Fdip align

.github/actions/build-dramsim/action.yml

@@ -12,6 +12,7 @@ runs:
         if [ ! -d "DRAMsim3" ]; then
           cp -r /nfs/home/share/gem5_ci/DRAMsim3 .
         fi
+        rm -rf DRAMsim3/build
         cd DRAMsim3 && mkdir -p build
         cd build
         cmake ..

configs/common/CacheConfig.py

@@ -69,6 +69,9 @@ def _get_cache_opts(cpu, level, options):
     prefetcher_attr = '{}_hwp_type'.format(level)
     if hasattr(options, prefetcher_attr) and (not options.no_pf):
         opts['prefetcher'] = create_prefetcher(cpu, level, options)
+        if level == 'l1i' and getattr(options, prefetcher_attr) == \
+                'FetchDirectedPrefetcher':
+            opts['demand_mshr_reserve'] = 2
 
     return opts
 

configs/common/PrefetcherConfig.py

@@ -29,7 +29,12 @@ def create_prefetcher(cpu, cache_level, options):
         return NULL
 
     if cpu != NULL:
-        prefetcher.registerTLB(cpu.mmu.dtb, cpu.mmu.functional)
+        prefetcher.registerTLB(
+            cpu.mmu.itb if cache_level == 'l1i' else cpu.mmu.dtb,
+            cpu.mmu.functional)
+
+    if prefetcher_name == 'FetchDirectedPrefetcher':
+        prefetcher.cpu = cpu
 
     if prefetcher_name == 'XSCompositePrefetcher':
         if options.l1d_enable_spp:

configs/example/kmhv3.py

@@ -176,6 +176,8 @@ def setKmhV3Params(args, system):
 
     # Enable prefetch buffers for all hardware prefetchers in this config.
     args.enable_pf_buffer = True
+    if not args.no_pf and args.l1i_hwp_type is None:
+        args.l1i_hwp_type = 'FetchDirectedPrefetcher'
 
     # Set default bp_type based on ideal_kmhv3 flag
     # If user didn't specify bp_type, set default based on ideal_kmhv3

src/cpu/o3/BaseO3CPU.py

@@ -255,6 +255,22 @@ def support_take_over(cls):
 
     store_prefetch_train = Param.Bool(True, "Training store prefetcher with store addresses")
 
+    fdip = Param.Bool(False, "Enable fetch-directed instruction prefetch")
+    fdipLookaheadTargets = Param.Unsigned(
+        4, "Number of future FSQ targets to scan for FDIP")
+    fdipMaxPrefetchesPerCycle = Param.Unsigned(
+        2, "Maximum FDIP prefetch requests generated per cycle")
+    fdipMaxBlocksPerTarget = Param.Unsigned(
+        1, "Maximum cache blocks to prefetch for each FDIP target")
+    fdipMinTargetDistance = Param.Unsigned(
+        1, "Minimum distance from the current fetch target to prefetch")
+    fdipMinTargetAgeCycles = Param.Unsigned(
+        0, "Minimum predicted-target age before FDIP prefetches it")
+    fdipSkipTargetStartBlock = Param.Bool(
+        False, "Skip the first cache block of each FDIP target")
+    fdipMaxPendingTranslations = Param.Unsigned(
+        32, "Maximum FDIP translations in flight")
+
     # value predictor
     valuePred = Param.ValuePredictor(NULL, "valuepred unit")
     enableSelectiveVPFlush = Param.Bool(False,

src/cpu/o3/cpu.cc

@@ -340,6 +340,12 @@ CPU::regProbePoints()
     ppDataAccessComplete = new ProbePointArg<
         std::pair<DynInstPtr, PacketPtr>>(
                 getProbeManager(), "DataAccessComplete");
+    ppFTQInsert =
+        new ProbePointArg<branch_prediction::btb_pred::FdipFetchTargetPtr>(
+            getProbeManager(), "FTQInsert");
+    ppFTQRemove =
+        new ProbePointArg<branch_prediction::btb_pred::FdipFetchTargetPtr>(
+            getProbeManager(), "FTQRemove");
 
     fetch.regProbePoints();
     rename.regProbePoints();

src/cpu/o3/cpu.hh

@@ -69,6 +69,7 @@
 #include "cpu/o3/rob.hh"
 #include "cpu/o3/scoreboard.hh"
 #include "cpu/o3/thread_state.hh"
+#include "cpu/pred/btb/fdip_target.hh"
 #include "cpu/simple_thread.hh"
 #include "cpu/timebuf.hh"
 #include "cpu/valuepred/valuepred_unit.hh"
@@ -193,6 +194,10 @@ class CPU : public BaseCPU
 
     ProbePointArg<PacketPtr> *ppInstAccessComplete;
     ProbePointArg<std::pair<DynInstPtr, PacketPtr> > *ppDataAccessComplete;
+    ProbePointArg<branch_prediction::btb_pred::FdipFetchTargetPtr>
+        *ppFTQInsert;
+    ProbePointArg<branch_prediction::btb_pred::FdipFetchTargetPtr>
+        *ppFTQRemove;
 
     /** Register probe points. */
     void regProbePoints() override;