Issues regarding Gem5 Indirect Memory Prefetcher

Question

I am trying to use IndirectMemoryPrefetcher() in my research project. In Proof of Concept when I am trying to train the Indirect Memory Prefetcher, I am not able to see any Prefetch hits at all. I’m expecting the IMP will get activated though:

printf("- IMP training Start\n");

for (int i = 0; i < trainSize; i++)

        tmp = dataArray[indices[i]];

printf("- IMP training End\n");

I suspect that Indirect Prefetcher Memory is not getting activated.

I’m using caches.py and two_level.py from the configs/learning_gem5/part1/ and I’ve added system.cpu = DerivO3CPU(branchPred=LTAGE()) to two_level.py and prefetcher details to caches.py as seen in below screenshot. IndirectMemoryPrefetcher() is used in L2Cache in a similar fashion.

It would be a great help if anyone would guide us with the instructions on using IMPs.enter image description here

I wrote a simple Proof of Concept(PoC) code to see if IMP is working or not. I tried debugging the performance with the debug flag ‘HWPrefetch’ to see if I am getting a prefetch hit/request.

while using the same PoC with StridePrefetcher() the indices array is getting prefetched in between the 2 print statements. For a successful IMP prefetching I’m expecting a Cache Hit on indexes beyond ‘trainSize‘ in my dataArray.

T1 = __rdtscp(&trash);

tmp = dataArray[trainSize + 64];

T2 = __rdtscp(&trash);