Improve jump codegen by setting uniform if jump's flag is workgroup/global uniform

under EU fusion.

It is off for now. And it can be turned on by setting key EnableWorkGroupUniformGoto to 1.

Author: jgu222

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -444,15 +444,17 @@ namespace IGC
         VISA_EMask_Ctrl emask = m_encoderState.m_noMask ? vISA_EMASK_M1_NM : vISA_EMASK_M1;
         VISA_Exec_Size execSize = visaExecSize(m_program->m_dispatchSize);
 
-        // As uniform maybe thread-uniform, it will not work for scalar jump
-        // for two threads whose flags are different. Once we improve uniform
-        // analysis to have global (uniform among all threads), not just local
-        // uniform (within a thread), we can still generate scalar jump when flag
-        // is a global uniform.
-        //
-        // For now, just disable uniform goto if EU fusion is on.
+        // visa and igc agreement.
+        //    goto (1) is used to tell visa the goto is uniform.
+        // Goto(1) is generated if
+        //   1. jump is unconditional, or
+        //   2. jump is uniform (thread uniform or above) and no EU fusion, or
+        //   3. jump is either workgroup or global uniform under EU fusion
+        //      (it is temporarily under key control for ease of debugging)
         if (flag == nullptr ||
-            (!m_program->m_Platform->hasFusedEU() && flag->IsUniform()))
+            (!m_program->m_Platform->hasFusedEU() && flag->IsUniform()) ||
+            (IGC_IS_FLAG_ENABLED(EnableWorkGroupUniformGoto) &&
+             m_program->m_Platform->hasFusedEU() && flag->IsWorkGroupOrGlobalUniform()))
         {
             execSize = EXEC_SIZE_1;
         }
@@ -4272,7 +4274,7 @@ namespace IGC
             {
                 SaveOption( vISA_LocalScheduling, false );
                 // temporarily enabling preRA_schedule due to the ACOdyssey regression in IGC-4149.  May be re-disenabled later
-                // SaveOption( vISA_preRA_Schedule, false ); 
+                // SaveOption( vISA_preRA_Schedule, false );
                 SaveOption( vISA_SpillSpaceCompression, false );
                 SaveOption( vISA_LVN, false );
                 SaveOption( vISA_QuickTokenAllocation, true );

IGC/common/igc_flags.def

@@ -89,6 +89,7 @@ DECLARE_IGC_REGKEY(bool, DisablePayloadCoalescing_RT,   false, "Setting this to
 DECLARE_IGC_REGKEY(bool, DisablePayloadCoalescing_Sample, false, "Setting this to 1/true adds a compiler switch to disable payload coalescing optimization for Samplers only", false)
 DECLARE_IGC_REGKEY(bool, DisablePayloadCoalescing_URB,  false, "Setting this to 1/true adds a compiler switch to disable payload coalescing optimization for URB writes only", false)
 DECLARE_IGC_REGKEY(bool, DisableUniformAnalysis,        false, "Setting this to 1/true adds a compiler switch to disable uniform_analysis", false)
+DECLARE_IGC_REGKEY(bool, EnableWorkGroupUniformGoto,    false, "Setting to 1 enables generating uniform goto for work group uniform [eu fusion only]", false)
 DECLARE_IGC_REGKEY(DWORD, DisablePushConstant,           0, "Bit mask to disable push constant per shader stages. bit0 = All, Bit 1 = VS, Bit 2 = HS, Bit 3 = DS, Bit 4 = GS, Bit 5 = PS", false)
 DECLARE_IGC_REGKEY(DWORD, DisableAttributePush,          0, "Bit mask to disable push Attribute per shader stages. bit0 = All, Bit 1 = VS, Bit 2 = HS, Bit 3 = DS, Bit 4 = GS", false)
 DECLARE_IGC_REGKEY(bool, DisableSimplePushWithDynamicUniformBuffers, false,"Disable Simple Push Constants Optimization for dynamic uniform buffers.", false)