same v0.6.5

same v0.6.5

下载 / Downloads

• Windows x64 and Linux x64: standard includes CPU and optional OpenCL iGPU; cuda additionally includes CUDA. macOS 14+ Apple Silicon has a standard package without CUDA.
• CUDA packages need a compatible NVIDIA driver, but users do not need to install the CUDA compiler. Ordinary CPU scanning does not require a GPU. OpenCL requires a compatible vendor driver.
• Linux packages target Ubuntu 24.04 / glibc 2.39+; macOS packages are not notarized. Archives include licenses and build-info.json. Verify the checksums in SHA256SUMS and the GitHub build provenance before use.

更新 / Changes since v0.6.0

• Every content worker can independently choose CPU SIMD, CUDA, or iGPU. The iGPU backend no longer holds a pool-wide admission lock for a whole file task: workers own separate OpenCL queues, kernels, and device buffers while sharing compiled program/context. Shared host and device memory budgets now account for all configured workers, and a recoverable iGPU failure retires only its owner. The first automatic GPU discovery/startup remains cautious to avoid delaying short CPU tasks.
• Online routing now fits its contention feature using overlap observed during execution rather than treating the selection-time peer count as the completed task's exposure. Prediction-error reporting still evaluates the estimate frozen when the decision was made. New iGPU peak/congestion diagnostics describe admitted file tasks, not guaranteed simultaneous GPU kernel execution or a speedup.
• Learned-state identity changes from same-learning-key-v1 to same-learning-key-v2 because the contention feature's training semantics changed. Existing v0.6.0 model rows remain in .same/model.db, but they are not mixed into the new model. The first v0.6.5 scan in an existing workspace therefore starts with fresh routing priors; later runs reuse v2 priors, or same train can calibrate them explicitly. Summary/TSV output reports the new model identity.
• same train, same merge, the command-line options, the default nonrecursive scan, and bytewise duplicate verification retain their v0.6.0 behavior. No migration or deletion of workspace data is required.

验证与边界 / Validation and limits

Windows MSVC Debug CTest passed 27/27, including cross-run migration isolation. A physical Windows OpenCL/CUDA smoke test with three distinct 32 MiB files and three workers observed igpu_peak_concurrency=3 without fallback; forced CPU, iGPU, and CUDA scans yielded identical BLAKE3 digests. The automatic scan preferred CPU on this cold, small corpus, which was faster than starting the accelerators. Existing v0.6.0 model rows were retained while new v2 rows were created and reused on a subsequent run. GitHub Actions also runs CPU, OpenCL-kernel, sanitizer, telemetry, and the five tested release packages across Windows, Linux, and macOS before this tag is published.

Multiple OpenCL queues permit independent submission but do not guarantee physical GPU parallelism or whole-scan speedup. Throughput depends on the device, driver, workload, memory contention, and cold-start cost. The online overlap feature is a proxy, not a measurement of overlap duration or memory bandwidth. Hosted CI has no physical GPU, so the hardware observation above comes from the local smoke test rather than every release platform.