Skip to content

GitLab

elpa
elpa
Commit 603ffc20 authored by Andreas Marek's avatar Andreas Marek
Browse files
Options

Some fixes for K-computer

parent 8666cdd0
Hide whitespace changes
Inline Side-by-side
Showing with 378 additions and 33 deletions
Makefile.am
View file @ 603ffc20
......@@ -61,9 +61,11 @@ libelpa@SUFFIX@_private_la_SOURCES = \
src/elpa1/elpa1.F90 \
src/elpa2/elpa2.F90 \
src/helpers/matrix_plot.F90 \
src/elpa_c_interface.c \
src/elpa_index.c
libelpa@SUFFIX@_private_la_SOURCES += src/elpa_c_interface.c
# elpa_utilities.F90 is private in new API, public in legacy
if ENABLE_LEGACY
libelpa@SUFFIX@_public_la_SOURCES += \
......@@ -554,7 +556,11 @@ elpa2_print_kernels@SUFFIX@_SOURCES = src/elpa2/elpa2_print_kernels.F90
elpa2_print_kernels@SUFFIX@_LDADD = libelpa@SUFFIX@.la
elpa2_print_kernels@SUFFIX@_FCFLAGS = $(AM_FCFLAGS) $(FC_MODINC)modules
include test_programs.am
include Fortran_test_programs.am
if ENABLE_C_TESTS
include C_test_programs.am
endif
if ENABLE_LEGACY
include legacy_test_programs.am
......
autogen.sh
View file @ 603ffc20
......@@ -5,5 +5,6 @@ mkdir -p m4/
test -n "$srcdir" || srcdir=`dirname "$0"`
test -n "$srcdir" || srcdir=.
$srcdir/generate_automake_test_programs.py > $srcdir/test_programs.am
$srcdir/generate_automake_Fortran_test_programs.py > $srcdir/Fortran_test_programs.am
$srcdir/generate_automake_C_test_programs.py > $srcdir/C_test_programs.am
autoreconf --force --install --verbose "$srcdir"
configure.ac
View file @ 603ffc20
......@@ -40,12 +40,24 @@ if test x$_cv_gnu_make_command = x ; then
AC_MSG_ERROR([Need GNU Make])
fi
AC_ARG_ENABLE([legacy],
AS_HELP_STRING([--disable-legacy],
AC_MSG_CHECKING(whether legacy interface should be provided)
AC_ARG_ENABLE([legacy-interface],
AS_HELP_STRING([--disable-legacy-interface],
[do not build legacy API, default yes]),
[],
[
if test x"$enableval" = x"yes"; then
enable_legayc=yes
else
enable_legacy=no
fi
],
[enable_legacy=yes])
AC_MSG_RESULT([${enable_legacy}])
AM_CONDITIONAL([ENABLE_LEGACY],[test x"$enable_legacy" = x"yes"])
if test x"${enable_legacy}" = x"yes"; then
AC_DEFINE([ENABLE_LEGACY], [1], [enable legacy interface])
fi
# gnu-make fortran module dependencies
m4_include([fdep/fortran_dependencies.m4])
......@@ -58,7 +70,13 @@ AC_MSG_CHECKING(whether --enable-openmp is specified)
AC_ARG_ENABLE([openmp],
AS_HELP_STRING([--enable-openmp],
[use OpenMP threading, default no.]),
[],
[
if test x"$enableval" = x"yes"; then
enable_openmp=yes
else
enable_openmp=no
fi
],
[enable_openmp=no])
AC_MSG_RESULT([${enable_openmp}])
AM_CONDITIONAL([WITH_OPENMP],[test x"$enable_openmp" = x"yes"])
......@@ -80,7 +98,13 @@ AC_MSG_CHECKING(whether --enable-scalapack-tests is specified)
AC_ARG_ENABLE([scalapack-tests],
AS_HELP_STRING([--enable-scalapack-tests],
[build SCALAPACK test cases for performance comparison, needs MPI, default no.]),
[],
[
if test x"$enableval" = x"yes"; then
enable_scalapack_tests=yes
else
enable_scalapack_tests=no
fi
],
[enable_scalapack_tests="no"])
AC_MSG_RESULT([$enable_scalapack_tests])
if test x"${enable_scalapack_tests}" = x"yes"; then
......@@ -163,7 +187,13 @@ AC_MSG_CHECKING(whether stdout/stderr file redirect should be enabled)
AC_ARG_ENABLE([redirect],
[AS_HELP_STRING([--enable-redirect],
[for test programs, allow redirection of stdout/stderr per MPI taks in a file (useful for timing), default no.])],
[],
[
if test x"$enableval" = x"yes"; then
enable_redirect=yes
else
enable_redirect=no
fi
],
[enable_redirect=no])
AC_MSG_RESULT([${enable_redirect}])
......@@ -188,7 +218,13 @@ dnl build with ftimings support
AC_ARG_ENABLE([timings],
[AS_HELP_STRING([--disable-timings],
[more detailed timing, default yes])],
[],
[
if test x"$enableval" = x"yes"; then
enable_timings=yes
else
enable_timings=no
fi
],
[enable_timings=yes])
if test x"${enable_timings}" = x"yes"; then
......@@ -201,7 +237,13 @@ AC_LANG_PUSH([C])
AC_ARG_WITH([papi],
[AS_HELP_STRING([--with-papi],
[Use PAPI to also measure flop count in the detailed timing (--enable-timing), disabled by default])],
[],
[
if test x"$enableval" = x"yes"; then
with_papi=yes
else
with_papi=no
fi
],
[with_papi="no"])
if test x"${enable_timings}" = x"yes"; then
if test x"$with_papi" = x"yes" ; then
......@@ -303,6 +345,7 @@ else
LIBS="-l${lib} ${old_LIBS}"
AC_MSG_CHECKING([whether -l${lib} already contains a BLACS implementation])
AC_LINK_IFELSE([AC_LANG_FUNC_LINK_TRY([blacs_gridinit])],[blacs_in_scalapack=yes],[blacs_in_scalapack=no])
AC_MSG_RESULT([${blacs_in_scalapack}])
if test x"${blacs_in_scalapack}" = x"yes"; then
break
......@@ -386,7 +429,13 @@ dnl check whether BAND_TO_FULL_BLOCKING is set
AC_MSG_CHECKING(whether BAND_TO_FLULL_BLOCKING is requested)
AC_ARG_ENABLE(band-to-full-blocking,[AS_HELP_STRING([--disable-band-to-full-blocking],
[build ELPA2 with blocking in band_to_full (default: enabled)])],
[],
[
if test x"$enableval" = x"yes"; then
enable_band_to_full_blocking=yes
else
enable_band_to_full_blocking=no
fi
],
[enable_band_to_full_blocking="yes"])
AC_MSG_RESULT([${enable_band_to_full_blocking}])
......@@ -430,7 +479,13 @@ if test x"${with_mpi}" = x"yes" ; then
AC_ARG_ENABLE([mpi-module],
AS_HELP_STRING([--disable-mpi-module],
[do not use the Fortran MPI module, get interfaces by 'include "mpif.h')]),
[],
[
if test x"$enableval" = x"yes"; then
enable_mpi_module=yes
else
enable_mpi_module=no
fi
],
[enable_mpi_module=yes])
if test x"${enable_mpi_module}" = x"yes" ; then
AC_MSG_CHECKING(whether Fortran mpi module can be used)
......@@ -1023,7 +1078,6 @@ DX_MAN_FEATURE(ON)
DX_HTML_FEATURE(ON)
DX_INIT_DOXYGEN([ELPA], [Doxyfile], [docs])
USE_ASSUMED_SIZE=yes
AC_MSG_CHECKING(whether assumed size Fortran arrays should be used)
AC_ARG_ENABLE([assumed-size],
AS_HELP_STRING([--disable-assumed-size],
......@@ -1040,7 +1094,6 @@ if test x"${USE_ASSUMED_SIZE}" = x"yes" ; then
AC_DEFINE([USE_ASSUMED_SIZE],[1],[for performance reasons use assumed size Fortran arrays, even if not debuggable])
fi
enable_fortran2008_features=yes
AC_MSG_CHECKING(whether Fortran2008 features should be enabled)
AC_ARG_ENABLE([Fortran2008-features],
AS_HELP_STRING([--enable-Fortran2008-features],
......@@ -1052,14 +1105,49 @@ AC_ARG_ENABLE([Fortran2008-features],
enable_fortran2008_features=no
fi
],
[])
[enable_fortran2008_features=yes])
AC_MSG_RESULT([${enable_fortran2008_features}])
AM_CONDITIONAL([USE_FORTRAN2008],[test x"$enable_fortran2008_features" = x"yes"])
if test x"${enable_fortran2008_features}" = x"yes"; then
AC_DEFINE([USE_FORTRAN2008], [1], [use some Fortran 2008 features])
fi
enable_kcomputer=no
AC_MSG_CHECKING(whether autotuning functionality should be enabled)
AC_ARG_ENABLE([autotuning],
AS_HELP_STRING([--enable-autotuning],
[enables autotuning functionality, default yes.]),
[
if test x"$enableval" = x"yes"; then
enable_autotuning=yes
else
enable_autotuning=no
fi
],
[enable_autotuning=yes])
AC_MSG_RESULT([${enable_autotuning}])
AM_CONDITIONAL([ENABLE_AUTOTUNING],[test x"$enable_autotuning" = x"yes"])
if test x"${enable_autotuning}" = x"yes"; then
AC_DEFINE([ENABLE_AUTOTUNING], [1], [enable autotuning functionality])
fi
AC_MSG_CHECKING(whether C tests should be provided)
AC_ARG_ENABLE([c-tests],
AS_HELP_STRING([--enable-c-tests],
[enables the C tests for elpa, default yes.]),
[
if test x"$enableval" = x"yes"; then
enable_c_tests=yes
else
enable_c_tests=no
fi
],
[enable_c_tests=yes])
AC_MSG_RESULT([${enable_c_tests}])
AM_CONDITIONAL([ENABLE_C_TESTS],[test x"$enable_c_tests" = x"yes"])
if test x"${enable_c_tests}" = x"yes"; then
AC_DEFINE([ENABLE_C_TESTS], [1], [enable C tests])
fi
AC_MSG_CHECKING(whether we build for K-computer)
AC_ARG_ENABLE([K-computer],
AS_HELP_STRING([--enable-K-computer],
......
generate_automake_test_programs.py → generate_automake_C_test_programs.py
View file @ 603ffc20
......@@ -3,7 +3,6 @@ from __future__ import print_function
from itertools import product
language_flag = {
"Fortran": "",
"C": "_c_version",
}
......@@ -213,6 +212,7 @@ for lang, p, d in product(sorted(language_flag.keys()), sorted(prec_flag.keys())
name = "test_autotune{langsuffix}_{d}_{p}".format(langsuffix=language_flag[lang], d=d, p=p)
print("if ENABLE_AUTOTUNING")
print("check_SCRIPTS += " + name + "_extended.sh")
print("noinst_PROGRAMS += " + name)
if lang == "Fortran":
......@@ -229,3 +229,5 @@ for lang, p, d in product(sorted(language_flag.keys()), sorted(prec_flag.keys())
domain_flag[d],
prec_flag[p]]))
print("endif\n" * endifs)
print("endif")
generate_automake_Fortran_test_programs.py 0 → 100755
View file @ 603ffc20
#!/usr/bin/env python
from __future__ import print_function
from itertools import product
language_flag = {
"Fortran": "",
}
domain_flag = {
"real": "-DTEST_REAL",
"complex": "-DTEST_COMPLEX",
}
prec_flag = {
"double": "-DTEST_DOUBLE",
"single": "-DTEST_SINGLE",
}
solver_flag = {
"1stage": "-DTEST_SOLVER_1STAGE",
"2stage": "-DTEST_SOLVER_2STAGE",
"scalapack_all": "-DTEST_SCALAPACK_ALL",
"scalapack_part": "-DTEST_SCALAPACK_PART",
}
gpu_flag = {
0: "-DTEST_GPU=0",
1: "-DTEST_GPU=1",
}
matrix_flag = {
"random": "-DTEST_MATRIX_RANDOM",
"analytic": "-DTEST_MATRIX_ANALYTIC",
"toeplitz": "-DTEST_MATRIX_TOEPLITZ",
"frank": "-DTEST_MATRIX_FRANK",
}
qr_flag = {
0: "-DTEST_QR_DECOMPOSITION=0",
1: "-DTEST_QR_DECOMPOSITION=1",
}
test_type_flag = {
"eigenvectors": "-DTEST_EIGENVECTORS",
"eigenvalues": "-DTEST_EIGENVALUES",
"solve_tridiagonal": "-DTEST_SOLVE_TRIDIAGONAL",
"cholesky": "-DTEST_CHOLESKY",
"hermitian_multiply": "-DTEST_HERMITIAN_MULTIPLY",
"generalized" : "-DTEST_GENERALIZED_EIGENPROBLEM",
"generalized_decomp": "-DTEST_GENERALIZED_DECOMP_EIGENPROBLEM",
}
layout_flag = {
"all_layouts": "-DTEST_ALL_LAYOUTS",
"square": ""
}
for lang, m, g, q, t, p, d, s, lay in product(sorted(language_flag.keys()),
sorted(matrix_flag.keys()),
sorted(gpu_flag.keys()),
sorted(qr_flag.keys()),
sorted(test_type_flag.keys()),
sorted(prec_flag.keys()),
sorted(domain_flag.keys()),
sorted(solver_flag.keys()),
sorted(layout_flag.keys())):
if lang == "C" and (m == "analytic" or m == "toeplitz" or m == "frank" or lay == "all_layouts"):
continue
# not implemented in the test.c file yet
if lang == "C" and (t == "cholesky" or t == "hermitian_multiply" or q == 1):
continue
# exclude some test combinations
# analytic tests only for "eigenvectors" and not on GPU
if(m == "analytic" and (g == 1 or t != "eigenvectors")):
continue
# Frank tests only for "eigenvectors" and eigenvalues and real double precision case
if(m == "frank" and ((t != "eigenvectors" or t != "eigenvalues") and (d != "real" or p != "double"))):
continue
if(s in ["scalapack_all", "scalapack_part"] and (g == 1 or t != "eigenvectors" or m != "analytic")):
continue
# do not test single-precision scalapack
if(s in ["scalapack_all", "scalapack_part"] and ( p == "single")):
continue
# solve tridiagonal only for real toeplitz matrix in 1stage
if (t == "solve_tridiagonal" and (s != "1stage" or d != "real" or m != "toeplitz")):
continue
# solve generalized only for random matrix in 1stage
if (t == "generalized" and (m != "random" or s == "2stage")):
continue
# solve generalized already decomposed only for random matrix in 1stage
# maybe this test should be further restricted, maybe not so important...
if (t == "generalized_decomp" and (m != "random" or s == "2stage")):
continue
# cholesky tests only 1stage and teoplitz or random matrix
if (t == "cholesky" and ((not (m == "toeplitz" or m == "random")) or s == "2stage")):
continue
if (t == "eigenvalues" and (m == "random")):
continue
if (t == "hermitian_multiply" and (s == "2stage")):
continue
if (t == "hermitian_multiply" and (m == "toeplitz")):
continue
# qr only for 2stage real
if (q == 1 and (s != "2stage" or d != "real" or t != "eigenvectors" or g == 1 or m != "random")):
continue
for kernel in ["all_kernels", "default_kernel"] if s == "2stage" else ["nokernel"]:
endifs = 0
extra_flags = []
if (t == "eigenvalues" and kernel == "all_kernels"):
continue
if (lang == "C" and kernel == "all_kernels"):
continue
if (g == 1):
print("if WITH_GPU_VERSION")
endifs += 1
if (lay == "all_layouts"):
print("if WITH_MPI")
endifs += 1
if (s in ["scalapack_all", "scalapack_part"]):
print("if WITH_SCALAPACK_TESTS")
endifs += 1
if kernel == "default_kernel":
extra_flags.append("-DTEST_KERNEL=ELPA_2STAGE_{0}_DEFAULT".format(d.upper()))
elif kernel == "all_kernels":
extra_flags.append("-DTEST_ALL_KERNELS")
if layout_flag[lay]:
extra_flags.append(layout_flag[lay])
if (p == "single"):
if (d == "real"):
print("if WANT_SINGLE_PRECISION_REAL")
elif (d == "complex"):
print("if WANT_SINGLE_PRECISION_COMPLEX")
else:
raise Exception("Oh no!")
endifs += 1
name = "test{langsuffix}_{d}_{p}_{t}_{s}{kernelsuffix}_{gpusuffix}{qrsuffix}{m}{layoutsuffix}".format(
langsuffix=language_flag[lang],
d=d, p=p, t=t, s=s,
kernelsuffix="" if kernel == "nokernel" else "_" + kernel,
gpusuffix="gpu_" if g else "",
qrsuffix="qr_" if q else "",
m=m,
layoutsuffix="_all_layouts" if lay == "all_layouts" else "")
print("if BUILD_KCOMPUTER")
print("bin_PROGRAMS += " + name)
print("else")
print("noinst_PROGRAMS += " + name)
print("endif")
if lay == "square":
print("check_SCRIPTS += " + name + "_default.sh")
elif lay == "all_layouts":
print("check_SCRIPTS += " + name + "_extended.sh")
else:
raise Exception("Unknown layout {0}".format(lay))
if lang == "Fortran":
print(name + "_SOURCES = test/Fortran/test.F90")
print(name + "_LDADD = $(test_program_ldadd)")
print(name + "_FCFLAGS = $(test_program_fcflags) \\")
elif lang == "C":
print(name + "_SOURCES = test/C/test.c")
print(name + "_LDADD = $(test_program_ldadd) $(FCLIBS)")
print(name + "_CFLAGS = $(test_program_cflags) \\")
print(" -DTEST_CASE=\\\"{0}\\\" \\".format(name))
print(" " + " \\\n ".join([
domain_flag[d],
prec_flag[p],
test_type_flag[t],
solver_flag[s],
gpu_flag[g],
qr_flag[q],
matrix_flag[m]] + extra_flags))
print("endif\n" * endifs)
for lang, p, d in product(sorted(language_flag.keys()), sorted(prec_flag.keys()), sorted(domain_flag.keys())):
endifs = 0
if (p == "single"):
if (d == "real"):
print("if WANT_SINGLE_PRECISION_REAL")
elif (d == "complex"):
print("if WANT_SINGLE_PRECISION_COMPLEX")
else:
raise Exception("Oh no!")
endifs += 1
name = "test_autotune{langsuffix}_{d}_{p}".format(langsuffix=language_flag[lang], d=d, p=p)
print("if ENABLE_AUTOTUNING")
print("check_SCRIPTS += " + name + "_extended.sh")
print("noinst_PROGRAMS += " + name)
if lang == "Fortran":
print(name + "_SOURCES = test/Fortran/test_autotune.F90")
print(name + "_LDADD = $(test_program_ldadd)")
print(name + "_FCFLAGS = $(test_program_fcflags) \\")
elif lang == "C":
print(name + "_SOURCES = test/C/test_autotune.c")
print(name + "_LDADD = $(test_program_ldadd) $(FCLIBS)")
print(name + "_CFLAGS = $(test_program_cflags) \\")
print(" " + " \\\n ".join([
domain_flag[d],
prec_flag[p]]))
print("endif\n" * endifs)
print("endif")
src/elpa.F90
View file @ 603ffc20
......@@ -189,6 +189,8 @@
!> \endcode
!>
!> \brief Fortran module to use the ELPA library. No other module shoule be used
#include "config-f90.h"
module elpa
use elpa_constants
use elpa_api
......@@ -219,7 +221,7 @@ module elpa
deallocate(obj)
end subroutine
#ifdef ENABLE_AUTOTUNING
!> \brief function to deallocate an ELPA autotune instance
!> Parameters
!> \details
......@@ -229,5 +231,6 @@ module elpa
call obj%destroy()
deallocate(obj)
end subroutine
#endif
end module
src/elpa_api.F90
View file @ 603ffc20
......@@ -153,10 +153,12 @@ module elpa_api
elpa_solve_tridiagonal_d, & !< matrix
elpa_solve_tridiagonal_f
#ifdef ENABLE_AUTOTUNING
! Auto-tune
procedure(elpa_autotune_setup_i), deferred, public :: autotune_setup !< method to prepare the ELPA autotuning
procedure(elpa_autotune_step_i), deferred, public :: autotune_step !< method to do an autotuning step
procedure(elpa_autotune_set_best_i), deferred, public :: autotune_set_best !< method to set the best options
#endif
!> \brief These method have to be public, in order to be overrideable in the extension types
procedure(elpa_set_integer_i), deferred, public :: elpa_set_integer
......@@ -204,7 +206,7 @@ module elpa_api
procedure(elpa_solve_tridiagonal_f_i), deferred, public :: elpa_solve_tridiagonal_f
end type elpa_t
#ifdef ENABLE_AUTOTUNING
!> \brief Abstract definition of the elpa_autotune type
type, abstract :: elpa_autotune_t
private
......@@ -212,7 +214,7 @@ module elpa_api
procedure(elpa_autotune_destroy_i), deferred, public :: destroy
procedure(elpa_autotune_print_i), deferred, public :: print
end type
#endif
!> \brief definition of helper function to get C strlen
!> Parameters
......@@ -243,7 +245,7 @@ module elpa_api
end function
end interface
#ifdef ENABLE_AUTOTUNING
!> \brief abstract definition of the autotune setup method
!> Parameters
!> \details
......@@ -298,7 +300,7 @@ module elpa_api
class(elpa_autotune_t), intent(in), target :: tune_state
end subroutine
end interface
#endif
!> \brief abstract definition of set method for integer values
!> Parameters
......@@ -558,7 +560,7 @@ module elpa_api
end subroutine
end interface
#ifdef ENABLE_AUTOTUNING
!> \brief abstract definition of interface to print the autotuning state
!> Parameters
!> \param self class(elpa_autotune_t): the ELPA autotune object
......@@ -581,7 +583,7 @@ module elpa_api
class(elpa_autotune_t), intent(inout) :: self
end subroutine
end interface
#endif
contains
......
src/elpa_autotune_impl.F90
View file @ 603ffc20
......@@ -4,7 +4,7 @@ module elpa_autotune_impl
use elpa_abstract_impl
use, intrinsic :: iso_c_binding
implicit none
#ifdef ENABLE_AUTOTUNING
type, extends(elpa_autotune_t) :: elpa_autotune_impl_t
class(elpa_abstract_impl_t), pointer :: parent => NULL()
integer :: i = 0
......@@ -37,5 +37,5 @@ module elpa_autotune_impl
class(elpa_autotune_impl_t), intent(inout) :: self
! nothing to do atm
end subroutine
#endif
end module
src/elpa_impl.F90
View file @ 603ffc20
......@@ -60,7 +60,9 @@ module elpa_impl
use elpa_utilities, only : error_unit
use elpa_abstract_impl
#ifdef ENABLE_AUTOTUNING
use elpa_autotune_impl
#endif
use, intrinsic :: iso_c_binding
implicit none
......@@ -154,10 +156,11 @@ module elpa_impl
procedure, private :: elpa_transform_back_generalized_fc
#endif
#ifdef ENABLE_AUTOTUNING
procedure, public :: autotune_setup => elpa_autotune_setup
procedure, public :: autotune_step => elpa_autotune_step
procedure, public :: autotune_set_best => elpa_autotune_set_best
#endif
procedure, private :: construct_scalapack_descriptor => elpa_construct_scalapack_descriptor
end type elpa_impl_t
......@@ -228,7 +231,7 @@ module elpa_impl
deallocate(self)
end subroutine
#ifdef ENABLE_AUTOTUNING
!c> /*! \brief C interface for the implementation of the elpa_autotune_deallocate method
!c> *
!c> * \param elpa_autotune_impl_t handle of ELPA autotune object to be deallocated
......@@ -244,7 +247,7 @@ module elpa_impl
call self%destroy()
deallocate(self)
end subroutine
#endif
!> \brief function to setup an ELPA object and to store the MPI communicators internally
!> Parameters
......@@ -734,7 +737,7 @@ module elpa_impl