Skip to content
GitLab
Commit 899bd7bf authored by Andreas Marek's avatar Andreas Marek
Browse files

Merge branch 'loh/autotuning' into master_pre_stage

parents b0c34c81 29bab835
Hide whitespace changes
Inline Side-by-side
Makefile.am
View file @ 899bd7bf
......@@ -35,6 +35,7 @@ noinst_LTLIBRARIES += libelpa@SUFFIX@_private.la
libelpa@SUFFIX@_private_la_FCFLAGS = $(AM_FCFLAGS) $(FC_MODOUT)private_modules $(FC_MODINC)private_modules
libelpa@SUFFIX@_private_la_SOURCES = \
src/elpa_impl.F90 \
src/elpa_autotune_impl.F90 \
src/elpa_abstract_impl.F90 \
src/helpers/mod_precision.F90 \
src/helpers/mod_mpi.F90 \
......
SWITCHING_TO_NEW_INTERFACE.md
View file @ 899bd7bf
......@@ -124,8 +124,6 @@ the user application):
"debug" can be one of { 0 | 1 }, will give more information case of an error if set to 1
"eigenvalues_only" can be one of { 0 | 1 }, if set to 1 solve will only compute the eigenvalues
6. do the desired task with the *ELPA* library, which could be
a) e%eigenvectors ! solve EV problem with solver as set by "set" method; computes eigenvalues AND eigenvectors
......
elpa/elpa.h
View file @ 899bd7bf
......@@ -7,6 +7,10 @@
struct elpa_struct;
typedef struct elpa_struct *elpa_t;
struct elpa_autotune_struct;
typedef struct elpa_autotune_struct *elpa_autotune_t;
#include <elpa/elpa_constants.h>
#include <elpa/elpa_generated.h>
#include <elpa/elpa_generic.h>
......
elpa/elpa_constants.h.in
View file @ 899bd7bf
......@@ -104,3 +104,21 @@ enum ELPA_CONSTANTS {
ELPA_2STAGE_NUMBER_OF_REAL_KERNELS = (0 ELPA_FOR_ALL_2STAGE_REAL_KERNELS(ELPA_ENUM_SUM)),
};
#define ELPA_FOR_ALL_AUTOTUNE_LEVELS(X, ...) \
X(ELPA_AUTOTUNE_NOT_TUNABLE, 0) \
X(ELPA_AUTOTUNE_FAST, 1) \
X(ELPA_AUTOTUNE_MEDIUM, 2)
enum ELPA_AUTOTUNE_LEVELS {
ELPA_FOR_ALL_AUTOTUNE_LEVELS(ELPA_ENUM_ENTRY)
};
#define ELPA_FOR_ALL_AUTOTUNE_DOMAINS(X, ...) \
X(ELPA_AUTOTUNE_DOMAIN_REAL, 1) \
X(ELPA_AUTOTUNE_DOMAIN_COMPLEX, 2) \
X(ELPA_AUTOTUNE_DOMAIN_ANY, 3)
enum ELPA_AUTOTUNE_DOMAINS {
ELPA_FOR_ALL_AUTOTUNE_DOMAINS(ELPA_ENUM_ENTRY)
};
generate_automake_test_programs.py
View file @ 899bd7bf
......@@ -2,6 +2,11 @@
from __future__ import print_function
from itertools import product
language_flag = {
"Fortran" : "Fortran",
"C" : "C",
}
domain_flag = {
"real" : "-DTEST_REAL",
"complex": "-DTEST_COMPLEX",
......@@ -45,7 +50,8 @@ layout_flag = {
"square" : ""
}
for m, g, q, t, p, d, s, l in product(
for lang, m, g, q, t, p, d, s, l in product(
sorted(language_flag.keys()),
sorted(matrix_flag.keys()),
sorted(gpu_flag.keys()),
sorted(qr_flag.keys()),
......@@ -55,6 +61,11 @@ for m, g, q, t, p, d, s, l in product(
sorted(solver_flag.keys()),
sorted(layout_flag.keys())):
if (lang == "C"):
continue
if (lang == "C" and ( m == "analytic" or l == "all_layouts")):
continue
# exclude some test combinations
# analytic tests only for "eigenvectors" and not on GPU
......@@ -126,30 +137,86 @@ for m, g, q, t, p, d, s, l in product(
raise Exception("Oh no!")
endifs += 1
name = "test_{0}_{1}_{2}_{3}{4}_{5}{6}{7}{8}".format(
d, p, t, s,
"" if kernel == "nokernel" else "_" + kernel,
"gpu_" if g else "",
"qr_" if q else "",
m,
"_all_layouts" if l == "all_layouts" else "")
print("if BUILD_KCOMPUTER")
print("bin_PROGRAMS += " + name)
print("else")
print("noinst_PROGRAMS += " + name)
print("endif")
print("check_SCRIPTS += " + name + ".sh")
print(name + "_SOURCES = test/Fortran/test.F90")
print(name + "_LDADD = $(test_program_ldadd)")
print(name + "_FCFLAGS = $(test_program_fcflags) \\")
print(" -DTEST_CASE=\\\"{0}\\\" \\".format(name))
print(" " + " \\\n ".join([
if (lang == "Fortran"):
name = "test_{0}_{1}_{2}_{3}{4}_{5}{6}{7}{8}".format(
d, p, t, s,
"" if kernel == "nokernel" else "_" + kernel,
"gpu_" if g else "",
"qr_" if q else "",
m,
"_all_layouts" if l == "all_layouts" else "")
print("if BUILD_KCOMPUTER")
print("bin_PROGRAMS += " + name)
print("else")
print("noinst_PROGRAMS += " + name)
print("endif")
print("check_SCRIPTS += " + name + ".sh")
print(name + "_SOURCES = test/Fortran/test.F90")
print(name + "_LDADD = $(test_program_ldadd)")
print(name + "_FCFLAGS = $(test_program_fcflags) \\")
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)
if (lang == "C"):
name = "test_c_version_{0}_{1}_{2}_{3}{4}_{5}{6}{7}{8}".format(
d, p, t, s,
"" if kernel == "nokernel" else "_" + kernel,
"gpu_" if g else "",
"qr_" if q else "",
m,
"_all_layouts" if l == "all_layouts" else "")
print("if BUILD_KCOMPUTER")
print("bin_PROGRAMS += " + name)
print("else")
print("noinst_PROGRAMS += " + name)
print("endif")
print("check_SCRIPTS += " + name + ".sh")
print(name + "_SOURCES = test/C/test.c")
print(name + "_LDADD = $(test_program_ldadd) $(FCLIBS)")
print(name + "_CFLAGS = $(test_program_fcflags) \\")
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 p, d in product(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_{0}_{1}".format(d, p)
print("noinst_PROGRAMS += " + name)
#if (p != "single"):
# print("check_SCRIPTS += " + name + ".sh")
print(name + "_SOURCES = test/Fortran/test_autotune.F90")
print(name + "_LDADD = $(test_program_ldadd)")
print(name + "_FCFLAGS = $(test_program_fcflags) \\")
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)
prec_flag[p]]))
print("endif\n" * endifs)
legacy_test_programs.am
View file @ 899bd7bf
noinst_PROGRAMS += \
real_2stage_c_version@SUFFIX@ \
autotune_c_version@SUFFIX@ \
legacy_real_1stage@SUFFIX@ \
legacy_complex_1stage@SUFFIX@ \
legacy_real_2stage@SUFFIX@ \
......@@ -72,6 +74,14 @@ noinst_PROGRAMS += \
endif
endif
real_2stage_c_version@SUFFIX@_SOURCES = test/C/real_2stage_c_version.c
real_2stage_c_version@SUFFIX@_LDADD = $(test_program_ldadd) $(FCLIBS)
real_2stage_c_version@SUFFIX@_FCFLAGS = $(test_program_fcflags)
autotune_c_version@SUFFIX@_SOURCES = test/C/autotune_c_version.c
autotune_c_version@SUFFIX@_LDADD = $(test_program_ldadd) $(FCLIBS)
autotune_c_version@SUFFIX@_FCFLAGS = $(test_program_fcflags)
legacy_real_1stage_c_version@SUFFIX@_SOURCES = test/C/elpa1/legacy_interface/legacy_real_1stage_c_version.c
legacy_real_1stage_c_version@SUFFIX@_LDADD = $(test_program_ldadd) $(FCLIBS)
legacy_real_1stage_c_version@SUFFIX@_FCFLAGS = $(test_program_fcflags)
......
src/elpa.F90
View file @ 899bd7bf
......@@ -208,14 +208,26 @@ module elpa
obj => elpa_impl_allocate()
end function
!> \brief function to deallocate an ELPA instance
!> Parameters
!> \details
!> \param obj class(elpa_t), pointer : pointer to object to be destroyed and deallocated
!> \param obj class(elpa_t), pointer : pointer to the ELPA object to be destroyed and deallocated
subroutine elpa_deallocate(obj)
class(elpa_t), pointer :: obj
call obj%destroy()
deallocate(obj)
end subroutine
!> \brief function to deallocate an ELPA autotune instance
!> Parameters
!> \details
!> \param obj class(elpa_autotune_t), pointer : pointer to the autotune object to be destroyed and deallocated
subroutine elpa_autotune_deallocate(obj)
class(elpa_autotune_t), pointer :: obj
call obj%destroy()
deallocate(obj)
end subroutine
end module
src/elpa_abstract_impl.F90
View file @ 899bd7bf
......@@ -72,28 +72,13 @@ module elpa_abstract_impl
type, abstract, extends(elpa_t) :: elpa_abstract_impl_t
#ifdef HAVE_DETAILED_TIMINGS
type(timer_t) :: timer
type(timer_t) :: autotune_timer
#else
type(timer_dummy_t) :: timer
#endif
type(c_ptr) :: index = C_NULL_PTR
logical :: eigenvalues_only
contains
! set private fields in the index
generic, public :: set_private => &
elpa_set_private_integer, &
elpa_set_private_double
generic, public :: get_private => &
elpa_get_private_integer, &
elpa_get_private_double
procedure, private :: elpa_set_private_integer
procedure, private :: elpa_set_private_double
procedure, private :: elpa_get_private_integer
procedure, private :: elpa_get_private_double
procedure, public :: elpa_set_integer !< private methods to implement the setting of an integer/double key/value pair
procedure, public :: elpa_set_double
......@@ -110,7 +95,7 @@ module elpa_abstract_impl
!> \param name string, the key
!> \param value integer, the value to be set
!> \result error integer, the error code
subroutine elpa_set_private_integer(self, name, value, error)
subroutine elpa_set_integer(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
......@@ -122,7 +107,7 @@ module elpa_abstract_impl
#endif
integer :: actual_error
actual_error = elpa_index_set_int_value_c(self%index, name // c_null_char, value, 0)
actual_error = elpa_index_set_int_value_c(self%index, name // c_null_char, value)
#ifdef USE_FORTRAN2008
if (present(error)) then
......@@ -143,7 +128,7 @@ module elpa_abstract_impl
!> \param name string, the key
!> \param value integer, the value of the key/vaue pair
!> \param error integer, optional, to store an error code
subroutine elpa_get_private_integer(self, name, value, error)
subroutine elpa_get_integer(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
......@@ -175,18 +160,19 @@ module elpa_abstract_impl
!> \param name string, the key
!> \param value double, the value to be set
!> \result error integer, the error code
subroutine elpa_set_private_double(self, name, value, error)
subroutine elpa_set_double(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
real(kind=c_double), intent(in) :: value
integer :: actual_error
#ifdef USE_FORTRAN2008
integer, optional :: error
#else
integer :: error
#endif
actual_error = elpa_index_set_double_value_c(self%index, name // c_null_char, value, 0)
actual_error = elpa_index_set_double_value_c(self%index, name // c_null_char, value)
#ifdef USE_FORTRAN2008
if (present(error)) then
......@@ -206,7 +192,7 @@ module elpa_abstract_impl
!> \param name string, the key
!> \param value double, the value of the key/vaue pair
!> \param error integer, optional, to store an error code
subroutine elpa_get_private_double(self, name, value, error)
subroutine elpa_get_double(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
......@@ -231,160 +217,4 @@ module elpa_abstract_impl
#endif
end subroutine
subroutine elpa_set_integer(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
integer(kind=c_int), intent(in) :: value
#ifdef USE_FORTRAN2008
integer, optional :: error
#else
integer :: error
#endif
integer :: actual_error
integer :: is_private
is_private = elpa_index_int_is_private_c(name // C_NULL_CHAR)
if (is_private == 0) then
call self%set_private(name, value, error)
else
if (is_private == 1) then
actual_error = ELPA_ERROR_ENTRY_NOT_FOUND
else
actual_error = is_private
endif
#ifdef USE_FORTRAN2008
if (present(error)) then
error = actual_error
else
write(error_unit,'(a)') "ELPA: Error setting option '" // name // "'" // &
" (got: " // elpa_strerr(actual_error) // ") and you did not check for errors!"
endif
#else
error = actual_error
#endif
endif
end subroutine
subroutine elpa_set_double(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
real(kind=c_double), intent(in) :: value
#ifdef USE_FORTRAN2008
integer, optional :: error
#else
integer :: error
#endif
integer :: actual_error
integer :: is_private
is_private = elpa_index_double_is_private_c(name // C_NULL_CHAR)
if (is_private == 0) then
call self%set_private(name, value, error)
else
if (is_private == 1) then
actual_error = ELPA_ERROR_ENTRY_NOT_FOUND
else
actual_error = is_private
endif
#ifdef USE_FORTRAN2008
if (present(error)) then
error = actual_error
else
write(error_unit,'(a)') "ELPA: Error setting option '" // name // "'" // &
" (got: " // elpa_strerr(actual_error) // ") and you did not check for errors!"
endif
#else
error = actual_error
#endif
endif
end subroutine
subroutine elpa_get_integer(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
integer(kind=c_int) :: value
#ifdef USE_FORTRAN2008
integer, optional, intent(out) :: error
#else
integer, intent(out) :: error
#endif
integer :: actual_error
integer :: is_private
is_private = elpa_index_int_is_private_c(name // C_NULL_CHAR)
if (is_private == 0) then
call self%get_private(name, value, error)
else
if (is_private == 1) then
actual_error = ELPA_ERROR_ENTRY_NOT_FOUND
else
actual_error = is_private
endif
#ifdef USE_FORTRAN2008
if (present(error)) then
error = actual_error
else
write(error_unit,'(a)') "ELPA: Error getting option '" // name // "'" // &
" (got: " // elpa_strerr(actual_error) // ") and you did not check for errors!"
endif
#else
error = actual_error
#endif
endif
end subroutine
subroutine elpa_get_double(self, name, value, error)
use iso_c_binding
class(elpa_abstract_impl_t) :: self
character(*), intent(in) :: name
real(kind=c_double) :: value
#ifdef USE_FORTRAN2008
integer, intent(out), optional :: error
#else
integer, intent(out) :: error
#endif
integer :: actual_error
integer :: is_private
is_private = elpa_index_double_is_private_c(name // C_NULL_CHAR)
if (is_private == 0) then
call self%get_private(name, value, error)
else
if (is_private == 1) then
actual_error = ELPA_ERROR_ENTRY_NOT_FOUND
else
actual_error = is_private
endif
#ifdef USE_FORTRAN2008
if (present(error)) then
error = actual_error
else
write(error_unit,'(a)') "ELPA: Error getting option '" // name // "'" // &
" (got: " // elpa_strerr(actual_error) // ") and you did not check for errors!"
endif
#else
error = actual_error
#endif
endif
end subroutine
end module
src/elpa_api.F90
View file @ 899bd7bf
......@@ -96,10 +96,10 @@ module elpa_api
procedure(elpa_can_set_i), deferred, public :: can_set !< method to check whether key/value can be set
! Timer
procedure(elpa_get_time_i), deferred, public :: get_time
procedure(elpa_print_times_i), deferred, public :: print_times
procedure(elpa_timer_start_i), deferred, public :: timer_start
procedure(elpa_timer_stop_i), deferred, public :: timer_stop
procedure(elpa_get_time_i), deferred, public :: get_time !< method to get the times from the timer object
procedure(elpa_print_times_i), deferred, public :: print_times !< method to print the timings tree
procedure(elpa_timer_start_i), deferred, public :: timer_start !< method to start a time measurement
procedure(elpa_timer_stop_i), deferred, public :: timer_stop !< method to stop a time measurement
! Actual math routines
......@@ -133,10 +133,14 @@ module elpa_api
elpa_invert_trm_dc, &
elpa_invert_trm_fc
generic, public :: solve_tridiagonal => & !< method to solve the eigenvalue problem for a tridiagonal
elpa_solve_tridiagonal_d, & !< matrix
generic, public :: solve_tridiagonal => & !< method to solve the eigenvalue problem for a tridiagonal
elpa_solve_tridiagonal_d, & !< matrix
elpa_solve_tridiagonal_f
! 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
!> \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
......@@ -175,6 +179,15 @@ module elpa_api
end type elpa_t
!> \brief Abstract definition of the elpa_autotune type
type, abstract :: elpa_autotune_t
private
contains
procedure(elpa_autotune_destroy_i), deferred, public :: destroy
procedure(elpa_autotune_print_i), deferred, public :: print
end type
!> \brief definition of helper function to get C strlen
!> Parameters
!> \details
......@@ -189,7 +202,8 @@ module elpa_api
end function
end interface
!> \brief abstract definition of setup method
!> \brief abstract definition of the ELPA setup method
!> Parameters
!> \details
!> \param self class(elpa_t): the ELPA object
......@@ -203,6 +217,58 @@ module elpa_api
end function
end interface
!> \brief abstract definition of the autotune setup method
!> Parameters
!> \details
!> \param self class(elpa_t): the ELPA object, which should be tuned
!> \param level integer: the level of "thoroughness" of the tuning steps
!> \param domain integer: domain (real/complex) which should be tuned
!> \result tune_state class(elpa_autotune_t): the autotuning object
abstract interface
function elpa_autotune_setup_i(self, level, domain) result(tune_state)
import elpa_t, elpa_autotune_t
implicit none
class(elpa_t), intent(inout), target :: self
integer, intent(in) :: level, domain
class(elpa_autotune_t), pointer :: tune_state
end function
end interface
!> \brief abstract definition of the autotune step method
!> Parameters
!> \details
!> \param self class(elpa_t): the ELPA object, which should be tuned
!> \param tune_state class(elpa_autotune_t): the autotuning object
!> \param unfinished logical: state whether tuning is unfinished or not
abstract interface
function elpa_autotune_step_i(self, tune_state) result(unfinished)
import elpa_t, elpa_autotune_t
implicit none
class(elpa_t), intent(inout) :: self
class(elpa_autotune_t), intent(inout), target :: tune_state
logical :: unfinished
end function
end interface
!> \brief abstract definition of the autotune set_best method
!> Parameters
!> \details
!> \param self class(elpa_t): the ELPA object, which should be tuned
!> \param tune_state class(elpa_autotune_t): the autotuning object
!> Sets the best combination of ELPA options
abstract interface
subroutine elpa_autotune_set_best_i(self, tune_state)
import elpa_t, elpa_autotune_t
implicit none
class(elpa_t), intent(inout) :: self
class(elpa_autotune_t), intent(in), target :: tune_state
end subroutine
end interface