This commit is contained in:
Ivan
2022-04-05 11:42:28 +03:00
commit 6dc0eb0fcf
5565 changed files with 1200500 additions and 0 deletions

View File

@@ -0,0 +1,3 @@
exports_files(glob(["*.BUILD"]))
exports_files(["build_defs.bzl"])

View File

@@ -0,0 +1,38 @@
load("//bindings/python:build_defs.bzl", "py_extension")
py_library(
name = "google_benchmark",
srcs = ["__init__.py"],
visibility = ["//visibility:public"],
deps = [
":_benchmark",
# pip; absl:app
],
)
py_extension(
name = "_benchmark",
srcs = ["benchmark.cc"],
copts = [
"-fexceptions",
"-fno-strict-aliasing",
],
features = ["-use_header_modules"],
deps = [
"//:benchmark",
"@pybind11",
"@python_headers",
],
)
py_test(
name = "example",
srcs = ["example.py"],
python_version = "PY3",
srcs_version = "PY3",
visibility = ["//visibility:public"],
deps = [
":google_benchmark",
],
)

View File

@@ -0,0 +1,158 @@
# Copyright 2020 Google Inc. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Python benchmarking utilities.
Example usage:
import google_benchmark as benchmark
@benchmark.register
def my_benchmark(state):
... # Code executed outside `while` loop is not timed.
while state:
... # Code executed within `while` loop is timed.
if __name__ == '__main__':
benchmark.main()
"""
from absl import app
from google_benchmark import _benchmark
from google_benchmark._benchmark import (
Counter,
kNanosecond,
kMicrosecond,
kMillisecond,
kSecond,
oNone,
o1,
oN,
oNSquared,
oNCubed,
oLogN,
oNLogN,
oAuto,
oLambda,
)
__all__ = [
"register",
"main",
"Counter",
"kNanosecond",
"kMicrosecond",
"kMillisecond",
"kSecond",
"oNone",
"o1",
"oN",
"oNSquared",
"oNCubed",
"oLogN",
"oNLogN",
"oAuto",
"oLambda",
]
__version__ = "0.2.0"
class __OptionMaker:
"""A stateless class to collect benchmark options.
Collect all decorator calls like @option.range(start=0, limit=1<<5).
"""
class Options:
"""Pure data class to store options calls, along with the benchmarked function."""
def __init__(self, func):
self.func = func
self.builder_calls = []
@classmethod
def make(cls, func_or_options):
"""Make Options from Options or the benchmarked function."""
if isinstance(func_or_options, cls.Options):
return func_or_options
return cls.Options(func_or_options)
def __getattr__(self, builder_name):
"""Append option call in the Options."""
# The function that get returned on @option.range(start=0, limit=1<<5).
def __builder_method(*args, **kwargs):
# The decorator that get called, either with the benchmared function
# or the previous Options
def __decorator(func_or_options):
options = self.make(func_or_options)
options.builder_calls.append((builder_name, args, kwargs))
# The decorator returns Options so it is not technically a decorator
# and needs a final call to @regiser
return options
return __decorator
return __builder_method
# Alias for nicer API.
# We have to instantiate an object, even if stateless, to be able to use __getattr__
# on option.range
option = __OptionMaker()
def register(undefined=None, *, name=None):
"""Register function for benchmarking."""
if undefined is None:
# Decorator is called without parenthesis so we return a decorator
return lambda f: register(f, name=name)
# We have either the function to benchmark (simple case) or an instance of Options
# (@option._ case).
options = __OptionMaker.make(undefined)
if name is None:
name = options.func.__name__
# We register the benchmark and reproduce all the @option._ calls onto the
# benchmark builder pattern
benchmark = _benchmark.RegisterBenchmark(name, options.func)
for name, args, kwargs in options.builder_calls[::-1]:
getattr(benchmark, name)(*args, **kwargs)
# return the benchmarked function because the decorator does not modify it
return options.func
def _flags_parser(argv):
argv = _benchmark.Initialize(argv)
return app.parse_flags_with_usage(argv)
def _run_benchmarks(argv):
if len(argv) > 1:
raise app.UsageError("Too many command-line arguments.")
return _benchmark.RunSpecifiedBenchmarks()
def main(argv=None):
return app.run(_run_benchmarks, argv=argv, flags_parser=_flags_parser)
# Methods for use with custom main function.
initialize = _benchmark.Initialize
run_benchmarks = _benchmark.RunSpecifiedBenchmarks

View File

@@ -0,0 +1,181 @@
// Benchmark for Python.
#include <map>
#include <string>
#include <vector>
#include "pybind11/operators.h"
#include "pybind11/pybind11.h"
#include "pybind11/stl.h"
#include "pybind11/stl_bind.h"
#include "benchmark/benchmark.h"
PYBIND11_MAKE_OPAQUE(benchmark::UserCounters);
namespace {
namespace py = ::pybind11;
std::vector<std::string> Initialize(const std::vector<std::string>& argv) {
// The `argv` pointers here become invalid when this function returns, but
// benchmark holds the pointer to `argv[0]`. We create a static copy of it
// so it persists, and replace the pointer below.
static std::string executable_name(argv[0]);
std::vector<char*> ptrs;
ptrs.reserve(argv.size());
for (auto& arg : argv) {
ptrs.push_back(const_cast<char*>(arg.c_str()));
}
ptrs[0] = const_cast<char*>(executable_name.c_str());
int argc = static_cast<int>(argv.size());
benchmark::Initialize(&argc, ptrs.data());
std::vector<std::string> remaining_argv;
remaining_argv.reserve(argc);
for (int i = 0; i < argc; ++i) {
remaining_argv.emplace_back(ptrs[i]);
}
return remaining_argv;
}
benchmark::internal::Benchmark* RegisterBenchmark(const char* name,
py::function f) {
return benchmark::RegisterBenchmark(
name, [f](benchmark::State& state) { f(&state); });
}
PYBIND11_MODULE(_benchmark, m) {
using benchmark::TimeUnit;
py::enum_<TimeUnit>(m, "TimeUnit")
.value("kNanosecond", TimeUnit::kNanosecond)
.value("kMicrosecond", TimeUnit::kMicrosecond)
.value("kMillisecond", TimeUnit::kMillisecond)
.value("kSecond", TimeUnit::kSecond)
.export_values();
using benchmark::BigO;
py::enum_<BigO>(m, "BigO")
.value("oNone", BigO::oNone)
.value("o1", BigO::o1)
.value("oN", BigO::oN)
.value("oNSquared", BigO::oNSquared)
.value("oNCubed", BigO::oNCubed)
.value("oLogN", BigO::oLogN)
.value("oNLogN", BigO::oLogN)
.value("oAuto", BigO::oAuto)
.value("oLambda", BigO::oLambda)
.export_values();
using benchmark::internal::Benchmark;
py::class_<Benchmark>(m, "Benchmark")
// For methods returning a pointer tor the current object, reference
// return policy is used to ask pybind not to take ownership oof the
// returned object and avoid calling delete on it.
// https://pybind11.readthedocs.io/en/stable/advanced/functions.html#return-value-policies
//
// For methods taking a const std::vector<...>&, a copy is created
// because a it is bound to a Python list.
// https://pybind11.readthedocs.io/en/stable/advanced/cast/stl.html
.def("unit", &Benchmark::Unit, py::return_value_policy::reference)
.def("arg", &Benchmark::Arg, py::return_value_policy::reference)
.def("args", &Benchmark::Args, py::return_value_policy::reference)
.def("range", &Benchmark::Range, py::return_value_policy::reference,
py::arg("start"), py::arg("limit"))
.def("dense_range", &Benchmark::DenseRange,
py::return_value_policy::reference, py::arg("start"),
py::arg("limit"), py::arg("step") = 1)
.def("ranges", &Benchmark::Ranges, py::return_value_policy::reference)
.def("args_product", &Benchmark::ArgsProduct,
py::return_value_policy::reference)
.def("arg_name", &Benchmark::ArgName, py::return_value_policy::reference)
.def("arg_names", &Benchmark::ArgNames,
py::return_value_policy::reference)
.def("range_pair", &Benchmark::RangePair,
py::return_value_policy::reference, py::arg("lo1"), py::arg("hi1"),
py::arg("lo2"), py::arg("hi2"))
.def("range_multiplier", &Benchmark::RangeMultiplier,
py::return_value_policy::reference)
.def("min_time", &Benchmark::MinTime, py::return_value_policy::reference)
.def("iterations", &Benchmark::Iterations,
py::return_value_policy::reference)
.def("repetitions", &Benchmark::Repetitions,
py::return_value_policy::reference)
.def("report_aggregates_only", &Benchmark::ReportAggregatesOnly,
py::return_value_policy::reference, py::arg("value") = true)
.def("display_aggregates_only", &Benchmark::DisplayAggregatesOnly,
py::return_value_policy::reference, py::arg("value") = true)
.def("measure_process_cpu_time", &Benchmark::MeasureProcessCPUTime,
py::return_value_policy::reference)
.def("use_real_time", &Benchmark::UseRealTime,
py::return_value_policy::reference)
.def("use_manual_time", &Benchmark::UseManualTime,
py::return_value_policy::reference)
.def(
"complexity",
(Benchmark * (Benchmark::*)(benchmark::BigO)) & Benchmark::Complexity,
py::return_value_policy::reference,
py::arg("complexity") = benchmark::oAuto);
using benchmark::Counter;
py::class_<Counter> py_counter(m, "Counter");
py::enum_<Counter::Flags>(py_counter, "Flags")
.value("kDefaults", Counter::Flags::kDefaults)
.value("kIsRate", Counter::Flags::kIsRate)
.value("kAvgThreads", Counter::Flags::kAvgThreads)
.value("kAvgThreadsRate", Counter::Flags::kAvgThreadsRate)
.value("kIsIterationInvariant", Counter::Flags::kIsIterationInvariant)
.value("kIsIterationInvariantRate",
Counter::Flags::kIsIterationInvariantRate)
.value("kAvgIterations", Counter::Flags::kAvgIterations)
.value("kAvgIterationsRate", Counter::Flags::kAvgIterationsRate)
.value("kInvert", Counter::Flags::kInvert)
.export_values()
.def(py::self | py::self);
py::enum_<Counter::OneK>(py_counter, "OneK")
.value("kIs1000", Counter::OneK::kIs1000)
.value("kIs1024", Counter::OneK::kIs1024)
.export_values();
py_counter
.def(py::init<double, Counter::Flags, Counter::OneK>(),
py::arg("value") = 0., py::arg("flags") = Counter::kDefaults,
py::arg("k") = Counter::kIs1000)
.def(py::init([](double value) { return Counter(value); }))
.def_readwrite("value", &Counter::value)
.def_readwrite("flags", &Counter::flags)
.def_readwrite("oneK", &Counter::oneK);
py::implicitly_convertible<py::float_, Counter>();
py::implicitly_convertible<py::int_, Counter>();
py::bind_map<benchmark::UserCounters>(m, "UserCounters");
using benchmark::State;
py::class_<State>(m, "State")
.def("__bool__", &State::KeepRunning)
.def_property_readonly("keep_running", &State::KeepRunning)
.def("pause_timing", &State::PauseTiming)
.def("resume_timing", &State::ResumeTiming)
.def("skip_with_error", &State::SkipWithError)
.def_property_readonly("error_occurred", &State::error_occurred)
.def("set_iteration_time", &State::SetIterationTime)
.def_property("bytes_processed", &State::bytes_processed,
&State::SetBytesProcessed)
.def_property("complexity_n", &State::complexity_length_n,
&State::SetComplexityN)
.def_property("items_processed", &State::items_processed,
&State::SetItemsProcessed)
.def("set_label", (void(State::*)(const char*)) & State::SetLabel)
.def("range", &State::range, py::arg("pos") = 0)
.def_property_readonly("iterations", &State::iterations)
.def_readwrite("counters", &State::counters)
.def_property_readonly("thread_index", &State::thread_index)
.def_property_readonly("threads", &State::threads);
m.def("Initialize", Initialize);
m.def("RegisterBenchmark", RegisterBenchmark,
py::return_value_policy::reference);
m.def("RunSpecifiedBenchmarks",
[]() { benchmark::RunSpecifiedBenchmarks(); });
};
} // namespace

View File

@@ -0,0 +1,136 @@
# Copyright 2020 Google Inc. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Example of Python using C++ benchmark framework.
To run this example, you must first install the `google_benchmark` Python package.
To install using `setup.py`, download and extract the `google_benchmark` source.
In the extracted directory, execute:
python setup.py install
"""
import random
import time
import google_benchmark as benchmark
from google_benchmark import Counter
@benchmark.register
def empty(state):
while state:
pass
@benchmark.register
def sum_million(state):
while state:
sum(range(1_000_000))
@benchmark.register
def pause_timing(state):
"""Pause timing every iteration."""
while state:
# Construct a list of random ints every iteration without timing it
state.pause_timing()
random_list = [random.randint(0, 100) for _ in range(100)]
state.resume_timing()
# Time the in place sorting algorithm
random_list.sort()
@benchmark.register
def skipped(state):
if True: # Test some predicate here.
state.skip_with_error("some error")
return # NOTE: You must explicitly return, or benchmark will continue.
... # Benchmark code would be here.
@benchmark.register
def manual_timing(state):
while state:
# Manually count Python CPU time
start = time.perf_counter() # perf_counter_ns() in Python 3.7+
# Something to benchmark
time.sleep(0.01)
end = time.perf_counter()
state.set_iteration_time(end - start)
@benchmark.register
def custom_counters(state):
"""Collect cutom metric using benchmark.Counter."""
num_foo = 0.0
while state:
# Benchmark some code here
pass
# Collect some custom metric named foo
num_foo += 0.13
# Automatic Counter from numbers.
state.counters["foo"] = num_foo
# Set a counter as a rate.
state.counters["foo_rate"] = Counter(num_foo, Counter.kIsRate)
# Set a counter as an inverse of rate.
state.counters["foo_inv_rate"] = Counter(num_foo, Counter.kIsRate | Counter.kInvert)
# Set a counter as a thread-average quantity.
state.counters["foo_avg"] = Counter(num_foo, Counter.kAvgThreads)
# There's also a combined flag:
state.counters["foo_avg_rate"] = Counter(num_foo, Counter.kAvgThreadsRate)
@benchmark.register
@benchmark.option.measure_process_cpu_time()
@benchmark.option.use_real_time()
def with_options(state):
while state:
sum(range(1_000_000))
@benchmark.register(name="sum_million_microseconds")
@benchmark.option.unit(benchmark.kMicrosecond)
def with_options2(state):
while state:
sum(range(1_000_000))
@benchmark.register
@benchmark.option.arg(100)
@benchmark.option.arg(1000)
def passing_argument(state):
while state:
sum(range(state.range(0)))
@benchmark.register
@benchmark.option.range(8, limit=8 << 10)
def using_range(state):
while state:
sum(range(state.range(0)))
@benchmark.register
@benchmark.option.range_multiplier(2)
@benchmark.option.range(1 << 10, 1 << 18)
@benchmark.option.complexity(benchmark.oN)
def computing_complexity(state):
while state:
sum(range(state.range(0)))
state.complexity_n = state.range(0)
if __name__ == "__main__":
benchmark.main()

View File

@@ -0,0 +1,20 @@
cc_library(
name = "pybind11",
hdrs = glob(
include = [
"include/pybind11/*.h",
"include/pybind11/detail/*.h",
],
exclude = [
"include/pybind11/common.h",
"include/pybind11/eigen.h",
],
),
copts = [
"-fexceptions",
"-Wno-undefined-inline",
"-Wno-pragma-once-outside-header",
],
includes = ["include"],
visibility = ["//visibility:public"],
)

View File

@@ -0,0 +1,6 @@
cc_library(
name = "python_headers",
hdrs = glob(["**/*.h"]),
includes = ["."],
visibility = ["//visibility:public"],
)

View File

@@ -0,0 +1,2 @@
absl-py>=0.7.1