Refactor use_padding from a global shape inference setting to a field of the `Conv_input` constructor

[lukstafi]

Good first step on the road to padding inference.

[lukstafi]

@claude-opus please refactor use_padding from a global shape inference setting to a field of the Conv_input constructor. Localize the choices based on use_padding to now use the field of the corresponding Conv_input -- if there is no Conv_input to deconstruct in the context of a use-case, point out such places in your report.

[claude[bot]]

Claude finished @lukstafi's task —— View job • claude/issue-386-20250918-1051 • Create PR ➔

---

Refactoring use_padding to Conv_input field ✅

Completed Tasks:

• [x] Search for current use_padding implementation
• [x] Locate Conv_input constructor definition
• [x] Add use_padding field to Conv_input in both dim and proj types
• [x] Update all Conv_input pattern matches in row.ml for both types
• [x] Update all Conv_input constructor calls to include use_padding
• [x] Build and verify changes compile successfully
• [x] Commit and push changes

Summary of Changes

Successfully refactored use_padding from a global shape inference setting to a field of the Conv_input constructor. The changes include:

1. Added use_padding: bool field to both:

   • The public dim type Conv_input constructor in row.ml and row.mli
   • The internal proj type Conv_input constructor in row.ml

2. Updated all pattern matches and constructor calls to include the use_padding field across row.ml

3. Kept the global use_padding variable for now because:

   • shape.ml creates Conv_input during einsum spec parsing
   • The use_padding value comes from the conv2d function parameter
   • Parsing happens independently of the conv2d call
   • Current workaround: nn_blocks.ml sets the global before einsum operations

Places without Conv_input context

The main location where we cannot directly use the Conv_input field is in shape.ml lines 420 and 1798 where Conv_input is constructed during einsum specification parsing. These locations currently use !Row.use_padding to get the value from the global reference.

The hack in nn_blocks.ml line 175 ([%oc Ocannl_tensor.Row.use_padding := use_padding]) is still necessary to pass the use_padding parameter to the shape inference system.

Next Steps

To fully eliminate the global variable would require one of:

1. Extending einsum specification syntax to include padding information
2. Adding a new API to configure padding at the operation level
3. Refactoring the shape inference system to thread padding context through

For now, the refactoring successfully localizes use_padding to the Conv_input constructor, making the dependency explicit even though the global is still needed as a communication mechanism.

---

|

[lukstafi]

This is fixed now.