UCT/CUDA: treat stitched VA as managed memory

[Akshay-Venkatesh]

What?

Detect VA ranges composed of multiple physical allocations and treat as managed memory to force pipeline protocols.

Initial tests on nvlink connected GPUs show protocols being selected correctly:

[1738101119.349005] [52879:0]   | ucp_context_0 inter-node cfg#2 | rendezvous data send(fast-completion|multi) from cuda-managed/GPU0 to cuda    |
[1738101119.349007] [52879:0]   +--------------------------------+---------------------------------------------------------------+---------------+
[1738101119.349010] [52879:0]   |                        0..8176 | fragmented copy-in copy-out                                   | tcp/enP2s2f1  |
[1738101119.349011] [52879:0]   |                       8177..4M | cuda_copy, flushed write to remote, frag cuda                 | cuda_ipc/cuda |
[1738101119.349012] [52879:0]   |                   4194305..inf | pipeline cuda_copy, write to remote, frag cuda                | cuda_ipc/cuda |
[1738101119.349013] [52879:0]   +--------------------------------+---------------------------------------------------------------+---------------+

[Akshay-Venkatesh]

@SeyedMir Can you please test these on x86 platforms and see if expected protocols are being chosen?

[brminich]

maybe goto out_default_range instead?
also some rephactoring could be done, e.g. to have a trace for all cases including the new one

[SeyedMir]

Is cuMemGetAddressRange support for VMM allocations documented somewhere? The driver API docs mention legacy allocators only.
Returns the base address in *pbase and size in *psize of the allocation by cuMemAlloc() or cuMemAllocPitch().

[Akshay-Venkatesh]

Indeed not well documented. The behavior we've seen is that it is able to provide base pointer and size corresponding to base address and length of mapped address space from the corresponding cuMemMap.

[yosefe]

Suggested change

	if ((uintptr_t)base_address + alloc_length < (uintptr_t)address + length) {
	if (UCS_PTR_BYTE_OFFSET(base_address, alloc_length) <
	UCS_PTR_BYTE_OFFSET(address, length)) {

[yosefe]

Some test failures seem relevant