diff --git a/bin/donkey_benchmark_compilation_wrapper.sh b/bin/donkey_benchmark_compilation_wrapper.sh
index e2691b5d71db2473b731fabcea6a1095ec2516d9..a786d111264ef0e67769100713f1776e3e002c3d 100755
--- a/bin/donkey_benchmark_compilation_wrapper.sh
+++ b/bin/donkey_benchmark_compilation_wrapper.sh
@@ -13,7 +13,4 @@ ml parmetis
("$@")
code=$?
-echo "Code: $code"
-sleep 0.1s
-
exit $code
diff --git a/bin/donkey_benchmark_execution_wrapper.py b/bin/donkey_benchmark_execution_wrapper.py
index 7951b8b06bfdeb217bd3328c9b3b88f229bbc606..d383963318291ae947c70b1f083c320a60250303 100755
--- a/bin/donkey_benchmark_execution_wrapper.py
+++ b/bin/donkey_benchmark_execution_wrapper.py
@@ -13,8 +13,3 @@ ret = subprocess.call(command)
# If that failed - fail!
if ret != 0:
sys.exit(ret)
-
-# If that was succesful, wait for the output file to be available on the filesystem
-# This step is necessary because the NFS synchronization is too slow for our workflow.
-while not os.path.isfile(sys.argv[2]):
- time.sleep(0.1)
diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt
index b96bd06cbce6361f1130e17bb7cb4d97efe08ec7..8881504f934e7b1d755580ffc46d47f2a1467346 100644
--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -25,5 +25,5 @@ add_executable(_autotune_target EXCLUDE_FROM_ALL _autotune.cc)
target_compile_options(_autotune_target PUBLIC -fno-strict-aliasing)
if(benchmark_FOUND)
- target_link_libraries(_autotune_target benchmark::benchmark)
+ target_link_libraries(_autotune_target benchmark)
endif()
diff --git a/python/dune/codegen/sumfact/autotune.py b/python/dune/codegen/sumfact/autotune.py
index d19fd7be5587258e8db1f3ada3301b1e7017f95e..a1d7442addc4e2d1d720b1cc81d8a914c1f75532 100644
--- a/python/dune/codegen/sumfact/autotune.py
+++ b/python/dune/codegen/sumfact/autotune.py
@@ -14,6 +14,7 @@ import re
import subprocess
import filelock
import hashlib
+import time
def get_cmake_cache_entry(entry):
@@ -94,10 +95,8 @@ def write_global_data(sf, filename):
def write_setup_code(sf, filename, define_thetas=True):
- opcounting = get_option("opcounter")
with open(filename, "a") as f:
# Setup a polynomial object (normally done in the LocalOperator members)
- set_option("opcounter", False)
from dune.codegen.loopy.target import type_floatingpoint
real = type_floatingpoint()
f.write(" using RF = {};\n".format(real))
@@ -116,11 +115,12 @@ def write_setup_code(sf, filename, define_thetas=True):
constructor_knl = lp.get_one_scheduled_kernel(constructor_knl)
# Allocate buffers
+ alignment = get_option("max_vector_width") // 8
size = max(product(m.quadrature_size for m in sf.matrix_sequence_quadrature_permuted) * sf.vector_width,
product(m.basis_size for m in sf.matrix_sequence_quadrature_permuted) * sf.vector_width)
size = int(size * (get_option("precision_bits") / 8))
- f.writelines([" char buffer0[{}] __attribute__ ((aligned (32)));\n".format(size),
- " char buffer1[{}] __attribute__ ((aligned (32)));\n".format(size),
+ f.writelines([" char buffer0[{}] __attribute__ ((aligned ({})));\n".format(size, alignment),
+ " char buffer1[{}] __attribute__ ((aligned ({})));\n".format(size, alignment),
])
# Setup fastdg inputs
@@ -129,7 +129,7 @@ def write_setup_code(sf, filename, define_thetas=True):
f.write("{} = 0;\n".format(arg))
else:
size = sf.interface.fastdg_interface_object_size
- f.write(" RF {}[{}] __attribute__ ((aligned (32)));\n".format(arg.split()[-1], size))
+ f.write(" RF {}[{}] __attribute__ ((aligned ({})));\n".format(arg.split()[-1], size, alignment))
# Write stuff into the input buffer
f.writelines([" {0} *input = ({0} *)buffer0;\n".format(real),
@@ -172,12 +172,15 @@ def write_setup_code(sf, filename, define_thetas=True):
" rng.seed(42);\n",
" std::uniform_int_distribution<> dis(0, {});\n".format(size / (get_option("precision_bits") / 8)),
])
- return opcounting
def generate_standalone_code_google_benchmark(sf, filename):
delete_cache_items("kernel_default")
+ # Turn off opcounting
+ opcounting = get_option("opcounter")
+ set_option("opcounter", False)
+
# Extract sum factorization kernel
from dune.codegen.pdelab.localoperator import extract_kernel_from_cache
knl = realize_sumfact_kernel_function(sf)
@@ -205,6 +208,10 @@ def generate_standalone_code_google_benchmark(sf, filename):
write_global_data(sf, filename)
with open(filename, "a") as f:
+ arguments = ', '.join(sf.interface.signature_args)
+ if len(arguments) > 0:
+ arguments = ', ' + arguments
+ arguments = 'const char* buffer0, const char* buffer1' + arguments
f.write("void sumfact_kernel({})\n".format(arguments))
for line in knl.member.lines[1:]:
f.write("{}\n".format(line))
@@ -212,7 +219,7 @@ def generate_standalone_code_google_benchmark(sf, filename):
f.write("\n\n")
f.write("static void BM_sumfact_kernel(benchmark::State& state){\n")
- opcounting = write_setup_code(sf, filename, define_thetas=False)
+ write_setup_code(sf, filename, define_thetas=False)
additional_arguments = [i.split()[-1] for i in sf.interface.signature_args]
additional_arguments = ', '.join(additional_arguments)
@@ -227,12 +234,18 @@ def generate_standalone_code_google_benchmark(sf, filename):
"\n",
"BENCHMARK_MAIN();"
])
+
+ # Maybe turn opcounting on again
set_option("opcounter", opcounting)
def generate_standalone_code(sf, filename):
delete_cache_items("kernel_default")
+ # Turn off opcounting
+ opcounting = get_option("opcounter")
+ set_option("opcounter", False)
+
# Extract sum factorization kernel
from dune.codegen.pdelab.localoperator import extract_kernel_from_cache
knl = realize_sumfact_kernel_function(sf)
@@ -257,7 +270,7 @@ def generate_standalone_code(sf, filename):
"{\n",
])
- opcounting = write_setup_code(sf, filename)
+ write_setup_code(sf, filename)
# Write measurement
with open(filename, "a") as f:
@@ -283,7 +296,9 @@ def generate_standalone_code(sf, filename):
" std::cout << accum;\n",
"}\n",
])
- set_option("opcounter", opcounting)
+
+ # Maybe turn opcounting on again
+ set_option("opcounter", opcounting)
def autotune_realization(sf):
@@ -320,6 +335,10 @@ def autotune_realization(sf):
if ret != 0:
raise CodegenAutotuneError("Compilation of autotune executable failed. Invocation: {}".format(" ".join(call)))
+ # File system synchronization!
+ while not os.path.exists(executable):
+ time.sleep(0.01)
+
# Check whether the user specified an execution wrapper
call = []
wrapper = get_cmake_cache_entry("DUNE_CODEGEN_BENCHMARK_EXECUTION_WRAPPER")
@@ -337,11 +356,19 @@ def autotune_realization(sf):
if ret != 0:
raise CodegenAutotuneError("Execution of autotune benchmark failed. Invocation: {}".format(" ".join(call)))
+ # File system synchronization!
+ while not os.path.exists(logname):
+ time.sleep(0.01)
+
# Extract the result form the log file
if get_option("autotune_google_benchmark"):
import json
with open(logname) as json_file:
- data = json.load(json_file)
- return data['benchmarks'][0]['cpu_time']
+ try:
+ data = json.load(json_file)
+ return data['benchmarks'][0]['cpu_time']
+ except Exception as e:
+ print("Error while loading file {}".format(logname))
+ raise e
else:
return float(next(iter(open(logname, "r")))) / 1000000