sunggg opened a new pull request, #15491: URL: https://github.com/apache/tvm/pull/15491
Currently, we have two paths for BYOC. Path1: `seq = [FuseOpsByPattern(patterns, annotate_codegen=False), MergeCompositeFunctions(), RunCodegen()]` Path2: `seq = [FuseOpsByPattern(patterns, annotate_codegen=True), RunCodegen()]` Path1 is used for inference engines such as TensorRT, DNNL, CoreML, etc. which are more efficient with larger subgraph and thus better off merge adjacent composite functions. In constrast, Path1 is used for kernel libraries, such as Cutlass, Cublas, and not need to merge. For better UX, this PR aims to make both paths have same interface with various BYOC codegens. In addition, it removes unnecessary `kPrimitive` attribute in the function that will be offloaded to external codegen. One remaining caveat is that two paths would have different function names of each composite function and thus structural equality check would fail between two paths. Instead, this PR compares each path to their expected forms which have the same interface except function names. cc. @masahi -- 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]
