pinpinpoo commented on issue #20401: URL: https://github.com/apache/tvm/issues/20401#issuecomment-6185035196
@Asthenia0412 Thanks for looking into this and offering to work on it! 1. **The small-reduction serial path sounds reasonable.** I would treat 256 as an initial heuristic. In my follow-up sweep on a shared RTX A6000, SERIAL was still faster at reduction size 289 (32.90 us versus 403.01 us for ON). 2. **For this case, I would start with `GeneralReduction`.** My tracing shows that `Reduction` rejects the `Max` update because its matcher currently recognizes additive accumulation only; `GeneralReduction` then accepts the fused function. One detail from the trace: at the output-block split in the quoted snippet, the loop extents are `[144, 32, 32, 1, 1]`, so `loops[-1]` is an inserted unit loop. The actual producer reduction is split later. A serial-per-output path would therefore need to coordinate output spatial tiling and producer placement as well as the reduction split. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
