nvbench

nvbench copied to clipboard

Custom CUPTI metrics

1

We would like to support additional CUPTI metric types. As I understand the code right now, it seems there is a fixed set of supported metrics. As noted in #91...

mmcloughlin

Implement and test NVRTC support

1

alliepiper

The markdown output can get very wide for benchmarks with a lot of parameters. Some ideas for saving horizontal space: - Drop redundant info - power of two expansions -...

alliepiper

Extend `nvbench::..._t` aliases to include CUDA specific types

1

E.g. `float16_t`, etc.

alliepiper

Use publicly visible CI runs

1

This would be a nice feature to have. Realistically, contributors will all use different OS and/or compiler configurations and versions. So what works for one contributor won't necessarily work for...

aaronmondal

When using `git log --graph`, one can see that PRs tend to be merged. This produces a git history with unnecessary branching. It also produces commits on the main branch...

aaronmondal

alliepiper

Change nvbench_compare statuses

2

Currently, when I compare a baseline with new result JSON using `nvbench_compare.py base.json new.json`, I get output like the following: | ValueType | Ref Time | Ref Noise | Cmp...

bernhardmgruber

NVBench uses a convention of `set_` and `get_` for interacting with object state, but the CUPTI metrics use a different convention of `collect_` and `is__collected`. The CUPTI methods should be...

alliepiper

alliepiper