Setting correct pushed inputs limit and use right max num GRFs

Basing on HW and vISA limitations

Author: skarczew

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -5967,8 +5967,11 @@ namespace IGC
     void CEncoder::DeclareInput(CVariable* var, uint offset, uint instance)
     {
         // Avoid declaring more inputs/outputs than available registers
-        if (offset + var->GetSize() >= vKernel->getNumRegTotal() * getGRFSize())
+        if (offset + var->GetSize() >= m_program->m_Platform->getMaxNumGRF(m_program->GetShaderType()) * getGRFSize())
+        {
+            IGC_ASSERT(0);
             return;
+        }
         V(vKernel->CreateVISAInputVar(
             var->visaGenVariable[instance],
             int_cast<unsigned short>(offset),
@@ -5978,8 +5981,11 @@ namespace IGC
     void CEncoder::DeclarePred(CVariable * var, uint offset)
     {
         // Avoid declaring more inputs/outputs than available registers
-        if (offset + var->GetSize() >= vKernel->getNumRegTotal() * getGRFSize())
+        if (offset + var->GetSize() >= m_program->m_Platform->getMaxNumGRF(m_program->GetShaderType()) * getGRFSize())
+        {
+            IGC_ASSERT(0);
             return;
+        }
         V(vKernel->CreateVISAPredVar(
             var->visaPredVariable,
             "",

IGC/Compiler/CISACodeGen/Platform.hpp

@@ -1213,6 +1213,18 @@ uint32_t getGRFSize() const
     return isCoreChildOf(IGFX_XE_HPC_CORE) ? 64 : 32;
 }
 
+uint32_t getMaxNumGRF(ShaderType type) const
+{
+    if (supportsVRT())
+    {
+        return 256;
+    }
+    else
+    {
+        return 128;
+    }
+}
+
 uint32_t getInlineDataSize() const
 {
     if (!supportInlineData()) return 0;

IGC/Compiler/CISACodeGen/PushAnalysis.cpp

@@ -1610,9 +1610,10 @@ namespace IGC
 
         if (m_context->platform.isCoreChildOf(IGFX_XE3_CORE))
         {
+            // VRT allows us to have 256GRFs, so we can double the number of pushed inputs
             MaxNumOfPushedInputs = 24 * 2;
-            m_pMaxNumOfVSPushedInputs = 24 * 2;
-            m_pMaxNumOfDSPushedInputs = 24 * 2;
+            m_pMaxNumOfVSPushedInputs = 30;
+            m_pMaxNumOfDSPushedInputs = 24 * 2; // 48 * vec4 == 192 GRF == 3/4 256 GRF
         }
         else
         {