"Hands on" project on GPU and vectorization: basis on CUDA and TensorFlow

30 sept. 2020, 14:00 → 15 oct. 2020, 18:00 Europe/Paris

Description of the project: Vectorised calculations and heterogeneous computing architectures receive a lot of attention in HEP. The aim of this project is to provide the introduction to massively parallel calculation and show how they can be useful in various stages of HEP data analysis. The project consists of two parts: 

• Introduction to CUDA programming (Dorothea): Basic principles of the GPU architecture and the programming language CUDA will be introduced. The participants can gain hands-on experience with several exercises implementing for example vector addition and matrix multiplication to be processed on a GPU.
• Introduction to TensorFlow framework and its applications in physics analysis (Anton). The tutorials will include exercises with both pure TensorFlow v2 (using python interface) and with the TensorFlowAnalysis package that provides HEP extension. The example applications we will deal with are toy MC generation, maximum likelihood fits and amplitude analyses. 

Preliminary requirements: Knowledge of C++ (CUDA) and Python (TensorFlow). Familiarity with numpy would be an advantage. 

• Hardware: 
  • CUDA introduction: access to machine with NVidia GPU
  • TensorFlow introduction: GPU access is not necessary, but would be an advantage. 
• Software: 
  • CUDA introduction: CUDA installation on the machine with Nvidia GPU
  • TensorFlow introduction: TensorFlow v2, sympy, uproot, iminuit. The configuration file for Conda package manager will be available at the dedicated GitHub page.  

Project coordinators:  Anton Poluektov (anton.poluektov@cern.ch), Dorothea Vom Bruch (dorothea.vom.bruch@cern.ch)

 

If you would like to participate in the CUDA hands-on project and do not have access to a server with CUDA installation and Nvidia GPU, please get in touch with Dorothea.

 

Mattermost channel for discussions: 

https://mattermost.web.cern.ch/gdr-inf/channels/gpu-vectorisation-hands-on-project

ZOOM connection details: 

https://cern.zoom.us/j/95974540487?pwd=VVY3OVBjY2ZNdmZlcHEwMGlGVStYZz09

Meeting ID: 959 7454 0487
Passcode: 023379
One tap mobile
+41432107042,,95974540487#,,,,,,0#,,023379# Switzerland
+41432107108,,95974540487#,,,,,,0#,,023379# Switzerland

Dial by your location
        +41 43 210 70 42 Switzerland
        +41 43 210 71 08 Switzerland
        +41 31 528 09 88 Switzerland
        +33 1 7037 9729 France
        +33 1 8699 5831 France
        +33 1 7037 2246 France
Meeting ID: 959 7454 0487
Passcode: 023379
Find your local number: https://cern.zoom.us/u/awXtof2cs

Join by SIP
95974540487@188.184.89.188

Join by H.323
188.184.89.188
Meeting ID: 959 7454 0487
Passcode: 023379

 

 

lun. 12 octobre

1 Introduction to the GPU-architecture and CUDA

Introductory lecture, followed by a description of the hands-on problems

Orateur: Dorothea vom Bruch

CUDA exercises

GDR_CUDA_exercises_vom_Bruch_October2020.pdf

GDR_CUDA_introduction_vom_Bruch_October2020.pdf

Video recording (both presentations).mp4

2 Introduction to TensorFlow and its use as a compute engine for HEP

Introductory lecture and presentation of hands-on project

Orateur: Anton Poluektov (Aix Marseille Univ, CNRS/IN2P3, CPPM, Marseille, France)

TensorFlow_Presentation.pdf

Tutorials in GitHub

mer. 14 octobre

3 Informal discussion round on TensorFlow projects

Orateur: Anton Poluektov (Aix Marseille Univ, CNRS/IN2P3, CPPM, Marseille, France)

4 Informal discussion round on CUDA projects

Orateur: Dorothea vom Bruch