masahi commented on PR #14440: URL: https://github.com/apache/tvm/pull/14440#issuecomment-1491209909
Hey I'm happy to report that I have pattern-based rewriter for dataflow block working. Compared to the existing call node based matching & rewriting that requires a common post-dominator in a pattern, it lets us match a tree structure and replace leaf nodes or branches with new expression. This can be immediately used for combining *any* number of multiple matmuls sharing the same LHS into one matmul. In Relay, we have `CombineParallelDense` pass for that purpose, but we can achieve the same thing via graph (tree) matching and rewrite. For example, in SD UNet we have many three parallel matmul for QKV projections. In addition, there are also highly non-obvious parallel matmuls consisting of 32 or 22 of them. Those patterns can all be matched and rewritten via the following generic pattern and rewriter. I'm very excited about this! https://github.com/masahi/web-stable-diffusion/blob/unet-opt/test.py#L37-L68 -- 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: commits-unsubscr...@tvm.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
