使用指南

KML_CONV只提供filter2D的核心算子实现，用户需按使用指南将核心算子注册进开源OpenCV，并调用开源OpenCV filter2d接口完成滤波计算。

算子注册

获取OpenCV源码。
下载地址：https://github.com/opencv/opencv/archive/refs/tags/4.9.0.tar.gz
解压OpenCV安装包。
```
tar -xvf opencv-4.9.0.tar.gz
```

注册kml_conv核心算子。

opencv/3rdparty 目录下新增名为“kconv”的文件夹。
```
cd your_path_to_opencv/opencv-4.9.0/3rdparty
mkdir -p kconv
```

拷贝头文件及相应依赖库至“opencv/3rdparty/kconv”文件夹下。

拷贝conv头文件。

cp cv.h your_path_to_opencv/opencv-4.9.0/3rdparty/kconv

拷贝CONV、BLAS、FFT依赖库。

cp libkconv_ext.so your_path_to_opencv/opencv-4.9.0/3rdparty/Kunpeng
cp libkblas.so your_path_to_opencv/opencv-4.9.0/3rdparty/Kunpeng
cp libkfftf.so your_path_to_opencv/opencv-4.9.0/3rdparty/Kunpeng
cp libkffth.so your_path_to_opencv/opencv-4.9.0/3rdparty/Kunpeng

“opencv/3rdparty/Kunpeng”文件夹下新增头文件与源文件。
- 新增头文件kml_cv.hpp。
```
#ifndef KML_CV_HPP
#define KML_CV_HPP
#include <opencv2/core/base.hpp>
#include "cv.h"
#endif
```
- 新增源文件kml_cv.cpp。
```
#include "kml_cv.hpp"
```

“opencv/3rdparty/kconv”文件夹下新增CMakeLists.txt。

project(kml_conv_ext CXX)
add_library(kml_conv_ext STATIC kml_cv.cpp)
target_link_libraries(kml_conv_ext ${CMAKE_CURRENT_SOURCE_DIR}/libkconv.so ${CMAKE_CURRENT_SOURCE_DIR}/libkblas.so ${CMAKE_CURRENT_SOURCE_DIR}/libkfftf.so ${CMAKE_CURRENT_SOURCE_DIR}/libkffth.so)
set(KML_HAL_VERSION "0.0.1" PARENT_SCOPE)
set(KML_HAL_LIBRARIES "kml_conv_ext" PARENT_SCOPE)
set(KML_HAL_HEADERS "kml_conv_ext.hpp" PARENT_SCOPE)
set(KML_HAL_INCLUDE_DIRS "${CMAKE_BINARY_DIR}" PARENT_SCOPE)
set_target_properties(kml_conv_ext PROPERTIES LINKER_LANGUAGE CXX)
set_target_properties(kml_conv_ext PROPERTIES ARCHIVE_OUTPUT_DIRECTORY ${3P_LIBRARY_OUTPUT_PATH})
if(NOT BUILD_SHARED_LIBS)
  ocv_install_target(kail_dnn_ext EXPORT OpenCVModules ARCHIVE DESTINATION ${OPENCV_3P_LIB_INSTALL_PATH} COMPONENT dev)
endif()
set(OPENCV_SRC_DIR "${CMAKE_SOURCE_DIR}")
target_include_directories(kml_conv_ext PRIVATE ${CMAKE_CURRENT_SOURCE_DIR} ${OPENCV_SRC_DIR}/modules/core/include)
configure_file("${CMAKE_CURRENT_SOURCE_DIR}/kml_cv.hpp" "${CMAKE_BINARY_DIR}/kml_cv.hpp" COPYONLY)
configure_file("${CMAKE_CURRENT_SOURCE_DIR}/cv.h" "${CMAKE_BINARY_DIR}/cv.h" COPYONLY)
set_target_properties(${KML_LIBRARIES} PROPERTIES
        OUTPUT_NAME ${KML_LIBRARIES}
        DEBUG_POSTFIX "${OPENCV_DEBUG_POSTFIX}"
        COMPILE_PDB_NAME ${KML_LIBRARIES}
        COMPILE_PDB_NAME_DEBUG "${KML_LIBRARIES}${OPENCV_DEBUG_POSTFIX}"
        ARCHIVE_OUTPUT_DIRECTORY ${3P_LIBRARY_OUTPUT_PATH}
    )

注册KML_CONV高性能filter2D核心算子。

修改“opencv-4.9.0/CMakeLists.txt”第965-995行。

set(OpenCV_HAL "kml_conv_ext")

foreach(hal ${OpenCV_HAL})
  if(hal STREQUAL "carotene")
    ...
  elseif(hal STREQUAL "openvx")
    add_subdirectory(3rdparty/openvx)
    ocv_hal_register(OPENVX_HAL_LIBRARIES OPENVX_HAL_HEADERS OPENVX_HAL_INCLUDE_DIRS)
    list(APPEND OpenCV_USED_HAL "openvx (ver ${OPENVX_HAL_VERSION})")
  elseif(hal STREQUAL "kml_conv_ext")
    add_subdirectory(3rdparty/kconv)
    ocv_hal_register(KML_HAL_LIBRARIES KML_HEADERS KML_INCLUDE_DIRS)
    list(APPEND OpenCV_USED_HAL "kml_conv_ext (ver ${KML_VERSION})")
  else()
    ocv_debug_message(STATUS "OpenCV HAL: ${hal} ...")
    ocv_clear_vars(OpenCV_HAL_LIBRARIES OpenCV_HAL_HEADERS OpenCV_HAL_INCLUDE_DIRS)
    find_package(${hal} NO_MODULE QUIET)
    if(${hal}_FOUND)
      ocv_hal_register(OpenCV_HAL_LIBRARIES OpenCV_HAL_HEADERS OpenCV_HAL_INCLUDE_DIRS)
      list(APPEND OpenCV_USED_HAL "${hal} (ver ${${hal}_VERSION})")
    endif()
endif()

编译安装OpenCV。

cd your_path_to_opencv/opencv-4.9.0
mkdir -p build
cd build
cmake .. -DWITH_ADE=OFF -DCMAKE_INSTALL_PREFIX=your_path_to_install_opencv # cmake到build文件夹下
make -j 
make install # 安装到your_path_to_install_opencv中

OpenCV filter2D接口定义

完成上述KML_CONV filter2D算子注册后，用户可直接调用OpenCV filter2D接口完成滤波计算。

OpenCV filter2D使用指南：

https://docs.opencv.org/4.9.0/d4/d86/group__imgproc__filter.html

C++ interface：

void cv::filter2D( InputArray src, OutputArray dst, int ddepth, InputArray kernel, Point anchor = Point(-1,-1), double delta = 0, int borderType = BORDER_DEFAULT)

参数列表：

参数名	类型	描述	输入/输出
src	InputArray	输入图像	输入
dst	OutputArray	输出图像	输出
int	ddepth	输出图像的深度	输入
kernel	InputArray	卷积核	输入
anchor	Point	卷积核锚点，表示滤波起始点在卷积核中的相对位置默认值为为Point(-1,-1),表示滤波起始点位于卷积核中心	输入
delta	0	偏置值，默认值为0	输入
int	borderType	边界填充类型，支持以下入参: BORDER_CONSTANT BORDER_REPLICATE BORDER_REFLECT BORDER_REFLECT_101 BORDER_REFLECT101 BORDER_DEFAULT（默认值） BORDER_ISOLATED	输入

完成上述KML_CONV算子注册后的OpenCV 将新增对CV_16F图像深度filter2D计算的功能支持。下表是OpenCV filter2D接口所支持的输入图像深度、输出图像深度组合。

输入图像深度	输出图像深度
CV_8U	-1/CV_16S/CV_32F/CV_64F （输出图像深度值为-1时，表示输出图像深度与输入图像深度相同）
CV_16U/CV_16S	-1/CV_32F/CV_64F
CV_32F	-1/CV_32F
CV_64F	-1/CV_64F
CV_16F	-1/CV_16F

当输入图像深度与输出图像深度均为CV_32F或CV_16F时，将调用KML_CONV高性能算子实现完成滤波计算。其他情况下，将调用OpenCV自身算子实现完成滤波计算。

依赖

#include <opencv2/imgproc/imgproc.hpp>

示例

C++ interface：

    cv::Mat src = (cv::Mat_<float>(6, 6) << 1, 2, 3, 4, 5, 6,
                                            7, 8, 9, 10, 11, 12,
                                            13, 14, 15, 16, 17, 18,
                                            19, 20, 21, 22, 23, 24,
                                            25, 26, 27, 28, 29, 30,
                                            31, 32, 33, 34, 35, 36);

    cv::Mat kernel = (cv::Mat_<float>(3, 3) << 0, -1, 0,
                                               -1, 5, -1,
                                               0, -1, 0);
    cv::Mat dst;
    cv::filter2D(src, dst, -1, kernel, cv::Point(-1, -1), 0, cv::BORDER_CONSTANT);
    /*
     * dst = [-4, -2, 0, 2, 4, 13;
     *        13, 8, 9, 10, 11, 25;
     *        25, 14, 15, 16, 17, 37;
     *        37, 20, 21, 22, 23, 49;
     *        49, 26, 27, 28, 29, 61;
     *        98, 70, 72, 74, 76, 115]
     */

父主题： 图像处理接口