Special memories only work with PyCUDA backend

[isazi]

The functions:

• copy_shared_memory_args
• copy_constant_memory_args
• copy_texture_memory_args
  In core.py check if self.lang == "CUDA", and not work with e.g. CuPy or other backends that implement the functionality.

Moreover, this is only checked if using run_kernel and not tune_kernel. I am waiting for some comments before fixing it, as I may have overlooked something.

[benvanwerkhoven]

I do know that constant memory works with the CuPy backend. But now that the number of backends is expanding, a table with feature completeness per backend would be nice to have.

In the meantime, I was also inspecting the code and I think you're onto something interesting here. There is some very old code in core.py that I've quoted below. This indeed includes the if's for lang == "CUDA".

The code in compile_and_benchmark is however not running into any issues since it skips these high-level variants and immediately calls the copy_constant_memory_args etc on self.dev instead of self. Basically, I think that the functions on these lines:

kernel_tuner/kernel_tuner/core.py

Lines 497 to 516 in b3ff4cd

	def copy_shared_memory_args(self, smem_args):
	"""adds shared memory arguments to the most recently compiled module, if using CUDA"""
	if self.lang == "CUDA":
	self.dev.copy_shared_memory_args(smem_args)
	else:
	raise RuntimeError("Error cannot copy shared memory arguments when language is not CUDA")
	def copy_constant_memory_args(self, cmem_args):
	"""adds constant memory arguments to the most recently compiled module, if using CUDA"""
	if self.lang == "CUDA":
	self.dev.copy_constant_memory_args(cmem_args)
	else:
	raise RuntimeError("Error cannot copy constant memory arguments when language is not CUDA")
	def copy_texture_memory_args(self, texmem_args):
	"""adds texture memory arguments to the most recently compiled module, if using CUDA"""
	if self.lang == "CUDA":
	self.dev.copy_texture_memory_args(texmem_args)
	else:
	raise RuntimeError("Error cannot copy texture memory arguments when language is not CUDA")

are actually not used at all and could be removed.

[benvanwerkhoven]

I've just made some extensions to the documentation on backends see pull request #213

[benvanwerkhoven]

It would appear that run_kernel maybe does use these methods. Will look into this soon.

[isazi]

I will fix this tomorrow. And base my "patch" on the mega pull-request from @fjwillemsen (unless that is merged into master already today or tomorrow).

[fjwillemsen]

The plan is indeed that #214 will be merged today or tomorrow, directly after Stein's pull request #189.
I'll notify you by posting in this thread once they have been merged.

[isazi]

Perfect, then I will just fix this issue on master after your merge, and avoid the conflicts :)

[fjwillemsen]

Both #189 and #214 have been merged into master!

[isazi]

Great, I will do this today. Will reach to you if something is not clear about the new development environment.

[isazi]

This should be fixed for now.