Torrance Hodgson
Torrance Hodgson
Please see latest commits. I've centralised all `ROCQueue` cleanup in `kill_queue!()`, and the finalizer for ROCQueue now simply calls out to `kill_queue()`. This ensures cleanup happens correctly either when ROCQueue...
An additional note: the tests fail with the same segfault when using the artifacts.
And as an aside, on Julia 1.7 and using `Pkg.add("AMDGPU")`, building without artifacts, and then running tests fails with `error: Invalid value (Producer: 'LLVM14.0.0git' Reader: 'LLVM 12.0.1jl')`. If I use...
I have downgraded to rocm-5.1.3, with the same results.
With Julia master, I no longer get the segfaults, however numerous test failures, most of which seem to be issues with the change in atomics in 1.9 (`ConcurrencyViolationError("setfield!: atomic field...
To port this into AMDGPU, we need to know the used scalar and vector general purpose registers (SGPR, VGPR) for a given kernel. These values seem to be provided by...
My current hack, btw, is to pass in `vector` and cast to `vector` in C+, which can't be the right way to do things: **C++** ``` mod.method("tableCommand", [](std::string command, std::vector...
I have found that I can create a `StdVectorAllocated{ConstCxxPtr{Table}}` object without providing an initial vector, and instead pushing to it, like so: ``` tables = StdVector{ConstCxxPtr{Table}}() push!(tables, Ref(ConstCxxPtr(table1))) [...] tableCommand("my...
Those snippets are copyto!(), not copy!(), right?
Ok thank you. Should the documentation be updated to reflect this?