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.

same v0.6.0

same v0.6.0

下载 / Downloads

• Windows x64 and Linux x64: standard includes CPU and optional OpenCL iGPU; cuda additionally includes CUDA.
• macOS Apple Silicon: standard includes CPU and optional OpenCL iGPU; CUDA is unavailable.
• CUDA packages require a compatible NVIDIA driver, not a local CUDA compiler. A GPU is not required for ordinary CPU scanning.
• Windows targets Windows 10/11 x64; Linux packages are built on Ubuntu 24.04 (glibc 2.39+); macOS targets macOS 14+ ARM64 and is not notarized.
• Verify SHA256SUMS and build provenance. Archives include licenses and build-info.json identifying the exact commit and backend configuration.

更新 / Changes

• Added same train <dirs> [options]: select a bounded, size-stratified corpus across multiple directories and measure CPU, CUDA and iGPU on each selected file. With no directory, the corpus defaults to .; the former --corpus=DIR option remains supported. Training updates the same online model subsequently used by automatic scans, not a separate training database. Missing or failing accelerators fail training rather than silently recording CPU fallback as GPU data.
• Added end-to-end training telemetry and an opt-in, sectioned train --summary. Successful training is quiet without --summary; warnings and failures remain visible. Help output also gains color-coded hierarchy in an interactive terminal without changing its plain-text meaning.
• Added same merge [dir] [-r] [--move] [--summary] for model contributions, scan cache, telemetry runs and missing ignore rules. The destination configuration remains unchanged. Descendant ignore rules are scoped to their subtree; rules from an unrelated directory apply at the destination root with a warning. Guarded --move removes only source .same state after successful import and an unchanged-state check.
• Model provenance makes remerging a source idempotent: when that source trains again, its newer contribution replaces its older one rather than double-counting history. Existing model databases migrate to the new schema. Scan cache gains absolute-path archival while ordinary duplicate results remain limited to the current scan scope; an imported digest is reused only after a complete file-stamp match, and exact duplicates still require bytewise verification.
• Telemetry runs are imported by run ID with dependent events, metrics and parameters. A repeated merge skips already imported runs. Imported telemetry remains subject to the destination's later retention policy.

验证与边界 / Validation and limits

Before tagging, local Windows CPU Debug CTest passed 27/27. An isolated Windows Release smoke used physical CUDA and iGPU backends to train three files from two corpus directories: each backend returned three verified digest samples, and all nine accepted training traces persisted. The resulting model and telemetry were imported into a fresh workspace using the Release merge command. This is functional evidence, not a controlled scan-speedup benchmark. Hosted GitHub Actions rebuilds and tests the five release packages across Windows, Linux and macOS before publishing; hosted runners without a physical GPU cannot establish physical GPU performance.

train requires all three backends and representative input; it is not useful on a CPU-only installation. Cold accelerator setup can exceed the cost of short tasks, so normal automatic scans retain their cautious startup policy. Cache stamps are not an adversarial filesystem snapshot. merge is independently transactional and retryable per database, not one atomic transaction across all databases; --move is not globally atomic across multiple sources. Review external ignore rules before scanning. Training calibration and model prediction error do not by themselves prove an end-to-end speedup.

Scanning never deletes input files and remains nonrecursive by default. Use same scan -r --summary to opt into recursion and inspect diagnostics.

same v0.5.1

same v0.5.1

下载 / Downloads

• Windows x64 / Linux x64: standard includes CPU + optional OpenCL iGPU; cuda additionally includes CUDA.
• macOS Apple Silicon: standard includes CPU + optional OpenCL iGPU. There is no CUDA support on macOS.
• CUDA packages require a compatible NVIDIA driver, not a locally installed CUDA compiler. Standard packages need no CUDA runtime.
• Windows packages target Windows 10/11 x64. Linux packages are built on Ubuntu 24.04 (glibc 2.39+). macOS packages target macOS 14+ ARM64.
• Verify SHA256SUMS and build provenance before use. Archives include licenses and build-info.json with the exact source commit.

更新 / Changes

• Reorganized the opt-in human --summary report into Results, Storage and I/O, Performance, Compute and routing, Online routing model, Workers, Model persistence, and Telemetry sections.
• Added redundant semantic ANSI highlighting for structure, healthy state, warnings, errors, timings, and routing evidence. Plain output retains all meaning; NO_COLOR, per-stream TTY detection, and explicit color policy remain intact.
• Preserved the complete current-run evidence, exact per-worker byte counters, legacy TSV results, raw key=value diagnostics, stream contracts, and exit codes.
• Windows launcher now enumerates Visual Studio installations and accepts a CUDA host environment only after the selected nvcc compiles and links the same native C++20 probe used by the build. It no longer assumes the newest Visual Studio is compatible.
• SAME_CUDA_MODE=auto visibly falls back to CPU only after all candidate host environments fail; strict on rejects the build. Explicit caller-owned toolchains remain untouched.

The Windows CUDA package path was locally validated with CUDA 12.8.93 and the Visual Studio 2022 x64 MSVC 19.44 host compiler; all 25 tests passed. Hosted release jobs independently rebuild and test every platform/flavor before publishing their artifacts. Hosted CI does not establish physical GPU performance.

扫描不删除文件；默认不递归。使用 same scan -r --summary 显式扫描子目录并查看完整分区报告。
Scanning never deletes input files; recursion is opt-in with same scan -r --summary.

same v0.5.0

same v0.5.0

下载 / Downloads

• Windows x64 / Linux x64: standard includes CPU + optional OpenCL iGPU; cuda additionally includes CUDA.
• macOS Apple Silicon: standard includes CPU + optional OpenCL iGPU. There is no CUDA support on macOS.
• CUDA packages require a compatible NVIDIA driver, not a locally installed CUDA compiler. Standard packages need no CUDA runtime.
• Windows packages target Windows 10/11 x64. Linux packages are built on Ubuntu 24.04 (glibc 2.39+). macOS packages target macOS 14+ ARM64.
• Verify SHA256SUMS and build provenance before use. Archives include licenses and build-info.json with the exact source commit.

更新 / Changes

• Native BLAKE3 iGPU support with optional dynamic OpenCL loading and exact CPU differential validation.
• CPU/CUDA/iGPU contextual cost prediction, thread-local profiling and cross-run learning in .same/model.db.
• Budgeted cold discovery/activation; no promise that GPU execution is always faster.
• Independent telemetry, safe CPU fallback and unchanged final byte-for-byte verification.
• Cross-platform CI now tests the actual Release package build, including required CUDA compilation on Windows and Linux.

Hosted CI has no NVIDIA GPU: it validates CUDA compilation, packaging and graceful fallback. Real CUDA and Intel Iris Xe execution were separately tested locally. macOS archives are not Apple-notarized; use only artifacts whose origin you have verified.

扫描不删除文件；默认不递归。使用 same scan -r --summary 显式扫描子目录并查看统计。
Scanning never deletes input files; recursion is opt-in with same scan -r --summary.