OpenCV算法加速（4）官方源码v4.5.5的默认并行和优化加速的编译选项是什么？请重点关注函数cv::getBuildInformation()的返回值
2024-05-26 06:50:28
举例opencv v4.5.5版本源码，Windows x64，VS2019，CMake
https://github.com/opencv/opencv/tree/4.5.5
https://sourceforge.net/projects/opencvlibrary/files/4.5.5/
下载安装opencv-4.5.5-vc14_vc15.exe，得到官方编译的动态库opencv_world455.dll
1、多线程并行库的使用情况
PPL
\opencv\sources\modules\core\src\parallel.cpp，里面默认使用的是微软PPL加速方式，
#if defined _MSC_VER && _MSC_VER >= 1600#define HAVE_CONCURRENCY
#endifcmake里面的编译选项，其他方式如TBB，HPX，OPENMP默认是关闭的。
\opencv\sources\x64\cvconfig.h，里面有宏定义的。
/* Halide support */
/* #undef HAVE_HALIDE *//* Intel Integrated Performance Primitives */
/* #undef HAVE_IPP */
/* #undef HAVE_IPP_ICV */
/* #undef HAVE_IPP_IW */
/* #undef HAVE_IPP_IW_LL *//* Intel Threading Building Blocks */
/* #undef HAVE_TBB *//* Ste||ar Group High Performance ParallelX */
/* #undef HAVE_HPX *//* OpenVX */
/* #undef HAVE_OPENVX */2、GPU之OpenCL
cmake里面的编译选项，默认是开启OpenCL的。
\opencv\sources\x64\cvconfig.h
/* OpenCL Support */
#define HAVE_OPENCL
/* #undef HAVE_OPENCL_STATIC */
/* #undef HAVE_OPENCL_SVM *//* NVIDIA OpenCL D3D Extensions support */
#define HAVE_OPENCL_D3D11_NVGPU之CUDA
cmake里面的编译选项，默认是禁用CUDA的。需要用户自己勾选，并下载库编译。
3、SIMD指令集优化选项
cmake编译选项，\opencv\sources\x64\cvconfig.h，宏定义CV_ENABLE_INTRINSICS默认是开启的
#ifndef OPENCV_CVCONFIG_H_INCLUDED
#define OPENCV_CVCONFIG_H_INCLUDED/* OpenCV compiled as static or dynamic libs */
#define BUILD_SHARED_LIBS/* OpenCV intrinsics optimized code */
#define CV_ENABLE_INTRINSICS/* OpenCV additional optimized code */
/* #undef CV_DISABLE_OPTIMIZATION */
cmake编译选项，\opencv\sources\x64\cv_cpu_config.h
// OpenCV CPU baseline features#define CV_CPU_COMPILE_SSE 1
#define CV_CPU_BASELINE_COMPILE_SSE 1#define CV_CPU_COMPILE_SSE2 1
#define CV_CPU_BASELINE_COMPILE_SSE2 1#define CV_CPU_COMPILE_SSE3 1
#define CV_CPU_BASELINE_COMPILE_SSE3 1#define CV_CPU_BASELINE_FEATURES 0 \, CV_CPU_SSE \, CV_CPU_SSE2 \, CV_CPU_SSE3 \// OpenCV supported CPU dispatched features#define CV_CPU_DISPATCH_COMPILE_SSE4_1 1
#define CV_CPU_DISPATCH_COMPILE_SSE4_2 1
#define CV_CPU_DISPATCH_COMPILE_FP16 1
#define CV_CPU_DISPATCH_COMPILE_AVX 1
#define CV_CPU_DISPATCH_COMPILE_AVX2 1
#define CV_CPU_DISPATCH_COMPILE_AVX512_SKX 1#define CV_CPU_DISPATCH_FEATURES 0 \, CV_CPU_SSE4_1 \, CV_CPU_SSE4_2 \, CV_CPU_FP16 \, CV_CPU_AVX \, CV_CPU_AVX2 \, CV_CPU_AVX512_SKX \关于OpenCV硬件加速，OpenCV Hardware Acceleration Layer(HAL)，请参考博文：
OpenCV中的HAL方法调用流程分析 - willhua - 博客园OpenCV中的HAL方法调用流程分析 在OpenCV中有一些所谓HAL(Hardware Acceleration Layer)实现，看名字好像和硬件相关，其实也不尽然，可以理解为比常规的OCV实现https://www.cnblogs.com/willhua/p/12521581.html
4、IPP加速
Intel® Integrated Performance Primitives (简称Intel® IPP)是一个软件库，提供了大量的函数。包括信号处理、图像处理、计算机视觉、数据压缩和字符串操作。通过对函数的优化，比如适配指令集操作等来提升运行效率。
完整版的IPP下载地址：
https://www.intel.com/content/www/us/en/developer/tools/oneapi/ipp.html
在 Learning OpenCV 这本书中，作者提到OpenCV可以利用Intel的IPP性能库来提升程序的运行速度，而这个IPP库是要另外进行购买的。实际上，Intel为当前的OpenCV免费提供了IPP加速库的一部分，在此我们称之为ippicv。ippicv会在cmake的时候自动从github上下载，但是在网络状况不佳的情况下会下载失败。这时候我们只能采用手动安装的方式。ippicv的下载地址其实就藏在ippicv.cmake文件中。
https://github.com/opencv/opencv/blob/4.5.5/3rdparty/ippicv/ippicv.cmake
里面的变量组合起来就是ippicv的下载网址：
https://raw.githubusercontent.com/opencv/opencv_3rdparty/a56b6ac6f030c312b2dce17430eef13aed9af274/ippicv/ippicv_2020_win_intel64_20191018_general.zip
或者直接访问这个网址也可以获得：
https://github.com/opencv/opencv_3rdparty/tree/ippicv/master_20191018/ippicv
如果在英特尔的处理器上使用，OpenCV就会自动使用一种免费的英特尔集成性能原语库（IPP）的子集，IPP 8.x(IPPICV)。IPPICV可以在编译阶段链接到OpenCV，这样一来，会替代相应的低级优化的C语言代码（在cmake中设置WITH_IPP=ON/OFF开启或者关闭这一功能，默认情况为开启）。使用IPP获得的速度提升非常可观，如下图显示了使用IPP之后得到的加速效果。
5、怎么知道当前OpenCV开启了哪些加速模式？
源码\opencv\sources\modules\core\src\system.cpp，有专门的函数设置和查看优化的情况：
volatile bool useOptimizedFlag = true;void setUseOptimized( bool flag )
{useOptimizedFlag = flag;currentFeatures = flag ? &featuresEnabled : &featuresDisabled;ipp::setUseIPP(flag);
#ifdef HAVE_OPENCLocl::setUseOpenCL(flag);
#endif
}bool useOptimized(void)
{return useOptimizedFlag;
}void setUseIPP(bool flag)
{CoreTLSData& data = getCoreTlsData();
#ifdef HAVE_IPPdata.useIPP = (getIPPSingleton().useIPP)?flag:false;
#elseCV_UNUSED(flag);data.useIPP = false;
#endif
}\opencv\sources\modules\core\src\ocl.cpp
void setUseOpenCL(bool flag)
{if (!flag)useOpenCL_ = 0;elseuseOpenCL_ = -1;
}已知宏HAVE_OPENCL是已定义的，HAVE_IPP也是已定义的。所以OpenCL和IPP的功能默认是启动的。
可以通过以下测试用例，来查询当前的opencv是否开启了OPENCL和IPP
#include <iostream>
#include <opencv2/core.hpp>
#include <opencv2/core/cuda.hpp>
#include <opencv2/core/ocl.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgcodecs.hpp>
#include <opencv2/imgproc.hpp>void checkIPP()
{bool use = cv::ipp::useIPP();bool useNE = cv::ipp::useIPP_NotExact();int status = cv::ipp::getIppStatus();cv::String ver = cv::ipp::getIppVersion();cv::String loc = cv::ipp::getIppErrorLocation();std::cout << "IPP use:" << use << std::endl;std::cout << "IPP useNE:" << useNE << std::endl;std::cout << "IPP status:" << status << std::endl;std::cout << "IPP ver:" << ver << std::endl;std::cout << "IPP loc:" << loc << std::endl;
}void checkOCL()
{//查询opencv当前是否开启了OpenCL功能bool ret1 = cv::ocl::haveOpenCL();bool ret2 = cv::ocl::useOpenCL();std::cout << "default haveOpenCL:" << ret1 << std::endl;std::cout << "default useOpenCL:" << ret2 << std::endl;
}6、最后，举例来看看canny算子的优化情况是什么？
\opencv\sources\modules\imgproc\src\canny.cpp，源码有宏定义#if CV_SIMD，
如果使能了宏定义CV_SIMD，就会开启指令集的优化源码。CV_SIMD默认是开启的，在头文件
\opencv\sources\modules\core\include\opencv2\core\hal\intrin.hpp，有定义
#if (CV_SSE2 || CV_NEON || CV_VSX || CV_MSA || CV_WASM_SIMD || CV_RVV071 || CV_RVV) && !defined(CV_FORCE_SIMD128_CPP)
#define CV__SIMD_FORWARD 128
#include "opencv2/core/hal/intrin_forward.hpp"
#endif#if CV_SSE2 && !defined(CV_FORCE_SIMD128_CPP)#include "opencv2/core/hal/intrin_sse_em.hpp"
#include "opencv2/core/hal/intrin_sse.hpp"namespace CV__SIMD_NAMESPACE {#define CV_SIMD CV_SIMD128另外，canny算子默认是开启了PPL并行计算的，并且也默认开启了对OpenCL和IPP的支持，一起来看看源码\opencv\sources\modules\imgproc\src\canny.cpp，
void Canny( InputArray _src, OutputArray _dst,double low_thresh, double high_thresh,int aperture_size, bool L2gradient )
{......CV_OCL_RUN(_dst.isUMat() && (_src.channels() == 1 || _src.channels() == 3),ocl_Canny<false>(_src, UMat(), UMat(), _dst, (float)low_thresh, (float)high_thresh, aperture_size, L2gradient, _src.channels(), size))......CALL_HAL(canny, cv_hal_canny, src.data, src.step, dst.data, dst.step, src.cols, src.rows, src.channels(),low_thresh, high_thresh, aperture_size, L2gradient);......CV_IPP_RUN_FAST(ipp_Canny(src, Mat(), Mat(), dst, (float)low_thresh, (float)high_thresh, L2gradient, aperture_size))......parallel_for_(Range(0, src.rows), parallelCanny(src, map, stack, low, high, aperture_size, L2gradient), numOfThreads);......
}void Canny( InputArray _dx, InputArray _dy, OutputArray _dst,double low_thresh, double high_thresh,bool L2gradient )
{......CV_OCL_RUN(_dst.isUMat(),ocl_Canny<true>(UMat(), _dx.getUMat(), _dy.getUMat(), _dst, (float)low_thresh, (float)high_thresh, 0, L2gradient, _dx.channels(), size))......CV_IPP_RUN_FAST(ipp_Canny(Mat(), dx, dy, dst, (float)low_thresh, (float)high_thresh, L2gradient, 0))......parallel_for_(Range(0, dx.rows), parallelCanny(dx, dy, map, stack, low, high, L2gradient), numOfThreads);......
}\opencv\sources\modules\imgproc\src\hal_replacement.hpp
inline int hal_ni_canny(const uchar* src_data, size_t src_step, uchar* dst_data, size_t dst_step,
int width, int height, int cn, double lowThreshold, double highThreshold, int ksize, bool L2gradient)
{ return CV_HAL_ERROR_NOT_IMPLEMENTED; }//! @cond IGNORED
#define cv_hal_canny hal_ni_canny
//! @endcond注意：
（1）CV_OCL_RUN有参数_dst.isUMat()，表明OpenCL需要使用UMat作为图像格式的输入，否则会失效；
（2）CALL_HAL是硬件加速，但是深入源码却发现返回值是CV_HAL_ERROR_NOT_IMPLEMENTED，表明该功能官方暂时未实现。
（3）以上两个如果都不执行，那么就执行CV_IPP_RUN_FAST(ipp_Canny...
\opencv\sources\modules\core\include\opencv2\core\private.hpp
#define CV_IPP_RUN_(condition, func, ...) \{ \if (cv::ipp::useIPP() && (condition)) \{ \CV__TRACE_REGION_("IPP:" #func, CV_TRACE_NS::details::REGION_FLAG_IMPL_IPP) \if(func) \{ \CV_IMPL_ADD(CV_IMPL_IPP); \} \else \{ \setIppErrorStatus(); \CV_Error(cv::Error::StsAssert, #func); \} \return __VA_ARGS__; \} \}
#else
#define CV_IPP_RUN_(condition, func, ...) \if (cv::ipp::useIPP() && (condition)) \{ \CV__TRACE_REGION_("IPP:" #func, CV_TRACE_NS::details::REGION_FLAG_IMPL_IPP) \if(func) \{ \CV_IMPL_ADD(CV_IMPL_IPP); \return __VA_ARGS__; \} \}
#endif
#else
#define CV_IPP_RUN_(condition, func, ...)
#endif#define CV_IPP_RUN_FAST(func, ...) CV_IPP_RUN_(true, func, __VA_ARGS__)
#define CV_IPP_RUN(condition, func, ...) CV_IPP_RUN_((condition), (func), __VA_ARGS__)x1、写个测试用例，请重点关注函数cv::getBuildInformation()
//OpenCV算法加速（4）官方源码v4.5.5的默认并行加速的编译选项是什么？
//https://libaineu2004.blog.csdn.net/article/details/122677388
//OpenCV 加速图像处理
//https://www.cnblogs.com/ybqjymy/p/13691132.html
//为什么OpenCV用GPU/cuda跑得比用CPU慢？
//https://blog.csdn.net/libaineu2004/article/details/129801112//执行这些简单算子,CPU比GPU更快
//cvtColor,GaussianBlur,Canny
//执行这些耗时算子,GPU比CPU更快
//HoughCircles,HoughLines,matchTemplate#include <iostream>
#include <opencv2/core.hpp>
#include <opencv2/core/cuda.hpp>
#include <opencv2/core/ocl.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgcodecs.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/cudaimgproc.hpp>
#include <opencv2/cudaarithm.hpp>
#include <opencv2/cudafilters.hpp>
#include <opencv2/cudawarping.hpp>#define IMAGE_TEST_PATHNAME "D:\\test_src.jpg"
#define IMAGE_SOURCE "D:\\test_src.jpg"
#define IMAGE_TEMPLATE "D:\\test_templ.jpg"void checkBuild()
{//查询opencv编译时配置std::cout << cv::getBuildInformation() << std::endl;
}void checkIPP()
{//Intel® Integrated Performance Primitives//https://www.intel.com/content/www/us/en/developer/tools/oneapi/ipp.htmlbool use = cv::ipp::useIPP();bool useNE = cv::ipp::useIPP_NotExact();int status = cv::ipp::getIppStatus();cv::String ver = cv::ipp::getIppVersion();cv::String loc = cv::ipp::getIppErrorLocation();std::cout << "IPP use:" << use << std::endl;std::cout << "IPP useNE:" << useNE << std::endl;std::cout << "IPP status:" << status << std::endl;std::cout << "IPP ver:" << ver << std::endl;std::cout << "IPP loc:" << loc << std::endl;
}void checkSimd()
{//查询opencv线程int numTh = cv::getNumThreads(); //默认值是cpu的逻辑线程数int numCore = cv::getNumberOfCPUs();std::cout << "getNumThreads=" << numTh << std::endl;std::cout << "getNumberOfCPUs=" << numCore << std::endl;//查询opencv当前是否开启了并行优化功能bool opt = cv::useOptimized(); //默认值是truestd::cout << "useOptimized=" << opt << std::endl;//查询opencv当前是否开启了OpenCL功能bool ret1 = cv::ocl::haveOpenCL();bool ret2 = cv::ocl::useOpenCL();std::cout << "default haveOpenCL:" << ret1 << std::endl;std::cout << "default useOpenCL:" << ret2 << std::endl;//查询opencv当前是否支持具体的CPU指令集bool check1 = cv::checkHardwareSupport(CV_CPU_SSE4_1);bool check2 = cv::checkHardwareSupport(CV_CPU_SSE4_2);bool check3 = cv::checkHardwareSupport(CV_CPU_AVX2);std::cout << "CV_CPU_SSE4_1=" << check1 << std::endl;std::cout << "CV_CPU_SSE4_2=" << check2 << std::endl;std::cout << "CV_CPU_AVX2=" << check3 << std::endl;//查询完整的硬件支持清单std::cout << "HardwareSupport:" << std::endl;std::cout << "CV_CPU_MMX: " << cv::checkHardwareSupport(CV_CPU_MMX) << std::endl;std::cout << "CV_CPU_SSE: " << cv::checkHardwareSupport(CV_CPU_SSE) << std::endl;std::cout << "CV_CPU_SSE2: " << cv::checkHardwareSupport(CV_CPU_SSE2) << std::endl;std::cout << "CV_CPU_SSE3: " << cv::checkHardwareSupport(CV_CPU_SSE3) << std::endl;std::cout << "CV_CPU_SSSE3: " << cv::checkHardwareSupport(CV_CPU_SSSE3) << std::endl;std::cout << "CV_CPU_SSE4_1: " << cv::checkHardwareSupport(CV_CPU_SSE4_1) << std::endl;std::cout << "CV_CPU_SSE4_2: " << cv::checkHardwareSupport(CV_CPU_SSE4_2) << std::endl;std::cout << "CV_CPU_POPCNT: " << cv::checkHardwareSupport(CV_CPU_POPCNT) << std::endl;std::cout << "CV_CPU_FP16: " << cv::checkHardwareSupport(CV_CPU_FP16) << std::endl;std::cout << "CV_CPU_AVX: " << cv::checkHardwareSupport(CV_CPU_AVX) << std::endl;std::cout << "CV_CPU_AVX2: " << cv::checkHardwareSupport(CV_CPU_AVX2) << std::endl;std::cout << "CV_CPU_FMA3: " << cv::checkHardwareSupport(CV_CPU_FMA3) << std::endl;std::cout << "CV_CPU_AVX_512F: " << cv::checkHardwareSupport(CV_CPU_AVX_512F) << std::endl;std::cout << "CV_CPU_AVX_512BW: " << cv::checkHardwareSupport(CV_CPU_AVX_512BW) << std::endl;std::cout << "CV_CPU_AVX_512CD: " << cv::checkHardwareSupport(CV_CPU_AVX_512CD) << std::endl;std::cout << "CV_CPU_AVX_512DQ: " << cv::checkHardwareSupport(CV_CPU_AVX_512DQ) << std::endl;std::cout << "CV_CPU_AVX_512ER: " << cv::checkHardwareSupport(CV_CPU_AVX_512ER) << std::endl;std::cout << "CV_CPU_AVX_512IFMA512: " << cv::checkHardwareSupport(CV_CPU_AVX_512IFMA512) << std::endl;std::cout << "CV_CPU_AVX_512IFMA: " << cv::checkHardwareSupport(CV_CPU_AVX_512IFMA) << std::endl;std::cout << "CV_CPU_AVX_512PF: " << cv::checkHardwareSupport(CV_CPU_AVX_512PF) << std::endl;std::cout << "CV_CPU_AVX_512VBMI: " << cv::checkHardwareSupport(CV_CPU_AVX_512VBMI) << std::endl;std::cout << "CV_CPU_AVX_512VL: " << cv::checkHardwareSupport(CV_CPU_AVX_512VL) << std::endl;std::cout << "CV_CPU_NEON: " << cv::checkHardwareSupport(CV_CPU_NEON) << std::endl;std::cout << "CV_CPU_VSX: " << cv::checkHardwareSupport(CV_CPU_VSX) << std::endl;std::cout << "CV_CPU_AVX512_SKX: " << cv::checkHardwareSupport(CV_CPU_AVX512_SKX) << std::endl;std::cout << "CV_HARDWARE_MAX_FEATURE: " << cv::checkHardwareSupport(CV_HARDWARE_MAX_FEATURE) << std::endl;std::cout << std::endl;//cv::setUseOptimized(false);//cv::setNumThreads(1);
}void checkCuda() //旧版本的是cv::gpu,#include <opencv2/gpu/gpu.hpp>,已弃用
{int64 begintime, endtime;int num_devices = cv::cuda::getCudaEnabledDeviceCount();if (num_devices <= 0){std::cerr << "There is no cuda device" << std::endl;return;}int enable_device_id = -1;for (int i = 0; i < num_devices; i++){cv::cuda::DeviceInfo dev_info(i);if (dev_info.isCompatible()){enable_device_id = i;}}if (enable_device_id < 0){std::cerr << "GPU module isn't built for GPU" << std::endl;return;}cv::cuda::setDevice(enable_device_id); //指定显卡//有一个问题是，一般使用GPU加速的话，第一次调用GPU，会很慢很慢，一条简单的语句都用了10多秒左右。//治标不治本的解决方法是在程序的开头加上一句cv::gpu::GpuMata(10, 10, CV_8U);//这样会令耗时的操作放在一开头，不那么影响后面的操作时间//为什么第一次函数调用很慢//那是因为初始化开销;在第一个GPU函数调用Cuda Runtime API被隐式初始化;cv::cuda::GpuMat(10, 10, CV_8U);//测试用例cv::Mat src_image = cv::imread(IMAGE_TEST_PATHNAME);cv::Mat dst_image;cv::cuda::GpuMat d_src_img(src_image); //upload src image to gpu//或者d_src_img.upload(src_image);cv::cuda::GpuMat d_dst_img;begintime = cv::getTickCount();cv::cuda::cvtColor(d_src_img, d_dst_img, cv::COLOR_BGR2GRAY); //cannyd_dst_img.download(dst_image); //download dst image to cpuendtime = cv::getTickCount();std::cerr << 1000 * (endtime - begintime) / cv::getTickFrequency() << std::endl;cv::namedWindow("checkCuda", cv::WINDOW_NORMAL);cv::imshow("checkCuda", dst_image);
}void checkOpenCL() //Open Computing Language:开放计算语言,可以附加在主机处理器的CPU或GPU上执行
{std::vector<cv::ocl::PlatformInfo> info;cv::ocl::getPlatfomsInfo(info);cv::ocl::PlatformInfo sdk = info.at(0);int number = sdk.deviceNumber();if (number < 1){std::cout << "Number of devices:" << number << std::endl;return;}std::cout << "***********SDK************" << std::endl;std::cout << "Name:" << sdk.name() << std::endl;std::cout << "Vendor:" << sdk.vendor() << std::endl;std::cout << "Version:" << sdk.version() << std::endl;std::cout << "Version:" << sdk.version() << std::endl;std::cout << "Number of devices:" << number << std::endl;for (int i = 0; i < number; i++){std::cout << std::endl;cv::ocl::Device device;sdk.getDevice(device, i);std::cout << "***********Device " << i + 1 << "***********" << std::endl;std::cout << "Vendor Id:" << device.vendorID() << std::endl;std::cout << "Vendor name:" << device.vendorName() << std::endl;std::cout << "Name:" << device.name() << std::endl;std::cout << "Driver version:" << device.vendorID() << std::endl;if (device.isAMD())std::cout << "Is AMD device" << std::endl;if (device.isIntel())std::cout << "Is Intel device" << std::endl;if (device.isNVidia())std::cout << "Is NVidia device" << std::endl;std::cout << "Global Memory size:" << device.globalMemSize() << std::endl;std::cout << "Memory cache size:" << device.globalMemCacheSize() << std::endl;std::cout << "Memory cache type:" << device.globalMemCacheType() << std::endl;std::cout << "Local Memory size:" << device.localMemSize() << std::endl;std::cout << "Local Memory type:" << device.localMemType() << std::endl;std::cout << "Max Clock frequency:" << device.maxClockFrequency() << std::endl;}
}void calcEdgesCPU()
{cv::ocl::setUseOpenCL(false);bool ret1 = cv::ocl::haveOpenCL();bool ret2 = cv::ocl::useOpenCL();std::cout << "haveOpenCL:" << ret1 << std::endl;std::cout << "useOpenCL:" << ret2 << std::endl;double start = cv::getTickCount();cv::Mat cpuGray, cpuBlur, cpuEdges;cv::Mat cpuFrame = cv::imread(IMAGE_TEST_PATHNAME);cvtColor(cpuFrame, cpuGray, cv::COLOR_BGR2GRAY);cv::GaussianBlur(cpuGray, cpuBlur, cv::Size(3, 3), 15, 15);cv::Canny(cpuBlur, cpuEdges, 50, 100, 3);std::vector<cv::Vec3f> vtCir;cv::HoughCircles(cpuBlur, vtCir, cv::HOUGH_GRADIENT_ALT, 1.5, 15, 300, 0.8, 1, 100);cv::HoughCircles(cpuBlur, vtCir, cv::HOUGH_GRADIENT, 1, 15, 100, 30, 1, 100);std::cout << "CPU cost time:(s)" << ((cv::getTickCount() - start) / cv::getTickFrequency()) << std::endl;cv::namedWindow("Canny Edges CPU", cv::WINDOW_NORMAL);cv::imshow("Canny Edges CPU", cpuEdges);
}void calcEdgesOpenCL()
{cv::ocl::setUseOpenCL(true);bool ret1 = cv::ocl::haveOpenCL();bool ret2 = cv::ocl::useOpenCL();std::cout << "haveOpenCL:" << ret1 << std::endl;std::cout << "useOpenCL:" << ret2 << std::endl;//通过使用UMat对象，OpenCV会自动在支持OpenCL的设备上使用GPU运算，在不支持OpenCL的设备仍然使用CPU运算，这样就避免了程序运行失败，而且统一了接口。double start = cv::getTickCount();cv::UMat gpuFrame, gpuGray, gpuBlur, gpuEdges;cv::Mat cpuFrame = cv::imread(IMAGE_TEST_PATHNAME);cpuFrame.copyTo(gpuFrame); //Mat与UMat相互转换cvtColor(gpuFrame, gpuGray, cv::COLOR_BGR2GRAY);cv::GaussianBlur(gpuGray, gpuBlur, cv::Size(3, 3), 15, 15);cv::Canny(gpuBlur, gpuEdges, 50, 100, 3);std::vector<cv::Vec3f> vtCir;cv::HoughCircles(gpuBlur, vtCir, cv::HOUGH_GRADIENT_ALT, 1.5, 15, 300, 0.8, 1, 100);cv::HoughCircles(gpuBlur, vtCir, cv::HOUGH_GRADIENT, 1, 15, 100, 30, 1, 100);std::cout << "OpenCL cost time:(s)" << ((cv::getTickCount() - start) / cv::getTickFrequency()) << std::endl;cv::Mat matResult = gpuEdges.getMat(cv::ACCESS_READ); //Mat与UMat相互转换cv::namedWindow("Canny Edges OpenCL1", cv::WINDOW_NORMAL);cv::imshow("Canny Edges OpenCL1", matResult);cv::namedWindow("Canny Edges OpenCL2", cv::WINDOW_NORMAL);cv::imshow("Canny Edges OpenCL2", gpuEdges);
}void calcEdgesCuda()
{cv::ocl::setUseOpenCL(false);double start = cv::getTickCount();cv::cuda::GpuMat gpuGray, gpuBlur, gpuEdges;cv::Mat cpuEdges;cv::Mat cpuFrame = cv::imread(IMAGE_TEST_PATHNAME);cv::cuda::registerPageLocked(cpuFrame); //锁页内存cv::cuda::GpuMat gpuFrame;gpuFrame.upload(cpuFrame);cv::cuda::cvtColor(gpuFrame, gpuGray, cv::COLOR_BGR2GRAY);cv::Ptr<cv::cuda::Filter> gaussFilter = cv::cuda::createGaussianFilter(CV_8UC1, CV_8UC1, cv::Size(3, 3), 15, 15);gaussFilter->apply(gpuGray, gpuBlur);cv::Ptr<cv::cuda::CannyEdgeDetector> cannyEdge = cv::cuda::createCannyEdgeDetector(50, 100, 3);cannyEdge->detect(gpuBlur, gpuEdges);cv::cuda::GpuMat gpuLines; //This should be GpuMat...
#if 0 //find linestd::vector<cv::Vec2f> vtLines;cv::Ptr<cv::cuda::HoughLinesDetector> hough = cv::cuda::createHoughLinesDetector(1, CV_PI / 180, 120);hough->detect(gpuEdges, gpuLines);hough->downloadResults(gpuLines, vtLines);
#elsecv::Ptr<cv::cuda::HoughCirclesDetector> hough1 = cv::cuda::createHoughCirclesDetector(1.5, 15, 300, 1, 1, 100);hough1->detect(gpuEdges, gpuLines);cv::Ptr<cv::cuda::HoughCirclesDetector> hough2 = cv::cuda::createHoughCirclesDetector(1, 15, 100, 30, 1, 100);hough2->detect(gpuEdges, gpuLines);
#endifgpuEdges.download(cpuEdges);cv::cuda::unregisterPageLocked(cpuFrame); //解除锁页std::cout << "Cuda cost time:(s)" << ((cv::getTickCount() - start) / cv::getTickFrequency()) << std::endl;cv::namedWindow("Canny Edges Cuda", cv::WINDOW_NORMAL);cv::imshow("Canny Edges Cuda", cpuEdges);
}//https://stackoverflow.com/questions/75571990/opencv-cuda-different-outcomes-on-cpu-vs-gpu
cv::cuda::GpuMat extractTargetsGPU(cv::cuda::GpuMat frame)
{std::vector<cv::cuda::GpuMat> channels;cv::cuda::split(frame, channels);cv::cuda::Stream bThresholdStream, gThresholdStream, rThresholdStream;cv::cuda::threshold(channels[0], channels[0], 127, 255, cv::THRESH_BINARY, bThresholdStream);cv::cuda::threshold(channels[1], channels[1], 127, 255, cv::THRESH_BINARY, gThresholdStream);cv::cuda::threshold(channels[2], channels[2], 127, 255, cv::THRESH_BINARY_INV, rThresholdStream);bThresholdStream.waitForCompletion();gThresholdStream.waitForCompletion();rThresholdStream.waitForCompletion();cv::cuda::bitwise_and(channels[1], channels[2], channels[1]);cv::cuda::bitwise_and(channels[0], channels[1], channels[0]);cv::Ptr<cv::cuda::Filter> gaussianFilter = cv::cuda::createGaussianFilter(CV_8UC1, CV_8UC1, cv::Size(9, 9), 1.0f);gaussianFilter->apply(channels[0], channels[0]);cv::cuda::GpuMat targets;cv::Ptr<cv::cuda::HoughCirclesDetector> houghCircles = cv::cuda::createHoughCirclesDetector(1, 20, 50, 30, 1, 75);houghCircles->detect(channels[0], targets);return targets;
}void matchTemplateCPU()
{cv::Mat src = cv::imread(IMAGE_SOURCE, cv::IMREAD_GRAYSCALE);cv::Mat templ = cv::imread(IMAGE_TEMPLATE, cv::IMREAD_GRAYSCALE);cv::Mat dst;double minVal = 0;double maxVal = 0;cv::Point minLoc;cv::Point maxLoc;double start = cv::getTickCount();cv::matchTemplate(src, templ, dst, cv::TM_CCOEFF_NORMED); //用6种匹配方式cv::normalize(dst, dst, 1, 0, cv::NORM_MINMAX);cv::minMaxLoc(dst, &minVal, &maxVal, &minLoc, &maxLoc); //找到最佳匹配点std::cout << "matchTemplateCPU cost time:(s)" << ((cv::getTickCount() - start) / cv::getTickFrequency()) << std::endl;cv::rectangle(src, cv::Rect(maxLoc.x, maxLoc.y, templ.cols, templ.rows), 1, 8, 0);cv::namedWindow("matchTemplateCPU", cv::WINDOW_NORMAL);cv::imshow("matchTemplateCPU", src);
}void matchTemplateCuda()
{cv::Mat src = cv::imread(IMAGE_SOURCE, cv::IMREAD_GRAYSCALE);cv::Mat templ = cv::imread(IMAGE_TEMPLATE, cv::IMREAD_GRAYSCALE);cv::Mat dst;double minVal = 0;double maxVal = 0;cv::Point minLoc;cv::Point maxLoc;cv::cuda::GpuMat gsrc, gtempl, gdst;double start = cv::getTickCount();gsrc.upload(src);gtempl.upload(templ);cv::Ptr<cv::cuda::TemplateMatching> matcher;matcher = cv::cuda::createTemplateMatching(CV_8U, cv::TM_CCOEFF_NORMED);matcher->match(gsrc, gtempl, gdst);cv::cuda::minMaxLoc(gdst, &minVal, &maxVal, &minLoc, &maxLoc);std::cout << "matchTemplateCuda cost time:(s)" << ((cv::getTickCount() - start) / cv::getTickFrequency()) << std::endl;cv::rectangle(src, cv::Rect(maxLoc.x, maxLoc.y, templ.cols, templ.rows), 1, 8, 0);cv::namedWindow("matchTemplateCuda", cv::WINDOW_NORMAL);cv::imshow("matchTemplateCuda", src);
}int main(int argc, char *argv[])
{checkBuild();checkIPP();checkSimd();checkCuda();checkOpenCL();calcEdgesCPU();calcEdgesOpenCL();calcEdgesCuda();matchTemplateCPU();matchTemplateCuda();cv::waitKey(0);return 0;
}cv::getBuildInformation()得到的结果是：
General configuration for OpenCV 4.5.5 =====================================Version control: 4.5.5Platform:Timestamp: 2021-12-25T04:33:15ZHost: Windows 10.0.19041 AMD64CMake: 3.16.4CMake generator: Visual Studio 15 2017CMake build tool: C:/Program Files (x86)/Microsoft Visual Studio/2017/Professional/MSBuild/15.0/Bin/MSBuild.exeMSVC: 1916Configuration: Debug ReleaseCPU/HW features:Baseline: SSE SSE2 SSE3requested: SSE3Dispatched code generation: SSE4_1 SSE4_2 FP16 AVX AVX2 AVX512_SKXrequested: SSE4_1 SSE4_2 AVX FP16 AVX2 AVX512_SKXSSE4_1 (16 files): + SSSE3 SSE4_1SSE4_2 (1 files): + SSSE3 SSE4_1 POPCNT SSE4_2FP16 (0 files): + SSSE3 SSE4_1 POPCNT SSE4_2 FP16 AVXAVX (4 files): + SSSE3 SSE4_1 POPCNT SSE4_2 AVXAVX2 (31 files): + SSSE3 SSE4_1 POPCNT SSE4_2 FP16 FMA3 AVX AVX2AVX512_SKX (5 files): + SSSE3 SSE4_1 POPCNT SSE4_2 FP16 FMA3 AVX AVX2 AVX_512F AVX512_COMMON AVX512_SKXC/C++:Built as dynamic libs?: YESC++ standard: 11C++ Compiler: C:/Program Files (x86)/Microsoft Visual Studio/2017/Professional/VC/Tools/MSVC/14.16.27023/bin/Hostx86/x64/cl.exe (ver 19.16.27042.0)C++ flags (Release): /DWIN32 /D_WINDOWS /W4 /GR /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi /fp:precise /EHa /wd4127 /wd4251 /wd4324 /wd4275 /wd4512 /wd4589 /MP2 /MD /O2 /Ob2 /DNDEBUGC++ flags (Debug): /DWIN32 /D_WINDOWS /W4 /GR /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi /fp:precise /EHa /wd4127 /wd4251 /wd4324 /wd4275 /wd4512 /wd4589 /MP2 /MDd /Zi /Ob0 /Od /RTC1C Compiler: C:/Program Files (x86)/Microsoft Visual Studio/2017/Professional/VC/Tools/MSVC/14.16.27023/bin/Hostx86/x64/cl.exeC flags (Release): /DWIN32 /D_WINDOWS /W3 /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi /fp:precise /MP2 /MD /O2 /Ob2 /DNDEBUGC flags (Debug): /DWIN32 /D_WINDOWS /W3 /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi /fp:precise /MP2 /MDd /Zi /Ob0 /Od /RTC1Linker flags (Release): /machine:x64 /INCREMENTAL:NOLinker flags (Debug): /machine:x64 /debug /INCREMENTALccache: NOPrecompiled headers: NOExtra dependencies:3rdparty dependencies:OpenCV modules:To be built: calib3d core dnn features2d flann gapi highgui imgcodecs imgproc ml objdetect photo stitching video videoio worldDisabled: python2 python3Disabled by dependency: -Unavailable: java tsApplications: appsDocumentation: NONon-free algorithms: NOWindows RT support: NOGUI:Win32 UI: YESVTK support: NOMedia I/O:ZLib: build (ver 1.2.11)JPEG: build-libjpeg-turbo (ver 2.1.2-62)WEBP: build (ver encoder: 0x020f)PNG: build (ver 1.6.37)TIFF: build (ver 42 - 4.2.0)JPEG 2000: build (ver 2.4.0)OpenEXR: build (ver 2.3.0)HDR: YESSUNRASTER: YESPXM: YESPFM: YESVideo I/O:DC1394: NOFFMPEG: YES (prebuilt binaries)avcodec: YES (58.134.100)avformat: YES (58.76.100)avutil: YES (56.70.100)swscale: YES (5.9.100)avresample: YES (4.0.0)GStreamer: NODirectShow: YESMedia Foundation: YESDXVA: YESParallel framework: ConcurrencyTrace: YES (with Intel ITT)Other third-party libraries:Intel IPP: 2020.0.0 Gold [2020.0.0]at: C:/build/master_winpack-build-win64-vc15/build/3rdparty/ippicv/ippicv_win/icvIntel IPP IW: sources (2020.0.0)at: C:/build/master_winpack-build-win64-vc15/build/3rdparty/ippicv/ippicv_win/iwEigen: NOCustom HAL: NOProtobuf: build (3.19.1)OpenCL: YES (NVD3D11)Include path: C:/build/master_winpack-build-win64-vc15/opencv/3rdparty/include/opencl/1.2Link libraries: Dynamic loadPython (for build): C:/utils/soft/python27-x64/python.exeJava:ant: C:/utils/soft/apache-ant-1.9.7/bin/ant.bat (ver 1.9.7)JNI: C:/Program Files/Java/jdk1.8.0_112/include C:/Program Files/Java/jdk1.8.0_112/include/win32 C:/Program Files/Java/jdk1.8.0_112/includeJava wrappers: NOJava tests: NOInstall to: C:/build/master_winpack-build-win64-vc15/install
-----------------------------------------------------------------x2、姊妹篇
一文彻底搞懂为什么OpenCV用GPU/cuda跑得比用CPU慢？_利白的博客-CSDN博客
OpenCV算法加速（4）官方源码v4.5.5的默认并行和优化加速的编译选项是什么？请重点关注函数cv::getBuildInformation()的返回值相关推荐
- 论文阅读: （ICDAR2021 海康威视）LGPMA（表格识别算法）及官方源码对应解读
目录 引言 2022-06-08 update LGPMA整体结构 训练阶段 Aligned Boudning Box Detection(对齐的包围框检测) Local Pyramid Mask A ...
- Python基于OpenCV的实时疲劳检测[源码＆演示视频＆部署教程]
1.图片演示 2.视频演示 [项目分享]Python基于OpenCV的实时疲劳检测[源码&演示视频&部署教程]_哔哩哔哩_bilibili 3.检测方法 1)方法 与用于计算眨眼的传统 ...
- 较高人工智能的人机博弈程序实现(多个算法结合)含C++源码
较高人工智能的人机博弈程序实现(多个算法结合)含C++源码 本文由恋花蝶最初发表于http://blog.csdn.net/lanphaday 上,您可以转载.引用.打印和分发等,但必须保留本文完整和 ...
- [HOW TO]-下载android官方源码
介绍下载android官方源码的方式: 使用每月更新的初始化包 传统初始化方法 1.使用每月更新的初始化包 下载初始化包->repo sync wget -c https://mirrors.t ...
- JavaScript实现唯一路径问题的动态编程方法的算法（附完整源码）
JavaScript实现唯一路径问题的动态编程方法的算法(附完整源码) dpUniquePaths.js完整源代码 dpUniquePaths.test.js完整源代码 dpUniquePaths.j ...
- JavaScript实现唯一路径问题的回溯方法的算法（附完整源码）
JavaScript实现唯一路径问题的回溯方法的算法(附完整源码) btUniquePaths.js完整源代码 btUniquePaths.test.js完整源代码 btUniquePaths.js完 ...
- JavaScript实现squareMatrixRotation方阵旋转算法（附完整源码）
JavaScript实现squareMatrixRotation方阵旋转算法(附完整源码) squareMatrixRotation.js完整源代码 squareMatrixRotation.test ...
- JavaScript实现递归楼梯问题（带记忆的递归解决方案）算法（附完整源码）
JavaScript实现递归楼梯问题(带记忆的递归解决方案)算法(附完整源码) recursiveStaircaseMEM.js完整源代码 recursiveStaircaseMEM.test.js完 ...
- JavaScript实现递归楼梯问题（迭代解决方案）算法（附完整源码）
JavaScript实现递归楼梯问题(迭代解决方案)算法(附完整源码) recursiveStaircaseIT.js完整源代码 recursiveStaircaseIT.test.js完整源代码 r ...
最新文章
- IDEA新建springboot项目发生错误
- niceyoo的2020年终总结-2021年Flag
- Flash--元件和实例
- 为双十二灵感设计屯好素材！
- 工作后辞职原因是什么？
- linux下limits.conf 修改不生效的原因
- C语言丨约瑟夫问题（约瑟夫环）
- convertTo函数的用法
- 汇编语言第一课作业1.1
- 计算机网络五层结构要点以及功能,试述五层协议的网络体系结构的要点，包括各层的主要功能...
- virtualhost配置
- laravel 项目笔记之SendCloud 驱动
- 顺序队列，链队列的基本操作
- 杭州出租车改革何以值得肯定
- 洛谷P5804 [SEERC2019]Absolute Game
- “68道 Redis+168道 MySQL”精品面试题（带解析），你背废了吗？
- 批量转换中文名称为英文名称（注:一般为转换格式拼音）
- 或是独体字吗_独体字结构 独体结构的字有哪些字？
- SourceTree提交代码报错git -c diff.mnemonicprefix=false -c core.quotepath=false commit -q -F C:\Users\honry
- 做产品和运营必须深参这5大人性弱点