You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@tvm.apache.org by aca88 via TVM Discuss <no...@discuss.tvm.ai> on 2020/04/06 08:38:49 UTC
[TVM Discuss] [Development/RFC] [Relay] Improved graph partitioning
algorithm
Hi,
Thanks for the nice animation and the work put into this.
I was wondering if the subgraph partitions for the same compiler (at this level) have any repercussions for downstream passes.
So for example assume that the last node of A1(called a1) is a Conv2d and the first of A3 (called a3) is an elemwise add.
Further assume that the "blue" compiler has a 'very fast' implementation of a Add(Tensor,Conv2d(Tensor)). This means that actually moving a1 into the A3 subgraph is a better option.
I guess this is similar to the previous concerns regarding the "optimal" subgraph partitioning, but without 'many' compilers. Since there is still only the "red" and "blue"
---
[Visit Topic](https://discuss.tvm.ai/t/relay-improved-graph-partitioning-algorithm/5830/17) to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/e6b3ee18897be3fb50f5f32fff0df65a2f909e318644c01c340353ede0c40974).
[TVM Discuss] [Development/RFC] [Relay] Improved graph partitioning
algorithm
Posted by Luke Hutton via TVM Discuss <no...@discuss.tvm.ai>.
Yes that's correct :slight_smile:
---
[Visit Topic](https://discuss.tvm.ai/t/relay-improved-graph-partitioning-algorithm/5830/20) to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/d7474dc4a85d8604bb3fad7f1ea1e637f800b52707c13cabe978a29b94517896).
[TVM Discuss] [Development/RFC] [Relay] Improved graph partitioning
algorithm
Posted by aca88 via TVM Discuss <no...@discuss.tvm.ai>.
>I believe your example would be taken care of using the Merge Composite pass before partitioning. You can imagine that after running this pass add+conv2d for the blue compiler would be represented by a single node.
Ok then I wasn't really sure of what passes come before and after. So you are saying that Merge Composite would create a singe "blue" node (out of a1 and a3) which has a two incoming edges:
1. from the previous "blue" parent of a1
2. from B2
?
---
[Visit Topic](https://discuss.tvm.ai/t/relay-improved-graph-partitioning-algorithm/5830/19) to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/8794df4bdc938f117988dcf749b4c8f01eda105c2e72b032122ca93e8fd218f6).
[TVM Discuss] [Development/RFC] [Relay] Improved graph partitioning
algorithm
Posted by Luke Hutton via TVM Discuss <no...@discuss.tvm.ai>.
Hi, @aca88 I believe your example would be taken care of using the Merge Composite pass before partitioning. You can imagine that after running this pass add+conv2d for the blue compiler would be represented by a single node. The partitioning would then happen as you described. However, this would force this result, there is no consideration for the 'optimal' partitioning for the time being.
Currently there is ongoing work in the graph partitioning area. I don't believe the new partitioning mechanism has support for Composite functions just yet, although this is something being looked into. @matt-arm @manupa-arm may be able to comment more.
---
[Visit Topic](https://discuss.tvm.ai/t/relay-improved-graph-partitioning-algorithm/5830/18) to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/1614dc2692e11cbc5cf1a5a83b200c86caa5aec29ad3261fee4313eff0d0706c).