[TVM][CUDA] NVIDIA GPU Int8 Support

[nishi-t]

This PR adds int8 support for nvidia gpus

[tqchen]

@vinx13 since you are working on related stuff, can you take a look?

[vinx13]

@tqchen do we need to support other lanes size? e.g. int4 if lanes == 16

[tqchen]

This, this would actually be very helpful, @vinx13 can you elaborate what are possible things we need to support and value translation rules in here?

[vinx13]

the rule here will be:

lanes() == 8 => int2 (aligned by 8)
lanes() == 16 => int4 (aligned by 16)

[tqchen]

@nishi-t can you act on @vinx13 's comment? being able to perform full vector load will be very helpful to get the full perf.

Specifically, we want to be able to load 4 words from memory at a time, unfortunately there is no char16 struct so we have two solutions:

• Use int2 and int4 for these vectors, while not support arithmatics for these as @vinx13 suggested.
• Try to deifne a char16 and support arithmetics

int2 and int4 might be a easier path for now if we only need save/load and get nvidia's native support

[nishi-t]

@vinx13 Thank you for the suggestion. Could you please review again?
@tqchen For now, I addressed the comment by first solution. And, I'll try second solution seperately in another PR, because it may take a little while. Is this okay?

[tqchen]

Sounds good, can you add test case to cover the load, possibly by a shared memory load vectorize?

[vinx13]

looks good from my side

[nishi-t]

@tqchen Yes, I'll address it.
@vinx13 Thank you.

[nishi-t]

@tqchen I addressed. Please review again.

[tqchen]

Thanks @nishi-t @vinx13 , this is now merged