4zhvml8 发表于 2024-7-9 14:53:36

直播回顾丨百度ACU软硬件结合优化的实践经验分享


    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://mmbiz.qpic.cn/mmbiz_jpg/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqib6Gz2GXz7WIl7xPxbLDGgDdPia9tDFtJgXfTLia0qoz198icHTOYk0wfg/640?wx_fmt=jpeg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">百度ACU(Apollo Computing Unit)</span></strong><span style="color: black;">是Apollo面向量产的自动驾驶车载计算单元,<span style="color: black;">按照</span><span style="color: black;">区别</span>需求场景的计算能力<span style="color: black;">需求</span>,分为多个系列<span style="color: black;">制品</span>。ACU-Advanced是行业首创的自主泊车<span style="color: black;">制品</span>Apollo Valet Parking专用车载计算平台,<span style="color: black;">日前</span>已量产下线用于客户的量产车型中。</span></p>
    <div style="color: black; text-align: left; margin-bottom: 10px;">
      <div style="color: black; text-align: left; margin-bottom: 10px;">
            <div style="color: black; text-align: left; margin-bottom: 10px;">
                <div style="color: black; text-align: left; margin-bottom: 10px;">
                  <div style="color: black; text-align: left; margin-bottom: 10px;">
                        <div style="color: black; text-align: left; margin-bottom: 10px;">关闭</div>
                        <div style="color: black; text-align: left; margin-bottom: 10px;"><strong style="color: blue;">观看<span style="color: black;">更加多</span></strong></div>
                        <div style="color: black; text-align: left; margin-bottom: 10px;"><span style="color: black;">更加多</span></div>
                  </div>
                </div>
            </div>
      </div>
      <div style="color: black; text-align: left; margin-bottom: 10px;">
            <div style="color: black; text-align: left; margin-bottom: 10px;">
                <div style="color: black; text-align: left; margin-bottom: 10px;">
                  <div style="color: black; text-align: left; margin-bottom: 10px;">
                        <div style="color: black; text-align: left; margin-bottom: 10px;">退出全屏</div>
                  </div>
                </div>
            </div>
      </div>
      <div style="color: black; text-align: left; margin-bottom: 10px;">
            <div style="color: black; text-align: left; margin-bottom: 10px;">
                <div style="color: black; text-align: left; margin-bottom: 10px;">
                  <div style="color: black; text-align: left; margin-bottom: 10px;">
                        <div style="color: black; text-align: left; margin-bottom: 10px;">
                            <div style="color: black; text-align: left; margin-bottom: 10px;">
                              <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">视频加载失败,请刷新页面再试</p><a style="color: black;"><img src="data:image/png;base64,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" style="width: 50%; margin-bottom: 20px;"> 刷新 </a>
                            </div><img src="data:image/png;base64,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" style="width: 50%; margin-bottom: 20px;">
                        </div>
                  </div>
                </div>
            </div>
      </div>
    </div>
    <div style="color: black; text-align: left; margin-bottom: 10px;">
      <div style="color: black; text-align: left; margin-bottom: 10px;"> <a style="color: black;"> 视频详情 </a> </div>
    </div>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;">ACU-Advanced软硬一体的<span style="color: black;">处理</span><span style="color: black;">方法</span>:</p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">能够支持5路摄像头,12路超声波雷达,预留毫米波雷达和激光雷达接口;</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">基于Xilinx ZU5(FPGA)设计,适配百度PaddlePaddle深度学习框架;</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">全面<span style="color: black;">安排</span>AUTOSAR,全车规<span style="color: black;">基本</span>软件<span style="color: black;">研发</span>;</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">严谨的软硬件安全合规设计,全面满足AVP场景需求(ASIL-B);</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">100%全车规元器件,抗冲击震动,耐电磁干扰,全面胜任车载环境;</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">-40℃~85℃超宽工作温度,自然风冷<span style="color: black;">没</span>需风扇或水冷;</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">设计流程,供应链和生产管控基于IATF16949质量管理体系。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">为了让客户、合作伙伴以及<span style="color: black;">研发</span>者能够更加深入<span style="color: black;">认识</span>百度ACU(Apollo Computing Unit),<span style="color: black;">这次</span>公开课邀请了百度ACU核心技术团队的工程师,为<span style="color: black;">大众</span>准备了4期公开课直播。</span></strong><span style="color: black;">今天为<span style="color: black;">大众</span>送上第二期直播内容回顾,<span style="color: black;">倘若</span>你<span style="color: black;">错失</span>了直播,<span style="color: black;">那样</span>这次的直播回顾你<span style="color: black;">必定</span><span style="color: black;">不可</span><span style="color: black;">错失</span>。阿波君给<span style="color: black;">大众</span>整理好了笔记,快来<span style="color: black;">一块</span>学习吧!</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">上期<span style="color: black;">咱们</span>为<span style="color: black;">大众</span>整理了</span><a style="color: black;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;">《</span></strong></span><strong style="color: blue;"><span style="color: black;">直播回顾丨百度ACU如</span></strong></a><strong style="color: blue;">何满足自动驾驶量产需求<strong style="color: blue;">》</strong></strong></span>,<span style="color: black;">本</span>期公开课由百度智能汽车事业部底层软件架构师——张扬为<span style="color: black;">咱们</span>讲解<strong style="color: blue;"><span style="color: black;">《百度ACU软硬件结合优化的实践经验分享》</span></strong>。</p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">ENJOY THE FOLLOWING &nbsp;</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">Apollo计算单元<span style="color: black;">重点</span>讲的是计算,Apollo自动驾驶对算力的需求,由需求引导至<span style="color: black;">为么</span><span style="color: black;">选择</span>AVP这款芯片,稍后会讲到技术架构方便大家理解,第四部分<span style="color: black;">便是</span>主题:在<span style="color: black;">有效</span>计算方面做了软硬件结合的实践,最后是总结。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSJPH6PX7cUeyNBYGvtk3Cadociaicyof3FKhMx0J5UgstJ1tgcic7gSMjpMNR6ILfVHialJ1uheNHXO7Q/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqf5fQ2G05WVxupk7YVb50YibrvB7rR3wp1r5eQckC5N3zjrUkeSea3mQ/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">自动驾驶的计算种类<span style="color: black;">包含</span>:<strong style="color: blue;"><span style="color: black;">神经网络、图像处理、3D重构</span></strong>与VSLAM和矩阵操作。自动驾驶的实现,依赖<span style="color: black;">各样</span>各样的算法。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">算力与帧率不必然成正比,有以下几种<span style="color: black;">原因</span>:<span style="color: black;">首要</span>是Elementwise算子带宽受限,其次是深度卷积算子<span style="color: black;">是不是</span>被优化,最后是<span style="color: black;">是不是</span>所有算子被加速。极端<span style="color: black;">状况</span>下,占<span style="color: black;">少于</span>1%计算量的算子,计算时间超过50%,<span style="color: black;">明显</span>影响帧率。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">简单<span style="color: black;">来讲</span>,<span style="color: black;">咱们</span>对硬件的需求是:算力强、带宽高、算子齐备、功耗低、符合车规。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSJPH6PX7cUeyNBYGvtk3CadeX9XxLnu1icFsUAjwjzxgRI1Y3sqn9DjjeFsD2MeTNOCLKOKjzUIibbw/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqQpicPIDq6DlIeVNunBvFa8FhY8nTMFp0ZK7um3QAxCkk612rezIe01Q/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">用什么芯片能满足自动驾驶的计算,能适应<span style="color: black;">区别</span>计算需求呢?<span style="color: black;">咱们</span><span style="color: black;">选择</span>了一款<strong style="color: blue;"><span style="color: black;">SOC——Xilinx ZU5</span></strong>,里面<span style="color: black;">包含</span>了FPGA,<span style="color: black;">首要</span>是比较灵活,有利于算法迭代;其次<span style="color: black;">供给</span>足够的算力,保持行驶速度,不是像蜗牛那样缓慢前行;<span style="color: black;">而后</span><span style="color: black;">便是</span><span style="color: black;">必须</span>满足85℃环境下正常<span style="color: black;">运用</span>的严苛车规<span style="color: black;">需求</span>。另一点看中的<span style="color: black;">便是</span>基于FPGA的SOC的<span style="color: black;">靠谱</span>性能,广泛应用于<span style="color: black;">靠谱</span>性<span style="color: black;">需求</span>比较高的行业,<span style="color: black;">例如</span>说通信行业、军工行业、航空航天行业,<span style="color: black;">亦</span>用于汽车行业。自动驾驶最担心的<span style="color: black;">便是</span>安全问题,我们致力于<span style="color: black;">保准</span>自动驾驶的安全。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSJPH6PX7cUeyNBYGvtk3CadU8uZdYlOicc9EINM1PA2E7hUxTYyoicpS0SfGM1QplenQQhyy4lQ0nng/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqahXPEAzT32NtBmPk6j23XeiboUoIBtYu88ic6ugOz3icpeBIzzfWhtnPg/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">从软件来看,<span style="color: black;">首要</span><span style="color: black;">研发</span>加速算子,使芯片<span style="color: black;">拥有</span>加速的能力;<span style="color: black;">而后</span>做<span style="color: black;">有些</span>驱动的程序来驱动,负责管理;最后基于PaddleLite做的深度学习模型编译及前向推测框架,<span style="color: black;">另一</span>是传统的CV算法。从硬件的<span style="color: black;">方向</span>来看,绿色的部分<span style="color: black;">表率</span>FPGA,<span style="color: black;">首要</span>有一个命令解析的模块,从ARM<span style="color: black;">经过</span>寄存器下发的命令解析出来,<span style="color: black;">而后</span>下发给相应的计算单元,<span style="color: black;">例如</span>说CNN、图像处理、SLAM等,<span style="color: black;">由于</span>深度学习和定位要并行,<span style="color: black;">因此</span>设计了两个<span style="color: black;">同期</span><span style="color: black;">拜访</span>内存的通道,<span style="color: black;">能够</span><span style="color: black;">保准</span>实时<span style="color: black;">有效</span>的运算。ARM<span style="color: black;">亦</span>是<span style="color: black;">经过</span>DDR<span style="color: black;">掌控</span>器和内存进行通信。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">重点</span>的计算都是在深度学习,用的都是CNN卷积神经网络,而<span style="color: black;">这里</span>过程中,CPU上对Paddle模型进行<strong style="color: blue;"><span style="color: black;">解析、算子合并、模型编译、内存开辟、部分算子</span></strong>计算,<span style="color: black;">经过</span>驱动程序对FPGA寄存器下发命令并管理共享的DDR内存,对FPGA而言,<span style="color: black;">必须</span>解析命令、<span style="color: black;">掌控</span><span style="color: black;">规律</span>、数据搬运、量化单元、算子加速以及管理内存的模块DMA。软硬结合的<span style="color: black;">最后</span><span style="color: black;">目的</span>是<strong style="color: blue;"><span style="color: black;"><span style="color: black;">加强</span>感知、定位、环境建模、规划<span style="color: black;">掌控</span>的处理帧率,<span style="color: black;">保准</span>数值精度和准召率。</span></strong>技术手段<span style="color: black;">第1</span>是<span style="color: black;">经过</span><span style="color: black;">加强</span>FPGA的算力,第二是<span style="color: black;">加强</span>FPGA对计算的覆盖度,第三是量化训练及计算仿真<span style="color: black;">工具</span>。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSJPH6PX7cUeyNBYGvtk3CadVcgBReEIjyrib0X62AcpYzudpMVbZzU98MxcjZIHkg3Hhps1DqNGKQQ/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqjLQJnFqSXrIxWQgRzpfbGhGZ4e6IuZyPQdfz2tXtK2iahk21kTObOCA/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;"><span style="color: black;">第1</span>点是进行了8bit的量化训练,</span></strong><span style="color: black;">与众<span style="color: black;">区别</span>的是<span style="color: black;">供给</span>跟FPGA计算相匹配的量化训练<span style="color: black;">工具</span>,实践证明,8bit计算后准召率下降,<span style="color: black;">同期</span>为了<span style="color: black;">保准</span>算法能够<span style="color: black;">运用</span>,<span style="color: black;">供给</span>了FPGA计算仿真器,计算结果<span style="color: black;">保准</span>与bit对齐。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">第二点是利用了PFGA这款DSP的特殊结构,</span></strong><span style="color: black;">这个计算单元<span style="color: black;">能够</span><span style="color: black;">同期</span>计算一个27位✖27位的,那<span style="color: black;">咱们</span>是8位✖8位的,就把一个乘数高8位,一个乘数低8位,<span style="color: black;">这般</span>乘出来的结果是<span style="color: black;">第1</span>个的乘积和第二个的乘积自然<span style="color: black;">便是</span><span style="color: black;">掰开</span>的。<span style="color: black;">针对</span>其他的<span style="color: black;">咱们</span>都采用<span style="color: black;">选定</span>化计算,<span style="color: black;">首要</span>要<span style="color: black;">保准</span>与浮点计算相比匹配度高于98%,在这种前提下,计算<span style="color: black;">过程</span><span style="color: black;">非常多</span>,<span style="color: black;">咱们</span>充分理解每一个计算环节,采用<span style="color: black;">选定</span>化要<span style="color: black;">保存</span>多少位来进行计算,<span style="color: black;">每一个</span>环节都采用变长计算,<span style="color: black;">保准</span>计算最<span style="color: black;">有效</span>。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqk2mSgfyB13OEvJticKTA0ribxZqHTzWdvcm8l1icCsqEZvr08OJaDC3TA/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">第三点定义广义算子,</span></strong><span style="color: black;">当某些计算类型比较类似,就抽象成一个计算形式。<span style="color: black;">这般</span>就<span style="color: black;">能够</span>复用FPGA资源,凡是被加速的算子<span style="color: black;">能够</span>得到足够多的资源<span style="color: black;">保准</span>算力。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqR71GP6pvqGaQEaC3tRPDIleCd8rhXvXB2lygOEG2YXE8xvOviazlssw/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">第四点是在计算中遇到的固定连接,</span></strong><span style="color: black;"><span style="color: black;">倘若</span>不做特殊处理,算完一个后<span style="color: black;">必须</span>把结果读到内存,下一次计算的时候<span style="color: black;">必须</span>从内存里取出来再次加工,这种存入读取的方式费时、效率低,<span style="color: black;">此时</span><span style="color: black;">能够</span>把这些当成一个计算来进行处理,像之前说的SDK对算子进行融合,减少带宽占用。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">还有一种<span style="color: black;">便是</span><span style="color: black;">针对</span>参数很大的卷积,要经过算子分裂,分裂成小卷积后再计算,能有效地降低FPGA片上存储资源的占用。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">针对</span>图像采集,采用R核图像驱动,共享内存实现0拷贝,<span style="color: black;">明显</span>降低了CPU占用和内存宽带,<span style="color: black;">加强</span>了图像帧率的稳定性。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">自动泊车有两种业务,一种是基于高精地图的PAVP,一个是基于学习建图的HAVP,FPGA加速算子,随着应用场景PAVP/HAVP切换,使得<span style="color: black;">每一个</span>场景都达到最高帧率。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSJPH6PX7cUeyNBYGvtk3Cad6cNKy7k17OjLkMjyRMrILUILoBTvLTFrtYHZuYEjSGjJic4z49fxtEA/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">自动驾驶的计算需求对承接计算加速<span style="color: black;">来讲</span>具有不确定性,在迭代的过程中,有些算子远<span style="color: black;">无</span>预期的效果,并非是只<span style="color: black;">处理</span>深度学习就能<span style="color: black;">处理</span>的问题,<span style="color: black;">因此</span><span style="color: black;">选取</span>了<span style="color: black;">运用</span>FPGA来做,<span style="color: black;">选择</span>ZU5芯片,这个<span style="color: black;">研发</span>是比较困难的。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gyAWOzWgf2e8JKJSH1AMicjqOrcDWbTtuD4qdStjk7It8GW7nbDI0ibgHTIIbWPHSLDnibQkbW5Z8INg/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">拓展:</span></strong><span style="color: black;">底层软件<span style="color: black;">包含</span>了<strong style="color: blue;"><span style="color: black;">模块<span style="color: black;">形成</span>——盖亚层</span></strong>,用来屏蔽硬件的差距,<span style="color: black;">亦</span><span style="color: black;">便是</span>说<span style="color: black;">此刻</span>用的ZU5芯片,将来可能会用TDS,<span style="color: black;">期盼</span>在硬件迁移的过程中,上层应用软件不<span style="color: black;">出现</span>改变,<span style="color: black;">能够</span>直接迁移。硬件迁移的过程被盖亚层所屏蔽掉。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">今天<span style="color: black;">触及</span>到模块有定制算子、硬件模式管理、传感器同步/预处理、DL/DV加速,还有硬件仿真器,用于X86等环境。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;">以上<span style="color: black;">便是</span><span style="color: black;">咱们</span>讲解的<strong style="color: blue;"><span style="color: black;">百度ACU软硬件结合优化的实践经验</span></strong>,<span style="color: black;">更加多</span>系列主题讲解,请<span style="color: black;">连续</span>关注Apollo<span style="color: black;">研发</span>者社区。</span></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">点击<span style="color: black;">文案</span>左下角</span></strong><strong style="color: blue;"><span style="color: black;">『阅读原文』</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><strong style="color: blue;"><span style="color: black;">可观看直播回放</span></strong></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSLstGztODuYQd2S0HcNUCce5jwk6tl6LW5ZUPdTb9xM0L9ARz8Q1TYzicmz6CN9MZGiaer67iaLgZuKw/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSLstGztODuYQd2S0HcNUCceQACbXzDQaovV9uOiaT6JlRXvViaoP5b7aWJ4nENQU7jnleRT4KDfL1QA/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><a style="color: black;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSJPH6PX7cUeyNBYGvtk3CadeI9jjwXIyHTe8crrDBV4IgRjibibiaibE3eKSypbKUsWNeT43tAsXPZayQ/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></a></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><a style="color: black;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/C4wVziccAsSIZhkrW19lT9CGamgGrFwYm9uPSmGzoaqRS2CYAVSyefquCWxsewJfY0gibYgK9lxic3OfMpqhRELRg/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></a></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><a style="color: black;"><span style="color: black;"><img src="https://mmbiz.qpic.cn/mmbiz_png/Lic3WIjno4gxXibRsyaQ65ZIjhH6kXqawg1XSia5RTDTWHjExBLpzZuEDv3BZWGicNw5TJwLHvuwCX6JbpWRkCicy7w/640?wx_fmt=png&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></span></a></p><img src="https://mmbiz.qpic.cn/mmbiz_jpg/C4wVziccAsSIS288mcoy3tGjZ4DB6YaPKJ5QSm1ibEd5Lv1jiao8DqL9j3uKkmyC6D4ibzcuTkVyu6bgkTyicjGF84w/640?wx_fmt=jpeg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;">
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://mmbiz.qpic.cn/mmbiz_jpg/C4wVziccAsSIS288mcoy3tGjZ4DB6YaPKggaDDyuWyWZwV7I0mqIEocLRU11OyYvsNASEB3JGfcrts94zRhvAKA/640?wx_fmt=jpeg&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1&amp;wx_co=1" style="width: 50%; margin-bottom: 20px;"></p>
    <p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><img src="https://mmbiz.qpic.cn/mmbiz_gif/Lic3WIjno4gzkSrGmcib4XfCq0MOolaiaq0TFmXrfAzbkB2Coic4XXdH1dooysL0ZlSxVf45k9zwoaFs9WeFoZjfMQ/640?wx_fmt=gif&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1" style="width: 50%; margin-bottom: 20px;"><img src="https://mmbiz.qpic.cn/mmbiz_gif/Lic3WIjno4gzdCRu197ulzGTmmZrt4tzexC3EdJLmW7CNWCjOjVNEHo57Hctgnn8U2RjhpGibEQZfCn68xrUefgA/640?wx_fmt=gif&amp;tp=webp&amp;wxfrom=5&amp;wx_lazy=1" style="width: 50%; margin-bottom: 20px;"></p>




流星的美 发表于 2024-8-22 09:24:11

我完全赞同你的观点,思考很有深度。

星☆雨 发表于 2024-8-25 02:22:04

你的话语如春风拂面,让我感到无比温暖。

m5k1umn 发表于 2024-9-27 21:59:31

论坛是一个舞台,让我们在这里尽情的释放自己。

nykek5i 发表于 2024-10-19 19:58:24

交流如星光璀璨,点亮思想夜空。

b1gc8v 发表于 2024-11-12 13:33:34

回顾历史,我们感慨万千;放眼未来,我们信心百倍。

qzmjef 发表于 4 天前

你的话语如春风拂面,让我心生暖意。
页: [1]
查看完整版本: 直播回顾丨百度ACU软硬件结合优化的实践经验分享