[Topi] Tensorcore support for Conv3D

[jwfromm]

This PR is a pretty direct port of the conv2d tensorcore schedules introduced in PR #5099 to conv3d. In my early testing I've found this new schedule to be up to 10X faster than the default conv3d schedule. I also snuck one little adjustment into the conv3d winograd schedule that helps for smaller workloads. Given that tensorcore support is currently only for NHWC and NDHWC layouts, I've also added support for converting to these layouts using the ConvertLayout pass.

[jwfromm]

@Shawn-Inspur, @Laurawly could you take a look at this PR?

[tqchen]

also cc @icemelon9 since it is related to stragey

[jwfromm]

Anyone know what the deal with the CI failure is? Seems to work fine on my local branch and has nothing to do with this PR.

Running './tests/scripts/task_sphinx_precheck.sh' inside tvmai/ci-gpu:v0.61...
mesg: ttyname failed: Inappropriate ioctl for device
Adding group `tvm' (GID 1000) ...
Done.
PreCheck sphinx doc generation WARNINGS..
rm -rf _build/*
rm -rf gen_modules
rm -rf tutorials
rm -rf vta/tutorials
python3 -m sphinx -b html -d _build/doctrees . _build/html
Running Sphinx v2.4.1
Makefile:70: recipe for target 'html' failed
script returned exit code 2

[tqchen]

The sphinx error is a known flaky case that we should look into, please push to retrigger

[FrozenGene]

@yangjunpro@minminsun Please also have a look, because you have done tensorcore support ever.

[icemelon]

LGTM. just some minor comments

[yangjunpro]

Nice work.
Just two small cents.
The schedule looks a little bit complicated.
Have your considered enriching the Auto TensorCore codegen pass to automate the TensorCore schedule generation process?
The previous work of Auto TensorCore codegen only covers GEMM and it does require additional work to add conv support. However with the addition of conv support into Auto TensorCore codegen pass, it may make the convolution tensorcore optimization more generic.

[jwfromm]

@yangjunpro, funnily enough I asked nearly the same question to the authors of the conv2d tensorcore schedule. You can read their answer here. The quick take is that it does not support conv (like you mentioned) but more importantly that it causes significant performance regression compared to a bespoke approach like this one.

[Laurawly]

LGTM

[tqchen]

Thanks @jwfromm@Laurawly @Shawn-Inspur @yangjunpro @icemelon9 !