Codegen fix for int16

Ensure that appropriate number of GRFs is allocated.
intel · Sep 3, 2024 · 507ff7f · 507ff7f
1 parent be3445e
commit 507ff7f
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 1 deletion.
diff --git a/IGC/Compiler/CISACodeGen/CISABuilder.cpp b/IGC/Compiler/CISACodeGen/CISABuilder.cpp
@@ -2273,7 +2273,7 @@ namespace IGC
         unsigned char sendc = isSendc ? 1 : 0;
         unsigned char src0Size = src0->GetSize() / getGRFSize();
         unsigned char src1Size = src1 ? src1->GetSize() / getGRFSize() : 0;
-        unsigned char dstSize = dst ? dst->GetSize() / getGRFSize() : 0;
+        unsigned char dstSize = dst ? (dst->GetSize() + getGRFSize() - 1) / getGRFSize() : 0;
         VISA_PredOpnd* predOpnd = GetFlagOperand(m_encoderState.m_flag);
         VISA_RawOpnd* srcOpnd0 = GetRawSource(src0);
         VISA_RawOpnd* srcOpnd1 = GetRawSource(src1);

diff --git a/IGC/Compiler/CISACodeGen/EmitVISAPass.cpp b/IGC/Compiler/CISACodeGen/EmitVISAPass.cpp
@@ -23805,6 +23805,7 @@ void EmitPass::emitTraceRay(TraceRayIntrinsic* I, bool RayQueryEnable)
         (m_currShader->m_SIMDSize == SIMDMode::SIMD32
             ) ? 2 : 1;
 
+
     for (uint32_t Cnt = 0; Cnt < NumSend; Cnt++)
     {
         // 'payload' contains bvhLevel, traceRayCtrl, and stackID as per
@@ -23944,6 +23945,7 @@ void EmitPass::emitTraceRay(TraceRayIntrinsic* I, bool RayQueryEnable)
             }
         }
 
+
         m_encoder->Sends(
             Dst,
             header,
@@ -23954,6 +23956,7 @@ void EmitPass::emitTraceRay(TraceRayIntrinsic* I, bool RayQueryEnable)
             false);
 
         m_encoder->Push();
+
     }
 
     // Insert a software fence after the send.rta so no IO operations get