[TKW] Move default scheduling params to common place

[Hardcode84]

Moved default scheduling params to get_default_scheduling_params().

Didn't updated gemm tests, as it uses VALU_UNITS: 8, SHUFFLE_UNITS: 8, unlike attention VALU_UNITS: 2, SHUFFLE_UNITS: 2.

[harsh-nod]

thanks for the cleanup! Was meaning to do this, but never got to it.

[raikonenfnu]

can't we just integrate into the backend directly? since we may use this area to override default variables

[raikonenfnu]

Would be good to add it here

iree-turbine/iree/turbine/kernel/wave/scheduling/resources.py

Lines 32 to 40 in e3b6c87

	def get_available_resources() -> list[int]:
	resources = [
	GLOBAL_MEMORY_UNITS,
	SHARED_MEMORY_UNITS,
	MMA_UNITS,
	VALU_UNITS,
	SHUFFLE_UNITS,
	]
	return np.array([int(subs_idxc(x)) for x in resources])

So instead of return np.array([int(subs_idxc(x)) for x in resources]) we can do something like:

default_resources =  get_default_scheduling_params()
return np.array([int(subs_idxc(x)) if x in subs_map else default_resources[x] for x in resources]) 

and similar method for the delays as well. I think if we can do this way we would not need to update hyperparameter for every kernel and it makes overwriting default values cleaner. :)

[Hardcode84]

Those values are not really 'default' as they can be different for different GPUs, it's more like get_default_device_scheduling_params() so I think it's better to set them explicitly.

[raikonenfnu]

'default' as they can be different for different GPUs

That is true, but this problem also exist in the current method introduced in this PR. Hence, we'd most likely need to add query for the gfx architecture in the get_default_device_scheduling_params() function.

so I think it's better to set them explicitly.

I think once we have querying gfx arch in get_default_device_scheduling_params then we can do this in resources.py as proposed in the root comment. We'd just have to check that the symbols for the resources or delays either exist in user specified context or get_default_device_scheduling_params(), if not, we will throw off some error saying that scheduling is not supported on the architecture by default and request user to specify values into context.

I think this is quite an important discussion, looking at it from kernel author's perspective, Wave programming would look quite a bit neater if we don't have to do get_default_device_scheduling_params() and manually set them in every kernel, it'd be just a matter of boolean turning scheduling on or off. :)