perf: improve vectorization optimizations and add function multi-versioning (#1271)

Signed-off-by: k4yt3x <i@k4yt3x.com>
2024-12-26 22:09:09 +00:00 · 2024-12-22 05:53:09 -05:00 · 2024-12-22 05:53:09 -05:00 · e1e8ed864d
commit e1e8ed864d
parent f68939c478
11 changed files with 116 additions and 52 deletions
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@ -49,11 +49,12 @@ jobs:
      - name: Build Video2X
        run: |
          mkdir -p /tmp/build /tmp/install
-          cmake -B /tmp/build -S . -DUSE_SYSTEM_NCNN=OFF -DUSE_SYSTEM_SPDLOG=OFF \
+          cmake -B /tmp/build -S . -DCMAKE_BUILD_TYPE=Debug \
            -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
-            -DCMAKE_BUILD_TYPE=Debug -DCMAKE_INSTALL_PREFIX=/tmp/install \
+            -DVIDEO2X_USE_EXTERNAL_NCNN=OFF -DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF \
            -DINSTALL_BIN_DESTINATION=. -DINSTALL_INCLUDE_DESTINATION=include \
-            -DINSTALL_LIB_DESTINATION=. -DINSTALL_MODEL_DESTINATION=.
+            -DINSTALL_LIB_DESTINATION=. -DINSTALL_MODEL_DESTINATION=. \
+            -DCMAKE_INSTALL_PREFIX=/tmp/install
          cmake --build /tmp/build --config Debug --target install

      - name: Upload artifacts
@ -97,7 +98,7 @@ jobs:
        shell: pwsh
        run: |
          cmake -S . -B build `
-            -DUSE_SYSTEM_NCNN=OFF -DUSE_SYSTEM_SPDLOG=OFF -DUSE_SYSTEM_BOOST=OFF `
+            -DVIDEO2X_USE_EXTERNAL_NCNN=OFF -DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF -DVIDEO2X_USE_EXTERNAL_BOOST=OFF `
            -DCMAKE_BUILD_TYPE=Debug -DCMAKE_INSTALL_PREFIX=build/video2x_install
          cmake --build build --config Debug --parallel --target install

--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@ -53,9 +53,10 @@ jobs:

      - name: Build Video2X
        run: |
-          cmake -B build -S . -DUSE_SYSTEM_NCNN=OFF -DUSE_SYSTEM_SPDLOG=OFF -DSPDLOG_NO_EXCEPTIONS=ON \
-              -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ \
-              -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=build/video2x-linux-ubuntu-2404-amd64/usr
+          cmake -B build -S . -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ \
+            -DVIDEO2X_USE_EXTERNAL_NCNN=OFF -DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF \
+            -DCMAKE_INSTALL_PREFIX=build/video2x-linux-ubuntu-2404-amd64/usr
          cmake --build build --config Release --target install --parallel
          mkdir -p build/video2x-linux-ubuntu-2404-amd64/DEBIAN
          cp packaging/debian/control build/video2x-linux-ubuntu-2404-amd64/DEBIAN/control
@ -107,7 +108,7 @@ jobs:
        shell: pwsh
        run: |
          cmake -S . -B build `
-            -DUSE_SYSTEM_NCNN=OFF -DUSE_SYSTEM_SPDLOG=OFF -DUSE_SYSTEM_BOOST=OFF `
+            -DVIDEO2X_USE_EXTERNAL_NCNN=OFF -DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF -DVIDEO2X_USE_EXTERNAL_BOOST=OFF `
            -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=build/video2x_install
          cmake --build build --config Release --parallel --target install

--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@ -5,6 +5,16 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

+## [Unreleased]
+
+### Added
+
+- Multi-versioning to critical functions to enhance performance in generic architecture builds.
+
+### Changed
+
+- Improve the CMake optimization flags and option names.
+
 ## [6.3.1] - 2024-12-21

 ### Fixed
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -16,6 +16,20 @@ if(NOT CMAKE_BUILD_TYPE)
    set(CMAKE_BUILD_TYPE Release)
 endif()

+# Build options
+option(BUILD_SHARED_LIBS "Build libvideo2x as a shared library" ON)
+option(VIDEO2X_BUILD_CLI "Build the video2x command line interface executable" ON)
+
+option(VIDEO2X_ENABLE_NATIVE "Enable native optimizations (-march=native)" OFF)
+option(VIDEO2X_ENABLE_X86_64_V4 "Enable x86-64-v4 optimizations (-march=x86-64-v4)" OFF)
+option(VIDEO2X_ENABLE_AVX512F "Enable AVX-512 foundation optimizations (-march=avx512f)" OFF)
+option(VIDEO2X_ENABLE_X86_64_V3 "Enable x86-64-v3 optimizations (-march=x86-64-v3)" OFF)
+option(VIDEO2X_ENABLE_AVX2 "Enable AVX2 optimizations (-march=avx2)" OFF)
+
+option(VIDEO2X_USE_EXTERNAL_NCNN "Use the system-provided ncnn library" ON)
+option(VIDEO2X_USE_EXTERNAL_SPDLOG "Use the system-provided spdlog library" ON)
+option(VIDEO2X_USE_EXTERNAL_BOOST "Use the system-provided Boost library" ON)
+
 # Set global compile options for all targets
 if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
    add_compile_options(/W4 /permissive-)
@ -25,30 +39,50 @@ endif()

 # Set the default optimization flags for Release builds
 if(CMAKE_BUILD_TYPE STREQUAL "Release")
+    # Set the optimization flags for each compiler
    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
        add_compile_options(/Ox /Ot /GL /DNDEBUG)
        add_link_options(/LTCG /OPT:REF /OPT:ICF)
    elseif(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
-        add_compile_options(-O3 -march=native -ffunction-sections -fdata-sections)
+        add_compile_options(-O3 -ffunction-sections -fdata-sections)
        add_link_options(-Wl,-s -flto -Wl,--gc-sections)
    endif()
 endif()

-# Build options
-option(BUILD_SHARED_LIBS "Build libvideo2x as a shared library" ON)
-option(BUILD_VIDEO2X_CLI "Build the video2x executable" ON)
-option(USE_SYSTEM_NCNN "Use system ncnn library" ON)
-option(USE_SYSTEM_SPDLOG "Use system spdlog library" ON)
-option(USE_SYSTEM_BOOST "Use system Boost library" ON)
+# Enable the requested architecture-specific optimizations
+if(VIDEO2X_ENABLE_NATIVE)
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/arch:NATIVE)
+    elseif(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+        add_compile_options(-march=native)
+    endif()
+elseif(VIDEO2X_ENABLE_X86_64_V4)
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/arch:AVX2)
+    elseif(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+        add_compile_options(-march=x86-64-v4)
+    endif()
+elseif(VIDEO2X_ENABLE_AVX512F)
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/arch:AVX512)
+    elseif(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+        add_compile_options(-mavx512f)
+    endif()
+elseif(VIDEO2X_ENABLE_X86_64_V3)
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/arch:AVX2)
+    elseif(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+        add_compile_options(-march=x86-64-v3)
+    endif()
+elseif(VIDEO2X_ENABLE_AVX2)
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/arch:AVX2)
+    elseif(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+        add_compile_options(-mavx2)
+    endif()
+endif()

-# Generate the version header file
-configure_file(
-    "${CMAKE_CURRENT_SOURCE_DIR}/include/libvideo2x/version.h.in"
-    "${CMAKE_CURRENT_BINARY_DIR}/libvideo2x/version.h"
-    @ONLY
-)
-
-# Find the required packages
+# Define lists to store include directories and libraries
 set(LIBVIDEO2X_INCLUDE_DIRS)
 set(LIBVIDEO2X_LIBS)
 set(VIDEO2X_INCLUDE_DIRS)
@ -109,7 +143,7 @@ else()
 endif() # WIN32

 # Find ncnn package
-if(USE_SYSTEM_NCNN)
+if(VIDEO2X_USE_EXTERNAL_NCNN)
    find_package(ncnn REQUIRED)
 else()
    option(NCNN_INSTALL_SDK "" OFF)
@ -208,12 +242,14 @@ else()
 endif()

 # spdlog
-if(USE_SYSTEM_SPDLOG)
+if(VIDEO2X_USE_EXTERNAL_SPDLOG)
    find_package(spdlog REQUIRED)
    list(APPEND LIBVIDEO2X_INCLUDE_DIRS ${spdlog_INCLUDE_DIRS})
    list(APPEND VIDEO2X_INCLUDE_DIRS ${spdlog_INCLUDE_DIRS})
    set(SPDLOG_LIB spdlog::spdlog)
 else()
+    # spdlog exceptions are incompatible with ncnn
+    option(SPDLOG_NO_EXCEPTIONS "" OFF)
    add_subdirectory(third_party/spdlog)
    set(SPDLOG_LIB spdlog::spdlog_header_only)
 endif()
@ -221,13 +257,13 @@ list(APPEND LIBVIDEO2X_LIBS ${SPDLOG_LIB})
 list(APPEND VIDEO2X_LIBS ${SPDLOG_LIB})

 # Find dependencies required for the CLI
-if(BUILD_VIDEO2X_CLI)
+if(VIDEO2X_BUILD_CLI)
    # Vulkan
    find_package(Vulkan REQUIRED)
    list(APPEND VIDEO2X_LIBS Vulkan::Vulkan)

    # Boost
-    if(USE_SYSTEM_BOOST)
+    if(VIDEO2X_USE_EXTERNAL_BOOST)
        find_package(Boost REQUIRED COMPONENTS program_options)
        list(APPEND LIBVIDEO2X_INCLUDE_DIRS ${Boost_INCLUDE_DIRS})
    else()
@ -255,7 +291,7 @@ ExternalProject_Add(
        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/realesrgan_install
        -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-        -DUSE_SYSTEM_NCNN=${USE_SYSTEM_NCNN}
+        -DUSE_SYSTEM_NCNN=${VIDEO2X_USE_EXTERNAL_NCNN}
    BUILD_ALWAYS ON
    INSTALL_COMMAND ${CMAKE_COMMAND} --build . --target install --config ${CMAKE_BUILD_TYPE}
 )
@ -268,7 +304,7 @@ ExternalProject_Add(
        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/realcugan_install
        -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-        -DUSE_SYSTEM_NCNN=${USE_SYSTEM_NCNN}
+        -DUSE_SYSTEM_NCNN=${VIDEO2X_USE_EXTERNAL_NCNN}
    BUILD_ALWAYS ON
    INSTALL_COMMAND ${CMAKE_COMMAND} --build . --target install --config ${CMAKE_BUILD_TYPE}
 )
@ -281,7 +317,7 @@ ExternalProject_Add(
        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/rife_install
        -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-        -DUSE_SYSTEM_NCNN=${USE_SYSTEM_NCNN}
+        -DUSE_SYSTEM_NCNN=${VIDEO2X_USE_EXTERNAL_NCNN}
    BUILD_ALWAYS ON
    INSTALL_COMMAND ${CMAKE_COMMAND} --build . --target install --config ${CMAKE_BUILD_TYPE}
 )
@ -305,6 +341,13 @@ endif()
 # Ensure that the shared library is built after the external projects
 add_dependencies(libvideo2x realesrgan realcugan rife)

+# Generate the version header file
+configure_file(
+    "${CMAKE_CURRENT_SOURCE_DIR}/include/libvideo2x/version.h.in"
+    "${CMAKE_CURRENT_BINARY_DIR}/libvideo2x/version.h"
+    @ONLY
+)
+
 # Include directories for the shared library
 target_include_directories(libvideo2x PRIVATE
    ${LIBVIDEO2X_INCLUDE_DIRS}
@ -335,7 +378,7 @@ list(APPEND LIBVIDEO2X_LIBS ${REALESRGAN_LIB} ${REALCUGAN_LIB} ${RIFE_LIB})
 target_link_libraries(libvideo2x PRIVATE ${LIBVIDEO2X_LIBS})

 if(NOT WIN32)
-    if(USE_SYSTEM_NCNN)
+    if(VIDEO2X_USE_EXTERNAL_NCNN)
        target_link_libraries(libvideo2x PUBLIC ncnn)
    else()
        target_link_libraries(libvideo2x PRIVATE ncnn)
@ -343,7 +386,7 @@ if(NOT WIN32)
 endif()

 # Create the executable 'video2x'
-if(BUILD_VIDEO2X_CLI)
+if(VIDEO2X_BUILD_CLI)
    file(GLOB VIDEO2X_SOURCES tools/video2x/src/*.cpp)
    add_executable(video2x ${VIDEO2X_SOURCES})
    set_target_properties(video2x PROPERTIES OUTPUT_NAME video2x)
@ -395,8 +438,8 @@ install(TARGETS libvideo2x
 # Install model files
 install(DIRECTORY ${CMAKE_SOURCE_DIR}/models DESTINATION ${INSTALL_MODEL_DESTINATION})

-# Install the executable if BUILD_VIDEO2X_CLI is enabled
-if(BUILD_VIDEO2X_CLI)
+# Install the executable if VIDEO2X_BUILD_CLI is enabled
+if(VIDEO2X_BUILD_CLI)
    install(TARGETS video2x RUNTIME DESTINATION ${INSTALL_BIN_DESTINATION})
 endif()

--- a/27
+++ b/27
@ -13,7 +13,8 @@ build:
 	cmake -S . -B $(BINDIR) \
 		-DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
 		-DCMAKE_CXX_COMPILER=$(CXX) \
-		-DCMAKE_BUILD_TYPE=Release
+		-DCMAKE_BUILD_TYPE=Release \
+		-DVIDEO2X_ENABLE_NATIVE=ON
 	cmake --build $(BINDIR) --config Release --parallel
 	cp $(BINDIR)/compile_commands.json .

@ -23,9 +24,9 @@ static:
 		-DCMAKE_CXX_COMPILER=$(CXX) \
 		-DCMAKE_BUILD_TYPE=Release \
 		-DBUILD_SHARED_LIBS=OFF \
-		-DUSE_SYSTEM_NCNN=OFF \
-		-DUSE_SYSTEM_SPDLOG=OFF \
-		-DUSE_SYSTEM_BOOST=OFF
+		-DVIDEO2X_USE_EXTERNAL_NCNN=OFF \
+		-DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF \
+		-DVIDEO2X_USE_EXTERNAL_BOOST=OFF
 	cmake --build $(BINDIR) --config Release --parallel
 	cp $(BINDIR)/compile_commands.json .

@ -39,18 +40,18 @@ debug:

 windows:
 	cmake -S . -B $(BINDIR) \
-		-DUSE_SYSTEM_NCNN=OFF \
-		-DUSE_SYSTEM_SPDLOG=OFF \
-		-DUSE_SYSTEM_BOOST=OFF \
+		-DVIDEO2X_USE_EXTERNAL_NCNN=OFF \
+		-DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF \
+		-DVIDEO2X_USE_EXTERNAL_BOOST=OFF \
 		-DCMAKE_BUILD_TYPE=Release \
 		-DCMAKE_INSTALL_PREFIX=build/libvideo2x-shared
 	cmake --build $(BINDIR) --config Release --parallel --target install

 windows-debug:
 	cmake -S . -B $(BINDIR) \
-		-DUSE_SYSTEM_NCNN=OFF \
-		-DUSE_SYSTEM_SPDLOG=OFF \
-		-DUSE_SYSTEM_BOOST=OFF \
+		-DVIDEO2X_USE_EXTERNAL_NCNN=OFF \
+		-DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF \
+		-DVIDEO2X_USE_EXTERNAL_BOOST=OFF \
 		-DCMAKE_BUILD_TYPE=Debug \
 		-DCMAKE_INSTALL_PREFIX=build/libvideo2x-shared
 	cmake --build $(BINDIR) --config Debug --parallel --target install
@ -70,7 +71,7 @@ debian:
 		libomp-dev \
 		libspdlog-dev \
 		libboost-program-options-dev
-	cmake -B /tmp/build -S . -DUSE_SYSTEM_NCNN=OFF -DCMAKE_CXX_COMPILER=$(CXX) \
+	cmake -B /tmp/build -S . -DVIDEO2X_USE_EXTERNAL_NCNN=OFF -DCMAKE_CXX_COMPILER=$(CXX) \
 		-DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/tmp/install \
 		-DINSTALL_BIN_DESTINATION=. -DINSTALL_INCLUDE_DESTINATION=include \
 		-DINSTALL_LIB_DESTINATION=. -DINSTALL_MODEL_DESTINATION=.
@ -90,7 +91,7 @@ ubuntu2404:
 		glslang-tools \
 		libomp-dev \
 		libboost-program-options-dev
-	cmake -B build -S . -DUSE_SYSTEM_NCNN=OFF -DUSE_SYSTEM_SPDLOG=OFF -DSPDLOG_NO_EXCEPTIONS=ON \
+	cmake -B build -S . -DVIDEO2X_USE_EXTERNAL_NCNN=OFF -DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF \
 		-DCMAKE_CXX_COMPILER=g++ -DCMAKE_BUILD_TYPE=Release \
 		-DCMAKE_INSTALL_PREFIX=video2x-linux-ubuntu-amd64/usr
 	cmake --build build --config Release --target install --parallel
@ -115,7 +116,7 @@ ubuntu2204:
 		glslang-tools \
 		libomp-dev \
 		libboost-program-options-dev
-	cmake -B build -S . -DUSE_SYSTEM_NCNN=OFF -DUSE_SYSTEM_SPDLOG=OFF -DSPDLOG_NO_EXCEPTIONS=ON \
+	cmake -B build -S . -DVIDEO2X_USE_EXTERNAL_NCNN=OFF -DVIDEO2X_USE_EXTERNAL_SPDLOG=OFF \
 		-DCMAKE_CXX_COMPILER=g++ -DCMAKE_BUILD_TYPE=Release \
 		-DCMAKE_INSTALL_PREFIX=video2x-linux-ubuntu-amd64/usr
 	cmake --build build --config Release --target install --parallel
--- a/packaging/arch/PKGBUILD
+++ b/packaging/arch/PKGBUILD
@ -1,12 +1,12 @@
 pkgname=video2x
-pkgver=r874.66c623f
+pkgver=r958.996b0bf
 pkgrel=1
-pkgdesc="A machine learning-based lossless video super resolution framework"
+pkgdesc="A machine learning-based video super resolution and frame interpolation framework"
 arch=('x86_64')
 url="https://github.com/k4yt3x/video2x"
 license=('AGPL3')
 depends=('ffmpeg' 'ncnn' 'vulkan-driver' 'spdlog' 'boost-libs')
-makedepends=('git' 'cmake' 'make' 'clang' 'pkgconf' 'vulkan-headers' 'openmp' 'boost')
+makedepends=('git' 'cmake' 'clang' 'vulkan-headers' 'openmp' 'boost')

 pkgver() {
    printf "r%s.%s" "$(git rev-list --count HEAD)" "$(git rev-parse --short HEAD)"
@ -17,7 +17,8 @@ prepare() {
 }

 build() {
-    cmake -B build -S .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr
+    cmake -B build -S .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr \
+        -DCMAKE_CXX_COMPILER=clang++ -DVIDEO2X_ENABLE_X86_64_V3=ON
    cmake --build build --config Release --parallel
 }

--- a/src/avutils.cpp
+++ b/src/avutils.cpp
@ -135,6 +135,7 @@ AVPixelFormat get_encoder_default_pix_fmt(const AVCodec *encoder, AVPixelFormat
    return best_pix_fmt;
 }

+[[gnu::target_clones("default", "avx2", "avx512f")]]
 float get_frame_diff(AVFrame *frame1, AVFrame *frame2) {
    if (!frame1 || !frame2) {
        logger()->error("Invalid frame(s) provided for comparison");
--- a/src/conversions.cpp
+++ b/src/conversions.cpp
@ -11,6 +11,7 @@ namespace video2x {
 namespace conversions {

 // Convert AVFrame format
+[[gnu::target_clones("default", "avx2", "avx512f")]]
 AVFrame *convert_avframe_pix_fmt(AVFrame *src_frame, AVPixelFormat pix_fmt) {
    AVFrame *dst_frame = av_frame_alloc();
    if (dst_frame == nullptr) {
@ -67,6 +68,7 @@ AVFrame *convert_avframe_pix_fmt(AVFrame *src_frame, AVPixelFormat pix_fmt) {
 }

 // Convert AVFrame to ncnn::Mat by copying the data
+[[gnu::target_clones("default", "avx2", "avx512f")]]
 ncnn::Mat avframe_to_ncnn_mat(AVFrame *frame) {
    AVFrame *converted_frame = nullptr;

@ -106,6 +108,7 @@ ncnn::Mat avframe_to_ncnn_mat(AVFrame *frame) {
 }

 // Convert ncnn::Mat to AVFrame with a specified pixel format (this part is unchanged)
+[[gnu::target_clones("default", "avx2", "avx512f")]]
 AVFrame *ncnn_mat_to_avframe(const ncnn::Mat &mat, AVPixelFormat pix_fmt) {
    int ret;

--- a/src/decoder.cpp
+++ b/src/decoder.cpp
@ -22,7 +22,7 @@ Decoder::~Decoder() {
    }
 }

-AVPixelFormat Decoder::get_hw_format(AVCodecContext *_, const AVPixelFormat *pix_fmts) {
+AVPixelFormat Decoder::get_hw_format(AVCodecContext *, const AVPixelFormat *pix_fmts) {
    for (const AVPixelFormat *p = pix_fmts; *p != AV_PIX_FMT_NONE; p++) {
        if (*p == hw_pix_fmt_) {
            return *p;
--- a/src/encoder.cpp
+++ b/src/encoder.cpp
@ -254,6 +254,7 @@ int Encoder::init(
    return 0;
 }

+[[gnu::target_clones("default", "avx2", "avx512f")]]
 int Encoder::write_frame(AVFrame *frame, int64_t frame_idx) {
    AVFrame *converted_frame = nullptr;
    int ret;
@ -325,6 +326,7 @@ int Encoder::write_frame(AVFrame *frame, int64_t frame_idx) {
    return 0;
 }

+[[gnu::target_clones("default", "avx2", "avx512f")]]
 int Encoder::flush() {
    int ret;
    AVPacket *enc_pkt = av_packet_alloc();
--- a/src/libvideo2x.cpp
+++ b/src/libvideo2x.cpp
@ -28,6 +28,7 @@ VideoProcessor::VideoProcessor(
      hw_device_type_(hw_device_type),
      benchmark_(benchmark) {}

+[[gnu::target_clones("default", "avx2", "avx512f")]]
 int VideoProcessor::process(
    const std::filesystem::path in_fname,
    const std::filesystem::path out_fname