Update DFG::combineForIter()

[MeowMJ]

I find two bugs in combineForIter() when I am debugging combineForUnroll() (because they share key codes).

1. isSuccessor() is used incorrectly. succNode->isSuccessor(headNode) means to tell whether succNode is a successor of headNode. But we actually want to find whether headNode is a successor of succNode to guarantee it is a cycle path.
2. currentFunc doesn't recover its original position when we finished finding the target pattern under specific dfgNode.

[MeowMJ]

I finished the DFG::combineForUnroll(). One problem, I can't new a DFG node by my code. So I'm not sure if the result fits your idea.

conv (unroll twice) before combineForUnroll: (0)phi->(11)add->(20)add->(0)phi

conv (unroll twice) after combineForUnroll: (0)phi<->(11)add (11)add->(20)add

[tancheng]

I didn't get the fusion about 0, 11, 20 DFG nodes. Can we have a sync on this?

[MeowMJ]

I have finished this pull request. Below is an example.

conv.c unroll 2 times before combineForUnroll()

conv.c unroll 2 times after combineForUnroll()

[tancheng]

Can this currently be used as a parameter/argument set inside the param.json?

[tancheng]

And what about unrolling factor = 3?

[tancheng]

And can you also provide a fused DFG that both the unrolling and the i++ ctrl flow are fused?

[tancheng]

I didn't see my previous comments are addressed?

[MeowMJ]

Can this currently be used as a parameter/argument set inside the param.json?

No. Below is an example using it (in DFG.cpp).

    list<string>* toConbinePattern = new list<string>[5];
    toConbinePattern->push_back("phi");
    toConbinePattern->push_back("add");
    toConbinePattern->push_back("add");
    toConbinePattern->push_back("add");
    toConbinePattern->push_back("add");
    combineForUnroll(toConbinePattern);

[MeowMJ]

And what about unrolling factor = 3?

And what about unrolling factor = 3?

conv.c unroll = 3 before conbineForUnroll()

conv.c unroll = 3 after conbineForUnroll()

[MeowMJ]

And can you also provide a fused DFG that both the unrolling and the i++ ctrl flow are fused?

conv.c unroll=2 after fuse and combineForIter() and combineForUnroll()

[tancheng]

For the conv.c unroll = 3 before combineForUnroll(), are we able to fuse 3 nodes out of the 4? Say fusing 28/0/11 without touching 22? There will be still one cycle with length of 2 after your merging, but would relief HW complexity.

And do you know why the edge from 25 to 0 changed its color from blue (ctrl) to red (data)? Can this be fixed?

[MeowMJ]

This is conv.c unroll = 3 fuse 3 nodes

I am trying to figure out why the control edge becomes a data edge.

[MeowMJ]

For the conv.c unroll = 3 before combineForUnroll(), are we able to fuse 3 nodes out of the 4? Say fusing 28/0/11 without touching 22? There will be still one cycle with length of 2 after your merging, but would relief HW complexity.

And do you know why the edge from 25 to 0 changed its color from blue (ctrl) to red (data)? Can this be fixed?

If we want to keep the blue edge, we should guarantee that the nodes of the blue edge are PatternRoot. Otherwise, the blue edge will be replaced by a red edge. See https://github.com/tancheng/CGRA-Mapper/blob/master/src/DFG.cpp#L76

To keep the blue edge, we may need to modify the replaceDFGEdge().

conv.c Unroll = 3 choose phi-add-add-add to fuse

conv.c Unroll = 3 choose add-phi-add-add to fuse

[tancheng]

ha, so for now, one ctrl edge out of the two edges will become data edge. Then do you think you can fix this in this PR?

[MeowMJ]

I hope I can fix it this week in this PR.

------------------ Original ------------------ From: Cheng Tan @.> Date: Tue,Jun 11,2024 6:10 PM To: tancheng/CGRA-Mapper @.> Cc: MeowMJ @.>, Author @.> Subject: Re: [tancheng/CGRA-Mapper] Update DFG::combineForIter() (PR #21)

ha, so for now, one ctrl edge out of the two edges will become data edge. Then do you think you can fix this in this PR?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[tancheng]

Thanks!

[MeowMJ]

I changed replaceDFGEdge() to make sure that every edge being replaced can keep their ctrl property.

conv.c unroll=3 after combineForUnroll()