allow main thread to exec instead of executor when calling clFinish
Explicit inorder dependency between commands in order to allow main thread to steal commands from the executor when main thread reach a clFinish call. It will reduce the latency of clFinish.
extract_cmds has a boolean to extract only non-batch commands. It will be very import when clvk will be able to use timeline semaphore to manage dependencies between commands
Do you prefer submitting yours, or should I update this one with your feedback?
I don't feel too strongly about this. You're likely to be faster than me so if landing this quickly(ish) matters to you, it'd probably be better for you to work on it. If not, I could take a crack at it.
Alright, I will submit an updated version of it in the coming days
no regression on the full cts on swiftshader and nvidia
no regression on the full cts with swiftshader and nvidia
As we discussed, let's focus first on merging #424. If you want a clearer view of this PR (without the diff of #424): https://github.com/rjodinchr/clvk/pull/3
I'll reopen one when #424 will be merged