diff --git a/python/dune/codegen/options.py b/python/dune/codegen/options.py
index 7455f1bb242fa042011f005e154e8088264ad47f..257da40169f578a4b64a367f6aaa4cd55729253d 100644
--- a/python/dune/codegen/options.py
+++ b/python/dune/codegen/options.py
@@ -55,6 +55,7 @@ class CodegenGlobalOptionsArray(ImmutableRecord):
target_name = CodegenOption(default=None, helpstr="The target name from CMake")
operator_to_build = CodegenOption(default=None, helpstr="The operators from the list that is about to be build now. CMake sets this one!!!")
debug_interpolate_input = CodegenOption(default=False, helpstr="Should the input for printresidual and printmatix be interpolated (instead of random input).")
+ use_likwid = CodegenOption(default=False, helpstr="Use likwid instead of own performance measurements.")
autotune_google_benchmark = CodegenOption(default=False, helpstr="Use google-benchmark library for autotuning (when autotuning is activated).")
# Arguments that are mainly to be set by logic depending on other options
diff --git a/python/dune/codegen/pdelab/driver/__init__.py b/python/dune/codegen/pdelab/driver/__init__.py
index 9e94c6fb55cf9cf23a4a0bc89b069f48f296115f..c33d3acebfe27686f35055801f1282dc1d2b1450 100644
--- a/python/dune/codegen/pdelab/driver/__init__.py
+++ b/python/dune/codegen/pdelab/driver/__init__.py
@@ -269,10 +269,9 @@ def generate_driver():
compare_L2_squared()
# Make sure that timestream is declared before retrieving chache items
- if get_option("instrumentation_level") >= 1:
- from dune.codegen.pdelab.driver.timings import setup_timer, name_timing_stream
+ if get_option("instrumentation_level") >= 1 and not get_option("use_likwid"):
+ from dune.codegen.pdelab.driver.timings import setup_timer
setup_timer()
- timestream = name_timing_stream()
from dune.codegen.pdelab.driver.error import return_statement
return_statement()
@@ -304,10 +303,13 @@ def generate_driver():
add_section("instat", "Set up instationary stuff...")
add_section("printing", "Maybe print residuals and matrices to stdout...")
add_section("error", "Maybe calculate errors for test results...")
+ add_section("end", "Stuff that should happen at the end...")
add_section("return_stmt", "Return statement...")
- if get_option("instrumentation_level") >= 1:
+ if get_option("instrumentation_level") >= 1 and not get_option("use_likwid"):
from dune.codegen.generation import post_include
+ from dune.codegen.pdelab.driver.timings import name_timing_stream
+ timestream = name_timing_stream()
post_include("HP_DECLARE_TIMER(driver);\n", filetag="driver")
contents.insert(0, Line(text="HP_TIMER_START(driver);\n"))
contents.insert(len(contents) - 2, Line(text="HP_TIMER_STOP(driver);\n"))
diff --git a/python/dune/codegen/pdelab/driver/timings.py b/python/dune/codegen/pdelab/driver/timings.py
index c6241c852b54379a9c39588542aebdd81b520cd1..a617b64ba9c767750050152e0eba3e167bd5fb30 100644
--- a/python/dune/codegen/pdelab/driver/timings.py
+++ b/python/dune/codegen/pdelab/driver/timings.py
@@ -109,29 +109,62 @@ def name_jacobian(form_ident):
return name
+@preamble(section="init")
+def init_likwid():
+ return ["LIKWID_MARKER_INIT;", "LIKWID_MARKER_THREADINIT;"]
+
+
+@preamble(section="end")
+def finalize_likwid():
+ return ["LIKWID_MARKER_CLOSE;"]
+
+
@cached
def setup_timer():
# TODO check that we are using YASP?
- if get_option('opcounter'):
- pre_include("#define ENABLE_COUNTER", filetag="driver")
- pre_include("#define ENABLE_HP_TIMERS", filetag="driver")
- include_file("dune/codegen/common/timer.hh", filetag="driver")
+ if get_option("use_likwid"):
+ pre_include("#define LIKWID_PERFMON", filetag="driver")
+ include_file("likwid.h", filetag="driver")
+ init_likwid()
+ finalize_likwid()
+ else:
+ from dune.codegen.loopy.target import type_floatingpoint
+ pre_include("#define HP_TIMER_OPCOUNTER {}".format(type_floatingpoint()), filetag="driver")
+ if get_option('opcounter'):
+ pre_include("#define ENABLE_COUNTER", filetag="driver")
+ pre_include("#define ENABLE_HP_TIMERS", filetag="driver")
+ include_file("dune/codegen/common/timer.hh", filetag="driver")
+
+
+
+@preamble(section="init")
+def init_likwid_timer(region):
+ return ["LIKWID_MARKER_REGISTER(\"{}\");".format(region)]
def init_region_timer(region):
setup_timer()
- from dune.codegen.generation import post_include
- post_include("HP_DECLARE_TIMER({});".format(region), filetag="driver")
+ if get_option("use_likwid"):
+ init_likwid_timer(region)
+ else:
+ from dune.codegen.generation import post_include
+ post_include("HP_DECLARE_TIMER({});".format(region), filetag="driver")
def start_region_timer(region):
- return ["HP_TIMER_START({});".format(region)]
+ if get_option("use_likwid"):
+ return ["LIKWID_MARKER_START(\"{}\");".format(region)]
+ else:
+ return ["HP_TIMER_START({});".format(region)]
def stop_region_timer(region):
- timestream = name_timing_stream()
- return ["HP_TIMER_STOP({});".format(region),
- "DUMP_TIMER({}, {}, {}, true);".format(get_option("instrumentation_level"), region, timestream)]
+ if get_option("use_likwid"):
+ return ["LIKWID_MARKER_STOP(\"{}\");".format(region)]
+ else:
+ timestream = name_timing_stream()
+ return ["HP_TIMER_STOP({});".format(region),
+ "DUMP_TIMER({}, {}, {}, true);".format(get_option("instrumentation_level"), region, timestream)]
def timed_region(region, actions):
@@ -140,21 +173,24 @@ def timed_region(region, actions):
assert(isinstance(actions, list))
- assembly = []
- print_times = []
+ if get_option('instrumentation_level') >= 2:
+ assembly = []
+ print_times = []
- init_region_timer(region)
+ init_region_timer(region)
- if get_option('instrumentation_level') >= 3:
- timestream = name_timing_stream()
- lop_name = name_localoperator(get_form_ident())
- print_times.append("{}.dump_timers({}, {}, true);".format(lop_name, timestream, name_timing_identifier()))
+ if get_option('instrumentation_level') >= 3:
+ timestream = name_timing_stream()
+ lop_name = name_localoperator(get_form_ident())
+ print_times.append("{}.dump_timers({}, {}, true);".format(lop_name, timestream, name_timing_identifier()))
- assembly += start_region_timer(region)
- assembly += actions
- assembly += stop_region_timer(region)
+ assembly += start_region_timer(region)
+ assembly += actions
+ assembly += stop_region_timer(region)
- return assembly + print_times
+ return assembly + print_times
+ else:
+ return actions
@preamble(section="timings")
@@ -165,10 +201,7 @@ def evaluate_residual_timer():
action = "{}.residual({}, {});".format(n_go, v, r)
- if get_option('instrumentation_level') >= 2:
- return timed_region("residual_evaluation", action)
- else:
- return action
+ return timed_region("residual_evaluation", action)
@preamble(section="timings")
@@ -185,10 +218,7 @@ def apply_jacobian_timer():
j1 = name_temporary_vector("j1", form)
action = "{}.nonlinear_jacobian_apply({}, {}, {});".format(n_go, v, j0, j1)
- if get_option('instrumentation_level') >= 2:
- return timed_region("apply_jacobian", action)
- else:
- return action
+ return timed_region("apply_jacobian", action)
@preamble(section="timings")
@@ -199,7 +229,4 @@ def assemble_matrix_timer():
action = "{}.jacobian({},{});".format(n_go, v, m)
- if get_option('instrumentation_level') >= 2:
- return timed_region("matrix_assembly", action)
- else:
- return [action]
+ return timed_region("matrix_assembly", action)