Optimize vector oword block load

Use vector copy instead of oword block load on specific device.

Author: ichenkai

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -17762,7 +17762,6 @@ void EmitPass::emitVectorLoad(LoadInst* inst, Value* offset, ConstantInt* immOff
         return;
     }
 
-
     bool bEmulateDWAligned = false;
 
     // generate oword-load if possible
@@ -17783,6 +17782,17 @@ void EmitPass::emitVectorLoad(LoadInst* inst, Value* offset, ConstantInt* immOff
             bool needTemp = (!destUniform || !IsGRFAligned(m_destination, EALIGN_GRF));
             CVariable * loadDest = m_destination;
 
+            // (needTemp = false) means GRF aligned as (destUniform = true)
+            if (!useDWAligned && useOWAligned && !needTemp &&
+                srcUniform && destUniform &&
+                destType == VISA_Type::ISA_TYPE_F &&
+                eOffset->GetType() == VISA_Type::ISA_TYPE_UD &&
+                m_currShader->m_Platform->allowFastestSIMDVectorLoad4Perf())
+            {
+                emitVectorCopy(loadDest, eOffset, elts);
+                return;
+            }
+
             if (useOWAligned)
             {
                 // Offset needs to be in OW!

IGC/Compiler/CISACodeGen/Platform.hpp

@@ -1787,5 +1787,11 @@ bool allowDivergentControlFlowRayQueryCheckRelease() const
     return m_WaTable.Wa_22019804511 != 0;
 }
 
+bool allowFastestSIMDVectorLoad4Perf() const
+{
+    return m_platformInfo.eProductFamily == IGFX_ALDERLAKE_P &&
+        m_platformInfo.usDeviceID == 0x4626;
+}
+
 };
 }//namespace IGC