Julian Samaroo

Integrate Dagger's event logging for timeline profiling

1

This would reduce the need for using `rocprof` or the Profile stdlib to observe kernel execution ordering and latency hiding efficiency.

...because it loads `libhsa-runtime64.so`, not `libhsa-runtime64.so.1`.

The current implementation has multiple flaws: - Resizing operations on `Array` are not thread-safe - wait-to-mark exhibits TOCTTOU races We need a solution for this that doesn't involve taking the...

Use refcounting for memory management

1

The current approach of escaping kernel inputs during kernel execution, and having finalizers directly free HSA memory allocations, is problematic when considering the potential benefits of https://github.com/JuliaLang/julia/pull/44056. We could instead...

When the GPU is under high load and spinning on `AMDGPU.device_signal_wait`, `hsa_executable_freeze` can hang as it tries to synchronize with the GPU. We should switch to using an `InterruptSignal` (exposed...

This shouldn't be necessary, but let's have CI confirm that for us.

Currently, a single thread serves hostcalls, but it should be possible to spawn multiple threads to service the same hostcall (especially once #50 is merged) concurrently. This would be useful...