improved matrix multiplication

Add missing `find_dependency` calls.

.github/workflows/cmake-multi-platform.yml

@@ -245,72 +245,37 @@ jobs:
         run: |
           ./out/Debug/unit_tests.exe
 
-      - name: Process Coverage (llvm-profdata & llvm-cov)
+      - name: Process and render coverage
         if: ${{ matrix.triplet == 'x64-windows' }}
         shell: pwsh
         run: |
-          $BUILD_DIR = "cmake-build/build/${{ matrix.preset }}"
+          $buildDir = "cmake-build/build/${{ matrix.preset }}"
-          $EXE_PATH = "./out/Debug/unit_tests.exe"
+          $exePath = "./out/Debug/unit_tests.exe"
+          $profile = "$buildDir/unit_tests.profdata"
+          $htmlDir = "$buildDir/coverage-html"
 
-          # 1. Merge raw profile data (essential step)
           & "C:/Program Files/LLVM/bin/llvm-profdata.exe" merge `
-            -sparse "$BUILD_DIR/unit_tests.profraw" `
+            -sparse "$buildDir/unit_tests.profraw" `
-            -o "$BUILD_DIR/unit_tests.profdata"
+            -o $profile
 
-          # 2. Export to LCOV format
+          if ($LASTEXITCODE -ne 0) {
-          #    NOTE: We explicitly ignore vcpkg_installed and system headers
+            exit $LASTEXITCODE
-          & "C:/Program Files/LLVM/bin/llvm-cov.exe" export "$EXE_PATH" `
-            -instr-profile="$BUILD_DIR/unit_tests.profdata" `
-            -format=lcov `
-            -ignore-filename-regex="vcpkg_installed|external|tests" `
-            > "$BUILD_DIR/lcov.info"
-
-          if (Test-Path "$BUILD_DIR/lcov.info") {
-            Write-Host "✅ LCOV info created at $BUILD_DIR/lcov.info"
-          } else {
-            Write-Error "Failed to create LCOV info"
-            exit 1
           }
 
-      - name: Install LCOV (for genhtml)
+          & "C:/Program Files/LLVM/bin/llvm-cov.exe" show $exePath `
-        if: ${{ matrix.triplet == 'x64-windows' }}
+            "-instr-profile=$profile" `
-        run: choco install lcov -y
+            "-format=html" `
+            "-output-dir=$htmlDir" `
+            "-ignore-filename-regex=vcpkg_installed|external|tests" `
+            "-show-branches=count"
 
-      - name: Generate HTML Report
+          if ($LASTEXITCODE -ne 0) {
-        if: ${{ matrix.triplet == 'x64-windows' }}
+            exit $LASTEXITCODE
-        shell: bash
+          }
-        run: |
-          BUILD_DIR="cmake-build/build/${{ matrix.preset }}"
-          LCOV_INFO="${BUILD_DIR}/lcov.info"
-          HTML_DIR="${BUILD_DIR}/coverage-html"
 
-          # Fix paths for genhtml (Perl hates backslashes)
+          if (-not (Test-Path "$htmlDir/index.html")) {
-          sed -i 's|\\|/|g' "${LCOV_INFO}"
+            throw "LLVM coverage report was not generated"
-          
+          }
-          # Locate genhtml provided by 'choco install lcov'
-          # It is typically in ProgramData/chocolatey/lib/lcov/tools/bin
-          GENHTML=$(find /c/ProgramData/chocolatey -name genhtml -print -quit)
-          
-          if [ -z "$GENHTML" ]; then
-            echo "Error: genhtml executable not found"
-            exit 1
-          fi
-          
-          echo "Using genhtml: $GENHTML"
-          mkdir -p "$HTML_DIR"
-          
-          # Run genhtml
-          # Added --demangle-cpp if your version supports it, otherwise remove it
-          perl "$GENHTML" \
-            "${LCOV_INFO}" \
-            --output-directory "$HTML_DIR" \
-            --title "OMath Coverage Report" \
-            --legend \
-            --show-details \
-            --branch-coverage \
-            --ignore-errors source
-            
-          echo "✅ LCOV HTML report generated at $HTML_DIR"
 
       - name: Upload Coverage (HTML Report)
         if: ${{ matrix.triplet == 'x64-windows' }}

VERSION

@@ -1 +1 @@
-5.4.0
+5.5.1

benchmark/benchmark_mat.cpp

@@ -2,11 +2,9 @@
 // Created by Vlad on 9/17/2025.
 //
 #include <benchmark/benchmark.h>
-
 #include <omath/omath.hpp>
 using namespace omath;
 
-
 void mat_float_multiplication_col_major(benchmark::State& state)
 {
     using MatType = Mat<128, 128, float, MatStoreType::COLUMN_MAJOR>;
@@ -15,9 +13,12 @@ void mat_float_multiplication_col_major(benchmark::State& state)
     a.set(3.f);
     b.set(7.f);
 
-
     for ([[maybe_unused]] const auto _ : state)
-        std::ignore = a * b;
+    {
+        benchmark::DoNotOptimize(a);
+        benchmark::DoNotOptimize(b);
+        benchmark::DoNotOptimize(a * b);
+    }
 }
 void mat_float_multiplication_row_major(benchmark::State& state)
 {
@@ -27,9 +28,12 @@ void mat_float_multiplication_row_major(benchmark::State& state)
     a.set(3.f);
     b.set(7.f);
 
-
     for ([[maybe_unused]] const auto _ : state)
-        std::ignore = a * b;
+    {
+        benchmark::DoNotOptimize(a);
+        benchmark::DoNotOptimize(b);
+        benchmark::DoNotOptimize(a * b);
+    }
 }
 
 void mat_double_multiplication_row_major(benchmark::State& state)
@@ -40,9 +44,12 @@ void mat_double_multiplication_row_major(benchmark::State& state)
     a.set(3.f);
     b.set(7.f);
 
-
     for ([[maybe_unused]] const auto _ : state)
-        std::ignore = a * b;
+    {
+        benchmark::DoNotOptimize(a);
+        benchmark::DoNotOptimize(b);
+        benchmark::DoNotOptimize(a * b);
+    }
 }
 
 void mat_double_multiplication_col_major(benchmark::State& state)
@@ -53,9 +60,12 @@ void mat_double_multiplication_col_major(benchmark::State& state)
     a.set(3.f);
     b.set(7.f);
 
-
     for ([[maybe_unused]] const auto _ : state)
-        std::ignore = a * b;
+    {
+        benchmark::DoNotOptimize(a);
+        benchmark::DoNotOptimize(b);
+        benchmark::DoNotOptimize(a * b);
+    }
 }
 
 BENCHMARK(mat_float_multiplication_col_major)->Iterations(5000);

cmake/omathConfig.cmake.in

@@ -6,6 +6,17 @@ if (@OMATH_IMGUI_INTEGRATION@)
     find_dependency(imgui CONFIG)
 endif()
 
+if (@OMATH_ENABLE_LUA@)
+    find_dependency(Lua)
+endif()
+
+if (@OMATH_ENABLE_HOOKING@)
+    find_dependency(safetyhook CONFIG)
+    if (NOT WIN32 AND (UNIX AND NOT APPLE))
+        find_dependency(OpenGL)
+    endif()
+endif()
+
 # Load the targets for the omath library
 include("${CMAKE_CURRENT_LIST_DIR}/omathTargets.cmake")
 check_required_components(omath)

examples/example_hud/gui/main_window.cpp

@@ -161,6 +161,12 @@ namespace imgui_desktop::gui
                                   ImGuiColorEditFlags_NoInputs);
                 ImGui::ColorEdit4("Gradient right##lbl", reinterpret_cast<float*>(&m_label_gradient_right),
                                   ImGuiColorEditFlags_NoInputs);
+                if (m_animate_gradients)
+                {
+                    int direction = static_cast<int>(m_label_gradient_direction);
+                    if (ImGui::Combo("Direction##lbl", &direction, "Right to left\0Left to right\0"))
+                        m_label_gradient_direction = static_cast<omath::hud::GradientDirection>(direction);
+                }
             }
             ImGui::SliderFloat("Offset##lbl", &m_label_offset, 0.f, 15.f);
             ImGui::Checkbox("Right##lbl", &m_show_right_labels);
@@ -309,7 +315,8 @@ namespace imgui_desktop::gui
         const Paint centered_label_color =
                 m_gradient_label ? Paint{omath::hud::Gradient{m_label_gradient_left, m_label_gradient_right,
                                                               m_label_gradient_right, m_label_gradient_left,
-                                                              m_animate_gradients}}
+                                                              m_animate_gradients, 4.f, 0.7f,
+                                                              m_label_gradient_direction}}
                                  : Paint{omath::Color::from_rgba(255, 255, 255, 255)};
 
         auto outline_helper = [](const bool is_outline) -> Outlined

examples/example_hud/gui/main_window.hpp

@@ -68,6 +68,7 @@ namespace imgui_desktop::gui
         omath::Color m_label_gradient_right{1.f, 0.2f, 0.7f, 1.f};
         bool m_outlined = true;
         bool m_gradient_label = true;
+        omath::hud::GradientDirection m_label_gradient_direction = omath::hud::GradientDirection::RightToLeft;
         bool m_show_right_labels = true, m_show_left_labels = true;
         bool m_show_top_labels = true, m_show_bottom_labels = true;
         bool m_show_centered_top = true, m_show_centered_bottom = true;

include/omath/hud/gradient.hpp

@@ -3,6 +3,12 @@
 
 namespace omath::hud
 {
+    enum class GradientDirection
+    {
+        RightToLeft,
+        LeftToRight,
+    };
+
     struct Gradient
     {
         Color top_left;
@@ -12,5 +18,6 @@ namespace omath::hud
         bool animated = false;
         float animation_speed = 4.f;
         float animation_spread = 0.7f;
+        GradientDirection direction = GradientDirection::RightToLeft;
     };
 } // namespace omath::hud

include/omath/linear_algebra/mat.hpp

@@ -186,7 +186,14 @@ namespace omath
                 else if constexpr (StoreType == MatStoreType::COLUMN_MAJOR)
                     return cache_friendly_multiply_col_major(other);
             }
-            if constexpr (StoreType == MatStoreType::ROW_MAJOR)
+            if constexpr (!std::is_same_v<Type, float> && !std::is_same_v<Type, double>)
+            {
+                if constexpr (StoreType == MatStoreType::ROW_MAJOR)
+                    return cache_friendly_multiply_row_major(other);
+                else if constexpr (StoreType == MatStoreType::COLUMN_MAJOR)
+                    return cache_friendly_multiply_col_major(other);
+            }
+            else if constexpr (StoreType == MatStoreType::ROW_MAJOR)
                 return avx_multiply_row_major(other);
             else if constexpr (StoreType == MatStoreType::COLUMN_MAJOR)
                 return avx_multiply_col_major(other);
@@ -429,13 +436,22 @@ namespace omath
         cache_friendly_multiply_row_major(const Mat<Columns, OtherColumns, Type, MatStoreType::ROW_MAJOR>& other) const
         {
             Mat<Rows, OtherColumns, Type, MatStoreType::ROW_MAJOR> result;
+            const Type* left_data = m_data.data();
+            const Type* right_data = other.raw_array().data();
+            Type* result_data = result.raw_array().data();
+
             for (std::size_t row_index = 0; row_index < Rows; ++row_index)
+            {
+                const Type* left_row = left_data + row_index * Columns;
+                Type* result_row = result_data + row_index * OtherColumns;
                 for (std::size_t column_index = 0; column_index < Columns; ++column_index)
                 {
-                    const Type& current_number = at(row_index, column_index);
+                    const Type current_number = left_row[column_index];
+                    const Type* right_row = right_data + column_index * OtherColumns;
                     for (std::size_t other_column = 0; other_column < OtherColumns; ++other_column)
-                        result.at(row_index, other_column) += current_number * other.at(column_index, other_column);
+                        result_row[other_column] += current_number * right_row[other_column];
                 }
+            }
             return result;
         }
 
@@ -444,13 +460,22 @@ namespace omath
                 const Mat<Columns, OtherColumns, Type, MatStoreType::COLUMN_MAJOR>& other) const
         {
             Mat<Rows, OtherColumns, Type, MatStoreType::COLUMN_MAJOR> result;
+            const Type* left_data = m_data.data();
+            const Type* right_data = other.raw_array().data();
+            Type* result_data = result.raw_array().data();
+
             for (std::size_t other_column = 0; other_column < OtherColumns; ++other_column)
+            {
+                const Type* right_column = right_data + other_column * Columns;
+                Type* result_column = result_data + other_column * Rows;
                 for (std::size_t column_index = 0; column_index < Columns; ++column_index)
                 {
-                    const Type& current_number = other.at(column_index, other_column);
+                    const Type current_number = right_column[column_index];
+                    const Type* left_column = left_data + column_index * Rows;
                     for (std::size_t row_index = 0; row_index < Rows; ++row_index)
-                        result.at(row_index, other_column) += at(row_index, column_index) * current_number;
+                        result_column[row_index] += left_column[row_index] * current_number;
                 }
+            }
             return result;
         }
 #ifdef OMATH_USE_AVX2
@@ -466,56 +491,92 @@ namespace omath
 
             if constexpr (std::is_same_v<Type, float>)
             {
-                // ReSharper disable once CppTooWideScopeInitStatement
                 constexpr std::size_t vector_size = 8;
+                constexpr std::size_t block_size = vector_size * 4;
                 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)
+                    std::size_t i = 0;
+                    for (; i + block_size <= Rows; i += block_size)
                     {
-                        const float bkj = reinterpret_cast<const float*>(other_mat_data)[k + j * Columns];
+                        __m256 cvec0 = _mm256_setzero_ps();
-                        const __m256 bkj_vec = _mm256_set1_ps(bkj);
+                        __m256 cvec1 = _mm256_setzero_ps();
-
+                        __m256 cvec2 = _mm256_setzero_ps();
-                        const auto* a_col_k = reinterpret_cast<const float*>(this_mat_data + k * Rows);
+                        __m256 cvec3 = _mm256_setzero_ps();
-
+                        for (std::size_t k = 0; k < Columns; ++k)
-                        std::size_t i = 0;
-                        for (; i + vector_size <= Rows; i += vector_size)
                         {
-                            __m256 cvec = _mm256_loadu_ps(c_col + i);
+                            const __m256 bkj_vec = _mm256_set1_ps(other_mat_data[k + j * Columns]);
+                            const auto* a_col_k = this_mat_data + k * Rows + i;
+                            cvec0 = _mm256_fmadd_ps(_mm256_loadu_ps(a_col_k), bkj_vec, cvec0);
+                            cvec1 = _mm256_fmadd_ps(_mm256_loadu_ps(a_col_k + vector_size), bkj_vec, cvec1);
+                            cvec2 = _mm256_fmadd_ps(_mm256_loadu_ps(a_col_k + vector_size * 2), bkj_vec, cvec2);
+                            cvec3 = _mm256_fmadd_ps(_mm256_loadu_ps(a_col_k + vector_size * 3), bkj_vec, cvec3);
+                        }
+                        _mm256_storeu_ps(c_col + i, cvec0);
+                        _mm256_storeu_ps(c_col + i + vector_size, cvec1);
+                        _mm256_storeu_ps(c_col + i + vector_size * 2, cvec2);
+                        _mm256_storeu_ps(c_col + i + vector_size * 3, cvec3);
+                    }
+                    for (; i + vector_size <= Rows; i += vector_size)
+                    {
+                        __m256 cvec = _mm256_setzero_ps();
+                        for (std::size_t k = 0; k < Columns; ++k)
+                        {
+                            const __m256 bkj_vec = _mm256_set1_ps(other_mat_data[k + j * Columns]);
+                            const auto* a_col_k = this_mat_data + k * Rows;
                             const __m256 a_vec = _mm256_loadu_ps(a_col_k + i);
                             cvec = _mm256_fmadd_ps(a_vec, bkj_vec, cvec);
-                            _mm256_storeu_ps(c_col + i, cvec);
                         }
-                        for (; i < Rows; ++i)
+                        _mm256_storeu_ps(c_col + i, cvec);
-                            c_col[i] += a_col_k[i] * bkj;
                     }
+                    for (; i < Rows; ++i)
+                        for (std::size_t k = 0; k < Columns; ++k)
+                            c_col[i] += this_mat_data[i + k * Rows] * other_mat_data[k + j * Columns];
                 }
             }
             else if (std::is_same_v<Type, double>)
-            { // double
+            {
-                // ReSharper disable once CppTooWideScopeInitStatement
                 constexpr std::size_t vector_size = 4;
+                constexpr std::size_t block_size = 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)
+                    std::size_t i = 0;
+                    for (; i + block_size <= Rows; i += block_size)
                     {
-                        const double bkj = reinterpret_cast<const double*>(other_mat_data)[k + j * Columns];
+                        __m256d cvec0 = _mm256_setzero_pd();
-                        const __m256d bkj_vec = _mm256_set1_pd(bkj);
+                        __m256d cvec1 = _mm256_setzero_pd();
-
+                        __m256d cvec2 = _mm256_setzero_pd();
-                        const auto* a_col_k = reinterpret_cast<const double*>(this_mat_data + k * Rows);
+                        __m256d cvec3 = _mm256_setzero_pd();
-
+                        for (std::size_t k = 0; k < Columns; ++k)
-                        std::size_t i = 0;
-                        for (; i + vector_size <= Rows; i += vector_size)
                         {
-                            __m256d cvec = _mm256_loadu_pd(c_col + i);
+                            const __m256d bkj_vec = _mm256_set1_pd(other_mat_data[k + j * Columns]);
+                            const auto* a_col_k = this_mat_data + k * Rows + i;
+                            cvec0 = _mm256_fmadd_pd(_mm256_loadu_pd(a_col_k), bkj_vec, cvec0);
+                            cvec1 = _mm256_fmadd_pd(_mm256_loadu_pd(a_col_k + vector_size), bkj_vec, cvec1);
+                            cvec2 = _mm256_fmadd_pd(_mm256_loadu_pd(a_col_k + vector_size * 2), bkj_vec, cvec2);
+                            cvec3 = _mm256_fmadd_pd(_mm256_loadu_pd(a_col_k + vector_size * 3), bkj_vec, cvec3);
+                        }
+                        _mm256_storeu_pd(c_col + i, cvec0);
+                        _mm256_storeu_pd(c_col + i + vector_size, cvec1);
+                        _mm256_storeu_pd(c_col + i + vector_size * 2, cvec2);
+                        _mm256_storeu_pd(c_col + i + vector_size * 3, cvec3);
+                    }
+                    for (; i + vector_size <= Rows; i += vector_size)
+                    {
+                        __m256d cvec = _mm256_setzero_pd();
+                        for (std::size_t k = 0; k < Columns; ++k)
+                        {
+                            const __m256d bkj_vec = _mm256_set1_pd(other_mat_data[k + j * Columns]);
+                            const auto* a_col_k = this_mat_data + k * Rows;
                             const __m256d a_vec = _mm256_loadu_pd(a_col_k + i);
                             cvec = _mm256_fmadd_pd(a_vec, bkj_vec, cvec);
-                            _mm256_storeu_pd(c_col + i, cvec);
                         }
-                        for (; i < Rows; ++i)
+                        _mm256_storeu_pd(c_col + i, cvec);
-                            c_col[i] += a_col_k[i] * bkj;
                     }
+                    for (; i < Rows; ++i)
+                        for (std::size_t k = 0; k < Columns; ++k)
+                            c_col[i] += this_mat_data[i + k * Rows] * other_mat_data[k + j * Columns];
                 }
             }
             else
@@ -536,56 +597,92 @@ namespace omath
 
             if constexpr (std::is_same_v<Type, float>)
             {
-                // ReSharper disable once CppTooWideScopeInitStatement
                 constexpr std::size_t vector_size = 8;
+                constexpr std::size_t block_size = vector_size * 4;
                 for (std::size_t i = 0; i < Rows; ++i)
                 {
-                    Type* c_row = result_mat_data + i * OtherColumns;
+                    auto* c_row = reinterpret_cast<float*>(result_mat_data + i * OtherColumns);
-                    for (std::size_t k = 0; k < Columns; ++k)
+                    std::size_t j = 0;
+                    for (; j + block_size <= OtherColumns; j += block_size)
                     {
-                        const auto aik = static_cast<float>(this_mat_data[i * Columns + k]);
+                        __m256 cvec0 = _mm256_setzero_ps();
-                        const __m256 aik_vec = _mm256_set1_ps(aik);
+                        __m256 cvec1 = _mm256_setzero_ps();
-                        const auto* b_row = reinterpret_cast<const float*>(other_mat_data + k * OtherColumns);
+                        __m256 cvec2 = _mm256_setzero_ps();
-
+                        __m256 cvec3 = _mm256_setzero_ps();
-                        std::size_t j = 0;
+                        for (std::size_t k = 0; k < Columns; ++k)
-                        for (; j + vector_size <= OtherColumns; j += vector_size)
                         {
-                            __m256 cvec = _mm256_loadu_ps(c_row + j);
+                            const __m256 aik_vec = _mm256_set1_ps(this_mat_data[i * Columns + k]);
+                            const auto* b_row = other_mat_data + k * OtherColumns + j;
+                            cvec0 = _mm256_fmadd_ps(_mm256_loadu_ps(b_row), aik_vec, cvec0);
+                            cvec1 = _mm256_fmadd_ps(_mm256_loadu_ps(b_row + vector_size), aik_vec, cvec1);
+                            cvec2 = _mm256_fmadd_ps(_mm256_loadu_ps(b_row + vector_size * 2), aik_vec, cvec2);
+                            cvec3 = _mm256_fmadd_ps(_mm256_loadu_ps(b_row + vector_size * 3), aik_vec, cvec3);
+                        }
+                        _mm256_storeu_ps(c_row + j, cvec0);
+                        _mm256_storeu_ps(c_row + j + vector_size, cvec1);
+                        _mm256_storeu_ps(c_row + j + vector_size * 2, cvec2);
+                        _mm256_storeu_ps(c_row + j + vector_size * 3, cvec3);
+                    }
+                    for (; j + vector_size <= OtherColumns; j += vector_size)
+                    {
+                        __m256 cvec = _mm256_setzero_ps();
+                        for (std::size_t k = 0; k < Columns; ++k)
+                        {
+                            const __m256 aik_vec = _mm256_set1_ps(this_mat_data[i * Columns + k]);
+                            const auto* b_row = other_mat_data + k * OtherColumns;
                             const __m256 b_vec = _mm256_loadu_ps(b_row + j);
                             cvec = _mm256_fmadd_ps(b_vec, aik_vec, cvec);
-
-                            _mm256_storeu_ps(c_row + j, cvec);
                         }
-                        for (; j < OtherColumns; ++j)
+                        _mm256_storeu_ps(c_row + j, cvec);
-                            c_row[j] += aik * b_row[j];
                     }
+                    for (; j < OtherColumns; ++j)
+                        for (std::size_t k = 0; k < Columns; ++k)
+                            c_row[j] += this_mat_data[i * Columns + k] * other_mat_data[k * OtherColumns + j];
                 }
             }
             else if (std::is_same_v<Type, double>)
-            { // double
+            {
-                // ReSharper disable once CppTooWideScopeInitStatement
                 constexpr std::size_t vector_size = 4;
+                constexpr std::size_t block_size = vector_size * 4;
                 for (std::size_t i = 0; i < Rows; ++i)
                 {
-                    Type* c_row = result_mat_data + i * OtherColumns;
+                    auto* c_row = reinterpret_cast<double*>(result_mat_data + i * OtherColumns);
-                    for (std::size_t k = 0; k < Columns; ++k)
+                    std::size_t j = 0;
+                    for (; j + block_size <= OtherColumns; j += block_size)
                     {
-                        const auto aik = static_cast<double>(this_mat_data[i * Columns + k]);
+                        __m256d cvec0 = _mm256_setzero_pd();
-                        const __m256d aik_vec = _mm256_set1_pd(aik);
+                        __m256d cvec1 = _mm256_setzero_pd();
-                        const auto* b_row = reinterpret_cast<const double*>(other_mat_data + k * OtherColumns);
+                        __m256d cvec2 = _mm256_setzero_pd();
-
+                        __m256d cvec3 = _mm256_setzero_pd();
-                        std::size_t j = 0;
+                        for (std::size_t k = 0; k < Columns; ++k)
-                        for (; j + vector_size <= OtherColumns; j += vector_size)
                         {
-                            __m256d cvec = _mm256_loadu_pd(c_row + j);
+                            const __m256d aik_vec = _mm256_set1_pd(this_mat_data[i * Columns + k]);
+                            const auto* b_row = other_mat_data + k * OtherColumns + j;
+                            cvec0 = _mm256_fmadd_pd(_mm256_loadu_pd(b_row), aik_vec, cvec0);
+                            cvec1 = _mm256_fmadd_pd(_mm256_loadu_pd(b_row + vector_size), aik_vec, cvec1);
+                            cvec2 = _mm256_fmadd_pd(_mm256_loadu_pd(b_row + vector_size * 2), aik_vec, cvec2);
+                            cvec3 = _mm256_fmadd_pd(_mm256_loadu_pd(b_row + vector_size * 3), aik_vec, cvec3);
+                        }
+                        _mm256_storeu_pd(c_row + j, cvec0);
+                        _mm256_storeu_pd(c_row + j + vector_size, cvec1);
+                        _mm256_storeu_pd(c_row + j + vector_size * 2, cvec2);
+                        _mm256_storeu_pd(c_row + j + vector_size * 3, cvec3);
+                    }
+                    for (; j + vector_size <= OtherColumns; j += vector_size)
+                    {
+                        __m256d cvec = _mm256_setzero_pd();
+                        for (std::size_t k = 0; k < Columns; ++k)
+                        {
+                            const __m256d aik_vec = _mm256_set1_pd(this_mat_data[i * Columns + k]);
+                            const auto* b_row = other_mat_data + k * OtherColumns;
                             const __m256d b_vec = _mm256_loadu_pd(b_row + j);
                             cvec = _mm256_fmadd_pd(b_vec, aik_vec, cvec);
-
-                            _mm256_storeu_pd(c_row + j, cvec);
                         }
-                        for (; j < OtherColumns; ++j)
+                        _mm256_storeu_pd(c_row + j, cvec);
-                            c_row[j] += aik * b_row[j];
                     }
+                    for (; j < OtherColumns; ++j)
+                        for (std::size_t k = 0; k < Columns; ++k)
+                            c_row[j] += this_mat_data[i * Columns + k] * other_mat_data[k * OtherColumns + j];
                 }
             }
             else

include/omath/utility/pattern_scan.hpp

@@ -54,9 +54,6 @@ namespace omath
                 return {value, Length - 1};
             }
         };
-        template<std::size_t N>
-        ConstevalPattern(const char (&)[N]) -> ConstevalPattern<N>;
-
         [[nodiscard]]
         static std::span<std::byte>::iterator scan_for_pattern(const std::span<std::byte>& range,
                                                                const std::string_view& pattern);

source/hud/renderer_realizations/imgui_renderer.cpp

@@ -103,6 +103,7 @@ namespace omath::hud
         if (gradient.animated)
         {
             const float animation_offset = static_cast<float>(ImGui::GetTime()) * gradient.animation_speed;
+            const float direction = gradient.direction == GradientDirection::RightToLeft ? 1.f : -1.f;
             float cursor_x = position.x;
             const char* glyph_start = text.data();
             const char* const text_end = text.data() + text.size();
@@ -117,7 +118,9 @@ namespace omath::hud
 
                 const char* const glyph_end = glyph_start + glyph_size;
                 const float blend =
-                        (std::sin(animation_offset + static_cast<float>(glyph_index) * gradient.animation_spread) + 1.f)
+                        (std::sin(animation_offset + direction * static_cast<float>(glyph_index)
+                                                             * gradient.animation_spread)
+                         + 1.f)
                         * 0.5f;
                 const auto color = gradient.top_left.value() * (1.f - blend) + gradient.top_right.value() * blend;
                 draw_list->AddText({cursor_x, position.y}, Color{color}.to_im_color(), glyph_start, glyph_end);

tests/general/unit_test_mat.cpp

@@ -16,6 +16,21 @@ namespace
         const float diff = actual - expected;
         return (diff < 0.0f ? -diff : diff) <= epsilon;
     }
+
+    template<size_t Rows, size_t Columns, size_t OtherColumns, class Type, MatStoreType StoreType>
+    void expect_multiplication_matches_scalar_reference(const Mat<Rows, Columns, Type, StoreType>& left,
+                                                        const Mat<Columns, OtherColumns, Type, StoreType>& right)
+    {
+        const auto result = left * right;
+        for (size_t row = 0; row < Rows; ++row)
+            for (size_t column = 0; column < OtherColumns; ++column)
+            {
+                Type expected{};
+                for (size_t shared_index = 0; shared_index < Columns; ++shared_index)
+                    expected += left.at(row, shared_index) * right.at(shared_index, column);
+                EXPECT_EQ(result.at(row, column), expected);
+            }
+    }
 } // namespace
 
 class UnitTestMat : public ::testing::Test
@@ -92,6 +107,45 @@ TEST_F(UnitTestMat, Operator_Multiplication_Matrix)
     EXPECT_FLOAT_EQ(m3.at(1, 1), 22.0f);
 }
 
+TEST(UnitTestMatStandalone, Operator_Multiplication_RowMajorSimdAndTail)
+{
+    Mat<3, 5, float, MatStoreType::ROW_MAJOR> left;
+    Mat<5, 33, float, MatStoreType::ROW_MAJOR> right;
+    for (size_t row = 0; row < left.row_count(); ++row)
+        for (size_t column = 0; column < left.columns_count(); ++column)
+            left.at(row, column) = static_cast<float>(row * 3 + column + 1);
+    for (size_t row = 0; row < right.row_count(); ++row)
+        for (size_t column = 0; column < right.columns_count(); ++column)
+            right.at(row, column) = static_cast<float>((row + 1) * (column % 5 + 1));
+
+    expect_multiplication_matches_scalar_reference(left, right);
+}
+
+TEST(UnitTestMatStandalone, Operator_Multiplication_ColumnMajorSimdAndTail)
+{
+    Mat<17, 5, double, MatStoreType::COLUMN_MAJOR> left;
+    Mat<5, 3, double, MatStoreType::COLUMN_MAJOR> right;
+    for (size_t row = 0; row < left.row_count(); ++row)
+        for (size_t column = 0; column < left.columns_count(); ++column)
+            left.at(row, column) = static_cast<double>(row * 3 + column + 1);
+    for (size_t row = 0; row < right.row_count(); ++row)
+        for (size_t column = 0; column < right.columns_count(); ++column)
+            right.at(row, column) = static_cast<double>((row + 1) * (column + 1));
+
+    expect_multiplication_matches_scalar_reference(left, right);
+}
+
+TEST(UnitTestMatStandalone, Operator_Multiplication_IntegerFallsBackFromAvx)
+{
+    constexpr Mat<2, 3, int> left{{1, 2, 3}, {4, 5, 6}};
+    constexpr Mat<3, 2, int> right{{7, 8}, {9, 10}, {11, 12}};
+    constexpr auto result = left * right;
+    static_assert(result.at(0, 0) == 58);
+    static_assert(result.at(1, 1) == 154);
+
+    expect_multiplication_matches_scalar_reference(left, right);
+}
+
 TEST_F(UnitTestMat, Operator_Multiplication_Scalar)
 {
     Mat<2, 2> m3 = m2 * 2.0f;