[Speculation] Automated buffering #629
Conversation
There was a problem hiding this comment.
Oh I didn't realize the merge of #552
before this latencies of beta-backend units were fixed, and speculation tests are benchmarked with those old latencies
| if (TailEn = '1') and (HeadEn = '0') then | ||
| -- if new tail index will reach head index | ||
| if ((Tail +1) mod {fifo_depth} = Head) then | ||
| if ((Tail + 2) mod {fifo_depth} = Head) then |
There was a problem hiding this comment.
This was the ring-buffer thing: we cannot fill up all slots to distinguish the empty and full state.
Full is latched so we should compare using Tail + 2, instead of + 1.
There was a problem hiding this comment.
do you mean buffered? A latched value is usually accidental https://vhdlwhiz.com/why-latches-are-bad/
| static inline bool canGoThroughOutsideBlocks(Operation *op) { | ||
| return isa<handshake::ForkOp, handshake::ExtUIOp, handshake::ExtSIOp, | ||
| handshake::TruncIOp>(op); | ||
| handshake::TruncIOp, handshake::BufferOp>(op); |
There was a problem hiding this comment.
to call isBackedge after buffering, we should include BufferOp
| let summary = ""; | ||
| let description = [{ | ||
| }]; |
There was a problem hiding this comment.
I'll add these once this design is approved
| for (auto branch : funcOp.getOps<handshake::SpeculatingBranchOp>()) { | ||
| if (branch->getAttr("specv1_branchDiscardCondNonSpec")) { | ||
| unsigned bb = getLogicBB(specOp1).value(); | ||
| ConditionalBranchOp controlBranch = findControlBranch(funcOp, bb); | ||
| if (controlBranch == nullptr) { | ||
| specOp1->emitError() | ||
| << "Could not find backedge within speculation bb.\n"; | ||
| return signalPassFailure(); | ||
| } | ||
| auto conditionOperand = controlBranch.getConditionOperand(); | ||
| branch->setOperand(0, conditionOperand); | ||
| branch->setOperand(1, conditionOperand); | ||
| } | ||
| } |
There was a problem hiding this comment.
The operands of this SpeculatingBranch didn't follow this figure actually.
The loop condition is not necessarily produced by the speculator.
It only becomes definite after placing speculator and save-commit units (because the loop condition may be generated by either of them), so I'm doing this at the end of the pass.
|
Thanks for this Shun! Could you put this into the google test integration tests instead of bringing the python script back? Could you also give an explanation of why we can't just run buffering after the previous spec pass? what is the problem with the save commit control signal? If you could also double check your explanations, they don't quite make sense currently, e.g.
I don't know what this means, is there a incorrect word in it maybe? |
|
also, what does
You are still manually buffering the circuit? |
To support the automated buffering for the speculation pass, I modified the speculation compilation flow.
The most important point is that now we have two passes for speculation which execute before and after buffering respectively.
The pre-buffering speculation pass places "pre-buffer" speculator units to preserve the join semantics:
Speculator is decomposed into SpecPreBuffer1 and 2, based on the join semantics. At this stage, SCCommitCtrl does not exist. It is only connected after buffering. Therefore, the control for save-commit units is connected temporarily. The network for SCCommitCtrl is partially constructed using SCIsMisspec and the loop condition.
After the buffering, two pre-buffering spec units are coalesced into a speculator. Moreover, SCCommitCtrl is connected and the control for save-commit units is final.
At this stage, we also place additional buffers to ensure high throughput and avoid deadlock.
I admit this method might not be ideal or optimal: because of the decomposition of speculator, the frequency regulation doesn't work. This is the quick workaround to make spec v1 comparable with new speculation and to mitigate the non-deterministic circuit construction.
Also, I fixed some bugs in speculator/save-commit implementations and speculation pass, which are needed to make the speculation integration tests work.