Benchmarks and reproduction
A benchmark is a comparison of specific image computations under stated conditions. It needs a numerical/output check as well as a timer. A faster run with a smaller source footprint, different stopping criterion or missing output is not necessarily an optimization of the same computation.
Existing reports
| Workload | Report |
|---|---|
| Segmentation pipelines and model-backed examples | Current benchmark overview |
| Finite-PSF RL/SHB, including original Deconwolf C | Deconvolution measurements and raw records |
| Earlier pipeline measurements | Archived results |
Use the reproduction commands and raw records in the report matching your workload. The site does not execute benchmarks during rendering or translate a hardware-specific speed ratio into a universal claim.
Establish the same task
Start with identical input pixels, selected channels/levels, physical spacing and numeric types. Match boundary and missing-data conventions. For a local filter, check the truncation radius and arithmetic definition. For segmentation, match connectivity, thresholds and overlap reconstruction. For model inference, match model weights, normalization, precision and acceleration route.
Compare useful outputs before comparing durations. Exact labels can be checked when identity construction and algorithms match; otherwise use explicit overlap matching and explain how splits and merges are counted. A matching object count alone can hide missed and spurious objects that cancel each other numerically.
For iterative reconstruction, fixed iteration throughput and time to a common quality target answer different questions. One algorithm may take a cheaper iteration but need more of them. The deconvolution report preserves both the iteration-cost comparison and reconstruction-quality curves.
Define the timed region
Separate fixture preparation, compilation, model download/loading, metadata opening, image reading, processing, compression and output publication. Choose the timed region that answers the intended question and state it. End-to-end runtime is valuable for users; isolated kernel throughput is valuable for understanding an optimization. Neither should silently stand in for the other.
Warm OS data is not the same as a warm decoded-chunk cache. A cached compressed file may still require decompression on every read. Block shape and halo width can multiply source traffic even when the number of output voxels is unchanged. Record cache policy, storage chunks, compute blocks and concurrency alongside the machine and build.
All Rust measurements use release builds. Inspect worker count and device use for every tool: matching a CPU count flag is insufficient when one backend threads inside kernels and another parallelizes blocks. Memory-limited admission can also reduce actual concurrency below the request.
A useful record for a new result
Save the source revision, effective configuration, machine description, input identity and output validation with the measurements. Use repeated runs on a reasonably quiet machine, retain individual samples, and distinguish cold from warm starts. Describe variability instead of choosing only the fastest sample.
For an execution change, vary the block grid while holding the mathematical computation fixed. Track traffic, active memory and CPU time when diagnosing a speed difference. A change can trade repeated halo arithmetic for saved intermediate I/O; wall time alone does not explain which trade paid off. ExecutionLog and Stats expose execution evidence, but their counters have different units: pixel-chain applications, fragment applications, tasks and spatial blocks are not interchangeable totals.
Once a matched comparison is established, link its report from the relevant example and preserve raw records. A later optimization should update the current result while keeping enough earlier data to explain what changed.