tree 0125c2bc8e6f37af7ce55ebc9cdcdc8036c4d1a9
parent 53dc6609adbdfc709b679a4f4ea5fb7f8eb17c6c
author Vivian Zhang <zhyuhang88@gmail.com> 1775673897 -0700
committer GitHub <noreply@github.com> 1775673897 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJp1qIpCRC1aQ7uu5UhlAAAJYAQAC58++quv2w3/+uuZm/wBU7h
 NxPQNlCjRtB2FwHLjZq0h1xSKKL+gEjKNrXZEuPnrtxYyJ0pcWXfvY0JrTiF4UAC
 S8x3Pf5wecjq+rlUjMvm7T0eByxKEjZ9Ey+uRCZwScFhBJCKm8QoLzX/lDvWPdWA
 Q7iSvQpGEX1Gi1JogoSYKpKeHKXIv4wyQgAXWBoPkeeKIORFcQFP+WkODC35czDj
 qQ1vhJ8hvLHCvWT9fA+/PH9qNlk6iLPSnQPBtAlkOaB3f+x67j/56BOcWhpgXxU9
 +hYyL3h+WwS1kJIEbHH0KxvxRgADyxL+V8NQJA86KxCOq7KZ770KTgI6zSozjpiw
 SpBkjq6N6/3nS9YhpdhcfpPWqijl35uHHrJqdQXg7CZxlVBkicqvyDpm4ClbAeKC
 dxUaI2J4cWnxkcLuLHKl18Jtjq77v957EDcKbZI0HYQe8Yl90yxGSJmJNAAVXYEk
 j6KfdOSoY0dmXITDXx9wCLC7Ktejn+VsTjDP32Bt8bHwbmDP0MVMKWTFG0nUTAq0
 DnLq9lc5SRSuZr6BZKxpm9e9mSyKIqtFCC2bxNtoIZwjR37p4nThz5ebjrMOlz6N
 I7BmpUklZNLjz20JDPU6ARMmFuuZ7X2BDcNkykEiGozX1UgrdGjRWg8VDfodYN3h
 XE8dFe9AmD9kuXNgEjXI
 =oU4i
 -----END PGP SIGNATURE-----
 

[Preprocessing] Add pass to convert strided insert_slice to linalg.generic (#23990)

Add `iree-preprocessing-convert-strided-insert-slice-to-generic` pass
that converts strided `tensor.insert_slice` into zero-constant
destinations to a `linalg.generic` with index arithmetic. This targets
backward data convolution patterns where the upstream gradient is
scattered into a zero buffer at strided positions.

The replacement generic computes the strided scatter in a single
dispatch, replacing the `Memset + slow_memcpy` pair. Power-of-2 strides
use bitwise ops instead of expensive integer division. The result is
selected via `arith.select` for branchless GPU execution.

This PR is a partial implementation for
https://github.com/iree-org/iree/issues/23976.

---------

Signed-off-by: yzhang93 <zhyuhang88@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>