transformations: (lower-csl-stencil) Optimise full-stencil access

[n-io]

For full-stencil accesses followed by an immediate reduction of a uniform reduction function, use the catch-all getRecvBufDsd API call instead of the neighbour-by-neighbour getRecvBufDsdByNeighbor function.

Iff this optimisation is applicable, the accumulator dsd might look like this:

  const accumulatorDsd = @get_dsd(mem4d_dsd, .{
    .tensor_access = |i,j,k|{directionCount, pattern, chunkSize} -> accumulator[k]
  });

This is currently implemented by setting the strides of i and j to 0, which might require a printer update (tbc).

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.98%. Comparing base (4a02c8a) to head (367c52d).
Report is 17 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3271      +/-   ##
==========================================
+ Coverage   89.92%   89.98%   +0.05%     
==========================================
  Files         442      444       +2     
  Lines       55489    55688     +199     
  Branches     8656     5362    -3294     
==========================================
+ Hits        49899    50110     +211     
- Misses       4161     4172      +11     
+ Partials     1429     1406      -23     

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

[dk949]

How does this pass interact with the "Option 3" method that used to be emitted?

If the optimization is performed, does Option 3 still apply?

[n-io]

How does this pass interact with the "Option 3" method that used to be emitted?

If the optimization is performed, does Option 3 still apply?

For csl_stencil.apply ops on which the optimisation is applied, I have extended the docstring to clarify this.

For csl_stencil.apply ops on which the optimisation is not applied, there is now a filecheck showing the lowering to individual per-neighbour @get_dsd ops, which is the known and trusted default behaviour. Commenting out the optimisation in the PatternRewriteWalker will also trigger the default behaviour, in case this is helpful for testing or running.