Skip to content

cern-nextgen/wp1.7-soa-wrapper

Repository files navigation

wp1.7-soa-wrapper

Wrapper class with common interface for different data layouts (SoA and AoS).

Requirements

  • A machine with an nvidia GPU and the necessary drivers.
  • Retrieve any further dependencies from this script.

Build and Run

git clone https://github.com/cern-nextgen/wp1.7-soa-wrapper.git
cd wp1.7-soa-wrapper
cmake -B build
cmake --build build
ctest --test-dir build --output-on-failure

Container

The following container contains the nvidia container toolkit and the dependencies to build the unit tests.

registry.cern.ch/ngt-wp1.7/wp1.7-soa-wrapper:latest

The corresponding Dockerfile can be found here wp1.7-soa-wrapper-image. Note that you still need to run the container on a machine with nvidia GPU. Once you are within the container, you can follow the steps of the section Build and Run. Pull the above image and run it with one of the following commands.

Docker

sudo docker run -it --rm --runtime=nvidia --gpus all wp1.7-soa-wrapper:latest bash

Podman

podman run -it --rm --security-opt=label=disable --device nvidia.com/gpu=all wp1.7-soa-wrapper:latest bash

Kubernetes Cluster of NGT

Follow the process described here, but in Step 6, use the following session.yaml instead.

apiVersion: v1
kind: Pod
metadata:
  name: soa-wrapper
  labels:
    mount-eos: "true"
    mount-cvmfs: "true"
spec:
  containers:
  - name: soa-wrapper
    image: registry.cern.ch/ngt-wp1.7/wp1.7-soa-wrapper:latest
    command: ["sleep", "infinity"]
    resources:
      limits:
        nvidia.com/gpu: 1
    securityContext:
      runAsUser: 0
      runAsGroup: 0
    volumeMounts:
    - name: home-volume
      mountPath: /root
  volumes:
  - name: home-volume
    persistentVolumeClaim:
      claimName: soa-wrapper-pvc
  nodeSelector:
    nvidia.com/gpu.product: NVIDIA-H100-NVL

---
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: soa-wrapper-pvc
spec:
  accessModes:
    - ReadWriteOnce
  resources:
    requests:
      storage: 1Gi

In Step 9, instead of running nvidia-smi, execute the commands in the section Build and Run.

TODO

  • Default initializations
  • Choose between uninitialized, default initialized, and custom initialized arrays
  • Allow putting default values in skeleton struct definition, e.g. by template spcialization
  • Aggregate initialization for arrays
  • Don't define closures twice
  • Return classical reference to struct in AoS case
  • Allow classical reference when passing to a function in AoS case

About

Wrapper class to interface with different data layouts (SoA and AoS).

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •