Apple's Metal GPU profiler is excellent. And it works even if you're using it through WebGPU from C++ compiled outside of XCode.

Among other things, when you click on a draw command it shows you a thumbnail of the changes it made on the screen. It does this by saving and re-running the entire window render, highlighting the one command.

Does someone knows of a software that can estimate pipeline flushes, execution unit saturation, memory bandwidth and latency limits etc. for a given cpu?

Like: simulate a.out

And it would give me a report for each line of code and tell me everything that happened on that line. Was it a cache miss? False sharing? Etc.

I would pay a lot for that software.

Have you tried perf mem [0]?

[0] https://access.redhat.com/documentation/en-us/red_hat_enterp...

Not sire if what you’re looking for, but cachgind does some measures in that sense. It’s from the same toolbox as valgrind.

Alternatively, the perf tools can show some values from the performance counters.

If you’re on intel, there is some interesting tooling in that sense (can’t recall the name now, sorry!). I’d love to see similar tooling for arm :(

intel vtune for everything native