On Monday, 21 December 2020 at 03:27:14 UTC, welkam wrote:
On Saturday, 19 December 2020 at 05:08:56 UTC, Max Haughton
wrote:
This will soon support Linux's perf_event so you will be able
to measure cache misses (and all the other thousands of pmc's
intel expose), use LBR msrs etc.
Are you going to read stdout from calling perf or are you going
to read perf.data file?
I already have code that uses perf_event_open directly, it's much
more efficient at the expense of having to decipher shit linux
documentation. I tried to get the bindings into druntime but they
weren't up to snuff apparently.
The architecture of the counters is designed around a stateless
widget giving you a stateful gadget specifically because this
makes using perf_event_open less of a minefield.
You can also enable userspace-rdpmc, which I will support
eventually (but you get a general protection fault if something
goes wrong so it's hard to debug).
