karolherbst OP , vor 24 Tagen

turns out, I've broken support for 32 bit architectures... again 🙃 oh well.. with that fixed it shouldn't look too bad actually.

fclc , vor 24 Tagen

@karolherbst I saw the second post without the RPI context and assumed x86, with a partially written “who even cares about 32 bit x86 anymore” ?

karolherbst OP , vor 24 Tagen

@fclc yeah... but the rpi is also cursed. The CPU side is all 64 bit, but not the GPU

Though with like 8GiB or RAM it also doesn't really matter all that much.

fclc , vor 24 Tagen

@karolherbst any better on the Pi5?

karolherbst OP , vor 24 Tagen

@fclc Not quite sure, but I don't think so.

karolherbst OP , vor 24 Tagen

sooo, with my own runner I have this result:

Pass 2333 Fails 114 Crashes 56

Though images are disabled for now due to driver internal reasons.

And most fails seem to be due to arithmetic precision requirements or something like that.

Also profiling doesn't properly work, but... I think at this point it's probably good enough to merge.

karolherbst OP , vor 24 Tagen

okay.. so there are a couple of interesting bugs left:

somehow mapped buffers don't get properly synchronized... I wonder if the barrier thing I have in place just doesn't work with the driver properly yet

profiling not working, which causes like a loot of fails

and a bunch of other fails, but if mapping doesn't properly work, then random fails might be expected anyway.

karolherbst OP , vor 23 Tagen

I think I've tracked down the last big issue 🙃

The driver replaced the backing buffer object of a host mapped memory object causing random faults and inconsistency issues.

Sadly the way mapping memory in OpenCL works is pretty annoying. I fear I'll have to rework those bits in #rusticl as well, because I'm way to optimistic on some assumptions. And I'll also have to fix contexts with multiple devices while at it.

karolherbst OP , vor 23 Tagen

"Pass 2419 Fails 71 Crashes 13" guess I'm getting close.

jhwgh1968 , vor 23 Tagen

@karolherbst 🥳

karolherbst OP , vor 23 Tagen

half the fails is really just v3d not supporting PIPE_CAP_QUERY_TIMESTAMP yet, which #rusticl relies on for even profiling.

Everything else is just random stuff, like not supporting unaligned load/stores yet, some arithmetic precision fails and random other stuff