https://111qqz.com/en/categories/%E4%BC%98%E5%8C%96/Recent content in 优化 on 111qqz's blogHugo -- gohugo.ioen© 2015-2026 111qqzSat, 21 Aug 2021 17:43:02 +0800https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E6%80%A7%E8%83%BD%E4%BC%98%E5%8C%96/2021-08-21-ska-flat-hash-map-notes/Sat, 21 Aug 2021 17:43:02 +0800https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E6%80%A7%E8%83%BD%E4%BC%98%E5%8C%96/2021-08-21-ska-flat-hash-map-notes/<h2 class="relative group">背景 <div id="背景" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e8%83%8c%e6%99%af" aria-label="Anchor">#</a> </span> </h2> <p>最近在调研各种hashmap.. 发现ska::flat hash map性能优秀。。于是来看看代码。。 发现最大的特点是,ska::flat_hash_map使用了带probe count上限的robin hood hashing</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E6%80%A7%E8%83%BD%E4%BC%98%E5%8C%96/2021-07-22-debugging-avx2-on-gcc4/Thu, 22 Jul 2021 20:01:50 +0800https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E6%80%A7%E8%83%BD%E4%BC%98%E5%8C%96/2021-07-22-debugging-avx2-on-gcc4/<h2 class="relative group">背景 <div id="背景" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e8%83%8c%e6%99%af" aria-label="Anchor">#</a> </span> </h2> <p>起因是同事在实现int4的功能，结果流水线有一条死活过不了(gcc版本为4.8.5),一直core dump 经过初步排查，找出了如下最小可以复现的代码:</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2020-03-24-tensorrt-model-compatibility-notes/Tue, 24 Mar 2020 12:26:01 +0800https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2020-03-24-tensorrt-model-compatibility-notes/<h2 class="relative group">名词说明 <div id="名词说明" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e5%90%8d%e8%af%8d%e8%af%b4%e6%98%8e" aria-label="Anchor">#</a> </span> </h2> <ul> <li>CUDA. 一般来说指的是CUDA SDK. 目前经常使用的是CUDA 8.0和CUDA 10.1两个版本. 8.0和10.1都是SDK的版本号.</li> <li>CUDNN. The NVIDIA CUDA® Deep Neural Network library (cuDNN). 是一个可以为神经网络提供GPU加速的库</li> <li>compute capability. 是GPU的固有参数,可以理解为GPU的版本.越新的显卡该数值往往越高.</li> <li>tensorRT.NVIDIA TensorRT™ is an SDK for high-performance deep learning inference. 是一个深度学习推理库,旨在提供高性能的推理速度.</li> <li>plan file,也称为 engine plan. 是生成的tensorRT 模型文件.</li> </ul> <h2 class="relative group">兼容性说明 <div id="兼容性说明" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#%e5%85%bc%e5%ae%b9%e6%80%a7%e8%af%b4%e6%98%8e" aria-label="Anchor">#</a> </span> </h2> <p><strong>Engine plan 的兼容性依赖于GPU的compute capability 和 TensorRT 版本, 不依赖于CUDA和CUDNN版本.</strong></p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E5%9B%BE%E5%83%8F%E5%A4%84%E7%90%86/2019-02-18-halide-notes/Mon, 18 Feb 2019 06:00:51 +0000https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E5%9B%BE%E5%83%8F%E5%A4%84%E7%90%86/2019-02-18-halide-notes/<blockquote><blockquote> <h3 class="relative group">**Halide is a programming language designed to make it easier to write high-performance image and array processing code on modern machines. ** <div id="halide-is-a-programming-language-designed-to-make-it-easier-to-write-high-performance-image-and-array-processing-code-on-modern-machines" class="anchor"></div> <span class="absolute top-0 w-6 transition-opacity opacity-0 -start-6 not-prose group-hover:opacity-100 select-none"> <a class="text-primary-300 dark:text-neutral-700 !no-underline" href="#halide-is-a-programming-language-designed-to-make-it-easier-to-write-high-performance-image-and-array-processing-code-on-modern-machines" aria-label="Anchor">#</a> </span> </h3> </blockquote></blockquote><p>halide有两个特性比较吸引人。一个是对于各种平台架构的支持。</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E6%80%A7%E8%83%BD%E4%BC%98%E5%8C%96/2019-01-23-loop-unrolling/Wed, 23 Jan 2019 11:51:46 +0000https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/%E6%80%A7%E8%83%BD%E4%BC%98%E5%8C%96/2019-01-23-loop-unrolling/<p><del>迫于生计，</del>最近要学习<a href="http://halide-lang.org/" target="_blank" rel="noreferrer">halide</a></p> <p>先去学习/复习一下常见的编译优化技巧。</p> <p>loop unrolling，也就是循环展开，顾名思义，就是把循环展开来写。</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-13-cuda-c-best-practices-guide-heterogeneous-computing/Tue, 13 Feb 2018 06:38:38 +0000https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-13-cuda-c-best-practices-guide-heterogeneous-computing/<p>CUDA 编程涉及到在不同的平台上同时运行代码:包含CPU的host 和包含GPU的device.</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-12-cuda-c-best-practices-parallel-computing-methodology/Mon, 12 Feb 2018 04:58:31 +0000https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-12-cuda-c-best-practices-parallel-computing-methodology/<p>APOD指的是Assess, Parallelize, Optimize, Deploy</p> <figure><img class="my-0 rounded-md" loading="lazy" decoding="async" fetchpriority="low" alt="Assess, Parallelize, Optimize, Deploy." src="http://docs.nvidia.com/cuda/cuda-c-best-practices-guide/graphics/apod-cycle.png" ></figure> <p>如图所示，APOD过程是一个循环的过程，每次只进行一部分，从A到P到O到D,然后再进行下一轮的APOD</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-10-cuda-thrust-notes/Sat, 10 Feb 2018 08:43:54 +0000https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-10-cuda-thrust-notes/<p>可以了解成并行版的STL(?</p> <p>过了一遍<a href="http://docs.nvidia.com/cuda/thrust/index.html" target="_blank" rel="noreferrer">nvidia的官方网文档</a></p> <p>发现如果熟悉STL的话,thrust没什么太多好说的,看起来很简单&hellip;</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-09-cuda-error-checking-notes/Fri, 09 Feb 2018 06:55:00 +0000https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-09-cuda-error-checking-notes/<p>由于发现cuda c++ 的 debug方式和c++ 差别很大,因此打算再开一篇,专门记录一些和error checking 以及debug有关的内容.</p>https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-01-cuda-notes/Thu, 01 Feb 2018 07:20:04 +0000https://111qqz.com/en/post/%E5%B7%A5%E7%A8%8B/cuda/2018-02-01-cuda-notes/<p>uodate:有毒吧。kernel中出问题原来是不会报错的。。。。</p>