transform: (stencil-tensorize-z-dimension) Tensorize arith.constant directly

[n-io]

Add support for directly tensorising arith.constant, e.g.

%0 = arith.constant dense<1.666600e-01> : tensor<510xf32>

The more generic way of tensorising any scalar operand (independent of its source) would be linalg.filling the value a newly created empty tensor. This still works as a generic way. However, in the special case that the scalar operand is a constant, this is much shorter.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.84%. Comparing base (09e3157) to head (781293e).
Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2970      +/-   ##
==========================================
- Coverage   89.86%   89.84%   -0.02%     
==========================================
  Files         409      409              
  Lines       51085    51167      +82     
  Branches     7923     7938      +15     
==========================================
+ Hits        45908    45972      +64     
- Misses       3925     3932       +7     
- Partials     1252     1263      +11     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[n-io]

This PR depends on #2969, re-run test when this is merged.

[superlopuh]

Code changes look good, but I'm not sure about the pass in general, and the test file. It seems like there's quite a lot of inspection and rewriting of operands, when the usual approach is to match directly on the uses. This matters a bit for composability of rewrites, and rewriting some of this tensorization as transformations on the uses would make it easier to read and to test. For example, the linalg.fill where the operand is a constant could be replaced with a new arith.constant of a tensor, leaving the original constant in place (and letting it be dce-ed away in a later stage in the pipeline). Similarly for the arith operations.

[n-io]

Code changes look good, but I'm not sure about the pass in general, and the test file. It seems like there's quite a lot of inspection and rewriting of operands, when the usual approach is to match directly on the uses. This matters a bit for composability of rewrites, and rewriting some of this tensorization as transformations on the uses would make it easier to read and to test. For example, the linalg.fill where the operand is a constant could be replaced with a new arith.constant of a tensor, leaving the original constant in place (and letting it be dce-ed away in a later stage in the pipeline). Similarly for the arith operations.

Thanks, done as discussed :)

[PapyChacal]

Nice!