[OpenCL] Skip opencl memory operations on empty tensors

[mbrookhart]

cc @masahi @tmoreau89

[masahi]

Since this is not a opencl specific problem, ideally we should think about how to handle this case in all backends. For now I think we should have the same change for cuda, rocm, and vk as well.

[mbrookhart]

I don't think cuda has this issue, I this cudamalloc does this under the hood, as does standard C malloc. I agree, though, it looks like we need it for rocm and vulkan.

[tmoreau89]

Can the runtime attempt to free a buffer that was allocated that is of size 0? Could that trigger a segfault?

[mbrookhart]

Feasibly. I didn't see such an issue with opencl, but indeed memory freeing doesn't pass in size, so I can't do the same check:

tvm/src/runtime/opencl/opencl_device_api.cc

Lines 129 to 136 in e697f03

	void OpenCLWorkspace::FreeDataSpace(TVMContext ctx, void* ptr) {
	// We have to make sure that the memory object is not in the command queue
	// for some OpenCL platforms.
	OPENCL_CALL(clFinish(this->GetQueue(ctx)));
	cl_mem mptr = static_cast<cl_mem>(ptr);
	OPENCL_CALL(clReleaseMemObject(mptr));
	}

[mbrookhart]

But it would be easy to check for null before freeing

[masahi]

Maybe we can implement all book keeping stuff in DeviceAPI base class? After we make sure size is valid, we forward calls to derived class.

[mbrookhart]

It's messy, but we could write a default method that checks for these things and then calls a helper function. We then mark the default function as final and change ALL of the device classes to override the helper functions.

[mbrookhart]

Closing in favor of #7691