diff --git a/cmake/modules/GeneratedSystemtests.cmake b/cmake/modules/GeneratedSystemtests.cmake
index 2e3de48d33d9939aed33ffbec4d90035f50eb42b..1803b1e508e43188c22ee1b846bd3d42abf0211f 100644
--- a/cmake/modules/GeneratedSystemtests.cmake
+++ b/cmake/modules/GeneratedSystemtests.cmake
@@ -49,7 +49,7 @@ function(dune_add_formcompiler_system_test)
add_generated_executable(TARGET ${tname}
UFLFILE ${SYSTEMTEST_UFLFILE}
- FORM_COMPILER_ARGS --ini-file ${inifile} --timer
+ FORM_COMPILER_ARGS --ini-file ${inifile}
)
# Exclude the target from all
diff --git a/python/dune/perftool/options.py b/python/dune/perftool/options.py
index 5f27ede95ccb273d089f231fc64083cb0c854d51..b521bf6dc61b99c0c614116fada26f476926e51a 100644
--- a/python/dune/perftool/options.py
+++ b/python/dune/perftool/options.py
@@ -39,8 +39,9 @@ def get_form_compiler_arguments():
parser.add_argument("--diagonal-transformation-matrix", action="store_true", help="set option if the jacobian of the transformation is diagonal (axiparallel grids)")
parser.add_argument("--constant-transformation-matrix", action="store_true", help="set option if the jacobian of the transformation is constant on a cell")
parser.add_argument("--ini-file", type=str, help="An inifile to use. A generated driver will be hard-coded to it, a [formcompiler] section will be used as default values to form compiler arguments (use snake case)")
- parser.add_argument("--timer", action="store_true", help="measure times")
- parser.add_argument("--opcounter", action="store_true", default=False, help="Count operations. Should only be used with yaspgrid. Timer should be set.")
+ parser.add_argument("--opcounter", action="store_true", help="Count operations. Note: In this case only oparor applications are generated since solving and operator counting does not work.")
+ parser.add_argument("--time-opcounter", action="store_true", help="Generate opcounter codepath. Can be used for timing opcounter programs without setting the opcounter option.")
+ parser.add_argument("--instrumentation-level", type=int, default=0, help="Control time/opcounter measurements. 0-do nothing, 1-measure program as a whole, 2-operator applications, 3-measure kernel (eg. alpha-volume, ...), 4-parts of kernel (eg. stage 1-3 of SF)")
parser.add_argument("--project-basedir", type=str, help="The base (build) directory of the dune-perftool project")
parser.add_argument("--fastdg", action="store_true", help="Use FastDGGridOperator from PDELab.")
# TODO at some point this help description should be updated
diff --git a/python/dune/perftool/pdelab/driver.py b/python/dune/perftool/pdelab/driver.py
index 62a72fcea2c6ebc82596be76fa17f1d4d19dc15c..775b3bda292210ab142d4c37dcfde16cf3852fe5 100644
--- a/python/dune/perftool/pdelab/driver.py
+++ b/python/dune/perftool/pdelab/driver.py
@@ -1188,10 +1188,10 @@ def dune_solve():
snp = name_stationarynonlinearproblemsolver(go_type, go)
solve = "{}.apply();".format(snp)
- if get_option('timer'):
+ if get_option('instrumentation_level') >= 2:
setup_timer()
from dune.perftool.generation import post_include
- post_include("HP_DECLARE_TIMER(total);", filetag="driver")
+ post_include("HP_DECLARE_TIMER(solve);", filetag="driver")
# Print times after solving
from dune.perftool.generation import get_global_context_value
@@ -1202,12 +1202,14 @@ def dune_solve():
lop_name = name_localoperator(formdata)
timestream = name_timing_stream()
print_times.append("{}.dump_timers({}, argv[0], true);".format(lop_name, timestream))
- solve = ["HP_TIMER_START(total);",
+
+ solve = ["HP_TIMER_START(solve);",
"{}".format(solve),
- "HP_TIMER_STOP(total);",
- "DUMP_TIMER(total, {}, true);".format(timestream),
+ "HP_TIMER_STOP(solve);",
+ "DUMP_TIMER(solve, {}, true);".format(timestream),
]
- solve.extend(print_times)
+ if get_option('instrumentation_level') >= 3:
+ solve.extend(print_times)
return solve
@@ -1391,7 +1393,7 @@ def name_test_fail_variable():
@cached
def setup_timer():
- assert(get_option('timer'))
+ assert(get_option('instrumentation_level') >= 1)
# Necessary includes and defines
from dune.perftool.generation import pre_include
@@ -1697,7 +1699,9 @@ def generate_driver(formdatas, data):
set_driver_data(formdatas, data)
# Entrypoint for driver generation
- if get_option("opcounter"):
+ if get_option("opcounter") or get_option("time_opcounter"):
+ if get_option("time_opcounter"):
+ assert(not get_option("opcounter"))
assert(any(_driver_data['form'].ufl_cell().cellname() in x for x in
["vertex", "interval", "quadrilateral", "hexahedron"]))
# In case of operator conunting we only assemble the matrix and evaluate the residual
@@ -1714,7 +1718,18 @@ def generate_driver(formdatas, data):
from dune.perftool.generation import retrieve_cache_items
from cgen import FunctionDeclaration, FunctionBody, Block, Value
driver_signature = FunctionDeclaration(Value('bool', 'driver'), [Value('int', 'argc'), Value('char**', 'argv')])
- driver_body = Block(contents=[i for i in retrieve_cache_items("preamble", make_generable=True)])
+ contents = [i for i in retrieve_cache_items("preamble", make_generable=True)]
+
+ from cgen import Line
+ if get_option("instrumentation_level") >= 1:
+ from dune.perftool.generation import post_include
+ post_include("HP_DECLARE_TIMER(driver);\n", filetag="driver")
+ contents.insert(0, Line(text="HP_TIMER_START(driver);\n"))
+ contents.insert(len(contents) - 1, Line(text="HP_TIMER_STOP(driver);\n"))
+ timestream = name_timing_stream()
+ contents.insert(len(contents) - 1, Line(text="DUMP_TIMER(driver, {}, true);\n".format(timestream)))
+ contents.insert(0, Line(text="\n"))
+ driver_body = Block(contents)
driver = FunctionBody(driver_signature, driver_body)
filename = get_option("driver_file")
diff --git a/python/dune/perftool/pdelab/localoperator.py b/python/dune/perftool/pdelab/localoperator.py
index 0686f44c49001424ef3cb819c9df134243ab5b18..839c57f8f5bef06cd85a241a283684ab87311f4b 100644
--- a/python/dune/perftool/pdelab/localoperator.py
+++ b/python/dune/perftool/pdelab/localoperator.py
@@ -556,7 +556,6 @@ class TimerMethod(ClassMember):
knl = name_example_kernel()
assert(knl is not None)
- # TODO: operator counting only works if alpha_volume_kernel exists
content = ["template <typename Stream>",
"void dump_timers(Stream& {}, char* {}, bool {})".format(os, ex, reset),
"{",
@@ -595,7 +594,7 @@ class LoopyKernelMethod(ClassMember):
content.append(' ' + p)
# Start timer
- if add_timings and get_option('timer'):
+ if add_timings and get_option('instrumentation_level') >= 3:
from dune.perftool.pdelab.signatures import assembler_routine_name
timer_name = assembler_routine_name() + '_kernel'
name_example_kernel(name=timer_name)
@@ -607,7 +606,7 @@ class LoopyKernelMethod(ClassMember):
content.extend(l for l in generate_body(kernel).split('\n')[1:-1])
# Stop timer
- if add_timings and get_option('timer'):
+ if add_timings and get_option('instrumentation_level') >= 3:
content.append(' ' + 'HP_TIMER_STOP({});'.format(timer_name))
content.append('}')
@@ -814,7 +813,7 @@ def generate_localoperator_file(formdata, kernels, filename):
for k in kernels.values():
operator_methods.extend(k)
- if get_option('timer'):
+ if get_option('instrumentation_level') >= 3:
include_file('dune/perftool/common/timer.hh', filetag='operatorfile')
operator_methods.append(TimerMethod())
diff --git a/python/dune/perftool/sumfact/sumfact.py b/python/dune/perftool/sumfact/sumfact.py
index 35d91cc0d553f803988c8c6e932aed18144c6337..216ce95c0f7a06ab99cedd7210f01b740b69999b 100644
--- a/python/dune/perftool/sumfact/sumfact.py
+++ b/python/dune/perftool/sumfact/sumfact.py
@@ -10,6 +10,7 @@ from dune.perftool.generation import (backend,
barrier,
built_instruction,
domain,
+ dump_accumulate_timer,
function_mangler,
generator_factory,
get_counter,
@@ -17,6 +18,7 @@ from dune.perftool.generation import (backend,
globalarg,
iname,
instruction,
+ post_include,
kernel_cached,
retrieve_cache_items,
silenced_warning,
@@ -32,6 +34,7 @@ from dune.perftool.loopy.buffer import (get_buffer_temporary,
from dune.perftool.sumfact.quadrature import nest_quadrature_loops
from dune.perftool.pdelab.localoperator import determine_accumulation_space
from dune.perftool.pdelab.restriction import restricted_name
+from dune.perftool.pdelab.signatures import assembler_routine_name
from dune.perftool.pdelab.spaces import (name_lfs,
name_lfs_bound,
)
@@ -178,6 +181,16 @@ def generate_accumulation_instruction(visitor, accterm, measure, subdomain_id):
replace_dict[Variable(iname)] = i
expression = substitute(pymbolic_expr, replace_dict)
+ # Write timing stuff for jacobian (for alpha methods it is done at the end of stage 1)
+ timer_dep = frozenset()
+ if get_option("instrumentation_level") >= 4:
+ timer_name = assembler_routine_name() + '_kernel' + '_quadratureloop'
+ post_include('HP_DECLARE_TIMER({});'.format(timer_name), filetag='operatorfile')
+ dump_accumulate_timer(timer_name)
+ if(visitor.inames):
+ timer_dep = frozenset({instruction(code="HP_TIMER_START({});".format(timer_name),
+ within_inames=frozenset(visitor.inames))})
+
# Determine dependencies
from loopy.match import Or, Writes
from loopy.symbolic import DependencyMapper
@@ -192,12 +205,17 @@ def generate_accumulation_instruction(visitor, accterm, measure, subdomain_id):
forced_iname_deps=frozenset(quadrature_inames() + visitor.inames),
forced_iname_deps_is_final=True,
tags=frozenset({"quadvec"}).union(vectag),
- depends_on=frozenset({deps})
+ depends_on=frozenset({deps}).union(timer_dep)
)
if insn_dep is None:
insn_dep = frozenset({contrib_dep})
+ if get_option("instrumentation_level") >= 4:
+ insn_dep = frozenset({instruction(code="HP_TIMER_STOP({});".format(timer_name),
+ depends_on=insn_dep,
+ within_inames=frozenset(visitor.inames))})
+
# Add a sum factorization kernel that implements the multiplication
# with the test function (stage 3)
pref_pos = i if accterm.argument.index else None
@@ -351,6 +369,14 @@ def sum_factorization_kernel(a_matrices,
ctags = ctags + ",vec"
vec_shape = (4,)
+ if get_option("instrumentation_level") >= 4:
+ timer_name = assembler_routine_name() + '_kernel' + '_stage{}'.format(stage)
+ post_include('HP_DECLARE_TIMER({});'.format(timer_name), filetag='operatorfile')
+ dump_accumulate_timer(timer_name)
+ insn_dep = frozenset({instruction(code="HP_TIMER_START({});".format(timer_name),
+ depends_on=insn_dep,
+ within_inames=additional_inames)})
+
insn_dep = frozenset({barrier(depends_on=insn_dep,
within_inames=additional_inames,
)})
@@ -432,6 +458,17 @@ def sum_factorization_kernel(a_matrices,
)
})
+ if get_option("instrumentation_level") >= 4:
+ insn_dep = frozenset({instruction(code="HP_TIMER_STOP({});".format(timer_name),
+ depends_on=insn_dep,
+ within_inames=additional_inames)})
+ if stage == 1:
+ qp_timer_name = assembler_routine_name() + '_kernel' + '_quadratureloop'
+ post_include('HP_DECLARE_TIMER({});'.format(timer_name), filetag='operatorfile')
+ dump_accumulate_timer(timer_name)
+ insn_dep = instruction(code="HP_TIMER_START({});".format(qp_timer_name),
+ depends_on=insn_dep)
+
if outshape is None:
assert stage == 3
outshape = tuple(mat.rows for mat in a_matrices)