Look ma, no CUDA! Programming GPUs with modern C++

This repo contains all the stuff used during the talk I gave at the Italian C++ Community meetup in Modena, 2019-05-09 about GPUs and how SYCL allows us to finally program them in a sane way. Please feel free to steal anything you want.

Contents

Path	What to expect
slides/	The slide deck. Nothing more, nothing less.
examples/	Show me the code.
cmake/	CMake helpers, mainly intended to support ComputeCpp and directly copied from its distribution.
external/	All git submodules related to external dependencies.

Getting Started

The CMakeLists.txt provided in this repo builds the following examples as separate executables:

Example	Depends on	WAT?
vector_add_sycl.cpp	SYCL	Parallel vector addition with linear accessors
vector_add_sycl_serial.cpp	SYCL	Serial vector addition with linear accessors
vector_add_tmp_sycl.cpp	SYCL	Templated (on value type) parallel vector addition with linear accessors
vector_add_tmp_sycl_serial.cpp	SYCL	Templated (on value type) serial vector addition with linear accessors
matrix_add_sycl.cpp	SYCL	Matrix (2-dimensions) parallel addition with 2-dimensional accessors
vector_add_opencl.cpp	OpenCL	Parallel vector addition using a linear kernel grid
vector_add_cuda.cu	CUDA	Parallel vector addition using a linear kernel grid

OpenCL and CUDA examples are built only if a suitable implementation can be found on the host while SYCL based ones are built anyway.

All of the SYCL based examples have been tested with the following implementations:

• triSYCL, the open source, standard reference implementation. Depends on boost and a C++ with OpenMP support;
• ComputeCpp, a commercial implementation by codeplay with a freely available community edition.

Please note that if SYCL cannot be found on the host, the triSYCL source tree will be automatically downloaded and included via git submodule.

Build configuration

The following CMake options are available:

Option	Default	Meaning
MEETUPCPP_SYCL	triSYCL	SYCL implementation to be used; allowed values are: triSYCL, ComputeCpp
MEETUPCPP_TRISYCL_DIR	<repo>/external/triSYCL	Path of the triSYCL source to be used when triSYCL is selected; defaults to vendored submodule
COMPUTECPP_BITCODE	ptx64	Bitcode generation target for ComputeCpp; defaults to NVIDIA PTX
TRISYCL_OPENCL	ON	Use OpenCL execution backend; ignored when MEETUPCPP_SYCL isn't set to triSYCL or OpenCL isn't available

Credits

A lot of excellent charts and pitches have been taken directly from publicly available talks across the net, I've strived really hard to put proper credits on the slides but if you notice that something's missing please open an issue on this repo.

Thanks everyone for your help in making my slides more clear and understandable. I'm particularly grateful to (in order of appearance):

• Karl Rupp
• Michael Wong
• Gordon Brown
• Aksel Alpay

Resources

• sycl.tech: a comprehensive community platform that collects articles, blog posts, talks and everything related to SYCL that shows up on the web.
• Accelerating your C++ on GPU with SYCL by Simon Brand - one of the nicest introductions around.
• Programming GPUs with SYCL by Gordon Brown - a great introduction to the whys and hows of SYCL.
• SYCL Developer Guide by Codeplay - the current lack of learning learning material about SYCL is appalling but this terse developer guide makes the situation a little better.
• CppCon 2018: Parallel Programming with Modern C++ by Gordon Brown - a great overview of parallel programming and modern C++.
• Modern C++ for accelerators: a SYCL deep dive by Andrew Richards - an excellent slide deck to be printed and studied.
• 2019 EuroLLVM Developers’ Meeting: A. Savonichev (Intel) "SYCL compiler: zero-cost abstraction and type safety for heterogeneous computing" - a nice insight on how Intel is working on his own SYCL implementation for LLVM; mandatory for compiler nerds.
• SYCL 1.2.1 API Reference Card - print and hang it on the wall next to the Picasso you just bought at Sotheby's.
• SYCL Standard Specification - to be a proper standard, you need a proper spec. Not so standardese (a traightforward and educational read, actually) but definitely not a novel.
• GCC support for offloading to PTX via OpenACC
• Compiling CUDA with clang - LLVM 9 documentation
• CppCon 2016: CUDA is a low-level language by Justin Lebar
• gpucc: An Open-Source GPGPU Compiler
• John Lawson et al., Cross-Platform Performance Portability Using Highly Parametrized SYCL Kernels - a very nice article about parametrizing SYCL kernels to achieve good performance portability across architectures.