Update gpurt from commit 80269d10

Remove unused path from BuildParallel
Set triangle 1 bits in triangle ID only when it is a pair compressed triangle
[Continuations] Change static ID handling for continuations
Fix tdr in rayquery apps when rdp attaches
Deprecate CmdBarrier() path
rename BuildBVHPLOC to BuildPLOC
Fix cmake defines and includes for validation
Add deviceSetting to disable compaction
Enable GPU debugging in all build stages
Move cpp-shared raytracingdefs.h into non-shared ShaderDefs.hlsli
Reduce CopyAS.hlsl dependencies
Move bit-related utils from Math to Bits.hlsli
Replace bufferView with typed and untyped BufferView
Continuations persistent launch support
[Continuations] Fix legacy compilation
Initialize parentId to -1 for ray history counter
Skip redundant copy in merge-sort iteration
Copy Indirect Args in InitExecuteIndirect
[Continuations] Remove redundant repacking of constant known ray flags
[Continuations] Add options to override TraceRay flags
Fix barrier corner cases
Softcode validation file extension and directory
[Continuations] Remove stack lowering guard
Add spirv pass to validation of clean shaders
Do not limit number of waves per simd for the encode path
Recompute the dispatchID when threadGroupSize != 32
Separate merge sort local/global dispatches

backends/pal/gpurtPalBackend.cpp

@@ -167,7 +167,9 @@ uint32 PalBackend::GetMaxDescriptorTableSize(
     ClientCmdBufferHandle cmdBuffer
     ) const
 {
-    const uint32 bufferSrdSizeDw = m_deviceProperties.gfxipProperties.srdSizes.bufferView / sizeof(uint32);
+    const uint32 bufferSrdSizeDw = Util::Max(m_deviceProperties.gfxipProperties.srdSizes.typedBufferView,
+                                             m_deviceProperties.gfxipProperties.srdSizes.untypedBufferView) /
+                                   sizeof(uint32);
     return GetCmdBuffer(cmdBuffer)->GetLargeEmbeddedDataLimit() / bufferSrdSizeDw;
 }
 
@@ -239,81 +241,39 @@ void PalBackend::InsertBarrier(
     const bool syncPostCpWrite  = flags & BarrierFlagSyncPostCpWrite;
 
     Pal::ICmdBuffer* pCmdBuffer = GetCmdBuffer(cmdBuffer);
-    if (m_deviceSettings.enableAcquireReleaseInterface)
-    {
-        Pal::AcquireReleaseInfo acqRelInfo  = {};
-        Pal::MemBarrier memoryBarrier       = {};
-
-        if (syncDispatch || syncIndirectArgs)
-        {
-            memoryBarrier.srcStageMask  = Pal::PipelineStageCs;
-            memoryBarrier.srcAccessMask = Pal::CoherShader;
-        }
-
-        if (syncPostCpWrite)
-        {
-            memoryBarrier.srcStageMask  |= Pal::PipelineStagePostPrefetch;
-            memoryBarrier.srcAccessMask |= Pal::CoherCp;
-        }
-
-        if (syncDispatch || syncPostCpWrite)
-        {
-            memoryBarrier.dstStageMask  = Pal::PipelineStageCs;
-            memoryBarrier.dstAccessMask = Pal::CoherShader;
-        }
-
-        if (syncIndirectArgs)
-        {
-            memoryBarrier.dstStageMask  |= Pal::PipelineStageFetchIndirectArgs;
-            memoryBarrier.dstAccessMask |= Pal::CoherIndirectArgs;
-        }
-
-        acqRelInfo.memoryBarrierCount = 1;
-        acqRelInfo.pMemoryBarriers    = &memoryBarrier;
-        acqRelInfo.reason             = m_deviceSettings.rgpBarrierReason;
-
-        pCmdBuffer->CmdReleaseThenAcquire(acqRelInfo);
-    }
-    else
-    {
-        Pal::BarrierInfo barrierInfo = {};
 
-        const uint32 pipePointCount  = (syncDispatch || syncIndirectArgs) ? 1 : 0;
-        Pal::HwPipePoint pipePoint   = Pal::HwPipePostCs;
+    Pal::AcquireReleaseInfo acqRelInfo  = {};
+    Pal::MemBarrier memoryBarrier       = {};
 
-        Pal::BarrierTransition transition = {};
-
-        if (syncDispatch)
-        {
-            transition.srcCacheMask = Pal::CoherShader;
-        }
-
-        if (syncPostCpWrite)
-        {
-            transition.srcCacheMask |= Pal::CoherCp;
-        }
+    if (syncDispatch || syncIndirectArgs)
+    {
+        memoryBarrier.srcStageMask  = Pal::PipelineStageCs;
+        memoryBarrier.srcAccessMask = Pal::CoherShader;
+    }
 
-        if (syncDispatch || syncPostCpWrite)
-        {
-            barrierInfo.waitPoint   = Pal::HwPipePreCs;
-            transition.dstCacheMask = Pal::CoherShader;
-        }
+    if (syncPostCpWrite)
+    {
+        memoryBarrier.srcStageMask  |= Pal::PipelineStagePostPrefetch;
+        memoryBarrier.srcAccessMask |= Pal::CoherCp;
+    }
 
-        if (syncIndirectArgs)
-        {
-            barrierInfo.waitPoint    = Pal::HwPipeTop;
-            transition.dstCacheMask |= Pal::CoherIndirectArgs;
-        }
+    if (syncDispatch || syncPostCpWrite)
+    {
+        memoryBarrier.dstStageMask  = Pal::PipelineStageCs;
+        memoryBarrier.dstAccessMask = Pal::CoherShader;
+    }
 
-        barrierInfo.pipePointWaitCount  = pipePointCount;
-        barrierInfo.pPipePoints         = &pipePoint;
-        barrierInfo.transitionCount     = 1;
-        barrierInfo.pTransitions        = &transition;
+    if (syncIndirectArgs)
+    {
+        memoryBarrier.dstStageMask  |= Pal::PipelineStageFetchIndirectArgs;
+        memoryBarrier.dstAccessMask |= Pal::CoherIndirectArgs;
+    }
 
-        barrierInfo.reason = m_deviceSettings.rgpBarrierReason;
+    acqRelInfo.memoryBarrierCount = 1;
+    acqRelInfo.pMemoryBarriers    = &memoryBarrier;
+    acqRelInfo.reason             = m_deviceSettings.rgpBarrierReason;
 
-        pCmdBuffer->CmdBarrier(barrierInfo);
-    }
+    pCmdBuffer->CmdReleaseThenAcquire(acqRelInfo);
 }
 
 // =====================================================================================================================
@@ -324,7 +284,11 @@ void PalBackend::CreateBufferViewSrds(
     bool                  isTyped
     ) const
 {
-    const uint32 bufferSrdSizeDw = m_deviceProperties.gfxipProperties.srdSizes.bufferView / sizeof(uint32);
+    const uint32 bufferSrdSizeDw = ((isTyped) ?
+                                    m_deviceProperties.gfxipProperties.srdSizes.typedBufferView :
+                                    m_deviceProperties.gfxipProperties.srdSizes.untypedBufferView)
+                                   / sizeof(uint32);
+
     const Pal::BufferViewInfo palBufferViewInfo = ConvertBufferViewToPalBufferView(bufferViewInfo);
     const void* pNullBuffer = m_deviceProperties.gfxipProperties.nullSrds.pNullBufferView;
 

gpurt/gpurt.h

@@ -285,7 +285,10 @@ enum class InternalRayTracingCsType : uint32
     BuildBVH,
     BuildBVHTD,
     BuildBVHTDTR,
-    BuildBVHPLOC,
+    BuildPLOC,
+#if GPURT_CLIENT_INTERFACE_MAJOR_VERSION < 52
+    BuildBVHPLOC = BuildPLOC,
+#endif
     UpdateQBVH,
     UpdateParallel,
     RefitBounds,
@@ -311,6 +314,9 @@ enum class InternalRayTracingCsType : uint32
     InitExecuteIndirect,
     PairCompression,
     MergeSort,
+    MergeSortLocal,
+    MergeSortGlobalIteration,
+    MergeSortCopyLastLevel,
     UpdateTriangles,
     UpdateAabbs,
     InitAccelerationStructure,
@@ -753,7 +759,9 @@ struct DeviceSettings
         uint32 enableParallelUpdate : 1;
         uint32 enableParallelBuild : 1;
         uint32 enablePrefixScanDLB : 1;
+#if GPURT_CLIENT_INTERFACE_MAJOR_VERSION < 51
         uint32 enableAcquireReleaseInterface : 1;
+#endif
         uint32 enableBuildAccelStructDumping : 1;
         uint32 enableBuildAccelStructScratchDumping : 1;
         uint32 enableBuildAccelStructStats : 1;
@@ -779,6 +787,7 @@ struct DeviceSettings
 
         uint32 enableRemapScratchBuffer : 1;                // Enable remapping bvh2 data from ScratchBuffer to ResultBuffer
         uint32 checkBufferOverlapsInBatch : 1;
+        uint32 disableCompaction : 1;                       // Reports and perform copy instead of compaction
     };
 
     uint64                      accelerationStructureUUID;  // Acceleration Structure UUID

gpurt/gpurtBuildSettings.h

@@ -69,7 +69,7 @@ struct CompileTimeBuildSettings
     uint32 enableTopDownBuild;
     uint32 useMortonCode30;
     uint32 enableMergeSort;
-    uint32 fastBuildThreshold;
+    uint32 unused14;
     uint32 enableFusedInstanceNode;
     float  tsPriority;
     uint32 numRebraidIterations;
@@ -99,7 +99,7 @@ struct CompileTimeBuildSettings
     uint32 unused11;
     uint32 unused12;
     uint32 unused13;
-    uint32 rebuildAccelStruct;
+    uint32 disableCompaction;
 };
 
 #define BUILD_SETTINGS_DATA_TOP_LEVEL_BUILD_ID                        0
@@ -119,7 +119,7 @@ struct CompileTimeBuildSettings
 #define BUILD_SETTINGS_DATA_ENABLE_TOP_DOWN_BUILD_ID                  14
 #define BUILD_SETTINGS_DATA_USE_MORTON_CODE_30_ID                     15
 #define BUILD_SETTINGS_DATA_ENABLE_MERGE_SORT_ID                      16
-#define BUILD_SETTINGS_DATA_FAST_BUILD_THRESHOLD_ID                   17
+// unused14 id                                                          17
 #define BUILD_SETTINGS_DATA_ENABLE_FUSED_INSTANCE_NODE_ID             18
 #define BUILD_SETTINGS_DATA_TS_PRIORITY_ID                            19
 #define BUILD_SETTINGS_DATA_NUM_REBRAID_ITERATIONS_ID                 20
@@ -135,7 +135,7 @@ struct CompileTimeBuildSettings
 #define BUILD_SETTINGS_DATA_ENCODE_ARRAY_OF_POINTERS_ID               41
 #define BUILD_SETTINGS_DATA_SCENE_BOUNDS_CALCULATION_TYPE_ID          42
 #define BUILD_SETTINGS_DATA_REBRAID_QUALITY_HEURISTIC_ID              43
-#define BUILD_SETTINGS_DATA_REBUILD_ACCELERATION_STRUCTURE_ID         47
+#define BUILD_SETTINGS_DATA_DISABLE_COMPACTION_ID                     47
 
 #ifdef __cplusplus
 } // namespace GpuRt

gpurt/gpurtDispatch.h

@@ -70,11 +70,11 @@ struct DispatchRaysConstantData
     uint32 missTableBaseAddressLo;      // Miss shader table base address low 32-bits
     uint32 missTableBaseAddressHi;      // Miss shader table base address high 32-bits
     uint32 missTableStrideInBytes;      // Miss shader table record byte stride
-    uint32 reserved0;                   // Reserved padding
+    uint32 rayDispatchMaxGroups;        // Max groups dispatched if persistent launch is enabled, else 0
     uint32 hitGroupTableBaseAddressLo;  // Hit group table base address low 32-bits
     uint32 hitGroupTableBaseAddressHi;  // Hit group table base address high 32-bits
     uint32 hitGroupTableStrideInBytes;  // Hit group table record byte stride
-    uint32 reserved1;                   // Reserved padding
+    uint32 reserved0;                   // Reserved padding
     uint32 callableTableBaseAddressLo;  // Callable shader table base address low 32-bits
     uint32 callableTableBaseAddressHi;  // Callable shader table base address high 32-bits
     uint32 callableTableStrideInBytes;  // Callable shader table byte stride
@@ -146,6 +146,7 @@ struct InitExecuteIndirectConstants
     uint32 rtThreadGroupSizeX;      // Internal RT threadgroup size X
     uint32 rtThreadGroupSizeY;      // Internal RT threadgroup size Y
     uint32 rtThreadGroupSizeZ;      // Internal RT threadgroup size Z
+    uint32 rayDispatchMaxGroups;    // Max groups dispatched if persistent launch is enabled, else 0
     uint32 counterMask;             // Mask for filtering ray history token
     uint32 pipelineCount;           // Number of pipelines to launch (1 for indirect launch, raygen count for unified)
     uint32 maxIterations;           // Max traversal interations for profiling
@@ -160,7 +161,6 @@ struct InitExecuteIndirectConstants
     uint32 counterRayIdRangeEnd;    // Counter ray ID range end
     uint32 cpsBackendStackSize;     // Scratch memory used by a compiler backend, start at offset 0
     uint32 padding0;                // Padding for 16-byte alignment
-    uint32 padding1;                // Padding for 16-byte alignment
 
 #if __cplusplus
      // Internal counter buffer SRDs

gpurt/gpurtLib.h

@@ -42,7 +42,7 @@ namespace GpuRt
 // update their definition of GPURT_CLIENT_INTERFACE_MAJOR_VERSION to indicate that they have made the required changes
 // to support a new version. When the client version is updated, the old interface will be compiled out and only the
 // new one will remain.
-#define GPURT_INTERFACE_MAJOR_VERSION 49
+#define GPURT_INTERFACE_MAJOR_VERSION 52
 
 #if GPURT_CLIENT_INTERFACE_MAJOR_VERSION < 44
 // Minor interface version. This number is incrememnted when a compatible interface change is made. Compatible changes

src/gpurtBvhBatcher.cpp

@@ -200,7 +200,14 @@ void BvhBatcher::BuildRaytracingAccelerationStructureBatch(
     }
     if ((updaters.IsEmpty() == false) || (builders.IsEmpty() == false))
     {
-        Barrier();
+        uint32 barrierFlags = BarrierFlagSyncDispatch;
+        if (updaters.IsEmpty() == false)
+        {
+            // Updates can be launched with indirect dispatch. We need to avoid fetching the indirect arguments
+            // from the header before they are written by a previous build/update/copy.
+            barrierFlags |= BarrierFlagSyncIndirectArg;
+        }
+        Barrier(barrierFlags);
     }
     RGP_POP_MARKER();
 
@@ -304,11 +311,72 @@ void BvhBatcher::BuildMultiDispatch(Util::Span<BvhBuilder> builders)
     if (PhaseEnabled(BuildPhaseFlags::MergeSort))
     {
         Barrier();
-        const uint32 wavesPerSimd = builders.size() == 1 ? 16U : 2U;
-        BuildFunction(BuildPhaseFlags::MergeSort, builders, [wavesPerSimd](BvhBuilder& builder)
+
+        if (builders.size() > 1)
         {
-            builder.MergeSort(wavesPerSimd);
-        });
+            const uint32 wavesPerSimd = 2U;
+            BuildFunction(BuildPhaseFlags::MergeSort, builders, [wavesPerSimd](BvhBuilder& builder)
+            {
+                builder.MergeSort(wavesPerSimd);
+            });
+        }
+        else
+        {
+            RGP_PUSH_MARKER("Merge Sort");
+
+            // Batch local sorts together.
+            BuildPhase("Merge Sort (Local)", builders, &BvhBuilder::MergeSortLocal);
+
+            Barrier();
+
+            // Batch global sort iterations together. Compute max iterations amongst the builder batch
+            uint32 maxMergeSortTreeLevel = 0;
+
+            bool batchNeedsLastLevelCopy = false;
+
+            for (const auto& builder : builders)
+            {
+                const uint32 mergeSortTreeLevel = builder.GetMaxMergeSortTreeLevel();
+                maxMergeSortTreeLevel = Util::Max(maxMergeSortTreeLevel, mergeSortTreeLevel);
+                batchNeedsLastLevelCopy |= ((mergeSortTreeLevel & 1) == 1);
+            }
+
+            if (maxMergeSortTreeLevel > 0)
+            {
+                RGP_PUSH_MARKER("Merge Sort (Global Iteration)");
+                for (uint32 level = 1; level <= maxMergeSortTreeLevel; level++)
+                {
+                    Barrier();
+
+                    BuildFunction(nullptr, builders, [level](BvhBuilder& builder)
+                    {
+                        if (level <= builder.GetMaxMergeSortTreeLevel())
+                        {
+                            builder.MergeSortGlobalIteration(level);
+                        }
+                    });
+                }
+                RGP_POP_MARKER();
+
+                if (batchNeedsLastLevelCopy)
+                {
+                    Barrier();
+
+                    RGP_PUSH_MARKER("Merge Sort (Copy Last Level)");
+                    BuildFunction(nullptr, builders, [](BvhBuilder& builder)
+                    {
+                        const uint32 mergeSortTreeLevel = builder.GetMaxMergeSortTreeLevel();
+                        if ((mergeSortTreeLevel & 1) == 1)
+                        {
+                            builder.MergeSortCopyLastLevel();
+                        }
+                    });
+                    RGP_POP_MARKER();
+                }
+            }
+
+            RGP_POP_MARKER();
+        }
     }
     if (PhaseEnabled(BuildPhaseFlags::RadixSort))
     {
@@ -327,13 +395,13 @@ void BvhBatcher::BuildMultiDispatch(Util::Span<BvhBuilder> builders)
         Barrier();
         BuildPhase(BuildPhaseFlags::BuildFastAgglomerativeLbvh, builders, &BvhBuilder::BuildFastAgglomerativeLbvh);
     }
-    if (PhaseEnabled(BuildPhaseFlags::BuildBVHPLOC))
+    if (PhaseEnabled(BuildPhaseFlags::BuildPLOC))
     {
         Barrier();
         const uint32 wavesPerSimd = builders.size() == 1 ? 8U : 1U;
-        BuildFunction(BuildPhaseFlags::BuildBVHPLOC, builders, [wavesPerSimd](BvhBuilder& builder)
+        BuildFunction(BuildPhaseFlags::BuildPLOC, builders, [wavesPerSimd](BvhBuilder& builder)
         {
-            builder.BuildBVHPLOC(wavesPerSimd);
+            builder.BuildPLOC(wavesPerSimd);
         });
     }
     if (PhaseEnabled(BuildPhaseFlags::RefitBounds))