[TVM Discuss] [Development/RFC] [RFC] Implement "add_to" semantic in TVM

[Bjarke Roune via TVM Discuss <no...@discuss.tvm.ai>]

In the first entry in this thread, it says as option 1 (if I'm understanding it correctly), to have Relay/TVM be able to see an op like A = Add(B, C), where this is the last use of e.g. B, and then decide to place B and A into the same memory buffer and have the implementation of Add be such that it still works when B and A point to the same memory. My question on that is whether this is a new feature that would have to be implemented or whether that is already something that TVM can and will do? (this is how it works in XLA, and it has been working fine there, so this is definitely a viable approach, though pushing this kind of optimization to its limits can get complicated in the face of nested loops, reuse of input and output parameter memory, other control flow etc.)





---
[Visit Topic](http://tracking.discuss.tvm.ai/tracking/click?d=1gm_cre9B1NJEXcmlvQA2fugS2q-fLz0Krx-70KtvuCbuU9dCeFez9en796SV6zJAhdbT0_7tZVQjbpW9MLfT3wo8xtSlbpvVxLd6_cPJ_nPIIke00X0YQc-98ey-X0Mp40WfN7snawEQV_Ts3wSIQ_zTjvgOJ7mbrvawFMSldvv7k-S54lQ5SCBaj5gcdRNnA2) to respond.

You are receiving this because you enabled mailing list mode.

To unsubscribe from these emails, [click here](http://tracking.discuss.tvm.ai/tracking/click?d=7cFgOaAA4XIBVlVKt_oyC07uihTjg4Q6cjeBRNRTiPqStckdoaKSEcoXRBR0kFI-q_p1iKJvmgG70xnrghJe9HWxQg80AbAT7qRPXPtOwsZFvZLWlp4QYc3QDCesW6ckZ6N_REnmj6u4CAKzsapEg_K0pg0kPVHTG9AtN3M43DESTP9OE1aA3b_jX8tdLc11tnKa2W2LK35kJ4HOa9YSX_x6kcVW0CnS6cbEPZ26-wiB0).

Tianqi Chen, UW, Seattle, WA, 98105, United States
http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=GivkeDIAGX328xmPiZuscA2