Update gpurt from commit ce49ca42

Support pull UserMarker into RRA
[Continuations] Remove _cont_SetupRayGen
Report 0 update scratch size for non-updatable AS
Add HLSL version of static_assert
BatchedBuilder: Remove TLAS serialization workaround
Add ganged queue check
Separate CPS implementation per RtIp level
BatchedBuilder: Support enableInsertBarriersInBuildAS setting
Disable triangle pairing of inactive triangles
BatchedBuilder: Add debug option to check for buffer overlaps
Use packedInstanceContribution for RT_20
Shader Cleanliness Enforcement Script
Enable merged EncodeUpdate for VK
Update RayQuery intrinsic in GPURT library
Unifying dxc options
Unify RootSignature
Canonicalize paths used in validation comparison
Enable dxc optimizations for VK/SPV
Remove glslang compilation path
Fix HLSL root signature mapping for compile time CBV
Fix refit issues with fast LBVH enabled
[Continuations] Restrict RtIp 1_1 and 2_0 Traversal variant compilation
Unify triangle data fetching code, add indirect builds support to trivial builder and STGB, and fixes
[Continuations] Optimize opaque RtIp2_0 Traversal path
[Continuations] Add CpsCandidatePrimitiveMode option
[Continuations] Implement deferred anyhit shading mode
BatchedBuilder: minor fixups
Fix issues with cputraversal and update scratch size

backends/pal/gpurtPalBridge.cpp

@@ -85,13 +85,20 @@ size_t GPURT_API_ENTRY GetDeviceSize()
 
 // =====================================================================================================================
 PipelineShaderCode GPURT_API_ENTRY GetShaderLibraryCode(
-    ShaderLibraryFeatureFlags flags)
+#if GPURT_CLIENT_INTERFACE_MAJOR_VERSION >= 48
+    const Pal::RayTracingIpLevel rayTracingIpLevel,
+#endif
+    ShaderLibraryFeatureFlags    flags)
 {
 #if GPURT_CLIENT_INTERFACE_MAJOR_VERSION < 40
     GPURT_EXPORT_UNMANGLED_SYMBOL_MSVC
 #endif
 
+#if GPURT_CLIENT_INTERFACE_MAJOR_VERSION >= 48
+    return Internal::GetShaderLibraryCode(rayTracingIpLevel, flags);
+#else
     return Internal::GetShaderLibraryCode(flags);
+#endif
 }
 
 // =====================================================================================================================

cmake/GpuRtGenerateShaders.cmake

@@ -131,6 +131,16 @@ if(GPURT_CLIENT_API STREQUAL "VULKAN")
             ${gpurtDxcCompiler}
             ${gpurtSpirvRemap}
 
+        COMMAND Python3::Interpreter "${gpurtCompileScript}"
+            --outputDir "${gpurtOutputDir}"
+            --validateShadersClean
+            ${COMPILER_ARGUMENT}
+            --defines "\"${gpurtDefines}\""
+            --includePaths "\"${gpurtIncludeDirectories}\""
+            "${gpurtDxilBvhShader}"
+            "${gpurtShadersSourceDir}"
+            "${gpurtSscStrict}"
+
         COMMAND Python3::Interpreter "${gpurtCompileScript}"
             --vulkan
             "${SPIRV_FLAG}"

gpurt/gpurt.h

@@ -373,6 +373,15 @@ enum class AccelStructBuilderType : uint32
 static_assert(uint32(AccelStructBuilderType::Gpu) == 0, "Enums encoded in the acceleration structure must not change.");
 static_assert(uint32(AccelStructBuilderType::Cpu) == 1, "Enums encoded in the acceleration structure must not change.");
 
+// Modes for when to perform a rebuild instead of an update
+enum ForceRebuildForUpdatesMode : uint32
+{
+    None        = 0x0,  // Disable build flag overrides
+    TopLevel    = 0x1,  // Override flags for top level acceleration structures only
+    BottomLevel = 0x2,  // Override flags for bottom level acceleration structures only
+    All         = 0x3,  // Override flags for all acceleration structure updates
+};
+
 // Modes for which interior box nodes in BLAS are written as fp16
 enum class Fp16BoxNodesInBlasMode : uint32
 {
@@ -732,6 +741,8 @@ struct DeviceSettings
     Fp16BoxNodesInBlasMode      fp16BoxNodesInBlasMode;               // Mode for which interior nodes in BLAS are FP16
     float                       fp16BoxModeMixedSaThresh;             // For fp16 mode "mixed", surface area threshold
 
+    ForceRebuildForUpdatesMode  forceRebuildForUpdates;               // When to perform a rebuild instead of an update
+
     Pal::RayTracingIpLevel      emulatedRtIpLevel;                    // Client request RTIP level, used to override IP level related GPURT settings.
 
     struct
@@ -767,6 +778,7 @@ struct DeviceSettings
         uint32 enableFastLBVH : 1;                          // Enable the Fast LBVH path
 
         uint32 enableRemapScratchBuffer : 1;                // Enable remapping bvh2 data from ScratchBuffer to ResultBuffer
+        uint32 checkBufferOverlapsInBatch : 1;
     };
 
     uint64                      accelerationStructureUUID;  // Acceleration Structure UUID
@@ -1019,6 +1031,8 @@ struct RtDispatchInfo
     ShaderTable missShaderTable;
     ShaderTable hitGroupTable;
     ShaderTable callableShaderTable;
+
+    uint64 userMarkerContext;
 };
 
 #if GPURT_DEVELOPER
@@ -1389,7 +1403,10 @@ size_t GPURT_API_ENTRY GetDeviceSize();
 // @return Shader code for the shader library
 //
 PipelineShaderCode GPURT_API_ENTRY GetShaderLibraryCode(
-    ShaderLibraryFeatureFlags flags);
+#if GPURT_CLIENT_INTERFACE_MAJOR_VERSION >= 48
+    const Pal::RayTracingIpLevel rayTracingIpLevel,
+#endif
+    ShaderLibraryFeatureFlags    flags);
 
 // =====================================================================================================================
 // Returns GPURT shader library function table for input ray tracing IP level.
@@ -1601,6 +1618,9 @@ class IDevice
     virtual const ClientCallbacks& GetClientCallbacks() const = 0;
     virtual const DeviceInitInfo& GetInitInfo() const = 0;
 
+    // Check if a build is a good candidate for ACE offload (typically barrier-free cases)
+    virtual bool ShouldUseGangedAceForBuild(const AccelStructBuildInputs& inputs) const = 0;
+
 protected:
 
     /// Client must create objects by explicitly calling CreateDevice method

gpurt/gpurtAccelStruct.h

@@ -48,13 +48,11 @@
 #define GPURT_ACCEL_STRUCT_MINOR_VERSION 3
 #define GPURT_ACCEL_STRUCT_VERSION       ((GPURT_ACCEL_STRUCT_MAJOR_VERSION << 16) | GPURT_ACCEL_STRUCT_MINOR_VERSION)
 
+#include "../src/shared/assert.h"
+
 #ifdef __cplusplus
 namespace GpuRt
 {
-
-#define GPURT_STATIC_ASSERT(condition, message) static_assert(condition, message)
-#else
-#define GPURT_STATIC_ASSERT(condition, message)
 #endif
 
 typedef uint32_t uint32;
@@ -76,10 +74,12 @@ struct AccelStructDataOffsets
 #define ACCEL_STRUCT_OFFSETS_PRIM_NODE_PTRS_OFFSET       12
 #define ACCEL_STRUCT_OFFSETS_SIZE                        16
 
+#ifdef __cplusplus
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_OFFSETS_INTERNAL_NODES_OFFSET == offsetof(AccelStructDataOffsets, internalNodes), "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_OFFSETS_LEAF_NODES_OFFSET == offsetof(AccelStructDataOffsets, leafNodes), "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_OFFSETS_GEOMETRY_INFO_OFFSET == offsetof(AccelStructDataOffsets, geometryInfo), "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_OFFSETS_PRIM_NODE_PTRS_OFFSET == offsetof(AccelStructDataOffsets, primNodePtrs), "");
+#endif
 GPURT_STATIC_ASSERT(sizeof(AccelStructDataOffsets) == 16,
     "AccelStructDataOffsets size cannot change because it is embedded in AccelStructHeader.");
 
@@ -105,10 +105,12 @@ struct AccelStructMetadataHeader
 #define ACCEL_STRUCT_METADATA_HEADER_SIZE               84
 
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_METADATA_HEADER_SIZE == sizeof(AccelStructMetadataHeader), "Acceleration structure header mismatch");
+#ifdef __cplusplus
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_METADATA_VA_LO_OFFSET == offsetof(AccelStructMetadataHeader, addressLo), "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_METADATA_VA_HI_OFFSET == offsetof(AccelStructMetadataHeader, addressHi), "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_METADATA_SIZE_OFFSET == offsetof(AccelStructMetadataHeader, sizeInBytes), "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_METADATA_TASK_COUNTER_OFFSET == offsetof(AccelStructMetadataHeader, taskCounter), "");
+#endif
 
 #ifdef __cplusplus
 // =====================================================================================================================
@@ -287,13 +289,14 @@ struct AccelStructHeader
 
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_SIZE == sizeof(AccelStructHeader),
     "Acceleration structure header mismatch");
+#ifdef __cplusplus
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_INFO_OFFSET                    == offsetof(AccelStructHeader, info),                 "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_METADATA_SIZE_OFFSET           == offsetof(AccelStructHeader, metadataSizeInBytes),  "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_BYTE_SIZE_OFFSET               == offsetof(AccelStructHeader, sizeInBytes),          "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_NUM_PRIMS_OFFSET               == offsetof(AccelStructHeader, numPrimitives),        "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_NUM_ACTIVE_PRIMS_OFFSET        == offsetof(AccelStructHeader, numActivePrims),       "");
 
-GPURT_STATIC_ASSERT(ACCEL_STRUCT_RESERVERED0_OFFSET                   == offsetof(AccelStructHeader, reserved0),             "");
+GPURT_STATIC_ASSERT(ACCEL_STRUCT_RESERVERED0_OFFSET                     == offsetof(AccelStructHeader, reserved0),             "");
 
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_NUM_DESCS_OFFSET               == offsetof(AccelStructHeader, numDescs),             "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_GEOMETRY_TYPE_OFFSET           == offsetof(AccelStructHeader, geometryType),         "");
@@ -309,6 +312,7 @@ GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_FP32_ROOT_BOX_OFFSET           == offset
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_INFO_2_OFFSET                  == offsetof(AccelStructHeader, info2),                "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_PACKED_FLAGS_OFFSET            == offsetof(AccelStructHeader, packedFlags),          "");
 GPURT_STATIC_ASSERT(ACCEL_STRUCT_HEADER_COMPACTED_BYTE_SIZE_OFFSET     == offsetof(AccelStructHeader, compactedSizeInBytes), "");
+#endif
 
 #ifdef __cplusplus
 // =====================================================================================================================

gpurt/gpurtBuildSettings.h

@@ -98,6 +98,7 @@ struct CompileTimeBuildSettings
     uint32 rebraidQualityHeuristic;
     uint32 unused11;
     uint32 unused12;
+    uint32 unused13;
 };
 
 #define BUILD_SETTINGS_DATA_TOP_LEVEL_BUILD_ID                        0

gpurt/gpurtCounter.h

@@ -531,6 +531,7 @@ enum class RayHistoryMetadataKind : uint32
     CounterInfo        = 1,
     Unused             = 2,
     TraversalFlags     = 3,
+    UserMarkerInfo     = 4
 };
 
 // ====================================================================================================================

gpurt/gpurtDispatch.h

@@ -41,12 +41,18 @@ typedef uint64_t uint64;
 #define constexpr static const
 #endif
 
+#if GPURT_CLIENT_INTERFACE_MAJOR_VERSION >= 47
+constexpr uint32 MaxBufferSrdSize = 8;
+#else
+constexpr uint32 MaxBufferSrdSize = 4;
+#endif
+
 // Dispatch rays arguments top-level descriptor table (GPU structure)
 struct DispatchRaysTopLevelData
 {
-    uint64 dispatchRaysConstGpuVa;   // DispatchRays info constant buffer GPU VA
-    uint32 internalUavBufferSrd[4];  // Internal UAV shader resource descriptor
-    uint32 accelStructTrackerSrd[4]; // Structured buffer SRD pointing to the accel struct tracker
+    uint64 dispatchRaysConstGpuVa;                  // DispatchRays info constant buffer GPU VA
+    uint32 internalUavBufferSrd[MaxBufferSrdSize];  // Internal UAV shader resource descriptor
+    uint32 accelStructTrackerSrd[MaxBufferSrdSize]; // Structured buffer SRD pointing to the accel struct tracker
 };
 
 // Dispatch rays constant buffer data (GPU structure). Note, using unaligned uint64_t in HLSL constant buffers requires
@@ -108,7 +114,6 @@ constexpr uint32 DispatchRaysConstantsDw = sizeof(DispatchRaysConstants) / sizeo
 #endif
 
 constexpr uint32 MaxSupportedIndirectCounters = 8;
-constexpr uint32 MaxBufferSrdSize = 4;
 
 // Resource bindings required for InitExecuteIndirect
 struct InitExecuteIndirectUserData
@@ -153,9 +158,9 @@ struct InitExecuteIndirectConstants
     uint32 counterMode;             // Counter mode
     uint32 counterRayIdRangeBegin;  // Counter ray ID range begin
     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
-    uint32 padding2;                // Padding for 16-byte alignment
 
 #if __cplusplus
      // Internal counter buffer SRDs
@@ -164,16 +169,19 @@ struct InitExecuteIndirectConstants
     // Internal acceleration structure tracker buffer SRD.
     uint32 accelStructTrackerSrd[MaxBufferSrdSize];
 #else
-    uint4 internalUavSrd[MaxSupportedIndirectCounters];
-    uint4 accelStructTrackerSrd;
+    uint4 internalUavSrd[MaxSupportedIndirectCounters][MaxBufferSrdSize / 4];
+    uint4 accelStructTrackerSrd[MaxBufferSrdSize / 4];
 #endif
-
 };
 
 constexpr uint32 InitExecuteIndirectConstantsDw = sizeof(InitExecuteIndirectConstants) / sizeof(uint32);
 
 #if __cplusplus
+#if GPURT_CLIENT_INTERFACE_MAJOR_VERSION >= 47
+static_assert((MaxBufferSrdSize == 8), "Buffer SRD size changed, affected shaders and constants need update");
+#else
 static_assert((MaxBufferSrdSize == 4), "Buffer SRD size changed, affected shaders and constants need update");
+#endif
 static_assert((sizeof(InitExecuteIndirectConstants) % sizeof(uint32)) == 0,
               "InitExecuteIndirectConstants is not dword-aligned");
 }

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 46
+#define GPURT_INTERFACE_MAJOR_VERSION 49
 
 #if GPURT_CLIENT_INTERFACE_MAJOR_VERSION < 44
 // Minor interface version. This number is incrememnted when a compatible interface change is made. Compatible changes