05:42fdobridge: <jekstrand> Only one way to find out! If they don't give me the numbers, I'm going to make a tool for auto-benchmarking and generating the table. We'll have half-decent numbers in this computer one way or another.
15:57fdobridge: <jekstrand> Well, I asked. Now we get to find out how open they really want to be. 😅 (It was hard to get those docs at Intel from inside the company!)
16:31HdkR: And then you hope someone throws the excel spreadsheeet over to you
16:57fdobridge: <karolherbst🐧🦀> I think Nvidia was open towards NDAs with specific contributors, so don't get disappointed if that's the endresult 😛
16:59fdobridge: <jekstrand> It's gonna end up in the code one way or another.
16:59fdobridge: <karolherbst🐧🦀> true
17:00fdobridge: <karolherbst🐧🦀> but...
17:00fdobridge: <jekstrand> It wouldn't be that hard to write a tiny little tool which passes SPIR-V to their driver, snags the pipeline cache, reads the stall info from instructions, and dumps out a table.
17:00fdobridge: <karolherbst🐧🦀> the ISA doc I've got is also under NDA although I'm free to use it for implementing stuff without restrictions....
17:00fdobridge: <karolherbst🐧🦀> sure, but that's not the point 😛
17:00fdobridge: <jekstrand> hehe
17:01fdobridge: <jekstrand> I'd like ISA docs at some point. Stuff like "here's what this instruction can do"
17:01fdobridge: <karolherbst🐧🦀> yeah
17:01fdobridge: <jekstrand> But that's not quite the same as a table that I'm literally going to copy+paste into the compiler.
17:01fdobridge: <karolherbst🐧🦀> that's exactly what the doc is telling
17:01fdobridge: <karolherbst🐧🦀> right..
17:01fdobridge: <jekstrand> But if they give it to me under NDA with the explicit (in writing!) expectation that I can copy+paste into the driver, that's probably ok.
17:02fdobridge: <jekstrand> And I'll likely R/E anyway just to make sure they aren't lying to me. 😅
17:02fdobridge: <karolherbst🐧🦀> yeah... so dunno. The ISA doc doesn't actually contain any "specifics" in the sense of "what bit do I have to touch" just what exists and what it's doing
17:03fdobridge: <karolherbst🐧🦀> it already specifics which instructions are variable latency tho
17:04fdobridge: <karolherbst🐧🦀> anyway.. it's 2 am here and I have to sleep 😄 good... morning or something 😄
17:18fdobridge: <jekstrand> hehe. Ok.
18:43fdobridge: <jekstrand> Contemplating how I want to handle predicates... For the actual "skip this instruction", I think I'm happy to make it a special case.
18:43fdobridge: <jekstrand> But there's like a half-dozen more ways to use them and I'm debating how to do that well.
21:30fdobridge: <jekstrand> @karolherbst🐧 Is there any particular reason why you made SM numbers in hex? I see nothing in the docs that indicates that NVIDIA thinks of them that way.