minor updates

liuyangzhuan · Dec 4, 2023 · 416774d · 416774d
1 parent 8187d48
commit 416774d
Show file tree

Hide file tree

Showing 36 changed files with 93 additions and 7 deletions.
diff --git a/EXAMPLE/RankBenchmark_Driver.f90 b/EXAMPLE/RankBenchmark_Driver.f90
@@ -117,7 +117,7 @@ PROGRAM ButterflyPACK_RankBenchmark
 	integer,allocatable::Permutation(:)
 	integer Nunk_loc,Nunk_m_loc, Nunk_n_loc
 	integer,allocatable::tree(:),tree_m(:),tree_n(:)
-	complex(kind=8),allocatable::rhs_glo(:,:),rhs_loc(:,:),x_glo(:,:),x_loc(:,:)
+	complex(kind=8),allocatable::rhs_glo(:,:),rhs_loc(:,:),x_glo(:,:),x_loc(:,:),xin_loc(:,:),xout_loc(:,:)
 	integer nrhs
 	type(z_matrixblock) ::blocks
 	character(len=1024)  :: strings,strings1
@@ -153,7 +153,7 @@ PROGRAM ButterflyPACK_RankBenchmark
 
 
 	quant%tst = 2
-
+    quant%wavelen = 0.25d0/8d0
 
 	nargs = iargc()
 	ii=1
@@ -170,6 +170,8 @@ PROGRAM ButterflyPACK_RankBenchmark
 						call getarg(ii,strings1)
 						if(trim(strings)=='--tst')then
 							read(strings1,*)quant%tst
+						elseif(trim(strings)=='--wavelen')then
+							read(strings1,*)quant%wavelen
 						else
 							if(ptree%MyID==Main_ID)write(*,*)'ignoring unknown quant: ', trim(strings)
 						endif
@@ -199,7 +201,7 @@ PROGRAM ButterflyPACK_RankBenchmark
 ! Read a full non-square matrix and do a BF compression
 
 	ppw=5
-    quant%wavelen = 0.25d0/8d0
+
     ds = quant%wavelen/ppw
     if(quant%tst==1)then ! two colinear plate
       Nperdim = z_ceiling_safe(1d0/ds)
@@ -298,7 +300,13 @@ PROGRAM ButterflyPACK_RankBenchmark
 	call MPI_Bcast(quant%Permutation_n,quant%Nunk_n,MPI_integer,0,ptree%comm,ierr)
 
 	call z_BF_Construct_Element_Compute(blocks, option, stats, msh, ker, ptree)
-
+	nrhs=1
+	allocate(xin_loc(Nunk_n_loc,nrhs))
+	xin_loc=1
+	allocate(xout_loc(Nunk_m_loc,nrhs))
+	call z_BF_Mult('N', xin_loc, xout_loc, Nunk_n_loc, Nunk_m_loc, nrhs, blocks, option, stats, ptree)
+	deallocate(xin_loc)
+	deallocate(xout_loc)
 !******************************************************************************!
 
 	!**** print statistics

diff --git a/SRC/BPACK_constr.f90 b/SRC/BPACK_constr.f90
@@ -1550,6 +1550,7 @@ subroutine BF_MD_Construct_Element_Compute(Ndim, blocks, option, stats, msh, ker
 
       ! !!!! the following functions have not been tensorized
       if (option%verbosity >= 0)call BF_MD_checkError(Ndim, blocks_1, option, msh, ker, stats, ptree, 0, option%verbosity)
+      stats%Mem_Comp_for=Memory
       ! call BF_ComputeMemory(blocks, stats%Mem_Comp_for)
 
 

diff --git a/SRC/Bplus_utilities.f90 b/SRC/Bplus_utilities.f90
@@ -7054,6 +7054,42 @@ subroutine BF_all2all_V_split(block_i, pgno_i, level_i, block_o, pgnos_o, level_
 
    end subroutine BF_all2all_V_split
 
+   subroutine BF_Mult(chara, xin, xout, Ninloc, Noutloc, Ncol, blocks, option, stats, ptree)
+      implicit none
+      real(kind=8) t1, t2
+      character chara
+      integer Ninloc, Noutloc, Ncol
+      DT::xin(Ninloc, Ncol), xout(Noutloc, Ncol)
+
+      type(Hstat)::stats
+      type(Hoption)::option
+      type(matrixblock)::blocks
+
+      type(proctree)::ptree
+
+      t1 = MPI_Wtime()
+
+      stats%Flop_Tmp = 0
+      stats%Flop_C_Mult = 0
+      stats%Time_C_Mult = 0
+      xout = 0
+      if (chara == 'N') then
+         call BF_block_MVP_dat(blocks, chara, Noutloc, Ninloc, Ncol, xin, Ninloc, xout,Noutloc, BPACK_cone, BPACK_czero, ptree, stats)
+      else
+         call BF_block_MVP_dat(blocks, chara, Ninloc, Noutloc, Ncol, xin, Ninloc, xout, Noutloc, BPACK_cone, BPACK_czero, ptree, stats)
+      endif
+
+      t2 = MPI_Wtime()
+
+      xout = xout/option%scale_factor
+
+      stats%Time_C_Mult = stats%Time_C_Mult + t2 - t1
+      stats%Flop_C_Mult = stats%Flop_C_Mult + stats%Flop_Tmp
+
+   end subroutine BF_Mult
+
+
+
 
    subroutine BF_block_MVP_dat(blocks, chara, M, N, Nrnd, random1, ldi, random2, ldo, a, b, ptree, stats)
       implicit none

diff --git a/example_scripts/pbs_script.sh b/example_scripts/pbs_script.sh
diff --git a/example_scripts/runYang_fkerreg.sh b/example_scripts/runYang_fkerreg.sh
diff --git a/example_scripts/run_cmake_build_gnu_osx_catalina.sh b/example_scripts/run_cmake_build_gnu_osx_catalina.sh
diff --git a/example_scripts/run_cmake_build_gnu_osx_static.sh b/example_scripts/run_cmake_build_gnu_osx_static.sh
diff --git a/example_scripts/run_cmake_build_gnu_osx_zsh.sh b/example_scripts/run_cmake_build_gnu_osx_zsh.sh
diff --git a/example_scripts/run_cmake_build_gnu_summit.sh b/example_scripts/run_cmake_build_gnu_summit.sh
diff --git a/example_scripts/run_cmake_build_gnu_ubuntu.sh b/example_scripts/run_cmake_build_gnu_ubuntu.sh
diff --git a/example_scripts/run_cmake_build_gnu_ubuntu_mpi4_gcc910.sh b/example_scripts/run_cmake_build_gnu_ubuntu_mpi4_gcc910.sh
@@ -32,10 +32,12 @@ cmake .. \
 	-DCMAKE_C_COMPILER=$MPICC \
 	-DCMAKE_INSTALL_PREFIX=. \
 	-DCMAKE_INSTALL_LIBDIR=./lib \
-	-DCMAKE_BUILD_TYPE=Debug\
+	-DCMAKE_BUILD_TYPE=Release\
 	-DTPL_ARPACK_LIBRARIES="/home/administrator/Desktop/Software/arpack-ng/build/lib/libarpack.so;/home/administrator/Desktop/Software/arpack-ng/build/lib/libparpack.so" \
 	-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON
 
+	chmod -R 777 ../SRC*
+
 	# -DTPL_ARPACK_LIBRARIES="/home/administrator/Desktop/Software/arpack-ng/build/lib/libarpack.so;/home/administrator/Desktop/Software/arpack-ng/build/lib/libparpack.so" \
 
 

diff --git a/example_scripts/run_cmake_build_intel_flux.sh b/example_scripts/run_cmake_build_intel_flux.sh
diff --git a/example_scripts/run_cmake_build_pgi_summit.sh b/example_scripts/run_cmake_build_pgi_summit.sh
diff --git a/example_scripts/run_cmake_build_pgi_ubuntu_mpi3_magma.sh b/example_scripts/run_cmake_build_pgi_ubuntu_mpi3_magma.sh
@@ -0,0 +1,38 @@
+module purge
+# module load pgi
+# module load scalapack-netlib/pgi-19.10/2.0.2
+module load PrgEnv-pgi
+d
+cd ..
+sed -i 's/^M$//' PrecisionPreprocessing.sh
+# bash PrecisionPreprocessing.sh
+mkdir -p build
+cd build
+rm -rf CMakeCache.txt
+rm -rf DartConfiguration.tcl
+rm -rf CTestTestfile.cmake
+rm -rf cmake_install.cmake
+rm -rf CMakeFiles
+cmake .. \
+	-DCMAKE_Fortran_FLAGS="" \
+	-DCMAKE_CXX_FLAGS="" \
+	-DBUILD_SHARED_LIBS=OFF \
+	-DTPL_BLAS_LIBRARIES="/opt/pgi/linux86-64-llvm/19.10/lib/libomp.so;/opt/pgi/linux86-64-llvm/19.10/lib/libblas.so" \
+	-DTPL_LAPACK_LIBRARIES="/opt/pgi/linux86-64-llvm/19.10/lib/liblapack.so" \
+	-DTPL_SCALAPACK_LIBRARIES="/opt/pgi/linux86-64-llvm/19.10/lib/scalapack/scalapack-2.0.2/openmpi-3.1.3/lib/libscalapack.a" \
+	-DCMAKE_Fortran_COMPILER=mpif90 \
+	-DCMAKE_CXX_COMPILER=mpicxx \
+	-DCMAKE_C_COMPILER=mpicc \
+	-DCMAKE_INSTALL_PREFIX=. \
+	-DCMAKE_BUILD_TYPE=Release \
+	-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON
+
+
+
+
+
+
+	#-DCMAKE_Fortran_FLAGS="-ftracer -funswitch-loops -ftree-vectorize -fimplicit-none -fno-range-check -finit-real=nan" \
+	# -DTPL_BLAS_LIBRARIES="" \
+	# -DTPL_LAPACK_LIBRARIES="/opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_gf_lp64.so;/opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_intel_thread.so;/opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_core.so;/opt/intel/compilers_and_libraries_2018.1.163/linux/compiler/lib/intel64/libiomp5.so" \
+	# -DTPL_SCALAPACK_LIBRARIES="/opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_blacs_intelmpi_lp64.so;/opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/lib/intel64/libmkl_scalapack_lp64.so" \
diff --git a/example_scripts/runit.sh b/example_scripts/runit.sh
diff --git a/example_scripts/runit_cpp.sh b/example_scripts/runit_cpp.sh
diff --git a/example_scripts/runit_em2d.sh b/example_scripts/runit_em2d.sh
diff --git a/example_scripts/runit_em2d_big.sh b/example_scripts/runit_em2d_big.sh
diff --git a/example_scripts/runit_em2d_big_corner.sh b/example_scripts/runit_em2d_big_corner.sh
diff --git a/example_scripts/runit_em2d_big_halfcircle.sh b/example_scripts/runit_em2d_big_halfcircle.sh
diff --git a/example_scripts/runit_em2d_big_halfcircle_100M.sh b/example_scripts/runit_em2d_big_halfcircle_100M.sh
diff --git a/example_scripts/runit_em2d_big_parallelstrip.sh b/example_scripts/runit_em2d_big_parallelstrip.sh
diff --git a/example_scripts/runit_em2d_big_rectangle.sh b/example_scripts/runit_em2d_big_rectangle.sh
diff --git a/example_scripts/runit_em2d_halfcircle_BACA.sh b/example_scripts/runit_em2d_halfcircle_BACA.sh
diff --git a/example_scripts/runit_em2d_parallel_test.sh b/example_scripts/runit_em2d_parallel_test.sh
diff --git a/example_scripts/runit_em2d_parallel_test_eigen.sh b/example_scripts/runit_em2d_parallel_test_eigen.sh
diff --git a/example_scripts/runit_em2d_parallel_test_hss.sh b/example_scripts/runit_em2d_parallel_test_hss.sh
diff --git a/example_scripts/runit_em2d_parallel_test_summit.sh b/example_scripts/runit_em2d_parallel_test_summit.sh
diff --git a/example_scripts/runit_em2d_parallel_test_summit_pgi.sh b/example_scripts/runit_em2d_parallel_test_summit_pgi.sh
diff --git a/example_scripts/runit_em3d.sh b/example_scripts/runit_em3d.sh
diff --git a/example_scripts/runit_em3d_bplus_cori.sh b/example_scripts/runit_em3d_bplus_cori.sh
diff --git a/example_scripts/runit_em3d_bplus_summit.sh b/example_scripts/runit_em3d_bplus_summit.sh
diff --git a/example_scripts/runit_em3d_sp.sh b/example_scripts/runit_em3d_sp.sh
diff --git a/example_scripts/runit_full_GO.sh b/example_scripts/runit_full_GO.sh
diff --git a/example_scripts/runit_krr_test_summit.sh b/example_scripts/runit_krr_test_summit.sh
diff --git a/example_scripts/runit_osx_ventura.sh b/example_scripts/runit_osx_ventura.sh
@@ -1,5 +1,6 @@
 export GPTUNEROOT=/Users/liuyangzhuan/Desktop/GPTune/
 export MPIRUN="$GPTUNEROOT/openmpi-4.1.5/bin/mpirun"
-$MPIRUN --allow-run-as-root -n 4 ../build/EXAMPLE/frankben_t -quant --tst 3 --wavelen 0.0625 -option --nmin_leaf 10 --lrlevel 100 --verbosity 1 --sample_para 1.0 --sample_para_outer 2.0 | tee a.out
-
+# $MPIRUN --allow-run-as-root -n 4 ../build/EXAMPLE/frankben_t -quant --tst 3 --wavelen 0.0625 -option --nmin_leaf 10 --lrlevel 100 --verbosity 1 --sample_para 1.0 --sample_para_outer 1.0 | tee a.out_tensor
+$MPIRUN --allow-run-as-root -n 4 ../build/EXAMPLE/frankben_t -quant --tst 2 --wavelen 0.03125 -option --nmin_leaf 10 --lrlevel 100 --verbosity 1 --sample_para 1.0 --sample_para_outer 1.0 | tee a.out_tensor
 
+$MPIRUN --allow-run-as-root -n 4 ../build/EXAMPLE/frankben -quant --tst 2 --wavelen 0.03125 -option --nmin_leaf 200 --lrlevel 100 --verbosity 1 --sample_para 2.0 --sample_para_outer 2.0 | tee a.out_matrix