# Roofline Analysis of the Standalone Application

## Build on Cori

The building process is detailed in https://github.com/hep-cce2/PPSwork/blob/master/Patatrack/StandaloneApplication.md


## Collect Roofline Data

Profiling 100 events for Roofline analysis:

```bash
module load cuda
module load esslurm
salloc -A m2845 -C gpu --nodes 1 --gres=gpu:1 --exclusive -t 04:00:00

# Copy profiling script to the working directory
cp profiling.sh /WORKING/DIRECTORY/pixeltrack-standalone
cd /WORKING/DIRECTORY/pixeltrack-standalone

# Source environment
source env.sh

# Run Nsight profiling
./profiling.sh
```
The output data file `patatrack_roofline.nsight-cuprof-report` can be found at `/global/cfs/cdirs/m2845/roofline/patatrack`

### Convert Nsight output to csv

```bash
module load cuda

cd /global/cfs/cdirs/m2845/roofline/patatrack
nv-nsight-cu-cli --csv -i patatrack_roofline.nsight-cuprof-report > patatrack_roofline.csv
```


## Roofline Analysis

1. [ ] `process_nsight.ipynb`: process raw profiling data, compute FLOPS & AI for Roofline plot
2. [ ]  `roofline_plot.ipynb`: plot L1, L2 and HBM Roofline for individual kernels & Hierarchical Roofline for the whole application
