GPGPU — General-purpose programming on GPU

48h — 6 CFU

en it

The course aims to introduce students to the use of graphic cards as high performance computing hardware (GPGPU: General-purpose Programming on GPU). Attendees are required to know at least one programming language, with a preference for C and/or C++.

The course will be held by Dr. Giuseppe Bilotta

Course outline

• C and C++ programming basics
• History of graphic cards and GPGPU
• GPGPU programming basics; introduction to CUDA and OpenCL
• Benchmarking, optimization and debugging
• GPU programming for embarrasingly parallel problems
• parallel scans and reductions

Calendar

Lessons will be held every monday and wednesday (5pm to 7pm) in Room 23. The Microsoft Team code for the course is 3icuged.

Lecture notes

• New lecture notes
• Lessons code

Laboratory

The course will have both theory (frontal) lessons and practice (lab) sessions. For the practice sessions, students may use their own laptops.

Students should have a C/C++ development environment in addition to the following software (see also the links below):

1. OpenCL SDK (runtime libraries and headers)
2. updated video drivers, including OpenCL support
3. CUDA toolkit, ROCm, or equivalent for their GPU (runtime library, include files and essential tools: compiler, debugger, profiler)

OpenCL support on Debian is distributed across several packages, to be installed depending on the available hardware:

• ocl-icd-opencl-dev for the OpenCL library and development headers;
• one or more ICD (Installable Client Drivers), such as:
  • beignet-opencl-icd for older Intel integrated GPUs
  • intel-opencl-icd for newer Intel integrated GPUs
  • nvidia-opencl-icd for NVIDIA GPUs
  • mesa-opencl-icd for AMD GPUs supported by MESA
  • rocm-opencl-icd for AMD GPUs supported by ROCm
  • pocl-opencl-icd for the CPU

Link

• hgpu.org — high-performance computing on graphics processing units
• OpenCL:
  • OpenCL (official website from Khronos Group)
  • Portable Computing Language, OpenCL implementation based on Clang and LLVM
  • for Intel hardware, the Intel® SDK for OpenCL™ Applications is available (may require registration; OpenCL-enaabled drivers (without development software) are available from here
  • ROCm, for AMD GPUs
  • NVIDIA OpenCL SDK Code Samples
  • Forum OpenCL — Khronos
  • Forum OpenCL — AMD
  • Forum OpenCL — Intel
• CUDA:
  • Getting started
  • CUDA downloads
  • List of CUDA-enabled GPUs
  • CUDA Toolkit Documentation
  • CUDA Developer zone

Affiliation

The course is offered by the Dipartimento di Matematica e Informatica of the Università di Catania in cooperation with the Osservatorio Etneo of the Istituto Nazionale di Geofisica e Vulcanologia (INGV).