Scratch space separation control

Add 'EnableSeparateScratchWA' regkey to control if we should apply the workaround when calculating slot0 and slot1 sizes
intel · Sep 13, 2023 · ea92b75 · ea92b75
1 parent 4a82e68
commit ea92b75
Show file tree

Hide file tree

Showing 6 changed files with 32 additions and 10 deletions.
diff --git a/IGC/Compiler/CISACodeGen/CISABuilder.cpp b/IGC/Compiler/CISACodeGen/CISABuilder.cpp
@@ -5323,7 +5323,8 @@ namespace IGC
         // slot1 is used for spilling only when SeparatingSpillAndPrivateScratchMemorySpace is on
         // and Slot0 is used for IGC private memory
         if (m_program->m_Platform->hasScratchSurface() &&
-            m_program->m_DriverInfo->supportsSeparatingSpillAndPrivateScratchMemorySpace())
+            (m_program->m_DriverInfo->supportsSeparatingSpillAndPrivateScratchMemorySpace() ||
+             m_program->GetContext()->getModuleMetaData()->compOpt.SeparateSpillPvtScratchSpace))
         {
             V(vKernel->AddKernelAttribute("SepSpillPvtSS", 0, nullptr));
         }

diff --git a/IGC/Compiler/CISACodeGen/CShader.cpp b/IGC/Compiler/CISACodeGen/CShader.cpp
@@ -72,11 +72,15 @@ CShader::CShader(Function* pFunc, CShaderProgram* pProgram)
     m_SavedFP = nullptr;
 
     bool SepSpillPvtSS = m_ctx->platform.hasScratchSurface() &&
-        m_ctx->m_DriverInfo.supportsSeparatingSpillAndPrivateScratchMemorySpace();
+        (m_ctx->m_DriverInfo.supportsSeparatingSpillAndPrivateScratchMemorySpace() ||
+         m_ctx->getModuleMetaData()->compOpt.SeparateSpillPvtScratchSpace);
+    bool SeparateScratchWA =
+        IGC_IS_FLAG_ENABLED(EnableSeparateScratchWA) &&
+        !m_ctx->getModuleMetaData()->compOpt.DisableSeparateScratchWA;
     m_simdProgram.init(!m_ctx->platform.hasScratchSurface(),
         m_ctx->platform.maxPerThreadScratchSpace(),
         GetContext()->getModuleMetaData()->compOpt.UseScratchSpacePrivateMemory,
-        SepSpillPvtSS);
+        SepSpillPvtSS, SeparateScratchWA);
 }
 
 void CShader::InitEncoder(SIMDMode simdSize, bool canAbortOnSpill, ShaderDispatchMode shaderMode)
@@ -3860,7 +3864,8 @@ bool CShader::CompileSIMDSizeInCommon(SIMDMode simdMode)
     m_simdProgram.setScratchSpaceUsedByShader(m_ScratchSpaceSize);
 
     if (m_ctx->platform.hasScratchSurface() &&
-        m_ctx->m_DriverInfo.supportsSeparatingSpillAndPrivateScratchMemorySpace())
+        (m_ctx->m_DriverInfo.supportsSeparatingSpillAndPrivateScratchMemorySpace() ||
+         m_ctx->getModuleMetaData()->compOpt.SeparateSpillPvtScratchSpace))
     {
         ret = ((m_simdProgram.getScratchSpaceUsageInSlot0() <= m_ctx->platform.maxPerThreadScratchSpace()) &&
             (m_simdProgram.getScratchSpaceUsageInSlot1() <= m_ctx->platform.maxPerThreadScratchSpace()));

diff --git a/IGC/Compiler/CodeGenPublic.h b/IGC/Compiler/CodeGenPublic.h
@@ -148,6 +148,7 @@ namespace IGC
         bool            m_roundPower2KBytes = false;
         bool            m_UseScratchSpacePrivateMemory = true;
         bool            m_SeparatingSpillAndPrivateScratchMemorySpace = false;
+        bool            m_EnableSeparateScratchWA = true;
         unsigned int m_scratchSpaceSizeLimit = 0;
         unsigned int m_numGRFTotal = 128;
         unsigned int m_numGRFSpillFill = 0;
@@ -195,12 +196,14 @@ namespace IGC
             }
         }
 
-        void init(bool roundPower2KBytes, unsigned int scratchSpaceSizeLimitT, bool useScratchSpacePrivateMemory, bool SepSpillPvtSS)
+        void init(bool roundPower2KBytes, unsigned int scratchSpaceSizeLimitT, bool useScratchSpacePrivateMemory, bool SepSpillPvtSS,
+            bool SeparateScratchWA)
         {
             m_roundPower2KBytes = roundPower2KBytes;
             m_scratchSpaceSizeLimit = scratchSpaceSizeLimitT;
             m_UseScratchSpacePrivateMemory = useScratchSpacePrivateMemory;
             m_SeparatingSpillAndPrivateScratchMemorySpace = SepSpillPvtSS;
+            m_EnableSeparateScratchWA = SeparateScratchWA;
         }
 
         // if IGC needs scratch for private memory, we use slot0 for private
@@ -218,7 +221,7 @@ namespace IGC
             {
                 result += (m_scratchSpaceUsedBySpills + m_scratchSpaceUsedByGtpin);
             }
-            else
+            else if (m_EnableSeparateScratchWA)
             {
                 // \TODO: doubts about driver-compiler interface, conservatively set the size
                 // to the max of two slots
@@ -236,9 +239,16 @@ namespace IGC
             unsigned int result = 0;
             if (m_SeparatingSpillAndPrivateScratchMemorySpace && slot0_offset > 0)
             {
-                // \TODO: doubts about driver-compiler interface, conservatively set the size
-                // to the max of two slots
-                result = std::max(slot0_offset, m_scratchSpaceUsedBySpills + m_scratchSpaceUsedByGtpin);
+                if (m_EnableSeparateScratchWA)
+                {
+                    // \TODO: doubts about driver-compiler interface, conservatively set the size
+                    // to the max of two slots
+                    result = std::max(slot0_offset, m_scratchSpaceUsedBySpills + m_scratchSpaceUsedByGtpin);
+                }
+                else
+                {
+                    result = m_scratchSpaceUsedBySpills + m_scratchSpaceUsedByGtpin;
+                }
             }
             result = roundSize(result);
             IGC_ASSERT(result <= m_scratchSpaceSizeLimit);

diff --git a/IGC/Compiler/ModuleAllocaAnalysis.cpp b/IGC/Compiler/ModuleAllocaAnalysis.cpp
@@ -201,7 +201,10 @@ bool ModuleAllocaAnalysis::safeToUseScratchSpace() const
 
         // if one API doesn't support stateless, we should try to use smallest dispatch mode
         // which can hold more pvt_data to avoid error out.
-        if (Ctx.platform.hasScratchSurface() && Ctx.m_DriverInfo.supportsSeparatingSpillAndPrivateScratchMemorySpace() && !supportsStatelessSpacePrivateMemory)
+        if (Ctx.platform.hasScratchSurface() &&
+            (Ctx.m_DriverInfo.supportsSeparatingSpillAndPrivateScratchMemorySpace() ||
+             Ctx.getModuleMetaData()->compOpt.SeparateSpillPvtScratchSpace) &&
+            !supportsStatelessSpacePrivateMemory)
             simd_size = numLanes(Ctx.platform.getMinDispatchMode());
 
         unsigned maxScratchSpaceBytes = Ctx.platform.maxPerThreadScratchSpace();

diff --git a/IGC/common/MDFrameWork.h b/IGC/common/MDFrameWork.h
@@ -387,6 +387,8 @@ namespace IGC
         bool DashGSpecified                             = false;
         bool FastCompilation                            = false;
         bool UseScratchSpacePrivateMemory               = true;
+        bool SeparateSpillPvtScratchSpace               = false;
+        bool DisableSeparateScratchWA                   = false;
         bool RelaxedBuiltins                            = false;
         bool SubgroupIndependentForwardProgressRequired = true;
         bool GreaterThan2GBBufferRequired               = true;

diff --git a/IGC/common/igc_flags.h b/IGC/common/igc_flags.h
@@ -99,6 +99,7 @@ DECLARE_IGC_REGKEY(bool, DPASTokenReduction,            false, "optimization to
 DECLARE_IGC_REGKEY(bool, EnableAdd3,                    true,  "Enable Add3. XeHP+ only", true)
 DECLARE_IGC_REGKEY(bool, EnableBfn,                     true,  "Enable Bfn. XeHP+ only", true)
 DECLARE_IGC_REGKEY(bool, SeparateSpillPvtScratchSpace,  false,  "Separate scratch spaces for spillfill and privatememory. XeHP and above only. Test only. Remove it once stabalized.", true)
+DECLARE_IGC_REGKEY(bool, EnableSeparateScratchWA,       true,  "Apply the workaround in slot0 and slot1 sizes when separating scratch spacesSeparate scratch space.", true)
 DECLARE_IGC_REGKEY(bool, DisableThreeALUPipes,          false,  "Disable three ALU Pipelines. XeHP only", true)
 DECLARE_IGC_REGKEY(bool, Enable16DWURBWrite,            false, "Enable 16 Dword URB Write messages", true)
 DECLARE_IGC_REGKEY(bool, Enable16OWSLMBlockRW,          true,  "Enable 16 OWord (8 GRF) SLM block read/write message", true)