tree 46aa0742523a2a5f9bba3ced736a774381cb24b2
parent 08b6af6b773558aaca14f7c8b783eb6d1f3622d0
author Quinn Dawkins <quinn.dawkins@gmail.com> 1764613830 -0500
committer GitHub <noreply@github.com> 1764613830 -0500
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJpLd7GCRC1aQ7uu5UhlAAAGhQQADinNTqehEt5PIgMNFnTQ4Hc
 TXY2IJHxbKzBhSaPm9O5MYpyop9gv+wpU7K3O1BGNJ1QcfVkWGO437o9yG/oncZ3
 dGy8EUMDAb8oS0zzEanoU5O5F7ugGae8UKVHML5xRM633/agxykSDJR8LWCaB8Yu
 BEjDjtdkvy2NvN8Wry0Y/lpw9RzwAwklFsKzLqgqHPtaM335R0Ar2nOZM5RKb6uK
 ihPoo/uW3aSWa/PEc1rch2op/P+w0MYF5YgYW0GVYmk7Y+ZXwyy7RYE4sfq+UU5y
 NI/3nea+KI6vJo4puRxfuw3owy+5Z0CdWvbV/gVj9TST2jADH6O9Kexo98pS+onB
 mpv6rsHDA4VZN78HmQ3lQa6+jPcwAbD+izFG+d1g2NRyiz2I9LJt30g1f/1zkcm+
 jUR8zKZkzEtl8xxXVK0j0Xxtli277kA8lOMmX8JK/R5w5qua58rLQg5eToWTzs9C
 liwvXU5K/+VQAjStp9VsyR8OaHAJTprzq/0942XIAyGabJHK27YEX4+L94DNWIhA
 rO26fwa6tq6LvT2RlFrUWV8Sgh+cFJmUE7NQrcg4XX8vK96nXWGofuDy9MyLe3wV
 uLJ12Gh8g9mUHhZ7sC9fyHsSU1+qEAfskWBS7l+nZOZiyC+6+QIfGtCmzt1LQzWa
 KJyYVjyXjqPQYWKDcnso
 =Mfnn
 -----END PGP SIGNATURE-----
 

[Codegen][GPU] Add fusion barrier after result promotion (#21709)

Currently result promotion relies on inserting two copies, one to
workgroup memory that is supposed to fuse with its producer, and one
with `derived_thread_config` that controls the distribution of the
consumer. This however has nothing forcing the intended fusion behavior,
and in all exercised cases today relies on reshapes missing fusion
patterns.

This patch introduces a fusion barrier op that sits between the two
copies to prevent fusion and gets dropped when bufferizing.