tkonolige opened a new pull request #7974: URL: https://github.com/apache/tvm/pull/7974
This PR makes the autoscheduler sparse matrix registration code more specific in regards to input shapes. Instead of using density (as a two digit float) to discriminate between different inputs, we use all the input shapes. This causes fewer collisions between specifier strings. However, there are still two sparse matrices in the `deploy_sparse` tutorial that have the same specifier string. Ideally we would be able to use the values of the sparse matrix indices to discriminate between these two (for example, by hashing the indices), but the way registration is current setup makes this impossible. Instead we just ignore conflicts and hope that this doesn't cause a performance regression. -- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org
