High Performance Computing 综述

Overview

• cggos

  • hpc
  • multicore_gpu_programming

• StreamHPC is a software development company in parallel software for many-core processors.

• The Supercomputing Blog

• FastC++: Coding Cpp Efficiently

Heterogeneous Computing

• Introduction to Parallel Computing

MPI (Message Passing Interface)

• MPI Forum: the standardization forum for MPI
• Open MPI: Open Source High Performance Computing.

Concurrency

• task switching
• hardware concurrency

Multi-Processing

IPC

• Interprocess Communications (Microsoft)

• Inter-Process Communication (IPC) Introduction and Sample Code

Multi-Threading

• POSIX C pthread
• boost::thread
• c++11 std::thread

CPU

• Intel® 64 and IA-32 Architectures Software Developer Manuals
• Hotspots, FLOPS, and uOps: To-The-Metal CPU Optimization

CPU Info

8 commands to check cpu information on Linux:

• /proc/cpuinfo: The /proc/cpuinfo file contains details about individual cpu cores.

• lscpu: simply print the cpu hardware details in a user-friendly format

• cpuid: fetches CPUID information about Intel and AMD x86 processors

• nproc: just prints out the number of processing units available, note that the number of processing units might not always be the same as number of cores

• dmidecode: displays some information about the cpu, which includes the socket type, vendor name and various flags

• hardinfo: would produce a large report about many hardware parts, by reading files from the /proc directory

• lshw -class processor: lshw by default shows information about various hardware parts, and the -class option can be used to pickup information about a specific hardware part

• inxi: a script that uses other programs to generate a well structured easy to read report about various hardware components on the system

ARM CPU features

• Runtime detection of CPU features on an ARMv8-A CPU

CPU Benchmark

Sysbench

Sysbench – Scriptable database and system performance benchmark, a cross-platform and multi-threaded benchmark tool

sysbench --test=cpu --cpu-max-prime=20000 --num-threads=4 run

htop

• htop - an interactive process viewer for Unix

• htop explained - Explanation of everything you can see in htop/top on Linux

压力测试

https://www.tecmint.com/linux-cpu-load-stress-test-with-stress-ng-tool/

cat /sys/class/thermal/thermal_zone0/temp

# stress
stress --cpu 4 --io 4 --vm 1 --vm-bytes 1G

CPU Instructions & Intrinsics

• Compiler Intrinsics

Assembly

• x86 Assembly

• winasm: The x86 Assembly community and official home of WinAsm Studio and HiEditor

• Easy Code Visual assembly IDE

• 0xAX/asm: Learning assembly for linux-x64

SIMD

Intel MMX & SSE

• Intel Intrinsics Guide

• SSE (Streaming SIMD Extentions)

  • C++ - Getting started with SSE
  • SSE - Vectorizing conditional code
  • SSE图像算法优化系列(cnblogs)

ARM NEON

Arm NEON technology is an advanced SIMD (single instruction multiple data) architecture extension for the Arm Cortex-A series and Cortex-R52 processors.

• NEON Intrinsics Reference
• ARM NEON Tutorial in C and Assembler
• ARM NEON编程初探——一个简单的BGR888转YUV444实例详解

Compiler Options:

• test ARM NEON

  gcc -dM -E -x c /dev/null | grep -i -E "(SIMD|NEON|ARM)"
  

• Raspberry Pi 3 Model B

  • g++ options

    -std=c++11 -O3 -march=native -mfpu=neon-vfpv4 -mfloat-abi=softfp -ffast-math
    

  • for the compilation error error: ‘vfmaq_f32’ was not declared in this scope, you might add the option -mfpu=neon-vfpv4 to enable __ARM_FEATURE_FMA in arm_neon.h

Reference Books:

• NEON Programmer’s Guide
• ARM® NEON Intrinsics Reference

Converter

• jratcliff63367/sse2neon
• From ARM NEON* to Intel® SSE - The Automatic Porting Solution, Tips and Tricks

OpenMP

The OpenMP API specification for parallel programming, an Application Program Interface (API) that may be used to explicitly direct multi-threaded, shared memory parallelism.

OpenMP有两种常用的并行开发形式: 一是通过简单的 fork/join 对串行程序并行化，二是采用 单程序多数据 对串行程序并行化。

• OpenMP Tutorials

• OpenMP in a nutshell

OpenMP in CMakeLists.txt:

find_package(OpenMP)
if (OPENMP_FOUND)
    set (CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
    set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
    set (CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
endif()

OpenACC

OpenACC is a user-driven directive-based performance-portable parallel programming model designed for scientists and engineers interested in porting their codes to a wide-variety of heterogeneous HPC hardware platforms and architectures with significantly less programming effort than required with a low-level model.

Intel TBB

Intel Threading Building Blocks (TBB) lets you easily write parallel C++ programs that take full advantage of multicore performance, that are portable and composable, and that have future-proof scalability.

Intel IPP

GPU

• GPU世界技术论坛
• General-Purpose Computation on Graphics Hardware

GPU Benchmark

• For the Raspberry Pi GPU benchmark, use the OpenGL 2.1 test that comes with GeeXLab

• msalvaris/gpu_monitor: Monitor your GPUs whether they are on a single computer or in a cluster

• Benchmark Your Graphics Card On Linux

watch -n 10 nvidia-smi       # 每隔10秒更新一下显卡

# on Android
watch -n 0.1 adb shell cat /sys/class/kgsl/kgsl-3d0/gpu_busy_percentage # 0.1s

Platforms

• ARM MALI GPU

• Nvidia GPU

Languages

• CUDA vs OpenCL: Which should I use?

OpenCL

OpenCL™ (Open Computing Language) is the open, royalty-free standard for cross-platform, parallel programming of diverse processors found in personal computers, servers, mobile devices and embedded platforms.

• install OpenCL

  # required: Ubuntu 16.04, nvidia GPU and nvidia driver installed
  sudo apt-get install nvidia-prime nvidia-modprobe nvidia-opencl-dev
  sudo ln -s /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 /usr/local/lib/libOpenCL.so
  

• build program

  g++ main.cpp -lOpenCL
  

CUDA

CUDA® is a parallel computing platform and programming model developed by NVIDIA for general computing on graphical processing units (GPUs).

• CUDA Zone
• ArchaeaSoftware/cudahandbook
• CUDA for ARM Platforms is Now Available

Thrust

• Thrust
• Thrust in CUDA Toolkit

DSP

Software Optimization