You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fix projection iterator assignment for Conv_input dimensions
The bug caused "Multiple constraints on the same Conv_input projection"
errors when using convolutions with use_padding=true.
Root cause: When processing `Iterated (Var v)` equations, fresh iterators
were immediately assigned to variables not yet in v_env. This happened
before other equations could establish that the variable should get its
index from a Conv_input affine expression instead.
Fix:
1. Defer `Iterated (Var v)` processing: collect such variables and process
them after all equations are handled, when their projections are known
2. Track Conv_input target projections and exclude them from early iterator
creation in product_dim processing
3. Create fresh iterators for remaining product dimensions only after
p_conv_input processing completes
This ensures projections that should get affine indices from Conv_input
don't conflict with prematurely assigned iterators.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
0 commit comments