Use sampler header to solve HW issue

Use sampler header to solve HW issue

Author: bcheng0127

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -4631,6 +4631,10 @@ namespace IGC
         {
             SaveOption(vISA_forceSamplerHeader, true);
         }
+        if (IGC_IS_FLAG_ENABLED(samplerHeaderWA))
+        {
+            SaveOption(vISA_samplerHeaderWA, true);
+        }
         if (IGC_IS_FLAG_ENABLED(EnableIGAEncoder))
         {
             SaveOption(vISA_IGAEncoder, true);

IGC/common/igc_flags.h

@@ -578,6 +578,7 @@ DECLARE_IGC_REGKEY(bool, LimitConstantBuffersPushed,    true, "Limit max number
 DECLARE_IGC_REGKEY(DWORD, MaxPreRASchedulerRegPressureThreshold, 60,  "Max PreRA Scheduler Threshold", false)
 DECLARE_IGC_REGKEY(bool, EnablePreRASampleCluster,      false, "Enabling helps cluster sample instructions with identical texture index which are ready to be scheduled, to be scheduled together", false)
 DECLARE_IGC_REGKEY(bool, forceSamplerHeader,            false, "force sampler messages to use header", false)
+DECLARE_IGC_REGKEY(bool, samplerHeaderWA,            false, "enable sampler header to solve HW WA", false)
 DECLARE_IGC_REGKEY(bool, VFPackingDisablePartialElements, false, "disable packing for partial vertex element as it causes performance drops", false)
 DECLARE_IGC_REGKEY(bool, cl_khr_srgb_image_writes,      false, "Enable cl_khr_srgb_image_writes extension", false)
 DECLARE_IGC_REGKEY(bool, MSAA16BitPayloadEnable,        true,  "Enable support for MSAA 16 bit payload , a hardware DCN supporting this from ICL+ to improve perf on MSAA workloads", false)

visa/HWCaps.inc

@@ -34,6 +34,11 @@ SPDX-License-Identifier: MIT
             (getPlatform() < GENX_ICLLP && m_options->getOption(vISA_enablePreemption));
     }
 
+    bool samplerHeaderPreemptionWA() const
+    {
+        return m_options->getOption(vISA_enablePreemption) && m_options->getOption(vISA_samplerHeaderWA);
+    }
+
     bool needsNoPreemptR2ForSend() const
     {
         return getPlatformGeneration() == PlatformGen::GEN11;

visa/VisaToG4/TranslateSend3D.cpp

@@ -1865,7 +1865,8 @@ int IR_Builder::translateVISASampler3DInst(
     bool simd16HFReturn = FP16Return && execSize == 16;
     if (needSamplerHeader(this, pixelNullMask, nonZeroAoffImmi, needHeaderForChannels,
         isBindlessSampler(sampler),
-        simd16HFReturn))
+        simd16HFReturn) ||
+        samplerHeaderPreemptionWA())
     {
         useHeader = true;
         ++numRows;
@@ -2119,7 +2120,8 @@ int IR_Builder::translateVISAGather3dInst(
 
     if (needSamplerHeader(this, pixelNullMask, nonZeroAoffImmi, needHeaderForChannels,
         isBindlessSampler(sampler),
-        simd16HFReturn))
+        simd16HFReturn) ||
+        samplerHeaderPreemptionWA())
     {
         useHeader = true;
         ++numRows;

visa/include/VISAOptionsDefs.h

@@ -111,6 +111,7 @@ DEF_VISA_OPTION(vISA_expandPlane,         ET_BOOL, "-expandPlane",        UNUSED
 DEF_VISA_OPTION(vISA_FImmToHFImm,         ET_BOOL, "-fiTohfi",    UNUSED, false)
 DEF_VISA_OPTION(vISA_cacheSamplerHeader,  ET_BOOL, "-noSamplerHeaderCache", UNUSED, true)
 DEF_VISA_OPTION(vISA_forceSamplerHeader,  ET_BOOL, "-forceSamplerHeader", UNUSED, false)
+DEF_VISA_OPTION(vISA_samplerHeaderWA,  ET_BOOL, "-samplerHeaderWA", UNUSED, false)
 DEF_VISA_OPTION(vISA_markSamplerMoves,    ET_BOOL, "-markSamplerMoves", UNUSED, false)
 DEF_VISA_OPTION(vISA_noncoherentStateless, ET_BOOL, "-ncstateless",       UNUSED, false)
 DEF_VISA_OPTION(vISA_enablePreemption,    ET_BOOL,  "-enablePreemption",  UNUSED, false)