commit 6fb08c61c5c49f4e72a8c908d61222332996c8eb
author Lei Zhang <antiagainst@google.com> Thu Mar 10 07:30:46 2022 -0500
committer GitHub <noreply@github.com> Thu Mar 10 07:30:46 2022 -0500
tree 9f361bf8378eb3ec40d99e67ecf3f12f69310f18
parent 1be371788923c1aacb8ad34a4a4d546120c84ba1

[spirv] Fuse tensor.pad ops into their consumer ops  (#8049)

It's typical for convolutions to require padding. Right now we handle
such padding with two separate steps, first performing `vkCmdFillBuffer`
to initialize the target buffer and then launch a specific kernel to
copy over the source content. These two steps also means two additional
pipeline barriers to sequence the execution.

This commit instead enables fusion of tensor.pad ops into the consumer
ops directly to avoid generating the buffer fill and copy. This is
expected to improve performance on mobile GPUs.

After fusing the tensor.pad op, we generate the kernel to contain both
a fast and a slow path. The fast path are for inner tiles that do not
need padding; the slow path is for boundary tiles that do need padding.
Both paths are vectorized properly.