cuda-qr

Make sure that G++10 is installed

sudo apt install g++-10

Inference

To test the QRx and QRX kernels directly, run:

cd individual-benchmarks
nvcc -arch=sm_80 -ccbin=g++-10 -std=c++20 benchmark.cu reference_kernels.cu -o benchmark -lcublas -lcurand -lcusolver
./benchmark --verbose   # --verbose is optional

To run project locally

cd src
./build_local.sh
cd build
../run.sh

QR kernels source code

The QR kernels sorurce code can be found in /final_project_devctr/src/qr_kernels.cu and is designed after https://onlinelibrary.wiley.com/doi/full/10.1002/cpe.13010. The base implementation is copied from https://github.com/evelyne-ringoot/Avocado-sandwich/tree/main/KAbasedSVD/src and the reference implementation is cusolver cusolverDnDgeqrf.

Docker build - Ubuntu

To test the docker build (based on https://github.com/accelerated-computing-class/final_project_devctr): Execute once:

./devtool build_devctr

Execute every change into src files:

./devtool build_project
py <path_to_telerun.py> submit build.tar

This will build a docker environment according to the /devctr/Dockerfile, build a tar file according to /src/build.sh and execute the /src/run.sh on telerun.

Windows build and run

On windows, make sure to install CUDA toolkit 12.4.1, Visual Studio with C++ integration and then run the following code:

 powershell -executionpolicy bypass -File .\buildandrun.ps1