1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
|
- It is intended that benchmark_plot.sh be used to drive benchmark.py and plot.py
- It is recommended you experiment with commands like
./benchmark_plot.sh -t my_tag --level1 true --level2 false --level3 false --benchmark true --plot false
./benchmark_plot.sh -t my_tag --level1 true --level2 false --level3 false --benchmark false --plot true
./benchmark_plot.sh -t my_tag --level1 true --level2 false --level3 false --benchmark false --plot true --theo_max true
look in directory blas1/my_tag for output .csv files and plots
- On a machine where you only want to run benchmarks, call benchmark_plot.sh with arguments
--benchmark true --plot false
- There are .yaml files for each function that specify input argument sizes
- rocblas-bench is called, and output will be in .csv files
- If you are in your fork (not ROCmSoftwarePlatform) the .csv files can be checked in to github
on a remote machine, and checked out for plotting on a local machine
- There is plotting of gflops vs size or gflops/theoretical_max_gflops vs size
- plots are in .png files
- For gfx90a here are checked in .csv result files and .png plot files
- to get the gfx90a .cvs and .png files the following thre calls were made to benchmark_plot.sh
./benchmark_plot.sh -t gfx90a --level1 true --level2 true --level3 true --benchmark true --plot false
./benchmark_plot.sh -t gfx90a --level1 true --level2 true --level3 true --benchmark false --plot true
./benchmark_plot.sh -t gfx90a --level1 true --level2 true --level3 true --benchmark false --plot true --theo_max true
|
