Enable GRF read delay of send stall instructions

Enable the vISA option -schedWithSendSrcReadCycle

Author: bcheng0127

visa/HWCaps.inc

@@ -845,6 +845,13 @@ bool needNopAfterCFInstWA() const {
   return false;
 }
 
+bool scheduleStoreAway() const {
+  const TARGET_PLATFORM p = getPlatform();
+  if (p >= Xe_PVCXT)
+    return getOption(vISA_schedWithSendSrcReadCycle);
+  return false;
+}
+
 bool hasReadSuppressionOrSharedLocalMemoryWAs() const {
   return getPlatform() < Xe2;
 }

visa/LocalScheduler/LocalScheduler_G4IR.cpp

@@ -2820,7 +2820,7 @@ uint32_t DDD::getEdgeLatency_old(Node *node, DepType depT) const {
 
   case WAR_MEMORY:
   case WAW_MEMORY:
-    if (kernel->getOption(vISA_schedWithSendSrcReadCycle)) {
+    if (kernel->fg.builder->scheduleStoreAway()) {
       latency = LT.getSendSrcReadLatency(inst);
     } else {
       latency = UNCOMPR_LATENCY;

visa/include/VISAOptionsDefs.h

@@ -435,7 +435,7 @@ DEF_VISA_OPTION(vISA_ScheduleACCDep, ET_BOOL, "-scheduleACCDep",
                 "Enable scheduling with accurate ACC dependence, instead of "
                 "coarse grained dependence",
                 false)
-DEF_VISA_OPTION(vISA_schedWithSendSrcReadCycle, ET_BOOL, "-schedWithSendSrcReadCycle", UNUSED, false)
+DEF_VISA_OPTION(vISA_schedWithSendSrcReadCycle, ET_BOOL_TRUE, "-schedWithSendSrcReadCycle", UNUSED, true)
 
 //=== SWSB options ===
 DEF_VISA_OPTION(vISA_USEL3HIT, ET_BOOL, "-SBIDL3Hit", UNUSED, false)