Programming GPUs with SYCL

• Introduction to GPGPU

  Why program GPUs
  CPU VS GPU architecture
  General GPU programming tips

• SYCL for OpenCL

  Overview
  Features

• SYCL example

  Vector add

Introduction to GPGPU

Why program GPUs

• Need for parallellism to gain performance

  “Free lunch” provided by Moore’s law is Over
  adding even more CPU cores is showing diminishing returns

• GPUs are extremely efficient for

  data parallel tasks
  Arithmetic heavy computations

CPU VS GPU architecture

CPU:

• task parallellism
• small number of large cores
• seperate instrustions on each core independently
• higher power consumption
• lower memory bandwidth
• random memory access

GPU:

• data parallellism
• large number of small execution Units
• single instrustions on all multiple execution units in lock-step
• lower power consumption
• higher memory bandwidth
• sequential memory access

common CPU Architecture

common GPU Architecture

common system architecture

GPUs execute in lock-step

GPUs access memory sequentially

General GPU programming tips

• ensure the task is suitable
  GPUs are most efficient for data parallel tasks
  performance gain from prforming computing > cost of moving data
• avoid branching
  waves of processing elements execute in lock-step
  both sides of branches execute with the other masked
• avoid non-coalesced memory access
  GPUs access memory more efficiently if accessed as contiguous blocks
• avoid exponsive data movement
  the bottleneck in GPU programming is data movement between CPU and GPU memory
  it’s important to have data as clse to the procesing as possible

SYCL for OpenCL

what is OpenCL

• allows you to write kernels that execute on accelerators
• allows you to copy data between the host CPU and accelerators
• supports a wide range of devices
• comes in two components
  Host side C API for en-queueing kernels and copying Data
  Device side OpenCL C language for writing kernels

Motivation of SYCL

• make heterogeneous programming more accessible
  provide a foundation for efficient and portable templeate algorithms
• create a C++ for OpenCL ecosystem
  define an open portable standard
  provide the performance and portability of OpenCL
  base only on standard C++
• provide a high-level shared source model
  provide a high-level abstraction over OpenCL boiler plate Code
  allow C++ template libraries to target OpenCL
  allow type safety across host and device

how does shared source work? Regular C++ (Single Source)

how does shared source work? OpenCL (separate Source)

how does shared source work? SYCL (shared Source)

Suported Subset of C++ in Device Code

Suported Features

• static polymorphism
• lambdas
• classes
• operator overloading
• templeates
• placement new

Non-Suported Features

• dynamic polymorphism
• dynamic allocation
• exception handling
• RTTI
• static variables
• function pointers

SYCL example

Vector add