Feedback on a new dev tool we're building to improve GPU programmer productivity

[adirajagopal]

Hey @vladotomov, it was great meeting you at the AI conference last week and thanks for the suggestion to post regarding the tool we're building at (https://www.ncompass.tech) here to see if it could be of use to the GPU kernel engineers on the team.

For any GPU kernel engineers in the community, we're building tools to make it easier and faster to profile, benchmark and build GPU kernels. We personally found the nsys + ncu + cuda-gdb flow a bit clunky and slow to dev with, which is the problem we're trying to solve. Ideas we're looking into are around:

• Lightweight profiling that doesn't generate massive trace files
• Being able to customize traces without having to heavily instrument any codebase
• Getting visualization of compiler graphs and multi-threaded code executions to aid debugging
• Building a trace to custom implementation UI that makes the turnaround time from profiling to getting performance results for a new implementation quickly.

I'm posting here to see if anyone in the chat would be open to having a quick call so we can better understand your workflow and identify the main pain points that we could help with solving. Your input here would directly shape our tool, so we'd love to give you access to the Alpha for free when we get there.

Thanks!