Add TGM fence workaround for Xe2

Add TGM fence workaround for Xe2 which needs to double every TMG
fence instruction if fence flush type is not NONE.

Author: fangliu2020

inc/common/sku_wa_defs.h

@@ -651,3 +651,9 @@ SPDX-License-Identifier: MIT
         WA_BUG_TYPE_UNKNOWN,
         WA_BUG_PERF_IMPACT_UNKNOWN, WA_COMPONENT_UNKNOWN)
 
+        WA_DECLARE(
+        Wa_14021891663,
+        "Workaround",
+        WA_BUG_TYPE_UNKNOWN,
+        WA_BUG_PERF_IMPACT_UNKNOWN, WA_COMPONENT_UNKNOWN)
+

visa/G4_SendDescs.cpp

@@ -700,6 +700,12 @@ LSC_DATA_ORDER G4_SendDescRaw::getLscDataOrder() const {
   }
 }
 
+LSC_FENCE_OP G4_SendDescRaw::getLscFenceOp() const {
+  vISA_ASSERT(isLscOp(), "must be LSC op");
+  vISA_ASSERT(isFence(), "must be fence op");
+  return static_cast<LSC_FENCE_OP>((desc.value >> 12) & 0x7);
+}
+
 int G4_SendDescRaw::getLscImmOff() const {
   vISA_ASSERT(isLscOp(), "must be LSC op");
   if (getSFID() == SFID::TGM)

visa/G4_SendDescs.hpp

@@ -508,6 +508,7 @@ class G4_SendDescRaw : public G4_SendDesc {
   LSC_ADDR_TYPE getLscAddrType() const;
   int getLscAddrSizeBytes() const; // e.g. a64 => 8
   LSC_DATA_ORDER getLscDataOrder() const;
+  LSC_FENCE_OP getLscFenceOp() const;
 
   int getLscImmOff() const;
   void setLscImmOff(int off);

visa/HWCaps.inc

@@ -893,4 +893,9 @@ uint32_t getSizeOfSendQueue() const {
 bool hasImmOnSrc0Src2ForAlign1Ternary() const {
   return getPlatformGeneration() >= PlatformGen::XE;
 }
+
+bool needTGMDoubleFenceWA() const {
+  return VISA_WA_CHECK(getPWaTable(), Wa_14021891663) &&
+      getOption(vISA_TGMDoubleFenceWA);
+}
 // end HW capabilities

visa/SWWA.cpp

@@ -3906,6 +3906,15 @@ void Optimizer::HWWorkaround() {
       if (inst->isIEEEExceptionTrap())
         expandIEEEExceptionTrap(ii, bb);
 
+      // Double up every TGM fence instruction if fenceOp is not
+      // LSC_FENCE_OP_NONE
+      if (builder.needTGMDoubleFenceWA() && inst->isSend() &&
+          inst->asSendInst()->isFence() &&
+          inst->asSendInst()->getMsgDesc()->getSFID() == SFID::TGM &&
+          inst->asSendInst()->getMsgDescRaw()->getLscFenceOp() !=
+              LSC_FENCE_OP_NONE)
+        bb->insertBefore(ii, inst->cloneInst());
+
       ii++;
     }
   }

visa/include/VISAOptionsDefs.h

@@ -741,6 +741,10 @@ DEF_VISA_OPTION(vISA_noIndirectSrcForCompressedInstWA, ET_BOOL,
                 "Disable WA for fixing compressed instructions with indirect "
                 "src0 and cross-grf dst",
                 false)
+DEF_VISA_OPTION(vISA_TGMDoubleFenceWA, ET_BOOL, "-tgmDoubleFenceWA",
+                "enable the WA to double any TGM fence instruction with "
+                "flush-type is not FLUSH_NONE",
+                false)
 
 //=== HW debugging options ===
 DEF_VISA_OPTION(vISA_GenerateDebugInfo, ET_BOOL, "-generateDebugInfo", UNUSED,