Changes in code.

intel · Feb 25, 2025 · b15c849 · b15c849
1 parent 34c8a74
commit b15c849
Show file tree

Hide file tree

Showing 5 changed files with 47 additions and 0 deletions.
diff --git a/visa/HWCaps.inc b/visa/HWCaps.inc
@@ -936,4 +936,10 @@ bool useDynamicAddrForExDesc() const {
   return getOption(vISA_dynamicAddrForExDescInLscSend);
 }
 
+
+bool src1Src2SwapForCompaction()
+{
+  return false;
+}
+
 // end HW capabilities
diff --git a/visa/Optimizer.cpp b/visa/Optimizer.cpp
@@ -691,6 +691,9 @@ void Optimizer::initOptimizations() {
                   TimerID::NUM_TIMERS);
   OPT_INITIALIZE_PASS(insertDummyCompactInst, vISA_InsertDummyCompactInst,
                   TimerID::NUM_TIMERS);
+  OPT_INITIALIZE_PASS(swapSrc1Src2OfMadForCompaction,
+                      vISA_SwapSrc1Src2OfMadForCompaction,
+                      TimerID::NUM_TIMERS);
   OPT_INITIALIZE_PASS(mergeScalarInst, vISA_MergeScalar, TimerID::OPTIMIZER);
   OPT_INITIALIZE_PASS(lowerMadSequence, vISA_EnableMACOpt, TimerID::OPTIMIZER);
   OPT_INITIALIZE_PASS(LVN, vISA_LVN, TimerID::OPTIMIZER);
@@ -1056,6 +1059,8 @@ int Optimizer::optimization() {
   // Insert a dummy compact instruction if requested for SKL+
   runPass(PI_insertDummyCompactInst);
 
+  runPass(PI_swapSrc1Src2OfMadForCompaction);
+
   runPass(PI_mapOrphans);
 
   runPass(PI_analyzeMove);

diff --git a/visa/Optimizer.h b/visa/Optimizer.h
@@ -305,6 +305,7 @@ class Optimizer {
   void insertDummyMovForHWRSWA();
   void insertHashMovs();
   void insertDummyCompactInst();
+  void swapSrc1Src2OfMadForCompaction();
   void removeLifetimeOps();
   void recomputeBound(std::unordered_set<G4_Declare *> &declares);
 
@@ -376,6 +377,7 @@ class Optimizer {
     PI_HWDebug,
     PI_insertDummyMovForHWRSWA,
     PI_insertDummyCompactInst,
+    PI_swapSrc1Src2OfMadForCompaction,
     PI_mergeScalarInst,
     PI_lowerMadSequence,
     PI_LVN,

diff --git a/visa/SWWA.cpp b/visa/SWWA.cpp
@@ -53,6 +53,38 @@ void Optimizer::insertDummyCompactInst() {
   bb->push_back(movInst);
 }
 
+void Optimizer::swapSrc1Src2OfMadForCompaction() {
+  if (!builder.src1Src2SwapForCompaction())
+    return;
+
+  BB_LIST_ITER ib, bend(fg.end());
+  for (ib = fg.begin(); ib != bend; ++ib) {
+    G4_BB *bb = (*ib);
+    INST_LIST_ITER ii = bb->begin();
+
+    while (ii != bb->end()) {
+      G4_INST *inst = *ii;
+      if (inst->opcode() == G4_mad) {
+        G4_Operand *src1 = inst->getSrc(1);
+        G4_Operand *src2 = inst->getSrc(2);
+        if (src1 && src2 && src1->getType() == src2->getType() &&
+            src1->isSrcRegRegion() &&
+            src2->isSrcRegRegion() &&
+            src1->getBase()->isRegVar() && src2->getBase()->isRegVar() &&
+            src1->getTopDcl()->getRegFile() == G4_GRF &&
+            src2->getTopDcl()->getRegFile() == G4_GRF) {
+          if (src1->asSrcRegRegion()->getRegion()->isScalar() &&
+              src2->asSrcRegRegion()->getRegion()->isFlatRegion()) {
+            inst->setSrc(src2, 1);
+            inst->setSrc(src1, 2);
+          }
+        }
+      }
+      ii++;
+    }
+  }
+}
+
 // add (1|M0)   null<1>:uw   null<0;1,0>:uw       0x0:uw
 void Optimizer::insertDummyAdd(G4_BB *bb, INST_LIST_ITER inst_it, int imm) {
   // Dst

diff --git a/visa/include/VISAOptionsDefs.h b/visa/include/VISAOptionsDefs.h
@@ -799,6 +799,8 @@ DEF_VISA_OPTION(vISA_InsertDummyMovForDPASRSWA, ET_BOOL,
                 "-insertDPASRSDummyMov", UNUSED, true)
 DEF_VISA_OPTION(vISA_InsertDummyCompactInst, ET_BOOL, "-insertDummyCompactInst",
                 UNUSED, false)
+DEF_VISA_OPTION(vISA_SwapSrc1Src2OfMadForCompaction, ET_BOOL, "-disableSwapSrc1Src2OfMadForCompaction",
+                UNUSED, true)
 DEF_VISA_OPTION(vISA_AsmFileNameOverridden, ET_BOOL, NULLSTR, UNUSED, false)
 DEF_VISA_OPTION(vISA_HashVal, ET_2xINT32, "-hashmovs",
                 "USAGE: -hashmovs hi32 lo32\n", 0)