Q OctoML Open Source at O〇ctoML TVM Meetup 11/8/2019 Jared Roesch OctoML is a new company building DL deployment solutions using the Apache (incubating) TVM project. A goal is to nurture the TVM community

阿里云云原生应用平台 阿里巴巴 k8s 超大规模实践 关注“阿里巴巴云原生”公众号 回复 1124 获取 PPT自我介绍 •曾凡松（逐灵），当前主要负责 k8s 在阿里巴巴场景中的规模化落地，将 k8s 应用于阿里最核心的业务，帮助客 户以云原生的方式管理应用并获得效率、 稳定性及成本的改善。 •汪萌海（木苏），经历了阿里巴巴集团 集群调度从自研 sigma 系统迁移到 k8s 体系的过程，目前主要负责解决阿 体系的过程，目前主要负责解决阿 里巴巴集团在大规模场景下使用 k8s 碰到的稳定性、容器编排质量和性能问 题。❖ 阿里巴巴容器的发展历程 ❖ 基于 k8s 云原生改造实践 ❖ k8s 规模及性能优化实践 ❖ 云原生应用管理演进路线 主要内容阿里巴巴容器的发展历程 2013 初步探索 使用容器的方式替换传统使用 VM 部署应用的，基于 lxc 自研 了 t4 容器并构建了 AI 集团管理 系统 统，并构建了集团统一资源池， 在此基础上发展出弹性、混部等 技术成果，大幅降低了数据中心 的资源成本 2019 全面拥抱云原生 阿里业务全面上云，运维体系全 面拥抱云原生，基于 k8s 生态在 阿里内部蓬勃发展。在 2019 双 11 中，k8s 体系支撑了阿里史上 规模最大的集群，并提供了极速 的应用扩容体验 2015 野蛮生长 体验到使用容器部署应用的优势 后，阿里内部发展出众多的运维 平台，包括

CppCon 2024: Improving our safety with a quantities and units library C++ developers needs help 8• Aerospace • Autonomous cars • Embedded industries CppCon 2024: Improving our safety with a quantities dimension T-1 angular_velocity\n(angular_displacement / duration)\n[rad/s, 1/s]\n{vector} frequency\n(1 / period_duration)\n[Hz, 1/s] rotational_frequency\n(rotation / duration)\n[1/s] angular_frequency\n(phase_angle mp_units::si { inline constexpr struct radian final : named_unit<"rad", metre / metre, kind_ofangular_measure>> {} radian; inline constexpr struct steradian final : named_unit<"sr", square(metre) / square(metre)

Intra-register sorting • Fast linear median-of-seven filter • Fast small-kernel convolution • Faster (?) UTF-8 to UTF-32 conversion (with AVX2) • No heavy code, but lots of pictures • Thinking "vertically" CppCon 2020 Bob Steagall K E W B C O M P U T I N G Convolution CppCon 2020 - Adventures in SIMD Thinking 8 S = s0 s1 s2 s3 s4 s5 s6 ... K = k0 k1 k2 r0 = ... r1 = s0k0 + s6 s7 s8 k1 k1 k1 k1 k1 k1 k1 k1 s2 s3 s4 s5 s6 s7 s8 s9 k2 k2 k2 k2 k2 k2 k2 k2 r1 r2 r3 r4 r5 r6 r7 r8Copyright © 2020

RAII •Beyond basic lifetime •Common pitfalls •Value categories •Unions/Arrays •Any assembly code 8 On this talk…• Principal Software Engineer @ Cadence Design Systems • Learning C++ since 2013 • Can’t olt.org/z/ffzG8hPT8 struct Foo { Foo() { std::cout << "Foo()" << std::endl; } ~Foo() { std::cout << "~Foo()" << std::endl; } }; int main() { Foo a; { Foo b; } } 1 2 3 4 5 6 7 8 9 10 11 21 Constructors olt.org/z/ffzG8hPT8 struct Foo { Foo() { std::cout << "Foo()" << std::endl; } ~Foo() { std::cout << "~Foo()" << std::endl; } }; int main() { Foo a; { Foo b; } } 1 2 3 4 5 6 7 8 9 10 11 22 Constructors

operation that’s not too boring, say matrix addition 6Back to high school 7Back to high school 8Back to C++ If the matrices are contiguous in memory, then we can treat them as ranges and simply use rhsRowStride, 5 float* dst, size_t dstRowStride 6 ) { 7 for (size_t i = 0; i < rows; ++i) 8 for (size_t j = 0; j < cols; ++j) 9 dst[dstRowStride * i + j] = 10 j]; 12 } 12add r10, r11 add r9, r8 cmp rcx, rdi je .LBB0_16 .LBB0_3: test rsi, rsi je .LBB0_2 cmp rsi, 8 jb .LBB0_11 mov r11, rbx

“people are not writing enough functions” 7Tony Van Eerd: “people are not writing enough functions” 8When people finally start writing more functions, we’d prefer to get only the well designed ones!When System V gABI psABI: ARM, x86, … C++ Itanium ABI Microsoft Windows ABIs C++ Standard 18armv8-a System V ABI - iPhone - M1 Mac and newer - Android smartphone x86-64 (AMD64) System V ABI - Linux x86 (IA-32) Microsoft ABI - ancient Windows device 19armv8-a System V ABI Procedure Call Standard for the Arm® 64-bit Architecture armv8-a clang 18.1.0 -O2 -std=c++20 x86-64 (AMD64) System V ABI

Intra-register sorting • Fast linear median-of-seven filter • Fast small-kernel convolution • Faster (?) UTF-8 to UTF-32 conversion (with AVX2) • No heavy code, but lots of pictures • Thinking "vertically" CppCon const* psrc) { return _mm512_loadu_ps(psrc); } m0 m1 m2 m3 m4 m5 m6 m7 m8 m9 m10 m11 m12 m13 m14 m15 m0 m1 m2 m3 m4 m5 m6 m7 m8 m9 m10 m11 m12 m13 m14 m15 psrcCopyright © 2020 Bob Steagall K E W B C m6 m7 m8 m9 m10 m11 m12 m13 m14 m15 f f m2 m3 f f m6 m7 f f f f m12 m13 m14 m15 f 0 0 1 1 0 0 1 1 0 0 0 0 1 1 1 1 fill mask psrcCopyright © 2020 Bob Steagall K E W B C O M P U T I N G 8 CppCon

Ben the count one digit at a time. 3. When done, yell "End of Line" Exercise 7async/concurrency 8async/concurrency GPIO Timer 1 Timer 2 I2C DMA Half-Duplex Serial core 9Sender World View 10Sender async::continue_on(s2) | async::then([] (int i) { return std::to_string(i); }) ; 1 2 3 4 5 6 7 8 auto r = comp | async::sync_wait(); 9 auto [str] = r.value_or(std::make_tuple(""_s)); 10 23Composition async::continue_on(s2) | async::then([] (int i) { return std::to_string(i); }) ; 1 2 3 4 5 6 7 8 auto r = comp | async::sync_wait(); 9 auto [str] = r.value_or(std::make_tuple(""_s)); 10 auto r =

1 MB = 33,725 devices >= 1 MB = 4,964 devicesMicrocontroller Constraints: Limited Storage & Memory 8 < 100kB = 31,098 devices >= 100kB = 7095With so little memory how do you prevent running out of std::vector std::unordered_map std::shared_ptr 111213 Lets not forget about ● std::inplace_vector: p0843r8 voted into C++26 @ St Louis 2024 ISO C++ meetingTIME SPACE So why do firmware developers avoid C++ the problem: Return mismatch 27 /// @returns false if this failed! bool io_expander_level(std::uint8_t p_pin_number, bool p_high_voltage) { // Implementation details... } void some_output_pin::level(bool