Advanced SIMD Algorithms in Pictures ADVANCED SIMD ALGORITHMS IN PICTURES Denis Yaroshevskiy hapsydenisyaroshevskiygithubialpresentatonsldyaroshev_presentatonsjconferance_ialksadvanced_simd_algorthms_in_pictures anced_simd_algorthms_in_pictures html Advanced SIMD Algorithms in Pictures 6 1015123. 10.53 PM Advanced SIMD Algorithms MEMCMP hpsyidenisyaroshevskiy gihub_olpresentatonsidyarashev_prasentatonslconferance_talksladvanced_simd_algorthms_m_pictures ntml 06 1015123. 10

Adventures in SIMD Thinking (Part 2 of 2) Bob Steagall CppCon 2020 K E W B C O M P U T I N GCopyright © 2020 Bob Steagall K E W B C O M P U T I N G Agenda • Learn a little about Intel's SIMD facilities I don't work for Intel) • Create some useful functions in terms of AVX-512 intrinsics • Try some SIMD-style thinking to tackle some interesting problems • Intra-register sorting • Fast linear median-of-seven Thinking "vertically" CppCon 2020 - Adventures in SIMD Thinking 2Copyright © 2020 Bob Steagall Small-Kernel Convolution 3 CppCon 2020 - Adventures in SIMD ThinkingCopyright © 2020 Bob Steagall K E W B

Adventures in SIMD Thinking (Part 1 of 2) Bob Steagall CppCon 2020 K E W B C O M P U T I N GCopyright © 2020 Bob Steagall K E W B C O M P U T I N G Agenda • Learn a little about Intel's SIMD facilities I don't work for Intel) • Create some useful functions in terms of AVX-512 intrinsics • Try some SIMD-style thinking to tackle some interesting problems • Intra-register sorting • Fast linear median-of-seven Thinking "vertically" CppCon 2020 - Adventures in SIMD Thinking 2Copyright © 2020 Bob Steagall K E W B C O M P U T I N G 3 CppCon 2020 - Adventures in SIMD Thinking Getting Started #include #include

(DLR) Institute of Software Methods for Product Virtualization VECTORIZING A CFD CODE WITH STD::SIMD SUPPLEMENTED BY (ALMOST) TRANSPARENT LOADING AND STORINGMotivation: The Origin of the Talk The load and store std::simd and scalar variables ▪ syntactically equalize scalar and vectorized code The talk: ▪ share experience with vectorization using std::simd ▪ introduce the SIMD_ACCESS library 3 multiple set of operands at once → Single Instruction Multiple Data (SIMD) For more details Matthias Kretz‘ Cppcon talk about std::simd: https://youtu.be/LAJ_hywLtMA 4 Olaf Krzikalla, DLR SP, 2024-09-17

std::simd how to express inherent parallelism efficiently via data-parallel types Dr. Matthias Kretz GSI Helmholtz Center for Heavy Ion Research CppCon ’23 @mkretz@floss.social github.com/mattkretzMotivation on std::simd Overview Example: Image Processing Programming Models Outlook Summary Goals and non-goals for this talk • This is not a tutorial! You won’t really know how to use the std::simd API after tangents to take; we don’t have that time. • My goal is to share my vision. Take your view from SIMD registers up to designing efficient software. • How to think / design… • I might have promised too

differ depending upon the target machine architecture. • Features may be hardware: CPU architecture, SIMD instruction set, DMA controller, GPIO module, etc. • Features may be software: OS, graphics API and File Structure Flat Deep plt/simd Simd.h Neon32.h Sse.h Sse2.h … plt/math Quat.h Quat_Common.h Quat_Neon32.h Quat_Sse.h Quat_Sse2.h … plt/simd Simd.h Neon32.h Sse.h Sse2.h … plt/math series of preprocessor macros handles generating the header file name to load • Example: INCLUDE_SIMD(Quat) becomes: ”Quat_SSE2.h”Header Inclusion Macros • #define INCLUDE_PLT(Feature, File) INCL

xmm1 xmm0 为什么需要 SIMD ？单个指令处理四个数据 • 这种单个指令处理多个数据的技术称为 SIMD （ single-instruction multiple-data ）。 • 他可以大大增加计算密集型程序的吞吐量。 • 因为 SIMD 把 4 个 float 打包到一个 xmm 寄存器里同时运算，很像数学中矢量的逐元 素加法。因此 SIMD 又被称为矢量，而原始的一次只能处理 在一定条件下，编译器能够把一个处理标量 float 的代码，转换成一个利用 SIMD 指令的 ，处理矢量 float 的代码，从而增强你程序的吞吐能力！ • 通常认为利用同时处理 4 个 float 的 SIMD 指令可以加速 4 倍。但是如果你的算法不 适合 SIMD ，则可能加速达不到 4 倍；也有因为 SIMD 让访问内存更有规律，节约了指 令解码和指令缓存的压力等原因，出现加速超过 4 将两个 int32 的写入合 并为一个 int64 的写入 。 合并写入：不能跳跃 但如果访问的两个元素地 址间有跳跃，就不能合并 了。 第 4 章：矢量化 更宽的合并写入：矢量化指令（ SIMD ） 两个 int32 可以合并为一个 int64 四个 int32 可以合并为一个 __m128 xmm0 由 SSE 引入，是个 128 位寄存 器 他可以一次存储 4 个 int ，或 4

............................................................. 80 8.9 RV32V，MIPS-32 MSA SIMD 和 x86-32 AVX SIMD 的比较 ......................................... 81 5 8.10 结束语 ..................... ............................................................. 113 11.7 “P”标准扩展：封装的单指令多数据（Packed-SIMD）指令 ................................... 114 11.8 “Q”标准扩展：四精度浮点 ................................. 2017]，这意味着x86指令的增长速率提高到了（在1978年到 2015年之内）每四天增长一条。我们是用汇编语言指令计算的，他们想必算入了机器语言指令。正如第八 章所解释的那样，这个增长的很大一部分是因为x86 ISA依赖于SIMD指令来实现数据级并行。 图1.3：x86-32 ASCII Adjust after Addition（aaa）指令的描述。它以二进制编码十进制数（BCD）形式 进行计算机运算，这

示例：用 RV32V 编写 DAXPY 程序 . . . . . . . . . . . . . . . . . . . 78 8.9 对比 RV32V、MIPS-32 MSA SIMD 和 x86-32 AVX SIMD . . . . . . . 79 8.10 结语 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 11.6 “N” 标准扩展：用户态中断 . . . . . . . . . . . . . . . . . . . . . . . . 119 11.7 “P” 标准扩展：紧缩 SIMD 指令 . . . . . . . . . . . . . . . . . . . . . 119 11.8 “Q” 标准扩展：四倍精度浮点 . . . . . . . . . . . . . 2017]，按这个数据，在 1978 到 2015 年期间，x86 指令 平均每 4 天增长 1 条。我们统计的是汇编语言指令，他们统计的也许是机器语言指令。正如第 8 章所介 绍，增长的主要原因是 x86 ISA 通过 SIMD 指令实现数据级并行。 AL 寄存器是默认的源寄存器和目的寄存器。 If AL 寄存器的低 4 位大于 9， 或辅助进位标志 AF 为 1， Then AL 的低 4 位加 6 且忽略溢出

Memory SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Unit Cache SIMD Exec Unit Cache SIMD Exec Unit Cache Texture Texture Texture Texture Tessellate Tessellate Culling Rasterizer Culling Rasterizer Culling Rasterizer Culling Rasterizer