Hi community,
Does AutoScheduler support BYOC? From the paper, it seems it should support? [2006.06762] Ansor : Generating High-Performance Tensor Programs for Deep Learning
Thanks!
cc: @merrymercy
Why do you mean by “support BYOC”? I think auto-tuning performs at the operator-level while BYOC is a subgraph substitution strategy at graph level that offloads execution of a subgraph to vendor libraries.
I understand that the BYOC is a subgraph substitution strategy. However I still have the same question as @fake_robot: Since AutoTVM uses a template bases aproach to optimize operations is there any possibility to use this attempt for BYOC as well. Assuming we have an accelerator / vendor libary that allows addaptation or variable parallization of task. Would it be possible to put this flexibility into a template and optimize the BYOC function with AutoTVM? Is there any way to apply this ML/grid-search approach and use the optimization benefits of AutoTVM also with BYOC? Did anyone tried something like this or figured out that it is definitely not working?
Does your accelerator codegen fully compatible with TE/TIR? If you just have some knobs to tune, it should be a lot easier to implement a simply benchmarker in your codegen such as CUTLASS and CuDNN.
Thank you @comaniac for your reply. Could you please specify a little more in detail what you mean with if my accelerator codegen is fully compatible with TE/TIR. I started a couple of month ago working with TVM but I still try to get the complete picture of BYOC to then use it to implement an accelarator. Does your question imply that there might be a way to use AutoTVM in the BYOC codegenerator? And also thank you for the tips for the other benchmarkers. I might have a look in it as soon as I get a first prototype of BYOC properly running.
PS: @comaniac: Since I have seen that you reply to many questions in this forum, also regaring BYOC. Is this a topic that you might help us with as well? [BYOC] no such function in module - #3 by slai-nick