add micro benchmarking project, setup number parsing benchmark

relating to previous work #422 Results are matching what was observed at the time ^^ was being worked on std::from_chars is included as the target to beat, but since only MSVC has it for floating point it's not hugely useful yet uniform real distribution is probably a horrible choice, and it might be good to randomise the number of sf in each string also (currently the y all end up at max length) Run on (4 X 3500 MHz CPU s) CPU Caches: L1 Data 32K (x4) L1 Instruction 32K (x4) L2 Unified 262K (x4) L3 Unified 6291K (x1) ---------------------------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ---------------------------------------------------------------------------------------------------------- RandFloats/double_from_string_sstream 804 ns 820 ns 896000 RandFloats/double_from_string_strtod 163 ns 162 ns 5973333 RandFloats/double_from_string_strtod_fixed 175 ns 172 ns 5352107 RandFloats/double_from_string_strtod_fixed_const_ref 150 ns 152 ns 5352107 RandFloats/double_from_string_std_from_chars 87.1 ns 88.3 ns 9557333 RandFloatsComma/double_from_string_strtod_fixed_comma_ref 172 ns 173 ns 5146257 RandFloatsComma/double_from_string_strtod_fixed_comma_const_ref 180 ns 175 ns 5352107
tfussell · Feb 29, 2020 · 0adb8a6 · 0adb8a6
1 parent ae6f9d2
commit 0adb8a6
Show file tree

Hide file tree

Showing 4 changed files with 262 additions and 0 deletions.
diff --git a/benchmarks/CMakeLists.txt b/benchmarks/CMakeLists.txt
@@ -41,3 +41,8 @@ foreach(BENCHMARK_SOURCE IN ITEMS ${BENCHMARK_SOURCES})
       $<TARGET_FILE_DIR:${BENCHMARK_EXECUTABLE}>)
   endif()
 endforeach()
+
+option(XLNT_MICROBENCH_ENABLED "Enable small benchmarks typically used for development" OFF)
+if (XLNT_MICROBENCH_ENABLED)
+	add_subdirectory(microbenchmarks)
+endif()
diff --git a/benchmarks/microbenchmarks/CMakeLists.txt b/benchmarks/microbenchmarks/CMakeLists.txt
@@ -0,0 +1,36 @@
+# FetchContent added in cmake v3.11
+# https://cmake.org/cmake/help/v3.11/module/FetchContent.html
+# this file is behind a feature flag (XLNT_MICROBENCH_ENABLED) so the primary build is not affected
+cmake_minimum_required(VERSION 3.11)
+project(xlnt_ubench)
+
+# acquire google benchmark dependency
+# disable generation of the various test projects
+set(BENCHMARK_ENABLE_TESTING OFF)
+# gtest not required
+set(BENCHMARK_ENABLE_GTEST_TESTS OFF)
+
+include(FetchContent)
+FetchContent_Declare(
+	googlebenchmark
+	GIT_REPOSITORY 	https://github.com/google/benchmark
+	GIT_TAG			v1.5.0
+)
+# download if not already present
+FetchContent_GetProperties(googlebenchmark)
+if(NOT googlebenchmark_POPULATED)
+	FetchContent_Populate(googlebenchmark)
+	add_subdirectory(${googlebenchmark_SOURCE_DIR} ${googlebenchmark_BINARY_DIR})
+endif()
+# equivalent of add_subdirectory, now available for use
+FetchContent_MakeAvailable(googlebenchmark)
+
+
+add_executable(xlnt_ubench)
+target_sources(xlnt_ubench
+	PRIVATE
+		string_to_double.cpp
+		double_to_string.cpp
+)
+target_link_libraries(xlnt_ubench benchmark_main xlnt)
+target_compile_features(xlnt_ubench PRIVATE cxx_std_17)
diff --git a/benchmarks/microbenchmarks/double_to_string.cpp b/benchmarks/microbenchmarks/double_to_string.cpp
@@ -0,0 +1,2 @@
+#include "benchmark/benchmark.h"
+
diff --git a/benchmarks/microbenchmarks/string_to_double.cpp b/benchmarks/microbenchmarks/string_to_double.cpp
@@ -0,0 +1,219 @@
+// A core part of the xlsx parsing routine is taking strings from the xml parser and parsing these to a double
+// this has a few requirements
+// - expect numbers in the form 1234.56 (i.e. no thousands seperator, '.' used for the decimal seperator)
+// - handles atleast 15 significant figures (excel only serialises numbers up to 15sf)
+
+#include <benchmark/benchmark.h>
+#include <locale>
+#include <random>
+
+// setup a large quantity of random doubles as strings
+template <bool Decimal_Locale = true>
+class RandomFloats : public benchmark::Fixture
+{
+    static constexpr size_t Number_of_Elements = 1 << 20;
+    static_assert(Number_of_Elements > 1'000'000, "ensure a decent set of random values is generated");
+
+    std::vector<std::string> inputs;
+
+    size_t index = 0;
+    const char *locale_str;
+
+public:
+    void SetUp(const ::benchmark::State &state)
+    {
+        if (Decimal_Locale)
+        {
+            locale_str = setlocale(LC_ALL, "C");
+        }
+        else
+        {
+            locale_str = setlocale(LC_ALL, "de-DE");
+        }
+        std::random_device rd; // obtain a seed for the random number engine
+        std::mt19937 gen(rd());
+        // doing full range is stupid (<double>::min/max()...), it just ends up generating very large numbers
+        // uniform is probably not the best distribution to use here, but it will do for now
+        std::uniform_real_distribution<double> dis(-1'000, 1'000);
+        // generate a large quantity of doubles to deserialise
+        inputs.reserve(Number_of_Elements);
+        for (int i = 0; i < Number_of_Elements; ++i)
+        {
+            double d = dis(gen);
+            char buf[16];
+            snprintf(buf, 16, "%.15f", d);
+            inputs.push_back(std::string(buf));
+        }
+    }
+
+    void TearDown(const ::benchmark::State &state)
+    {
+        // restore locale
+        setlocale(LC_ALL, locale_str);
+        // gbench is keeping the fixtures alive somewhere, need to clear the data...
+        inputs = std::vector<std::string>{};
+    }
+
+    std::string &get_rand()
+    {
+        return inputs[++index & Number_of_Elements];
+    }
+};
+
+// method used by xlsx_consumer.cpp in commit - ba01de47a7d430764c20ec9ac9600eec0eb38bcf
+// std::istringstream with the locale set to "C"
+#include <sstream>
+struct number_converter
+{
+    number_converter()
+    {
+        stream.imbue(std::locale("C"));
+    }
+
+    double stold(const std::string &s)
+    {
+        stream.str(s);
+        stream.clear();
+        stream >> result;
+        return result;
+    }
+
+    std::istringstream stream;
+    double result;
+};
+
+using RandFloats = RandomFloats<true>;
+
+BENCHMARK_F(RandFloats, double_from_string_sstream)
+(benchmark::State &state)
+{
+    number_converter converter;
+    while (state.KeepRunning())
+    {
+        benchmark::DoNotOptimize(
+            converter.stold(get_rand()));
+    }
+}
+
+// using strotod
+// https://en.cppreference.com/w/cpp/string/byte/strtof
+// this naive usage is broken in the face of locales (fails condition 1)
+#include <cstdlib>
+BENCHMARK_F(RandFloats, double_from_string_strtod)
+(benchmark::State &state)
+{
+    while (state.KeepRunning())
+    {
+        benchmark::DoNotOptimize(
+            strtod(get_rand().c_str(), nullptr));
+    }
+}
+
+// to resolve the locale issue with strtod, a little preprocessing of the input is required
+struct number_converter_mk2
+{
+    explicit number_converter_mk2()
+        : should_convert_to_comma(std::use_facet<std::numpunct<char>>(std::locale{}).decimal_point() == ',')
+    {
+    }
+
+    double stold(std::string &s) const noexcept
+    {
+        assert(!s.empty());
+        if (should_convert_to_comma)
+        {
+            auto decimal_pt = std::find(s.begin(), s.end(), '.');
+            if (decimal_pt != s.end())
+            {
+                *decimal_pt = ',';
+            }
+        }
+        return strtod(s.c_str(), nullptr);
+    }
+
+    double stold(const std::string &s) const
+    {
+        assert(!s.empty());
+        if (!should_convert_to_comma)
+        {
+            return strtod(s.c_str(), nullptr);
+        }
+        std::string copy(s);
+        auto decimal_pt = std::find(copy.begin(), copy.end(), '.');
+        if (decimal_pt != copy.end())
+        {
+            *decimal_pt = ',';
+        }
+        return strtod(copy.c_str(), nullptr);
+    }
+
+private:
+    bool should_convert_to_comma = false;
+};
+
+BENCHMARK_F(RandFloats, double_from_string_strtod_fixed)
+(benchmark::State &state)
+{
+    number_converter_mk2 converter;
+    while (state.KeepRunning())
+    {
+        benchmark::DoNotOptimize(
+            converter.stold(get_rand()));
+    }
+}
+
+BENCHMARK_F(RandFloats, double_from_string_strtod_fixed_const_ref)
+(benchmark::State &state)
+{
+    number_converter_mk2 converter;
+    while (state.KeepRunning())
+    {
+        const std::string &inp = get_rand();
+        benchmark::DoNotOptimize(
+            converter.stold(inp));
+    }
+}
+
+// locale names are different between OS's, and std::from_chars is only complete in MSVC
+#ifdef _MSC_VER
+
+#include <charconv>
+BENCHMARK_F(RandFloats, double_from_string_std_from_chars)
+(benchmark::State &state)
+{
+    while (state.KeepRunning())
+    {
+        const std::string &input = get_rand();
+        double output;
+        benchmark::DoNotOptimize(
+            std::from_chars(input.data(), input.data() + input.size(), output));
+    }
+}
+
+// not using the standard "C" locale with '.' seperator
+// german locale uses ',' as the seperator
+using RandFloatsComma = RandomFloats<false>;
+BENCHMARK_F(RandFloatsComma, double_from_string_strtod_fixed_comma_ref)
+(benchmark::State &state)
+{
+    number_converter_mk2 converter;
+    while (state.KeepRunning())
+    {
+        benchmark::DoNotOptimize(
+            converter.stold(get_rand()));
+    }
+}
+
+BENCHMARK_F(RandFloatsComma, double_from_string_strtod_fixed_comma_const_ref)
+(benchmark::State &state)
+{
+    number_converter_mk2 converter;
+    while (state.KeepRunning())
+    {
+        const std::string &inp = get_rand();
+        benchmark::DoNotOptimize(
+            converter.stold(inp));
+    }
+}
+
+#endif