Generate unnecessarily fancy profiling plots for CUDA kernels by parsing output from NVIDIA Nsight compute CLI.
Note: architectural limits are currently specified for NVIDIA V100 GPUs. Automatic parametrization of plots is not currently supported. You can adapt plots to your favorite GPU by editing values labeled as
Device parametersin.jinjafiles.
$ ncu $(./ncu-args.sh) myapp.x > myapp.ncu
Make sure the .ncu output file is in the root of the repository:
$ ls -1 *.ncu
myapp.ncu
You can generate a chart for each type by using proper make targets:
$ make myapp.roofline-fp.pdf
$ make myapp.roofline-inst.pdf
$ make myapp.roofline-shared.pdf
$ make myapp.hist-instmix.pdf
$ make myapp.hist-occupancy.pdf
$ make myapp.hist-predication.pdf
The previous commands generate charts considering all kernels in the profiling data.
This can be confusing when the ncu output comes from a real-world application with tens of kernels.
To select and plot only specific kernels:
$ make ... KERNELS="KernelA KernelB"
Only kernels whose declaration contains at least one of the strings in KERNELS will be included in plots.
Note:
gnuplotscripts are sometimes picky about thegnuplotrunning them; the currently tested version isgnuplot 6.0 patchlevel 2.
All roofline charts are inspired by:
N. Ding and S. Williams, “An Instruction Roofline Model for GPUs,” in 2019 IEEE/ACM Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems (PMBS), Denver, CO, USA: IEEE, Nov. 2019, pp. 7–18. doi: 10.1109/PMBS49563.2019.00007.
make target: <basename>.roofline-inst.pdf
make target: <basename>.roofline-fp.pdf
make target: <basename>.roofline-shared.pdf
make target: <basename>.hist-instmix.pdf
make target: <basename>.hist-occupancy.pdf
make target: <basename>.hist-predication.pdf
