Ben Vanik
Ben Vanik
As a data point I've used sccache locally and it worked as expected for our cmake builds.
PTAL; fixed the leak and made it non-fatal (it'll just print a disassembly failed line).
` %1 = torch.operator "onnx.Concat"(%0#0, %0#1, %0#2, %0#3, %0#4, %0#5, %0#6, %0#7) {torch.onnx.axis = 3 : si64} : (!torch.vtensor, !torch.vtensor, !torch.vtensor, !torch.vtensor, !torch.vtensor, !torch.vtensor, !torch.vtensor, !torch.vtensor) -> !torch.vtensor` axis 3...
yeah we can't be polluting the source folder - that'll have to change. I'll take a look at the rest in the morning - a quick glance there's some style...
yeah, that path is wrong - no clue why `/lib/` is in there, but the `lib` prefix is something we need to drop - here's what we do elsewhere: ```...
yeah, that loop is going to be a problem (it's going to run entirely on the host in the VM interpreter) - tensorizing may help a bit but it really...
I'd start by looking at converting the ops if there are equivalents - it's absolutely required that those loops end up inside a dispatch region, and doing that with linalg...
(this looks like a great win but has gone stale - please reopen/rebase if it's still needed!)
The buffer usage flags have been reworked and `IREE_HAL_BUFFER_USAGE_SHARING_IMMUTABLE` is likely the thing to key off. We may also want to require `IREE_HAL_BUFFER_USAGE_SHARING_CONCURRENT`. The use case for this is wanting...