tqchen commented on PR #20423: URL: https://github.com/apache/tvm/pull/20423#issuecomment-5815793710
A parsing-performance data point for the Blackwell TIRx `tcgen05` CTA-group-1 GEMM kernel: at 7505c9a88d43b97931cfbb254b0b42cefa09ca44, median CPU parsing and native IR construction time with source spans enabled improves from **23.59 ms to 15.34 ms** — **1.54× faster**, or approximately **35% less time**. | Parser | Median time | | --- | ---: | | Original parser (`5b4b7535`) | 23.59 ms | | New parser (`7505c9a8`) | 15.34 ms | This uses the same 57-line kernel source, with 63 timed samples per parser across three fresh processes each, excluding warmups. The resulting IR was checked for structural equality against the original parser output, and spans were enabled in both variants. This measures CPU parsing/IR construction for this kernel. GPU compilation and execution are outside the measurement; the result does not establish a workload-wide speedup. -- 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]
