Tantalus13A98B5F commented on pull request #8605: URL: https://github.com/apache/tvm/pull/8605#issuecomment-1039458148
Hi @mengbingrock Not sure how long have you been going, but `bsr_conv2d.convert2` is merely for graph & param transformation, and it assumes well-padded shapes (0 for 1x1 and 1 for 3x3). Maybe you will want to tweak the implementation of this. `spconv2d_3x3_nhwc` is the target-specialized implementation for the operator. In TVM, for every operator, we will want to [register it with Relay](https://tvm.apache.org/docs/dev/how_to/relay_add_op.html) so that we can use the operator in a graph. For the case of sparse operators, we would also like to provide transformations like `bsr_conv2d.convert2` to generate sparse graphs out of dense graphs. If we want our implementation to be visible for users, we have to fix all the code along the way. For early-stage testing purposes, it's not necessary to do all the stuff at once. You can directly pass ndarrays to a scheduled template; if you have finished the Relay integration you can then manually compose a sparse graph for testing; once you've done the former steps you can consider tweaking the convert function. -- 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]
