Clarify the use of descriptor set indices in the PAL metadata

include/vkgcDefs.h

@@ -240,8 +240,9 @@ struct PipelineShaderOptions
     /// Use the LLVM backend's SI scheduler instead of the default scheduler.
     bool      useSiScheduler;
 
-    // Whether update descriptor root offset in ELF
-    bool      updateDescInElf;
+    /// Emit descriptor set indices instead of userdata table indices in the PAL userdata mapping metadata. For use with
+    /// pre-compilation: the userdata mapping must be adjusted later.
+    bool      emitDescriptorSetIndexInMetadata;
 
 #if LLPC_CLIENT_INTERFACE_MAJOR_VERSION >= 35
     /// Disable the the LLVM backend's LICM pass.

llpc/builder/llpcBuilderImplInOut.cpp

@@ -32,6 +32,7 @@
 #include "llpcBuilderImpl.h"
 #include "llpcInternal.h"
 #include "llpcPipelineState.h"
+#include "llpcUtil.h"
 
 #define DEBUG_TYPE "llpc-builder-impl-inout"
 

llpc/builder/llpcPipeline.h

@@ -139,8 +139,9 @@ struct ShaderOptions
     // Use the LLVM backend's SI scheduler instead of the default scheduler.
     bool      useSiScheduler;
 
-    // Whether update descriptor root offset in ELF
-    bool      updateDescInElf;
+    /// Emit descriptor set indices instead of userdata table indices in the PAL userdata mapping metadata. For use with
+    /// pre-compilation: the userdata mapping must be adjusted later.
+    bool      emitDescriptorSetIndexInMetadata;
 
     /// Default unroll threshold for LLVM.
     uint32_t  unrollThreshold;

llpc/builder/llpcPipelineState.cpp

@@ -37,6 +37,7 @@
 #include "llpcPatch.h"
 #include "llpcPipelineState.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Module.h"

llpc/builder/llpcResourceUsage.h

@@ -88,7 +88,7 @@ struct FsInterpInfo
 };
 
 // Invalid interpolation info
-static const FsInterpInfo InvalidFsInterpInfo = { InvalidValue, false, false, false };
+static const FsInterpInfo InvalidFsInterpInfo = { ~0u, false, false, false };
 
 // Enumerate the workgroup layout options.
 enum class WorkgroupLayout : uint32_t
@@ -453,7 +453,7 @@ struct InterfaceData
     static const uint32_t MaxEsGsOffsetCount = 6;
     static const uint32_t MaxCsUserDataCount = 10;
     static const uint32_t CsStartUserData     = 2;
-    static const uint32_t UserDataUnmapped = InvalidValue;
+    static const uint32_t UserDataUnmapped = ~0u;
 
     uint32_t                    userDataCount;                    // User data count
     uint32_t                    userDataMap[MaxUserDataCount];    // User data map (from SGPR No. to API logical ID)

llpc/context/llpcCompiler.cpp

@@ -1270,14 +1270,6 @@ Result Compiler::BuildPipelineInternal(
         graphicsShaderCacheChecker.UpdateAndMerge(result, pPipelineElf);
     }
 
-    if ((result == Result::Success) &&
-        pFragmentShaderInfo &&
-        pFragmentShaderInfo->options.updateDescInElf &&
-        (pContext->GetShaderStageMask() & ShaderStageToMask(ShaderStageFragment)))
-    {
-        graphicsShaderCacheChecker.UpdateRootUserDateOffset(pPipelineElf);
-    }
-
     pContext->setDiagnosticHandlerCallBack(nullptr);
 
     return result;
@@ -1334,19 +1326,6 @@ uint32_t GraphicsShaderCacheChecker::Check(
     return stageMask;
 }
 
-// =====================================================================================================================
-// Update root level descriptor offset for graphics pipeline.
-void GraphicsShaderCacheChecker::UpdateRootUserDateOffset(
-    ElfPackage*       pPipelineElf)   // [In, Out] ELF that could be from compile or merged
-{
-    ElfWriter<Elf64> writer(m_pContext->GetGfxIpVersion());
-    // Load ELF binary
-    auto result = writer.ReadFromBuffer(pPipelineElf->data(), pPipelineElf->size());
-    assert(result == Result::Success);
-    (void(result)); // unused
-    writer.UpdateElfBinary(m_pContext, pPipelineElf);
-}
-
 // =====================================================================================================================
 // Update shader caches for graphics pipeline from compile result, and merge ELF outputs if necessary.
 void GraphicsShaderCacheChecker::UpdateAndMerge(
@@ -1371,7 +1350,7 @@ void GraphicsShaderCacheChecker::UpdateAndMerge(
             auto result = writer.ReadFromBuffer(nonFragmentPipelineElf.pCode, nonFragmentPipelineElf.codeSize);
             assert(result == Result::Success);
             (void(result)); // unused
-            writer.MergeElfBinary(m_pContext, &m_fragmentElf, pPipelineElf);
+            writer.MergeFragmentShader(m_pContext, &m_fragmentElf, pPipelineElf);
 
             pipelineElf.codeSize = pPipelineElf->size();
             pipelineElf.pCode = pPipelineElf->data();
@@ -1400,7 +1379,7 @@ void GraphicsShaderCacheChecker::UpdateAndMerge(
             assert(result == Result::Success);
             (void(result)); // unused
 
-            writer.MergeElfBinary(m_pContext, &fragmentPipelineElf, pPipelineElf);
+            writer.MergeFragmentShader(m_pContext, &fragmentPipelineElf, pPipelineElf);
 
             pipelineElf.codeSize = pPipelineElf->size();
             pipelineElf.pCode = pPipelineElf->data();
@@ -1419,7 +1398,7 @@ void GraphicsShaderCacheChecker::UpdateAndMerge(
         auto result = writer.ReadFromBuffer(m_nonFragmentElf.pCode, m_nonFragmentElf.codeSize);
         assert(result == Result::Success);
         (void(result)); // unused
-        writer.MergeElfBinary(m_pContext, &m_fragmentElf, pPipelineElf);
+        writer.MergeFragmentShader(m_pContext, &m_fragmentElf, pPipelineElf);
     }
 
     // Whole pipeline is compiled

llpc/context/llpcCompiler.h

@@ -74,7 +74,6 @@ class GraphicsShaderCacheChecker
 
     // Update shader caches with results of compile, and merge ELF outputs if necessary.
     void UpdateAndMerge(Result result, ElfPackage* pPipelineElf);
-    void UpdateRootUserDateOffset(ElfPackage* pPipelineElf);
 
 private:
     Compiler* m_pCompiler;

llpc/context/llpcPipelineContext.cpp

@@ -358,7 +358,7 @@ void PipelineContext::SetOptionsInPipeline(
 #endif
 
             shaderOptions.useSiScheduler = EnableSiScheduler || pShaderInfo->options.useSiScheduler;
-            shaderOptions.updateDescInElf = pShaderInfo->options.updateDescInElf;
+            shaderOptions.emitDescriptorSetIndexInMetadata = pShaderInfo->options.emitDescriptorSetIndexInMetadata;
             shaderOptions.unrollThreshold = pShaderInfo->options.unrollThreshold;
 
             pPipeline->SetShaderOptions(static_cast<ShaderStage>(stage), shaderOptions);

llpc/patch/gfx6/chip/llpcGfx6ConfigBuilder.cpp

@@ -34,6 +34,7 @@
 #include "llpcGfx6ConfigBuilder.h"
 #include "llpcPipelineState.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 #include "llvm/Support/CommandLine.h"
 
 #define DEBUG_TYPE "llpc-gfx6-config-builder"
@@ -1391,7 +1392,7 @@ void ConfigBuilder::BuildUserDataConfig(
             if (pIntfData->userDataMap[i] != InterfaceData::UserDataUnmapped)
             {
                 AppendConfig(startUserData + i, pIntfData->userDataMap[i]);
-                if ((pIntfData->userDataMap[i] & DescRelocMagicMask) != DescRelocMagic)
+                if (pIntfData->userDataMap[i] < VkDescriptorSetIndexLow)
                 {
                     userDataLimit = std::max(userDataLimit, pIntfData->userDataMap[i] + 1);
                 }

llpc/patch/gfx9/chip/llpcGfx9ConfigBuilder.cpp

@@ -33,6 +33,7 @@
 #include "llpcGfx9ConfigBuilder.h"
 #include "llpcPipelineState.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 #include "llvm/Support/CommandLine.h"
 
 #define DEBUG_TYPE "llpc-gfx9-config-builder"
@@ -2839,7 +2840,7 @@ void ConfigBuilder::BuildUserDataConfig(
             if (pIntfData1->userDataMap[i] != InterfaceData::UserDataUnmapped)
             {
                 AppendConfig(startUserData + i, pIntfData1->userDataMap[i]);
-                if ((pIntfData1->userDataMap[i] & DescRelocMagicMask) != DescRelocMagic)
+                if (pIntfData1->userDataMap[i] < VkDescriptorSetIndexLow)
                 {
                     userDataLimit = std::max(userDataLimit, pIntfData1->userDataMap[i] + 1);
                 }

llpc/patch/gfx9/llpcNggLdsManager.h

@@ -35,6 +35,7 @@
 #include "llvm/IR/IRBuilder.h"
 
 #include "llpcInternal.h"
+#include "llpcUtil.h"
 
 namespace Llpc
 {

llpc/patch/llpcPatchCopyShader.cpp

@@ -40,6 +40,7 @@
 #include "llpcPipelineShaders.h"
 #include "llpcPipelineState.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 
 #define DEBUG_TYPE "llpc-patch-copy-shader"
 

llpc/patch/llpcPatchDescriptorLoad.cpp

@@ -35,6 +35,7 @@
 
 #include "llpcPatchDescriptorLoad.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 
 #define DEBUG_TYPE "llpc-patch-descriptor-load"
 

llpc/patch/llpcPatchEntryPointMutate.cpp

@@ -39,6 +39,7 @@
 #include "llpcPatchEntryPointMutate.h"
 #include "llpcPipelineShaders.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 
 #define DEBUG_TYPE "llpc-patch-entry-point-mutate"
 
@@ -674,15 +675,17 @@ FunctionType* PatchEntryPointMutate::GenerateEntryPointType(
                     assert(pNode->sizeInDwords == 1);
 
                     auto pShaderOptions = &m_pPipelineState->GetShaderOptions(m_shaderStage);
-                    if (pShaderOptions->updateDescInElf && (m_shaderStage == ShaderStageFragment))
+                    if (pShaderOptions->emitDescriptorSetIndexInMetadata && (m_shaderStage == ShaderStageFragment))
                     {
-                        // Put set number to register first, will update offset after merge ELFs
-                        // For partial pipeline compile, only fragment shader needs to adjust offset of root descriptor
-                        // If there are more individual shader compile in future, we can add more stages here
-                        pIntfData->userDataMap[userDataIdx] = DescRelocMagic | pNode->innerTable[0].set;
+                        // Record the descriptor set index into the mapping now instead of. This is used during partial
+                        // pipeline compile, when the final offset of descriptor set VA pointers within the PAL
+                        // userdata table are not yet known.
+                        assert(pNode->innerTable[0].set <= VkDescriptorSetIndexHigh - VkDescriptorSetIndexLow);
+                        pIntfData->userDataMap[userDataIdx] = VkDescriptorSetIndexLow + pNode->innerTable[0].set;
                     }
                     else
                     {
+                        // Record the offset into the PAL userdata table.
                         pIntfData->userDataMap[userDataIdx] = pNode->offsetInDwords;
                     }
 

llpc/patch/llpcPatchInOutImportExport.cpp

@@ -42,6 +42,7 @@
 #include "llpcPatchInOutImportExport.h"
 #include "llpcPipelineShaders.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 #include "llpcVertexFetch.h"
 
 #define DEBUG_TYPE "llpc-patch-in-out-import-export"

llpc/patch/llpcPatchNullFragShader.cpp

@@ -39,6 +39,7 @@
 #include "llpcInternal.h"
 #include "llpcPatch.h"
 #include "llpcPipelineState.h"
+#include "llpcUtil.h"
 
 #define DEBUG_TYPE "llpc-patch-null-frag-shader"
 

llpc/patch/llpcPatchPushConstOp.cpp

@@ -37,6 +37,7 @@
 #include "llpcPatchPushConstOp.h"
 #include "llpcPipelineShaders.h"
 #include "llpcPipelineState.h"
+#include "llpcUtil.h"
 
 #define DEBUG_TYPE "llpc-patch-push-const"
 

llpc/patch/llpcSystemValues.cpp

@@ -35,6 +35,7 @@
 #include "llpcPipelineState.h"
 #include "llpcSystemValues.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 
 #define DEBUG_TYPE "llpc-system-values"
 

llpc/patch/llpcVertexFetch.cpp

@@ -35,6 +35,7 @@
 #include "llpcPipelineState.h"
 #include "llpcSystemValues.h"
 #include "llpcTargetInfo.h"
+#include "llpcUtil.h"
 #include "llpcVertexFetch.h"
 
 #define DEBUG_TYPE "llpc-vertex-fetch"

llpc/tool/vfx/vfxSection.h

@@ -1163,7 +1163,7 @@ class SectionShaderOption : public Section
 
         INIT_STATE_MEMBER_NAME_TO_ADDR(SectionShaderOption, forceLoopUnrollCount, MemberTypeInt, false);
         INIT_STATE_MEMBER_NAME_TO_ADDR(SectionShaderOption, useSiScheduler, MemberTypeBool, false);
-        INIT_STATE_MEMBER_NAME_TO_ADDR(SectionShaderOption, updateDescInElf, MemberTypeBool, false);
+        INIT_STATE_MEMBER_NAME_TO_ADDR(SectionShaderOption, emitDescriptorSetIndexInMetadata, MemberTypeBool, false);
         INIT_STATE_MEMBER_NAME_TO_ADDR(SectionShaderOption, allowVaryWaveSize, MemberTypeBool, false);
 #if LLPC_CLIENT_INTERFACE_MAJOR_VERSION >= 33
         INIT_STATE_MEMBER_NAME_TO_ADDR(SectionShaderOption, enableLoadScalarizer, MemberTypeBool, false);