Changes in code.

Change-Id: I68a598a82f10474a26a56597ce9388dfdcdb7f61

Author: jgu222

IGC/Compiler/CISACodeGen/DriverInfo.hpp

@@ -252,6 +252,9 @@ namespace IGC
         /// Allow aggressive vector value aliasing
         virtual bool EnableVecAliasing() const { return false; }
 
+        /// Prevent MAD optimization if result used in Floor
+        virtual bool PreventMadforRound() const { return false; }
+
 
 
     };

IGC/Compiler/CISACodeGen/PatternMatchPass.cpp

@@ -1797,7 +1797,7 @@ namespace IGC
             return I.getType()->isFloatingPointTy();
         };
 
-        if (isFpMad(I) && (m_ctx->getModuleMetaData()->isPrecise || m_ctx->getModuleMetaData()->compOpt.disableMathRefactoring))
+        if (isFpMad(I) && m_ctx->getModuleMetaData()->isPrecise)
         {
             return false;
         }
@@ -1819,6 +1819,12 @@ namespace IGC
             return false;
         }
 
+        bool preventMadForRound = false;
+        if (IGC_IS_FLAG_ENABLED(EnableMadRoundDepCheck) || m_ctx->m_DriverInfo.PreventMadforRound())
+        {
+            preventMadForRound = true;
+        }
+
         bool found = false;
         llvm::Value* sources[3];
         e_modifier src_mod[3];
@@ -1900,6 +1906,47 @@ namespace IGC
                 return false;
         }
 
+        if (found && preventMadForRound)
+        {
+            /*=============================================================================
+            This checks for pattern
+
+            mul x a b
+            add y x c
+            floor z y
+
+            ----or---
+
+            mul x a b
+            add y x c
+            floor z abs(y)
+
+            If the case falls in either of the two cases, Mad optimisation is skipped because
+            small precision difference between Mul+add and Mad can be extrapolated by floor
+            (currently just enabled for floor operation, if required add for ceiling in future)
+            =============================================================================*/
+
+            for (auto iter = I.user_begin(); iter != I.user_end(); iter++)
+            {
+                if (IntrinsicInst * source = dyn_cast<IntrinsicInst>(*iter))
+                {
+                    if (source->getIntrinsicID() == Intrinsic::fabs)
+                    {
+                        for (auto it = source->user_begin(); it != source->user_end(); it++)
+                        {
+                            IntrinsicInst* source_fabs = dyn_cast<IntrinsicInst>(*it);
+                            if (source_fabs && source_fabs->getIntrinsicID() == Intrinsic::floor)
+                                return false;
+                        }
+                    }
+                    if (source->getIntrinsicID() == Intrinsic::floor)
+                    {
+                        return false;
+                    }
+                }
+            }
+        }
+
         if (found)
         {
             MadPattern* pattern = new (m_allocator) MadPattern();

IGC/common/MDFrameWork.h

@@ -201,7 +201,6 @@ namespace IGC
         bool UniformWGS                                 = false;
         bool disableVertexComponentPacking              = false;
         bool PreferBindlessImages                       = false;
-        bool disableMathRefactoring                     = false;
     };
 
     struct ComputeShaderInfo

IGC/common/igc_flags.def

@@ -186,6 +186,8 @@ DECLARE_IGC_REGKEY(DWORD, RetryManagerFirstStateId,     0,     "For debugging pu
 DECLARE_IGC_REGKEY(bool, DisableSendSrcDstOverlapWA,    false, "Disable Send Source/destination overlap WA which is enabled for GEN10/GEN11 and whenever Wddm2Svm is set in WATable")
 DECLARE_IGC_REGKEY(debugString, DisablePassToggles,     0,     "Disable each IGC pass by setting the bit. HEXADECIMAL ONLY!. Ex: C0 is to disable pass 6 and pass 7.")
 DECLARE_IGC_REGKEY(bool, ForceStatelessForQueueT,       true,  "In OCL, force to use stateless memory to hold queue_t*. This is a legacy feature to be removed.")
+DECLARE_IGC_REGKEY(bool, EnableMadRoundDepCheck,        false,  "Enable Floor Dependency on Mad Check, Check was added since the floor was extrapolating the mad precision difference.")
+
 
 DECLARE_IGC_GROUP("Shader dumping")
 DECLARE_IGC_REGKEY(bool, EnableCosDump, false, "Enable cos dump")
@@ -290,7 +292,7 @@ DECLARE_IGC_REGKEY(DWORD, ConstantPromotionCmpSelSize,  4, "Array size threshold
 DECLARE_IGC_REGKEY(bool, EnableVariableReuse,           true, "Enable local variable reuse")
 DECLARE_IGC_REGKEY(bool, EnableVariableAlias,           true, "Enable variable aliases (part of VariableReuse Pass, but separate functionality)")
 DECLARE_IGC_REGKEY(DWORD, EnableVATemp,                 0, "[temp]Enable variable aliases sub-optimization, once it is stable, remove this key")
-DECLARE_IGC_REGKEY(DWORD, VATemp,                       1, "[temp]New code to replace code under EnableVATemp. Once stable, remove this.")
+DECLARE_IGC_REGKEY(DWORD, VATemp,                       0, "[temp]New code to replace code under EnableVATemp. Once stable, remove this.")
 DECLARE_IGC_REGKEY(bool, EnableExtractMask,             false, "When enabled, it is mostly for reducing response size of send messages.")
 DECLARE_IGC_REGKEY(DWORD, VariableReuseByteSize,        64, "The byte size threshold for variable reuse")
 DECLARE_IGC_REGKEY(bool, EnableGather4cpoWA,            true, "Enable WA transforming gather4cpo/gather4po into gather4c/gather4")