Merge pull request #10 from stevenewald/bench

Added benchmarks and ARM NEON support

CMakeLists.txt

@@ -14,8 +14,8 @@ include(cmake/project-is-top-level.cmake)
 include(cmake/variables.cmake)
 
 set(CMAKE_INTERPROCEDURAL_OPTIMIZATION ON)
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=x86-64 -mavx512f -mavx512dq -mavx512vl -mavx512bf16")
-add_compile_options(-fno-inline -fno-omit-frame-pointer)
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
+# add_compile_options(-fno-inline -fno-omit-frame-pointer)
 
 
 # ---- Declare library ----
@@ -27,17 +27,37 @@ add_library(
     source/mandelbrot/mandelbrot_window.cpp
     source/graphics/color_conversions/color_conversions.cpp
     source/graphics/aspect_ratio/aspect_ratio.cpp
-    source/mandelbrot/equations_simd.cpp
-    source/mandelbrot/equations.cpp
+    source/units/coordinates.cpp
 )
 
+include(CheckCXXCompilerFlag)
+check_cxx_compiler_flag("-mavx512f -mavx512dq -mavx512vl -mavx512bf16" HAS_AVX512)
+check_cxx_source_compiles("
+        #include <arm_neon.h>
+        int main() {
+            float32x4_t vec = vdupq_n_f32(0.0f);
+            return 0;
+		}" HAS_NEON)
+
+if (HAS_ALL_AVX512)
+    message(STATUS "AVX-512 is supported by the compiler.")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -mavx512dq -mavx512vl -mavx512bf16")
+	target_sources(fractal-generator_lib PRIVATE source/mandelbrot/equations_simd.cpp)
+elseif(HAS_NEON)
+	message(STATUS "ARM NEON is supported by the compiler.")
+	target_sources(fractal-generator_lib PRIVATE source/mandelbrot/equations_neon.cpp)
+else()
+	message(STATUS "SIMD is not fully supported by the compiler. SIMD will not be enabled.")
+	target_sources(fractal-generator_lib PRIVATE source/mandelbrot/equations_compat.cpp)
+endif()
+
 target_include_directories(
     fractal-generator_lib ${warning_guard}
     PUBLIC
     "\$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/source>"
 )
 
-target_compile_features(fractal-generator_lib PUBLIC cxx_std_20)
+target_compile_features(fractal-generator_lib PUBLIC cxx_std_23)
 
 find_package(fmt REQUIRED)
 find_package(argparse REQUIRED)
@@ -55,8 +75,9 @@ add_executable(fractal-generator::exe ALIAS fractal-generator_exe)
 
 set_property(TARGET fractal-generator_exe PROPERTY OUTPUT_NAME fractal-generator)
 
-target_compile_features(fractal-generator_exe PRIVATE cxx_std_20)
+target_compile_features(fractal-generator_exe PRIVATE cxx_std_23)
 
+target_link_libraries(fractal-generator_exe PRIVATE fmt::fmt)
 target_link_libraries(fractal-generator_exe PRIVATE fractal-generator_lib)
 target_link_libraries(fractal-generator_exe PRIVATE argparse::argparse)
 target_link_libraries(fractal-generator_exe PRIVATE sfml-graphics)

CMakePresets.json

@@ -50,7 +50,7 @@
       "hidden": true,
       "cacheVariables": {
         "CMAKE_CXX_EXTENSIONS": "OFF",
-        "CMAKE_CXX_STANDARD": "20",
+        "CMAKE_CXX_STANDARD": "23",
         "CMAKE_CXX_STANDARD_REQUIRED": "ON"
       }
     },
@@ -59,7 +59,7 @@
       "description": "These flags are supported by both GCC and Clang",
       "hidden": true,
       "cacheVariables": {
-        "CMAKE_CXX_FLAGS": "-D_GLIBCXX_ASSERTIONS=1 -fstack-protector-strong -fcf-protection=full -fstack-clash-protection -Wall -Wextra -Wpedantic -Wconversion -Wsign-conversion -Wcast-qual -Wformat=2 -Wundef -Werror=float-equal -Wshadow -Wcast-align -Wunused -Wnull-dereference -Wdouble-promotion -Wimplicit-fallthrough -Wextra-semi -Woverloaded-virtual -Wnon-virtual-dtor -Wold-style-cast",
+        "CMAKE_CXX_FLAGS": "-D_GLIBCXX_ASSERTIONS=1 -Wall -Wextra -Wpedantic -Wconversion -Wsign-conversion -Wcast-qual -Wformat=2 -Wundef -Werror=float-equal -Wshadow -Wcast-align -Wunused -Wnull-dereference -Wdouble-promotion -Wimplicit-fallthrough -Wextra-semi -Woverloaded-virtual -Wnon-virtual-dtor -Wold-style-cast",
         "CMAKE_EXE_LINKER_FLAGS": "-Wl,--allow-shlib-undefined,--as-needed,-z,noexecstack,-z,relro,-z,now,-z,nodlopen",
         "CMAKE_SHARED_LINKER_FLAGS": "-Wl,--allow-shlib-undefined,--as-needed,-z,noexecstack,-z,relro,-z,now,-z,nodlopen"
       }
@@ -68,7 +68,7 @@
       "name": "flags-appleclang",
       "hidden": true,
       "cacheVariables": {
-        "CMAKE_CXX_FLAGS": "-fstack-protector-strong -Wall -Wextra -Wpedantic -Wconversion -Wsign-conversion -Wcast-qual -Wformat=2 -Wundef -Werror=float-equal -Wshadow -Wcast-align -Wunused -Wnull-dereference -Wdouble-promotion -Wimplicit-fallthrough -Wextra-semi -Woverloaded-virtual -Wnon-virtual-dtor -Wold-style-cast"
+        "CMAKE_CXX_FLAGS": "-Wall -Wextra -Wpedantic -Wconversion -Wsign-conversion -Wcast-qual -Wformat=2 -Wundef -Werror=float-equal -Wshadow -Wcast-align -Wunused -Wnull-dereference -Wdouble-promotion -Wimplicit-fallthrough -Wextra-semi -Woverloaded-virtual -Wnon-virtual-dtor -Wold-style-cast"
       }
     },
     {

Taskfile.yml

@@ -2,8 +2,8 @@ version: '3'
 
 vars:
   NAME: fractal-generator
-  RELEASE_BUILD: "false"
-  RELEASE_PRESET: "ci-ubuntu"
+  RELEASE_BUILD: "true"
+  RELEASE_PRESET: "ci-macos"
 
 tasks:
   deps:
@@ -14,12 +14,11 @@ tasks:
 
   init:
     dir: .
-    vars:
-      PRESET: '{{if eq .RELEASE_BUILD "true"}}{{.RELEASE_PRESET}}{{else}}dev{{end}}'
     preconditions:
       - test -f CMakeUserPresets.json
     cmds:
-      - cmake --preset={{.PRESET}}
+      - cmake --preset=dev
+      - cmake --preset={{.RELEASE_PRESET}} -DCMAKE_BUILD_TYPE=Release
 
 
   build:
@@ -57,6 +56,12 @@ tasks:
       - task: build
       - ctest {{.TEST_FLAGS}}
 
+  bench:
+    dir: .
+    cmds:
+      - task: build
+      - ./build/benchmark/fractal_benchmarks --benchmark_repetitions=5 --benchmark_min_warmup_time=0.5 --benchmark_report_aggregates_only=true
+
   fmt:
     dir: .
     cmds:

benchmark/CMakeLists.txt

@@ -0,0 +1,37 @@
+project(FractalBenchmarks LANGUAGES CXX)
+if(NOT CMAKE_BUILD_TYPE)
+  set(CMAKE_BUILD_TYPE Debug)
+endif()
+# ---- Dependencies ----
+
+find_package(benchmark REQUIRED)
+
+# ---- Benchmarks ----
+add_executable(fractal_benchmarks
+  source/bench.cpp
+  source/main.cpp
+)
+
+target_include_directories(
+  fractal_benchmarks ${warning_guard}
+  PUBLIC
+  "$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/source>"
+)
+
+target_link_libraries(
+  fractal_benchmarks PRIVATE
+  fractal-generator_lib
+  sfml-graphics
+  benchmark::benchmark
+)
+
+# Enable lto for release builds
+include(CheckIPOSupported)
+check_ipo_supported(RESULT lto_supported OUTPUT error)
+if(lto_supported)
+  set_target_properties(fractal_benchmarks PROPERTIES INTERPROCEDURAL_OPTIMIZATION_RELEASE TRUE)
+else()
+    message(WARNING "LTO is not supported: ${error}")
+endif()
+
+target_compile_features(fractal_benchmarks PRIVATE cxx_std_23)

benchmark/source/bench.cpp

@@ -0,0 +1,35 @@
+#include "config.hpp"
+#include "units/coordinates.hpp"
+#include "graphics/display_to_complex.hpp"
+#include "mandelbrot/equations.hpp"
+#include "mandelbrot/equations.hpp"
+#include "units/units.hpp"
+
+#include <benchmark/benchmark.h>
+
+using namespace fractal;
+
+static void BM_GenerateMandelbrotSimd(benchmark::State& state)
+{
+    DisplayDomain display = {
+        {0,   0  },
+        {800, 800}
+    };
+    complex_domain complex = START_COMPLEX_DOMAIN;
+    avx512_complex start{};
+    DisplayToComplexCoordinates t{display, complex};
+
+    std::int64_t prox = 0;
+    for (auto _ : state) {
+        for (auto it = display.begin(); it != display.end(); it += 8) {
+            benchmark::DoNotOptimize(
+                compute_iterations(start, t.to_complex_projections(*it), 100)
+            );
+        }
+        prox += display.size();
+    }
+    state.SetItemsProcessed(prox);
+}
+
+BENCHMARK(BM_GenerateMandelbrotSimd)->Arg(0);
+// BENCHMARK(BM_GenerateMandelbrotSimd)->Arg(100000);

benchmark/source/main.cpp

@@ -0,0 +1,2 @@
+#include <benchmark/benchmark.h>
+BENCHMARK_MAIN();

cmake/dev-mode.cmake

@@ -2,6 +2,7 @@ include(cmake/folders.cmake)
 
 include(CTest)
 add_subdirectory(test)
+add_subdirectory(benchmark)
 
 add_custom_target(
     run-exe

conanfile.py

@@ -19,3 +19,4 @@ def configure(self):
 
     def build_requirements(self):
         self.test_requires("catch2/3.7.0")
+        self.test_requires("benchmark/1.9.0");

source/config.hpp

@@ -1,21 +1,16 @@
 #pragma once
 
-#include "coordinates.hpp"
-#include "units.hpp"
+#include "units/coordinates.hpp"
+#include "units/units.hpp"
 
 #include <cstddef>
 
 namespace fractal {
 
-constexpr std::size_t WINDOW_WIDTH = 5120;
-constexpr std::size_t WINDOW_HEIGHT = 1440;
+constexpr std::size_t WINDOW_WIDTH = 800UZ;
+constexpr std::size_t WINDOW_HEIGHT = 600UZ;
 constexpr std::size_t FRAME_RATE = 60UZ;
 
-constexpr display_domain DISPLAY_DOMAIN{
-    {0,                0                },
-    {WINDOW_WIDTH - 1, WINDOW_HEIGHT - 1}
-};
-
 constexpr complex_domain START_COMPLEX_DOMAIN{
     {complex_underlying{-1.402}, complex_underlying{-.001}},
     {complex_underlying{-1.400}, complex_underlying{.001} }

source/graphics/aspect_ratio/aspect_ratio.cpp

@@ -1,13 +1,15 @@
 #include "aspect_ratio.hpp"
 
+#include "units/display_domain.hpp"
+
 namespace fractal {
 
-display_coordinate calculate_rectangle_end_point(
+DisplayDomain calculate_rectangle_end_points(
     display_coordinate start, display_coordinate current, float target_aspect_ratio
 )
 {
-    auto width = static_cast<float>(std::abs(current.first - start.first));
-    auto height = static_cast<float>(std::abs(current.second - start.second));
+    auto width = static_cast<float>(std::abs(current.x - start.x));
+    auto height = static_cast<float>(std::abs(current.y - start.y));
 
     // Adjust the dimensions to maintain the target aspect ratio
     if (width / height > target_aspect_ratio) {
@@ -19,18 +21,30 @@ display_coordinate calculate_rectangle_end_point(
         height = width / target_aspect_ratio;
     }
 
-    auto x = static_cast<float>(std::min(current.first, start.first));
-    auto y = static_cast<float>(std::min(current.second, start.second));
+    auto x_pos1 = static_cast<float>(start.x);
+    auto y_pos1 = static_cast<float>(start.y);
+    auto x_pos2 = x_pos1 + width;
+    auto y_pos2 = y_pos1 + height;
+
+    bool flipped_horizontal = start.x > current.x;
+    bool flipped_vertical = start.y > current.y;
 
-    // Adjust the top-left corner based on new dimensions
-    if (current.first < start.first) {
-        x = static_cast<float>(start.first) - width;
+    if (flipped_vertical) {
+        y_pos1 -= height;
+        y_pos2 -= height;
     }
-    if (current.second < start.second) {
-        y = static_cast<float>(start.second) - height;
+    if (flipped_horizontal) {
+        x_pos1 -= width;
+        x_pos2 -= width;
     }
 
-    // Return the top-left and bottom-right corners as a pair of sf::Vector2f
-    return {x + width, y + height};
+    display_coordinate top_left{
+        static_cast<uint16_t>(x_pos1), static_cast<uint16_t>(y_pos1)
+    };
+    display_coordinate bottom_right{
+        static_cast<uint16_t>(x_pos2), static_cast<uint16_t>(y_pos2)
+    };
+
+    return {top_left, bottom_right};
 }
 } // namespace fractal

source/graphics/aspect_ratio/aspect_ratio.hpp

@@ -1,10 +1,11 @@
 #pragma once
 
 #include "config.hpp"
-#include "coordinates.hpp"
+#include "units/coordinates.hpp"
+#include "units/display_domain.hpp"
 
 namespace fractal {
-display_coordinate calculate_rectangle_end_point(
+DisplayDomain calculate_rectangle_end_points(
     display_coordinate start, display_coordinate current,
     float target_aspect_ratio = static_cast<float>(WINDOW_WIDTH) / WINDOW_HEIGHT
 );

source/graphics/color_conversions/color_conversions.cpp

@@ -7,10 +7,10 @@
 #include <stdexcept>
 
 namespace fractal {
-color hsv_to_rgb(float hue, float saturation, float value)
+color hsv_to_rgb(Hue hue, Percentage saturation, Percentage value)
 {
-    float chroma = value * saturation;
-    float hue_prime = hue / 60.0f;
+    float chroma = value.get_percentage() * saturation.get_percentage();
+    float hue_prime = hue.get_hue() / 60.0f;
     float uint16_termediate =
         chroma * (1 - static_cast<float>(std::fabs(std::fmod(hue_prime, 2) - 1)));
     float red_temp = 0.0f;
@@ -42,7 +42,7 @@ color hsv_to_rgb(float hue, float saturation, float value)
         blue_temp = uint16_termediate;
     }
 
-    float match_value = value - chroma;
+    float match_value = value.get_percentage() - chroma;
     float red = red_temp + match_value;
     float green = green_temp + match_value;
     float blue = blue_temp + match_value;
@@ -54,15 +54,11 @@ color hsv_to_rgb(float hue, float saturation, float value)
     return {red_int, green_int, blue_int};
 }
 
-color ratio_to_rgb(float ratio)
+color ratio_to_rgb(Percentage ratio)
 {
-    if (ratio < 0 || ratio > 1) [[unlikely]] {
-        throw std::out_of_range(fmt::format("Ratio out of range: {}", ratio));
-    }
-
-    float hue = ratio * 360.0f;
-    float saturation = 1.0f;
-    float value = 1.0f;
+    Hue hue{ratio.get_percentage() * 360.0f};
+    Percentage saturation{1.0f};
+    Percentage value{1.0f};
 
     return hsv_to_rgb(hue, saturation, value);
 }

source/graphics/color_conversions/color_conversions.hpp

@@ -1,9 +1,13 @@
 #pragma once
 
-#include "units.hpp"
+#include "units/units.hpp"
+#include "util.hpp"
+
+#include <cassert>
 
 namespace fractal {
-color hsv_to_rgb(float hue, float saturation, float value);
 
-color ratio_to_rgb(float ratio);
+color hsv_to_rgb(Hue hue, Percentage saturation, Percentage value);
+
+color ratio_to_rgb(Percentage ratio);
 } // namespace fractal

source/graphics/display/display.cpp

@@ -41,7 +41,7 @@ void PixelDisplay::handle_event_(const sf::Event& event)
     }
 }
 
-void PixelDisplay::add_observer(std::unique_ptr<DisplayEventObserver> observer)
+void PixelDisplay::add_drawable(std::unique_ptr<DisplayEventObserver> observer)
 {
     observers_.push_back(std::move(observer));
 }

source/graphics/display/display.hpp

@@ -21,7 +21,7 @@ class PixelDisplay {
 public:
     explicit PixelDisplay();
 
-    void add_observer(std::unique_ptr<DisplayEventObserver> observer);
+    void add_drawable(std::unique_ptr<DisplayEventObserver> observer);
 
     void poll_window_events();
     void display_window();