guberti opened a new issue, #13528: URL: https://github.com/apache/tvm/issues/13528
Once #13242 lands, the convolution schedules for microTVM will have a tunable parameter `num_outputs`: https://github.com/apache/tvm/blob/f11243a0eec0bae9a8f9b4a6d4ad6152de4e3fc9/python/tvm/topi/arm_cpu/qnn.py#L235-L239 As the comments states, picking this value is important for performance. It would be awesome to be able to autotune this - the correct value is very dependent on the exact parameters of the convolution, picking correctly will have a >10% impact on performance, and predicting it without autotuning would be challenging (though theoretically possible). Note that this value is used in the _compute function_, not in the _scheduling function_, which makes autotuning harder. -- 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]
