Distinguish PHI_START and PHI_CONST

[Jackcuii]

Hi Sir Tan @tancheng ~ I fully manually verified the whole compiled gemv kernel and figure out the problem of PHI_START and PHI_CONST mentioned before.

First we define the two clearly.

PHI_START: in first run, it chooses src0 and does not consume src1. In others, it consumes both and chooses according to the predicates. predicates can be both F (output value is arbitary) or one T one F (normal). Both T is not allowed.

PHI_CONST: in first run, it chooses src0 and does not consume src1. In others, it consumes both and always choose src1.

1. It seems by now all the generated PHI_X are PHI_START in literal.
2. Some of the PHI_X can be PHI_CONST (eg. the leftmost node on the graph below), but according to my observation, these positions could also be PHI_START, since src0's predicates are always guaranteed by GRANT_ONCE. I am quite not sure about that,but I did not think of a counter example.

FYI, I paste a verified fully-annotated DFG of gemv here. It shows all the nodes' function in this kernel. They are manually verifed.

Thanks @ShangkunLi for sync with me~

[tancheng]

@Jackcuii what are the action items? Anything wrong with the mapper? @ShangkunLi can help fix?

[ShangkunLi]

@Jackcuii what are the action items? Anything wrong with the mapper? @ShangkunLi can help fix?

Sure, I can help. @Jackcuii could you plz elaborate the problem/bug more clearly? I think currently we only introduce phi and phi_start in our neura dialect.

We do not introduce the phi_const here.

[Jackcuii]

@Jackcuii what are the action items? Anything wrong with the mapper? @ShangkunLi can help fix?

Sure, I can help. @Jackcuii could you plz elaborate the problem/bug more clearly? I think currently we only introduce phi and phi_start in our neura dialect.

We do not introduce the phi_const here.

Cool! So we don't have PHI_CONST now, then there is no need to modify the compiler😃. But is VectorCGRA also like this? I think VectorCGRA used to have PHI_CONST only🤔

[ShangkunLi]

@Jackcuii what are the action items? Anything wrong with the mapper? @ShangkunLi can help fix?

Sure, I can help. @Jackcuii could you plz elaborate the problem/bug more clearly? I think currently we only introduce phi and phi_start in our neura dialect.
We do not introduce the phi_const here.

Cool! So we don't have PHI_CONST now, then there is no need to modify the compiler😃. But is VectorCGRA also like this? I think VectorCGRA used to have PHI_CONST only🤔

The phi_const in VectorCGRA has similar functionality as phi_start; the difference between them is that:

• phi_const: in the first run, it chooses the src0 from the tile const queue
• phi_start: in the first run, it chooses the src0 from the function unit input