yuv420转yuv422使用neno加速的示例代码

2024-05-29 18:36
文章标签 代码 使用 加速 示例 yuv422 yuv420 neno

本文主要是介绍yuv420转yuv422使用neno加速的示例代码，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

一、使用none实现yuv420快速转换为yuv422

#include <arm_neon.h>

static vs_int32_t yuv420_yuv422_conver(vs_uint8_t *p_yuv420_y, vs_uint8_t *p_yuv420_uv, vs_uint32_t yuv420_y_stride, vs_uint32_t yuv420_uv_stride, vs_int32_t width, vs_int32_t height, vs_uint8_t* p_yuv422_virt_addr)
{
    vs_int32_t w = 0, h = 0, remain = 0;
    vs_uint8_t* p_yuv422_out = NULL;
    uint8x8_t u8rslt;
    uint8x8x2_t u8yuvdata;
    uint8x8x2_t u8trans_idx;
    vs_int32_t p_yuv422_stride = (width << 1);
    const vs_uint8_t data_idx[2][8] = { { 0, 8, 1, 9, 2, 10, 3, 11 },
                                         { 4, 12, 5, 13, 6, 14, 7, 15 } };

    //width must be even
    if (width % 2 != 0) {
        VS_UVC_LOG_E("width[%u] must be even!\n", width);
        return VS_FAILED;
    }

    p_yuv422_out = p_yuv422_virt_addr;
    u8trans_idx.val[0] = vld1_u8(data_idx[0]);
    u8trans_idx.val[1] = vld1_u8(data_idx[1]);

    h = height;
    for (; h > 0; h--) {

        w = (width >> 3);
        remain = width - (w << 3);
        for (; w > 0; w--) {

            u8yuvdata.val[0] = vld1_u8(p_yuv420_y);
            u8yuvdata.val[1] = vld1_u8(p_yuv420_uv);
            u8rslt = vtbl2_u8(u8yuvdata, u8trans_idx.val[0]);
            vst1_u8(p_yuv422_out, u8rslt);

            u8rslt = vtbl2_u8(u8yuvdata, u8trans_idx.val[1]);

            vst1_u8(p_yuv422_out + 8, u8rslt);
            p_yuv420_y += 8;
            p_yuv420_uv += 8;
            p_yuv422_out += 16;
        }

        for (; remain > 0; remain -= 2) {

            p_yuv422_out[0] = p_yuv420_y[0];
            p_yuv422_out[1] = p_yuv420_uv[1];
            p_yuv422_out[2] = p_yuv420_y[1];
            p_yuv422_out[3] = p_yuv420_uv[0];
            p_yuv420_y += 2;
            p_yuv420_uv += 2;
            p_yuv422_out += 4;
        }

        if (h & 1) {
            p_yuv420_uv += (yuv420_uv_stride - width);
        } else {
            p_yuv420_uv -= width;
        }
        p_yuv420_y += (yuv420_y_stride - width);
        p_yuv422_out += (p_yuv422_stride - (width << 1));
    }

    return VS_SUCCESS;
}

static vs_int32_t frame_yuv420_yuv422_conver(vs_video_frame_s* p_yuv420_frame, vs_uint8_t* p_yuv422_virt_addr, vs_int32_t yuv422_w, vs_int32_t yuv422_h, FILE *p_yuv420_file)
{
    vs_int32_t ret = VS_SUCCESS;
    vs_uint32_t yuv420_frame_size = 0;
    vs_uint8_t *p_yuv420_virt_addr = NULL, *p_yuv420_y = NULL, *p_yuv420_uv = NULL;


    if(p_yuv420_frame->width != yuv422_w || p_yuv420_frame->height != yuv422_h) {
        VS_UVC_LOG_E("p_yuv420_frame->width != yuv422_w\n");
        return VS_FAILED;
    }

    yuv420_frame_size = p_yuv420_frame->stride[0] * p_yuv420_frame->height * 3 / 2;
    VS_UVC_LOG_I("===========p_yuv420_frame->stride[0] = %d =============\n", p_yuv420_frame->stride[0]);
    VS_UVC_LOG_I("===========p_yuv420_frame->stride[1] = %d =============\n", p_yuv420_frame->stride[1]);
    p_yuv420_virt_addr = (vs_uint8_t*)vs_mal_sys_mmap_cached(p_yuv420_frame->phys_addr[0], yuv420_frame_size);
    if (!p_yuv420_virt_addr) {
        VS_UVC_LOG_E("vs_mal_sys_mmap_cached failed\n");
        return VS_FAILED;
    }


    p_yuv420_y = p_yuv420_virt_addr;
    p_yuv420_uv = p_yuv420_y + p_yuv420_frame->stride[0] * p_yuv420_frame->height;

    ret = yuv420_yuv422_conver(p_yuv420_y, p_yuv420_uv, p_yuv420_frame->stride[0], p_yuv420_frame->stride[1], yuv422_w, yuv422_h, p_yuv422_virt_addr);
    if (ret != VS_SUCCESS) {
        VS_UVC_LOG_E("yuv420_2_yuv422_conver failed\n");
        vs_mal_sys_unmap(p_yuv420_virt_addr, yuv420_frame_size);
        return ret;
    }

    fwrite(p_yuv420_virt_addr, 1, yuv420_frame_size, p_yuv420_file);
    fflush(p_yuv420_file);

    vs_mal_sys_unmap(p_yuv420_virt_addr, yuv420_frame_size);

    return VS_SUCCESS;
}

static vs_int32_t frame_yuv420_witre_2_yuv422file(vs_video_frame_s* p_yuv420_frame)
{
    vs_int32_t ret = VS_SUCCESS;
    static vs_int32_t savefile_index = 0;
    vs_uint32_t yuv422_frame_size = 0;
    vs_uint8_t* p_yuv422_virt_addr = VS_NULL;
    FILE *p_yuv420_file = VS_NULL, *p_yuv422_file = VS_NULL;
    vs_char_t yuv420_file_name[64], yuv422_file_name[64];

    //yuv420_frame_size = p_yuv420_frame->stride[0] * p_yuv420_frame->height * 3 / 2;
    yuv422_frame_size = p_yuv420_frame->stride[0] * p_yuv420_frame->height * 2;
    p_yuv422_virt_addr = malloc(yuv422_frame_size);
    if (p_yuv422_virt_addr == VS_NULL) {
        VS_UVC_LOG_E("malloc size[%u] failed!\n", yuv422_frame_size);
        return VS_FAILED;
    }

    memset(yuv420_file_name, 0, sizeof(yuv420_file_name));
    memset(yuv422_file_name, 0, sizeof(yuv422_file_name));
    sprintf((char *)yuv420_file_name, "yuv420_idx%d.yuv", savefile_index);
    sprintf((char *)yuv422_file_name, "yuv422_idx%d.yuv", savefile_index);
    savefile_index++;
    p_yuv420_file = fopen((char *)yuv420_file_name, "wb");
    if (VS_NULL == p_yuv420_file) {
        VS_UVC_LOG_E("fopen file[%s] failed\n", yuv420_file_name);
        free(p_yuv422_virt_addr);
        return VS_FAILED;
    }
    p_yuv422_file = fopen((char *)yuv422_file_name, "wb");
    if (VS_NULL == p_yuv422_file) {
        VS_UVC_LOG_E("fopen file[%s] failed\n", yuv422_file_name);
        free(p_yuv422_virt_addr);
        fclose(p_yuv420_file);
        return VS_FAILED;
    }

    ret = frame_yuv420_yuv422_conver(p_yuv420_frame, p_yuv422_virt_addr, p_yuv420_frame->width, p_yuv420_frame->height, p_yuv420_file);
    if (ret != VS_SUCCESS) {
        VS_UVC_LOG_E("yuv420_2_yuv422_conver failed\n");
        free(p_yuv422_virt_addr);
        fclose(p_yuv420_file);
        fclose(p_yuv422_file);
        return ret;
    }

    fwrite(p_yuv422_virt_addr, 1, yuv422_frame_size, p_yuv422_file);
    fflush(p_yuv422_file);

    free(p_yuv422_virt_addr);
    fclose(p_yuv420_file);
    fclose(p_yuv422_file);

    return VS_SUCCESS;
}

二、使用none实现yuv420快速转换为nv12

static vs_int32_t yuv420_nv12_conver(vs_uint8_t* p_yuv420_virt_addr, vs_uint32_t yuv420_y_stride, vs_uint32_t yuv420_uv_stride, vs_int32_t width, vs_int32_t height, vs_uint8_t* p_nv12_virt_addr)
{
    vs_int32_t w, h;
    vs_uint8_t* p_yuv420_in = p_yuv420_virt_addr, *p_nv12_out = p_nv12_virt_addr;
    vs_int32_t remain = 0;
    uint8x16_t u8data;

    //width must be even
    if (width % 2 != 0) {
        VS_UVC_LOG_E("width[%u] must be even!\n", width);
        return VS_FAILED;
    }

    // for y
    h = height;
    for (; h > 0; h--) {
        w = (width >> 4);
        remain = width % 16;
        for (; w > 0; w--) {
            u8data = vld1q_u8(p_yuv420_in);
            vst1q_u8(p_nv12_out, u8data);
            p_nv12_out += 16;
            p_yuv420_in += 16;
        }
        for (; remain > 0; remain--) {
            *p_nv12_out = *p_yuv420_in;
            p_nv12_out++;
            p_yuv420_in++;
        }
        p_yuv420_in += (yuv420_y_stride - width);
    }

    // for yv
    h = (height >> 1);
    for (; h > 0; h--) {
        w = (width >> 4);
        remain = width % 16;
        for (; w > 0; w--) {
            u8data = vld1q_u8(p_yuv420_in);
            //u8data = vrev16q_u8(u8data);
            vst1q_u8(p_nv12_out, u8data);
            p_nv12_out += 16;
            p_yuv420_in += 16;
        }

        for (; remain > 0; remain -= 2) {
            p_nv12_out[0] = p_yuv420_in[0];
            p_nv12_out[1] = p_yuv420_in[1];
            p_nv12_out += 2;
            p_yuv420_in += 2;
        }
        p_yuv420_in += (yuv420_uv_stride - width);
    }

    return  VS_SUCCESS;
}

static vs_int32_t frame_yuv420_nv12_conver(vs_video_frame_s* p_yuv420_frame, vs_uint8_t* p_nv12_virt_addr,vs_int32_t nv12_w,vs_int32_t nv12_h, FILE *p_yuv420_file)
{
    vs_int32_t ret = VS_SUCCESS;
    vs_uint32_t yuv420_frame_size = 0;
    vs_uint8_t *p_yuv420_virt_addr = NULL;

    if (p_yuv420_frame->width != nv12_w || p_yuv420_frame->height != nv12_h) {
        VS_UVC_LOG_E("p_yuv420_frame->width != nv12_w\n");
        return 0;
    }

    yuv420_frame_size = p_yuv420_frame->stride[0] * p_yuv420_frame->height * 3 / 2;
    VS_UVC_LOG_I("===========p_yuv420_frame->stride[0] = %d =============\n", p_yuv420_frame->stride[0]);
    VS_UVC_LOG_I("===========p_yuv420_frame->stride[1] = %d =============\n", p_yuv420_frame->stride[1]);
    p_yuv420_virt_addr = (vs_uint8_t*)vs_mal_sys_mmap_cached(p_yuv420_frame->phys_addr[0], yuv420_frame_size);
    if (!p_yuv420_virt_addr) {
        VS_UVC_LOG_E(" vs_mal_sys_mmap_cached failed\n");
        return 0;
    }

    ret = yuv420_nv12_conver(p_yuv420_virt_addr, p_yuv420_frame->stride[0], p_yuv420_frame->stride[1], nv12_w, nv12_h, p_nv12_virt_addr);
    if (ret != VS_SUCCESS) {
        VS_UVC_LOG_E("yuv420_2_yuv422_conver failed\n");
        vs_mal_sys_unmap(p_yuv420_virt_addr, yuv420_frame_size);
        return ret;
    }

    fwrite(p_yuv420_virt_addr, 1, yuv420_frame_size, p_yuv420_file);
    fflush(p_yuv420_file);

    vs_mal_sys_unmap(p_yuv420_virt_addr, yuv420_frame_size);

    return VS_SUCCESS;
}

static vs_int32_t frame_yuv420_witre_2_nv12file(vs_video_frame_s* p_yuv420_frame)
{
    vs_int32_t ret = VS_SUCCESS;
    static vs_int32_t savefile_index = 0;
    vs_uint32_t nv12_frame_size = 0;
    vs_uint8_t* p_yuv422_virt_addr = VS_NULL;
    FILE *p_yuv420_file = VS_NULL, *p_nv12_file = VS_NULL;
    vs_char_t yuv420_file_name[64], nv12_file_name[64];

    //yuv420_frame_size = p_yuv420_frame->stride[0] * p_yuv420_frame->height * 3 / 2;
    nv12_frame_size = p_yuv420_frame->stride[0] * p_yuv420_frame->height * 3 / 2;
    p_yuv422_virt_addr = malloc(nv12_frame_size);
    if (p_yuv422_virt_addr == VS_NULL) {
        VS_UVC_LOG_E("malloc size[%u] failed!\n", nv12_frame_size);
        return VS_FAILED;
    }

    memset(yuv420_file_name, 0, sizeof(yuv420_file_name));
    memset(nv12_file_name, 0, sizeof(nv12_file_name));
    sprintf((char *)yuv420_file_name, "yuv420_idx%d.yuv", savefile_index);
    sprintf((char *)nv12_file_name, "nv12_idx%d.yuv", savefile_index);
    savefile_index++;
    p_yuv420_file = fopen((char *)yuv420_file_name, "wb");
    if (VS_NULL == p_yuv420_file) {
        VS_UVC_LOG_E("fopen file[%s] failed\n", yuv420_file_name);
        free(p_yuv422_virt_addr);
        return VS_FAILED;
    }
    p_nv12_file = fopen((char *)nv12_file_name, "wb");
    if (VS_NULL == p_nv12_file) {
        VS_UVC_LOG_E("fopen file[%s] failed\n", nv12_file_name);
        free(p_yuv422_virt_addr);
        fclose(p_yuv420_file);
        return VS_FAILED;
    }

    ret = frame_yuv420_nv12_conver(p_yuv420_frame, p_yuv422_virt_addr, p_yuv420_frame->width, p_yuv420_frame->height, p_yuv420_file);
    if (ret != VS_SUCCESS) {
        VS_UVC_LOG_E("yuv420_2_yuv422_conver failed\n");
        free(p_yuv422_virt_addr);
        fclose(p_yuv420_file);
        fclose(p_nv12_file);
        return ret;
    }

    fwrite(p_yuv422_virt_addr, 1, nv12_frame_size, p_nv12_file);
    fflush(p_nv12_file);

    free(p_yuv422_virt_addr);
    fclose(p_yuv420_file);
    fclose(p_nv12_file);

    return VS_SUCCESS;
}
 

这篇关于yuv420转yuv422使用neno加速的示例代码的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！


http://www.chinasem.cn/article/1014316。

相关文章

Python如何去除图片干扰代码示例

Python如何去除图片干扰代码示例

《Python如何去除图片干扰代码示例》图片降噪是一个广泛应用于图像处理的技术,可以提高图像质量和相关应用的效果,：本文主要介绍Python如何去除图片干扰的相关资料,文中通过代码介绍的非常详细,... 目录一、噪声去除1. 高斯噪声（像素值正态分布扰动）2. 椒盐噪声（随机黑白像素点）3. 复杂噪声（如伪
阅读更多...
Java Spring ApplicationEvent 代码示例解析

Java Spring ApplicationEvent 代码示例解析

《JavaSpringApplicationEvent代码示例解析》本文解析了Spring事件机制,涵盖核心概念（发布-订阅/观察者模式）、代码实现（事件定义、发布、监听）及高级应用（异步处理、... 目录一、Spring 事件机制核心概念1. 事件驱动架构模型2. 核心组件二、代码示例解析1. 事件定义
阅读更多...
python使用库爬取m3u8文件的示例

python使用库爬取m3u8文件的示例

《python使用库爬取m3u8文件的示例》本文主要介绍了python使用库爬取m3u8文件的示例,可以使用requests、m3u8、ffmpeg等库,实现获取、解析、下载视频片段并合并等步骤,具有... 目录一、准备工作二、获取m3u8文件内容三、解析m3u8文件四、下载视频片段五、合并视频片段六、错误
阅读更多...
HTML5 getUserMedia API网页录音实现指南示例小结

HTML5 getUserMedia API网页录音实现指南示例小结

《HTML5getUserMediaAPI网页录音实现指南示例小结》本教程将指导你如何利用这一API，结合WebAudioAPI，实现网页录音功能,从获取音频流到处理和保存录音，整个过程将逐步... 目录1. html5 getUserMedia API简介1.1 API概念与历史1.2 功能与优势1.3
阅读更多...
gitlab安装及邮箱配置和常用使用方式

gitlab安装及邮箱配置和常用使用方式

《gitlab安装及邮箱配置和常用使用方式》：本文主要介绍gitlab安装及邮箱配置和常用使用方式,具有很好的参考价值,希望对大家有所帮助,如有错误或未考虑完全的地方,望不吝赐教... 目录1.安装GitLab2.配置GitLab邮件服务3.GitLab的账号注册邮箱验证及其分组4.gitlab分支和标签的
阅读更多...
spring中的ImportSelector接口示例详解

spring中的ImportSelector接口示例详解

《spring中的ImportSelector接口示例详解》Spring的ImportSelector接口用于动态选择配置类,实现条件化和模块化配置,关键方法selectImports根据注解信息返回... 目录一、核心作用二、关键方法三、扩展功能四、使用示例五、工作原理六、应用场景七、自定义实现Impor
阅读更多...
SpringBoot3应用中集成和使用Spring Retry的实践记录

SpringBoot3应用中集成和使用Spring Retry的实践记录

《SpringBoot3应用中集成和使用SpringRetry的实践记录》SpringRetry为SpringBoot3提供重试机制,支持注解和编程式两种方式,可配置重试策略与监听器,适用于临时性故... 目录1. 简介2. 环境准备3. 使用方式3.1 注解方式 基础使用自定义重试策略失败恢复机制注意事项
阅读更多...
nginx启动命令和默认配置文件的使用

nginx启动命令和默认配置文件的使用

《nginx启动命令和默认配置文件的使用》：本文主要介绍nginx启动命令和默认配置文件的使用,具有很好的参考价值,希望对大家有所帮助,如有错误或未考虑完全的地方,望不吝赐教... 目录常见命令nginx.conf配置文件location匹配规则图片服务器总结常见命令# 默认配置文件启动./nginx
阅读更多...
在Windows上使用qemu安装ubuntu24.04服务器的详细指南

在Windows上使用qemu安装ubuntu24.04服务器的详细指南

《在Windows上使用qemu安装ubuntu24.04服务器的详细指南》本文介绍了在Windows上使用QEMU安装Ubuntu24.04的全流程：安装QEMU、准备ISO镜像、创建虚拟磁盘、配置... 目录1. 安装QEMU环境2. 准备Ubuntu 24.04镜像3. 启动QEMU安装Ubuntu4
阅读更多...
mysql中insert into的基本用法和一些示例

mysql中insert into的基本用法和一些示例

《mysql中insertinto的基本用法和一些示例》INSERTINTO用于向MySQL表插入新行,支持单行/多行及部分列插入,下面给大家介绍mysql中insertinto的基本用法和一些示例... 目录基本语法插入单行数据插入多行数据插入部分列的数据插入默认值注意事项在mysql中，INSERT I
阅读更多...

Copyright China编程 23002807@qq.com

沪ICP备2023033072号-2