Choose timezone
Your profile timezone:
Alice Weekly Meeting: Software for Hardware Accelerators
→
Europe/Zurich
ALICE GPU Meeting
-
-
Speaker: Christian Sonnabend (CERN)
Neural network seeding vector (https://github.com/AliceO2Group/AliceO2/pull/15756)
- Found a bug in the implementation (missing setting during application, seeding vectors were never used) -> Fixed
- Results don't change much. 0-100% centrality is mostly unchanged, but one improvement can be seen in 0-5% centrality enforced simulation
Is it worth the complications? 🤷♂️
ROOT: TMVA SOFIE inference of .onnx models on the EPNs
- Discussion during ROOT Q3 meeting last week: ALICE has great interest to give SOFIE a try. Meeting with SOFIE devs (Lorenzo and Sanjban) was very productive
- One night of synergetic work between me and ChatGPT Codex: Huston, we have liftoff!
- SOFIE PR: https://github.com/ML4EP/SOFIE/pull/63
- O2 PR: https://github.com/AliceO2Group/AliceO2/pull/15913
- alidist PR: https://github.com/alisw/alidist/pull/6369
This brings...
- Support for cuda_ / hip_stream
- GPU code generation and on-the-fly compilation of .onnx models. Inflight-CCDB-loaded models are supported
- Support for FP16 and FP32
- Support for dense layers and ReLU operations
- Support for memory arena pointer (provided by reco framework as scratch allocation)
- Direct interface with SOFIE's generalised kernel generation code
Benchmarks (ONNX to be redone with new software. Old benchmark from over several months ago).
- SOFIE
- ONNX
10% improvement in TF reconstruction time on single GPU benchmark! -> probably 2-3x faster NN inference by SOFIE
AOB
- Brevitas presentation
- Slides attached, please have a look
- Presentation is next Monday at 14:30, 30 mins including questions
- New physics accurate event display rendering with blender and video support
- Testing kubernetes setup with Giada on test nodes
- O2 analysis tutorial
- Will now continue with Sandro's simulation project on the track pruning
-
Speakers: Felix Schlepper (CERN, Heidelberg University (DE)), Gabriele Cimador (CERN, Università and INFN Torino), Matteo Concas (CERN)
Gabriele: Working on optimzing vertexer kernels
E.g.: fitPeaksKernels
Before: one thread per peak
- Each GPU thread handles one vertex candidate (peak)
- Collects the lines inside the peak's z-window that are time-compatible with the peak line
- Fits a seed vertex: the point closest to all those lines (least squares)
- Keeps only the lines passing within
pairCutof the seed - Refits the kept lines → final vertex; applies the cuts (min. contributors, distance to beam)
- Computes the vertex quality: mean residual components and mean squared distance of the lines to the vertex
- Problem: a busy peak (thousands of lines) is walked three times by a single thread, and the whole kernel waits for it
After: one warp per peak
- Each warp handles one peak
- The window's lines are split across the lanes: lane i takes lines i, i+64, i+128, …
- Each lane adds its lines into a partial sum of the fit equations (9 numbers)
- A warp reduction (register shuffles: no shared memory, no synchronisation) adds up the partial sums, so every lane ends up with the full total
- Every lane solves the same 3×3 system → the seed vertex is in every lane's registers
- Lines are reselected around the seed: again partial sums, reduce, solve → final vertex
- Residuals and distances to the vertex: partial sums per lane, reduce, divide by the number of lines
- Lane 0 writes the result
- The work per peak is split 64 ways, so busy peaks no longer hold up the whole kernel
Next:
- Optimise other kernels
- Tune
-