Merge branch 'master' of https://github.com/reger-men/HPL_GPU

CMakeLists.txt

@@ -148,7 +148,7 @@ src/pgesv/HPL_pdupdateNT.cpp
 src/pgesv/HPL_pdupdateNN.cpp src/pgesv/HPL_pdupdateTN.cpp src/pgesv/HPL_pdupdateTT.cpp
 src/pgesv/HPL_pdupdateTN.cpp src/pgesv/HPL_pdupdateTT.cpp
 src/pauxil/HPL_pwarn.cpp src/comm/HPL_bcast.cpp src/comm/HPL_blong.cpp
-src/comm/HPL_blonM.cpp src/comm/HPL_1ring.cpp src/comm/HPL_2ring.cpp
+src/comm/HPL_blonM.cpp src/comm/HPL_bidir.cpp src/comm/HPL_1ring.cpp src/comm/HPL_2ring.cpp
 src/comm/HPL_1rinM.cpp src/comm/HPL_2rinM.cpp src/comm/HPL_packL.cpp
 src/comm/HPL_sdrv.cpp src/comm/HPL_send.cpp src/pgesv/HPL_pdlaswp00N.cpp
 src/comm/HPL_recv.cpp src/grid/HPL_reduce.cpp src/comm/HPL_binit.cpp

include/hpl_comm.h

@@ -65,7 +65,8 @@ typedef enum
    HPL_2RING_M       = 404,           /* Increasing 2-ring (modified) */
    HPL_BLONG         = 405,                         /* long broadcast */
    HPL_BLONG_M       = 406,               /* long broadcast (modified) */
-   HPL_IBCAST       = 407               /* default bcast in MPI */
+   HPL_IBCAST        = 407,               /* default bcast in MPI */
+   HPL_BIDIR         = 408,               /* bidirectional */
 } HPL_T_TOP;
 /*
  * ---------------------------------------------------------------------
@@ -160,6 +161,10 @@ int HPL_binit_blonM STDC_ARGS( ( HPL_T_panel *        ) );
 int HPL_bcast_blonM STDC_ARGS( ( HPL_T_panel *, int * ) );
 int HPL_bwait_blonM STDC_ARGS( ( HPL_T_panel *        ) );
 
+int HPL_binit_bidir STDC_ARGS( ( HPL_T_panel *        ) );
+int HPL_bcast_bidir STDC_ARGS( ( HPL_T_panel *, int * ) );
+int HPL_bwait_bidir STDC_ARGS( ( HPL_T_panel *        ) );
+
 #endif
 /*
  * End of hpl_comm.h

makes/Make.comm

@@ -56,7 +56,7 @@ INCdep           = \
 #
 HPL_comobj       = \
    HPL_1ring.o            HPL_1rinM.o            HPL_2ring.o            \
-   HPL_2rinM.o            HPL_blong.o            HPL_blonM.o            \
+   HPL_2rinM.o            HPL_blong.o            HPL_bidir.o           HPL_blonM.o            \
    HPL_packL.o            HPL_copyL.o            HPL_binit.o            \
    HPL_bcast.o            HPL_bwait.o            HPL_send.o             \
    HPL_recv.o             HPL_sdrv.o
@@ -84,6 +84,8 @@ HPL_2rinM.o            : ../HPL_2rinM.cpp            $(INCdep)
 	$(CC) -o $@ -c $(CCFLAGS)  ../HPL_2rinM.cpp
 HPL_blong.o            : ../HPL_blong.cpp            $(INCdep)
 	$(CC) -o $@ -c $(CCFLAGS)  ../HPL_blong.cpp
+HPL_bidir.o            : ../HPL_bidir.cpp            $(INCdep)
+	$(CC) -o $@ -c $(CCFLAGS)  ../HPL_bidir.cpp
 HPL_blonM.o            : ../HPL_blonM.cpp            $(INCdep)
 	$(CC) -o $@ -c $(CCFLAGS)  ../HPL_blonM.cpp
 HPL_packL.o            : ../HPL_packL.cpp            $(INCdep)

scripts/config/HPL_128GPU.dat

@@ -8,8 +8,8 @@ HPL.out      output file name (if any)
 384          NBs
 0            PMAP process mapping (0=Row-,1=Column-major)
 1            # of process grids (P x Q)
-8            Ps
-16           Qs
+16           Ps
+8            Qs
 16.0         threshold
 1            # of panel fact
 2            PFACTs (0=left, 1=Crout, 2=Right)

scripts/config/HPL_2048GPU.dat

@@ -0,0 +1,31 @@
+HPLinpack benchmark input file
+Innovative Computing Laboratory, University of Tennessee
+HPL.out      output file name (if any)
+0            device out (6=stdout,7=stderr,file)
+1            # of problems sizes (N)
+4104192      N
+1            # of NBs
+384          NBs
+0            PMAP process mapping (0=Row-,1=Column-major)
+1            # of process grids (P x Q)
+64           Ps
+32           Qs
+16.0         threshold
+1            # of panel fact
+2            PFACTs (0=left, 1=Crout, 2=Right)
+1            # of recursive stopping criterium
+2            NBMINs (>= 1)
+1            # of panels in recursion
+2            NDIVs
+1            # of recursive panel fact.
+2            RFACTs (0=left, 1=Crout, 2=Right)
+1            # of broadcast
+6            BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=ibcast)
+1            # of lookahead depth
+1            DEPTHs (>=0)
+1            SWAP (0=bin-exch,1=long,2=mix)
+64           swapping threshold
+1            L1 in (0=transposed,1=no-transposed) form
+0            U  in (0=transposed,1=no-transposed) form
+0            Equilibration (0=no,1=yes)
+8            memory alignment in double (> 0)

scripts/config/HPL_32GPU.dat

@@ -8,8 +8,8 @@ HPL.out      output file name (if any)
 384          NBs
 0            PMAP process mapping (0=Row-,1=Column-major)
 1            # of process grids (P x Q)
-4            Ps
-8            Qs
+8            Ps
+4            Qs
 16.0         threshold
 1            # of panel fact
 2            PFACTs (0=left, 1=Crout, 2=Right)

scripts/config/HPL_4096GPU.dat

@@ -0,0 +1,31 @@
+HPLinpack benchmark input file
+Innovative Computing Laboratory, University of Tennessee
+HPL.out      output file name (if any)
+0            device out (6=stdout,7=stderr,file)
+1            # of problems sizes (N)
+5786112      N
+1            # of NBs
+384          NBs
+0            PMAP process mapping (0=Row-,1=Column-major)
+1            # of process grids (P x Q)
+64           Ps
+64           Qs
+16.0         threshold
+1            # of panel fact
+2            PFACTs (0=left, 1=Crout, 2=Right)
+1            # of recursive stopping criterium
+2            NBMINs (>= 1)
+1            # of panels in recursion
+2            NDIVs
+1            # of recursive panel fact.
+2            RFACTs (0=left, 1=Crout, 2=Right)
+1            # of broadcast
+6            BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=ibcast)
+1            # of lookahead depth
+1            DEPTHs (>=0)
+1            SWAP (0=bin-exch,1=long,2=mix)
+64           swapping threshold
+1            L1 in (0=transposed,1=no-transposed) form
+0            U  in (0=transposed,1=no-transposed) form
+0            Equilibration (0=no,1=yes)
+8            memory alignment in double (> 0)

scripts/config/HPL_512GPU.dat

@@ -8,8 +8,8 @@ HPL.out      output file name (if any)
 384          NBs
 0            PMAP process mapping (0=Row-,1=Column-major)
 1            # of process grids (P x Q)
-16           Ps
-32           Qs
+32           Ps
+16           Qs
 16.0         threshold
 1            # of panel fact
 2            PFACTs (0=left, 1=Crout, 2=Right)

scripts/config/HPL_8GPU.dat

@@ -3,13 +3,13 @@ Innovative Computing Laboratory, University of Tennessee
 HPL.out      output file name (if any)
 0            device out (6=stdout,7=stderr,file)
 1            # of problems sizes (N)
-256000       N
+256512       N
 1            # of NBs
 384          NBs
 0            PMAP process mapping (0=Row-,1=Column-major)
 1            # of process grids (P x Q)
-2            Ps
-4            Qs
+4            Ps
+2            Qs
 16.0         threshold
 1            # of panel fact
 2            PFACTs (0=left, 1=Crout, 2=Right)
@@ -19,8 +19,8 @@ HPL.out      output file name (if any)
 2            NDIVs
 1            # of recursive panel fact.
 2            RFACTs (0=left, 1=Crout, 2=Right)
-1            # of broadcast
-6            BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=ibcast)
+8            # of broadcast
+6            BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=ibcast,7=BiDir)
 1            # of lookahead depth
 1            DEPTHs (>=0)
 1            SWAP (0=bin-exch,1=long,2=mix)

scripts/env.crusher.sh

@@ -0,0 +1,20 @@
+# modules
+module reset
+
+module load craype-accel-amd-gfx90a
+module load PrgEnv-amd
+module load amd/5.2.0
+module load rocm/5.2.0
+module load cray-mpich/8.1.17
+module load openblas/0.3.17-omp
+
+#
+# env
+#
+export LD_LIBRARY_PATH="${CRAY_LD_LIBRARY_PATH}:${LD_LIBRARY_PATH}"
+# enable GPU aware MPI
+export MPICH_GPU_SUPPORT_ENABLED=1
+# to work around the OFI registration cache issue for > 8 nodes
+export FI_MR_CACHE_MAX_COUNT=0
+#export MPICH_SMP_SINGLE_COPY_MODE=NONE # does not work
+export MPICH_RANK_REORDER_DISPLAY=1

scripts/env.lumi.sh

@@ -0,0 +1,22 @@
+# Any other commands must follow the #SBATCH directives
+module load LUMI/22.06 partition/G
+#module load rocm/5.1.4
+
+module use /project/project_462000075/paklui/modulefiles
+#module load rocm/5.3.0-10584
+#module load rocm/5.3.0-10619
+module load rocm/5.3.0-10670
+#module load openblas/0.3.17-omp
+#module load cce/14.0.2
+#module load cray-libsci/22.08.1.1
+module load cray-mpich/8.1.18
+#module load craype/2.7.17
+
+#
+# env
+#
+# enable GPU aware MPI
+export MPICH_GPU_SUPPORT_ENABLED=1
+# to work around the OFI registration cache issue for > 8 nodes
+#export FI_MR_CACHE_MAX_COUNT=0
+export MPICH_RANK_REORDER_DISPLAY=1

scripts/run_hpl.slurm

@@ -3,14 +3,16 @@
 #SBATCH -N 1
 ##SBATCH -n 8
 #SBATCH -c 8
-#SBATCH -t 1:00:00
-#SBATCH -A VEN114
+#SBATCH -t 2:00:00
+##SBATCH -A VEN114
+#SBATCH -A project_462000075
 #SBATCH -J xhplhip
 #SBATCH --gpu-bind=closest
 #SBATCH --ntasks-per-node=8
 #SBATCH --gpus-per-node=8
 
-source ../env/env.crusher.sh
+#source ../env/env.crusher.sh
+source ../env/env.lumi.sh
 
 NP=$SLURM_NPROCS
 NODES=$SLURM_NNODES
@@ -34,9 +36,17 @@ CMD+="-o $LOG -e $LOG "
 #CMD+="${HOME}/mpich_bind.sh "
 CMD+="$EXE"
 
+if [ $NODES -gt 8 ]; then
+    echo "export FI_MR_CACHE_MAX_COUNT=0"
+          export FI_MR_CACHE_MAX_COUNT=0
+else
+    echo "unset FI_MR_CACHE_MAX_COUNT"
+          unset FI_MR_CACHE_MAX_COUNT
+fi
+
 #export MPICH_SMP_SINGLE_COPY_MODE=NONE # does not work
-export FI_MR_CACHE_MAX_COUNT=0
-export MPICH_RANK_REORDER_DISPLAY=1
+#export FI_MR_CACHE_MAX_COUNT=0 # workaround for failed to allocate memory
+#export MPICH_RANK_REORDER_DISPLAY=1
 
 echo $CMD >> $LOG
 echo $CMD 2>&1 | tee -a $LOG

src/Makefile.am

@@ -8,7 +8,7 @@ blas/HPL_dscal.c blas/HPL_dtrsm.c blas/HPL_dtrsv.c blas/HPL_idamax.c \
 blas/HPL_dgemv.c blas/HPL_dscal.c blas/HPL_daxpy.c \
 blas/HPL_dcopy.c blas/HPL_dgemm.c blas/HPL_dgemv.c blas/HPL_dger.c \
 comm/HPL_sdrv.c comm/HPL_send.c comm/HPL_recv.c comm/HPL_bcast.c \
-comm/HPL_binit.c comm/HPL_bwait.c comm/HPL_blong.c comm/HPL_1ring.c \
+comm/HPL_binit.c comm/HPL_bwait.c comm/HPL_blong.c comm/HPL_bidir.c comm/HPL_1ring.c \
 comm/HPL_1rinM.c comm/HPL_2rinM.c comm/HPL_2ring.c comm/HPL_blonM.c comm/HPL_packL.c \
 grid/HPL_reduce.c grid/HPL_sum.c grid/HPL_grid_info.c grid/HPL_grid_init.c \
 grid/HPL_all_reduce.c grid/HPL_broadcast.c grid/HPL_grid_exit.c grid/HPL_max.c \

src/Makefile.in

@@ -147,6 +147,7 @@ am_libhpl_a_OBJECTS = auxil/HPL_dlatcpy.$(OBJEXT) \
 	comm/HPL_recv.$(OBJEXT) comm/HPL_bcast.$(OBJEXT) \
 	comm/HPL_binit.$(OBJEXT) comm/HPL_bwait.$(OBJEXT) \
 	comm/HPL_blong.$(OBJEXT) comm/HPL_1ring.$(OBJEXT) \
+	comm/HPL_bidir.$(OBJEXT) \
 	comm/HPL_1rinM.$(OBJEXT) comm/HPL_2rinM.$(OBJEXT) \
 	comm/HPL_2ring.$(OBJEXT) comm/HPL_blonM.$(OBJEXT) \
 	comm/HPL_packL.$(OBJEXT) grid/HPL_reduce.$(OBJEXT) \
@@ -229,6 +230,7 @@ am__depfiles_remade = auxil/$(DEPDIR)/HPL_dlacpy.Po \
 	comm/$(DEPDIR)/HPL_2rinM.Po comm/$(DEPDIR)/HPL_2ring.Po \
 	comm/$(DEPDIR)/HPL_bcast.Po comm/$(DEPDIR)/HPL_binit.Po \
 	comm/$(DEPDIR)/HPL_blonM.Po comm/$(DEPDIR)/HPL_blong.Po \
+	comm/$(DEPDIR)/HPL_bidir.Po \
 	comm/$(DEPDIR)/HPL_bwait.Po comm/$(DEPDIR)/HPL_packL.Po \
 	comm/$(DEPDIR)/HPL_recv.Po comm/$(DEPDIR)/HPL_sdrv.Po \
 	comm/$(DEPDIR)/HPL_send.Po grid/$(DEPDIR)/HPL_all_reduce.Po \
@@ -428,7 +430,7 @@ blas/HPL_dscal.c blas/HPL_dtrsm.c blas/HPL_dtrsv.c blas/HPL_idamax.c \
 blas/HPL_dgemv.c blas/HPL_dscal.c blas/HPL_daxpy.c \
 blas/HPL_dcopy.c blas/HPL_dgemm.c blas/HPL_dgemv.c blas/HPL_dger.c \
 comm/HPL_sdrv.c comm/HPL_send.c comm/HPL_recv.c comm/HPL_bcast.c \
-comm/HPL_binit.c comm/HPL_bwait.c comm/HPL_blong.c comm/HPL_1ring.c \
+comm/HPL_binit.c comm/HPL_bwait.c comm/HPL_blong.c comm/HPL_bidir.c comm/HPL_1ring.c \
 comm/HPL_1rinM.c comm/HPL_2rinM.c comm/HPL_2ring.c comm/HPL_blonM.c comm/HPL_packL.c \
 grid/HPL_reduce.c grid/HPL_sum.c grid/HPL_grid_info.c grid/HPL_grid_init.c \
 grid/HPL_all_reduce.c grid/HPL_broadcast.c grid/HPL_grid_exit.c grid/HPL_max.c \
@@ -583,6 +585,8 @@ comm/HPL_bwait.$(OBJEXT): comm/$(am__dirstamp) \
 	comm/$(DEPDIR)/$(am__dirstamp)
 comm/HPL_blong.$(OBJEXT): comm/$(am__dirstamp) \
 	comm/$(DEPDIR)/$(am__dirstamp)
+comm/HPL_bidir.$(OBJEXT): comm/$(am__dirstamp) \
+	comm/$(DEPDIR)/$(am__dirstamp)
 comm/HPL_1ring.$(OBJEXT): comm/$(am__dirstamp) \
 	comm/$(DEPDIR)/$(am__dirstamp)
 comm/HPL_1rinM.$(OBJEXT): comm/$(am__dirstamp) \
@@ -818,6 +822,7 @@ distclean-compile:
 @AMDEP_TRUE@@am__include@ @am__quote@comm/$(DEPDIR)/HPL_binit.Po@am__quote@ # am--include-marker
 @AMDEP_TRUE@@am__include@ @am__quote@comm/$(DEPDIR)/HPL_blonM.Po@am__quote@ # am--include-marker
 @AMDEP_TRUE@@am__include@ @am__quote@comm/$(DEPDIR)/HPL_blong.Po@am__quote@ # am--include-marker
+@AMDEP_TRUE@@am__include@ @am__quote@comm/$(DEPDIR)/HPL_bidir.Po@am__quote@ # am--include-marker
 @AMDEP_TRUE@@am__include@ @am__quote@comm/$(DEPDIR)/HPL_bwait.Po@am__quote@ # am--include-marker
 @AMDEP_TRUE@@am__include@ @am__quote@comm/$(DEPDIR)/HPL_packL.Po@am__quote@ # am--include-marker
 @AMDEP_TRUE@@am__include@ @am__quote@comm/$(DEPDIR)/HPL_recv.Po@am__quote@ # am--include-marker
@@ -1086,6 +1091,7 @@ distclean: distclean-am
 	-rm -f comm/$(DEPDIR)/HPL_binit.Po
 	-rm -f comm/$(DEPDIR)/HPL_blonM.Po
 	-rm -f comm/$(DEPDIR)/HPL_blong.Po
+	-rm -f comm/$(DEPDIR)/HPL_bidir.Po
 	-rm -f comm/$(DEPDIR)/HPL_bwait.Po
 	-rm -f comm/$(DEPDIR)/HPL_packL.Po
 	-rm -f comm/$(DEPDIR)/HPL_recv.Po
@@ -1233,6 +1239,7 @@ maintainer-clean: maintainer-clean-am
 	-rm -f comm/$(DEPDIR)/HPL_binit.Po
 	-rm -f comm/$(DEPDIR)/HPL_blonM.Po
 	-rm -f comm/$(DEPDIR)/HPL_blong.Po
+	-rm -f comm/$(DEPDIR)/HPL_bidir.Po
 	-rm -f comm/$(DEPDIR)/HPL_bwait.Po
 	-rm -f comm/$(DEPDIR)/HPL_packL.Po
 	-rm -f comm/$(DEPDIR)/HPL_recv.Po

src/comm/HPL_bcast.cpp

@@ -109,6 +109,7 @@ int HPL_bcast
       case HPL_BLONG_M : ierr = HPL_bcast_blonM( PANEL, IFLAG ); break;
       case HPL_BLONG   : ierr = HPL_bcast_blong( PANEL, IFLAG ); break;
       case HPL_IBCAST  : ierr = HPL_BE_bcast_ibcast( PANEL, IFLAG, HPL_TR); break;
+      case HPL_BIDIR   : ierr = HPL_bcast_bidir( PANEL, IFLAG ); break;
       default          : ierr = HPL_SUCCESS;
    }