AdePT - Accelerated demonstrator of electromagnetic Particle Transport
Standalone GPU simulation prototype
Main goal:
- Understand the technical possibilities and limitations of GPU usage for full simulation and develop adapted solutions, using a prototype demonstrating a complete simulation workflow rather than testing only individual components. Make an assessment in a time scale of 6 month - 1 yr on the feasibility, development effort and performance expectations of a large scale GPU simulation project targeting EM shower simulation confined to GPU.
Prototype:
- e+/e-/gamma transport featuring track-level parallelism
- calorimeter setup (eg ATLAS EM calo)
- magnetic field
- VecGeom-based transport manager (in a first phase, possible replacement with Optiks-based engine added later on)
- a set of (simplified) physics processes allowing to simulate EM showers: pair production, brems, absorption at threshold in a first phase, more realistic later on
- simple data structure for scoring, transferred to host at the end of events
Technical objectives:
- develop a first implementation of a GPU-friendly transport engine, allowing dynamic scheduling of kernels and efficient state data management on device
- adapt/develop/optimize GPU-friendly transport components, starting with mock-up versions improved gradually, to allow developing in parallel the transport engine. Understand/adopt the design changes needed for gradually achieving the physics performance equivalent to CPU
- understand the constraints and hard limits for performance, data handling, memory management, single precision usage, kernel scheduling, finding/implementing solutions where possible
Repository: https://github.com/apt-sim/AdePT