bioimage-cpp 0.3.0__tar.gz → 0.4.0__tar.gz

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/AGENTS.md

@@ -24,10 +24,10 @@ Required in the environment: `scikit-build-core`, `cmake>=3.21`, `ninja`, `nanob
 
 Before introducing union-finds, priority queues, edge hashing, stride math, threading helpers, or label relabeling, check `include/bioimage_cpp/detail/`:
 
-- `detail/union_find.hxx` — `UnionFind` (path compression + union by rank). `find`, `merge`, `merge_to`, `unite_roots`.
+- `util/union_find.hxx` — `UnionFind` (namespace `bioimage_cpp::util`, not `detail`; path compression + union by rank). `find`, `merge`, `merge_to`, `unite_roots`.
 - `detail/indexed_heap.hxx` — addressable max-heap with mutable priorities. `DenseIndexedHeap` (vector-backed locator, integer keys in `[0, N)`) and `SparseIndexedHeap` (hashmap-backed, arbitrary keys).
 - `detail/edge_hash.hxx` — `Edge`, `edge_key`, `EdgeHash` for hashing unordered node pairs.
-- `detail/grid.hxx` — `c_order_strides`, `valid_offset_target`, `is_valid_grid_edge` for row-major grid offsets.
+- `detail/grid.hxx` — `number_of_elements`, `c_order_strides`, `valid_offset_target`, `is_valid_grid_edge` for row-major grid shapes/offsets.
 - `detail/relabel.hxx` — `dense_relabel` to map arbitrary labels onto `[0, k)` preserving first-occurrence order.
 - `detail/threading.hxx` — `normalize_thread_count`, `parallel_for_chunks(n_threads, n_items, chunk)`.
 

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: bioimage-cpp
-Version: 0.3.0
+Version: 0.4.0
 Summary: Dependency-light C++ bioimage analysis algorithms with Python bindings
 Author: bioimage-cpp contributors
 License: MIT
@@ -31,6 +31,11 @@ Description-Content-Type: text/markdown
 
 # bioimage-cpp
 
+[![Documentation](https://img.shields.io/badge/docs-pdoc-blue)](https://computational-cell-analytics.github.io/bioimage-cpp/)
+[![Build Status](https://github.com/computational-cell-analytics/bioimage-cpp/actions/workflows/tests.yml/badge.svg)](https://github.com/computational-cell-analytics/bioimage-cpp/actions/workflows/tests.yml)
+[![PyPI](https://img.shields.io/pypi/v/bioimage-cpp)](https://pypi.org/project/bioimage-cpp/)
+[![conda-forge](https://img.shields.io/conda/vn/conda-forge/bioimage-cpp)](https://anaconda.org/conda-forge/bioimage-cpp)
+
 Image processing and segmentation functionality in C++ with light-weight python bindings through nanobind and minimal dependencies to enable distribution via pip.
 
 The `bioimage_cpp` python library can be installed via pip:
@@ -38,4 +43,9 @@ The `bioimage_cpp` python library can be installed via pip:
 pip install bioimage-cpp
 ```
 
+Or via conda-forge:
+```bash
+conda install -c conda-forge bioimage-cpp
+```
+
 Please refer to [the documentation](https://computational-cell-analytics.github.io/bioimage-cpp/) for details.

bioimage_cpp-0.4.0/README.md

@@ -0,0 +1,20 @@
+# bioimage-cpp
+
+[![Documentation](https://img.shields.io/badge/docs-pdoc-blue)](https://computational-cell-analytics.github.io/bioimage-cpp/)
+[![Build Status](https://github.com/computational-cell-analytics/bioimage-cpp/actions/workflows/tests.yml/badge.svg)](https://github.com/computational-cell-analytics/bioimage-cpp/actions/workflows/tests.yml)
+[![PyPI](https://img.shields.io/pypi/v/bioimage-cpp)](https://pypi.org/project/bioimage-cpp/)
+[![conda-forge](https://img.shields.io/conda/vn/conda-forge/bioimage-cpp)](https://anaconda.org/conda-forge/bioimage-cpp)
+
+Image processing and segmentation functionality in C++ with light-weight python bindings through nanobind and minimal dependencies to enable distribution via pip.
+
+The `bioimage_cpp` python library can be installed via pip:
+```bash
+pip install bioimage-cpp
+```
+
+Or via conda-forge:
+```bash
+conda install -c conda-forge bioimage-cpp
+```
+
+Please refer to [the documentation](https://computational-cell-analytics.github.io/bioimage-cpp/) for details.

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/development/flow/PERFORMANCE_NOTES.md

@@ -199,6 +199,65 @@ wheel still loads on SSE2-only CPUs. Expected speedup at 1T: ~20–30 %, less
 at 8T due to closer-to-bandwidth saturation. Estimated effort: half a day
 plus a careful microbenchmark of `vpgatherdd` cost on the target CPUs.
 
+## Interleaved (channel-last) layout + hand-written AVX2 FMA (rolled back, 2026-06-12)
+
+This implemented the two top "remaining" ideas below and measured them. **Net
+result: no win; small regression. Rolled back.** Documented here so the avenue
+is not re-attempted without new information.
+
+The idea, in two composable stages:
+
+1. **Channel-last (interleaved) flow buffer.** The input is channel-first
+   (`channels[axis] = flow.data + axis*n_pixels`), so the `D` components at one
+   position are gathers from regions `n_pixels` floats apart — in theory `D`
+   distinct cache lines per corner. A one-time `O(N)` transpose into a
+   `std::vector<float>` where a voxel's components sit at `[v*VS + axis]` makes
+   each corner touch one line and yield all `D` components. `compute_corners`
+   is unchanged (its offset is already the voxel flat index); only the sampler
+   multiplies by the voxel stride `VS`.
+2. **Hand-written AVX2+FMA sampler for 3D** on that interleaved buffer, with
+   `VS = 4` padding so each corner loads as one 128-bit vector. Lane-wise FMA
+   accumulates the weighted `D`-vector across the 8 corners — **no gather**
+   (the rejected approach above). Runtime-dispatched via
+   `__builtin_cpu_supports`, per-function `__attribute__((target("avx2,fma")))`,
+   scalar fallback for MSVC/arm64/non-AVX2. No CMake/global-arch-flag change.
+
+The codegen was confirmed ideal via `objdump`: 8× `vfmadd132ps` each fusing a
+128-bit load (`(%rdi,%rax,1)`) with a `vbroadcastss` weight, zero gather
+instructions. So the SIMD path was real and optimal — and still no faster than
+scalar.
+
+Measurements (3D fixture, warm, min of 8 runs, same session, same machine as
+the headline numbers):
+
+```
+variant                          1T min     8T min
+channel-first (baseline)         5.33 s     1.357 s
+interleaved VS=3 (no pad, scalar) 5.44 s    1.46 s
+interleaved VS=4 (scalar)        5.56 s     1.42 s
+interleaved VS=4 + AVX2 FMA      5.58 s     1.46 s
+```
+
+Conclusions:
+
+- **The loop is memory-latency bound, as the roofline diagnosis said.**
+  Cutting the corner sum from 24 scalar mul-adds to 8 packed FMAs changes the
+  ALU count, not the load latency, so it does nothing. The 8 corner loads are
+  independent, so the out-of-order engine already overlaps them (high MLP);
+  channel-first's three independent channel loads overlap too.
+- **Channel-first already has good locality.** A 64-byte line holds 16
+  consecutive-x voxels of one channel; a trilinear cell's x-pair lives on one
+  line. Interleaving with `VS=4` cuts that to 4 voxels/line — *worse* density —
+  which is why VS=4 lost ~0.12 s/1T to VS=3. But even VS=3 (no waste, same
+  144 MB footprint) did not beat channel-first.
+- **Padding `VS=4` also costs +33 % flow memory** (144 → 192 MB for the 3D
+  fixture) for the AVX2 path's aligned 4-wide load.
+- The differences are partly inside this laptop's **±~8 % thermal-throttle
+  noise** (baseline 1T alone ranged 5.33–5.85 s across the session, 8T
+  1.357–1.59 s). That noise floor exceeds the expected gain of any remaining
+  micro-optimization, so small wins cannot be validated on this host — they
+  need a fixed-clock machine with `perf` counters.
+
 ## What was tried and rejected
 
 - **Per-thread density scatter buffers** — scatter is <1 % of runtime,
@@ -212,25 +271,36 @@ plus a careful microbenchmark of `vpgatherdd` cost on the target CPUs.
 - **Half-precision (fp16) flow storage** — would halve memory traffic, but
   the kernel isn't bandwidth-bound (see diagnosis), and it's a breaking API
   change.
+- **Interleaved channel-last layout + hand-written AVX2 FMA** — implemented and
+  measured 2026-06-12 (see the dedicated section above). No win, small
+  regression, rolled back. The SIMD path was confirmed optimal in codegen yet
+  did not beat scalar — the loop is load-latency bound, not ALU bound.
 
 ## What remains worth trying
 
-In rough order of expected return on effort:
-
-1. **Hand-written AVX2 (and SSE2-fallback) intrinsics for the corner sum**
-   inside a runtime-dispatched specialization. Plausible 20–30 % at 1T. The
-   plumbing scaffolding is documented above in the SIMD section.
-2. **SoA position layout** (`vector<float> pos_x, pos_y, pos_z` instead of
-   `vector<array<float,3>>`). Lets the autovectorizer batch the clip + step
-   + convergence check across N particles. Pairs naturally with (1).
-3. **Reduce gather miss latency** via batch prefetching: prefetch the
-   next-K particles' corner cache lines before computing the current one.
-   Likely modest gain; needs profiling.
+In rough order of expected return on effort. **Note (2026-06-12):** items 1 and
+2 below were implemented and measured — they gave no speedup (see "Interleaved
+layout + AVX2 FMA" above). They are kept here only with that caveat; the live
+candidate is now (3), and (4) is the prerequisite for evaluating it.
+
+1. ~~Hand-written AVX2 intrinsics for the corner sum~~ — **tried, no win.** The
+   corner sum on an interleaved buffer compiled to 8 packed FMAs with no
+   gathers, and was no faster than scalar: the loop waits on the 8 corner
+   loads, not the arithmetic.
+2. ~~SoA position layout~~ — its only rationale was enabling the SIMD in (1);
+   with (1) shown not to help, SoA's batched clip/step/convergence has nothing
+   to pay for the extra coordinate gather/scatter. Not pursued separately.
+3. **Reduce load latency** via software prefetching: prefetch the next-K
+   particles' corner cache lines before computing the current one. This is the
+   one lever that targets the actual (latency) bottleneck, and it works on the
+   existing channel-first layout — no relayout needed. Expected gain is modest
+   and **cannot be validated on the current laptop** (its ±~8 % thermal noise
+   floor swamps it); needs a fixed-clock host.
 4. **A real `perf stat`** run on a host where `linux-perf-tools` is
    installed, to confirm where cycles actually go (front-end, back-end
-   memory, back-end core, retired-FMA-ratio). The streaming-probe upper
-   bound is necessary but not sufficient — counter data would localise the
-   bottleneck precisely.
+   memory, back-end core, retired-FMA-ratio) and to provide a low-noise
+   measurement environment in which (3) could actually be evaluated. The
+   streaming-probe upper bound is necessary but not sufficient.
 
 ## Reproducing these measurements
 

bioimage_cpp-0.4.0/development/graph/benchmark_accumulate_labels.py

@@ -0,0 +1,82 @@
+"""Benchmark RAG label accumulation (``accumulate_labels``).
+
+The code review replaced the per-thread ``n_threads x n_nodes`` map-of-maps with
+a single combined-key ``(node, other)`` histogram per thread, and made the
+per-node argmax single-threaded. This script times ``accumulate_labels`` across
+thread counts (to confirm the allocation win and rule out a high-thread-count
+regression from the now-sequential argmax) and across ``other``-label
+cardinalities.
+
+Inputs: a deterministic block-wise over-segmentation (each node is a contiguous
+block, as in a real over-segmentation), so the RAG adjacency is local.
+
+Not part of the test suite. Run::
+
+    python development/graph/benchmark_accumulate_labels.py --repeats 11
+"""
+from __future__ import annotations
+
+import argparse
+from statistics import median
+from time import perf_counter
+
+import numpy as np
+
+import bioimage_cpp as bic
+
+
+def _timeit(fn, repeats: int, warmup: int = 1) -> dict:
+    for _ in range(warmup):
+        fn()
+    timings = []
+    for _ in range(repeats):
+        t0 = perf_counter()
+        fn()
+        timings.append(perf_counter() - t0)
+    return {"median": median(timings), "min": min(timings), "n": repeats}
+
+
+def _make_labels(shape=(40, 256, 256), coarsen=(2, 4, 4)) -> np.ndarray:
+    coarse_shape = tuple(s // c for s, c in zip(shape, coarsen))
+    n_nodes = int(np.prod(coarse_shape))
+    coarse = np.arange(n_nodes, dtype=np.uint64).reshape(coarse_shape)
+    block = np.ones(coarsen, dtype=np.uint64)
+    return np.ascontiguousarray(np.kron(coarse, block))
+
+
+def run(repeats: int = 11) -> dict:
+    labels = _make_labels()
+    rag = bic.graph.region_adjacency_graph(labels)
+    n_nodes = int(rag.number_of_nodes)
+    rng = np.random.default_rng(0)
+
+    results: dict[str, dict] = {}
+    for n_other, tag in ((8, "dense8"), (2000, "sparse2000")):
+        other = rng.integers(0, n_other, size=labels.shape).astype(np.uint64)
+        other = np.ascontiguousarray(other)
+        for threads in (1, 4, 8):
+            def fn(t=threads, o=other):
+                bic.graph.features.accumulate_labels(rag, labels, o, number_of_threads=t)
+
+            key = f"accumulate_labels_{tag}_t{threads}"
+            results[key] = _timeit(fn, repeats)
+            results[key]["meta"] = {
+                "n_nodes": n_nodes,
+                "n_pixels": int(labels.size),
+                "n_other": n_other,
+                "threads": threads,
+            }
+    return results
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--repeats", type=int, default=11)
+    args = parser.parse_args()
+    results = run(repeats=args.repeats)
+    for name, r in results.items():
+        print(f"{name:<32} median={r['median'] * 1e3:8.3f} ms  min={r['min'] * 1e3:8.3f} ms")
+
+
+if __name__ == "__main__":
+    main()

bioimage_cpp-0.4.0/development/graph/benchmark_bfs.py

@@ -0,0 +1,70 @@
+"""Benchmark the workspace-reused BFS path (``lifted_edges_from_node_labels``).
+
+The code review changed ``BfsWorkspace::reset`` from an O(N) clear of the
+visited/distance buffers to an O(1) generation-stamp bump. That only pays off
+when one workspace is reset across many sources, which is exactly what
+``lifted_edges_from_node_labels`` does (one workspace per chunk, reset per
+source). The single-call ``breadth_first_search`` builds a fresh workspace each
+call and would NOT show the change, so we benchmark the lifted-edge path.
+
+Single-threaded so the per-source reset cost is not hidden by parallelism. The
+node count is swept to expose the previous O(N^2)-of-memset behavior.
+
+Not part of the test suite. Run::
+
+    python development/graph/benchmark_bfs.py --repeats 5
+"""
+from __future__ import annotations
+
+import argparse
+from statistics import median
+from time import perf_counter
+
+import numpy as np
+
+import bioimage_cpp as bic
+
+
+def _timeit(fn, repeats: int, warmup: int = 1) -> dict:
+    for _ in range(warmup):
+        fn()
+    timings = []
+    for _ in range(repeats):
+        t0 = perf_counter()
+        fn()
+        timings.append(perf_counter() - t0)
+    return {"median": median(timings), "min": min(timings), "n": repeats}
+
+
+def run(repeats: int = 5, depth: int = 2) -> dict:
+    shapes = [(100, 100), (160, 160), (220, 220)]
+    results: dict[str, dict] = {}
+    for shape in shapes:
+        n_nodes = int(np.prod(shape))
+        graph = bic.graph.grid_graph(shape)
+        rng = np.random.default_rng(0)
+        node_labels = rng.integers(0, 50, size=(n_nodes,), dtype=np.uint64)
+
+        def fn(g=graph, nl=node_labels, d=depth):
+            bic.graph.lifted_multicut.lifted_edges_from_node_labels(
+                g, nl, graph_depth=d, number_of_threads=1
+            )
+
+        key = f"bfs_lifted_{n_nodes}nodes_d{depth}"
+        results[key] = _timeit(fn, repeats)
+        results[key]["meta"] = {"n_nodes": n_nodes, "shape": shape, "depth": depth}
+    return results
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--repeats", type=int, default=5)
+    parser.add_argument("--depth", type=int, default=2)
+    args = parser.parse_args()
+    results = run(repeats=args.repeats, depth=args.depth)
+    for name, r in results.items():
+        print(f"{name:<28} median={r['median'] * 1e3:9.3f} ms  min={r['min'] * 1e3:9.3f} ms")
+
+
+if __name__ == "__main__":
+    main()

bioimage_cpp-0.4.0/development/segmentation/benchmark_mutex_watershed.py

@@ -0,0 +1,89 @@
+"""Benchmark the mutex-watershed grid kernel (bic only).
+
+The code review routed the neighbor computation in ``mutex_watershed_grid``
+through ``detail::valid_offset_target`` (per-axis bounds check) instead of a
+single precomputed flat-offset add. This script times ``mutex_watershed`` on the
+ISBI affinities for 2D and 3D so we can A/B that inner-loop change against a
+pre-review build. No external (affogato) dependency — bic only.
+
+Not part of the test suite. Run::
+
+    python development/segmentation/benchmark_mutex_watershed.py --repeats 5
+"""
+from __future__ import annotations
+
+import argparse
+from statistics import median
+from time import perf_counter
+
+import numpy as np
+
+import bioimage_cpp as bic
+from bioimage_cpp._data import load_isbi_affinities
+
+
+def _timeit(fn, repeats: int, warmup: int = 1) -> dict:
+    for _ in range(warmup):
+        fn()
+    timings = []
+    for _ in range(repeats):
+        t0 = perf_counter()
+        fn()
+        timings.append(perf_counter() - t0)
+    return {"median": median(timings), "min": min(timings), "n": repeats}
+
+
+def _attractive_flip(affs: np.ndarray, n_attractive: int) -> np.ndarray:
+    # Match the convention in the equivalence checker: attractive channels are
+    # turned into merge affinities (1 - aff).
+    out = affs.copy()
+    out[:n_attractive] *= -1
+    out[:n_attractive] += 1
+    return out
+
+
+def run(repeats: int = 5) -> dict:
+    affinities, offsets = load_isbi_affinities()
+    affinities = np.ascontiguousarray(affinities)
+    offsets = [tuple(o) for o in offsets]
+    results: dict[str, dict] = {}
+
+    # --- 2D: in-plane channels of a single z slice ---
+    channels_2d = [i for i, o in enumerate(offsets) if o[0] == 0]
+    aff2d = np.ascontiguousarray(affinities[channels_2d, 0, :256, :256])
+    offsets_2d = [offsets[i][1:] for i in channels_2d]
+    aff2d_flipped = _attractive_flip(aff2d, 2)
+
+    def mws_2d():
+        bic.segmentation.mutex_watershed(
+            aff2d_flipped, offsets_2d, number_of_attractive_channels=2
+        )
+
+    results["mws_2d_256x256"] = _timeit(mws_2d, repeats)
+    results["mws_2d_256x256"]["meta"] = {"shape": list(aff2d.shape)}
+
+    # --- 3D: small crop, all offsets ---
+    aff3d = np.ascontiguousarray(affinities[:, :6, :256, :256])
+    aff3d_flipped = _attractive_flip(aff3d, 3)
+
+    def mws_3d():
+        bic.segmentation.mutex_watershed(
+            aff3d_flipped, offsets, number_of_attractive_channels=3
+        )
+
+    results["mws_3d_6x256x256"] = _timeit(mws_3d, repeats)
+    results["mws_3d_6x256x256"]["meta"] = {"shape": list(aff3d.shape)}
+    return results
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--repeats", type=int, default=5)
+    args = parser.parse_args()
+    results = run(repeats=args.repeats)
+    for name, r in results.items():
+        print(f"{name:<20} median={r['median'] * 1e3:9.3f} ms  min={r['min'] * 1e3:9.3f} ms")
+
+
+if __name__ == "__main__":
+    main()

bioimage_cpp-0.4.0/development/utils/benchmark_union_find.py

@@ -0,0 +1,70 @@
+"""Micro-benchmark for the UnionFind Python bindings.
+
+Times the bulk ``merge((N, 2) edges)`` and ``find(node_array)`` entry points,
+which the code review changed to validate every node id against ``uf.size``
+before touching the C++ structure (an extra O(N) pre-pass). This script lets us
+A/B that pre-pass against a pre-review build.
+
+Not part of the test suite. Run::
+
+    python development/utils/benchmark_union_find.py --repeats 11
+"""
+from __future__ import annotations
+
+import argparse
+from statistics import median
+from time import perf_counter
+
+import numpy as np
+
+import bioimage_cpp as bic
+
+
+def _timeit(fn, repeats: int, warmup: int = 1) -> dict:
+    for _ in range(warmup):
+        fn()
+    timings = []
+    for _ in range(repeats):
+        t0 = perf_counter()
+        fn()
+        timings.append(perf_counter() - t0)
+    return {"median": median(timings), "min": min(timings), "n": repeats}
+
+
+def run(repeats: int = 11, n_nodes: int = 1_000_000, n_edges: int = 2_000_000) -> dict:
+    rng = np.random.default_rng(0)
+    edges = rng.integers(0, n_nodes, size=(n_edges, 2), dtype=np.uint64)
+    query = rng.integers(0, n_nodes, size=(n_edges,), dtype=np.uint64)
+
+    # Pre-merged structure for the find benchmark (find does not mutate the
+    # partition, so it can be reused across repeats).
+    merged = bic.utils.UnionFind(n_nodes)
+    merged.merge(edges)
+
+    def bulk_merge():
+        uf = bic.utils.UnionFind(n_nodes)
+        uf.merge(edges)
+
+    def bulk_find():
+        merged.find(query)
+
+    results = {
+        "uf_bulk_merge": _timeit(bulk_merge, repeats),
+        "uf_bulk_find": _timeit(bulk_find, repeats),
+    }
+    for r in results.values():
+        r["meta"] = {"n_nodes": n_nodes, "n_edges": n_edges}
+    return results
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--repeats", type=int, default=11)
+    args = parser.parse_args()
+    results = run(repeats=args.repeats)
+    for name, r in results.items():
+        print(f"{name:<16} median={r['median'] * 1e3:8.3f} ms  min={r['min'] * 1e3:8.3f} ms")
+
+
+if __name__ == "__main__":
+    main()

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/include/bioimage_cpp/detail/grid.hxx

@@ -6,6 +6,16 @@
 
 namespace bioimage_cpp::detail {
 
+// Total number of elements in a row-major array of the given shape (the product
+// of the per-axis extents). Shape entries are assumed non-negative.
+inline std::size_t number_of_elements(const std::vector<std::ptrdiff_t> &shape) {
+    std::size_t total = 1;
+    for (const auto extent : shape) {
+        total *= static_cast<std::size_t>(extent);
+    }
+    return total;
+}
+
 // C-order strides for a row-major array of the given shape, in units of array
 // elements (not bytes). The innermost (last) axis has stride 1.
 inline std::vector<std::ptrdiff_t> c_order_strides(const std::vector<std::ptrdiff_t> &shape) {

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/include/bioimage_cpp/detail/profile.hxx

@@ -89,7 +89,12 @@ inline ProfileTimerNull make_profile_timer(NullProfiler &profiler, const char *n
 } // namespace bioimage_cpp::detail
 
 #define BIOIMAGE_PROFILE_INIT(var) ::bioimage_cpp::detail::Profiler var;
-#define BIOIMAGE_PROFILE_SCOPE(var, name) auto _bp_##__LINE__ = ::bioimage_cpp::detail::make_profile_timer(var, name);
+// Two-level indirection so __LINE__ is expanded before the token paste; without
+// it every BIOIMAGE_PROFILE_SCOPE in a translation unit would declare the same
+// identifier `_bp___LINE__`, breaking two scopes in one block.
+#define BIOIMAGE_PROFILE_CONCAT_(a, b) a##b
+#define BIOIMAGE_PROFILE_CONCAT(a, b) BIOIMAGE_PROFILE_CONCAT_(a, b)
+#define BIOIMAGE_PROFILE_SCOPE(var, name) auto BIOIMAGE_PROFILE_CONCAT(_bp_, __LINE__) = ::bioimage_cpp::detail::make_profile_timer(var, name);
 #define BIOIMAGE_PROFILE_REPORT(var) (var).report();
 
 #else

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/include/bioimage_cpp/flow/flow_density.hxx

@@ -109,7 +109,10 @@ inline std::ptrdiff_t round_to_flat_index(
         } else if (clipped > grid.upper[axis]) {
             clipped = grid.upper[axis];
         }
-        const auto coord = static_cast<std::ptrdiff_t>(std::nearbyint(clipped));
+        // Round half up, matching the nearest-neighbor convention in
+        // transformation/affine.hxx and segmentation/watershed.hxx.
+        // std::nearbyint would honor the FP rounding mode (round-half-to-even).
+        const auto coord = static_cast<std::ptrdiff_t>(std::floor(clipped + 0.5f));
         flat += coord * grid.strides[axis];
     }
     return flat;

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/include/bioimage_cpp/graph/agglomeration/detail.hxx

@@ -93,8 +93,9 @@ inline std::size_t agglo_merge_dynamic_nodes(
             if (new_priority != current_priority) {
                 edge.weight = new_priority;
                 // The edge may have been previously popped via kRejectEdge
-                // (GASP cannot-link), in which case it is no longer in the
-                // heap. Use push_or_change to handle both cases.
+                // (GASP cannot-link). In that case it must stay out of the heap
+                // (re-adding it would let it be re-popped and merged), so only
+                // update the priority when the edge is still present.
                 if (heap.contains(removed_edge_id)) {
                     heap.change(removed_edge_id, new_priority);
                 }

{bioimage_cpp-0.3.0 → bioimage_cpp-0.4.0}/include/bioimage_cpp/graph/breadth_first_search.hxx

@@ -2,6 +2,7 @@
 
 #include "bioimage_cpp/graph/undirected_graph.hxx"
 
+#include <algorithm>
 #include <cstddef>
 #include <cstdint>
 #include <limits>
@@ -19,23 +20,44 @@ struct BfsEntry {
 };
 
 // Reusable scratch state for `breadth_first_search`. Reset via `reset(graph)`
-// before each call so the visited-flag and distance buffers grow once and stay
+// before each call so the visited and distance buffers grow once and stay
 // allocated across many BFS invocations on the same graph.
+//
+// Visited tracking uses a per-call generation stamp rather than a boolean
+// buffer cleared each call: `reset` just increments the generation (O(1)),
+// doing a full clear only on the rare 32-bit wraparound. This matters for the
+// "k-hop neighborhood from every node" pattern, where an O(N)-per-call clear
+// would otherwise make the whole sweep O(N^2). `distance_` is sized but not
+// cleared; it is written before being read for every visited node.
 class BfsWorkspace {
 public:
     BfsWorkspace() = default;
 
     void reset(const UndirectedGraph &graph) {
         const auto n_nodes = static_cast<std::size_t>(graph.number_of_nodes());
-        visited_.assign(n_nodes, 0);
-        distance_.assign(n_nodes, 0);
+        if (visited_.size() != n_nodes) {
+            visited_.assign(n_nodes, 0);
+            generation_ = 0;
+        }
+        distance_.resize(n_nodes);
+        if (generation_ == std::numeric_limits<std::uint32_t>::max()) {
+            std::fill(visited_.begin(), visited_.end(), std::uint32_t{0});
+            generation_ = 0;
+        }
+        ++generation_;
     }
 
-    [[nodiscard]] std::vector<unsigned char> &visited() { return visited_; }
+    [[nodiscard]] bool is_visited(const std::uint64_t node) const {
+        return visited_[static_cast<std::size_t>(node)] == generation_;
+    }
+    void mark_visited(const std::uint64_t node) {
+        visited_[static_cast<std::size_t>(node)] = generation_;
+    }
     [[nodiscard]] std::vector<std::uint64_t> &distance() { return distance_; }
 
 private:
-    std::vector<unsigned char> visited_;
+    std::vector<std::uint32_t> visited_;
+    std::uint32_t generation_ = 0;
     std::vector<std::uint64_t> distance_;
 };
 
@@ -71,13 +93,12 @@ inline std::vector<BfsEntry> breadth_first_search(
         );
     }
     workspace.reset(graph);
-    auto &visited = workspace.visited();
     auto &distance = workspace.distance();
 
     std::vector<BfsEntry> result;
     std::queue<std::uint64_t> queue;
     queue.push(source);
-    visited[static_cast<std::size_t>(source)] = 1;
+    workspace.mark_visited(source);
     distance[static_cast<std::size_t>(source)] = 0;
     if (include_source) {
         result.push_back({source, 0});
@@ -93,10 +114,10 @@ inline std::vector<BfsEntry> breadth_first_search(
         const auto next_distance = node_distance + 1;
         for (const auto adjacency : graph.node_adjacency(node)) {
             const auto neighbor = adjacency.node;
-            if (visited[static_cast<std::size_t>(neighbor)] != 0) {
+            if (workspace.is_visited(neighbor)) {
                 continue;
             }
-            visited[static_cast<std::size_t>(neighbor)] = 1;
+            workspace.mark_visited(neighbor);
             distance[static_cast<std::size_t>(neighbor)] = next_distance;
             result.push_back({neighbor, next_distance});
             queue.push(neighbor);