Disable dynamic rayquery management if we have to emulate SIMD32 for rayqueries

Disable dynamic rayquery management if we have to emulate SIMD32 for rayqueries

Author: jaladreips

IGC/AdaptorCommon/RayTracing/DynamicRayManagementPass.cpp

@@ -115,9 +115,11 @@ bool DynamicRayManagementPass::runOnFunction(Function& F)
     // 1. RayTracing is not supported on this platform.
     // 2. Shader does not use RayQuery at all.
     // 3. There are more than 1 exit block.
+    // 4. RayQuery needs splitting due to forced SIMD32
     if ((m_CGCtx->platform.supportRayTracing() == false) ||
         (!m_CGCtx->hasSyncRTCalls()) ||
-        (getNumberOfExitBlocks(F) > 1))
+        (getNumberOfExitBlocks(F) > 1) ||
+        m_CGCtx->syncRTCallsNeedSplitting())
     {
         return false;
     }

IGC/AdaptorCommon/RayTracing/TraceRayInlineLoweringPass.cpp

@@ -1040,31 +1040,6 @@ class RTGlobalsPointerLoweringPass : public FunctionPass
 
 char RTGlobalsPointerLoweringPass::ID = 0;
 
-bool RTGlobalsPointerLoweringPass::needsSplitting(const CodeGenContext* Ctx)
-{
-    // In general, we don't want to compile SIMD32 for rayquery.
-    // Determine if we are forced to do so.
-
-    if (Ctx->type != ShaderType::COMPUTE_SHADER)
-        return false;
-
-    auto& csInfo = Ctx->getModuleMetaData()->csInfo;
-
-    if (IGC_IS_FLAG_ENABLED(ForceCSSIMD32) ||
-        IGC_GET_FLAG_VALUE(ForceCSSimdSize4RQ) == 32)
-        return true;
-    if (IGC_IS_FLAG_ENABLED(ForceCSSIMD16))
-        return false;
-    if (csInfo.forcedSIMDSize == 32)
-        return true;
-    if (csInfo.forcedSIMDSize == 16)
-        return false;
-    if (csInfo.waveSize == 32)
-        return true;
-
-    return false;
-}
-
 bool RTGlobalsPointerLoweringPass::runOnFunction(Function& F)
 {
     m_CGCtx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
@@ -1085,7 +1060,7 @@ bool RTGlobalsPointerLoweringPass::runOnFunction(Function& F)
     IGC_ASSERT_MESSAGE(nullptr != modMD,
         "Invalid Module Metadata in RTGlobalsPointerLoweringPass");
 
-    const bool NeedsSplitting = needsSplitting(m_CGCtx);
+    const bool NeedsSplitting = m_CGCtx->syncRTCallsNeedSplitting();
 
     for (auto* GBP : globalBuffPtrs)
     {

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -23671,9 +23671,6 @@ void EmitPass::emitRayQueryCheckRelease(
 
     m_encoder->Lifetime(LIFETIME_START, dummySource);
 
-    const uint32_t NumSend =
-        (m_currShader->m_SIMDSize == SIMDMode::SIMD32
-            ) ? 2 : 1;
 
     CVariable* flag = nullptr;
 
@@ -23683,28 +23680,17 @@ void EmitPass::emitRayQueryCheckRelease(
         flag = GetSymbol(I->getOperand(0));
     }
 
-    for (uint32_t Cnt = 0; Cnt < NumSend; Cnt++)
-    {
-        if (m_currShader->m_SIMDSize == SIMDMode::SIMD32
-            )
-        {
-            m_encoder->SetSimdSize(SIMDMode::SIMD16);
-            if (Cnt == 1)
-                m_encoder->SetMask(EMASK_H2);
-        }
-
-        m_encoder->SetPredicate(flag);
+    m_encoder->SetPredicate(flag);
 
-        m_encoder->Sends(
-            nullptr,
-            header,
-            dummySource,
-            extDescriptor,
-            exDesc,
-            pMessDesc);
+    m_encoder->Sends(
+        nullptr,
+        header,
+        dummySource,
+        extDescriptor,
+        exDesc,
+        pMessDesc);
 
-        m_encoder->Push();
-    }
+    m_encoder->Push();
 
     // Insert a software fence after the send.rta so no IO operations get
     // scheduled across the send from below.  We should be able to remove this

IGC/Compiler/CodeGenPublic.h

@@ -1240,6 +1240,31 @@ namespace IGC
 
         // For remarks
         void initializeRemarkEmitter(const ShaderHash& hash);
+
+        bool syncRTCallsNeedSplitting()
+        {
+            // In general, we don't want to compile SIMD32 for rayquery.
+            // Determine if we are forced to do so.
+
+            if (type != ShaderType::COMPUTE_SHADER)
+                return false;
+
+            auto& csInfo = getModuleMetaData()->csInfo;
+
+            if (IGC_IS_FLAG_ENABLED(ForceCSSIMD32) ||
+                IGC_GET_FLAG_VALUE(ForceCSSimdSize4RQ) == 32)
+                return true;
+            if (IGC_IS_FLAG_ENABLED(ForceCSSIMD16))
+                return false;
+            if (csInfo.forcedSIMDSize == 32)
+                return true;
+            if (csInfo.forcedSIMDSize == 16)
+                return false;
+            if (csInfo.waveSize == 32)
+                return true;
+
+            return false;
+        }
     };
 
     struct SComputeShaderSecondCompileInput