Skip to content

AUDIT.md

Latest commit

 

History

History
87 lines (71 loc) · 5.15 KB

AUDIT.md

File metadata and controls

87 lines (71 loc) · 5.15 KB

⚠️ Historischer Auditbericht – Befunde ohne aktuellen Codebeleg mit <!-- TODO: add source file evidence --> markieren. Veraltete Befunde entfernen.

Audit Report — Acceleration Module

Last Audit: 2026-04-19 Auditor: Copilot Status: ✅ Pass

Summary

Metric Result
Build System Registration ✅ Verified
Source Files 26 (.cpp in src/acceleration/)
Test Coverage ✅ > 80% (confirmed by maintainer, Issue #1398)
Open TODOs 23 files contain TODOs (primarily hardware-conditional stubs)
Open Stubs 1 (CUDAGraphBackend GPU graph traversal — CPU fallback active)
Security Issues None (security hardening pass complete, Issue #1394)

Build System

  • All acceleration source files are registered in cmake/CMakeLists.txt and cmake/ModularBuild.cmake.
  • Conditional compilation guards: THEMIS_ENABLE_CUDA, THEMIS_ENABLE_HIP, THEMIS_ENABLE_VULKAN, THEMIS_ENABLE_OPENCL, THEMIS_ENABLE_OPENGL.
  • BACKEND_CONTRACT_VERSION = 100 guarantees ABI stability — verified in tests/test_backend_api_stability.cpp.
  • CI benchmark gates enforced via .github/workflows/acceleration-benchmark-ci.yml.

Source Files Audited

File Purpose
ai_hardware_dispatcher.cpp AI workload dispatch across heterogeneous hardware
backend_registry.cpp Runtime backend selection and capability negotiation
compute_backend.cpp Base compute backend interface
cpu_backend.cpp Single-threaded CPU fallback backend
cpu_backend_mt.cpp Multi-threaded CPU backend
cpu_backend_tbb.cpp TBB-parallelized CPU backend
cuda/ (6 files) CUDA kernels: ANN, geo, vector, HNSW, graph, Tensor Core
cuda_backend.cpp CUDA backend coordinator with graph capture
device_manager.cpp Device discovery with 60s TTL cache
directx_backend_full.cpp DirectX 12 compute backend
faiss_gpu_backend.cpp FAISS GPU ANN integration
geo_acceleration_bridge.cpp Geo module GPU dispatch bridge
graphics_backends.cpp Vulkan and OpenGL compute backends
hip/ (2 files) ROCm/HIP kernels (ANN and geo)
hip_backend.cpp HIP backend coordinator
multi_gpu_backend.cpp Multi-GPU sharding and fan-out search
nccl_vector_backend.cpp NCCL-based multi-GPU vector operations
oneapi_backend.cpp Intel oneAPI compute backend
opencl_backend.cpp OpenCL fallback backend
plugin_loader.cpp Dynamic backend plugin loading
plugin_security.cpp Plugin signature verification and sandboxing
rccl_vector_backend.cpp RCCL-based multi-GPU vector operations
shader_integrity.cpp GPU shader binary integrity verification
tensor_core_matmul.cpp Tensor Core FP16/BF16 matrix operations
vec_knn.cpp Vectorised k-nearest-neighbour primitives
vllm_resource_manager.cpp vLLM-compatible resource lifecycle management
vulkan_backend_full.cpp Full Vulkan compute backend
zluda_backend.cpp ZLUDA compatibility shim backend

Test Coverage

  • Unit tests: tests/test_acceleration.cpp, tests/test_cuda_ann_search.cpp, tests/test_multi_gpu_backend.cpp, tests/test_cuda_graph_capture.cpp, tests/test_backend_selection_matrix.cpp (65 tests), tests/test_backend_registry_startup.cpp, tests/test_opengl_backend.cpp, tests/test_plugin_security_audit.cpp, tests/test_backend_api_stability.cpp
  • CPU/GPU parity tests: tests/test_cpu_gpu_parity.cpp — GPU paths skip gracefully when hardware is absent
  • Regression tests: null-pointer, zero-dim, k-clamp inputs covered in test_cuda_ann_search.cpp
  • Benchmark harness: benchmarks/bench_cuda_vs_cpu.cpp with JSON output baseline

Findings

Resolved

  • Shell injection surface in plugin signature verification — verifyGPGSignature and verifyMacOSCodeSignature used popen/shell; replaced with posix_spawn+execv and SecStaticCodeCheckValidity (Issue #1394).
  • RTLD_LAZY in plugin loader — replaced with RTLD_NOW to catch unresolved symbols at load time.
  • Missing supportedPrecisions/supportedMetrics in HIPVectorBackend/ZLUDAVectorBackend — fields added; no ABI break.
  • Group/world-writable plugin bypass — file permission check added to PluginLoader::loadPlugin().

Open

  • CUDAGraphBackend GPU graph traversal stub — BFS and Bellman-Ford kernels exist in cuda/graph_kernels.cu but the CUDAGraphBackend for the graph module still delegates to CPU; acceptable because CPU fallback is always active.
  • WASM kernel sandbox — infrastructure is complete but requires a concrete WasmRuntime (Wasmtime/WasmEdge) registration before untrusted kernel execution is operational (Issue #1572).
  • Windows plugin signature verification — not yet implemented; unsigned plugins are accepted on Windows.

Compliance

  • No PII processed directly; vector embeddings are treated as opaque numeric data.
  • VRAM tenant quotas support multi-tenant isolation requirements.
  • Plugin loading hardening aligns with supply-chain security best practices (signed binaries, integrity hashes).