<endl;
#endif
}
void processMatchedVerticesAndSendMessagesS(
MilanLongInt NLVer,
vector<MilanLongInt> &UChunkBeingProcessed,
vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *NumMessagesBundled,
MilanLongInt *SPtr,
MilanLongInt *verLocPtr,
MilanLongInt *verLocInd,
MilanLongInt *verDistance,
MilanLongInt *PCounter,
vector<MilanLongInt> &Counter,
MilanInt myRank,
MilanInt numProcs,
MilanLongInt *candidateMate,
vector<MilanLongInt> &GMate,
MilanLongInt *Mate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
MilanFloat *edgeLocWeight,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner,
MPI_Comm comm,
MilanLongInt *msgActual,
vector<MilanLongInt> &Message)
{
MilanLongInt initialSize = QLocalVtx.size();
MilanLongInt adj1, adj2, adj11, adj12, k, k1, v = -1, w = -1, ghostOwner;
int option;
MilanLongInt mateVal;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef COUNT_LOCAL_VERTEX
MilanLongInt localVertices = 0;
#endif
//#pragma omp parallel private(k, w, v, k1, adj1, adj2, adj11, adj12, ghostOwner, option) \
firstprivate(Message, privateU, StartIndex, EndIndex, privateQLocalVtx, privateQGhostVtx,\
privateQMsgType, privateQOwner, UChunkBeingProcessed) default(shared) \
num_threads(NUM_THREAD) \
reduction(+ \
: msgInd[:1], PCounter \
[:numProcs], myCard \
[:1], NumMessagesBundled \
[:1], msgActual \
[:1])
{
while (!U.empty()) {
extractUChunk(UChunkBeingProcessed, U, privateU);
for (MilanLongInt u : UChunkBeingProcessed) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")u: " << u;
fflush(stdout);
#endif
if ((u >= StartIndex) && (u <= EndIndex)) { // Process Only the Local Vertices
#ifdef COUNT_LOCAL_VERTEX
localVertices++;
#endif
// Get the Adjacency list for u
adj1 = verLocPtr[u - StartIndex]; // Pointer
adj2 = verLocPtr[u - StartIndex + 1];
for (k = adj1; k < adj2; k++) {
option = -1;
v = verLocInd[k];
if ((v >= StartIndex) && (v <= EndIndex)) { // If Local Vertex:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")v: " << v << " c(v)= " << candidateMate[v - StartIndex] << " Mate[v]: " << Mate[v];
fflush(stdout);
#endif
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
#pragma omp critical
{
#pragma omp atomic read
mateVal = Mate[v - StartIndex];
// If the current vertex is pointing to a matched vertex and is not matched
if (mateVal < 0) {
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMateS(verLocPtr[v - StartIndex],
verLocPtr[v - StartIndex + 1],
edgeLocWeight, 0,
verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) { // A ghost
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message:";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs);
#endif
option = 2;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
option = 1;
Mate[v - StartIndex] = w; // v is a local vertex
GMate[Ghost2LocalMap[w]] = v; // w is a ghost vertex
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
option = 3;
Mate[v - StartIndex] = w; // v is a local vertex
Mate[w - StartIndex] = v; // w is a local vertex
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else
option = 4; // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of If (candidateMate[v-StartIndex] == u
}
} // End of task
} // mateval < 0
} // End of if ( (v >= StartIndex) && (v <= EndIndex) ) //If Local Vertex:
else { // Neighbor is a ghost vertex
#pragma omp critical
{
if (candidateMate[NLVer + Ghost2LocalMap[v]] == u)
candidateMate[NLVer + Ghost2LocalMap[v]] = -1;
if (v != Mate[u - StartIndex])
option = 5; // u is local
} // End of critical
} // End of Else //A Ghost Vertex
switch (option)
{
case -1:
// No things to do
break;
case 1:
// Found a dominating edge, it is a ghost and candidateMate[NLVer + Ghost2LocalMap[w]] == v
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") ";
fflush(stdout);
#endif
// Decrement the counter:
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], SPtr);
case 2:
// Found a dominating edge, it is a ghost
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(REQUEST);
privateQOwner.push_back(ghostOwner);
break;
case 3:
privateU.push_back(v);
privateU.push_back(w);
(*myCard)++;
break;
case 4:
// Could not find a dominating vertex
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 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);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = v; // LOCAL
// Message[1] = w; // GHOST
// Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(v);
privateQGhostVtx.push_back(w);
privateQMsgType.push_back(FAILURE);
privateQOwner.push_back(ghostOwner);
} // End of if(GHOST)
} // End of for loop
break;
case 5:
default:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a success message: ";
cout << "\n(" << myRank << ")Ghost is " << v << " Owner is: " << findOwnerOfGhost(v, verDistance, myRank, numProcs) << "\n";
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(v, verDistance, myRank, numProcs);
// Build the Message Packet:
// Message[0] = u; // LOCAL
// Message[1] = v; // GHOST
// Message[2] = SUCCESS; // TYPE
// Send a Request (Asynchronous)
// MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgActual)++;
(*msgInd)++;
privateQLocalVtx.push_back(u);
privateQGhostVtx.push_back(v);
privateQMsgType.push_back(SUCCESS);
privateQOwner.push_back(ghostOwner);
break;
} // End of switch
} // End of inner for
}
} // End of outer for
queuesTransfer(U, privateU, QLocalVtx,
QGhostVtx,
QMsgType, QOwner, privateQLocalVtx,
privateQGhostVtx,
privateQMsgType,
privateQOwner);
} // End of while ( !U.empty() )
#ifdef COUNT_LOCAL_VERTEX
printf("Count local vertexes: %ld for thread %d of processor %d\n",
localVertices, mp_get_thread_num(), myRank);
#endif
} // End of parallel region
// Send the messages
#ifdef DEBUG_HANG_
cout << myRank<<" Sending: "<<QOwner.size()-initialSize<<" messages" <<endl;
#endif
for (int i = initialSize; i < QOwner.size(); i++) {
Message[0] = QLocalVtx[i];
Message[1] = QGhostVtx[i];
Message[2] = QMsgType[i];
ghostOwner = QOwner[i];
//MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
}
#ifdef DEBUG_HANG_
cout << myRank<<" Done sending messages"<<endl;
#endif
}
+631
View File
@@ -0,0 +1,631 @@
#include "MatchBoxPC.h"
//#define DEBUG_HANG_
void processMessagesD(
MilanLongInt NLVer,
MilanLongInt *Mate,
MilanLongInt *candidateMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
vector<MilanLongInt> &GMate,
vector<MilanLongInt> &Counter,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *msgActual,
MilanReal *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *verLocPtr,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanInt numProcs,
MilanInt myRank,
MPI_Comm comm,
vector<MilanLongInt> &Message,
MilanLongInt numGhostEdges,
MilanLongInt u,
MilanLongInt v,
MilanLongInt *S,
vector<MilanLongInt> &U)
{
//#define PRINT_DEBUG_INFO_
MilanInt Sender;
MPI_Status computeStatus;
MilanLongInt bundleSize, w;
MilanLongInt adj11, adj12, k1;
MilanLongInt ghostOwner;
int error_codeC;
error_codeC = MPI_Comm_set_errhandler(MPI_COMM_WORLD, MPI_ERRORS_RETURN);
char error_message[MPI_MAX_ERROR_STRING];
int message_length;
MilanLongInt message_type = 0;
// Buffer to receive bundled messages
// Maximum messages that can be received from any processor is
// twice the edge cut: REQUEST; REQUEST+(FAILURE/SUCCESS)
vector<MilanLongInt> ReceiveBuffer;
try
{
ReceiveBuffer.reserve(numGhostEdges * 2 * 3); // Three integers per cross edge
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesMessageBundling: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
#ifdef PRINT_DEBUG_INFO_
cout
<< "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")About to begin Message processing phase ... *S=" << *S << endl;
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
// BLOCKING RECEIVE:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << " Waiting for blocking receive..." << endl;
fflush(stdout);
fflush(stdout);
#endif
//cout << myRank<<" Receiving ...";
error_codeC = MPI_Recv(&Message[0], 3, TypeMap<MilanLongInt>(), MPI_ANY_SOURCE, ComputeTag, comm, &computeStatus);
if (error_codeC != MPI_SUCCESS)
{
MPI_Error_string(error_codeC, error_message, &message_length);
cout << "\n*Error in call to MPI_Receive on Slave: " << error_message << "\n";
fflush(stdout);
}
Sender = computeStatus.MPI_SOURCE;
//cout << " ...from "<<Sender << endl;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received message from Process " << Sender << " Type= " << Message[2] << endl;
fflush(stdout);
#endif
if (Message[2] == SIZEINFO) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received bundled message from Process " << Sender << " Size= " << Message[0] << endl;
fflush(stdout);
#endif
bundleSize = Message[0]; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message Bundle Before: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
#endif
// Receive the message
//cout << myRank<<" Receiving from "<<Sender<<endl;
error_codeC = MPI_Recv(&ReceiveBuffer[0], bundleSize, TypeMap<MilanLongInt>(), Sender, BundleTag, comm, &computeStatus);
if (error_codeC != MPI_SUCCESS) {
MPI_Error_string(error_codeC, error_message, &message_length);
cout << "\n*Error in call to MPI_Receive on processor " << myRank << " Error: " << error_message << "\n";
fflush(stdout);
}
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message Bundle After: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
#endif
} else { // Just a single message:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received regular message from Process " << Sender << " u= " << Message[0] << " v= " << Message[1] << endl;
fflush(stdout);
#endif
// Add the current message to Queue:
bundleSize = 3; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
ReceiveBuffer[0] = Message[0]; // u
ReceiveBuffer[1] = Message[1]; // v
ReceiveBuffer[2] = Message[2]; // message_type
}
#ifdef DEBUG_GHOST_
if ((v < StartIndex) || (v > EndIndex)) {
cout << "\n(" << myRank << ") From ReceiveBuffer: This should not happen: u= " << u << " v= " << v << " Type= " << message_type << " StartIndex " << StartIndex << " EndIndex " << EndIndex << endl;
fflush(stdout);
}
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Processing message: u= " << u << " v= " << v << " Type= " << message_type << endl;
fflush(stdout);
#endif
// Most of the time bundleSize == 3, thus, it's not worth parallelizing thi loop
for (MilanLongInt bundleCounter = 3; bundleCounter < bundleSize + 3; bundleCounter += 3) {
u = ReceiveBuffer[bundleCounter - 3]; // GHOST
v = ReceiveBuffer[bundleCounter - 2]; // LOCAL
message_type = ReceiveBuffer[bundleCounter - 1]; // TYPE
// CASE I: REQUEST
if (message_type == REQUEST) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is REQUEST" << endl;
fflush(stdout);
#endif
#ifdef DEBUG_GHOST_
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer)) {
cout << "\n(" << myRank << ") case 1 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
#endif
if (Mate[v - StartIndex] == -1) {
// Process only if not already matched (v is local)
candidateMate[NLVer + Ghost2LocalMap[u]] = v; // Set CandidateMate for the ghost
if (candidateMate[v - StartIndex] == u) {
GMate[Ghost2LocalMap[u]] = v; // u is ghost
Mate[v - StartIndex] = u; // v is local
U.push_back(v);
U.push_back(u);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << u << ") " << endl;
fflush(stdout);
#endif
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
} // End of if ( candidateMate[v-StartIndex] == u )e
} // End of if ( Mate[v] == -1 )
} // End of REQUEST
else { // CASE II: SUCCESS
if (message_type == SUCCESS) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is SUCCESS" << endl;
fflush(stdout);
#endif
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process it again
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
#ifdef DEBUG_GHOST_
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer)) {
cout << "\n(" << myRank << ") case 2 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
#endif
if (Mate[v - StartIndex] == -1) {
// Process only if not already matched ( v is local)
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMateD(verLocPtr[v - StartIndex], verLocPtr[v - StartIndex + 1],
edgeLocWeight, k,verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w << endl;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) {
// w is a ghost
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
Mate[v - StartIndex] = w; // v is local
GMate[Ghost2LocalMap[w]] = v; // w is ghost
U.push_back(v);
U.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
#endif
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], S);
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
Mate[v - StartIndex] = w; // v is local
Mate[w - StartIndex] = v; // w is local
// Q.push_back(u);
U.push_back(v);
U.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else { // No dominant edge found
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 1];
for (k1 = adj11; k1 < adj12; k1++) {
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex)) {
// A ghost
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
} // End of if(GHOST)
} // End of for loop
} // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of if ( candidateMate[v-StartIndex] == u )
} // End of if ( Mate[v] == -1 )
} // End of if ( message_type == SUCCESS )
else {
// CASE III: FAILURE
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is FAILURE" << endl;
fflush(stdout);
#endif
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process this anymore
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S); // Decrease the counter
} // End of else: CASE III
} // End of else: CASE I
}
return;
}
void processMessagesS(
MilanLongInt NLVer,
MilanLongInt *Mate,
MilanLongInt *candidateMate,
map<MilanLongInt, MilanLongInt> &Ghost2LocalMap,
vector<MilanLongInt> &GMate,
vector<MilanLongInt> &Counter,
MilanLongInt StartIndex,
MilanLongInt EndIndex,
MilanLongInt *myCard,
MilanLongInt *msgInd,
MilanLongInt *msgActual,
MilanFloat *edgeLocWeight,
MilanLongInt *verDistance,
MilanLongInt *verLocPtr,
MilanLongInt k,
MilanLongInt *verLocInd,
MilanInt numProcs,
MilanInt myRank,
MPI_Comm comm,
vector<MilanLongInt> &Message,
MilanLongInt numGhostEdges,
MilanLongInt u,
MilanLongInt v,
MilanLongInt *S,
vector<MilanLongInt> &U)
{
//#define PRINT_DEBUG_INFO_
MilanInt Sender;
MPI_Status computeStatus;
MilanLongInt bundleSize, w;
MilanLongInt adj11, adj12, k1;
MilanLongInt ghostOwner;
int error_codeC;
error_codeC = MPI_Comm_set_errhandler(MPI_COMM_WORLD, MPI_ERRORS_RETURN);
char error_message[MPI_MAX_ERROR_STRING];
int message_length;
MilanLongInt message_type = 0;
// Buffer to receive bundled messages
// Maximum messages that can be received from any processor is
// twice the edge cut: REQUEST; REQUEST+(FAILURE/SUCCESS)
vector<MilanLongInt> ReceiveBuffer;
try
{
ReceiveBuffer.reserve(numGhostEdges * 2 * 3); // Three integers per cross edge
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesMessageBundling: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
#ifdef PRINT_DEBUG_INFO_
cout
<< "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")About to begin Message processing phase ... *S=" << *S << endl;
fflush(stdout);
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << "=========================************===============================" << endl;
fflush(stdout);
fflush(stdout);
#endif
// BLOCKING RECEIVE:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << " Waiting for blocking receive..." << endl;
fflush(stdout);
fflush(stdout);
#endif
//cout << myRank<<" Receiving ...";
error_codeC = MPI_Recv(&Message[0], 3, TypeMap<MilanLongInt>(), MPI_ANY_SOURCE, ComputeTag, comm, &computeStatus);
if (error_codeC != MPI_SUCCESS)
{
MPI_Error_string(error_codeC, error_message, &message_length);
cout << "\n*Error in call to MPI_Receive on Slave: " << error_message << "\n";
fflush(stdout);
}
Sender = computeStatus.MPI_SOURCE;
//cout << " ...from "<<Sender << endl;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received message from Process " << Sender << " Type= " << Message[2] << endl;
fflush(stdout);
#endif
if (Message[2] == SIZEINFO) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received bundled message from Process " << Sender << " Size= " << Message[0] << endl;
fflush(stdout);
#endif
bundleSize = Message[0]; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message Bundle Before: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
#endif
// Receive the message
//cout << myRank<<" Receiving from "<<Sender<<endl;
error_codeC = MPI_Recv(&ReceiveBuffer[0], bundleSize, TypeMap<MilanLongInt>(), Sender, BundleTag, comm, &computeStatus);
if (error_codeC != MPI_SUCCESS) {
MPI_Error_string(error_codeC, error_message, &message_length);
cout << "\n*Error in call to MPI_Receive on processor " << myRank << " Error: " << error_message << "\n";
fflush(stdout);
}
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message Bundle After: " << endl;
for (int i = 0; i < bundleSize; i++)
cout << ReceiveBuffer[i] << ",";
cout << endl;
fflush(stdout);
#endif
} else { // Just a single message:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Received regular message from Process " << Sender << " u= " << Message[0] << " v= " << Message[1] << endl;
fflush(stdout);
#endif
// Add the current message to Queue:
bundleSize = 3; //#of integers in the message
// Build the Message Buffer:
if (!ReceiveBuffer.empty())
ReceiveBuffer.clear(); // Empty it out first
ReceiveBuffer.resize(bundleSize, -1); // Initialize
ReceiveBuffer[0] = Message[0]; // u
ReceiveBuffer[1] = Message[1]; // v
ReceiveBuffer[2] = Message[2]; // message_type
}
#ifdef DEBUG_GHOST_
if ((v < StartIndex) || (v > EndIndex)) {
cout << "\n(" << myRank << ") From ReceiveBuffer: This should not happen: u= " << u << " v= " << v << " Type= " << message_type << " StartIndex " << StartIndex << " EndIndex " << EndIndex << endl;
fflush(stdout);
}
#endif
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Processing message: u= " << u << " v= " << v << " Type= " << message_type << endl;
fflush(stdout);
#endif
// Most of the time bundleSize == 3, thus, it's not worth parallelizing thi loop
for (MilanLongInt bundleCounter = 3; bundleCounter < bundleSize + 3; bundleCounter += 3) {
u = ReceiveBuffer[bundleCounter - 3]; // GHOST
v = ReceiveBuffer[bundleCounter - 2]; // LOCAL
message_type = ReceiveBuffer[bundleCounter - 1]; // TYPE
// CASE I: REQUEST
if (message_type == REQUEST) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is REQUEST" << endl;
fflush(stdout);
#endif
#ifdef DEBUG_GHOST_
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer)) {
cout << "\n(" << myRank << ") case 1 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
#endif
if (Mate[v - StartIndex] == -1) {
// Process only if not already matched (v is local)
candidateMate[NLVer + Ghost2LocalMap[u]] = v; // Set CandidateMate for the ghost
if (candidateMate[v - StartIndex] == u) {
GMate[Ghost2LocalMap[u]] = v; // u is ghost
Mate[v - StartIndex] = u; // v is local
U.push_back(v);
U.push_back(u);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << u << ") " << endl;
fflush(stdout);
#endif
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
} // End of if ( candidateMate[v-StartIndex] == u )e
} // End of if ( Mate[v] == -1 )
} // End of REQUEST
else { // CASE II: SUCCESS
if (message_type == SUCCESS) {
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is SUCCESS" << endl;
fflush(stdout);
#endif
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process it again
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S);
#ifdef DEBUG_GHOST_
if ((v < 0) || (v < StartIndex) || ((v - StartIndex) > NLVer)) {
cout << "\n(" << myRank << ") case 2 Bad address " << v << " " << StartIndex << " " << v - StartIndex << " " << NLVer << endl;
fflush(stdout);
}
#endif
if (Mate[v - StartIndex] == -1) {
// Process only if not already matched ( v is local)
if (candidateMate[v - StartIndex] == u) {
// Start: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
w = computeCandidateMateS(verLocPtr[v - StartIndex], verLocPtr[v - StartIndex + 1],
edgeLocWeight, k,verLocInd, StartIndex, EndIndex,
GMate, Mate, Ghost2LocalMap);
candidateMate[v - StartIndex] = w;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")" << v << " Points to: " << w << endl;
fflush(stdout);
#endif
// If found a dominating edge:
if (w >= 0) {
if ((w < StartIndex) || (w > EndIndex)) {
// w is a ghost
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = REQUEST; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a request message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
if (candidateMate[NLVer + Ghost2LocalMap[w]] == v) {
Mate[v - StartIndex] = w; // v is local
GMate[Ghost2LocalMap[w]] = v; // w is ghost
U.push_back(v);
U.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
#endif
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[w]], S);
} // End of if CandidateMate[w] = v
} // End of if a Ghost Vertex
else { // w is a local vertex
if (candidateMate[w - StartIndex] == v) {
Mate[v - StartIndex] = w; // v is local
Mate[w - StartIndex] = v; // w is local
// Q.push_back(u);
U.push_back(v);
U.push_back(w);
(*myCard)++;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")MATCH: (" << v << "," << w << ") " << endl;
fflush(stdout);
#endif
} // End of if(CandidateMate(w) = v
} // End of Else
} // End of if(w >=0)
else { // No dominant edge found
adj11 = verLocPtr[v - StartIndex];
adj12 = verLocPtr[v - StartIndex + 1];
for (k1 = adj11; k1 < adj12; k1++) {
w = verLocInd[k1];
if ((w < StartIndex) || (w > EndIndex)) {
// A ghost
// Build the Message Packet:
Message[0] = v; // LOCAL
Message[1] = w; // GHOST
Message[2] = FAILURE; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending a failure message: ";
cout << "\n(" << myRank << ")Ghost is " << w << " Owner is: " << findOwnerOfGhost(w, verDistance, myRank, numProcs) << endl;
fflush(stdout);
#endif
ghostOwner = findOwnerOfGhost(w, verDistance, myRank, numProcs);
//assert(ghostOwner != -1);
//assert(ghostOwner != myRank);
//cout << myRank<<" Sending to "<<ghostOwner<<endl;
MPI_Bsend(&Message[0], 3, TypeMap<MilanLongInt>(), ghostOwner, ComputeTag, comm);
(*msgInd)++;
(*msgActual)++;
} // End of if(GHOST)
} // End of for loop
} // End of Else: w == -1
// End: PARALLEL_PROCESS_EXPOSED_VERTEX_B(v)
} // End of if ( candidateMate[v-StartIndex] == u )
} // End of if ( Mate[v] == -1 )
} // End of if ( message_type == SUCCESS )
else {
// CASE III: FAILURE
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Message type is FAILURE" << endl;
fflush(stdout);
#endif
GMate[Ghost2LocalMap[u]] = EndIndex + 1; // Set a Dummy Mate to make sure that we do not (u is a ghost) process this anymore
PROCESS_CROSS_EDGE(&Counter[Ghost2LocalMap[u]], S); // Decrease the counter
} // End of else: CASE III
} // End of else: CASE I
}
return;
}
+32
View File
@@ -0,0 +1,32 @@
#include "MatchBoxPC.h"
void queuesTransfer(vector<MilanLongInt> &U,
vector<MilanLongInt> &privateU,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MilanLongInt> &privateQLocalVtx,
vector<MilanLongInt> &privateQGhostVtx,
vector<MilanLongInt> &privateQMsgType,
vector<MilanInt> &privateQOwner)
{
#pragma omp critical(U)
{
U.insert(U.end(), privateU.begin(), privateU.end());
//#pragma omp critical(sendMessageTransfer)
//{
QLocalVtx.insert(QLocalVtx.end(), privateQLocalVtx.begin(), privateQLocalVtx.end());
QGhostVtx.insert(QGhostVtx.end(), privateQGhostVtx.begin(), privateQGhostVtx.end());
QMsgType.insert(QMsgType.end(), privateQMsgType.begin(), privateQMsgType.end());
QOwner.insert(QOwner.end(), privateQOwner.begin(), privateQOwner.end());
}
privateU.clear();
privateQLocalVtx.clear();
privateQGhostVtx.clear();
privateQMsgType.clear();
privateQOwner.clear();
}
@@ -0,0 +1,208 @@
#include "MatchBoxPC.h"
void sendBundledMessages(MilanLongInt *numGhostEdges,
MilanInt *BufferSize,
MilanLongInt *Buffer,
vector<MilanLongInt> &PCumulative,
vector<MilanLongInt> &PMessageBundle,
vector<MilanLongInt> &PSizeInfoMessages,
MilanLongInt *PCounter,
MilanLongInt NumMessagesBundled,
MilanLongInt *msgActual,
MilanLongInt *msgInd,
MilanInt numProcs,
MilanInt myRank,
MPI_Comm comm,
vector<MilanLongInt> &QLocalVtx,
vector<MilanLongInt> &QGhostVtx,
vector<MilanLongInt> &QMsgType,
vector<MilanInt> &QOwner,
vector<MPI_Request> &SRequest,
vector<MPI_Status> &SStatus)
{
MilanLongInt myIndex = 0, numMessagesToSend;
MilanInt i = 0, OneMessageSize = 0;
#ifdef DEBUG_HANG_
if (myRank == 0)
cout << "\n(" << myRank << ") Send Bundles" << endl;
fflush(stdout);
#endif
#pragma omp parallel private(i) default(shared) num_threads(NUM_THREAD)
{
#pragma omp master
{
// Data structures for Bundled Messages:
#pragma omp task depend(inout \
: PCumulative, PMessageBundle, PSizeInfoMessages) depend(in \
: NumMessagesBundled, numProcs)
{
try {
PMessageBundle.reserve(NumMessagesBundled * 3); // Three integers per message
PCumulative.reserve(numProcs + 1); // Similar to Row Pointer vector in CSR data structure
PSizeInfoMessages.reserve(numProcs * 3); // Buffer to hold the Size info message packets
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesMessageBundling: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
PMessageBundle.resize(NumMessagesBundled * 3, -1); // Initialize
PCumulative.resize(numProcs + 1, 0); // Only initialize the counter variable
PSizeInfoMessages.resize(numProcs * 3, 0);
}
#pragma omp task depend(inout \
: PCumulative) depend(in \
: PCounter)
{
for (i = 0; i < numProcs; i++)
PCumulative[i + 1] = PCumulative[i] + PCounter[i];
}
#pragma omp task depend(inout \
: PCounter)
{
// Reuse PCounter to keep track of how many messages were inserted:
for (MilanInt i = 0; i < numProcs; i++) // Changed by Fabio to be an integer, addresses needs to be integers!
PCounter[i] = 0;
}
// Build the Message Bundle packet:
#pragma omp task depend(in \
: PCounter, QLocalVtx, QGhostVtx, QMsgType, QOwner, PMessageBundle, PCumulative) depend(out \
: myIndex, PMessageBundle, PCounter)
{
for (i = 0; i < NumMessagesBundled; i++) {
myIndex = (PCumulative[QOwner[i]] + PCounter[QOwner[i]]) * 3;
PMessageBundle[myIndex + 0] = QLocalVtx[i];
PMessageBundle[myIndex + 1] = QGhostVtx[i];
PMessageBundle[myIndex + 2] = QMsgType[i];
PCounter[QOwner[i]]++;
}
}
// Send the Bundled Messages: Use ISend
#pragma omp task depend(out \
: SRequest, SStatus)
{
try
{
SRequest.reserve(numProcs * 2); // At most two messages per processor
SStatus.reserve(numProcs * 2); // At most two messages per processor
}
catch (length_error)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesLinearSearchImmediateSend: \n";
cout << "Not enough memory to allocate the internal variables \n";
exit(1);
}
}
// Send the Messages
#pragma omp task depend(inout \
: SRequest, PSizeInfoMessages, PCumulative) depend(out \
: *msgActual, *msgInd)
{
for (i = 0; i < numProcs; i++) { // Changed by Fabio to be an integer, addresses needs to be integers!
if (i == myRank) // Do not send anything to yourself
continue;
// Send the Message with information about the size of next message:
// Build the Message Packet:
PSizeInfoMessages[i * 3 + 0] = (PCumulative[i + 1] - PCumulative[i]) * 3; // # of integers in the next message
PSizeInfoMessages[i * 3 + 1] = -1; // Dummy packet
PSizeInfoMessages[i * 3 + 2] = SIZEINFO; // TYPE
// Send a Request (Asynchronous)
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Sending bundled message to process " << i << " size: " << PSizeInfoMessages[i * 3 + 0] << endl;
fflush(stdout);
#endif
if (PSizeInfoMessages[i * 3 + 0] > 0)
{ // Send only if it is a nonempty packet
MPI_Isend(&PSizeInfoMessages[i * 3 + 0], 3, TypeMap<MilanLongInt>(), i, ComputeTag, comm,
&SRequest[(*msgInd)]);
(*msgActual)++;
(*msgInd)++;
// Now Send the message with the data packet:
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")SendiFFng Bundle to : " << i << endl;
for (k = (PCumulative[i] * 3); k < (PCumulative[i] * 3 + PSizeInfoMessages[i * 3 + 0]); k++)
cout << PMessageBundle[k] << ",";
cout << endl;
fflush(stdout);
#endif
MPI_Isend(&PMessageBundle[PCumulative[i] * 3], PSizeInfoMessages[i * 3 + 0],
TypeMap<MilanLongInt>(), i, BundleTag, comm, &SRequest[(*msgInd)]);
(*msgInd)++;
} // End of if size > 0
}
}
#pragma omp task depend(inout \
: PCumulative, QLocalVtx, QGhostVtx, QMsgType, QOwner)
{
// Free up temporary memory:
PCumulative.clear();
QLocalVtx.clear();
QGhostVtx.clear();
QMsgType.clear();
QOwner.clear();
}
#pragma omp task depend(inout : OneMessageSize, *BufferSize) depend(out : numMessagesToSend) depend(in : *numGhostEdges)
{
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Number of Ghost edges = " << *numGhostEdges;
cout << "\n(" << myRank << ")Total number of potential message X 2 = " << *numGhostEdges * 2;
cout << "\n(" << myRank << ")Number messages already sent in bundles = " << NumMessagesBundled;
if (*numGhostEdges > 0)
{
cout << "\n(" << myRank << ")Percentage of total = " << ((double)NumMessagesBundled / (double)(*numGhostEdges * 2)) * 100.0 << "% \n";
}
fflush(stdout);
#endif
// Allocate memory for MPI Send messages:
/* WILL COME BACK HERE - NO NEED TO STORE ALL THIS MEMORY !! */
OneMessageSize = 0;
MPI_Pack_size(3, TypeMap<MilanLongInt>(), comm, &OneMessageSize); // Size of one message packet
// How many messages to send?
// Potentially three kinds of messages will be sent/received:
// Request, Success, Failure.
// But only two will be sent from a given processor.
// Substract the number of messages that have already been sent as bundled messages:
numMessagesToSend = (*numGhostEdges) * 2 - NumMessagesBundled;
*BufferSize = (OneMessageSize + MPI_BSEND_OVERHEAD) * numMessagesToSend;
}
#pragma omp task depend(out : Buffer) depend(in : *BufferSize)
{
Buffer = 0;
#ifdef PRINT_DEBUG_INFO_
cout << "\n(" << myRank << ")Size of One Message from PACK= " << OneMessageSize;
cout << "\n(" << myRank << ")Size of Message overhead = " << MPI_BSEND_OVERHEAD;
cout << "\n(" << myRank << ")Number of Ghost edges = " << *numGhostEdges;
cout << "\n(" << myRank << ")Number of remaining message = " << numMessagesToSend;
cout << "\n(" << myRank << ")BufferSize = " << (*BufferSize);
cout << "\n(" << myRank << ")Attaching Buffer on.. ";
fflush(stdout);
#endif
if ((*BufferSize) > 0)
{
Buffer = (MilanLongInt *)malloc((*BufferSize)); // Allocate memory
if (Buffer == 0)
{
cout << "Error in function algoDistEdgeApproxDominatingEdgesLinearSearch: \n";
cout << "Not enough memory to allocate for send buffer on process " << myRank << "\n";
exit(1);
}
MPI_Buffer_attach(Buffer, *BufferSize); // Attach the Buffer
}
}
}
}
}
@@ -507,6 +507,71 @@ subroutine amg_c_hierarchy_bld(a,desc_a,prec,info)
return
contains
#if ( __GNUC__ == 13 && __GNUC_MINOR__ == 3)
! gfortran 13.3.0 generates a strange error here with MOLD
! moving to SOURCE but only for this version, since it's heavier
subroutine save_smoothers(level,save1, save2,info)
type(amg_c_onelev_type), intent(inout) :: level
class(amg_c_base_smoother_type), allocatable , intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(save1)) then
call save1%free(info)
if (info == 0) deallocate(save1,stat=info)
if (info /= 0) return
end if
if (allocated(save2)) then
call save2%free(info)
if (info == 0) deallocate(save2,stat=info)
if (info /= 0) return
end if
allocate(save1, source=level%sm,stat=info)
if (info == 0) call level%sm%clone_settings(save1,info)
if ((info == 0).and.allocated(level%sm2a)) then
allocate(save2, source=level%sm2a,stat=info)
if (info == 0) call level%sm2a%clone_settings(save2,info)
end if
return
end subroutine save_smoothers
subroutine restore_smoothers(level,save1, save2,info)
type(amg_c_onelev_type), intent(inout), target :: level
class(amg_c_base_smoother_type), allocatable, intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(level%sm)) then
if (info == 0) call level%sm%free(info)
if (info == 0) deallocate(level%sm,stat=info)
end if
if (allocated(save1)) then
if (info == 0) allocate(level%sm,source=save1,stat=info)
if (info == 0) call save1%clone_settings(level%sm,info)
end if
if (info /= 0) return
if (allocated(level%sm2a)) then
if (info == 0) call level%sm2a%free(info)
if (info == 0) deallocate(level%sm2a,stat=info)
end if
if (allocated(save2)) then
if (info == 0) allocate(level%sm2a,source=save2,stat=info)
if (info == 0) call save2%clone_settings(level%sm2a,info)
if (info == 0) level%sm2 => level%sm2a
else
if (allocated(level%sm)) level%sm2 => level%sm
end if
return
end subroutine restore_smoothers
#else
subroutine save_smoothers(level,save1, save2,info)
type(amg_c_onelev_type), intent(inout) :: level
class(amg_c_base_smoother_type), allocatable , intent(inout) :: save1, save2
@@ -565,5 +630,5 @@ contains
return
end subroutine restore_smoothers
#endif
end subroutine amg_c_hierarchy_bld
+2 -2
View File
@@ -186,8 +186,8 @@ subroutine amg_c_smoothers_bld(a,desc_a,prec,info,amold,vmold,imold)
& ' but it has been changed to distributed.'
end if
case(psb_ilu_n_, psb_ilu_t_,psb_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= psb_ilu_n_) then
case(amg_ilu_n_, amg_ilu_t_,amg_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= amg_ilu_n_) then
write(psb_err_unit,*) &
& 'AMG4PSBLAS: Warning: original coarse solver was requested as ',&
& amg_fact_names(coarse_solve_id)
+34 -18
View File
@@ -84,6 +84,7 @@ subroutine amg_ccprecseti(p,what,val,info,ilev,ilmax,pos,idx)
use amg_c_diag_solver
use amg_c_l1_diag_solver
use amg_c_ilu_solver
use amg_c_jac_solver
use amg_c_id_solver
use amg_c_gs_solver
use amg_c_ainv_solver
@@ -193,6 +194,15 @@ subroutine amg_ccprecseti(p,what,val,info,ilev,ilmax,pos,idx)
end if
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
case('COARSE_INVFILL')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
& ': Error: Inconsistent specification of WHAT vs. ILEV'
info = -2
return
end if
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
case('BJAC_ITRACE')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
@@ -242,6 +252,11 @@ subroutine amg_ccprecseti(p,what,val,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
end if
case('COARSE_INVFILL')
if (nlev_ > 1) then
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
end if
case('BJAC_ITRACE')
if (nlev_ > 1) then
call p%precv(nlev_)%set('SMOOTHER_ITRACE',val,info,pos=pos)
@@ -308,6 +323,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
use amg_c_diag_solver
use amg_c_l1_diag_solver
use amg_c_ilu_solver
use amg_c_jac_solver
use amg_c_id_solver
use amg_c_gs_solver
use amg_c_ainv_solver
@@ -469,7 +485,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -483,7 +499,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -506,7 +522,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -514,21 +530,21 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -543,7 +559,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -566,7 +582,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -590,7 +606,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -698,7 +714,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -711,7 +727,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -734,7 +750,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -742,21 +758,21 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -771,7 +787,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -794,7 +810,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -818,7 +834,7 @@ subroutine amg_ccprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
+1 -1
View File
@@ -170,7 +170,7 @@ subroutine amg_cfile_prec_descr(prec,info,iout,root, verbosity,prefix)
call prec%precv(1)%sm%descr(info,iout=iout_,prefix=prefix)
nswps = prec%precv(1)%parms%sweeps_pre
end if
write(iout_,*) trim(prefix_), ' Number of sweeps : ',nswps
write(iout_,*) trim(prefix_), ' Number of sweeps/degree : ',nswps
write(iout_,*) trim(prefix_)
else if (nlev > 1) then
+152
View File
@@ -0,0 +1,152 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
! File: amg_dfile_prec_memory_use.f90
!
!
! Subroutine: amg_file_prec_memory_use
! Version: complex
!
! This routine prints a memory_useiption of the preconditioner to the standard
! output or to a file. It must be called after the preconditioner has been
! built by amg_precbld.
!
! Arguments:
! p - type(amg_Tprec_type), input.
! The preconditioner data structure to be printed out.
! info - integer, output.
! error code.
! iout - integer, input, optional.
! The id of the file where the preconditioner description
! will be printed. If iout is not present, then the standard
! output is condidered.
! root - integer, input, optional.
! The id of the process printing the message; -1 acts as a wildcard.
! Default is psb_root_
!
!
!
! verbosity:
! <0: suppress all messages
! 0: normal
! >1: increased details
!
subroutine amg_cfile_prec_memory_use(prec,info,iout,root, verbosity,prefix,global)
use psb_base_mod
use amg_c_prec_mod, amg_protect_name => amg_cfile_prec_memory_use
use amg_c_inner_mod
use amg_c_gs_solver
implicit none
! Arguments
class(amg_cprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
! Local variables
integer(psb_ipk_) :: ilev, nlev, ilmin, nswps
type(psb_ctxt_type) :: ctxt
integer(psb_ipk_) :: me, np
logical :: is_symgs
character(len=20), parameter :: name='amg_file_prec_memory_use'
integer(psb_ipk_) :: iout_, root_, verbosity_
logical :: global_
character(1024) :: prefix_
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
end if
if (iout_ < 0) iout_ = psb_out_unit
if (present(verbosity)) then
verbosity_ = verbosity
else
verbosity_ = 0
end if
if (verbosity_ < 0) goto 9998
ctxt = prec%ctxt
call psb_info(ctxt,me,np)
prefix_ = ""
if (verbosity_ == 0) then
if (present(prefix)) then
prefix_ = prefix
end if
else if (verbosity_ > 0) then
if (present(prefix)) then
write(prefix_,'(a,a,i5,a)') prefix,' from process ',me,': '
else
write(prefix_,'(a,i5,a)') 'Process ',me,': '
end if
end if
if (present(root)) then
root_ = root
else
root_ = psb_root_
end if
if (root_ == -1) root_ = me
if (allocated(prec%precv)) then
if (verbosity_ >=0) then
if (me == root_) then
write(iout_,*)
write(iout_,'(a,1x,a)') trim(prefix_),'Preconditioner memory usage'
end if
nlev = size(prec%precv)
do ilev=1,nlev
call prec%precv(ilev)%memory_use(ilev,nlev,ilmin,info, &
& iout=iout_,verbosity=verbosity_,prefix=trim(prefix_),global=global)
end do
else
write(iout_,*) trim(name), &
& ': Error: no base preconditioner available, something is wrong!'
info = -2
return
endif
end if
9998 continue
end subroutine amg_cfile_prec_memory_use
+1 -1
View File
@@ -98,6 +98,7 @@ subroutine amg_cprecinit(ctxt,prec,ptype,info)
use amg_c_diag_solver
use amg_c_ilu_solver
use amg_c_gs_solver
#if defined(HAVE_SLU_)
use amg_c_slu_solver
#endif
@@ -152,7 +153,6 @@ subroutine amg_cprecinit(ctxt,prec,ptype,info)
if (info /= psb_success_) return
allocate(amg_c_diag_solver_type :: prec%precv(ilev_)%sm%sv, stat=info)
call prec%precv(ilev_)%default()
case ('L1-DIAG','L1-JACOBI','L1_DIAG','L1_JACOBI')
nlev_ = 1
ilev_ = 1
@@ -507,6 +507,71 @@ subroutine amg_d_hierarchy_bld(a,desc_a,prec,info)
return
contains
#if ( __GNUC__ == 13 && __GNUC_MINOR__ == 3)
! gfortran 13.3.0 generates a strange error here with MOLD
! moving to SOURCE but only for this version, since it's heavier
subroutine save_smoothers(level,save1, save2,info)
type(amg_d_onelev_type), intent(inout) :: level
class(amg_d_base_smoother_type), allocatable , intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(save1)) then
call save1%free(info)
if (info == 0) deallocate(save1,stat=info)
if (info /= 0) return
end if
if (allocated(save2)) then
call save2%free(info)
if (info == 0) deallocate(save2,stat=info)
if (info /= 0) return
end if
allocate(save1, source=level%sm,stat=info)
if (info == 0) call level%sm%clone_settings(save1,info)
if ((info == 0).and.allocated(level%sm2a)) then
allocate(save2, source=level%sm2a,stat=info)
if (info == 0) call level%sm2a%clone_settings(save2,info)
end if
return
end subroutine save_smoothers
subroutine restore_smoothers(level,save1, save2,info)
type(amg_d_onelev_type), intent(inout), target :: level
class(amg_d_base_smoother_type), allocatable, intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(level%sm)) then
if (info == 0) call level%sm%free(info)
if (info == 0) deallocate(level%sm,stat=info)
end if
if (allocated(save1)) then
if (info == 0) allocate(level%sm,source=save1,stat=info)
if (info == 0) call save1%clone_settings(level%sm,info)
end if
if (info /= 0) return
if (allocated(level%sm2a)) then
if (info == 0) call level%sm2a%free(info)
if (info == 0) deallocate(level%sm2a,stat=info)
end if
if (allocated(save2)) then
if (info == 0) allocate(level%sm2a,source=save2,stat=info)
if (info == 0) call save2%clone_settings(level%sm2a,info)
if (info == 0) level%sm2 => level%sm2a
else
if (allocated(level%sm)) level%sm2 => level%sm
end if
return
end subroutine restore_smoothers
#else
subroutine save_smoothers(level,save1, save2,info)
type(amg_d_onelev_type), intent(inout) :: level
class(amg_d_base_smoother_type), allocatable , intent(inout) :: save1, save2
@@ -565,5 +630,5 @@ contains
return
end subroutine restore_smoothers
#endif
end subroutine amg_d_hierarchy_bld
+2 -2
View File
@@ -186,8 +186,8 @@ subroutine amg_d_smoothers_bld(a,desc_a,prec,info,amold,vmold,imold)
& ' but it has been changed to distributed.'
end if
case(psb_ilu_n_, psb_ilu_t_,psb_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= psb_ilu_n_) then
case(amg_ilu_n_, amg_ilu_t_,amg_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= amg_ilu_n_) then
write(psb_err_unit,*) &
& 'AMG4PSBLAS: Warning: original coarse solver was requested as ',&
& amg_fact_names(coarse_solve_id)
+34 -18
View File
@@ -84,6 +84,7 @@ subroutine amg_dcprecseti(p,what,val,info,ilev,ilmax,pos,idx)
use amg_d_diag_solver
use amg_d_l1_diag_solver
use amg_d_ilu_solver
use amg_d_jac_solver
use amg_d_id_solver
use amg_d_gs_solver
use amg_d_ainv_solver
@@ -199,6 +200,15 @@ subroutine amg_dcprecseti(p,what,val,info,ilev,ilmax,pos,idx)
end if
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
case('COARSE_INVFILL')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
& ': Error: Inconsistent specification of WHAT vs. ILEV'
info = -2
return
end if
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
case('BJAC_ITRACE')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
@@ -248,6 +258,11 @@ subroutine amg_dcprecseti(p,what,val,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
end if
case('COARSE_INVFILL')
if (nlev_ > 1) then
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
end if
case('BJAC_ITRACE')
if (nlev_ > 1) then
call p%precv(nlev_)%set('SMOOTHER_ITRACE',val,info,pos=pos)
@@ -314,6 +329,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
use amg_d_diag_solver
use amg_d_l1_diag_solver
use amg_d_ilu_solver
use amg_d_jac_solver
use amg_d_id_solver
use amg_d_gs_solver
use amg_d_ainv_solver
@@ -483,7 +499,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -499,7 +515,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -522,7 +538,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -530,21 +546,21 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -559,7 +575,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -589,7 +605,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -627,7 +643,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -737,7 +753,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -752,7 +768,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -775,7 +791,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -783,21 +799,21 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -812,7 +828,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -842,7 +858,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -880,7 +896,7 @@ subroutine amg_dcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
+1 -1
View File
@@ -170,7 +170,7 @@ subroutine amg_dfile_prec_descr(prec,info,iout,root, verbosity,prefix)
call prec%precv(1)%sm%descr(info,iout=iout_,prefix=prefix)
nswps = prec%precv(1)%parms%sweeps_pre
end if
write(iout_,*) trim(prefix_), ' Number of sweeps : ',nswps
write(iout_,*) trim(prefix_), ' Number of sweeps/degree : ',nswps
write(iout_,*) trim(prefix_)
else if (nlev > 1) then
+152
View File
@@ -0,0 +1,152 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
! File: amg_dfile_prec_memory_use.f90
!
!
! Subroutine: amg_file_prec_memory_use
! Version: real
!
! This routine prints a memory_useiption of the preconditioner to the standard
! output or to a file. It must be called after the preconditioner has been
! built by amg_precbld.
!
! Arguments:
! p - type(amg_Tprec_type), input.
! The preconditioner data structure to be printed out.
! info - integer, output.
! error code.
! iout - integer, input, optional.
! The id of the file where the preconditioner description
! will be printed. If iout is not present, then the standard
! output is condidered.
! root - integer, input, optional.
! The id of the process printing the message; -1 acts as a wildcard.
! Default is psb_root_
!
!
!
! verbosity:
! <0: suppress all messages
! 0: normal
! >1: increased details
!
subroutine amg_dfile_prec_memory_use(prec,info,iout,root, verbosity,prefix,global)
use psb_base_mod
use amg_d_prec_mod, amg_protect_name => amg_dfile_prec_memory_use
use amg_d_inner_mod
use amg_d_gs_solver
implicit none
! Arguments
class(amg_dprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
! Local variables
integer(psb_ipk_) :: ilev, nlev, ilmin, nswps
type(psb_ctxt_type) :: ctxt
integer(psb_ipk_) :: me, np
logical :: is_symgs
character(len=20), parameter :: name='amg_file_prec_memory_use'
integer(psb_ipk_) :: iout_, root_, verbosity_
logical :: global_
character(1024) :: prefix_
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
end if
if (iout_ < 0) iout_ = psb_out_unit
if (present(verbosity)) then
verbosity_ = verbosity
else
verbosity_ = 0
end if
if (verbosity_ < 0) goto 9998
ctxt = prec%ctxt
call psb_info(ctxt,me,np)
prefix_ = ""
if (verbosity_ == 0) then
if (present(prefix)) then
prefix_ = prefix
end if
else if (verbosity_ > 0) then
if (present(prefix)) then
write(prefix_,'(a,a,i5,a)') prefix,' from process ',me,': '
else
write(prefix_,'(a,i5,a)') 'Process ',me,': '
end if
end if
if (present(root)) then
root_ = root
else
root_ = psb_root_
end if
if (root_ == -1) root_ = me
if (allocated(prec%precv)) then
if (verbosity_ >=0) then
if (me == root_) then
write(iout_,*)
write(iout_,'(a,1x,a)') trim(prefix_),'Preconditioner memory usage'
end if
nlev = size(prec%precv)
do ilev=1,nlev
call prec%precv(ilev)%memory_use(ilev,nlev,ilmin,info, &
& iout=iout_,verbosity=verbosity_,prefix=trim(prefix_),global=global)
end do
else
write(iout_,*) trim(name), &
& ': Error: no base preconditioner available, something is wrong!'
info = -2
return
endif
end if
9998 continue
end subroutine amg_dfile_prec_memory_use
+10 -1
View File
@@ -98,6 +98,8 @@ subroutine amg_dprecinit(ctxt,prec,ptype,info)
use amg_d_diag_solver
use amg_d_ilu_solver
use amg_d_gs_solver
use amg_d_poly_smoother
#if defined(HAVE_UMF_)
use amg_d_umf_solver
#endif
@@ -155,7 +157,14 @@ subroutine amg_dprecinit(ctxt,prec,ptype,info)
if (info /= psb_success_) return
allocate(amg_d_diag_solver_type :: prec%precv(ilev_)%sm%sv, stat=info)
call prec%precv(ilev_)%default()
case ('POLY')
nlev_ = 1
ilev_ = 1
allocate(prec%precv(nlev_),stat=info)
allocate(amg_d_poly_smoother_type :: prec%precv(ilev_)%sm, stat=info)
if (info /= psb_success_) return
allocate(amg_d_l1_diag_solver_type :: prec%precv(ilev_)%sm%sv, stat=info)
call prec%precv(ilev_)%default()
case ('L1-DIAG','L1-JACOBI','L1_DIAG','L1_JACOBI')
nlev_ = 1
ilev_ = 1
@@ -507,6 +507,71 @@ subroutine amg_s_hierarchy_bld(a,desc_a,prec,info)
return
contains
#if ( __GNUC__ == 13 && __GNUC_MINOR__ == 3)
! gfortran 13.3.0 generates a strange error here with MOLD
! moving to SOURCE but only for this version, since it's heavier
subroutine save_smoothers(level,save1, save2,info)
type(amg_s_onelev_type), intent(inout) :: level
class(amg_s_base_smoother_type), allocatable , intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(save1)) then
call save1%free(info)
if (info == 0) deallocate(save1,stat=info)
if (info /= 0) return
end if
if (allocated(save2)) then
call save2%free(info)
if (info == 0) deallocate(save2,stat=info)
if (info /= 0) return
end if
allocate(save1, source=level%sm,stat=info)
if (info == 0) call level%sm%clone_settings(save1,info)
if ((info == 0).and.allocated(level%sm2a)) then
allocate(save2, source=level%sm2a,stat=info)
if (info == 0) call level%sm2a%clone_settings(save2,info)
end if
return
end subroutine save_smoothers
subroutine restore_smoothers(level,save1, save2,info)
type(amg_s_onelev_type), intent(inout), target :: level
class(amg_s_base_smoother_type), allocatable, intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(level%sm)) then
if (info == 0) call level%sm%free(info)
if (info == 0) deallocate(level%sm,stat=info)
end if
if (allocated(save1)) then
if (info == 0) allocate(level%sm,source=save1,stat=info)
if (info == 0) call save1%clone_settings(level%sm,info)
end if
if (info /= 0) return
if (allocated(level%sm2a)) then
if (info == 0) call level%sm2a%free(info)
if (info == 0) deallocate(level%sm2a,stat=info)
end if
if (allocated(save2)) then
if (info == 0) allocate(level%sm2a,source=save2,stat=info)
if (info == 0) call save2%clone_settings(level%sm2a,info)
if (info == 0) level%sm2 => level%sm2a
else
if (allocated(level%sm)) level%sm2 => level%sm
end if
return
end subroutine restore_smoothers
#else
subroutine save_smoothers(level,save1, save2,info)
type(amg_s_onelev_type), intent(inout) :: level
class(amg_s_base_smoother_type), allocatable , intent(inout) :: save1, save2
@@ -565,5 +630,5 @@ contains
return
end subroutine restore_smoothers
#endif
end subroutine amg_s_hierarchy_bld
+2 -2
View File
@@ -186,8 +186,8 @@ subroutine amg_s_smoothers_bld(a,desc_a,prec,info,amold,vmold,imold)
& ' but it has been changed to distributed.'
end if
case(psb_ilu_n_, psb_ilu_t_,psb_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= psb_ilu_n_) then
case(amg_ilu_n_, amg_ilu_t_,amg_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= amg_ilu_n_) then
write(psb_err_unit,*) &
& 'AMG4PSBLAS: Warning: original coarse solver was requested as ',&
& amg_fact_names(coarse_solve_id)
+34 -18
View File
@@ -84,6 +84,7 @@ subroutine amg_scprecseti(p,what,val,info,ilev,ilmax,pos,idx)
use amg_s_diag_solver
use amg_s_l1_diag_solver
use amg_s_ilu_solver
use amg_s_jac_solver
use amg_s_id_solver
use amg_s_gs_solver
use amg_s_ainv_solver
@@ -193,6 +194,15 @@ subroutine amg_scprecseti(p,what,val,info,ilev,ilmax,pos,idx)
end if
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
case('COARSE_INVFILL')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
& ': Error: Inconsistent specification of WHAT vs. ILEV'
info = -2
return
end if
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
case('BJAC_ITRACE')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
@@ -242,6 +252,11 @@ subroutine amg_scprecseti(p,what,val,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
end if
case('COARSE_INVFILL')
if (nlev_ > 1) then
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
end if
case('BJAC_ITRACE')
if (nlev_ > 1) then
call p%precv(nlev_)%set('SMOOTHER_ITRACE',val,info,pos=pos)
@@ -308,6 +323,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
use amg_s_diag_solver
use amg_s_l1_diag_solver
use amg_s_ilu_solver
use amg_s_jac_solver
use amg_s_id_solver
use amg_s_gs_solver
use amg_s_ainv_solver
@@ -469,7 +485,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -483,7 +499,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -506,7 +522,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -514,21 +530,21 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -543,7 +559,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -566,7 +582,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -590,7 +606,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -698,7 +714,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -711,7 +727,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -734,7 +750,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -742,21 +758,21 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -771,7 +787,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -794,7 +810,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -818,7 +834,7 @@ subroutine amg_scprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
+1 -1
View File
@@ -170,7 +170,7 @@ subroutine amg_sfile_prec_descr(prec,info,iout,root, verbosity,prefix)
call prec%precv(1)%sm%descr(info,iout=iout_,prefix=prefix)
nswps = prec%precv(1)%parms%sweeps_pre
end if
write(iout_,*) trim(prefix_), ' Number of sweeps : ',nswps
write(iout_,*) trim(prefix_), ' Number of sweeps/degree : ',nswps
write(iout_,*) trim(prefix_)
else if (nlev > 1) then
+152
View File
@@ -0,0 +1,152 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
! File: amg_dfile_prec_memory_use.f90
!
!
! Subroutine: amg_file_prec_memory_use
! Version: real
!
! This routine prints a memory_useiption of the preconditioner to the standard
! output or to a file. It must be called after the preconditioner has been
! built by amg_precbld.
!
! Arguments:
! p - type(amg_Tprec_type), input.
! The preconditioner data structure to be printed out.
! info - integer, output.
! error code.
! iout - integer, input, optional.
! The id of the file where the preconditioner description
! will be printed. If iout is not present, then the standard
! output is condidered.
! root - integer, input, optional.
! The id of the process printing the message; -1 acts as a wildcard.
! Default is psb_root_
!
!
!
! verbosity:
! <0: suppress all messages
! 0: normal
! >1: increased details
!
subroutine amg_sfile_prec_memory_use(prec,info,iout,root, verbosity,prefix,global)
use psb_base_mod
use amg_s_prec_mod, amg_protect_name => amg_sfile_prec_memory_use
use amg_s_inner_mod
use amg_s_gs_solver
implicit none
! Arguments
class(amg_sprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
! Local variables
integer(psb_ipk_) :: ilev, nlev, ilmin, nswps
type(psb_ctxt_type) :: ctxt
integer(psb_ipk_) :: me, np
logical :: is_symgs
character(len=20), parameter :: name='amg_file_prec_memory_use'
integer(psb_ipk_) :: iout_, root_, verbosity_
logical :: global_
character(1024) :: prefix_
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
end if
if (iout_ < 0) iout_ = psb_out_unit
if (present(verbosity)) then
verbosity_ = verbosity
else
verbosity_ = 0
end if
if (verbosity_ < 0) goto 9998
ctxt = prec%ctxt
call psb_info(ctxt,me,np)
prefix_ = ""
if (verbosity_ == 0) then
if (present(prefix)) then
prefix_ = prefix
end if
else if (verbosity_ > 0) then
if (present(prefix)) then
write(prefix_,'(a,a,i5,a)') prefix,' from process ',me,': '
else
write(prefix_,'(a,i5,a)') 'Process ',me,': '
end if
end if
if (present(root)) then
root_ = root
else
root_ = psb_root_
end if
if (root_ == -1) root_ = me
if (allocated(prec%precv)) then
if (verbosity_ >=0) then
if (me == root_) then
write(iout_,*)
write(iout_,'(a,1x,a)') trim(prefix_),'Preconditioner memory usage'
end if
nlev = size(prec%precv)
do ilev=1,nlev
call prec%precv(ilev)%memory_use(ilev,nlev,ilmin,info, &
& iout=iout_,verbosity=verbosity_,prefix=trim(prefix_),global=global)
end do
else
write(iout_,*) trim(name), &
& ': Error: no base preconditioner available, something is wrong!'
info = -2
return
endif
end if
9998 continue
end subroutine amg_sfile_prec_memory_use
+10 -1
View File
@@ -98,6 +98,8 @@ subroutine amg_sprecinit(ctxt,prec,ptype,info)
use amg_s_diag_solver
use amg_s_ilu_solver
use amg_s_gs_solver
use amg_s_poly_smoother
#if defined(HAVE_SLU_)
use amg_s_slu_solver
#endif
@@ -152,7 +154,14 @@ subroutine amg_sprecinit(ctxt,prec,ptype,info)
if (info /= psb_success_) return
allocate(amg_s_diag_solver_type :: prec%precv(ilev_)%sm%sv, stat=info)
call prec%precv(ilev_)%default()
case ('POLY')
nlev_ = 1
ilev_ = 1
allocate(prec%precv(nlev_),stat=info)
allocate(amg_s_poly_smoother_type :: prec%precv(ilev_)%sm, stat=info)
if (info /= psb_success_) return
allocate(amg_s_l1_diag_solver_type :: prec%precv(ilev_)%sm%sv, stat=info)
call prec%precv(ilev_)%default()
case ('L1-DIAG','L1-JACOBI','L1_DIAG','L1_JACOBI')
nlev_ = 1
ilev_ = 1
@@ -507,6 +507,71 @@ subroutine amg_z_hierarchy_bld(a,desc_a,prec,info)
return
contains
#if ( __GNUC__ == 13 && __GNUC_MINOR__ == 3)
! gfortran 13.3.0 generates a strange error here with MOLD
! moving to SOURCE but only for this version, since it's heavier
subroutine save_smoothers(level,save1, save2,info)
type(amg_z_onelev_type), intent(inout) :: level
class(amg_z_base_smoother_type), allocatable , intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(save1)) then
call save1%free(info)
if (info == 0) deallocate(save1,stat=info)
if (info /= 0) return
end if
if (allocated(save2)) then
call save2%free(info)
if (info == 0) deallocate(save2,stat=info)
if (info /= 0) return
end if
allocate(save1, source=level%sm,stat=info)
if (info == 0) call level%sm%clone_settings(save1,info)
if ((info == 0).and.allocated(level%sm2a)) then
allocate(save2, source=level%sm2a,stat=info)
if (info == 0) call level%sm2a%clone_settings(save2,info)
end if
return
end subroutine save_smoothers
subroutine restore_smoothers(level,save1, save2,info)
type(amg_z_onelev_type), intent(inout), target :: level
class(amg_z_base_smoother_type), allocatable, intent(inout) :: save1, save2
integer(psb_ipk_), intent(out) :: info
info = 0
if (allocated(level%sm)) then
if (info == 0) call level%sm%free(info)
if (info == 0) deallocate(level%sm,stat=info)
end if
if (allocated(save1)) then
if (info == 0) allocate(level%sm,source=save1,stat=info)
if (info == 0) call save1%clone_settings(level%sm,info)
end if
if (info /= 0) return
if (allocated(level%sm2a)) then
if (info == 0) call level%sm2a%free(info)
if (info == 0) deallocate(level%sm2a,stat=info)
end if
if (allocated(save2)) then
if (info == 0) allocate(level%sm2a,source=save2,stat=info)
if (info == 0) call save2%clone_settings(level%sm2a,info)
if (info == 0) level%sm2 => level%sm2a
else
if (allocated(level%sm)) level%sm2 => level%sm
end if
return
end subroutine restore_smoothers
#else
subroutine save_smoothers(level,save1, save2,info)
type(amg_z_onelev_type), intent(inout) :: level
class(amg_z_base_smoother_type), allocatable , intent(inout) :: save1, save2
@@ -565,5 +630,5 @@ contains
return
end subroutine restore_smoothers
#endif
end subroutine amg_z_hierarchy_bld
+2 -2
View File
@@ -186,8 +186,8 @@ subroutine amg_z_smoothers_bld(a,desc_a,prec,info,amold,vmold,imold)
& ' but it has been changed to distributed.'
end if
case(psb_ilu_n_, psb_ilu_t_,psb_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= psb_ilu_n_) then
case(amg_ilu_n_, amg_ilu_t_,amg_milu_n_)
if (prec%precv(iszv)%sm%sv%get_id() /= amg_ilu_n_) then
write(psb_err_unit,*) &
& 'AMG4PSBLAS: Warning: original coarse solver was requested as ',&
& amg_fact_names(coarse_solve_id)
+34 -18
View File
@@ -84,6 +84,7 @@ subroutine amg_zcprecseti(p,what,val,info,ilev,ilmax,pos,idx)
use amg_z_diag_solver
use amg_z_l1_diag_solver
use amg_z_ilu_solver
use amg_z_jac_solver
use amg_z_id_solver
use amg_z_gs_solver
use amg_z_ainv_solver
@@ -199,6 +200,15 @@ subroutine amg_zcprecseti(p,what,val,info,ilev,ilmax,pos,idx)
end if
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
case('COARSE_INVFILL')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
& ': Error: Inconsistent specification of WHAT vs. ILEV'
info = -2
return
end if
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
case('BJAC_ITRACE')
if (ilev_ /= nlev_) then
write(psb_err_unit,*) name,&
@@ -248,6 +258,11 @@ subroutine amg_zcprecseti(p,what,val,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('SUB_FILLIN',val,info,pos=pos)
end if
case('COARSE_INVFILL')
if (nlev_ > 1) then
call p%precv(nlev_)%set('INV_FILLIN',val,info,pos=pos)
end if
case('BJAC_ITRACE')
if (nlev_ > 1) then
call p%precv(nlev_)%set('SMOOTHER_ITRACE',val,info,pos=pos)
@@ -314,6 +329,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
use amg_z_diag_solver
use amg_z_l1_diag_solver
use amg_z_ilu_solver
use amg_z_jac_solver
use amg_z_id_solver
use amg_z_gs_solver
use amg_z_ainv_solver
@@ -483,7 +499,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -499,7 +515,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -522,7 +538,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -530,21 +546,21 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -559,7 +575,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -589,7 +605,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -627,7 +643,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -737,7 +753,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -752,7 +768,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#elif defined(HAVE_MUMPS_)
call p%precv(nlev_)%set('SUB_SOLVE',amg_mumps_,info,pos=pos)
#else
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
#endif
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
@@ -775,7 +791,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -783,21 +799,21 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
#endif
case('ILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('ILUT')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_t_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_t_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
call p%precv(nlev_)%set('COARSE_MAT',amg_repl_mat_,info,pos=pos)
case('MILU')
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_milu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_milu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_repl_mat_)
@@ -812,7 +828,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -842,7 +858,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
@@ -880,7 +896,7 @@ subroutine amg_zcprecsetc(p,what,string,info,ilev,ilmax,pos,idx)
call p%precv(nlev_)%set('COARSE_MAT',amg_distr_mat_,info,pos=pos)
#else
call p%precv(nlev_)%set('SMOOTHER_TYPE',amg_bjac_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',psb_ilu_n_,info,pos=pos)
call p%precv(nlev_)%set('SUB_SOLVE',amg_ilu_n_,info,pos=pos)
if (hier_asb) &
& call amg_warn_coarse_mat(p%precv(nlev_)%parms%get_coarse_mat(),&
& amg_distr_mat_)
+1 -1
View File
@@ -170,7 +170,7 @@ subroutine amg_zfile_prec_descr(prec,info,iout,root, verbosity,prefix)
call prec%precv(1)%sm%descr(info,iout=iout_,prefix=prefix)
nswps = prec%precv(1)%parms%sweeps_pre
end if
write(iout_,*) trim(prefix_), ' Number of sweeps : ',nswps
write(iout_,*) trim(prefix_), ' Number of sweeps/degree : ',nswps
write(iout_,*) trim(prefix_)
else if (nlev > 1) then
+152
View File
@@ -0,0 +1,152 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
! File: amg_dfile_prec_memory_use.f90
!
!
! Subroutine: amg_file_prec_memory_use
! Version: complex
!
! This routine prints a memory_useiption of the preconditioner to the standard
! output or to a file. It must be called after the preconditioner has been
! built by amg_precbld.
!
! Arguments:
! p - type(amg_Tprec_type), input.
! The preconditioner data structure to be printed out.
! info - integer, output.
! error code.
! iout - integer, input, optional.
! The id of the file where the preconditioner description
! will be printed. If iout is not present, then the standard
! output is condidered.
! root - integer, input, optional.
! The id of the process printing the message; -1 acts as a wildcard.
! Default is psb_root_
!
!
!
! verbosity:
! <0: suppress all messages
! 0: normal
! >1: increased details
!
subroutine amg_zfile_prec_memory_use(prec,info,iout,root, verbosity,prefix,global)
use psb_base_mod
use amg_z_prec_mod, amg_protect_name => amg_zfile_prec_memory_use
use amg_z_inner_mod
use amg_z_gs_solver
implicit none
! Arguments
class(amg_zprec_type), intent(in) :: prec
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
integer(psb_ipk_), intent(in), optional :: root
integer(psb_ipk_), intent(in), optional :: verbosity
character(len=*), intent(in), optional :: prefix
logical, intent(in), optional :: global
! Local variables
integer(psb_ipk_) :: ilev, nlev, ilmin, nswps
type(psb_ctxt_type) :: ctxt
integer(psb_ipk_) :: me, np
logical :: is_symgs
character(len=20), parameter :: name='amg_file_prec_memory_use'
integer(psb_ipk_) :: iout_, root_, verbosity_
logical :: global_
character(1024) :: prefix_
info = psb_success_
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
end if
if (iout_ < 0) iout_ = psb_out_unit
if (present(verbosity)) then
verbosity_ = verbosity
else
verbosity_ = 0
end if
if (verbosity_ < 0) goto 9998
ctxt = prec%ctxt
call psb_info(ctxt,me,np)
prefix_ = ""
if (verbosity_ == 0) then
if (present(prefix)) then
prefix_ = prefix
end if
else if (verbosity_ > 0) then
if (present(prefix)) then
write(prefix_,'(a,a,i5,a)') prefix,' from process ',me,': '
else
write(prefix_,'(a,i5,a)') 'Process ',me,': '
end if
end if
if (present(root)) then
root_ = root
else
root_ = psb_root_
end if
if (root_ == -1) root_ = me
if (allocated(prec%precv)) then
if (verbosity_ >=0) then
if (me == root_) then
write(iout_,*)
write(iout_,'(a,1x,a)') trim(prefix_),'Preconditioner memory usage'
end if
nlev = size(prec%precv)
do ilev=1,nlev
call prec%precv(ilev)%memory_use(ilev,nlev,ilmin,info, &
& iout=iout_,verbosity=verbosity_,prefix=trim(prefix_),global=global)
end do
else
write(iout_,*) trim(name), &
& ': Error: no base preconditioner available, something is wrong!'
info = -2
return
endif
end if
9998 continue
end subroutine amg_zfile_prec_memory_use
+1 -1
View File
@@ -98,6 +98,7 @@ subroutine amg_zprecinit(ctxt,prec,ptype,info)
use amg_z_diag_solver
use amg_z_ilu_solver
use amg_z_gs_solver
#if defined(HAVE_UMF_)
use amg_z_umf_solver
#endif
@@ -155,7 +156,6 @@ subroutine amg_zprecinit(ctxt,prec,ptype,info)
if (info /= psb_success_) return
allocate(amg_z_diag_solver_type :: prec%precv(ilev_)%sm%sv, stat=info)
call prec%precv(ilev_)%default()
case ('L1-DIAG','L1-JACOBI','L1_DIAG','L1_JACOBI')
nlev_ = 1
ilev_ = 1
+4
View File
@@ -15,6 +15,7 @@ amg_c_base_onelev_csetc.o \
amg_c_base_onelev_cseti.o \
amg_c_base_onelev_csetr.o \
amg_c_base_onelev_descr.o \
amg_c_base_onelev_memory_use.o \
amg_c_base_onelev_dump.o \
amg_c_base_onelev_free.o \
amg_c_base_onelev_free_smoothers.o \
@@ -31,6 +32,7 @@ amg_d_base_onelev_csetc.o \
amg_d_base_onelev_cseti.o \
amg_d_base_onelev_csetr.o \
amg_d_base_onelev_descr.o \
amg_d_base_onelev_memory_use.o \
amg_d_base_onelev_dump.o \
amg_d_base_onelev_free.o \
amg_d_base_onelev_free_smoothers.o \
@@ -47,6 +49,7 @@ amg_s_base_onelev_csetc.o \
amg_s_base_onelev_cseti.o \
amg_s_base_onelev_csetr.o \
amg_s_base_onelev_descr.o \
amg_s_base_onelev_memory_use.o \
amg_s_base_onelev_dump.o \
amg_s_base_onelev_free.o \
amg_s_base_onelev_free_smoothers.o \
@@ -63,6 +66,7 @@ amg_z_base_onelev_csetc.o \
amg_z_base_onelev_cseti.o \
amg_z_base_onelev_csetr.o \
amg_z_base_onelev_descr.o \
amg_z_base_onelev_memory_use.o \
amg_z_base_onelev_dump.o \
amg_z_base_onelev_free.o \
amg_z_base_onelev_free_smoothers.o \
@@ -46,6 +46,7 @@ subroutine amg_c_base_onelev_csetc(lv,what,val,info,pos,idx)
use amg_c_as_smoother
use amg_c_diag_solver
use amg_c_l1_diag_solver
use amg_c_jac_solver
use amg_c_ilu_solver
use amg_c_id_solver
use amg_c_gs_solver
@@ -78,6 +79,8 @@ subroutine amg_c_base_onelev_csetc(lv,what,val,info,pos,idx)
type(amg_c_as_smoother_type) :: amg_c_as_smoother_mold
type(amg_c_diag_solver_type) :: amg_c_diag_solver_mold
type(amg_c_l1_diag_solver_type) :: amg_c_l1_diag_solver_mold
type(amg_c_jac_solver_type) :: amg_c_jac_solver_mold
type(amg_c_l1_jac_solver_type) :: amg_c_l1_jac_solver_mold
type(amg_c_ilu_solver_type) :: amg_c_ilu_solver_mold
type(amg_c_id_solver_type) :: amg_c_id_solver_mold
type(amg_c_gs_solver_type) :: amg_c_gs_solver_mold
@@ -185,10 +188,10 @@ subroutine amg_c_base_onelev_csetc(lv,what,val,info,pos,idx)
case ('NONE','NOPREC','FACT_NONE')
call lv%set(amg_c_id_solver_mold,info,pos=pos)
case ('DIAG')
case ('DIAG','JACOBI')
call lv%set(amg_c_diag_solver_mold,info,pos=pos)
case ('L1-DIAG')
case ('L1-DIAG','L1-JACOBI')
call lv%set(amg_c_l1_diag_solver_mold,info,pos=pos)
case ('GS','FGS','FWGS')
@@ -243,7 +246,7 @@ subroutine amg_c_base_onelev_csetc(lv,what,val,info,pos,idx)
end if
select case(val)
case('DEC')
case('DEC','DECOUPLED')
allocate(amg_c_dec_aggregator_type :: lv%aggr, stat=info)
case('SYMDEC')
allocate(amg_c_symdec_aggregator_type :: lv%aggr, stat=info)
@@ -164,7 +164,7 @@ subroutine amg_c_base_onelev_cseti(lv,what,val,info,pos,idx)
case (amg_bwgs_)
call lv%set(amg_c_bwgs_solver_mold,info,pos=pos)
case (psb_ilu_n_,psb_milu_n_,psb_ilu_t_)
case (amg_ilu_n_,amg_milu_n_,amg_ilu_t_)
call lv%set(amg_c_ilu_solver_mold,info,pos=pos)
if (info == 0) then
if ((ipos_==amg_smooth_pre_) .or.(ipos_==amg_smooth_both_)) then
@@ -109,6 +109,8 @@ subroutine amg_c_base_onelev_mat_asb(lv,a,desc_a,ilaggr,nlaggr,t_prol,info)
type(psb_cspmat_type) :: ac, op_restr, op_prol
integer(psb_ipk_) :: nzl, inl
integer(psb_ipk_) :: debug_level, debug_unit
integer(psb_ipk_), save :: idx_matbld=-1, idx_matasb=-1, idx_mapbld=-1
logical, parameter :: do_timings=.false.
name='amg_c_onelev_mat_asb'
call psb_erractionsave(err_act)
@@ -120,6 +122,12 @@ subroutine amg_c_base_onelev_mat_asb(lv,a,desc_a,ilaggr,nlaggr,t_prol,info)
info = psb_success_
ctxt = desc_a%get_context()
call psb_info(ctxt,me,np)
if ((do_timings).and.(idx_matbld==-1)) &
& idx_matbld = psb_get_timer_idx("LEV_MASB: mat_bld")
if ((do_timings).and.(idx_matasb==-1)) &
& idx_matasb = psb_get_timer_idx("LEV_MASB: mat_asb")
if ((do_timings).and.(idx_mapbld==-1)) &
& idx_mapbld = psb_get_timer_idx("LEV_MASB: map_bld")
call amg_check_def(lv%parms%aggr_prol,'Smoother',&
& amg_smooth_prol_,is_legal_ml_aggr_prol)
@@ -139,9 +147,10 @@ subroutine amg_c_base_onelev_mat_asb(lv,a,desc_a,ilaggr,nlaggr,t_prol,info)
! the mapping defined by amg_aggrmap_bld and applying the aggregation
! algorithm specified by lv%iprcparm(amg_aggr_prol_)
!
if (do_timings) call psb_tic(idx_matbld)
call lv%aggr%mat_bld(lv%parms,a,desc_a,ilaggr,nlaggr,&
& lv%ac,lv%desc_ac,op_prol,op_restr,t_prol,info)
if (do_timings) call psb_toc(idx_matbld)
if(info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='amg_aggrmat_asb')
goto 9999
@@ -151,14 +160,17 @@ subroutine amg_c_base_onelev_mat_asb(lv,a,desc_a,ilaggr,nlaggr,t_prol,info)
! Now build its descriptor and convert global indices for
! ac, op_restr and op_prol
!
if (do_timings) call psb_tic(idx_matasb)
if (info == psb_success_) &
& call lv%aggr%mat_asb(lv%parms,a,desc_a,&
& lv%ac,lv%desc_ac,op_prol,op_restr,info)
if (do_timings) call psb_toc(idx_matasb)
if (do_timings) call psb_tic(idx_mapbld)
if (info == psb_success_) call lv%ac%cscnv(info,type='csr',dupl=psb_dupl_add_)
if (info == psb_success_) call lv%aggr%bld_map(desc_a, lv%desc_ac,&
& ilaggr,nlaggr,op_restr,op_prol,lv%linmap,info)
if (do_timings) call psb_toc(idx_mapbld)
if(info /= psb_success_) then
call psb_errpush(psb_err_from_subroutine_,name,a_err='mat_asb/map_bld')
goto 9999
@@ -0,0 +1,150 @@
!
!
! AMG4PSBLAS version 1.0
! Algebraic Multigrid Package
! based on PSBLAS (Parallel Sparse BLAS version 3.7)
!
! (C) Copyright 2021
!
! Salvatore Filippone
! Pasqua D'Ambra
! Fabio Durastante
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
! 1. Redistributions of source code must retain the above copyright
! notice, this list of conditions and the following disclaimer.
! 2. Redistributions in binary form must reproduce the above copyright
! notice, this list of conditions, and the following disclaimer in the
! documentation and/or other materials provided with the distribution.
! 3. The name of the AMG4PSBLAS group or the names of its contributors may
! not be used to endorse or promote products derived from this
! software without specific written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
! TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
! PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AMG4PSBLAS GROUP OR ITS CONTRIBUTORS
! BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
! CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
! SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
! INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
! CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
! ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
!
!
!
! verbosity:
! <0: suppress all messages
! 0: normal
! >1: increased details
!
subroutine amg_c_base_onelev_memory_use(lv,il,nl,ilmin,info,iout,verbosity,prefix,global)
use psb_base_mod
use amg_c_onelev_mod, amg_protect_name => amg_c_base_onelev_memory_use
Implicit None
! Arguments
class(amg_c_onelev_type), intent(in) :: lv
integer(psb_ipk_), intent(in) :: il,nl,ilmin
integer(psb_ipk_), intent(out) :: info
integer(psb_ipk_), intent(in), optional :: iout
character(len=*), intent(in), optional :: prefix
integer(psb_ipk_), intent(in), optional :: verbosity
logical, intent(in), optional :: global
! Local variables
type(psb_ctxt_type) :: ctxt
integer(psb_ipk_) :: err_act ,me, np
character(len=20), parameter :: name='amg_c_base_onelev_memory_use'
integer(psb_ipk_) :: iout_, verbosity_
logical :: coarse, global_
character(1024) :: prefix_
integer(psb_epk_), allocatable :: sz(:)
call psb_erractionsave(err_act)
ctxt = lv%base_desc%get_ctxt()
call psb_info(ctxt,me,np)
coarse = (il==nl)
if (present(iout)) then
iout_ = iout
else
iout_ = psb_out_unit
end if
if (present(verbosity)) then
verbosity_ = verbosity
else
verbosity_ = 0
end if
if (verbosity_ < 0) goto 9998
if (present(global)) then
global_ = global
else
global_ = .true.
end if
if (present(prefix)) then
prefix_ = prefix
else
prefix_ = ""
end if
if ((me == 0).or.(verbosity_>0)) write(iout_,*) trim(prefix_)
if (global_) then
allocate(sz(6))
sz(:) = 0
sz(1) = lv%base_a%sizeof()
sz(2) = lv%base_desc%sizeof()
if (il >1) sz(3) = lv%linmap%sizeof()
if (allocated(lv%sm)) sz(4) = lv%sm%sizeof()
if (allocated(lv%sm2a)) sz(5) = lv%sm2a%sizeof()
if (allocated(lv%wrk)) sz(6) = lv%wrk%sizeof()
call psb_sum(ctxt,sz)
if (me == 0) then
if (coarse) then
write(iout_,*) trim(prefix_), ' Level ',il,' (coarse)'
else
write(iout_,*) trim(prefix_), ' Level ',il
end if
write(iout_,*) trim(prefix_), ' Matrix:', sz(1)
write(iout_,*) trim(prefix_), ' Descriptor:', sz(2)
if (il >1) write(iout_,*) trim(prefix_), ' Linear map:', sz(3)
if (allocated(lv%sm)) write(iout_,*) trim(prefix_), ' Smoother:', sz(4)
if (allocated(lv%sm2a)) write(iout_,*) trim(prefix_), ' Smoother 2:', sz(5)
if (allocated(lv%wrk)) write(iout_,*) trim(prefix_), ' Workspace:', sz(6)
end if
else
if ((me == 0).or.(verbosity_>0)) then
if (coarse) then
write(iout_,*) trim(prefix_), ' Level ',il,' (coarse)'
else
write(iout_,*) trim(prefix_), ' Level ',il
end if
write(iout_,*) trim(prefix_), ' Matrix:', lv%base_a%sizeof()
write(iout_,*) trim(prefix_), ' Descriptor:', lv%base_desc%sizeof()
if (il >1) write(iout_,*) trim(prefix_), ' Linear map:', lv%linmap%sizeof()
if (allocated(lv%sm)) write(iout_,*) trim(prefix_), ' Smoother:', lv%sm%sizeof()
if (allocated(lv%sm2a)) write(iout_,*) trim(prefix_), ' Smoother 2:', lv%sm2a%sizeof()
if (allocated(lv%wrk)) write(iout_,*) trim(prefix_), ' Workspace:', lv%wrk%sizeof()
end if
endif
9998 continue
call psb_erractionrestore(err_act)
return
9999 call psb_error_handler(err_act)
return
end subroutine amg_c_base_onelev_memory_use
+11 -3
View File
@@ -43,10 +43,12 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
use amg_d_dec_aggregator_mod
use amg_d_symdec_aggregator_mod
use amg_d_parmatch_aggregator_mod
use amg_d_poly_smoother
use amg_d_jac_smoother
use amg_d_as_smoother
use amg_d_diag_solver
use amg_d_l1_diag_solver
use amg_d_jac_solver
use amg_d_ilu_solver
use amg_d_id_solver
use amg_d_gs_solver
@@ -85,6 +87,8 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
type(amg_d_as_smoother_type) :: amg_d_as_smoother_mold
type(amg_d_diag_solver_type) :: amg_d_diag_solver_mold
type(amg_d_l1_diag_solver_type) :: amg_d_l1_diag_solver_mold
type(amg_d_jac_solver_type) :: amg_d_jac_solver_mold
type(amg_d_l1_jac_solver_type) :: amg_d_l1_jac_solver_mold
type(amg_d_ilu_solver_type) :: amg_d_ilu_solver_mold
type(amg_d_id_solver_type) :: amg_d_id_solver_mold
type(amg_d_gs_solver_type) :: amg_d_gs_solver_mold
@@ -92,6 +96,7 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
type(amg_d_ainv_solver_type) :: amg_d_ainv_solver_mold
type(amg_d_invk_solver_type) :: amg_d_invk_solver_mold
type(amg_d_invt_solver_type) :: amg_d_invt_solver_mold
type(amg_d_poly_smoother_type) :: amg_d_poly_smoother_mold
#if defined(HAVE_UMF_)
type(amg_d_umf_solver_type) :: amg_d_umf_solver_mold
#endif
@@ -153,6 +158,9 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
call lv%set(amg_d_as_smoother_mold,info,pos=pos)
if (info == 0) call lv%set(amg_d_ilu_solver_mold,info,pos=pos)
case ('POLY')
call lv%set(amg_d_poly_smoother_mold,info,pos=pos)
if (info == 0) call lv%set(amg_d_l1_diag_solver_mold,info,pos=pos)
case ('GS','FWGS')
call lv%set(amg_d_jac_smoother_mold,info,pos='pre')
if (info == 0) call lv%set(amg_d_gs_solver_mold,info,pos='pre')
@@ -198,10 +206,10 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
case ('NONE','NOPREC','FACT_NONE')
call lv%set(amg_d_id_solver_mold,info,pos=pos)
case ('DIAG')
case ('DIAG','JACOBI')
call lv%set(amg_d_diag_solver_mold,info,pos=pos)
case ('L1-DIAG')
case ('L1-DIAG','L1-JACOBI')
call lv%set(amg_d_l1_diag_solver_mold,info,pos=pos)
case ('GS','FGS','FWGS')
@@ -264,7 +272,7 @@ subroutine amg_d_base_onelev_csetc(lv,what,val,info,pos,idx)
end if
select case(val)
case('DEC')
case('DEC','DECOUPLED')
allocate(amg_d_dec_aggregator_type :: lv%aggr, stat=info)
case('SYMDEC')
allocate(amg_d_symdec_aggregator_type :: lv%aggr, stat=info)

Some files were not shown because too many files have changed in this diff Show More