rutkoor commented on code in PR #15679:
URL: https://github.com/apache/tvm/pull/15679#discussion_r1324286476
##########
python/tvm/relax/transform/legalize_ops/manipulate.py:
##########
@@ -205,3 +213,16 @@ def te_layout_transform(data, name):
output_dtype = call_args[0].struct_info.dtype
output_sinfo = [TensorStructInfo(output_shape, output_dtype)]
return call_tir(gvar, call_args, output_sinfo, tir_vars)
+
+
+@register_legalize("relax.remove_pad")
+def _remove_pad(bb: BlockBuilder, call: Call) -> Expr:
+ orig_shape = call.attrs.orig_shape
+
+ def te_remove_pad(data):
+ """
+ Returns a new compute that restrict the original expression to the
shape of orig_shape
+ """
+ return te.compute(orig_shape, data, name="te_remove_pad")
Review Comment:
Hi @masahi, currently the elements are padded in the end of the original
buffer. So, when we iterate over orig_shape(dimensions of non-padded buffer),
we can be sure that we will be iterating over non-padded entries.
The one assumption that I have taken here is that padding will always be
happening in the end of the buffer.
In future, if AlterOp introduces padding in-between original buffer then we
should modify remove_pad operator to reflect the same.
--
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]
