Merge pull request #73 from orange-cpp/featore/performance_tests

added performance folder

.gitmodules

@@ -1,3 +1,6 @@
 [submodule "extlibs/googletest"]
 	path = extlibs/googletest
 	url = https://github.com/google/googletest.git
+[submodule "extlibs/benchmark"]
+	path = extlibs/benchmark
+	url = https://github.com/google/benchmark.git

.idea/vcs.xml

@@ -2,6 +2,7 @@
 <project version="4">
   <component name="VcsDirectoryMappings">
     <mapping directory="" vcs="Git" />
+    <mapping directory="$PROJECT_DIR$/extlibs/benchmark" vcs="Git" />
     <mapping directory="$PROJECT_DIR$/extlibs/googletest" vcs="Git" />
   </component>
 </project>

CMakeLists.txt

@@ -6,6 +6,7 @@ include(CMakePackageConfigHelpers)
 
 
 option(OMATH_BUILD_TESTS "Build unit tests" ${PROJECT_IS_TOP_LEVEL})
+option(OMATH_BUILD_BENCHMARK "Build benchmarks" ${PROJECT_IS_TOP_LEVEL})
 option(OMATH_THREAT_WARNING_AS_ERROR "Set highest level of warnings and force compiler to treat them as errors" ON)
 option(OMATH_BUILD_AS_SHARED_LIBRARY "Build Omath as .so or .dll" OFF)
 option(OMATH_USE_AVX2 "Omath will use AVX2 to boost performance" ON)
@@ -16,9 +17,10 @@ option(OMATH_SUPRESS_SAFETY_CHECKS "Supress some safety checks in release build
 option(OMATH_USE_UNITY_BUILD "Will enable unity build to speed up compilation" OFF)
 option(OMATH_ENABLE_LEGACY "Will enable legacy classes that MUST be used ONLY for backward compatibility" OFF)
 
-message(STATUS "[${PROJECT_NAME}]: Building on ${CMAKE_HOST_SYSTEM_NAME}")
+message(STATUS "[${PROJECT_NAME}]: Building on ${CMAKE_HOST_SYSTEM_NAME}, compiler ${CMAKE_CXX_COMPILER_ID}")
 message(STATUS "[${PROJECT_NAME}]: Warnings as errors ${OMATH_THREAT_WARNING_AS_ERROR}")
 message(STATUS "[${PROJECT_NAME}]: Build unit tests ${OMATH_BUILD_TESTS}")
+message(STATUS "[${PROJECT_NAME}]: Build benchmark ${OMATH_BUILD_BENCHMARK}")
 message(STATUS "[${PROJECT_NAME}]: As dynamic library ${OMATH_BUILD_AS_SHARED_LIBRARY}")
 message(STATUS "[${PROJECT_NAME}]: Static C++ runtime ${OMATH_STATIC_MSVC_RUNTIME_LIBRARY}")
 message(STATUS "[${PROJECT_NAME}]: CMake unity build ${OMATH_USE_UNITY_BUILD}")
@@ -90,19 +92,25 @@ if (OMATH_STATIC_MSVC_RUNTIME_LIBRARY)
     )
 endif ()
 
-if (CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+if (OMATH_USE_AVX2 AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-    target_compile_options(${PROJECT_NAME} PRIVATE -mavx2 -mfma)
+    target_compile_options(${PROJECT_NAME} PUBLIC -mavx2 -mavx -mfma)
 endif ()
 
 target_compile_features(${PROJECT_NAME} PUBLIC cxx_std_23)
 
+if (OMATH_BUILD_TESTS OR OMATH_BUILD_BENCHMARK)
+    add_subdirectory(extlibs)
+endif ()
 
 if (OMATH_BUILD_TESTS)
-    add_subdirectory(extlibs)
     add_subdirectory(tests)
     target_compile_definitions(${PROJECT_NAME} PUBLIC OMATH_BUILD_TESTS)
 endif ()
 
+if (OMATH_BUILD_BENCHMARK)
+    add_subdirectory(benchmark)
+endif ()
+
 if (OMATH_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif ()

benchmark/CMakeLists.txt

@@ -0,0 +1,15 @@
+project(omath_benchmark)
+
+
+file(GLOB_RECURSE OMATH_BENCHMARK_SOURCES CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/*.cpp")
+add_executable(${PROJECT_NAME} ${OMATH_BENCHMARK_SOURCES})
+
+set_target_properties(${PROJECT_NAME} PROPERTIES
+        ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_SOURCE_DIR}/out/${CMAKE_BUILD_TYPE}"
+        LIBRARY_OUTPUT_DIRECTORY "${CMAKE_SOURCE_DIR}/out/${CMAKE_BUILD_TYPE}"
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_SOURCE_DIR}/out/${CMAKE_BUILD_TYPE}"
+        CXX_STANDARD 23
+        CXX_STANDARD_REQUIRED ON)
+
+
+target_link_libraries(${PROJECT_NAME} PRIVATE benchmark::benchmark omath)

benchmark/benchmark_mat.cpp

@@ -0,0 +1,66 @@
+//
+// Created by Vlad on 9/17/2025.
+//
+#include <benchmark/benchmark.h>
+
+#include <omath/omath.hpp>
+#include <chrono>
+using namespace omath;
+
+
+void mat_float_multiplication_col_major(benchmark::State& state)
+{
+    using MatType = Mat<128, 128, float, MatStoreType::COLUMN_MAJOR>;
+    MatType a;
+    MatType b;
+    a.set(3.f);
+    b.set(7.f);
+
+
+    for (auto _ : state)
+        std::ignore = a * b;
+}
+void mat_float_multiplication_row_major(benchmark::State& state)
+{
+    using MatType = Mat<128, 128, float, MatStoreType::ROW_MAJOR>;
+    MatType a;
+    MatType b;
+    a.set(3.f);
+    b.set(7.f);
+
+
+    for (auto _ : state)
+        std::ignore = a * b;
+}
+
+void mat_double_multiplication_row_major(benchmark::State& state)
+{
+    using MatType = Mat<128, 128, double, MatStoreType::ROW_MAJOR>;
+    MatType a;
+    MatType b;
+    a.set(3.f);
+    b.set(7.f);
+
+
+    for (auto _ : state)
+        std::ignore = a * b;
+}
+
+void mat_double_multiplication_col_major(benchmark::State& state)
+{
+    using MatType = Mat<128, 128, double, MatStoreType::COLUMN_MAJOR>;
+    MatType a;
+    MatType b;
+    a.set(3.f);
+    b.set(7.f);
+
+
+    for (auto _ : state)
+        std::ignore = a * b;
+}
+
+BENCHMARK(mat_float_multiplication_col_major)->Iterations(5000);
+BENCHMARK(mat_float_multiplication_row_major)->Iterations(5000);
+
+BENCHMARK(mat_double_multiplication_col_major)->Iterations(5000);
+BENCHMARK(mat_double_multiplication_row_major)->Iterations(5000);

benchmark/main.cpp

@@ -0,0 +1,5 @@
+//
+// Created by Vlad on 9/17/2025.
+//
+#include <benchmark/benchmark.h>
+BENCHMARK_MAIN();

extlibs/CMakeLists.txt

@@ -1 +1,2 @@
 add_subdirectory(googletest)
+add_subdirectory(benchmark)

include/omath/linear_algebra/mat.hpp

@@ -10,7 +10,7 @@
 #include <sstream>
 #include <stdexcept>
 #include <utility>
-
+#include <immintrin.h>
 
 namespace omath
 {
@@ -155,10 +155,17 @@ namespace omath
         constexpr Mat<Rows, OtherColumns, Type, StoreType>
         operator*(const Mat<Columns, OtherColumns, Type, StoreType>& other) const
         {
+#ifdef OMATH_USE_AVX2
+            if constexpr (StoreType == MatStoreType::ROW_MAJOR)
+                return avx_multiply_row_major(other);
+            if constexpr (StoreType == MatStoreType::COLUMN_MAJOR)
+                return avx_multiply_col_major(other);
+#else
             if constexpr (StoreType == MatStoreType::ROW_MAJOR)
                 return cache_friendly_multiply_row_major(other);
             if constexpr (StoreType == MatStoreType::COLUMN_MAJOR)
                 return cache_friendly_multiply_col_major(other);
+#endif
             std::unreachable();
         }
 
@@ -391,6 +398,160 @@ namespace omath
                 }
             return result;
         }
+#ifdef OMATH_USE_AVX2
+        template<size_t OtherColumns> [[nodiscard]]
+        constexpr Mat<Rows, OtherColumns, Type, MatStoreType::COLUMN_MAJOR>
+        avx_multiply_col_major(const Mat<Columns, OtherColumns, Type, MatStoreType::COLUMN_MAJOR>& other) const
+        {
+            Mat<Rows, OtherColumns, Type, MatStoreType::COLUMN_MAJOR> result;
+
+            const Type* this_mat_data = this->raw_array().data();
+            const Type* other_mat_data = other.raw_array().data();
+            Type* result_mat_data = result.raw_array().data();
+
+            if constexpr (std::is_same_v<Type, float>)
+            {
+                // ReSharper disable once CppTooWideScopeInitStatement
+                constexpr std::size_t vector_size = 8;
+                for (std::size_t j = 0; j < OtherColumns; ++j)
+                {
+                    auto* c_col = reinterpret_cast<float*>(result_mat_data + j * Rows);
+                    for (std::size_t k = 0; k < Columns; ++k)
+                    {
+                        const float bkj = reinterpret_cast<const float*>(other_mat_data)[k + j * Columns];
+                        __m256 bkjv = _mm256_set1_ps(bkj);
+
+                        const auto* a_col_k = reinterpret_cast<const float*>(this_mat_data + k * Rows);
+
+                        std::size_t i = 0;
+                        for (; i + vector_size <= Rows; i += vector_size)
+                        {
+                            __m256 cvec = _mm256_loadu_ps(c_col + i);
+                            __m256 avec = _mm256_loadu_ps(a_col_k + i);
+#if defined(__FMA__)
+                            cvec = _mm256_fmadd_ps(avec, bkjv, cvec);
+#else
+                            cvec = _mm256_add_ps(cvec, _mm256_mul_ps(avec, bkjv));
+#endif
+                            _mm256_storeu_ps(c_col + i, cvec);
+                        }
+                        for (; i < Rows; ++i)
+                            c_col[i] += a_col_k[i] * bkj;
+                    }
+                }
+            }
+            else if (std::is_same_v<Type, double>)
+            { // double
+                // ReSharper disable once CppTooWideScopeInitStatement
+                constexpr std::size_t vector_size = 4;
+                for (std::size_t j = 0; j < OtherColumns; ++j)
+                {
+                    auto* c_col = reinterpret_cast<double*>(result_mat_data + j * Rows);
+                    for (std::size_t k = 0; k < Columns; ++k)
+                    {
+                        const double bkj = reinterpret_cast<const double*>(other_mat_data)[k + j * Columns];
+                        __m256d bkjv = _mm256_set1_pd(bkj);
+
+                        const auto* a_col_k = reinterpret_cast<const double*>(this_mat_data + k * Rows);
+
+                        std::size_t i = 0;
+                        for (; i + vector_size <= Rows; i += vector_size)
+                        {
+                            __m256d cvec = _mm256_loadu_pd(c_col + i);
+                            __m256d avec = _mm256_loadu_pd(a_col_k + i);
+#if defined(__FMA__)
+                            cvec = _mm256_fmadd_pd(avec, bkjv, cvec);
+#else
+                            cvec = _mm256_add_pd(cvec, _mm256_mul_pd(avec, bkjv));
+#endif
+                            _mm256_storeu_pd(c_col + i, cvec);
+                        }
+                        for (; i < Rows; ++i)
+                            c_col[i] += a_col_k[i] * bkj;
+                    }
+                }
+            }
+            else
+                std::unreachable();
+
+            return result;
+        }
+
+        template<size_t OtherColumns> [[nodiscard]]
+        constexpr Mat<Rows, OtherColumns, Type, MatStoreType::ROW_MAJOR>
+        avx_multiply_row_major(const Mat<Columns, OtherColumns, Type, MatStoreType::ROW_MAJOR>& other) const
+        {
+            Mat<Rows, OtherColumns, Type, MatStoreType::ROW_MAJOR> result;
+
+            const Type* this_mat_data = this->raw_array().data();
+            const Type* other_mat_data = other.raw_array().data();
+            Type* result_mat_data = result.raw_array().data();
+
+            if constexpr (std::is_same_v<Type, float>)
+            {
+                // ReSharper disable once CppTooWideScopeInitStatement
+                constexpr std::size_t vector_size = 8;
+                for (std::size_t i = 0; i < Rows; ++i)
+                {
+                    Type* c_row = result_mat_data + i * OtherColumns;
+                    for (std::size_t k = 0; k < Columns; ++k)
+                    {
+                        const auto aik = static_cast<float>(this_mat_data[i * Columns + k]);
+                        __m256 aikv = _mm256_set1_ps(aik);
+                        const auto* b_row = reinterpret_cast<const float*>(other_mat_data + k * OtherColumns);
+
+                        std::size_t j = 0;
+                        for (; j + vector_size <= OtherColumns; j += vector_size)
+                        {
+                            __m256 cvec = _mm256_loadu_ps(c_row + j);
+                            __m256 bvec = _mm256_loadu_ps(b_row + j);
+#if defined(__FMA__)
+                            cvec = _mm256_fmadd_ps(bvec, aikv, cvec);
+#else
+                            cvec = _mm256_add_ps(cvec, _mm256_mul_ps(bvec, aikv));
+#endif
+                            _mm256_storeu_ps(c_row + j, cvec);
+                        }
+                        for (; j < OtherColumns; ++j)
+                            c_row[j] += aik * b_row[j];
+                    }
+                }
+            }
+            else if (std::is_same_v<Type, double>)
+            { // double
+                // ReSharper disable once CppTooWideScopeInitStatement
+                constexpr std::size_t vector_size = 4;
+                for (std::size_t i = 0; i < Rows; ++i)
+                {
+                    Type* c_row = result_mat_data + i * OtherColumns;
+                    for (std::size_t k = 0; k < Columns; ++k)
+                    {
+                        const auto aik = static_cast<double>(this_mat_data[i * Columns + k]);
+                        __m256d aikv = _mm256_set1_pd(aik);
+                        const auto* b_row = reinterpret_cast<const double*>(other_mat_data + k * OtherColumns);
+
+                        std::size_t j = 0;
+                        for (; j + vector_size <= OtherColumns; j += vector_size)
+                        {
+                            __m256d cvec = _mm256_loadu_pd(c_row + j);
+                            __m256d bvec = _mm256_loadu_pd(b_row + j);
+#if defined(__FMA__)
+                            cvec = _mm256_fmadd_pd(bvec, aikv, cvec);
+#else
+                            cvec = _mm256_add_pd(cvec, _mm256_mul_pd(bvec, aikv));
+#endif
+                            _mm256_storeu_pd(c_row + j, cvec);
+                        }
+                        for (; j < OtherColumns; ++j)
+                            c_row[j] += aik * b_row[j];
+                    }
+                }
+            }
+            else
+                std::unreachable();
+            return result;
+        }
+#endif
     };
 
     template<class Type = float, MatStoreType St = MatStoreType::ROW_MAJOR> [[nodiscard]]