Skip to content

Conversation

@ftynse
Copy link
Contributor

@ftynse ftynse commented Nov 28, 2025

needs to be augmented with tests and sliced properly, but feel free to review with that in mind

@ftynse
Copy link
Contributor Author

ftynse commented Nov 28, 2025

The analysis result do match for matmuls, but sympy is unable to prove equality between:

$WG0*BLOCK_M + Piecewise((Mod($T0, 16), ~$MMA_ACC), (4*floor((Mod($T0, 64))/16), $MMA_ACC)) and

Piecewise(($WG0*BLOCK_M + Mod($T0, 16), ~$MMA_ACC), ($WG0*BLOCK_M + 4*floor((Mod($T0, 64))/16), $MMA_ACC))

even with simplification. Short of finding the way to extract the common component from the piecewise generally, we may need some tweaking in reconstruction to automate the tests.

@ftynse ftynse force-pushed the users/ftynse/python-connection branch from 65e948c to 1e1a4de Compare December 1, 2025 10:13
This mirrors the index symbol attribute and lets us differentiate induction
variables used in indexing expressions from regular symbols with the same name
without resorting to magic naming schemes in the C++ code. Since we are reusing
affine expression parser that expects a bare keyword, we still have to resort
to the magic naming scheme prefixing with `_Iter` in the textual format of the
IR. Harden the verifier to complain about regular symbol names starting with an
underscore and make them reserved.

Signed-off-by: Alex Zinenko <[email protected]>
Signed-off-by: Alex Zinenko <[email protected]>
Signed-off-by: Alex Zinenko <[email protected]>
Signed-off-by: Alex Zinenko <[email protected]>
@ftynse ftynse force-pushed the users/ftynse/python-connection branch from 1e1a4de to 6a6279d Compare December 8, 2025 17:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants