fnhirwa commented on PR #19763:
URL: https://github.com/apache/tvm/pull/19763#issuecomment-4721994733

   Thanks for the feedback.
   
   **Double lowering fix**: Removed the `complex64 -> float32[..., 2]` 
transformation from `_input_type()`. The lowering now happens exactly once in 
`from_tflite()` after user overrides are merged. Verified that `shape_dict` and 
`dtype_dict` overrides no longer produce double-packed shapes.
   
   **Constant tensor support**: Added `COMPLEX64` to 
`get_tensor_type_as_numpy`, `get_tensor_type_str`, and `get_tensor_expr`. 
Constant complex tensors are now reinterpreted as `float32[..., 2]` by reading 
the buffer as `np.complex64` and calling `.view(np.float32).reshape(shape + 
(2,))`.
   
   **RFFT2D**: Removed from this PR entirely. An O(N²) matmul decomposition 
using existing Relax ops (`matmul`, `permute_dims`, `concat`) is feasible as an 
interim implementation and will be contributed in a follow-up PR. The correct 
long-term fix is a native `relax.op.signal.rfft2d` with a C++ registered 
backend, tracked in #19764.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to