Fix compile time explosion for minkowski distance

[ahendriksen]

As explained in #1246 (comment), ptxas chokes on the minkowski distance when VecLen==4 and IdxT==uint32_t.

This PR removes the veclen == 4 specialization for the minkowski distance.

Follow up to: #1239

[codecov-commenter]

Codecov Report

Base: 87.99% // Head: 87.99% // No change to project coverage 👍

Coverage data is based on head (12dd68a) compared to base (bd0d86b).
Patch coverage: 100.00% of modified lines in pull request are covered.

Additional details and impacted files

@@              Coverage Diff              @@
##           branch-23.04    #1254   +/-   ##
=============================================
  Coverage         87.99%   87.99%           
=============================================
  Files                21       21           
  Lines               483      483           
=============================================
  Hits                425      425           
  Misses               58       58           

Impacted Files	Coverage Δ
python/pylibraft/pylibraft/cluster/__init__.py	100.00% <100.00%> (ø)

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[cjnolet]

Not sure how I missed this one originally but I'd like to keep the (modern) naming conventions more consistent across the APIs.

Also, I want to make sure the seemingly lower compile times aren't the result of sccache (which is suspect they are) and giving us false hope here. This is what happened before the release (we saw 45 minute compile times because of sccache, got excited, and then realized after the sccsche reset that the compile times were still taking >6 hours. I'd like to do some more profiling of this before merging.

[ahendriksen]

Also, I want to make sure the seemingly lower compile times aren't the result of sccache

Did you reset sccache in commit 585fc47 or do we still have to test this?

[cjnolet]

I tried to but that didn't seem to work. I'm going to build locally without sccache in the meantime since my workstation specs are similar to CI.

[cjnolet]

Here's the ninja trace log:

The trace above is from building RAFT on my local machine with the --allgpuarch flag.

Looks like the high offenders are now (in no particular order):

1. linkage.cu googletest
2. ivfpq_search runtime wrapper (this is weird because it means it's not properly using the specialization)
3. All of the ivf-flat googletests
4. All of the ivf-flat benchmarks
5. eigen_solvers googletest
6. refine specializations (expected)
7. ball_cover_knn_query.cu specialization
8. canberra specialization (expected.. but we should still figure this out)
9. fused_l2_knn googletest

I notice a lot of the source files that take a long time to compile are taking in upwards of 1.5 hours or more.

[cjnolet]

I'm going ahead and approving this guy because I think the build issues related to these changes have been fixed. The other build issues predate these changes, I believe.

[cjnolet]

/merge

[cjnolet]

@ahendriksen just leaving a note for later that we should pull the masked_l2_1nn_params out from masked_nn.cuh and put it in a file called mask_nn_types.hpp. This allows users to expose that through their own APIs (knowing it doesn't require nvcc). We can do this in a follow-on.

[sean-frye]

LGTM