Feedback Same keyword in different Subgraph doesn't seem to be recognized

Hi, I assume this is known may even be fixed in SG 10.x, but for SG 7.x (2019.4 LTS):

- it seems even if we use the exact same keyword in 2 different subgraph, the shader variant still explode as if they are different keywords.


I can certainly up the shader variant limit, but I am at 65536 (2^16):

- with 5 boolean keywords in 2 subgraphs.
- and 2 enum keywords each with 5 entries.
- my main graph now says I am over variant limits.
- but if I put 2 of the same subgraph, then everything is fine.

But in either case, our main graph really only need 5 boolean + 2 * 5 enums in total, which should be below variant limit.


I guess I don't fully understand know keyword in subgraph works with main graph:

- I would hope it does nothing unless the main graph defines the keyword.
- which means the variant limit should only count the main graph (we need to copy them to main graph anyway).

Evidently, it wasn't the case?

 

The shader graph also become terribly slow to operate, by default, an empty HDRP Lit node has:

- 114 variants if you skip unused shader_features, and 99.72K variants if you don't.

With just the keywords added, without ANY connection at all, the shader graph becomes:

- 114 variants if you skip unused shader_features, but now 159.54M variants if you don't.

With all the connection linked:

- 435 variants if you skip unused shader_features, but now 319.08M variants if you don't.

I don't necessarily know how SG compile and strip its shader_features, but I have a strong feeling the time increase (from a few seconds to almost a minute), the compile time increase is closer to an exponential increase than a linear increase.

I know SG has its own continuous integration tests, does it have performance benchmark on 7.x??

 

To prove how slow it is, here is a pretty simply graph:




Here is how long it took to compile a single connection change on HDRP master node (around 20 seconds before unity ui unfreeze.)



Saving asset takes even longer:



We REALLY need a disable auto-compile option, even if we can't get any compile speed improvement in SG 7.x