[CUDA] [Codegen] Ensuring atleast one thread block to handle empty tensor

[anijain2305]

Topk was failing on CUDA when k is a var and its value is 0 at runtime. At closer inspection I found that there are 0 thread blocks at runtime. This PR ensures that there is atleast 1 thread block.

[anijain2305]

@kevinthesun @masahi @mbrookhart @zhiics @trevor-m Please review.

[masahi]

hmm, I think I've already added a fix for such cases, here:

tvm/python/tvm/tir/ir_builder.py

Lines 205 to 206 in 82942fb

	if attr_key == "thread_extent":
	value = op.max(1, value)

Do you know why it is not working? cc @mbrookhart

[anijain2305]

hmm, I think I've already added a fix for such cases, here:

tvm/python/tvm/tir/ir_builder.py

Lines 205 to 206 in 82942fb

	if attr_key == "thread_extent":
	value = op.max(1, value)

Do you know why it is not working? cc @mbrookhart

Is this because this the lines that you suggested are specific to IR Builder, while the failure that I see is for injective schedule? My failures was coming for an injective schedule.

[mbrookhart]

Yeah, I think this change catches it at a lower level. We might not need the ir_builder change after this.

[kevinthesun]

LGTM

[mbrookhart]

LGTM

[masahi]

I see, thanks.