basic implement for estride value predictor #201
Conversation
This flag may later be used by the value predictor Change-Id: Ib15eaf4cb488b4458ace63a336cd4a761807dc88
Change-Id: I4a7d33e6c46503d9a08e0ec678e08665ce7f04d8
Change-Id: I8b3ca27605e921edaaea779a1f0ff32462d4a62d
src/cpu/o3/cpu.cc
Outdated
| @@ -1251,6 +1251,10 @@ CPU::instDone(ThreadID tid, const DynInstPtr &inst) | |||
| thread[tid]->threadStats.numInsts++; | |||
| cpuStats.committedInsts[tid]++; | |||
|
|
|||
| if (thread[tid]->numInst % 10000 == 0) { | |||
| fprintf(stderr, "commit %lu insts\n", thread[tid]->numInst); | |||
There was a problem hiding this comment.
We always prefer DPRINTF rather than fprintf in GEM5
There was a problem hiding this comment.
I've fixed the problem with this part of the code and I'll start running the scores after I get issue1 finished.
src/cpu/valuepred/enhanced_stride.cc
Outdated
| for (size_t i = 0; i < indexEachWays.size(); i++) { | ||
| indexEachWays[i] = pcHashToWayIndex(esPredMetaData->pc, i); | ||
| indexMarker += std::to_string(indexEachWays[i]); | ||
| indexMarker += " "; |
There was a problem hiding this comment.
I'm not sure how such string operation affect performance. If performance is degraded, I suggest to put strong operations into if block like if (debug::xxx) {indexMarker =
There was a problem hiding this comment.
I've fixed the problem with this part of the code and I'll start running the scores after I get issue1 finished.
|
I don't have a way to turn off the value predictor yet, the only way to turn off the value predictor right now is to checkout to xs-dev and turn off the value predictor by changing the branch to cancel the value predictor related changes. I think eventually I'll implement turning off the value predictor in code. |
Change-Id: Ice141526e5a67c61149cb95b42e289c3d0b6eae8
also fixed the value predictor's handling of fault. Change-Id: I288b6fd358f67ff1e381bbe46b03a64775597bf4
Change-Id: I9384405f20f999c0baaeaca458429c48a420eb6d
related roadmap: #188
This PR is solely intended to track code related to value prediction and does not currently need to be merged into xs-dev.
Implemented features in the current code:
Known issues:
spec benchmark scoring is still ongoing.