OpenCV RVV

OpenCV 在 RVV（RISC-V Vector）上的使用与 x86 平台基本一致，主要是库的差异。

什么是OpenCV？

OpenCV（Open Source Computer Vision Library） 是一个开源的计算机视觉和机器学习软件库，由英特尔公司发起并得到社区的广泛支持。它提供了一个跨平台的编程框架，用于实时的计算机视觉应用开发。常用于:

• 图像处理和分析
• 人脸检测和识别
• 物体检测和跟踪
• 机器学习应用
• 视频分析
• 相机标定和 3D 重建

C++使用

本节对比带 RVV 和不带 RVV的 resize 函数的性能表现

安装必要依赖

sudo apt install wget cmake gcc-15 g++-15

安装 opencv-spacemit

opencv-spacemit 包跟踪上游最新的RVV优化，提供在 SpacemiT riscv64 平台上的最佳性能

sudo apt update
sudo apt install opencv-spacemit

安装完成后终端打印：

==============================
How to use this custom OpenCV
==============================

Method 1: Specify OpenCV_DIR in CMakeLists.txt

    set(OpenCV_DIR "/opt/opencv-spacemit/lib/cmake/opencv4")
    find_package(OpenCV REQUIRED)

Method 2: Set CMAKE_PREFIX_PATH when running cmake

    cmake -DCMAKE_PREFIX_PATH=/opt/opencv-spacemit ..

================================

测试图片下载

cd ~
wget https://archive.spacemit.com/spacemit-ai/BRDK/Model_Zoo/Datasets/test/obb_demo.jpg

项目结构

➜  cv_test tree -L 2 .
.
├── CMakeLists.txt
└── src
    └── main.cpp # 测试程序

main.cpp

#include <opencv2/opencv.hpp>

#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <string>

namespace {

void printUsage(const char* program) {
    std::cout << "Usage: " << program << " [image_path] [iterations] [width] [height]\n"
              << "Defaults:\n"
              << "  image_path : obb_demo.jpg\n"
              << "  iterations : 1000\n"
              << "  width      : 640\n"
              << "  height     : 480\n";
}

int parsePositiveInt(const char* value, const char* name) {
    char* end = nullptr;
    const long parsed = std::strtol(value, &end, 10);
    if (*value == '\0' || *end != '\0' || parsed <= 0) {
        std::cerr << "Invalid " << name << ": " << value << '\n';
        std::exit(1);
    }
    return static_cast<int>(parsed);
}

}  // namespace

int main(int argc, char** argv) {
    if (argc > 1 && (std::string(argv[1]) == "-h" || std::string(argv[1]) == "--help")) {
        printUsage(argv[0]);
        return 0;
    }

    const std::string imagePath = argc > 1 ? argv[1] : "obb_demo.jpg";
    const int iterations = argc > 2 ? parsePositiveInt(argv[2], "iterations") : 1000;
    const int targetWidth = argc > 3 ? parsePositiveInt(argv[3], "width") : 640;
    const int targetHeight = argc > 4 ? parsePositiveInt(argv[4], "height") : 480;

    if (argc > 5) {
        printUsage(argv[0]);
        return 1;
    }

    const cv::Mat input = cv::imread(imagePath, cv::IMREAD_COLOR);
    if (input.empty()) {
        std::cerr << "Failed to read image: " << imagePath << '\n';
        return 1;
    }

    cv::Mat output;
    const cv::Size targetSize(targetWidth, targetHeight);

    for (int i = 0; i < 10; ++i) {
        cv::resize(input, output, targetSize, 0.0, 0.0, cv::INTER_LINEAR);
    }

    const int64 start = cv::getTickCount();
    for (int i = 0; i < iterations; ++i) {
        cv::resize(input, output, targetSize, 0.0, 0.0, cv::INTER_LINEAR);
    }
    const int64 end = cv::getTickCount();

    const double totalMs = (end - start) * 1000.0 / cv::getTickFrequency();
    const double avgMs = totalMs / iterations;
    const double fps = 1000.0 / avgMs;

    std::cout << std::fixed << std::setprecision(3)
              << "Input image : " << imagePath << " (" << input.cols << "x" << input.rows << ")\n"
              << "Output size : " << targetWidth << "x" << targetHeight << '\n'
              << "Iterations  : " << iterations << '\n'
              << "Total time  : " << totalMs << " ms\n"
              << "Average     : " << avgMs << " ms/resize\n"
              << "Throughput  : " << fps << " resize/s\n";

    return 0;
}

CMakeLists.txt

cmake_minimum_required(VERSION 3.10)
project(opencv_resize_benchmark LANGUAGES CXX)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

find_package(OpenCV REQUIRED)

add_executable(resize_benchmark src/main.cpp)
target_link_libraries(resize_benchmark PRIVATE ${OpenCV_LIBS})
target_include_directories(resize_benchmark PRIVATE ${OpenCV_INCLUDE_DIRS})

本文顶层目录路径默认为 ~/cv_test

带 RVV 编译测试

cd ~/cv_test
mkdir build && cd build
cmake -DCMAKE_PREFIX_PATH=/opt/opencv-spacemit ..

终端打印如下：

cmake -DCMAKE_PREFIX_PATH=/opt/opencv-spacemit ..
-- The CXX compiler identification is GNU 15.2.0
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found OpenCV: /opt/opencv-spacemit (found version "4.14.0")
-- Configuring done (0.6s)
-- Generating done (0.0s)
-- Build files have been written to: /home/bianbu26/cv_test/build

编译：

make

开启测试：

./resize_benchmark ~/obb_demo.jpg

终端打印如下：

不带 RVV 编译测试

安装系统 OpenCV库

sudo apt install libopencv-dev

测试完成后，如需恢复为优先使用 opencv-spacemit 的环境，可卸载系统 OpenCV：

sudo apt remove libopencv-dev
sudo apt autoremove

配置

cd ~/cv_test
rm -rf build
mkdir build && cd build
cmake ..

终端打印

cmake ..
-- The CXX compiler identification is GNU 15.2.0
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found OpenCV: /usr (found version "4.10.0")
-- Configuring done (0.6s)
-- Generating done (0.0s)
-- Build files have been written to: /home/bianbu26/cv_test/build

编译

make

开启测试

./resize_benchmark ~/obb_demo.jpg

终端打印如下

可以看出，RVV 加速的 OpenCV 库性能相比于系统自带的 OpenCV 库有明显提升（1.293ms -> 0.478ms）

Python使用

安装必要依赖

sudo apt install python3-venv python3-pip

设置 SpacemiT 源

pip config set global.index-url https://mirrors.aliyun.com/pypi/simple/
pip config set global.extra-index-url https://git.spacemit.com/api/v4/projects/33/packages/pypi/simple

创建虚拟环境

python3 -m venv ~/cv_venv

安装 opencv-python

source ~/cv_venv/bin/activate
pip install opencv-python

终端打印

➜  ~ source ~/cv_venv/bin/activate
pip install opencv-python
Looking in indexes: https://mirrors.aliyun.com/pypi/simple/, https://git.spacemit.com/api/v4/projects/33/packages/pypi/simple
Collecting opencv-python
  Downloading https://git.spacemit.com/api/v4/projects/33/packages/pypi/files/21718d6ed9cc3558fa3cc3dde26496ff8739e58d5b7b1d4e7b4f7a34e8ca947c/opencv_python-4.13.0.92-cp39-abi3-linux_riscv64.whl (91.2 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 91.2/91.2 MB 588.5 kB/s eta 0:00:00
Collecting numpy>=2 (from opencv-python)
  Downloading https://git.spacemit.com/api/v4/projects/33/packages/pypi/files/6121ae6d6bbaf9b1f26d98af46100529679020759bbd1c1cb8d0f29fdf2ec941/numpy-2.4.6-cp314-cp314-manylinux_2_38_riscv64.manylinux_2_41_riscv64.whl (11.1 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 11.1/11.1 MB 657.3 kB/s eta 0:00:00
Installing collected packages: numpy, opencv-python
Successfully installed numpy-2.4.6 opencv-python-4.13.0.92

测试

测试代码：

main.py

import argparse
import sys
import time

import cv2

def positive_int(value: str) -> int:
    try:
        parsed = int(value)
    except ValueError:
        raise argparse.ArgumentTypeError(f"invalid positive int value: {value}")

    if parsed <= 0:
        raise argparse.ArgumentTypeError(f"invalid positive int value: {value}")
    return parsed

def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="Benchmark OpenCV resize performance.",
        formatter_class=argparse.RawTextHelpFormatter,
    )
    parser.add_argument("image_path", nargs="?", default="obb_demo.jpg")
    parser.add_argument("iterations", nargs="?", type=positive_int, default=1000)
    parser.add_argument("width", nargs="?", type=positive_int, default=640)
    parser.add_argument("height", nargs="?", type=positive_int, default=480)
    return parser.parse_args()

def main() -> int:
    args = parse_args()

    input_image = cv2.imread(args.image_path, cv2.IMREAD_COLOR)
    if input_image is None:
        print(f"Failed to read image: {args.image_path}", file=sys.stderr)
        return 1

    target_size = (args.width, args.height)
    output_image = None

    for _ in range(10):
        output_image = cv2.resize(input_image, target_size, interpolation=cv2.INTER_LINEAR)

    start = time.perf_counter()
    for _ in range(args.iterations):
        output_image = cv2.resize(input_image, target_size, interpolation=cv2.INTER_LINEAR)
    end = time.perf_counter()

    if output_image is None:
        return 1

    total_ms = (end - start) * 1000.0
    avg_ms = total_ms / args.iterations
    fps = 1000.0 / avg_ms

    height, width = input_image.shape[:2]
    print(f"Input image : {args.image_path} ({width}x{height})")
    print(f"Output size : {args.width}x{args.height}")
    print(f"Iterations  : {args.iterations}")
    print(f"Total time  : {total_ms:.3f} ms")
    print(f"Average     : {avg_ms:.3f} ms/resize")
    print(f"Throughput  : {fps:.3f} resize/s")

    return 0

if __name__ == "__main__":
    raise SystemExit(main())

source ~/cv_venv/bin/activate
python main.py ~/obb_demo.jpg

结果如下：

更多OpenCV性能测试数据

单核对比

• 使用 OPENCV_FOR_THREADS_NUM=1 OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 taskset -c 7 限制线程和核数
• 计时策略是跑 100 次取平均值，预热10次
• no_rvv_avg_ms 表示的是系统 libopencv-dev 包的表现
• rvv_avg_ms 表示的是 opencv-spacemit 包的表现

module	function	api	input_size	output_size	no_rvv_avg_ms	rvv_avg_ms	speedup
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	224x224	0.8742	0.4927	1.7743x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	320x320	1.7556	1.0153	1.7291x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	640x640	5.5175	3.7299	1.4793x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	512x512	3.8247	2.4210	1.5798x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	1024x1024	11.8105	9.4665	1.2476x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	224x224	0.9117	0.5334	1.7092x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	320x320	1.8685	1.1515	1.6227x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	640x640	6.5235	4.0568	1.6080x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	512x512	4.5591	2.8163	1.6188x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	1024x1024	14.2114	9.5484	1.4884x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	224x224	0.9762	0.6885	1.4179x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	320x320	1.9186	1.3032	1.4722x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	640x640	7.1410	4.5058	1.5848x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	512x512	4.6476	2.9959	1.5513x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	1024x1024	15.7861	9.8249	1.6067x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	224x224	1.1642	1.0324	1.1277x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	320x320	2.0680	1.6827	1.2290x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	640x640	7.2973	4.6838	1.5580x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	512x512	4.7658	3.2211	1.4796x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	1024x1024	18.5301	10.9281	1.6956x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	224x224	0.8616	0.4740	1.8177x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	320x320	1.5443	0.9449	1.6344x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	640x640	4.6122	3.6942	1.2485x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	512x512	3.3079	2.3730	1.3940x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	1024x1024	10.0030	9.4498	1.0585x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	224x224	224x224	0.0956	0.0165	5.7939x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	320x320	320x320	0.1944	0.0280	6.9429x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	640x640	640x640	0.7615	0.1061	7.1772x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	512x512	512x512	0.4897	0.0683	7.1698x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	1024x1024	1024x1024	1.9531	0.3598	5.4283x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	224x224	224x224	0.1119	0.0487	2.2977x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	320x320	320x320	0.2275	0.0911	2.4973x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	640x640	640x640	0.8893	0.3537	2.5143x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	512x512	512x512	0.5715	0.2277	2.5099x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	1024x1024	1024x1024	2.2687	0.9184	2.4703x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	224x224	224x224	0.1562	0.0459	3.4031x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	320x320	320x320	0.3212	0.0876	3.6667x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	640x640	640x640	1.2323	0.3478	3.5431x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	512x512	512x512	0.7879	0.2232	3.5300x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	1024x1024	1024x1024	3.1761	0.9304	3.4137x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	224x224	1x3x224x224	0.8813	0.1660	5.3090x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	320x320	1x3x320x320	1.7852	0.3357	5.3178x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	640x640	1x3x640x640	7.3417	1.3883	5.2883x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	512x512	1x3x512x512	4.6450	0.8978	5.1738x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	1024x1024	1x3x1024x1024	35.3432	3.5119	10.0638x
core	mean_std_normalize	cv::subtract + cv::divide	224x224	224x224	1.8795	1.8057	1.0409x
core	mean_std_normalize	cv::subtract + cv::divide	320x320	320x320	3.8272	3.6704	1.0427x
core	mean_std_normalize	cv::subtract + cv::divide	640x640	640x640	15.3598	14.7864	1.0388x
core	mean_std_normalize	cv::subtract + cv::divide	512x512	512x512	9.8466	9.4438	1.0427x
core	mean_std_normalize	cv::subtract + cv::divide	1024x1024	1024x1024	39.2297	37.8176	1.0373x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	224x224	224x224	0.6891	0.1476	4.6687x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	320x320	320x320	1.3929	0.2931	4.7523x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	640x640	640x640	5.5161	1.0480	5.2635x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	512x512	512x512	3.5383	0.7893	4.4828x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	1024x1024	1024x1024	14.1017	4.1439	3.4030x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	224x224	288x288	0.0302	0.0219	1.3790x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	320x320	384x384	0.0498	0.0373	1.3351x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	640x640	704x704	0.1448	0.1187	1.2199x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	512x512	576x576	0.1006	0.0802	1.2544x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	1024x1024	1088x1088	0.4003	0.3729	1.0735x
core	split_channels	cv::split	224x224	224x224	0.1056	0.0154	6.8571x
core	split_channels	cv::split	320x320	320x320	0.2096	0.0265	7.9094x
core	split_channels	cv::split	640x640	640x640	0.7822	0.0989	7.9090x
core	split_channels	cv::split	512x512	512x512	0.5051	0.0639	7.9045x
core	split_channels	cv::split	1024x1024	1024x1024	1.9977	0.3546	5.6337x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	224x224	224x224	0.2112	0.1309	1.6134x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	320x320	320x320	0.4248	0.2568	1.6542x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	640x640	640x640	1.6296	0.9885	1.6486x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	512x512	512x512	1.0522	0.6298	1.6707x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	1024x1024	1024x1024	4.1545	2.6075	1.5933x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	224x224	224x224	2.2849	0.1739	13.1392x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	320x320	320x320	4.6580	0.3523	13.2217x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	640x640	640x640	18.8223	1.4077	13.3710x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	512x512	512x512	12.0154	0.9013	13.3312x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	1024x1024	1024x1024	48.0701	3.6437	13.1927x
core	cartToPolar	cv::cartToPolar	224x224	224x224	0.7694	0.2275	3.3820x
core	cartToPolar	cv::cartToPolar	320x320	320x320	1.5677	0.4609	3.4014x
core	cartToPolar	cv::cartToPolar	640x640	640x640	6.2492	1.8439	3.3891x
core	cartToPolar	cv::cartToPolar	512x512	512x512	3.9960	1.1781	3.3919x
core	cartToPolar	cv::cartToPolar	1024x1024	1024x1024	15.8954	4.7343	3.3575x
core	compare	cv::compare(CMP_GT)	224x224	224x224	0.0517	0.0122	4.2377x
core	compare	cv::compare(CMP_GT)	320x320	320x320	0.0968	0.0202	4.7921x
core	compare	cv::compare(CMP_GT)	640x640	640x640	0.3740	0.0642	5.8255x
core	compare	cv::compare(CMP_GT)	512x512	512x512	0.2408	0.0429	5.6131x
core	compare	cv::compare(CMP_GT)	1024x1024	1024x1024	0.9486	0.1509	6.2863x
core	divide	cv::divide(Scalar)	224x224	224x224	1.5925	1.6317	0.9760x
core	divide	cv::divide(Scalar)	320x320	320x320	3.2487	3.3241	0.9773x
core	divide	cv::divide(Scalar)	640x640	640x640	13.0132	13.3281	0.9764x
core	divide	cv::divide(Scalar)	512x512	512x512	8.3188	8.5162	0.9768x
core	divide	cv::divide(Scalar)	1024x1024	1024x1024	33.2666	34.0628	0.9766x
core	dot	Mat::dot	224x224	1x1	0.2014	0.0347	5.8040x
core	dot	Mat::dot	320x320	1x1	0.4110	0.0711	5.7806x
core	dot	Mat::dot	640x640	1x1	1.6571	0.5697	2.9087x
core	dot	Mat::dot	512x512	1x1	1.0530	0.2462	4.2770x
core	dot	Mat::dot	1024x1024	1x1	4.2366	1.5042	2.8165x
core	dft	cv::dft	224x224	224x224	1.0208	2.3128	0.4414x
core	dft	cv::dft	320x320	320x320	1.7618	2.0457	0.8612x
core	dft	cv::dft	640x640	640x640	7.8431	8.1615	0.9610x
core	dft	cv::dft	512x512	512x512	4.7081	5.0837	0.9261x
core	dft	cv::dft	1024x1024	1024x1024	35.0114	40.7955	0.8582x
core	exp	cv::exp	224x224	224x224	3.5022	0.2809	12.4678x
core	exp	cv::exp	320x320	320x320	7.1433	0.5707	12.5167x
core	exp	cv::exp	640x640	640x640	28.5879	2.3018	12.4198x
core	exp	cv::exp	512x512	512x512	18.2907	1.4683	12.4571x
core	exp	cv::exp	1024x1024	1024x1024	73.1500	5.8740	12.4532x
core	flip_horizontal	cv::flip(1)	224x224	224x224	0.1385	0.0225	6.1556x
core	flip_horizontal	cv::flip(1)	320x320	320x320	0.2779	0.0345	8.0551x
core	flip_horizontal	cv::flip(1)	640x640	640x640	1.0931	0.1331	8.2126x
core	flip_horizontal	cv::flip(1)	512x512	512x512	0.7038	0.0857	8.2124x
core	flip_horizontal	cv::flip(1)	1024x1024	1024x1024	2.8300	0.4153	6.8144x
core	LUT_invert	cv::LUT	224x224	224x224	0.2077	0.0361	5.7535x
core	LUT_invert	cv::LUT	320x320	320x320	0.4216	0.0731	5.7674x
core	LUT_invert	cv::LUT	640x640	640x640	1.6924	0.2834	5.9718x
core	LUT_invert	cv::LUT	512x512	512x512	1.0834	0.1820	5.9527x
core	LUT_invert	cv::LUT	1024x1024	1024x1024	4.3393	0.7384	5.8766x
core	minMaxLoc	cv::minMaxLoc	224x224	1x1	0.0921	0.0041	22.4634x
core	minMaxLoc	cv::minMaxLoc	320x320	1x1	0.1879	0.0081	23.1975x
core	minMaxLoc	cv::minMaxLoc	640x640	1x1	0.7472	0.0300	24.9067x
core	minMaxLoc	cv::minMaxLoc	512x512	1x1	0.4793	0.0195	24.5795x
core	minMaxLoc	cv::minMaxLoc	1024x1024	1x1	1.9131	0.0754	25.3727x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	224x224	224x224	2.1725	0.6094	3.5650x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	320x320	320x320	4.4258	1.2401	3.5689x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	640x640	640x640	17.7909	6.5314	2.7239x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	512x512	512x512	11.3623	3.3567	3.3850x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	1024x1024	1024x1024	45.5470	16.7297	2.7225x
core	transpose	cv::transpose	224x224	224x224	0.0865	0.0647	1.3369x
core	transpose	cv::transpose	320x320	320x320	0.1771	0.1349	1.3128x
core	transpose	cv::transpose	640x640	640x640	1.0396	0.8684	1.1971x
core	transpose	cv::transpose	512x512	512x512	0.7604	0.6568	1.1577x
core	transpose	cv::transpose	1024x1024	1024x1024	5.0411	4.7557	1.0600x
imgproc	Canny	cv::Canny	224x224	224x224	1.0812	0.6529	1.6560x
imgproc	Canny	cv::Canny	320x320	320x320	1.9907	1.1120	1.7902x
imgproc	Canny	cv::Canny	640x640	640x640	6.6657	3.0625	2.1766x
imgproc	Canny	cv::Canny	512x512	512x512	4.4574	2.1652	2.0587x
imgproc	Canny	cv::Canny	1024x1024	1024x1024	15.6314	6.4734	2.4147x
imgproc	erode_dilate	cv::erode + cv::dilate	224x224	224x224	0.5417	0.1744	3.1061x
imgproc	erode_dilate	cv::erode + cv::dilate	320x320	320x320	1.0222	0.2862	3.5716x
imgproc	erode_dilate	cv::erode + cv::dilate	640x640	640x640	3.6923	0.7448	4.9574x
imgproc	erode_dilate	cv::erode + cv::dilate	512x512	512x512	2.4159	0.5288	4.5686x
imgproc	erode_dilate	cv::erode + cv::dilate	1024x1024	1024x1024	8.9698	2.2336	4.0158x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	224x224	224x224	0.6036	0.4676	1.2908x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	320x320	320x320	1.2193	0.8677	1.4052x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	640x640	640x640	4.7939	3.4166	1.4031x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	512x512	512x512	3.0798	2.1936	1.4040x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	1024x1024	1024x1024	12.2247	8.7092	1.4037x

八核对比

• 不限制线程数和核数
• 计时策略是跑 100 次取平均值，预热10次
• no_rvv_avg_ms 表示的是系统 libopencv-dev 包的表现
• rvv_avg_ms 表示的是 opencv-spacemit 包的表现

module	function	api	input_size	output_size	no_rvv_avg_ms	rvv_avg_ms	speedup
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	224x224	0.8725	0.1081	8.0712x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	320x320	0.9257	0.1826	5.0696x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	640x640	1.0397	0.5110	2.0346x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	512x512	1.0322	0.4255	2.4259x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1280x720	1024x1024	1.7110	1.2817	1.3349x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	224x224	0.9079	0.1230	7.3813x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	320x320	0.9523	0.2013	4.7308x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	640x640	1.2041	0.5653	2.1300x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	512x512	1.2184	0.4141	2.9423x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	1920x1080	1024x1024	2.2713	1.3177	1.7237x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	224x224	0.9704	0.1516	6.4011x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	320x320	0.9757	0.2183	4.4695x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	640x640	1.3831	1.2363	1.1187x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	512x512	1.2882	0.7792	1.6532x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	2560x1440	1024x1024	2.3685	1.4875	1.5923x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	224x224	1.1752	0.2099	5.5989x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	320x320	1.1078	0.3085	3.5909x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	640x640	1.4702	1.0798	1.3615x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	512x512	1.3615	0.7537	1.8064x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	3840x2160	1024x1024	2.8575	1.8892	1.5125x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	224x224	0.8618	0.0911	9.4599x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	320x320	0.8084	0.1479	5.4659x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	640x640	0.8757	0.5049	1.7344x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	512x512	0.9078	0.3330	2.7261x
imgproc	resize_linear	cv::resize(INTER_LINEAR)	640x480	1024x1024	1.4604	1.3587	1.0749x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	224x224	224x224	0.0957	0.0170	5.6294x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	320x320	320x320	0.1065	0.0285	3.7368x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	640x640	640x640	0.1611	0.1086	1.4834x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	512x512	512x512	0.1372	0.0685	2.0029x
imgproc	cvtColor_bgr2rgb	cv::cvtColor(COLOR_BGR2RGB)	1024x1024	1024x1024	0.3892	0.4042	0.9629x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	224x224	224x224	0.1124	0.0310	3.6258x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	320x320	320x320	0.1210	0.0395	3.0633x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	640x640	640x640	0.1712	0.0756	2.2646x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	512x512	512x512	0.1553	0.0466	3.3326x
imgproc	cvtColor_bgr2gray	cv::cvtColor(COLOR_BGR2GRAY)	1024x1024	1024x1024	0.4499	0.1528	2.9444x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	224x224	224x224	0.1608	0.0465	3.4581x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	320x320	320x320	0.3130	0.0884	3.5407x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	640x640	640x640	1.2758	0.3500	3.6451x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	512x512	512x512	0.8213	0.2256	3.6405x
core	normalize_uint8_to_fp32	Mat::convertTo(CV_32FC3)	1024x1024	1024x1024	3.0313	0.8945	3.3888x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	224x224	1x3x224x224	0.8829	0.1663	5.3091x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	320x320	1x3x320x320	1.7879	0.3361	5.3195x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	640x640	1x3x640x640	7.4718	1.3843	5.3975x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	512x512	1x3x512x512	4.6766	0.9020	5.1847x
dnn	blobFromImage_bgr_fp32	cv::dnn::blobFromImage	1024x1024	1x3x1024x1024	35.5356	3.5207	10.0933x
core	mean_std_normalize	cv::subtract + cv::divide	224x224	224x224	1.8786	1.8028	1.0420x
core	mean_std_normalize	cv::subtract + cv::divide	320x320	320x320	3.8291	3.6625	1.0455x
core	mean_std_normalize	cv::subtract + cv::divide	640x640	640x640	15.6279	14.7644	1.0585x
core	mean_std_normalize	cv::subtract + cv::divide	512x512	512x512	10.2764	9.4336	1.0893x
core	mean_std_normalize	cv::subtract + cv::divide	1024x1024	1024x1024	39.5829	37.7746	1.0479x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	224x224	224x224	0.1167	0.0357	3.2689x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	320x320	320x320	0.2155	0.0710	3.0352x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	640x640	640x640	0.7413	0.2251	3.2932x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	512x512	512x512	0.9123	0.1678	5.4368x
imgproc	gaussian_blur_3x3	cv::GaussianBlur(Size(3	1024x1024	1024x1024	1.8435	1.0925	1.6874x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	224x224	288x288	0.0303	0.0223	1.3587x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	320x320	384x384	0.0496	0.0375	1.3227x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	640x640	704x704	0.1484	0.1188	1.2492x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	512x512	576x576	0.1002	0.0806	1.2432x
core	copy_make_border_32	cv::copyMakeBorder(BORDER_CONSTANT)	1024x1024	1088x1088	0.4016	0.3486	1.1520x
core	split_channels	cv::split	224x224	224x224	0.0981	0.0154	6.3701x
core	split_channels	cv::split	320x320	320x320	0.1966	0.0274	7.1752x
core	split_channels	cv::split	640x640	640x640	0.7809	0.0977	7.9928x
core	split_channels	cv::split	512x512	512x512	0.5008	0.0628	7.9745x
core	split_channels	cv::split	1024x1024	1024x1024	1.9996	0.3540	5.6486x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	224x224	224x224	0.2115	0.0697	3.0344x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	320x320	320x320	0.2892	0.0905	3.1956x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	640x640	640x640	0.6844	0.2053	3.3337x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	512x512	512x512	0.5092	0.1409	3.6139x
imgproc	threshold_otsu_gray	cv::threshold(THRESH_OTSU)	1024x1024	1024x1024	1.5612	0.5983	2.6094x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	224x224	224x224	2.3038	0.1759	13.0972x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	320x320	320x320	4.6996	0.3544	13.2607x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	640x640	640x640	18.7963	1.4091	13.3392x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	512x512	512x512	12.0276	0.9007	13.3536x
core	addWeighted_segmentation_mask	cv::addWeighted(mask overlay)	1024x1024	1024x1024	48.1972	3.6353	13.2581x
core	cartToPolar	cv::cartToPolar	224x224	224x224	0.7703	0.2276	3.3844x
core	cartToPolar	cv::cartToPolar	320x320	320x320	1.5760	0.4614	3.4157x
core	cartToPolar	cv::cartToPolar	640x640	640x640	6.3527	1.8492	3.4354x
core	cartToPolar	cv::cartToPolar	512x512	512x512	4.1158	1.1776	3.4951x
core	cartToPolar	cv::cartToPolar	1024x1024	1024x1024	16.0252	4.7353	3.3842x
core	compare	cv::compare(CMP_GT)	224x224	224x224	0.0514	0.0123	4.1789x
core	compare	cv::compare(CMP_GT)	320x320	320x320	0.0966	0.0201	4.8060x
core	compare	cv::compare(CMP_GT)	640x640	640x640	0.3733	0.0621	6.0113x
core	compare	cv::compare(CMP_GT)	512x512	512x512	0.2401	0.0415	5.7855x
core	compare	cv::compare(CMP_GT)	1024x1024	1024x1024	0.9441	0.1491	6.3320x
core	divide	cv::divide(Scalar)	224x224	224x224	1.5929	1.6311	0.9766x
core	divide	cv::divide(Scalar)	320x320	320x320	3.2471	3.3224	0.9773x
core	divide	cv::divide(Scalar)	640x640	640x640	13.0019	13.3133	0.9766x
core	divide	cv::divide(Scalar)	512x512	512x512	8.3138	8.5093	0.9770x
core	divide	cv::divide(Scalar)	1024x1024	1024x1024	33.3617	34.0391	0.9801x
core	dot	Mat::dot	224x224	1x1	0.2032	0.0356	5.7079x
core	dot	Mat::dot	320x320	1x1	0.4119	0.0710	5.8014x
core	dot	Mat::dot	640x640	1x1	1.6660	0.6130	2.7178x
core	dot	Mat::dot	512x512	1x1	1.0615	0.2744	3.8684x
core	dot	Mat::dot	1024x1024	1x1	4.2606	1.6210	2.6284x
core	dft	cv::dft	224x224	224x224	1.0249	2.3067	0.4443x
core	dft	cv::dft	320x320	320x320	1.7625	2.0284	0.8689x
core	dft	cv::dft	640x640	640x640	7.8298	8.0986	0.9668x
core	dft	cv::dft	512x512	512x512	4.7497	5.0603	0.9386x
core	dft	cv::dft	1024x1024	1024x1024	32.9397	42.4881	0.7753x
core	exp	cv::exp	224x224	224x224	3.5000	0.2806	12.4733x
core	exp	cv::exp	320x320	320x320	7.1405	0.5709	12.5074x
core	exp	cv::exp	640x640	640x640	28.5876	2.3021	12.4181x
core	exp	cv::exp	512x512	512x512	18.2888	1.4667	12.4694x
core	exp	cv::exp	1024x1024	1024x1024	73.1620	5.8793	12.4440x
core	flip_horizontal	cv::flip(1)	224x224	224x224	0.1373	0.0221	6.2127x
core	flip_horizontal	cv::flip(1)	320x320	320x320	0.2778	0.0344	8.0756x
core	flip_horizontal	cv::flip(1)	640x640	640x640	1.0956	0.1342	8.1639x
core	flip_horizontal	cv::flip(1)	512x512	512x512	0.7046	0.0859	8.2026x
core	flip_horizontal	cv::flip(1)	1024x1024	1024x1024	2.8236	0.4305	6.5589x
core	LUT_invert	cv::LUT	224x224	224x224	0.2081	0.0361	5.7645x
core	LUT_invert	cv::LUT	320x320	320x320	0.4214	0.0372	11.3280x
core	LUT_invert	cv::LUT	640x640	640x640	0.3100	0.0816	3.7990x
core	LUT_invert	cv::LUT	512x512	512x512	0.2877	0.0595	4.8353x
core	LUT_invert	cv::LUT	1024x1024	1024x1024	0.6399	0.1755	3.6462x
core	minMaxLoc	cv::minMaxLoc	224x224	1x1	0.0923	0.0042	21.9762x
core	minMaxLoc	cv::minMaxLoc	320x320	1x1	0.1888	0.0084	22.4762x
core	minMaxLoc	cv::minMaxLoc	640x640	1x1	0.7480	0.0303	24.6865x
core	minMaxLoc	cv::minMaxLoc	512x512	1x1	0.4793	0.0197	24.3299x
core	minMaxLoc	cv::minMaxLoc	1024x1024	1x1	1.9119	0.0757	25.2563x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	224x224	224x224	2.1189	0.3691	5.7407x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	320x320	320x320	2.1759	0.5212	4.1748x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	640x640	640x640	3.0862	1.0133	3.0457x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	512x512	512x512	2.8669	0.6035	4.7505x
imgproc	remap_identity	cv::remap(INTER_LINEAR)	1024x1024	1024x1024	6.1271	2.3530	2.6040x
core	transpose	cv::transpose	224x224	224x224	0.0867	0.0652	1.3298x
core	transpose	cv::transpose	320x320	320x320	0.1763	0.1345	1.3108x
core	transpose	cv::transpose	640x640	640x640	1.0095	0.8645	1.1677x
core	transpose	cv::transpose	512x512	512x512	0.7859	0.6602	1.1904x
core	transpose	cv::transpose	1024x1024	1024x1024	5.0581	4.6902	1.0784x
imgproc	Canny	cv::Canny	224x224	224x224	0.3560	0.2436	1.4614x
imgproc	Canny	cv::Canny	320x320	320x320	0.5887	0.4115	1.4306x
imgproc	Canny	cv::Canny	640x640	640x640	1.4790	1.1927	1.2400x
imgproc	Canny	cv::Canny	512x512	512x512	1.0579	0.8249	1.2825x
imgproc	Canny	cv::Canny	1024x1024	1024x1024	3.2367	2.4633	1.3140x
imgproc	erode_dilate	cv::erode + cv::dilate	224x224	224x224	0.5410	0.0708	7.6412x
imgproc	erode_dilate	cv::erode + cv::dilate	320x320	320x320	1.0229	0.1043	9.8073x
imgproc	erode_dilate	cv::erode + cv::dilate	640x640	640x640	3.6897	0.2643	13.9603x
imgproc	erode_dilate	cv::erode + cv::dilate	512x512	512x512	2.4183	0.2005	12.0613x
imgproc	erode_dilate	cv::erode + cv::dilate	1024x1024	1024x1024	8.9730	0.6168	14.5477x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	224x224	224x224	0.4020	0.2386	1.6848x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	320x320	320x320	0.7846	0.4526	1.7335x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	640x640	640x640	2.9932	1.7714	1.6897x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	512x512	512x512	2.0003	1.1914	1.6789x
imgproc	pyrDown_pyrUp	cv::pyrDown + cv::pyrUp	1024x1024	1024x1024	7.7747	4.4134	1.7616x