Add VF immediate support for divm/sqrtm/pow

weiyu-chen · sys_zuul · commit 350842a2e84c · 2020-03-06T14:15:08.000-08:00
Change-Id: I49e63a8b4bee9b1179e9a4ad5c31ae6ff70a5d8b
diff --git a/visa/Gen4_IR.hpp b/visa/Gen4_IR.hpp
@@ -2303,6 +2303,21 @@ class G4_Operand
     {
         return Rel_disjoint;
     }
+
+    static G4_Type GetNonVectorImmType(G4_Type type)
+    {
+        switch (type)
+        {
+            case Type_V:
+                return Type_W;
+            case Type_UV:
+                return Type_UW;
+            case Type_VF:
+                return Type_F;
+            default:
+                return type;
+        }
+    }
 };
 
 class G4_VarBase
diff --git a/visa/HWConformity.cpp b/visa/HWConformity.cpp
@@ -588,7 +588,7 @@ bool HWConformity::fixMathInst(INST_LIST_ITER it, G4_BB *bb)
         bool needsSrc0Mov = needsMove(0, src0_type);
         if (needsSrc0Mov)
         {
-            inst->setSrc(insertMovBefore(it, 0, src0->isImm() ? getNonVectorType(src0_type) : src0_type, bb), 0);
+            inst->setSrc(insertMovBefore(it, 0, src0->isImm() ? G4_Operand::GetNonVectorImmType(src0_type) : src0_type, bb), 0);
             src0 = inst->getSrc(0);
         }
     }
@@ -609,7 +609,7 @@ bool HWConformity::fixMathInst(INST_LIST_ITER it, G4_BB *bb)
             }
             else
             {
-                inst->setSrc(insertMovBefore(it, 1, src1->isImm() ? getNonVectorType(src1_type) : src1_type, bb), 1);
+                inst->setSrc(insertMovBefore(it, 1, src1->isImm() ? G4_Operand::GetNonVectorImmType(src1_type) : src1_type, bb), 1);
             }
             src1 = inst->getSrc(1);
         }
@@ -752,7 +752,7 @@ void HWConformity::fixImmAndARFSrc(INST_LIST_ITER it, G4_BB *bb)
 
                 if (needConstMov)
                 {
-                    G4_Type tmpType = getNonVectorType(src0->getType());
+                    G4_Type tmpType = G4_Operand::GetNonVectorImmType(src0->getType());
 
                     G4_Operand* newSrc0 = insertMovBefore(it, 0, tmpType, bb);
                     inst->setSrc(newSrc0, 0);
@@ -859,7 +859,7 @@ void HWConformity::fixImmAndARFSrc(INST_LIST_ITER it, G4_BB *bb)
                 //   mov(8) TV0(0, 0)<1> : uw 0xeca86420 : uv{ Align1 }
                 //   add(8) A0(0, 0)<1> : uw &V36 + 0 TV0(0, 0)<8; 8, 1> : uw{ Align1, Q1 }
                 G4_Type type = src1->getType();
-                inst->setSrc(insertMovBefore(it, 1, getNonVectorType(type), bb), 1);
+                inst->setSrc(insertMovBefore(it, 1, G4_Operand::GetNonVectorImmType(type), bb), 1);
                 // And we swap addr expr and the new variable
                 //   add(8) A0(0, 0)<1> : uw TV0(0, 0)<8; 8, 1> : uw &V36 + 0 {Align1, Q1}
                 // The final code sequence is
@@ -872,7 +872,7 @@ void HWConformity::fixImmAndARFSrc(INST_LIST_ITER it, G4_BB *bb)
             else
             {
                 G4_Type newSrcType = inst->needsDWType() ? (IS_UNSIGNED_INT(src0->getType()) ? Type_UD : Type_D) :
-                    getNonVectorType(src0->getType());
+                    G4_Operand::GetNonVectorImmType(src0->getType());
                 inst->setSrc(insertMovBefore(it, 0, newSrcType, bb), 0);
             }
         }
diff --git a/visa/HWConformity.h b/visa/HWConformity.h
@@ -101,7 +101,6 @@ namespace vISA
         void fixPackedHFConversions(INST_LIST_ITER it, G4_BB* bb);
 
         bool fixIndirectOpnd(INST_LIST_ITER i, G4_BB *bb);
-        G4_Type getNonVectorType(G4_Type srcType);
         void fix3SrcInst(INST_LIST_ITER i, G4_BB* bb);
         void fixAlign13SrcInst(INST_LIST_ITER i, G4_BB* bb);
         void fixCompareInst(INST_LIST_ITER i, G4_BB *bb, G4_Type exType, int dst_elsize);
diff --git a/visa/IsaVerification.cpp b/visa/IsaVerification.cpp
@@ -1604,12 +1604,12 @@ static void verifyInstructionArith(
                       "mul/div does not support Q/UQ types for src%d", i);
                   break;
             case ISA_DIVM:
-                  REPORT_INSTRUCTION(options,srcType == ISA_TYPE_F || srcType == ISA_TYPE_DF,
-                      "ieee div does not support types for src%d, other than F/DF", i);
+                  REPORT_INSTRUCTION(options,srcType == ISA_TYPE_F || srcType == ISA_TYPE_DF || srcType == ISA_TYPE_VF,
+                      "ieee div does not support types for src%d, other than F/DF/VF", i);
                   break;
             case ISA_SQRTM:
-                  REPORT_INSTRUCTION(options,srcType == ISA_TYPE_F || srcType == ISA_TYPE_DF,
-                      "ieee sqrt does not support types for src%d, other than F/DF", i);
+                  REPORT_INSTRUCTION(options,srcType == ISA_TYPE_F || srcType == ISA_TYPE_DF || srcType == ISA_TYPE_VF,
+                      "ieee sqrt does not support types for src%d, other than F/DF/VF", i);
                   break;
             case ISA_ADDC:
             case ISA_SUBB:
diff --git a/visa/ReduceExecSize.cpp b/visa/ReduceExecSize.cpp
@@ -31,23 +31,6 @@ using namespace vISA;
 
 using std::printf;
 
-G4_Type HWConformity::getNonVectorType( G4_Type srcType )
-{
-    if( srcType == Type_V )
-    {
-        return Type_W;
-    }
-    else if( srcType == Type_UV )
-    {
-        return Type_UW;
-    }
-    else if( srcType == Type_VF )
-    {
-        return Type_F;
-    }
-    return srcType;
-}
-
 uint8_t HWConformity::checkMinExecSize( G4_opcode op )
 {
     if( op == G4_dp2 ||
@@ -1514,7 +1497,7 @@ bool HWConformity::checkSrcDstOverlap( INST_LIST_ITER iter, G4_BB* bb, bool comp
             if (useTmp)
             {
                 // insert mov
-                inst->setSrc(insertMovBefore(iter, i, getNonVectorType(srcs[i]->getType()), bb), i);
+                inst->setSrc(insertMovBefore(iter, i, G4_Operand::GetNonVectorImmType(srcs[i]->getType()), bb), i);
                 srcs[i] = inst->getSrc(i);
                 // reducing exec size for new MOV
                 INST_LIST_ITER newMovIter = iter;
@@ -1549,7 +1532,7 @@ void HWConformity::moveSrcToGRF( INST_LIST_ITER it, uint32_t srcNum, uint16_t nu
     G4_INST* def_inst = NULL;
     def_inst = checkSrcDefInst(inst, def_inst, srcNum);
 
-    G4_Type tmpType = getNonVectorType(src->getType());
+    G4_Type tmpType = G4_Operand::GetNonVectorImmType(src->getType());
 
     if( def_inst && def_inst->getDst()->getType() == tmpType &&
         ( def_inst->getExecSize() == execSize ) &&
diff --git a/visa/TranslationInterface.cpp b/visa/TranslationInterface.cpp
@@ -252,7 +252,7 @@ void IR_Builder::expandPow(uint8_t exsize, G4_Predicate *predOpnd, bool saturate
     // math.log tmp abs(src0)
     // mul dst tmp tmp src1
     // math.exp dst tmp
-    G4_Type mathType = src0Opnd->getType();
+    G4_Type mathType = G4_Operand::GetNonVectorImmType(src0Opnd->getType());
     G4_Declare* tmpVar = createTempVar(exsize, mathType, Any);
     G4_DstRegRegion* logDst = Create_Dst_Opnd_From_Dcl(tmpVar, 1);
     G4_Operand* logSrc = src0Opnd;
@@ -299,6 +299,9 @@ void IR_Builder::expandPow(uint8_t exsize, G4_Predicate *predOpnd, bool saturate
                 }
                 break;
             }
+            case Type_VF:
+                // ToDo: what if VF contains negative values?
+                break;
             default:
                 assert(false && "unexpected src0 type for pow");
         }
@@ -441,7 +444,7 @@ static G4_SrcRegRegion* operandToDirectSrcRegRegion(
     {
         //src is an immediate
         MUST_BE_TRUE(src->isImm(), "expect immediate operand");
-        G4_Declare *tmpSrc = builder.createTempVarWithNoSpill(exsize, src->getType(), Any);
+        G4_Declare *tmpSrc = builder.createTempVarWithNoSpill(exsize, G4_Operand::GetNonVectorImmType(src->getType()), Any);
         builder.Create_MOV_Inst(tmpSrc, 0, 0, exsize, nullptr, nullptr, src, true);
         return builder.Create_Src_Opnd_From_Dcl(tmpSrc, builder.getRegionStride1());
     }

Original file line number	Diff line number	Diff line change
`@@ -588,7 +588,7 @@ bool HWConformity::fixMathInst(INST_LIST_ITER it, G4_BB *bb)`
`588`	`588`	`bool needsSrc0Mov = needsMove(0, src0_type);`
`589`	`589`	`if (needsSrc0Mov)`
`590`	`590`	`{`
`591`		`- inst->setSrc(insertMovBefore(it, 0, src0->isImm() ? getNonVectorType(src0_type) : src0_type, bb), 0);`
	`591`	`+ inst->setSrc(insertMovBefore(it, 0, src0->isImm() ? G4_Operand::GetNonVectorImmType(src0_type) : src0_type, bb), 0);`
`592`	`592`	`src0 = inst->getSrc(0);`
`593`	`593`	`}`
`594`	`594`	`}`
`@@ -609,7 +609,7 @@ bool HWConformity::fixMathInst(INST_LIST_ITER it, G4_BB *bb)`
`609`	`609`	`}`
`610`	`610`	`else`
`611`	`611`	`{`
`612`		`- inst->setSrc(insertMovBefore(it, 1, src1->isImm() ? getNonVectorType(src1_type) : src1_type, bb), 1);`
	`612`	`+ inst->setSrc(insertMovBefore(it, 1, src1->isImm() ? G4_Operand::GetNonVectorImmType(src1_type) : src1_type, bb), 1);`
`613`	`613`	`}`
`614`	`614`	`src1 = inst->getSrc(1);`
`615`	`615`	`}`
`@@ -752,7 +752,7 @@ void HWConformity::fixImmAndARFSrc(INST_LIST_ITER it, G4_BB *bb)`
`752`	`752`
`753`	`753`	`if (needConstMov)`
`754`	`754`	`{`
`755`		`- G4_Type tmpType = getNonVectorType(src0->getType());`
	`755`	`+ G4_Type tmpType = G4_Operand::GetNonVectorImmType(src0->getType());`
`756`	`756`
`757`	`757`	`G4_Operand* newSrc0 = insertMovBefore(it, 0, tmpType, bb);`
`758`	`758`	`inst->setSrc(newSrc0, 0);`
`@@ -859,7 +859,7 @@ void HWConformity::fixImmAndARFSrc(INST_LIST_ITER it, G4_BB *bb)`
`859`	`859`	`// mov(8) TV0(0, 0)<1> : uw 0xeca86420 : uv{ Align1 }`
`860`	`860`	`// add(8) A0(0, 0)<1> : uw &V36 + 0 TV0(0, 0)<8; 8, 1> : uw{ Align1, Q1 }`
`861`	`861`	`G4_Type type = src1->getType();`
`862`		`- inst->setSrc(insertMovBefore(it, 1, getNonVectorType(type), bb), 1);`
	`862`	`+ inst->setSrc(insertMovBefore(it, 1, G4_Operand::GetNonVectorImmType(type), bb), 1);`
`863`	`863`	`// And we swap addr expr and the new variable`
`864`	`864`	`// add(8) A0(0, 0)<1> : uw TV0(0, 0)<8; 8, 1> : uw &V36 + 0 {Align1, Q1}`
`865`	`865`	`// The final code sequence is`
`@@ -872,7 +872,7 @@ void HWConformity::fixImmAndARFSrc(INST_LIST_ITER it, G4_BB *bb)`
`872`	`872`	`else`
`873`	`873`	`{`
`874`	`874`	`G4_Type newSrcType = inst->needsDWType() ? (IS_UNSIGNED_INT(src0->getType()) ? Type_UD : Type_D) :`
`875`		`- getNonVectorType(src0->getType());`
	`875`	`+ G4_Operand::GetNonVectorImmType(src0->getType());`
`876`	`876`	`inst->setSrc(insertMovBefore(it, 0, newSrcType, bb), 0);`
`877`	`877`	`}`
`878`	`878`	`}`
Original file line number	Diff line number	Diff line change
`@@ -252,7 +252,7 @@ void IR_Builder::expandPow(uint8_t exsize, G4_Predicate *predOpnd, bool saturate`
`252`	`252`	`// math.log tmp abs(src0)`
`253`	`253`	`// mul dst tmp tmp src1`
`254`	`254`	`// math.exp dst tmp`
`255`		`- G4_Type mathType = src0Opnd->getType();`
	`255`	`+ G4_Type mathType = G4_Operand::GetNonVectorImmType(src0Opnd->getType());`
`256`	`256`	`G4_Declare* tmpVar = createTempVar(exsize, mathType, Any);`
`257`	`257`	`G4_DstRegRegion* logDst = Create_Dst_Opnd_From_Dcl(tmpVar, 1);`
`258`	`258`	`G4_Operand* logSrc = src0Opnd;`
`@@ -299,6 +299,9 @@ void IR_Builder::expandPow(uint8_t exsize, G4_Predicate *predOpnd, bool saturate`
`299`	`299`	`}`
`300`	`300`	`break;`
`301`	`301`	`}`
	`302`	`+ case Type_VF:`
	`303`	`+ // ToDo: what if VF contains negative values?`
	`304`	`+ break;`
`302`	`305`	`default:`
`303`	`306`	`assert(false && "unexpected src0 type for pow");`
`304`	`307`	`}`
`@@ -441,7 +444,7 @@ static G4_SrcRegRegion* operandToDirectSrcRegRegion(`
`441`	`444`	`{`
`442`	`445`	`//src is an immediate`
`443`	`446`	`MUST_BE_TRUE(src->isImm(), "expect immediate operand");`
`444`		`- G4_Declare *tmpSrc = builder.createTempVarWithNoSpill(exsize, src->getType(), Any);`
	`447`	`+ G4_Declare *tmpSrc = builder.createTempVarWithNoSpill(exsize, G4_Operand::GetNonVectorImmType(src->getType()), Any);`
`445`	`448`	`builder.Create_MOV_Inst(tmpSrc, 0, 0, exsize, nullptr, nullptr, src, true);`
`446`	`449`	`return builder.Create_Src_Opnd_From_Dcl(tmpSrc, builder.getRegionStride1());`
`447`	`450`	`}`