BLAS and TLAS — Acceleration Structures

• Hardware-accelerated BVH for Vulkan ray tracing.
• Parent: PathTracer Learning

Conceptual Overview

• Two-level hierarchy
  • BLAS (Bottom-Level AS): geometry BVH — triangles or AABBs
  • TLAS (Top-Level AS): instance BVH — references to BLASes with transforms
• Why two levels?
  • Allows instancing: one BLAS, many TLAS instances (e.g., 1000 trees from 1 mesh)
  • TLAS can be rebuilt cheaply (just transforms) while BLASes stay static
  • Hardware traversal handles both levels transparently
• Analogy to software BVH
  • BLAS ≈ per-mesh BVH (built offline or at load time)
  • TLAS ≈ scene BVH (rebuilt every frame)

BLAS Construction

• Input: vertex buffer + index buffer (triangle list)
• VkAccelerationStructureGeometryKHR

  geometry.geometryType = VK_GEOMETRY_TYPE_TRIANGLES_KHR;
  geometry.geometry.triangles.vertexFormat = VK_FORMAT_R32G32B32_SFLOAT;
  geometry.geometry.triangles.vertexData.deviceAddress = vertexBufferAddress;
  geometry.geometry.triangles.vertexStride = sizeof(Vertex);
  geometry.geometry.triangles.maxVertex = vertexCount - 1;
  geometry.geometry.triangles.indexType = VK_INDEX_TYPE_UINT32;
  geometry.geometry.triangles.indexData.deviceAddress = indexBufferAddress;
  

• Build flags
  • VK_BUILD_ACCELERATION_STRUCTURE_PREFER_FAST_TRACE_BIT_KHR — optimize for traversal (static geometry)
  • VK_BUILD_ACCELERATION_STRUCTURE_PREFER_FAST_BUILD_BIT_KHR — optimize for build speed (dynamic geometry)
  • VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_UPDATE_BIT_KHR — allow incremental updates (skinned meshes)
  • VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_COMPACTION_BIT_KHR — compact after build (saves memory)
• Memory requirements
  • Query with vkGetAccelerationStructureBuildSizesKHR
  • Returns: accelerationStructureSize, buildScratchSize, updateScratchSize
  • Scratch buffer is temporary — can be reused after build completes
• Compaction (important for memory)
  • Build with ALLOW_COMPACTION flag
  • Query compacted size with VK_QUERY_TYPE_ACCELERATION_STRUCTURE_COMPACTED_SIZE_KHR
  • Copy to smaller buffer with vkCmdCopyAccelerationStructureKHR
  • Typical compaction ratio: 50-70% size reduction

TLAS Construction

• Input: array of VkAccelerationStructureInstanceKHR
• Instance struct fields
  • transform — 3×4 row-major affine transform matrix
  • instanceCustomIndex — 24-bit value accessible as gl_InstanceCustomIndexEXT in shaders
  • mask — 8-bit visibility mask (AND with ray mask in traceRayEXT)
  • instanceShaderBindingTableRecordOffset — SBT hit group offset for this instance
  • flags — VK_GEOMETRY_INSTANCE_TRIANGLE_FACING_CULL_DISABLE_BIT_KHR etc.
  • accelerationStructureReference — device address of the BLAS
• TLAS is rebuilt every frame
  • Upload new instance buffer with updated transforms
  • vkCmdBuildAccelerationStructuresKHR with VK_BUILD_ACCELERATION_STRUCTURE_MODE_BUILD_KHR
  • Or use MODE_UPDATE_KHR for incremental update (faster but lower quality BVH)

Memory Layout and Device Addresses

• PathTracer Learning - Concept - Device Address Bit
• All AS input buffers need VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT
• AS buffer needs VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_STORAGE_BIT_KHR
• Scratch buffer needs VK_BUFFER_USAGE_STORAGE_BIT | VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT

Traversal in Shaders

• traceRayEXT(tlas, rayFlags, cullMask, sbtOffset, sbtStride, missIndex, origin, tMin, dir, tMax, payloadLocation)
• Ray flags
  • gl_RayFlagsOpaqueEXT — skip any-hit shaders (faster)
  • gl_RayFlagsTerminateOnFirstHitEXT — stop at first hit (shadow rays)
  • gl_RayFlagsSkipClosestHitShaderEXT — don’t call closest-hit (occlusion only)
• Built-in variables in hit shaders
  • gl_HitTEXT — distance to hit
  • gl_PrimitiveID — triangle index within the BLAS geometry
  • gl_InstanceID — instance index in the TLAS
  • gl_InstanceCustomIndexEXT — custom index from instance struct
  • gl_WorldRayOriginEXT, gl_WorldRayDirectionEXT — ray in world space
  • gl_ObjectRayOriginEXT, gl_ObjectRayDirectionEXT — ray in object space
  • gl_ObjectToWorldEXT, gl_WorldToObjectEXT — transform matrices