Merge pull request #33 from fastfloat/some_tuning

Some tuning

Author: lemire

benchmarks/CMakeLists.txt

@@ -3,7 +3,6 @@ add_executable(benchmark
 )
 add_library(benchmark_deps INTERFACE)
 add_library(ieeeToString ieeeToString.cpp)
-target_include_directories(ieeeToString PRIVATE ${ryu_SOURCE_DIR})
 target_link_libraries(benchmark_deps INTERFACE ieeeToString)
 include(CheckSourceCompiles)
 check_source_compiles(CXX "

benchmarks/algorithms.h

@@ -82,7 +82,7 @@ struct BenchArgs {
   std::string name{};
   int (*func)(T, std::span<char>&){};
   bool used{};
-  unsigned char testRepeat{100};
+  size_t testRepeat{100};
 };
 
 template<arithmetic_float T>

benchmarks/benchmark.cpp

@@ -8,6 +8,7 @@
  */
 
 #include "algorithms.h"
+#include <vector>
 #define IEEE_8087
 #include "benchutil.h"
 #include "cxxopts.hpp"
@@ -22,24 +23,37 @@
 #include <string>
 #include <variant>
 #include <fast_float/fast_float.h>
+#include <fmt/core.h>
 
 using Benchmarks::arithmetic_float;
 using Benchmarks::BenchArgs;
 
+bool is_matched(const std::string &str, const std::span<std::string> filter) {
+  if (filter.empty()) {
+    return true;
+  }
+  for (const auto &f : filter) {
+    if (str.find(f) != std::string::npos) {
+      return true;
+    }
+  }
+  return false;
+}
+
 template <arithmetic_float T>
 void evaluateProperties(const std::vector<T> &lines,
-                        const std::array<BenchArgs<T>, Benchmarks::COUNT> &args, const std::string& filter = "") {
+                        const std::array<BenchArgs<T>, Benchmarks::COUNT> &args, const std::span<std::string> filter = {}) {
   constexpr auto precision = std::numeric_limits<T>::digits10;
   fmt::println("{:20} {:20}", "Algorithm", "Valid round-trip");
 
   for (const auto &algo : args) {
     if (!algo.used) {
-      std::cout << "# skipping " << algo.name << std::endl;
+      fmt::println("# skipping {}", algo.name);
       continue;
     }
     // Apply filter if provided
-    if (!filter.empty() && std::string(filter).find(algo.name) == std::string::npos) {
-      std::cout << "# filtered out " << algo.name << std::endl;
+    if (!is_matched(algo.name, filter)) {
+      fmt::println("# filtered out {}", algo.name);
       continue;
     }
     char buf1[100], buf2[100];
@@ -71,17 +85,42 @@ void evaluateProperties(const std::vector<T> &lines,
   }
 }
 
+struct diy_float_t {
+		uint64_t	significand;
+		int							exponent;
+		bool						is_negative;
+};
+
 template <arithmetic_float T>
 void process(const std::vector<T> &lines,
-             const std::array<BenchArgs<T>, Benchmarks::COUNT> &args, const std::string& filter = "") {
+             const std::array<BenchArgs<T>, Benchmarks::COUNT> &args, const std::span<std::string> filter = {}) {
+  // We have a special algorithm for the string generation:
+  std::string just_string = "just_string";
+  if (is_matched(just_string, filter)) {
+    std::vector<diy_float_t> parsed;
+    for(auto d : lines) {
+      auto v = jkj::grisu_exact(d);
+      parsed.emplace_back(v.significand, v.exponent, v.is_negative);
+    }
+    pretty_print(parsed, just_string, [](const std::vector<diy_float_t>& parsed) -> int {
+      int volume = 0;
+      char buf[100];
+      std::span<char> bufspan(buf, sizeof(buf));
+      for (const auto v : parsed)
+        volume +=  to_chars(v.significand, v.exponent, v.is_negative, bufspan.data());
+      return volume;
+    }, 100);
+  } else {
+    fmt::println("# skipping {}", just_string);
+  }
   for (const auto &algo : args) {
     if (!algo.used) {
-      std::cout << "# skipping " << algo.name << std::endl;
+      fmt::println("# skipping {}", algo.name);
       continue;
     }
     // Apply filter if provided
-    if (!filter.empty() && std::string(filter).find(algo.name) == std::string::npos) {
-      std::cout << "# filtered out " << algo.name << std::endl;
+    if (!is_matched(algo.name, filter)) {
+      fmt::println("# filtered out {}", algo.name);
       continue;
     }
     pretty_print(lines, algo.name, [&algo](const std::vector<T> &lines) -> int {
@@ -93,13 +132,14 @@ void process(const std::vector<T> &lines,
       return volume;
     }, algo.testRepeat);
   }
+
 }
 
 template <typename T>
 std::vector<T> fileload(const std::string &filename) {
   std::ifstream inputfile(filename);
   if (!inputfile) {
-    std::cerr << "can't open " << filename << std::endl;
+    fmt::print(stderr, "can't open {}\n", filename);
     return {};
   }
 
@@ -110,24 +150,21 @@ std::vector<T> fileload(const std::string &filename) {
       lines.push_back(std::is_same_v<T, float> ? std::stof(line)
                                                : std::stod(line));
     } catch (...) {
-      std::cerr << "problem with " << line << "\n"
-                << "We expect floating-point numbers (one per line)."
-                << std::endl;
+      fmt::print(stderr, "problem with {}\nWe expect floating-point numbers (one per line).\n", line);
       std::abort();
     }
   }
-  std::cout << "# read " << lines.size() << " lines " << std::endl;
+  fmt::println("# read {} lines", lines.size());
   return lines;
 }
 
 template <typename T>
 std::vector<T> get_random_numbers(size_t howmany,
                                   const std::string &random_model) {
-  std::cout << "# parsing random numbers" << std::endl;
+  fmt::println("# parsing random numbers");
   std::vector<T> lines;
   auto g = get_generator_by_name<T>(random_model);
-  std::cout << "model: " << g->describe() << "\n"
-            << "volume: " << howmany << " floats" << std::endl;
+  fmt::print("model: {}\nvolume: {} floats\n", g->describe(), howmany);
   lines.reserve(howmany); // let us reserve plenty of memory.
   for (size_t i = 0; i < howmany; i++) {
     const T line = g->new_float();
@@ -155,20 +192,21 @@ int main(int argc, char **argv) {
         cxxopts::value<bool>()->default_value("false"))
         ("e,errol", "Enable errol3 (current impl. returns invalid values, e.g., for 0).",
         cxxopts::value<bool>()->default_value("false"))
-        ("a,algo-filter", "Filter algorithms by name substring.",
-          cxxopts::value<std::string>()->default_value(""))
+        ("a,algo-filter", "Filter algorithms by name substring: you can use multiple filters separated by commas.",
+        cxxopts::value<std::vector<std::string>>()->default_value(""))
+        ("r,repeat", "Force a number of repetitions.",
+        cxxopts::value<size_t>()->default_value("0"))
         ("h,help", "Print usage.");
     const auto result = options.parse(argc, argv);
 
     if (result["help"].as<bool>()) {
-      std::cout << options.help() << std::endl;
+      fmt::print("{}\n", options.help());
       return EXIT_SUCCESS;
     }
-
+    const size_t repeat = result["repeat"].as<size_t>();
     const bool single = result["single"].as<bool>();
-    const std::string filter = result["algo-filter"].as<std::string>();
-    std::cout << "number type: binary"
-              << (single ? "32 (float)" : "64 (double)") << std::endl;
+    std::vector<std::string> filter = result["algo-filter"].as<std::vector<std::string>>();
+    fmt::println("number type: binary{}", (single ? "32 (float)" : "64 (double)"));
 
     std::variant<std::vector<float>, std::vector<double>> numbers;
     const auto filename = result["file"].as<std::string>();
@@ -179,9 +217,7 @@ int main(int argc, char **argv) {
         numbers = get_random_numbers<float>(volume, model);
       else
         numbers = get_random_numbers<double>(volume, model);
-      std::cout << "# You can also provide a filename (with the -f flag): "
-                   "it should contain one string per line corresponding to a number"
-                << std::endl;
+      fmt::println("# You can also provide a filename (with the -f flag): it should contain one string per line corresponding to a number");
     }
     else {
       if (single)
@@ -198,6 +234,14 @@ int main(int argc, char **argv) {
     else
       algorithms = Benchmarks::initArgs<double>(errol);
 
+    if(repeat > 0) {
+      fmt::println("# forcing repeat count to {}", repeat);
+      std::visit([repeat](auto &args) {
+        for (auto &arg : args)
+          arg.testRepeat = repeat;
+      }, algorithms);
+    }
+
     const bool test = result["test"].as<bool>();
     std::visit([test,&filter](const auto &lines, const auto &args) {
       using T1 = typename std::decay_t<decltype(lines)>::value_type;
@@ -210,7 +254,20 @@ int main(int argc, char **argv) {
       }
     }, numbers, algorithms);
   } catch (const std::exception &e) {
-    std::cout << "error parsing options: " << e.what() << std::endl;
+    fmt::println("Error parsing options: {}", e.what());
+    fmt::println("\nUSAGE GUIDE:");
+    fmt::println("  ./benchmark [OPTIONS]");
+    fmt::println("\nCOMMAND SUMMARY:");
+    fmt::println("  The benchmark tool evaluates the performance of different floating-point to string");
+    fmt::println("  conversion algorithms. It can use either synthetic data or a file containing");
+    fmt::println("  floating-point numbers (one per line).");
+    fmt::println("\nEXAMPLES:");
+    fmt::println("  ./benchmark --single                    # Run benchmark with single precision (float)");
+    fmt::println("  ./benchmark --file=data/canada.txt      # Run benchmark using numbers from a file");
+    fmt::println("  ./benchmark --test                      # Test correctness instead of performance");
+    fmt::println("  ./benchmark --volume=1000 --model=uniform # Generate 1000 uniform random numbers");
+    fmt::println("  ./benchmark --algo-filter=ryu,grisu     # Only test algorithms containing 'ryu' or 'grisu'");
+    fmt::println("\nFor full options list, run: ./benchmark --help");
     return EXIT_FAILURE;
   }
 }

benchmarks/exhaustivefloat32.cpp

@@ -8,6 +8,7 @@
 #include <iostream>
 #include <string_view>
 #include <charconv>
+#include <vector>
 
 #include "algorithms.h"
 #include "cxxopts.hpp"
@@ -47,19 +48,19 @@ std::optional<float> parse_float(std::string_view sv) {
   float result;
   const char* begin = sv.data();
   const char* end = sv.data() + sv.size();
-  
+
   auto [ptr, ec] = std::from_chars(begin, end, result);
-  
+
   // Check if parsing succeeded and consumed the entire string
   if (ec == std::errc{} && ptr == end) {
       return result;
   }
-  
+
   // Return nullopt if parsing failed or didn't consume all input
   return std::nullopt;
 }
 
-void run_exhaustive32(bool errol) {
+void run_exhaustive32(bool errol, const std::vector<std::string>& algo_filter = {}) {
   constexpr auto precision = std::numeric_limits<float>::digits10;
   fmt::println("{:20} {:20}", "Algorithm", "Valid shortest serialization");
 
@@ -75,6 +76,22 @@ void run_exhaustive32(bool errol) {
       fmt::print("# skipping {} because it is the reference.\n", algo.name);
       continue;
     }
+
+    // Apply filter if provided
+    if (!algo_filter.empty()) {
+      bool matched = false;
+      for (const auto &f : algo_filter) {
+        if (algo.name.find(f) != std::string::npos) {
+          matched = true;
+          break;
+        }
+      }
+      if (!matched) {
+        fmt::print("# filtered out {}\n", algo.name);
+        continue;
+      }
+    }
+
     bool incorrect = false;
     char buf1[100], buf2[100];
     std::span<char> bufRef(buf1, sizeof(buf1)), bufAlgo(buf2, sizeof(buf2));
@@ -149,15 +166,21 @@ int main(int argc, char **argv) {
     options.add_options()(
         "e,errol",
         "Enable errol3 (current impl. returns invalid values, e.g., for 0).",
-        cxxopts::value<bool>()->default_value("false"))("h,help",
-                                                        "Print usage.");
+        cxxopts::value<bool>()->default_value("false"))(
+        "a,algorithm",
+        "Specify which algorithm(s) to test (comma-separated).",
+        cxxopts::value<std::vector<std::string>>()->default_value({}))(
+        "h,help",
+        "Print usage.");
     const auto result = options.parse(argc, argv);
 
     if (result["help"].as<bool>()) {
       fmt::print("{}\n", options.help());
       return EXIT_SUCCESS;
     }
-    run_exhaustive32(result["errol"].as<bool>());
+
+    auto algo_filter = result["algorithm"].as<std::vector<std::string>>();
+    run_exhaustive32(result["errol"].as<bool>(), algo_filter);
   } catch (const std::exception &e) {
     fmt::print("error parsing options: {}\n", e.what());
     return EXIT_FAILURE;