lhutton1 opened a new pull request #9560: URL: https://github.com/apache/tvm/pull/9560
Fixes an issue that causes strides to be incorrectly calculated when the number of channels in the input is less than 16 and involves a conversion from NHWC to NHCWB16. This is due to TVM being 'too smart' when analyzing generated TE and removing compute that is deemed unnecessary. Consequently, strides over data are incorrectly calculated leading to an output mismatch. The PR uses a reduce sum operation to trick TE's data dependency analyzer into looping over a whole block (16), rather than the number of channels actually used (< 16). This causes the calculated strides to be a multiple of 16 which is required for NHCWB16 format. cc @mbaret @ekalda @manupa-arm @NicolaLancellotti @dchauhan-arm -- 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]
