Add registry key for adding send GRF read cycle in scheduling

Add the key so that scheduler can consider the GRF read latency for send
instructions
_OS_DESCRIPTION

Author: bcheng0127

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -5152,6 +5152,11 @@ namespace IGC
             SaveOption(vISA_EnableGroupScheduleForBC, true);
         }
 
+        if (IGC_IS_FLAG_ENABLED(SchedWithSendSrcReadCycle))
+        {
+            SaveOption(vISA_schedWithSendSrcReadCycle, true);
+        }
+
         if (IGC_IS_FLAG_ENABLED(CopyA0ToDBG0))
         {
             SaveOption(vISA_CopyA0ToDBG0, true);

IGC/common/igc_flags.h

@@ -89,6 +89,7 @@ DECLARE_IGC_REGKEY(DWORD,EnableIndirectInstEnd,         0,     "Enable the indir
 DECLARE_IGC_REGKEY(DWORD,EnableSWSBTokenBarrier,        0,     "Enable force specific instruction as a barrier for software scoreboard generation", true)
 DECLARE_IGC_REGKEY(DWORD,SWSBTokenNum,                  0,     "Total tokens used for SWSB.", true)
 DECLARE_IGC_REGKEY(bool, EnableGroupScheduleForBC,      false,  "Enable bank conflict reduction in scheduling.", true)
+DECLARE_IGC_REGKEY(bool, SchedWithSendSrcReadCycle,     false,  "Scheduling with GRF read cycle from send.", true)
 DECLARE_IGC_REGKEY(bool, EnableIGASWSB,                 false,  "Use IGA for SWSB", true)
 DECLARE_IGC_REGKEY(bool, EnableSWSBStitch,              false,  "Insert dependence resolve for kernel stitching", true)
 DECLARE_IGC_REGKEY(bool, DisableRegDistDep,             false,  "distable regDist dependence", true)

visa/LocalScheduler/LocalScheduler_G4IR.cpp

@@ -2985,6 +2985,9 @@ uint32_t DDD::getEdgeLatency_old(Node *node, DepType depT) const {
   }
 
   if (depT <= NODEP || depT >= CONTROL_FLOW_BARRIER) {
+    if (kernel->fg.builder->modelSendSrcReadLatency() && inst->isSend()) {
+      return LT.getSendSrcReadLatency(inst);
+    }
     return node->getOccupancy();
   }