dialects: (csl) Data structure descriptors

[n-io]

A basic implementation of dsd types and the @get_dsd. Custom array index expressions are not yet supported, and the current version should be understood (and printed) as accessing A[i,j,k] according to the number of dimensions and generate induction vars accordingly.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.62%. Comparing base (6a1643d) to head (7c30803).
Report is 19 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2648      +/-   ##
==========================================
+ Coverage   89.60%   89.62%   +0.01%     
==========================================
  Files         360      364       +4     
  Lines       46179    46658     +479     
  Branches     6978     7057      +79     
==========================================
+ Hits        41378    41816     +438     
- Misses       3726     3748      +22     
- Partials     1075     1094      +19     

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

[dk949]

Would fabin_dsd and fabout_dsd be part of the same type and op, or should they have their own?

[n-io]

Would fabin_dsd and fabout_dsd be part of the same type and op, or should they have their own?

Since they need different fields specified, making it a separate op is actually a viable option.

[dk949]

Would fabin_dsd and fabout_dsd be part of the same type and op, or should they have their own?

Since they need different fields specified, making it a separate op is actually a viable option.

In that case it would make sense to rename this op (and class) something like csl.get_mem_dsd.

I think it makes sense for both ops to still return a csl.dsd, since all the maths builtins accept either a mem dsd or a fab dsd.

EDIT: by both ops I mean this and the hypothetical future csl.get_fab_dsd.

[AntonLydike]

If this is all that is needed to get DSDs supported in xDSL then that's amazing! I'm afraid that I don't know much about the internals of DSDs though, so for that @dk949 is probably the better conversation partner!

From my side this looks good!

[AntonLydike]

Two questions here:

1. Can we not just read the sizes from the memref type?
2. Can we have two sizes passed to the get_dsd if the underlying array is 1D?

[n-io]

1. What if you don't want to iterate over the entire buffer?
2. It will fail to verify.

[n-io]

Would fabin_dsd and fabout_dsd be part of the same type and op, or should they have their own?

Since they need different fields specified, making it a separate op is actually a viable option.

In that case it would make sense to rename this op (and class) something like csl.get_mem_dsd.

I think it makes sense for both ops to still return a csl.dsd, since all the maths builtins accept either a mem dsd or a fab dsd.

EDIT: by both ops I mean this and the hypothetical future csl.get_fab_dsd.

I've added fabin_dsd and fabout_dsd into the same get_dsd op. It ended up with a lot of the fields being optional and needs verification that all (+only) the right fields are populated.

[n-io]

GetDsdOp is now an abstract base class and there are subclasses for csl.get_mem_dsd and csl.get_fab_dsd. Overall, this looks a lot cleaner.