docs: known-issues — cite PRs #30, #31, #32

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

BENCHMARKS.md

@@ -601,6 +601,42 @@ about 150 per dataset here) against a Fixed Array's few blocks. Writing costs th
 same; files grow by about 36 bytes per chunk. The default therefore stays
 the 1.10 format; the 1.8 format is opt-in.
 
+### Streaming `FileBuilder::write` (2026-09-29, tank)
+
+`FileBuilder::write` now streams the file through a buffered writer
+instead of building it in memory first (`feat/huge-chunk-dims`, the change
+that removes copies of 4 GiB unfiltered chunks). Measured 2026-09-29 on
+tank (AMD Ryzen 7 7800X3D), idle (1-minute load average 1.66 to 1.83 at the
+start of each run): `main` at `4260af4` against the stacked branch at
+`549e442`, alternating, three full runs each; median (min-max) of
+criterion's estimate, ms.
+
+> **Run:** `cargo bench -p clawhdf5-bench --bench h5bench_write -- --noplot --warm-up-time 1 --measurement-time 3`
+
+The first candidate used a 1 MiB write buffer and was 1.35x to 1.83x slower
+on every 512 x 512 (1 MiB) chunked case when the full suite ran (e.g.
+`write_2d_chunked/512x512` 0.454 -> 0.754 ms): the buffer is above glibc's
+128 KiB mmap threshold, and after the smaller cases it was mapped afresh on
+each write (1 629 368 minor page faults over the `write_2d_chunked` group
+against 12 888 on `main`). Run alone, the case showed no difference. With a
+64 KiB buffer (`549e442`):
+
+| benchmark | main | candidate | ratio |
+|---|---:|---:|---:|
+| write_1d_contiguous/100000 | 0.216 (0.215-0.218) | 0.206 (0.204-0.207) | 0.96x |
+| write_2d_chunked/32x32 | 0.062 (0.061-0.063) | 0.061 (0.061-0.064) | 0.99x |
+| write_2d_chunked/128x128 | 0.102 (0.100-0.103) | 0.101 (0.099-0.102) | 0.98x |
+| write_2d_chunked/512x512 | 0.460 (0.457-0.463) | 0.466 (0.463-0.477) | 1.01x |
+| write_2d_chunked_zstd/deflate-6/512x512 | 0.457 (0.450-0.465) | 0.465 (0.458-0.477) | 1.02x |
+| write_2d_chunked_zstd/zstd-3/512x512 | 0.376 (0.369-0.381) | 0.387 (0.379-0.390) | 1.03x |
+| write_2d_chunked_pcodec/pcodec/512x512 | 0.906 (0.895-0.908) | 0.901 (0.898-0.914) | 0.99x |
+| write_multi_dataset/64 | 0.136 (0.135-0.138) | 0.136 (0.135-0.136) | 1.00x |
+| write_with_attrs/64 | 0.063 (0.062-0.063) | 0.063 (0.062-0.063) | 1.00x |
+
+The other 18 cases are within 2%. Minor page faults per full run: 59 602 to
+64 486 on both sides. The write path is as fast as before; what changes is
+peak memory for large unfiltered chunks (see CHANGELOG, 2026-09-29).
+
 ### HDF5 1.8 format: version-1 B-tree chunk indexes, idle re-run (2026-09-28, tank)
 
 Measured 2026-09-28 on tank (AMD Ryzen 7 7800X3D), idle: 1-minute load

CHANGELOG.md

@@ -56,6 +56,128 @@
   and `docs/known-issues.md` (tank, 2026-09-29, netCDF4-python 1.7.4).
   Affected v2.1.0 to v2.7.0.
 
+### HDF5 2.0 native complex is its own type on read; Python `libver=` (2026-09-29)
+- **`Datatype::parse` returns `Datatype::Complex { size, base_type }`** for
+  a class-11 message, also as a compound member, array base or
+  variable-length base. It used to return the equivalent `{r, i}`
+  compound (`docs/known-issues.md`, "HDF5 2.0 native complex numbers read
+  as a `{r, i}` compound"). **Breaking** for code that matched the
+  compound view of a native complex type: `Dataset::raw_datatype()` and
+  attribute datatypes are now `Datatype::Complex`;
+  `Datatype::complex_as_compound(size, base)` still gives the `{r, i}` view,
+  and `data_read::read_compound_fields` accepts a `Complex` directly.
+  Re-serializing a parsed type (e.g. copying it to another file) now writes
+  class 11 again instead of a compound.
+- **Facade:** `DType::Complex(Box<DType>)` (new variant, **breaking** for
+  exhaustive `match`es over `DType`): `Complex(F32)` is numpy `complex64`,
+  `Complex(F64)` `complex128`, a binary16 part `Complex(Other("float16"))`
+  (the same `Other` name small floats already use); `Display` prints
+  `complex<f64>`. h5py's own complex encoding, the compound `{r, i}`, is
+  still `DType::Compound`. `read_complex_f64`/`read_complex_f32` read both.
+- **`h5rs`:** `dump` prints native complex as h5dump 2.2.0 does:
+  `H5T_COMPLEX_IEEE_F{16,32,64}{LE,BE}` (else `H5T_COMPLEX { <base> }`), in
+  arrays and compounds too, and values as `1.5-2i` (`%g%+gi`; for binary16
+  parts, which h5dump has no C type for, `1+-2i` as it prints them). `ls`
+  shows `complex64`, `complex128-be`, `complex32` (and `complex<part>` for
+  non-IEEE parts); `ls -v` shows h5ls 2.2.0's `complex number of` /
+  `IEEE 64-bit big-endian float`. `diff` compares complex values part by
+  part and prints the difference as h5diff 2.2.0 does (`1+0i`); a native
+  complex and an `{r, i}` compound are no longer comparable (different
+  classes, as in h5diff). `dump --json` keeps the `{r, i}` compound and
+  `[re, im]` values: hdf5-json (h5json 2.0.0) has no complex class.
+- **Browser (`clawhdf5-wasm`):** native complex datasets are readable, as
+  `[re, im]` pairs: `info().elementShape` ends in `2`, `read()` returns the
+  parts interleaved in a `Float32Array`/`Float64Array` (binary16 parts
+  widened to f32), and `dtype` is `complex<f64>` (`complex<f32
+  (big-endian)>`, `array[2]<complex<f32>>`). The viewer shows each element
+  as `[re, im]` with no change to the page. h5py's `{r, i}` compound is
+  still refused, as every compound is.
+- **Python:** `clawhdf5.File(path, 'w', libver=...)` takes h5py's values:
+  `'v108'`, `'v110'`, `'v112'`, `'v114'`, `'v200'`, `'latest'` (the low
+  bound; high `'latest'`, as h5py) or a `(low, high)` tuple, mapped to
+  `FileBuilder::libver_bounds`. `'earliest'` as the low bound writes the
+  1.8 format with a `UserWarning` (clawhdf5 cannot write the pre-1.8
+  format); as the high bound it is a `ValueError`, as are unknown names and
+  a low bound above the high one. It is ignored for `'r'` and refused
+  (`NotImplementedError`) for `'r+'`/`'a'`. Native complex already read as
+  numpy `complex64`/`complex128` and still does
+  (`test_read_native_complex_from_h5py`).
+- Tests (tank, 2026-09-29): `crates/clawhdf5-tools/tests/native_complex_dump.rs`
+  compares `h5rs dump` with h5dump 2.2.0's output stored next to a fixture
+  written by h5py 3.16 / libhdf5 2.0.0
+  (`crates/clawhdf5/tests/fixtures/gen_native_complex.py`:
+  `F32LE`/`F64LE`/`F64BE`/`F16LE`, scalar, compound member, array, root
+  attribute), line for line outside the values and value for value within
+  them, and `ls`/`diff` with h5ls/h5diff 2.2.0;
+  `integration_tests::complex_datasets_and_attributes_round_trip`
+  (`DType::Complex`, `raw_datatype()`); `clawhdf5-wasm` unit tests on the
+  fixture, `h5py_interop` and `examples/wasm-viewer/test` (Node:
+  `test.mjs`, 274 + 1324 checks; the page in headless Chromium for
+  `/native_c128`, `/native_c64_be`, `/pairs`) with two native complex
+  datasets added to `make_fixture.py` (when h5py's libhdf5 is 2.0+);
+  `crates/clawhdf5-py/tests/test_libver.py` (every bound; `'v108'` output
+  opens in HDF5 1.8.23's h5dump and in h5py). `test.mjs` no longer
+  hard-codes the fixture's length.
+
+### Chunk dimensions of 2^32 or more; 4 GiB chunks written without copies (2026-09-29)
+- **Chunk dimensions of 2^32 or more** (libhdf5 2.x writes them with
+  layout message version 5, up to 8 bytes per dimension; a 4 GiB chunk of
+  1-byte elements needs one) are read and written. They were refused
+  (`InvalidChunkDimensions`). **Public API changes:**
+  `DataLayout::Chunked::chunk_dimensions` is `Vec<u64>` (was `Vec<u32>`),
+  and the `chunk_dimensions` argument of `chunked_read::{collect_chunk_info_checked,
+  collect_chunk_info_checked_in, generate_implicit_chunks_in_grid}`,
+  `fixed_array::{read_fixed_array_chunks, read_fixed_array_chunks_in}`,
+  `extensible_array::{read_extensible_array_chunks,
+  read_extensible_array_chunks_in}`, and the `chunk_dims` argument of
+  `chunked_write::serialize_v4_single_chunk_pub`, are `&[u64]` (were
+  `&[u32]`). A chunk whose size overflows 64 bits is refused when a dataset is read
+  (`InvalidChunkDimensions`), and the writer refuses a chunk dimension of
+  0 up front. A version-3 layout (4-byte dimensions) is never written for
+  such a chunk: over 4 GiB, it always takes version 5, as in libhdf5.
+- **Writing without copies:** `FileWriter` no longer copies chunks into a
+  buffer per pass and then into the file's buffer. A chunk that is a
+  contiguous run of the dataset's data (a dataset stored as one chunk of
+  its shape, blocks of whole rows) is borrowed when unfiltered and
+  compressed straight from it when filtered, and contiguous datasets are
+  not copied. New `FileWriter::finish_with(put)` hands the file out piece
+  by piece; `FileBuilder::write` uses it (through a 1 MiB `BufWriter`,
+  still to a temporary file renamed into place), so a file is never
+  assembled in memory. New `DatasetBuilder::with_u8_data_owned(Vec<u8>)`
+  takes the data without the copy `with_u8_data` makes. Peak resident
+  memory of one unfiltered 1 GiB chunk of `u8`
+  (`cargo run --release -p clawhdf5 --example write_one_chunk --
+  1073741824 OUT owned|slice`, tank, 2026-09-29): 5.0 GiB before (the
+  writer of `4260af4`), 1.0 GiB after with `owned`; 6.0 and 2.0 GiB with
+  `slice`; 4 GiB + 7 bytes `owned`, 4.00 GiB. Same bytes written. Write
+  speed was not re-measured (the machine was not idle): the A/B of
+  `FileBuilder::write` and `finish` on small and large files is still to
+  be run.
+- `chunked_write::extract_chunk` (behind `split_into_chunks`) no longer
+  panics when the data is shorter than the shape; the missing part is
+  zeros, as it was meant to be.
+- Tests: `crates/clawhdf5/tests/fixtures/huge_chunk_dims.h5` (31 KB,
+  libhdf5 2.0.0 via h5py 3.16, `gen_huge_chunks.py dims`: `u8` chunks of
+  2^32 + 7 deflated twice, Single Chunk and Extensible Array), and in
+  `huge_chunks_interop`: `huge_chunk_dims_list` (always), and with
+  `CLAWHDF5_HUGE_CHUNKS=1` `huge_chunk_dims_read`,
+  `huge_chunk_dims_unfiltered_read` (a sparse file h5py writes, Single
+  Chunk and Fixed Array, mmap and positioned reads),
+  `writer_huge_chunk_dims_round_trip` (deflated, read back by clawhdf5,
+  h5py 3.16 and — layouts — h5dump 2.2.0),
+  `writer_unfiltered_huge_chunk_round_trip` (an unfiltered 4 GiB + 7 byte
+  chunk: clawhdf5, h5py, and h5dump 2.2.0 printing values past 2^32) and
+  `writer_huge_chunks_lz4_zstd_round_trip` (with the `lz4`/`zstd`
+  features; h5py through hdf5plugin 7.1.0). The whole opt-in suite
+  (`cargo test --release -p clawhdf5 --features lz4,zstd --test
+  huge_chunks_interop -- --test-threads=1`, h5py and h5dump 2.2.0
+  included) passed with a peak of 10.04 GiB resident (tank, 2026-09-29,
+  commit `0065c6b`). Unit tests: 5-byte dimension encoding, contiguous
+  chunk detection, `finish_with` against `finish`. The wasm package test
+  lists the new fixture and gets `FormatError::Overflow` reading it.
+  `docs/known-issues.md`: two fixed entries; the open "Chunks of 4 GiB or
+  more: limits" keeps the refused filters, the editor and memory.
+
 ### NetCDF-4: variables' dimensions come from the file (2026-09-28)
 - `clawhdf5-netcdf4` gave each variable the first unused dimension of
   equal size (else an anonymous `dim_<n>`), so a variable on an unlimited

README.md

@@ -102,7 +102,8 @@ writes 85.2 µs vs 877 µs (10.3x); 64 group creates 130 µs vs 1.37 ms
 (10.6x); a 512×512 `f32` chunked deflate-6 write 1.44 ms vs 65.0 ms
 (re-measured 2026-09-23 with the pure-Rust deflate: 1.46 ms vs 51.4 ms,
 35x); a 100K `f32` sequential write is a tie. The writer (`FileBuilder`)
-assembles a file in memory and writes it once, which is part of that
+lays out the whole file before writing it (when these were measured, in
+one buffer in memory), which is part of that
 difference; read the caveats in [BENCHMARKS.md](BENCHMARKS.md#caveats)
 before quoting these.
 
@@ -115,12 +116,12 @@ Limits and open issues, with dates, are in
 |---|---|---|---|
 | **File format** | Superblock v0–v3, user blocks, v1/v2 object headers; writing the HDF5 1.10 format (default) or, with `libver_bounds(LibVer::V18, LibVer::V18)`, files HDF5 1.8 reads (checked with HDF5 1.8.23) | Metadata cache images | Writing the pre-1.8 format (version-0 superblock, symbol-table groups) |
 | **Groups and links** | Symbol-table, compact and dense groups (tested to 100 000 links), creation order, soft and hard links; writing external links | | Following external links (explicit error); user-defined links are skipped |
-| **Datatypes** | Integers and IEEE floats of every width and byte order (incl. `f16`), enums, compounds (every version, incl. HDF5 2.0's v5), arrays, fixed-length strings, opaque, complex: h5py's `{r, i}` compound (`with_complex_f64_data`) and HDF5 2.0's native class 11 (`with_native_complex_f64_data`, opt-in: only libhdf5 2.0+ reads it; reads surface it as `{r, i}`) | Variable-length strings and sequences, object references; HDF5 2.x's small floats (bfloat16, FP8 E4M3/E5M2, FP6 E2M3/E3M2, FP4 E2M1: every bit pattern decoded as libhdf5 2.2.0 decodes it) and other non-IEEE floats up to 64 bits | Writing variable-length data; writing non-IEEE floats; decoding region and attribute references; x87 long double and binary128 |
-| **Layouts and chunk indexes** | Compact, contiguous and chunked; chunk indexes single chunk, Fixed Array, Extensible Array and v2 B-tree (the writer picks one as libhdf5 does), and v1 B-tree (every chunked dataset under the 1.8 bound, split as libhdf5 splits it); chunks of 4 GiB or more (HDF5 2.0's layout message version 5); fill values; resizable datasets; virtual datasets (read limits in known-issues) | The implicit chunk index (the editor also changes it) | External raw data files (explicit error); chunk dimensions of 2^32 or more |
+| **Datatypes** | Integers and IEEE floats of every width and byte order (incl. `f16`), enums, compounds (every version, incl. HDF5 2.0's v5), arrays, fixed-length strings, opaque, complex: h5py's `{r, i}` compound (`with_complex_f64_data`) and HDF5 2.0's native class 11 (`with_native_complex_f64_data`, opt-in: only libhdf5 2.0+ reads it; read as its own type, `DType::Complex`, and printed by `h5rs` as h5dump 2.x prints it) | Variable-length strings and sequences, object references; HDF5 2.x's small floats (bfloat16, FP8 E4M3/E5M2, FP6 E2M3/E3M2, FP4 E2M1: every bit pattern decoded as libhdf5 2.2.0 decodes it) and other non-IEEE floats up to 64 bits | Writing variable-length data; writing non-IEEE floats; decoding region and attribute references; x87 long double and binary128 |
+| **Layouts and chunk indexes** | Compact, contiguous and chunked; chunk indexes single chunk, Fixed Array, Extensible Array and v2 B-tree (the writer picks one as libhdf5 does), and v1 B-tree (every chunked dataset under the 1.8 bound, split as libhdf5 splits it); chunks of 4 GiB or more and chunk dimensions of 2^32 or more (HDF5 2.0's layout message version 5); fill values; resizable datasets; virtual datasets (read limits in known-issues) | The implicit chunk index (the editor also changes it) | External raw data files (explicit error) |
 | **Filters** | deflate (pure-Rust zlib-rs), shuffle, Fletcher-32, LZ4 (opt-in), Zstd (C, opt-in); plugins LZF, bitshuffle, bzip2, Blosc 1 | N-Bit, scale-offset, SZIP (C, opt-in); plugins Blosc2 and ZFP | Other filter IDs, unless you register a codec (`filter_registry::register_filter`) |
 | **Editing in place** | `FileEditor`: overwrite values, grow and shrink chunked datasets (every index), set attributes (compact and dense), in files from h5py or clawhdf5 | | Creating or deleting objects in an existing file; deleting attributes; new chunks in implicit indexes; VL data; rewriting chunks of 4 GiB or more; filters this build cannot encode (refused before any write) |
 | **Access** | Local files (mmap or buffered), bytes in memory, any `Storage` backend, HTTP(S) and S3/GCS/Azure via `clawhdf5-remote`, SWMR reading (`File::open_swmr`, `Dataset::refresh`) | Remote files and the browser are read-only | SWMR writing; remote SWMR; MPI collective I/O (`clawhdf5-io`'s `mpi-io` reads on one rank and broadcasts) |
-| **Bindings** | Python (read, `'w'` for numeric and complex arrays, `'r+'` editing, URLs), NetCDF-4 (CF scale/offset/fill) | WebAssembly (`open(bytes)`, `openUrl`); no Zstd/SZIP/pcodec, no compound, reference, opaque, bitfield, time or VL-sequence datasets | Node.js (the package does not work; see known-issues) |
+| **Bindings** | Python (read, `'w'` for numeric and complex arrays with h5py's `libver=`, `'r+'` editing, URLs), NetCDF-4 (CF scale/offset/fill) | WebAssembly (`open(bytes)`, `openUrl`; HDF5 2.0 native complex as `[re, im]` pairs); no Zstd/SZIP/pcodec, no compound (h5py's `{r, i}` complex included), reference, opaque, bitfield, time or VL-sequence datasets | Node.js (the package does not work; see known-issues) |
 
 Plugin filters other than LZF are cargo features (`bitshuffle`, `bzip2`,
 `blosc`, `blosc2`, `zfp`, or `plugin-filters` for all of them), all pure

crates/clawhdf5-format/src/chunked_read.rs

@@ -555,7 +555,7 @@ pub(crate) fn checked_byte_len(elements: u64, elem_size: usize) -> Result<usize,
 /// HDF5 2.0 writes them with layout version 5. A zero chunk dimension used
 /// to read as all fill values, and a huge one to hang the reader.
 pub(crate) fn chunk_geometry(
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     layout_version: u8,
     dataspace: &Dataspace,
     elem_size: usize,
@@ -576,15 +576,26 @@ pub(crate) fn chunk_geometry(
             "chunk size must be > 0, dim = {d}"
         )));
     }
-    let bytes = spatial
-        .iter()
-        .fold(elem_size as u128, |acc, &c| acc * u128::from(c));
+    // Saturating: 33 dimensions of up to 64 bits each can overflow u128.
+    let bytes = spatial.iter().fold(elem_size as u128, |acc, &c| {
+        acc.saturating_mul(u128::from(c))
+    });
+    // So the indexes can compute a chunk's size in 64 bits.
+    if bytes > u128::from(u64::MAX) {
+        return Err(FormatError::InvalidChunkDimensions(format!(
+            "chunk size overflows 64 bits (chunk {spatial:?} of {elem_size}-byte elements)"
+        )));
+    }
     if layout_version < 4 && bytes > u128::from(u32::MAX) {
         return Err(FormatError::InvalidChunkDimensions(format!(
             "chunk size must be < 4GB with v1 b-tree index (chunk {spatial:?} of {elem_size}-byte elements)"
         )));
     }
-    Ok((rank, spatial.iter().map(|&c| c as usize).collect()))
+    let spatial = spatial
+        .iter()
+        .map(|&c| to_usize(c))
+        .collect::<Result<Vec<usize>, _>>()?;
+    Ok((rank, spatial))
 }
 
 /// The size of one element of `dt` as stored in the file: a
@@ -625,7 +636,7 @@ pub fn check_chunk_element_size(
         return Ok(());
     };
     let expected = stored_element_size(datatype, offset_size);
-    if u64::from(stored) != expected {
+    if stored != expected {
         return Err(FormatError::InvalidChunkDimensions(format!(
             "stored datatype size in chunk layout does not match datatype description \
              (layout {stored} bytes, datatype {expected})"
@@ -759,7 +770,7 @@ pub fn collect_chunk_info_in<S: Storage + ?Sized>(
 pub fn collect_chunk_info_checked(
     file_data: &[u8],
     btree_address: u64,
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     offset_size: u8,
     length_size: u8,
 ) -> Result<Vec<ChunkInfo>, FormatError> {
@@ -776,7 +787,7 @@ pub fn collect_chunk_info_checked(
 pub fn collect_chunk_info_checked_in<S: Storage + ?Sized>(
     file_data: &S,
     btree_address: u64,
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     offset_size: u8,
     length_size: u8,
 ) -> Result<Vec<ChunkInfo>, FormatError> {
@@ -808,7 +819,7 @@ pub fn collect_chunk_info_checked_in<S: Storage + ?Sized>(
             .iter_mut()
             .zip(chunk.offsets.iter().zip(chunk_dimensions))
         {
-            *w = o / u64::from(d);
+            *w = o / d;
         }
         wanted[ndims - 1] = 0;
         if let Some(i) = root.find(&wanted) {
@@ -849,9 +860,9 @@ impl ChunkNode {
         }
     }
 
-    fn scale_keys(&mut self, dims: &[u32]) {
+    fn scale_keys(&mut self, dims: &[u64]) {
         for (k, &d) in self.keys.iter_mut().zip(dims.iter().cycle()) {
-            *k /= u64::from(d);
+            *k /= d;
         }
         if let ChunkChildren::Nodes(nodes) = &mut self.children {
             for n in nodes {
@@ -919,9 +930,9 @@ fn btree_cmp3(lt: &[u64], scaled: &[u64], rt: &[u64]) -> core::cmp::Ordering {
 
 /// Check one v1 B-tree chunk key's offsets (see
 /// [`collect_chunk_info_checked`]).
-fn check_key_offsets(offsets: &[u64], chunk_dimensions: &[u32]) -> Result<(), FormatError> {
+fn check_key_offsets(offsets: &[u64], chunk_dimensions: &[u64]) -> Result<(), FormatError> {
     for (&offset, &dim) in offsets.iter().zip(chunk_dimensions) {
-        if dim == 0 || offset % u64::from(dim) != 0 {
+        if dim == 0 || offset % dim != 0 {
             return Err(FormatError::ChunkedReadError(format!(
                 "bad coordinate offset {offsets:?} for chunk dimensions {chunk_dimensions:?}"
             )));
@@ -937,7 +948,7 @@ fn read_key_offsets(
     file_data: &[u8],
     pos: usize,
     ndims: usize,
-    chunk_dimensions: Option<&[u32]>,
+    chunk_dimensions: Option<&[u64]>,
     out: &mut Vec<u64>,
 ) -> Result<(), FormatError> {
     let start = out.len();
@@ -966,7 +977,7 @@ fn parse_chunk_node<S: Storage + ?Sized>(
     file_data: &S,
     btree_address: u64,
     ndims: usize,
-    chunk_dimensions: Option<&[u32]>,
+    chunk_dimensions: Option<&[u64]>,
     offset_size: u8,
     depth: usize,
     stored: &mut Vec<ChunkInfo>,
@@ -1075,7 +1086,7 @@ fn parse_chunk_node<S: Storage + ?Sized>(
 pub fn generate_implicit_chunks(
     base_address: u64,
     dataset_dims: &[u64],
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     element_size: u32,
 ) -> Vec<ChunkInfo> {
     generate_implicit_chunks_in_grid(
@@ -1097,17 +1108,18 @@ pub fn generate_implicit_chunks_in_grid(
     base_address: u64,
     dataset_dims: &[u64],
     max_dims: &[u64],
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     element_size: u32,
 ) -> Vec<ChunkInfo> {
     let rank = chunk_dimensions.len();
-    let chunk_byte_size: u64 =
-        chunk_dimensions.iter().map(|&d| d as u64).product::<u64>() * element_size as u64;
+    let chunk_byte_size: u64 = chunk_dimensions
+        .iter()
+        .fold(u64::from(element_size), |acc, &d| acc.saturating_mul(d));
 
     let mut num_chunks_per_dim = Vec::with_capacity(rank);
     let mut grid_per_dim = Vec::with_capacity(rank);
     for d in 0..rank {
-        let ch = chunk_dimensions[d] as u64;
+        let ch = chunk_dimensions[d];
         let n = dataset_dims[d].div_ceil(ch);
         num_chunks_per_dim.push(n);
         grid_per_dim.push(max_dims.get(d).map_or(n, |m| m.div_ceil(ch)).max(n));
@@ -1125,7 +1137,7 @@ pub fn generate_implicit_chunks_in_grid(
             let nchunks = num_chunks_per_dim[d];
             let chunk_idx = remaining % nchunks;
             remaining /= nchunks;
-            offsets[d] = chunk_idx * chunk_dimensions[d] as u64;
+            offsets[d] = chunk_idx * chunk_dimensions[d];
             grid_idx = grid_idx.saturating_add(chunk_idx.saturating_mul(down));
             down = down.saturating_mul(grid_per_dim[d]);
         }
@@ -1350,7 +1362,7 @@ pub fn list_chunks_in<S: Storage + ?Sized>(
         }
         (4, Some(2)) => {
             // Implicit index — use spatial chunk dims only
-            let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
+            let spatial_chunk_dims: &[u64] = &chunk_dimensions[..rank];
             generate_implicit_chunks_in_grid(
                 addr,
                 &dataspace.dimensions,
@@ -1364,7 +1376,7 @@ pub fn list_chunks_in<S: Storage + ?Sized>(
         }
         (4, Some(3)) => {
             // Fixed Array — use spatial chunk dims only
-            let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
+            let spatial_chunk_dims: &[u64] = &chunk_dimensions[..rank];
             let header = FixedArrayHeader::parse_in(
                 file_data,
                 checked_addr(addr)?,
@@ -1384,7 +1396,7 @@ pub fn list_chunks_in<S: Storage + ?Sized>(
         }
         (4, Some(4)) => {
             // Extensible Array — use spatial chunk dims only
-            let spatial_chunk_dims: &[u32] = &chunk_dimensions[..rank];
+            let spatial_chunk_dims: &[u64] = &chunk_dimensions[..rank];
             let header = ExtensibleArrayHeader::parse_in(
                 file_data,
                 checked_addr(addr)?,
@@ -2435,12 +2447,12 @@ mod tests {
     /// A leaf whose final key is the one libhdf5 writes past the last chunk:
     /// each coordinate of the last chunk plus its chunk dimension (the
     /// element size last).
-    fn build_chunk_btree_leaf_dims(chunks: &[ChunkInfo], dims: &[u32], offset_size: u8) -> Vec<u8> {
+    fn build_chunk_btree_leaf_dims(chunks: &[ChunkInfo], dims: &[u64], offset_size: u8) -> Vec<u8> {
         let last = &chunks.last().expect("a chunk").offsets;
         let end: Vec<u64> = dims
             .iter()
             .enumerate()
-            .map(|(d, &c)| last.get(d).copied().unwrap_or(0) + u64::from(c))
+            .map(|(d, &c)| last.get(d).copied().unwrap_or(0) + c)
             .collect();
         build_chunk_btree_leaf_to(chunks, &end, offset_size)
     }
@@ -2771,7 +2783,7 @@ mod tests {
         }
 
         // Build B-tree at offset 0x100
-        let dims = [chunk_size_elems as u32, elem_size as u32];
+        let dims = [chunk_size_elems as u64, elem_size as u64];
         let btree = build_chunk_btree_leaf_dims(&chunk_infos, &dims, os);
         let btree_addr = 0x100usize;
         file_data[btree_addr..btree_addr + btree.len()].copy_from_slice(&btree);
@@ -3145,7 +3157,7 @@ mod tests {
             data_offset += compressed.len() + 16; // some padding
         }
 
-        let dims = [chunk_elems as u32, elem_size as u32];
+        let dims = [chunk_elems as u64, elem_size as u64];
         let btree = build_chunk_btree_leaf_dims(&chunk_infos, &dims, os);
         let btree_addr = 0x100usize;
         file_data[btree_addr..btree_addr + btree.len()].copy_from_slice(&btree);
@@ -3228,7 +3240,7 @@ mod tests {
             }
         }
 
-        let dims = [chunk_dims[0] as u32, chunk_dims[1] as u32, elem_size as u32];
+        let dims = [chunk_dims[0] as u64, chunk_dims[1] as u64, elem_size as u64];
         let btree = build_chunk_btree_leaf_dims(&chunk_infos, &dims, os);
         let btree_addr = 0x100usize;
         file_data[btree_addr..btree_addr + btree.len()].copy_from_slice(&btree);
@@ -3408,7 +3420,7 @@ mod tests {
         }
 
         let layout = DataLayout::Chunked {
-            chunk_dimensions: vec![chunk_elems as u32, elem_size as u32],
+            chunk_dimensions: vec![chunk_elems as u64, elem_size as u64],
             btree_address: Some(data_addr as u64),
             version: 4,
             chunk_index_type: Some(1),

crates/clawhdf5-format/src/chunked_write.rs

@@ -5,12 +5,14 @@ extern crate alloc;
 
 use crate::addr::saturating_usize;
 #[cfg(not(feature = "std"))]
-use alloc::{format, vec, vec::Vec};
+use alloc::{borrow::Cow, format, vec, vec::Vec};
+#[cfg(feature = "std")]
+use std::borrow::Cow;
 
 use crate::btree_v1_write;
 use crate::btree_v2_write::{BTreeV2Params, build_btree_v2};
 use crate::checksum::jenkins_lookup3;
-use crate::chunk_cache::{CACHE_LINE_SIZE, align_to_cache_line};
+use crate::chunk_cache::CACHE_LINE_SIZE;
 use crate::chunk_grid::ChunkGrid;
 use crate::ea_writer;
 use crate::error::FormatError;
@@ -464,7 +466,9 @@ fn extract_chunk(
         let dst: usize = (0..rank).map(|d| idx[d] * chunk_strides[d]).sum::<usize>() * element_size;
         // Whole elements only, as far as `raw_data` reaches.
         let n = row.min(raw_data.len().saturating_sub(src)) / element_size * element_size;
+        if n > 0 {
             chunk[dst..dst + n].copy_from_slice(&raw_data[src..src + n]);
+        }
         // Next row: advance every dimension but the last.
         let mut d = rank - 1;
         loop {
@@ -481,6 +485,66 @@ fn extract_chunk(
     }
 }
 
+/// The bytes of the `linear_idx`-th chunk (see [`extract_chunk`]),
+/// borrowed from `raw_data` when the chunk is one contiguous run of it: a
+/// chunk inside the dataset (not past its edge) whose dimensions after the
+/// first one above 1 equal the dataset's. A dataset stored as one chunk of
+/// its own shape is such a run, so its chunk is never copied. Other chunks
+/// are copied out, as [`extract_chunk`] does.
+fn chunk_bytes_of<'a>(
+    raw_data: &'a [u8],
+    shape: &[u64],
+    chunk_dims: &[u64],
+    element_size: usize,
+    linear_idx: u64,
+) -> Cow<'a, [u8]> {
+    if shape.is_empty() {
+        return Cow::Borrowed(raw_data);
+    }
+    if let Some(range) = contiguous_chunk(shape, chunk_dims, element_size, linear_idx)
+        && let (Ok(start), Ok(end)) = (usize::try_from(range.start), usize::try_from(range.end))
+        && end <= raw_data.len()
+    {
+        return Cow::Borrowed(&raw_data[start..end]);
+    }
+    Cow::Owned(extract_chunk(raw_data, shape, chunk_dims, element_size, linear_idx).1)
+}
+
+/// The byte range of the `linear_idx`-th chunk in the row-major dataset,
+/// when the chunk is a contiguous run of it (see [`chunk_bytes_of`]).
+fn contiguous_chunk(
+    shape: &[u64],
+    chunk_dims: &[u64],
+    element_size: usize,
+    linear_idx: u64,
+) -> Option<core::ops::Range<u64>> {
+    let rank = shape.len();
+    // Past the first chunk dimension above 1, every chunk dimension must
+    // span the dataset's.
+    let first_wide = chunk_dims.iter().position(|&c| c > 1).unwrap_or(rank);
+    if (first_wide + 1..rank).any(|d| chunk_dims[d] != shape[d]) {
+        return None;
+    }
+    let mut remaining = linear_idx;
+    let mut start = 0u64;
+    let mut stride = element_size as u64;
+    for d in (0..rank).rev() {
+        let n = shape[d].div_ceil(chunk_dims[d]);
+        let offset = (remaining % n).checked_mul(chunk_dims[d])?;
+        remaining /= n;
+        // A chunk past the dataset's edge is padded: not a run of it.
+        if offset.checked_add(chunk_dims[d])? > shape[d] {
+            return None;
+        }
+        start = start.checked_add(offset.checked_mul(stride)?)?;
+        stride = stride.checked_mul(shape[d])?;
+    }
+    let len = chunk_dims
+        .iter()
+        .try_fold(element_size as u64, |acc, &c| acc.checked_mul(c))?;
+    Some(start..start.checked_add(len)?)
+}
+
 /// Parallel compression threshold: use rayon when chunk count exceeds this.
 ///
 /// Lowered to 2 to enable parallel compression for typical 4-chunk workloads
@@ -508,25 +572,21 @@ const PARALLEL_COMPRESS_MAX_CHUNK_BYTES: u64 = 64 << 20;
 /// at most [`PARALLEL_COMPRESS_MAX_CHUNK_BYTES`], compression runs across
 /// rayon threads; otherwise it is sequential. Output order matches chunk
 /// order, so per-chunk bytes are identical to the sequential path.
-fn compress_all_chunks(
-    raw_data: &[u8],
+fn compress_all_chunks<'a>(
+    raw_data: &'a [u8],
     shape: &[u64],
     chunk_dims: &[u64],
     element_size: usize,
     chunk_bytes: u64,
     pipeline: &Option<FilterPipeline>,
-) -> Result<Vec<(u64, Vec<u8>, u32)>, FormatError> {
-    let one = |i: u64| -> Result<(u64, Vec<u8>, u32), FormatError> {
-        let (_, raw) = if shape.is_empty() {
-            (Vec::new(), raw_data.to_vec())
-        } else {
-            extract_chunk(raw_data, shape, chunk_dims, element_size, i)
-        };
+) -> Result<Vec<PreparedChunk<'a>>, FormatError> {
+    let one = |i: u64| -> Result<PreparedChunk<'a>, FormatError> {
+        let raw = chunk_bytes_of(raw_data, shape, chunk_dims, element_size, i);
         let raw_size = raw.len() as u64;
         match pipeline {
             Some(pl) => {
                 let (stored, mask) = compress_chunk_masked(&raw, pl, element_size as u32)?;
-                Ok((raw_size, stored, mask))
+                Ok((raw_size, Cow::Owned(stored), mask))
             }
             None => Ok((raw_size, raw, 0)),
         }
@@ -557,7 +617,7 @@ fn compress_all_chunks(
 /// layout/pipeline messages. `base_address` is where the blob will be placed in the file.
 /// Serialize a v4 single chunk layout message (public for OH size estimation).
 pub fn serialize_v4_single_chunk_pub(
-    chunk_dims: &[u32],
+    chunk_dims: &[u64],
     chunk_address: u64,
     filtered_size: Option<u64>,
     filter_mask: Option<u32>,
@@ -578,7 +638,7 @@ pub fn serialize_v4_single_chunk_pub(
 /// Serialize a v4 (or, for a chunk of 4 GiB or more, v5) single chunk
 /// layout message.
 fn serialize_v4_single_chunk(
-    chunk_dims: &[u32],
+    chunk_dims: &[u64],
     chunk_address: u64,
     filtered_size: Option<u64>,
     filter_mask: Option<u32>,
@@ -622,7 +682,7 @@ fn serialize_v4_single_chunk(
 
 /// Serialize a v4 Fixed Array layout message.
 fn serialize_v4_fixed_array(
-    chunk_dims: &[u32],
+    chunk_dims: &[u64],
     fixed_array_address: u64,
     offset_size: u8,
     element_size: u32,
@@ -653,7 +713,8 @@ fn serialize_v4_fixed_array(
 /// dimensions, then the element size). Each takes the fewest bytes that hold
 /// the largest, as libhdf5 computes it (`H5D__chunk_set_sizes`:
 /// `(log2(dim) + 8) / 8`); HDF5 2.0.0 refuses any other width.
-pub(crate) fn push_v4_chunk_dims(buf: &mut Vec<u8>, chunk_dims: &[u32], element_size: u32) {
+pub(crate) fn push_v4_chunk_dims(buf: &mut Vec<u8>, chunk_dims: &[u64], element_size: u32) {
+    let element_size = u64::from(element_size);
     let max_dim = chunk_dims
         .iter()
         .copied()
@@ -661,14 +722,14 @@ pub(crate) fn push_v4_chunk_dims(buf: &mut Vec<u8>, chunk_dims: &[u32], element_
         .max()
         .unwrap_or(1)
         .max(1);
-    let width = (32 - max_dim.leading_zeros()).div_ceil(8) as usize;
+    let width = (64 - max_dim.leading_zeros()).div_ceil(8) as usize;
     buf.push(width as u8);
     for &d in chunk_dims.iter().chain(core::iter::once(&element_size)) {
         buf.extend_from_slice(&d.to_le_bytes()[..width]);
     }
 }
 
-fn layout_v4_chunked_prefix(chunk_dims: &[u32], element_size: u32, version: u8) -> Vec<u8> {
+fn layout_v4_chunked_prefix(chunk_dims: &[u64], element_size: u32, version: u8) -> Vec<u8> {
     let mut buf = Vec::new();
     buf.push(version);
     buf.push(2); // class = chunked
@@ -884,7 +945,66 @@ pub fn precompress_chunks(
     element_size: usize,
     options: &ChunkOptions,
 ) -> Result<PrecompressedChunks, FormatError> {
-    let (_, chunk_bytes) = checked_chunk_dims(chunk_dims, element_size)?;
+    let p = prepare_chunks(raw_data, shape, chunk_dims, element_size, options)?;
+    Ok(PrecompressedChunks {
+        chunks: p
+            .chunks
+            .into_iter()
+            .map(|(raw, stored, mask)| (raw, stored.into_owned(), mask))
+            .collect(),
+        has_filters: p.has_filters,
+        element_size: p.element_size,
+        shape: p.shape,
+        chunk_dims: p.chunk_dims,
+        pipeline_message: p.pipeline_message,
+    })
+}
+
+/// One chunk of [`PreparedChunks`]: its raw size, its stored bytes and its
+/// filter mask.
+pub(crate) type PreparedChunk<'a> = (u64, Cow<'a, [u8]>, u32);
+
+/// [`PrecompressedChunks`] whose stored bytes may borrow the dataset's data:
+/// an unfiltered chunk that is a contiguous run of it (a dataset stored as
+/// one chunk, say) is not copied. What the file writer lays out.
+pub(crate) struct PreparedChunks<'a> {
+    pub(crate) chunks: Vec<PreparedChunk<'a>>,
+    pub(crate) has_filters: bool,
+    pub(crate) element_size: usize,
+    pub(crate) shape: Vec<u64>,
+    pub(crate) chunk_dims: Vec<u64>,
+    pub(crate) pipeline_message: Option<Vec<u8>>,
+}
+
+impl PrecompressedChunks {
+    /// A view of these chunks as [`PreparedChunks`] (their bytes borrowed).
+    fn prepared(&self) -> PreparedChunks<'_> {
+        PreparedChunks {
+            chunks: self
+                .chunks
+                .iter()
+                .map(|(raw, stored, mask)| (*raw, Cow::Borrowed(&stored[..]), *mask))
+                .collect(),
+            has_filters: self.has_filters,
+            element_size: self.element_size,
+            shape: self.shape.clone(),
+            chunk_dims: self.chunk_dims.clone(),
+            pipeline_message: self.pipeline_message.clone(),
+        }
+    }
+}
+
+/// [`precompress_chunks`], borrowing what it can (see [`PreparedChunks`]).
+/// A filtered chunk that is a contiguous run of `raw_data` is compressed
+/// straight from it, without a copy of the raw chunk.
+pub(crate) fn prepare_chunks<'a>(
+    raw_data: &'a [u8],
+    shape: &[u64],
+    chunk_dims: &[u64],
+    element_size: usize,
+    options: &ChunkOptions,
+) -> Result<PreparedChunks<'a>, FormatError> {
+    let chunk_bytes = checked_chunk_dims(chunk_dims, element_size)?;
     if chunk_bytes > MAX_V4_CHUNK_BYTES {
         check_huge_chunk_filters(options, chunk_bytes)?;
     }
@@ -902,7 +1022,7 @@ pub fn precompress_chunks(
         &pipeline,
     )?;
 
-    Ok(PrecompressedChunks {
+    Ok(PreparedChunks {
         chunks,
         has_filters,
         element_size,
@@ -912,26 +1032,87 @@ pub fn precompress_chunks(
     })
 }
 
-/// The chunk dimensions as the layout message stores them (each below
-/// 2^32), and one chunk's size in bytes. A chunk dimension of 2^32 or more
-/// (which HDF5 2.0 can store, in wider fields) is refused, as it is when
-/// read: clawhdf5 holds chunk dimensions as `u32`. So is a chunk whose size
-/// overflows 64 bits, or that this platform cannot hold in memory (a chunk
-/// of 4 GiB or more on a 32-bit target).
-fn checked_chunk_dims(
-    chunk_dims: &[u64],
-    element_size: usize,
-) -> Result<(Vec<u32>, u64), FormatError> {
-    let dims = chunk_dims
-        .iter()
-        .map(|&d| {
-            u32::try_from(d).map_err(|_| {
-                FormatError::InvalidChunkDimensions(format!(
-                    "chunk dimension {d} is 2^32 or more, which clawhdf5 does not support"
-                ))
-            })
-        })
-        .collect::<Result<Vec<u32>, _>>()?;
+/// A piece of a chunked dataset's data blob as laid out by
+/// [`place_chunked_data`]: zero padding, a chunk's stored bytes (by index
+/// into [`PreparedChunks::chunks`]), or index structures.
+#[derive(Debug)]
+pub(crate) enum Piece {
+    Zeros(usize),
+    Chunk(usize),
+    Bytes(Vec<u8>),
+}
+
+/// A chunked dataset laid out at an address, without its chunk bytes
+/// copied: the pieces of its data blob, in order, and its messages.
+pub(crate) struct ChunkedPlacement {
+    pub(crate) pieces: Vec<Piece>,
+    /// Length of the data blob in bytes.
+    pub(crate) len: u64,
+    pub(crate) layout_message: Vec<u8>,
+    pub(crate) pipeline_message: Option<Vec<u8>>,
+}
+
+impl ChunkedPlacement {
+    /// The data blob as one buffer (the chunks copied in).
+    fn into_result(self, pre: &PreparedChunks<'_>) -> ChunkedDataResult {
+        let mut data_bytes = Vec::with_capacity(saturating_usize(self.len));
+        for piece in &self.pieces {
+            match piece {
+                Piece::Zeros(n) => data_bytes.resize(data_bytes.len() + n, 0),
+                Piece::Chunk(i) => data_bytes.extend_from_slice(&pre.chunks[*i].1),
+                Piece::Bytes(b) => data_bytes.extend_from_slice(b),
+            }
+        }
+        ChunkedDataResult {
+            data_bytes,
+            layout_message: self.layout_message,
+            pipeline_message: self.pipeline_message,
+        }
+    }
+}
+
+/// Builds the pieces of a data blob, tracking its length.
+#[derive(Default)]
+struct Pieces {
+    pieces: Vec<Piece>,
+    len: u64,
+}
+
+impl Pieces {
+    /// Pad to the next cache line.
+    fn align(&mut self) {
+        let aligned = align_chunk_offset(self.len);
+        if aligned > self.len {
+            self.pieces
+                .push(Piece::Zeros(saturating_usize(aligned - self.len)));
+            self.len = aligned;
+        }
+    }
+
+    fn chunk(&mut self, i: usize, len: usize) {
+        self.pieces.push(Piece::Chunk(i));
+        self.len += len as u64;
+    }
+
+    fn bytes(&mut self, b: Vec<u8>) {
+        self.len += b.len() as u64;
+        self.pieces.push(Piece::Bytes(b));
+    }
+}
+
+/// One chunk's size in bytes. A chunk dimension of 0 is refused, and so is
+/// a chunk whose size overflows 64 bits, or that this platform cannot hold
+/// in memory (a chunk of 4 GiB or more on a 32-bit target). Chunk
+/// dimensions of 2^32 or more are allowed, as in libhdf5 2.x: such a chunk
+/// is over 4 GiB, so it takes layout message version 5, whose dimension
+/// fields are up to 8 bytes wide (a version-3 layout, with 4-byte fields,
+/// is never written for it).
+fn checked_chunk_dims(chunk_dims: &[u64], element_size: usize) -> Result<u64, FormatError> {
+    if chunk_dims.contains(&0) {
+        return Err(FormatError::InvalidChunkDimensions(format!(
+            "chunk dimensions {chunk_dims:?} include 0"
+        )));
+    }
     let bytes = chunk_dims
         .iter()
         .try_fold(element_size as u64, |acc, &d| acc.checked_mul(d))
@@ -941,7 +1122,7 @@ fn checked_chunk_dims(
                 "a chunk of {chunk_dims:?} x {element_size} bytes exceeds this platform's address space"
             ))
         })?;
-    Ok((dims, bytes))
+    Ok(bytes)
 }
 
 /// The filters clawhdf5 can apply to a chunk of more than `u32::MAX` bytes:
@@ -1010,7 +1191,22 @@ pub fn build_chunked_data_from_precompressed_libver(
     low: LibVer,
     high: LibVer,
 ) -> Result<ChunkedDataResult, FormatError> {
-    let (_, chunk_bytes) = checked_chunk_dims(&pre.chunk_dims, pre.element_size)?;
+    let pre = pre.prepared();
+    let placement = place_chunked_data(&pre, base_address, maxshape, low, high)?;
+    Ok(placement.into_result(&pre))
+}
+
+/// [`build_chunked_data_from_precompressed_libver`] without copying the
+/// chunks into one buffer: the data blob as [`Piece`]s referring to
+/// `pre`'s chunks. The file writer writes the pieces out one by one.
+pub(crate) fn place_chunked_data(
+    pre: &PreparedChunks<'_>,
+    base_address: u64,
+    maxshape: Option<&[u64]>,
+    low: LibVer,
+    high: LibVer,
+) -> Result<ChunkedPlacement, FormatError> {
+    let chunk_bytes = checked_chunk_dims(&pre.chunk_dims, pre.element_size)?;
     if chunk_bytes > MAX_V4_CHUNK_BYTES {
         if high < LibVer::V200 {
             return Err(FormatError::LibverBound {
@@ -1020,7 +1216,7 @@ pub fn build_chunked_data_from_precompressed_libver(
             });
         }
     } else if low < LibVer::V110 {
-        return build_btree_v1_chunked_data(pre, base_address, maxshape);
+        return place_btree_v1_chunked_data(pre, base_address, maxshape);
     }
     let index = ChunkIndexPlan::new(&pre.shape, maxshape, &pre.chunk_dims)?;
     let offset_size: u8 = 8;
@@ -1028,17 +1224,14 @@ pub fn build_chunked_data_from_precompressed_libver(
     let num_chunks = pre.chunks.len();
     let element_size = pre.element_size;
 
-    let mut data_buf = Vec::new();
+    let mut blob = Pieces::default();
     let mut written_chunks = Vec::with_capacity(num_chunks);
 
-    for (raw_size, compressed, filter_mask) in &pre.chunks {
-        let aligned_offset = align_to_cache_line(data_buf.len());
-        if aligned_offset > data_buf.len() {
-            data_buf.resize(aligned_offset, 0u8);
-        }
-        let address = base_address + data_buf.len() as u64;
+    for (i, (raw_size, compressed, filter_mask)) in pre.chunks.iter().enumerate() {
+        blob.align();
+        let address = base_address + blob.len;
         let compressed_size = compressed.len() as u64;
-        data_buf.extend_from_slice(compressed);
+        blob.chunk(i, compressed.len());
         written_chunks.push(WrittenChunk {
             address,
             compressed_size,
@@ -1047,28 +1240,22 @@ pub fn build_chunked_data_from_precompressed_libver(
         });
     }
 
-    let (chunk_dims_u32, chunk_bytes) = checked_chunk_dims(&pre.chunk_dims, element_size)?;
     let version = layout_version_for(chunk_bytes);
-
-    let aligned_idx = align_to_cache_line(data_buf.len());
-    if aligned_idx > data_buf.len() {
-        data_buf.resize(aligned_idx, 0u8);
-    }
+    blob.align();
 
     let layout_message = match &index {
         ChunkIndexPlan::ExtensibleArray(grid) => {
-            let ea_address = base_address + data_buf.len() as u64;
+            let ea_address = base_address + blob.len;
             let slots = index_slots(grid, &pre.shape, &pre.chunk_dims, &written_chunks, None)?;
-            let ea_bytes = ea_writer::build_extensible_array_at(
+            blob.bytes(ea_writer::build_extensible_array_at(
                 &slots,
                 offset_size,
                 length_size,
                 pre.has_filters,
                 ea_address,
-            );
-            data_buf.extend_from_slice(&ea_bytes);
+            ));
             ea_writer::serialize_v4_extensible_array(
-                &chunk_dims_u32,
+                &pre.chunk_dims,
                 ea_address,
                 offset_size,
                 element_size as u32,
@@ -1084,7 +1271,7 @@ pub fn build_chunked_data_from_precompressed_libver(
             };
             let filter_mask = pre.has_filters.then_some(written_chunks[0].filter_mask);
             serialize_v4_single_chunk(
-                &chunk_dims_u32,
+                &pre.chunk_dims,
                 chunk_addr,
                 filtered_size,
                 filter_mask,
@@ -1094,7 +1281,7 @@ pub fn build_chunked_data_from_precompressed_libver(
             )
         }
         ChunkIndexPlan::FixedArray(grid, nslots) => {
-            let fa_address = base_address + data_buf.len() as u64;
+            let fa_address = base_address + blob.len;
             let slots = index_slots(
                 grid,
                 &pre.shape,
@@ -1102,16 +1289,15 @@ pub fn build_chunked_data_from_precompressed_libver(
                 &written_chunks,
                 Some(*nslots),
             )?;
-            let fa_bytes = build_fixed_array_at(
+            blob.bytes(build_fixed_array_at(
                 &slots,
                 offset_size,
                 length_size,
                 pre.has_filters,
                 fa_address,
-            );
-            data_buf.extend_from_slice(&fa_bytes);
+            ));
             serialize_v4_fixed_array(
-                &chunk_dims_u32,
+                &pre.chunk_dims,
                 fa_address,
                 offset_size,
                 element_size as u32,
@@ -1120,7 +1306,7 @@ pub fn build_chunked_data_from_precompressed_libver(
             )
         }
         ChunkIndexPlan::BTreeV2 => {
-            let bt_address = base_address + data_buf.len() as u64;
+            let bt_address = base_address + blob.len;
             let records: Vec<(Vec<u64>, &WrittenChunk)> = written_chunks
                 .iter()
                 .enumerate()
@@ -1134,9 +1320,9 @@ pub fn build_chunked_data_from_precompressed_libver(
                 pre.has_filters,
                 bt_address,
             )?;
-            data_buf.extend_from_slice(&bt_bytes);
+            blob.bytes(bt_bytes);
             serialize_v4_btree_v2(
-                &chunk_dims_u32,
+                &pre.chunk_dims,
                 bt_address,
                 offset_size,
                 element_size as u32,
@@ -1146,8 +1332,9 @@ pub fn build_chunked_data_from_precompressed_libver(
         }
     };
 
-    Ok(ChunkedDataResult {
-        data_bytes: data_buf,
+    Ok(ChunkedPlacement {
+        pieces: blob.pieces,
+        len: blob.len,
         layout_message,
         pipeline_message: pre.pipeline_message.clone(),
     })
@@ -1156,11 +1343,11 @@ pub fn build_chunked_data_from_precompressed_libver(
 /// Lay out precompressed chunks at `base_address` followed by a version-1
 /// B-tree chunk index, with a version-3 layout message: what libhdf5 writes
 /// for a chunked dataset under a low bound of 1.8.
-fn build_btree_v1_chunked_data(
-    pre: &PrecompressedChunks,
+fn place_btree_v1_chunked_data(
+    pre: &PreparedChunks<'_>,
     base_address: u64,
     maxshape: Option<&[u64]>,
-) -> Result<ChunkedDataResult, FormatError> {
+) -> Result<ChunkedPlacement, FormatError> {
     if let Some(ms) = maxshape {
         let bad = |what: &str| FormatError::ChunkedReadError(format!("maxshape: {what}"));
         if ms.len() != pre.shape.len() {
@@ -1171,20 +1358,17 @@ fn build_btree_v1_chunked_data(
         }
     }
     let offset_size: u8 = 8;
-    let mut data_buf = Vec::new();
+    let mut blob = Pieces::default();
     let mut entries = Vec::with_capacity(pre.chunks.len());
     for (i, (_raw_size, stored, filter_mask)) in pre.chunks.iter().enumerate() {
-        let aligned_offset = align_to_cache_line(data_buf.len());
-        if aligned_offset > data_buf.len() {
-            data_buf.resize(aligned_offset, 0u8);
-        }
+        blob.align();
         entries.push(btree_v1_write::ChunkEntry {
             scaled: scaled_coords(&pre.shape, &pre.chunk_dims, i),
             nbytes: stored.len() as u64,
             filter_mask: *filter_mask,
-            address: base_address + data_buf.len() as u64,
+            address: base_address + blob.len,
         });
-        data_buf.extend_from_slice(stored);
+        blob.chunk(i, stored.len());
     }
     let element_size = u32::try_from(pre.element_size)
         .map_err(|_| FormatError::Overflow("element size".into()))?;
@@ -1193,11 +1377,8 @@ fn build_btree_v1_chunked_data(
     let btree_address = if entries.is_empty() {
         u64::MAX
     } else {
-        let aligned_idx = align_to_cache_line(data_buf.len());
-        if aligned_idx > data_buf.len() {
-            data_buf.resize(aligned_idx, 0u8);
-        }
-        let addr = base_address + data_buf.len() as u64;
+        blob.align();
+        let addr = base_address + blob.len;
         let tree = btree_v1_write::build_chunk_btree_v1_at(
             &entries,
             &pre.chunk_dims,
@@ -1205,13 +1386,14 @@ fn build_btree_v1_chunked_data(
             addr,
             offset_size,
         )?;
-        data_buf.extend_from_slice(&tree);
+        blob.bytes(tree);
         addr
     };
     let layout_message =
         serialize_v3_chunked(&pre.chunk_dims, btree_address, offset_size, element_size)?;
-    Ok(ChunkedDataResult {
-        data_bytes: data_buf,
+    Ok(ChunkedPlacement {
+        pieces: blob.pieces,
+        len: blob.len,
         layout_message,
         pipeline_message: pre.pipeline_message.clone(),
     })
@@ -1424,7 +1606,7 @@ fn build_btree_v2_chunk_index_at(
 
 /// Serialize a v4 layout message for a version-2 B-tree chunk index.
 fn serialize_v4_btree_v2(
-    chunk_dims: &[u32],
+    chunk_dims: &[u64],
     btree_address: u64,
     offset_size: u8,
     element_size: u32,
@@ -2184,7 +2366,7 @@ mod tests {
     /// filtered index element stores the chunk's size in 8 bytes.
     #[test]
     fn huge_chunk_layout_messages_and_index_elements() {
-        let dims = [HUGE_DIM as u32];
+        let dims = [HUGE_DIM];
         let parsed = |msg: &[u8]| {
             assert_eq!(msg[0], 5, "layout message version");
             // Class chunked, then (after the flags) 2 dimensions of 4 bytes.
@@ -2197,7 +2379,7 @@ mod tests {
                     chunk_index_type,
                     ..
                 } => {
-                    assert_eq!(chunk_dimensions, vec![HUGE_DIM as u32, 8]);
+                    assert_eq!(chunk_dimensions, vec![HUGE_DIM, 8]);
                     chunk_index_type.unwrap()
                 }
                 other => panic!("{other:?}"),
@@ -2245,22 +2427,89 @@ mod tests {
         assert_eq!(u16::from_le_bytes([bt[10], bt[11]]), 28);
     }
 
+    /// A chunk dimension of 2^32 or more (a 1-byte element, 2^32 + 7 per
+    /// chunk): the dimensions take 5 bytes each, as libhdf5 2.x writes
+    /// them (`(log2(dim) + 8) / 8`), and parse back whole.
+    #[test]
+    fn chunk_dimensions_of_2_pow_32_or_more_take_5_bytes() {
+        const DIM: u64 = (1 << 32) + 7;
+        let msg = serialize_v4_single_chunk(&[DIM], 0x800, None, None, 8, 1, 5);
+        assert_eq!(&msg[..5], &[5, 2, 0, 2, 5]);
+        assert_eq!(&msg[5..10], &[7, 0, 0, 0, 1]);
+        assert_eq!(&msg[10..15], &[1, 0, 0, 0, 0]);
+        assert_eq!(msg[15], 1, "Single Chunk index");
+        match DataLayout::parse(&msg, 8, 8).unwrap() {
+            DataLayout::Chunked {
+                chunk_dimensions, ..
+            } => assert_eq!(chunk_dimensions, [DIM, 1]),
+            other => panic!("{other:?}"),
+        }
+        // The largest dimension takes all 8 bytes.
+        let mut buf = Vec::new();
+        push_v4_chunk_dims(&mut buf, &[u64::MAX, 3], 4);
+        assert_eq!(buf[0], 8);
+        assert_eq!(buf.len(), 1 + 3 * 8);
+        // A version-3 layout has 4-byte dimensions: refused, not truncated.
+        assert!(serialize_v3_chunked(&[DIM], 0x800, 8, 1).is_err());
+    }
+
+    /// Which chunks are contiguous runs of the dataset's data (borrowed,
+    /// not copied, when unfiltered).
+    #[test]
+    fn contiguous_chunks_are_borrowed() {
+        // Rows of a 2-D dataset: chunk (2, 5) of shape (5, 5).
+        assert_eq!(contiguous_chunk(&[5, 5], &[2, 5], 4, 0), Some(0..40));
+        assert_eq!(contiguous_chunk(&[5, 5], &[2, 5], 4, 1), Some(40..80));
+        // The last row chunk runs past the edge: padded, so copied.
+        assert_eq!(contiguous_chunk(&[5, 5], &[2, 5], 4, 2), None);
+        // Columns split: not contiguous.
+        assert_eq!(contiguous_chunk(&[4, 6], &[2, 3], 1, 0), None);
+        // Leading dimensions of 1, then a partial row: contiguous.
+        assert_eq!(contiguous_chunk(&[2, 3, 8], &[1, 1, 4], 2, 5), Some(40..48));
+        // The whole dataset as one chunk.
+        assert_eq!(contiguous_chunk(&[3, 4], &[3, 4], 8, 0), Some(0..96));
+
+        let raw: Vec<u8> = (0..100).collect();
+        let whole = chunk_bytes_of(&raw, &[10, 10], &[10, 10], 1, 0);
+        assert!(matches!(whole, Cow::Borrowed(b) if b.as_ptr() == raw.as_ptr()));
+        let rows = chunk_bytes_of(&raw, &[10, 10], &[3, 10], 1, 1);
+        assert!(matches!(rows, Cow::Borrowed(b) if b == &raw[30..60]));
+        // Data shorter than the shape is copied (and zero-padded) as before.
+        let short = chunk_bytes_of(&raw[..50], &[10, 10], &[10, 10], 1, 0);
+        assert!(matches!(&short, Cow::Owned(v) if v.len() == 100));
+        // Borrowed or copied, the bytes are what extract_chunk gives.
+        for (shape, chunk) in [
+            (&[10u64, 10][..], &[3u64, 10][..]),
+            (&[10, 10], &[3, 4]),
+            (&[4, 5, 5], &[1, 5, 5]),
+            (&[4, 5, 5], &[1, 2, 5]),
+        ] {
+            for i in 0..chunk_count(shape, chunk) {
+                let got = chunk_bytes_of(&raw, shape, chunk, 1, i);
+                assert_eq!(&got[..], &extract_chunk(&raw, shape, chunk, 1, i).1[..]);
+            }
+        }
+    }
+
     #[test]
     fn huge_chunks_refused_where_unsupported() {
-        // A chunk dimension of 2^32 or more.
+        // A chunk dimension of 0.
         assert!(matches!(
-            checked_chunk_dims(&[1 << 32], 1),
-            Err(FormatError::InvalidChunkDimensions(m)) if m.contains("2^32")
+            checked_chunk_dims(&[4, 0], 1),
+            Err(FormatError::InvalidChunkDimensions(_))
         ));
+        // A chunk dimension of 2^32 or more is fine (on a 64-bit target).
+        #[cfg(target_pointer_width = "64")]
+        assert_eq!(
+            checked_chunk_dims(&[(1 << 32) + 5], 1).unwrap(),
+            (1 << 32) + 5
+        );
         // A chunk size that overflows u64.
         assert!(matches!(
             checked_chunk_dims(&[u32::MAX.into(), u32::MAX.into(), 2], 8),
             Err(FormatError::Overflow(_))
         ));
-        assert_eq!(
-            checked_chunk_dims(&[HUGE_DIM], 8).unwrap(),
-            (vec![HUGE_DIM as u32], HUGE_BYTES)
-        );
+        assert_eq!(checked_chunk_dims(&[HUGE_DIM], 8).unwrap(), HUGE_BYTES);
         // Filters that cannot take a chunk that large.
         for plugin in [
             PluginFilter::Lzf,

crates/clawhdf5-format/src/data_layout.rs

@@ -34,7 +34,7 @@ const MAX_LAYOUT_NDIMS: usize = 33;
 /// (`H5O__layout_decode`): at most [`MAX_LAYOUT_NDIMS`], no dimension 0, and
 /// before version 4 at least one dataspace dimension plus the element size.
 /// A zero chunk dimension used to read the dataset as all fill values.
-fn check_chunk_dims(dims: Vec<u32>, layout_version: u8) -> Result<Vec<u32>, FormatError> {
+fn check_chunk_dims(dims: Vec<u64>, layout_version: u8) -> Result<Vec<u64>, FormatError> {
     if dims.len() > MAX_LAYOUT_NDIMS {
         return Err(FormatError::InvalidChunkDimensions(
             "dimensionality is too large".into(),
@@ -72,7 +72,7 @@ pub enum DataLayout {
     /// Chunked: data stored in chunks via a B-tree.
     Chunked {
         /// Chunk dimension sizes.
-        chunk_dimensions: Vec<u32>,
+        chunk_dimensions: Vec<u64>,
         /// B-tree address, or `None` if undefined.
         btree_address: Option<u64>,
         /// Layout version (3 or 4). Version 1/2 messages (HDF5 1.4/1.6-era)
@@ -404,11 +404,11 @@ impl DataLayout {
             _ => return Err(FormatError::InvalidLayoutClass(layout_class)),
         };
         ensure_len(data, p, dimensionality * 4)?;
-        let dims: Vec<u32> = data[p..p + dimensionality * 4]
+        let dims: Vec<u64> = data[p..p + dimensionality * 4]
             .as_chunks::<4>()
             .0
             .iter()
-            .map(|c| u32::from_le_bytes(*c))
+            .map(|c| u64::from(u32::from_le_bytes(*c)))
             .collect();
         p += dimensionality * 4;
         match layout_class {
@@ -424,7 +424,7 @@ impl DataLayout {
             1 => {
                 let size = dims
                     .iter()
-                    .try_fold(1u64, |acc, &d| acc.checked_mul(d as u64))
+                    .try_fold(1u64, |acc, &d| acc.checked_mul(d))
                     .ok_or_else(|| {
                         FormatError::Overflow(format!("contiguous layout size {dims:?}"))
                     })?;
@@ -489,7 +489,12 @@ impl DataLayout {
                 ensure_len(data, p, dimensionality * 4)?;
                 let mut chunk_dimensions = Vec::with_capacity(dimensionality);
                 for _ in 0..dimensionality {
-                    let dim = u32::from_le_bytes([data[p], data[p + 1], data[p + 2], data[p + 3]]);
+                    let dim = u64::from(u32::from_le_bytes([
+                        data[p],
+                        data[p + 1],
+                        data[p + 2],
+                        data[p + 3],
+                    ]));
                     chunk_dimensions.push(dim);
                     p += 4;
                 }
@@ -565,14 +570,8 @@ impl DataLayout {
                         .iter()
                         .rev()
                         .fold(0u64, |acc, &b| (acc << 8) | u64::from(b));
-                    // Chunk dimensions are held as u32; HDF5 2.0 can write
-                    // larger ones (layout version 5), which are refused
-                    // rather than truncated.
-                    let val = u32::try_from(val).map_err(|_| {
-                        FormatError::InvalidChunkDimensions(format!(
-                            "chunk dimension {val} is larger than 2^32 - 1, which is not supported"
-                        ))
-                    })?;
+                    // HDF5 2.0 writes dimensions of 2^32 or more (layout
+                    // version 5, chunks over 4 GiB) in 5 to 8 bytes.
                     chunk_dimensions.push(val);
                     p += dim_size_encoded_length;
                 }
@@ -858,7 +857,7 @@ mod tests {
                 .unwrap_or_else(|e| panic!("width {width}: {e:?}"));
             assert!(
                 matches!(&layout, DataLayout::Chunked { chunk_dimensions, .. }
-                    if chunk_dimensions.iter().map(|&d| u64::from(d)).eq(dims)),
+                    if chunk_dimensions[..] == dims),
                 "width {width}: {layout:?}"
             );
         }
@@ -871,13 +870,14 @@ mod tests {
                 )
             );
         }
-        // A dimension past u32 cannot be represented and is refused, not
-        // truncated.
+        // HDF5 2.0 writes dimensions of 2^32 or more (up to 8 bytes each).
+        for (width, dim) in [(5u8, (1u64 << 32) + 3), (8, u64::MAX)] {
             assert!(matches!(
-            DataLayout::parse(&v4_chunked_msg(5, &[1 << 32, 8]), 8, 8),
-            Err(FormatError::InvalidChunkDimensions(m)) if m.contains("2^32")
+                DataLayout::parse(&v4_chunked_msg(width, &[dim, 1]), 8, 8),
+                Ok(DataLayout::Chunked { chunk_dimensions, .. }) if chunk_dimensions == [dim, 1]
             ));
         }
+    }
 
     #[test]
     fn v1v2_rejects_bad_class_dimensionality_and_truncation() {

crates/clawhdf5-format/src/datatype.rs

@@ -145,13 +145,16 @@ pub enum Datatype {
     /// imaginary, in rectangular form. `size` is twice the base size and the
     /// base is an IEEE float.
     ///
-    /// This variant exists for **writing** (see
+    /// [`Datatype::parse`] returns this variant for every class-11 message,
+    /// also inside compounds, arrays and variable-length types. Readers that
+    /// want the member view (`{r, i}`, as h5py writes complex numbers by
+    /// default) use [`Datatype::complex_as_compound`]; `read_compound_fields`
+    /// accepts this variant directly.
+    ///
+    /// Writing it is opt-in (see
     /// `type_builders::make_native_complex_f64_type`): only libhdf5 2.0 and
-    /// newer can read class 11, so it is opt-in and h5py's compound `{r, i}`
-    /// stays the default complex encoding. [`Datatype::parse`] still
-    /// surfaces a class-11 message as that equivalent `{r, i}` compound, so
-    /// every compound reader handles both encodings; parsing what this
-    /// variant serializes therefore yields a `Compound`, not a `Complex`.
+    /// newer can read class 11, so h5py's compound `{r, i}` stays the
+    /// default complex encoding.
     Complex { size: u32, base_type: Box<Datatype> },
 }
 
@@ -857,9 +860,7 @@ impl Datatype {
                 // Complex number (HDF5 2.0, datatype version 5). The properties
                 // are a single base floating-point datatype message; an element
                 // is two consecutive base-type values (real, imaginary). There
-                // is no member list. Surface it as the equivalent two-member
-                // compound `{r, i}` — the same shape h5py writes for numpy
-                // complex dtypes — so downstream compound readers work as-is.
+                // is no member list.
                 if version != 5 {
                     return Err(FormatError::InvalidDatatypeVersion {
                         class: class_id,
@@ -875,7 +876,13 @@ impl Datatype {
                         actual: size as usize,
                     });
                 }
-                Ok((Self::complex_as_compound(size, &base_type), pos))
+                Ok((
+                    Datatype::Complex {
+                        size,
+                        base_type: Box::new(base_type),
+                    },
+                    pos,
+                ))
             }
             _ => Err(FormatError::InvalidDatatypeClass(class_id)),
         }
@@ -1174,9 +1181,8 @@ impl Datatype {
 
     /// The `{r, i}` compound equivalent to a native complex type of `size`
     /// bytes over `base_type`: `r` at offset 0, `i` right after it — the
-    /// shape h5py writes for numpy complex dtypes, and what [`Self::parse`]
-    /// returns for a class-11 message. Readers that meet a
-    /// [`Datatype::Complex`] handle it through this view.
+    /// shape h5py writes for numpy complex dtypes. Readers that want a
+    /// [`Datatype::Complex`] as members handle it through this view.
     pub fn complex_as_compound(size: u32, base_type: &Datatype) -> Datatype {
         let base_size = base_type.type_size();
         Datatype::Compound {
@@ -1849,21 +1855,24 @@ mod tests {
     fn test_complex_v5_from_hdf5_2_0() {
         let (dt, consumed) = Datatype::parse(&COMPLEX_F64_HDF5_2_0).unwrap();
         assert_eq!(consumed, COMPLEX_F64_HDF5_2_0.len());
-        match dt {
-            Datatype::Compound { size, members } => {
-                assert_eq!(size, 16);
-                assert_eq!(members.len(), 2);
-                assert_eq!((members[0].name.as_str(), members[0].byte_offset), ("r", 0));
-                assert_eq!((members[1].name.as_str(), members[1].byte_offset), ("i", 8));
-                for m in &members {
+        match &dt {
+            Datatype::Complex { size, base_type } => {
+                assert_eq!(*size, 16);
                 assert!(matches!(
-                        m.datatype,
+                    base_type.as_ref(),
                     Datatype::FloatingPoint { size: 8, .. }
                 ));
             }
+            other => panic!("expected Complex, got {other:?}"),
         }
-            other => panic!("expected Compound, got {other:?}"),
-        }
+        // The member view is h5py's `{r, i}` compound.
+        let Datatype::Compound { members, .. } =
+            Datatype::complex_as_compound(16, &crate::type_builders::make_f64_type())
+        else {
+            unreachable!()
+        };
+        assert_eq!((members[0].name.as_str(), members[0].byte_offset), ("r", 0));
+        assert_eq!((members[1].name.as_str(), members[1].byte_offset), ("i", 8));
     }
 
     #[test]
@@ -1887,7 +1896,7 @@ mod tests {
                 assert_eq!(members.len(), 2);
                 assert!(matches!(
                     &members[0].datatype,
-                    Datatype::Compound { size: 16, members } if members.len() == 2
+                    Datatype::Complex { size: 16, .. }
                 ));
                 assert_eq!(
                     (members[1].name.as_str(), members[1].byte_offset),
@@ -1918,12 +1927,9 @@ mod tests {
         assert_eq!(dt.serialize(), COMPLEX_F64_HDF5_2_0);
         assert_eq!(dt.type_size(), 16);
         dt.check_encodable().unwrap();
-        // Parsing surfaces class 11 as the equivalent `{r, i}` compound.
+        // Parsing returns the native complex type unchanged.
         let (parsed, _) = Datatype::parse(&dt.serialize()).unwrap();
-        assert_eq!(
-            parsed,
-            Datatype::complex_as_compound(16, &crate::type_builders::make_f64_type())
-        );
+        assert_eq!(parsed, dt);
 
         let f32c = make_native_complex_f32_type().serialize();
         assert_eq!(

crates/clawhdf5-format/src/ea_writer.rs

@@ -14,7 +14,7 @@ use crate::chunked_write::{
 
 /// Serialize a v4 Extensible Array layout message.
 pub(crate) fn serialize_v4_extensible_array(
-    chunk_dims: &[u32],
+    chunk_dims: &[u64],
     ea_address: u64,
     offset_size: u8,
     element_size: u32,

crates/clawhdf5-format/src/extensible_array.rs

@@ -454,7 +454,7 @@ pub fn read_extensible_array_chunks(
     header: &ExtensibleArrayHeader,
     dataset_dims: &[u64],
     max_dims: Option<&[u64]>,
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     element_size: u32,
     offset_size: u8,
     length_size: u8,
@@ -480,7 +480,7 @@ pub fn read_extensible_array_chunks_in<S: Storage + ?Sized>(
     header: &ExtensibleArrayHeader,
     dataset_dims: &[u64],
     max_dims: Option<&[u64]>,
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     element_size: u32,
     offset_size: u8,
     _length_size: u8,
@@ -489,12 +489,12 @@ pub fn read_extensible_array_chunks_in<S: Storage + ?Sized>(
 
     // Linear indexes follow the maximum dimensions, with the unlimited
     // dimension swizzled to the slowest position (see `chunk_grid`).
-    let dims_u64: Vec<u64> = chunk_dimensions.iter().map(|&d| d as u64).collect();
-    let grid = ChunkGrid::extensible_array(dataset_dims, max_dims, &dims_u64)?;
+    let grid = ChunkGrid::extensible_array(dataset_dims, max_dims, chunk_dimensions)?;
     let grid = &grid;
 
-    let chunk_byte_size: u64 =
-        chunk_dimensions.iter().map(|&d| d as u64).product::<u64>() * element_size as u64;
+    let chunk_byte_size: u64 = chunk_dimensions
+        .iter()
+        .fold(u64::from(element_size), |acc, &d| acc.saturating_mul(d));
 
     // Parse index block (EAIB): signature(4) + version(1) + client_id(1)
     // + header address(offset_size), then the inline elements, then the
@@ -930,7 +930,7 @@ mod tests {
 
         let header = ExtensibleArrayHeader::parse(&file_data, aehd_offset, os, ls).unwrap();
         let ds_dims = vec![40u64]; // 2 chunks × 20 elements
-        let chunk_dims = vec![20u32];
+        let chunk_dims = vec![20u64];
         let chunks = read_extensible_array_chunks(
             &file_data,
             &header,
@@ -1057,7 +1057,7 @@ mod tests {
         let file_data = build_inline_plus_data_blocks();
         let header = ExtensibleArrayHeader::parse(&file_data, 0x100, os, ls).unwrap();
         let ds_dims = vec![40u64];
-        let chunk_dims = vec![10u32];
+        let chunk_dims = vec![10u64];
         let chunks = read_extensible_array_chunks(
             &file_data,
             &header,

crates/clawhdf5-format/src/file_writer.rs

@@ -3,15 +3,14 @@
 //! Produces valid HDF5 files with v3 superblock, v2 object headers,
 //! link messages, contiguous datasets, inline and dense attributes.
 
-use crate::addr::saturating_usize;
+use crate::addr::{saturating_usize, to_usize};
 #[cfg(not(feature = "std"))]
 use alloc::{format, string::String, vec, vec::Vec};
 
 use crate::attribute::AttributeMessage;
 use crate::btree_v2_write::{BTreeV2Params, build_btree_v2};
 use crate::chunked_write::{
-    ChunkOptions, PrecompressedChunks, build_chunked_data_from_precompressed_libver,
-    precompress_chunks,
+    ChunkOptions, Piece, PreparedChunks, place_chunked_data, prepare_chunks,
 };
 use crate::data_layout::VdsMapping;
 use crate::dataspace::{Dataspace, DataspaceType};
@@ -1612,7 +1611,30 @@ impl FileWriter {
         self
     }
 
+    /// The file, in memory. Its datasets' data is copied into the buffer
+    /// once (a dataset stored as one chunk of its own shape, or unfiltered
+    /// chunks that are contiguous runs of its data, are not copied before);
+    /// to write a file without holding it all, use [`Self::finish_with`].
     pub fn finish(self) -> Result<Vec<u8>, FormatError> {
+        let mut buf = Vec::new();
+        self.finish_into(&mut buf)?;
+        Ok(buf)
+    }
+
+    /// The file, handed to `put` in order, a piece at a time, without
+    /// assembling it in memory: an unfiltered chunk or contiguous dataset
+    /// is passed straight from the data the dataset was given. Writing a
+    /// dataset stored as one unfiltered chunk this way needs little memory
+    /// beyond the dataset's own data. `put`'s first error stops the write
+    /// and is returned.
+    pub fn finish_with<E: From<FormatError>>(
+        self,
+        put: impl FnMut(&[u8]) -> Result<(), E>,
+    ) -> Result<(), E> {
+        self.finish_into(&mut FnSink(put))
+    }
+
+    fn finish_into<S: Sink>(self, sink: &mut S) -> Result<(), S::Error> {
         let page_size = self.page_size;
         if let Some(ps) = page_size
             && !(MIN_FILE_SPACE_PAGE_SIZE..=MAX_FILE_SPACE_PAGE_SIZE).contains(&ps)
@@ -1620,7 +1642,8 @@ impl FileWriter {
             return Err(FormatError::SerializationError(format!(
                 "file space page size {ps} is outside libhdf5's \
                  {MIN_FILE_SPACE_PAGE_SIZE}..={MAX_FILE_SPACE_PAGE_SIZE} bytes"
-            )));
+            ))
+            .into());
         }
 
         let (low, high) = (self.low, self.high);
@@ -1774,15 +1797,16 @@ impl FileWriter {
             })
             .collect::<Result<_, _>>()?;
 
-        struct DataBlob {
+        struct DataBlob<'a> {
             data: Vec<u8>,
             oh_bytes: Vec<u8>,
             /// Cached compressed chunks for chunked datasets; reused in Pass 2
-            /// to avoid re-compressing the same data.
-            precompressed: Option<PrecompressedChunks>,
+            /// to avoid re-compressing the same data. Unfiltered chunks
+            /// borrow the dataset's data where they can.
+            precompressed: Option<PreparedChunks<'a>>,
         }
 
-        let mut dummy_blobs: Vec<DataBlob> = Vec::new();
+        let mut dummy_blobs: Vec<DataBlob<'_>> = Vec::new();
         let mut dummy_cursor = 0u64;
         for (i, d) in all_ds.iter().enumerate() {
             let dense_blob = ds_dense[i]
@@ -1820,21 +1844,16 @@ impl FileWriter {
                     .resolve_chunk_dims_for(&d.ds.dimensions, elem_size);
                 // Compress once in Pass 1; cache the result so Pass 2 can skip
                 // re-compression and just rebuild the index with real addresses.
-                let pre = precompress_chunks(
+                let pre = prepare_chunks(
                     &d.raw,
                     &d.ds.dimensions,
                     &chunk_dims,
                     elem_size,
                     &d.chunk_options,
                 )?;
-                let result = build_chunked_data_from_precompressed_libver(
-                    &pre,
-                    dummy_cursor,
-                    d.maxshape.as_deref(),
-                    low,
-                    high,
-                )?;
-                dummy_cursor += result.data_bytes.len() as u64;
+                let result =
+                    place_chunked_data(&pre, dummy_cursor, d.maxshape.as_deref(), low, high)?;
+                dummy_cursor += result.len;
                 let oh = build_chunked_dataset_oh(
                     &d.dt,
                     &d.ds,
@@ -1849,7 +1868,7 @@ impl FileWriter {
                     d.refcount,
                 )?;
                 dummy_blobs.push(DataBlob {
-                    data: result.data_bytes,
+                    data: Vec::new(),
                     oh_bytes: oh,
                     precompressed: Some(pre),
                 });
@@ -1955,7 +1974,7 @@ impl FileWriter {
             })
             .collect::<Result<_, FormatError>>()?;
 
-        let mut ds_blobs2: Vec<DataBlob> = Vec::new();
+        let mut ds_blobs2: Vec<OutBlob<'_>> = Vec::new();
         let global_align_threshold = self.alignment_threshold;
         let global_align_bytes = self.alignment_bytes;
         for (i, d) in all_ds.iter().enumerate() {
@@ -1977,26 +1996,21 @@ impl FileWriter {
                     &d.fill_message,
                     d.refcount,
                 )?;
-                ds_blobs2.push(DataBlob {
-                    data: gcol_bytes.clone(),
+                ds_blobs2.push(OutBlob {
+                    data: vec![Out::Slice(gcol_bytes)],
                     oh_bytes: oh,
-                    precompressed: None,
                 });
             } else if is_chunked[i] {
                 let base_address = cursor2 as u64;
                 // Reuse precompressed chunks from Pass 1 — avoids re-compressing
                 // the same data a second time.
-                let result = build_chunked_data_from_precompressed_libver(
-                    dummy_blobs[i]
+                let pre = dummy_blobs[i]
                     .precompressed
                     .as_ref()
-                        .expect("chunked dataset missing precompressed cache"),
-                    base_address,
-                    d.maxshape.as_deref(),
-                    low,
-                    high,
-                )?;
-                cursor2 += result.data_bytes.len();
+                    .expect("chunked dataset missing precompressed cache");
+                let result =
+                    place_chunked_data(pre, base_address, d.maxshape.as_deref(), low, high)?;
+                cursor2 += to_usize(result.len)?;
                 let oh = build_chunked_dataset_oh(
                     &d.dt,
                     &d.ds,
@@ -2010,11 +2024,16 @@ impl FileWriter {
                     &d.fill_message,
                     d.refcount,
                 )?;
-                ds_blobs2.push(DataBlob {
-                    data: result.data_bytes,
-                    oh_bytes: oh,
-                    precompressed: None,
-                });
+                let data = result
+                    .pieces
+                    .into_iter()
+                    .map(|p| match p {
+                        Piece::Zeros(n) => Out::Zeros(n),
+                        Piece::Chunk(c) => Out::Slice(&pre.chunks[c].1),
+                        Piece::Bytes(b) => Out::Bytes(b),
+                    })
+                    .collect();
+                ds_blobs2.push(OutBlob { data, oh_bytes: oh });
             } else if is_compact[i] {
                 // Compact: data is inline in the object header, no external blob
                 let oh = build_compact_dataset_oh(
@@ -2030,10 +2049,9 @@ impl FileWriter {
                     d.refcount,
                     layout_version,
                 )?;
-                ds_blobs2.push(DataBlob {
+                ds_blobs2.push(OutBlob {
                     data: vec![],
                     oh_bytes: oh,
-                    precompressed: None,
                 });
             } else {
                 // Determine alignment: per-dataset overrides global
@@ -2060,13 +2078,10 @@ impl FileWriter {
                     d.refcount,
                     layout_version,
                 )?;
-                let mut data = vec![0u8; padding];
-                data.extend_from_slice(&d.raw);
                 cursor2 += d.raw.len();
-                ds_blobs2.push(DataBlob {
-                    data,
+                ds_blobs2.push(OutBlob {
+                    data: vec![Out::Zeros(padding), Out::Slice(&d.raw)],
                     oh_bytes: oh,
-                    precompressed: None,
                 });
             }
         }
@@ -2080,7 +2095,8 @@ impl FileWriter {
             cursor2 = cursor2.next_multiple_of(ps as usize);
         }
         let eof_addr2 = cursor2 as u64;
-        let mut buf = Vec::with_capacity(cursor2);
+        sink.reserve(cursor2);
+        let mut out = Counted { sink, written: 0 };
 
         let sb = Superblock {
             version: superblock_version,
@@ -2103,9 +2119,9 @@ impl FileWriter {
             checksum: None,
             page_size: None,
         };
-        buf.extend_from_slice(&sb.serialize());
+        out.put(&sb.serialize())?;
         if let Some(ref ext) = sb_ext {
-            buf.extend_from_slice(ext);
+            out.put(ext)?;
         }
 
         // Group OHs + dense blobs (link blob, then attr blob, matching pass 2)
@@ -2132,32 +2148,120 @@ impl FileWriter {
                 g.refcount,
             )?;
             debug_assert_eq!(oh.len(), group_oh_sizes[gi]);
-            debug_assert_eq!(buf.len() as u64, group_addrs2[gi]);
-            buf.extend_from_slice(&oh);
+            debug_assert_eq!(out.written as u64, group_addrs2[gi]);
+            out.put(&oh)?;
             if let Some(ref b) = link_blob {
-                buf.extend_from_slice(&b.blob);
+                out.put(&b.blob)?;
             }
             if let Some(ref blob) = group_dense_blobs[gi] {
-                buf.extend_from_slice(&blob.blob);
+                out.put(&blob.blob)?;
             }
         }
 
         // Dataset OHs + dense blobs
         for (i, blob) in ds_blobs2.iter().enumerate() {
-            buf.extend_from_slice(&blob.oh_bytes);
+            out.put(&blob.oh_bytes)?;
             if let Some(ref dense) = ds_dense_blobs[i] {
-                buf.extend_from_slice(&dense.blob);
+                out.put(&dense.blob)?;
             }
         }
 
         // Data
         for blob in &ds_blobs2 {
-            buf.extend_from_slice(&blob.data);
+            for piece in &blob.data {
+                match piece {
+                    Out::Bytes(b) => out.put(b)?,
+                    Out::Slice(s) => out.put(s)?,
+                    Out::Zeros(n) => out.zeros(*n)?,
+                }
+            }
         }
 
-        debug_assert_eq!(buf.len(), data_end);
-        buf.resize(cursor2, 0);
-        Ok(buf)
+        debug_assert_eq!(out.written, data_end);
+        out.zeros(cursor2 - data_end)?;
+        Ok(())
+    }
+}
+
+/// A piece of a dataset's data as [`FileWriter`] writes it out.
+enum Out<'a> {
+    Bytes(Vec<u8>),
+    /// Borrowed from the dataset's data or its prepared chunks.
+    Slice(&'a [u8]),
+    Zeros(usize),
+}
+
+/// A dataset's object header and the pieces of its data.
+struct OutBlob<'a> {
+    oh_bytes: Vec<u8>,
+    data: Vec<Out<'a>>,
+}
+
+/// Where [`FileWriter`] puts a file's bytes, in order.
+trait Sink {
+    type Error: From<FormatError>;
+
+    /// The file will be `_len` bytes long.
+    fn reserve(&mut self, _len: usize) {}
+
+    fn put(&mut self, bytes: &[u8]) -> Result<(), Self::Error>;
+
+    fn zeros(&mut self, n: usize) -> Result<(), Self::Error> {
+        const ZEROS: [u8; 4096] = [0; 4096];
+        let mut left = n;
+        while left > 0 {
+            let k = left.min(ZEROS.len());
+            self.put(&ZEROS[..k])?;
+            left -= k;
+        }
+        Ok(())
+    }
+}
+
+impl Sink for Vec<u8> {
+    type Error = FormatError;
+
+    fn reserve(&mut self, len: usize) {
+        self.reserve_exact(len.saturating_sub(self.len()));
+    }
+
+    fn put(&mut self, bytes: &[u8]) -> Result<(), FormatError> {
+        self.extend_from_slice(bytes);
+        Ok(())
+    }
+
+    fn zeros(&mut self, n: usize) -> Result<(), FormatError> {
+        self.resize(self.len() + n, 0);
+        Ok(())
+    }
+}
+
+/// A [`Sink`] calling a function ([`FileWriter::finish_with`]).
+struct FnSink<F>(F);
+
+impl<E: From<FormatError>, F: FnMut(&[u8]) -> Result<(), E>> Sink for FnSink<F> {
+    type Error = E;
+
+    fn put(&mut self, bytes: &[u8]) -> Result<(), E> {
+        (self.0)(bytes)
+    }
+}
+
+/// A [`Sink`] and the number of bytes put into it.
+struct Counted<'s, S> {
+    sink: &'s mut S,
+    written: usize,
+}
+
+impl<S: Sink> Counted<'_, S> {
+    fn put(&mut self, bytes: &[u8]) -> Result<(), S::Error> {
+        self.written += bytes.len();
+        self.sink.put(bytes)
+    }
+
+    fn zeros(&mut self, n: usize) -> Result<(), S::Error> {
+        self.written += n;
+        self.sink.zeros(n)
     }
 }
 
@@ -3015,4 +3119,75 @@ mod tests {
         assert_eq!(Superblock::parse(&bytes, 0).unwrap().version, 3);
         assert_eq!(layout_of(&bytes, "x")[0], 3);
     }
+
+    /// `finish_with` hands out exactly the bytes `finish` returns, for every
+    /// storage (contiguous, compact, chunked filtered and not, chunks
+    /// borrowed and copied, a version-1 B-tree, a paged file), and stops at
+    /// the sink's first error.
+    #[test]
+    fn finish_with_matches_finish() {
+        let build = |low: LibVer, paged: bool| {
+            let mut fw = FileWriter::new();
+            fw.libver_bounds(low, LibVer::Latest);
+            if paged {
+                fw.with_page_size(4096);
+            }
+            let v: Vec<f64> = (0..300).map(f64::from).collect();
+            fw.create_dataset("contig").with_f64_data(&v);
+            fw.create_dataset("compact")
+                .with_f64_data(&[1.0, 2.0])
+                .compact();
+            fw.create_dataset("one_chunk")
+                .with_f64_data(&v)
+                .with_shape(&[30, 10])
+                .with_chunks(&[30, 10]);
+            fw.create_dataset("rows")
+                .with_f64_data(&v)
+                .with_shape(&[30, 10])
+                .with_chunks(&[7, 10]);
+            fw.create_dataset("tiles")
+                .with_f64_data(&v)
+                .with_shape(&[30, 10])
+                .with_chunks(&[7, 3]);
+            fw.create_dataset("deflated")
+                .with_f64_data(&v)
+                .with_chunks(&[64])
+                .with_shuffle()
+                .with_deflate(4);
+            fw.create_dataset("grows")
+                .with_u8_data_owned((0..=255).collect())
+                .with_chunks(&[100])
+                .with_maxshape(&[u64::MAX]);
+            fw
+        };
+        for (low, paged) in [
+            (LibVer::V18, false),
+            (LibVer::V110, false),
+            (LibVer::Latest, true),
+        ] {
+            let bytes = build(low, paged).finish().unwrap();
+            let mut streamed = Vec::new();
+            build(low, paged)
+                .finish_with(|b: &[u8]| {
+                    streamed.extend_from_slice(b);
+                    Ok::<(), FormatError>(())
+                })
+                .unwrap();
+            assert_eq!(streamed, bytes, "{low:?} paged {paged}");
+        }
+
+        let mut calls = 0;
+        let err = build(LibVer::V110, false)
+            .finish_with(|_| {
+                calls += 1;
+                if calls == 3 {
+                    Err(FormatError::SerializationError("full".into()))
+                } else {
+                    Ok(())
+                }
+            })
+            .unwrap_err();
+        assert_eq!(err, FormatError::SerializationError("full".into()));
+        assert_eq!(calls, 3);
+    }
 }

crates/clawhdf5-format/src/fixed_array.rs

@@ -155,7 +155,7 @@ pub fn read_fixed_array_chunks(
     header: &FixedArrayHeader,
     dataset_dims: &[u64],
     max_dims: Option<&[u64]>,
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     element_size: u32,
     offset_size: u8,
     length_size: u8,
@@ -180,7 +180,7 @@ pub fn read_fixed_array_chunks_in<S: Storage + ?Sized>(
     header: &FixedArrayHeader,
     dataset_dims: &[u64],
     max_dims: Option<&[u64]>,
-    chunk_dimensions: &[u32],
+    chunk_dimensions: &[u64],
     element_size: u32,
     offset_size: u8,
     _length_size: u8,
@@ -227,11 +227,11 @@ pub fn read_fixed_array_chunks_in<S: Storage + ?Sized>(
 
     // The index is laid out over the chunk grid of the *maximum* dimensions
     // (row-major), so a dataset smaller than its maxshape has gaps.
-    let dims_u64: Vec<u64> = chunk_dimensions.iter().map(|&d| d as u64).collect();
-    let grid = ChunkGrid::fixed_array(dataset_dims, max_dims, &dims_u64)?;
+    let grid = ChunkGrid::fixed_array(dataset_dims, max_dims, chunk_dimensions)?;
 
-    let chunk_byte_size: u64 =
-        chunk_dimensions.iter().map(|&d| d as u64).product::<u64>() * element_size as u64;
+    let chunk_byte_size: u64 = chunk_dimensions
+        .iter()
+        .fold(u64::from(element_size), |acc, &d| acc.saturating_mul(d));
 
     let mut chunks = Vec::new();
     // `rel` is relative to the data block, whose bytes are in `w`.
@@ -670,7 +670,7 @@ mod tests {
         let header =
             FixedArrayHeader::parse(&file_data, fahd_offset, offset_size, length_size).unwrap();
         let ds_dims = vec![100u64];
-        let chunk_dims = vec![20u32];
+        let chunk_dims = vec![20u64];
         let chunks = read_fixed_array_chunks(
             &file_data,
             &header,
@@ -747,7 +747,7 @@ mod tests {
         let header =
             FixedArrayHeader::parse(&file_data, fahd_offset, offset_size, length_size).unwrap();
         let ds_dims = vec![60u64];
-        let chunk_dims = vec![20u32];
+        let chunk_dims = vec![20u64];
         let chunks = read_fixed_array_chunks(
             &file_data,
             &header,
@@ -848,7 +848,7 @@ mod tests {
         assert_eq!(header.num_elements, 11);
 
         let ds_dims = vec![11u64 * 20];
-        let chunk_dims = vec![20u32];
+        let chunk_dims = vec![20u64];
         let chunks = read_fixed_array_chunks(
             &file_data,
             &header,

crates/clawhdf5-format/src/type_builders.rs

@@ -663,11 +663,17 @@ impl DatasetBuilder {
     }
 
     pub fn with_u8_data(&mut self, data: &[u8]) -> &mut Self {
+        self.with_u8_data_owned(data.to_vec())
+    }
+
+    /// [`Self::with_u8_data`] without the copy: the builder takes the
+    /// vector. For large data this halves the memory a write needs.
+    pub fn with_u8_data_owned(&mut self, data: Vec<u8>) -> &mut Self {
         self.datatype = Some(make_u8_type());
-        self.data = Some(data.to_vec());
         if self.shape.is_none() {
             self.shape = Some(vec![data.len() as u64]);
         }
+        self.data = Some(data);
         self
     }
 

crates/clawhdf5-py/README.md

@@ -104,7 +104,22 @@ writes `float64`, `float32`, `int64`, `int32`, `uint8`, `complex64` and
 `complex128` arrays; the file is written on `close()`. Complex arrays are
 stored as h5py stores them, a compound `{r, i}` that every libhdf5 reads
 (not HDF5 2.0's native complex type, which only libhdf5 2.0+ reads; the
-Rust API writes that on request).
+Rust API writes that on request). Both forms read back as numpy
+`complex64`/`complex128`.
+
+`libver=` sets the library version bounds as in h5py: `'v108'`,
+`'v110'`, `'v112'`, `'v114'`, `'v200'` or `'latest'` (the low bound; the
+high bound is then `'latest'`), or a `(low, high)` tuple. With `'v108'`
+the file is written in the HDF5 1.8 format (version-2 superblock,
+version-1 B-tree chunk indexes), which HDF5 1.8 reads; the default is the
+HDF5 1.10 format. `'earliest'` as the low bound writes the 1.8 format too,
+with a `UserWarning`: clawhdf5 cannot write the pre-1.8 format. The
+argument is ignored when reading and refused for `'r+'`/`'a'`.
+
+```python
+with clawhdf5.File("old.h5", "w", libver="v108") as f:   # HDF5 1.8 reads it
+    f.create_dataset("x", data=np.arange(10.0), chunks=(5,), compression="gzip")
+```
 
 ## Editing a file in place
 

crates/clawhdf5-py/src/file.rs

@@ -13,10 +13,13 @@ use crate::attrs::PyAttrs;
 use crate::group::{PyGroup, ReadGroup, WriteGroupState, finalize_write_group};
 use crate::handle::Handle;
 use crate::{DatasetSpec, OwnedAttrValue, apply_dataset_spec, extract_numpy_data, to_py_err};
+use clawhdf5_rs::LibVer;
 
 /// Internal state for write mode.
 struct WriteState {
     path: PathBuf,
+    /// `libver=` as (low, high); `None` keeps the writer's default.
+    libver: Option<(LibVer, LibVer)>,
     root_datasets: Vec<DatasetSpec>,
     root_attrs: Arc<Mutex<Vec<(String, OwnedAttrValue)>>>,
     groups: Vec<Arc<Mutex<WriteGroupState>>>,
@@ -80,10 +83,32 @@ impl PyFile {
     ///     `az://`; which schemes work depends on how the wheel was built)
     ///     to read the file remotely with default options (see `open_url`)
     ///   mode: 'r' for read (default), 'w' for write
+    ///   libver: library version bounds for mode 'w', as h5py's: one of
+    ///     'earliest', 'v108', 'v110', 'v112', 'v114', 'v200', 'latest' (the
+    ///     low bound; the high bound is then 'latest') or a (low, high)
+    ///     tuple of them. The low bound is the oldest HDF5 release whose
+    ///     format the file uses ('v108': HDF5 1.8 can read it); the high
+    ///     bound the newest whose features it may use. clawhdf5 cannot write
+    ///     the pre-1.8 format, so a low bound of 'earliest' writes the 1.8
+    ///     format (with a warning) and a high bound of 'earliest' is an
+    ///     error. Default (None): the HDF5 1.10 format clawhdf5 has always
+    ///     written. Ignored for reading; not supported with 'r+' / 'a'.
     #[new]
-    #[pyo3(signature = (path, mode="r"))]
-    fn new(py: Python<'_>, path: &str, mode: &str) -> PyResult<Self> {
+    #[pyo3(signature = (path, mode="r", libver=None))]
+    fn new(
+        py: Python<'_>,
+        path: &str,
+        mode: &str,
+        libver: Option<&Bound<'_, PyAny>>,
+    ) -> PyResult<Self> {
         let filename = path.to_string();
+        let libver = libver.map(|v| parse_libver(py, v)).transpose()?;
+        if libver.is_some() && matches!(mode, "r+" | "a") {
+            return Err(PyNotImplementedError::new_err(format!(
+                "libver with mode '{mode}': clawhdf5's in-place editor keeps the format \
+                 versions the file already uses"
+            )));
+        }
         if is_url(path) {
             if mode != "r" {
                 return Err(PyValueError::new_err(format!(
@@ -113,6 +138,7 @@ impl PyFile {
                     // Absolute now: the file is written at close, possibly
                     // after the working directory changed.
                     path: std::path::absolute(path).unwrap_or_else(|_| PathBuf::from(path)),
+                    libver,
                     root_datasets: Vec::new(),
                     root_attrs: Arc::new(Mutex::new(Vec::new())),
                     groups: Vec::new(),
@@ -460,10 +486,71 @@ fn parse_compression(
     }
 }
 
+/// One of h5py's `libver` names as a bound; `high` says which end it is.
+/// `Ok(None)` is 'earliest' as the low bound: the pre-1.8 format, which
+/// clawhdf5 cannot write.
+fn libver_name(name: &str, high: bool) -> PyResult<Option<LibVer>> {
+    Ok(Some(match name {
+        "earliest" if high => {
+            return Err(PyValueError::new_err(
+                "libver high bound 'earliest' (the pre-1.8 format) cannot be written by \
+                 clawhdf5; the oldest format it writes is 'v108'",
+            ));
+        }
+        "earliest" => return Ok(None),
+        "v108" => LibVer::V18,
+        "v110" => LibVer::V110,
+        "v112" => LibVer::V112,
+        "v114" => LibVer::V114,
+        "v200" => LibVer::V200,
+        "latest" => LibVer::Latest,
+        other => {
+            return Err(PyValueError::new_err(format!(
+                "unknown libver '{other}'; expected 'earliest', 'v108', 'v110', 'v112', \
+                 'v114', 'v200' or 'latest'"
+            )));
+        }
+    }))
+}
+
+/// h5py's `libver=`: a name (the low bound, high bound 'latest') or a
+/// `(low, high)` pair.
+fn parse_libver(py: Python<'_>, v: &Bound<'_, PyAny>) -> PyResult<(LibVer, LibVer)> {
+    let (low, high): (String, String) = match v.extract::<String>() {
+        Ok(name) => (name, "latest".into()),
+        Err(_) => v.extract().map_err(|_| {
+            PyValueError::new_err("libver must be a string or a (low, high) tuple of strings")
+        })?,
+    };
+    let Some(high) = libver_name(&high, true)? else {
+        unreachable!("a high bound is never None")
+    };
+    let low = match libver_name(&low, false)? {
+        Some(low) => low,
+        None => {
+            PyModule::import(py, "warnings")?.getattr("warn")?.call1((
+                "libver 'earliest': clawhdf5 cannot write the pre-1.8 format; the file \
+                 is written in the HDF5 1.8 format ('v108') instead",
+                py.get_type::<pyo3::exceptions::PyUserWarning>(),
+            ))?;
+            LibVer::V18
+        }
+    };
+    if low > high {
+        return Err(PyValueError::new_err(format!(
+            "libver low bound {low} is newer than the high bound {high}"
+        )));
+    }
+    Ok((low, high))
+}
+
 /// Build and write the HDF5 file from accumulated write state.
 fn finalize_write(state: WriteState) -> PyResult<()> {
     crate::no_panic(|| {
         let mut builder = clawhdf5_rs::FileBuilder::new();
+        if let Some((low, high)) = state.libver {
+            builder.libver_bounds(low, high);
+        }
 
         // Root attributes
         let root_attrs = state.root_attrs.lock().unwrap_or_else(|e| e.into_inner());
@@ -520,6 +607,7 @@ mod tests {
 
         let state = WriteState {
             path: path.clone(),
+            libver: None,
             root_datasets: vec![DatasetSpec {
                 name: "data".into(),
                 data: crate::DatasetData::F64(vec![1.0, 2.0, 3.0]),

crates/clawhdf5-py/src/node.rs

@@ -189,12 +189,7 @@ pub(crate) fn chunk_shape(file: &File, hdr: &ObjectHeader, rank: usize) -> Optio
     {
         clawhdf5_format::data_layout::DataLayout::Chunked {
             chunk_dimensions, ..
-        } if chunk_dimensions.len() >= rank => Some(
-            chunk_dimensions[..rank]
-                .iter()
-                .map(|&d| u64::from(d))
-                .collect(),
-        ),
+        } if chunk_dimensions.len() >= rank => Some(chunk_dimensions[..rank].to_vec()),
         _ => None,
     }
 }

crates/clawhdf5-py/tests/test_libver.py

@@ -0,0 +1,143 @@
+"""`clawhdf5.File(path, 'w', libver=...)`: h5py's library version bounds.
+
+A file written with libver='v108' must open in HDF5 1.8. Its h5dump is
+found through CLAWHDF5_H5DUMP18 or at ~/.cache/hdf5-1.8.23/bin/h5dump
+(scripts/build-hdf5-1.8.sh builds it); without it that check is skipped.
+"""
+
+import os
+import subprocess
+
+import numpy as np
+import pytest
+
+import clawhdf5
+
+
+def superblock_version(path):
+    with open(path, "rb") as f:
+        head = f.read(9)
+    assert head[:8] == b"\x89HDF\r\n\x1a\n"
+    return head[8]
+
+
+def h5dump18():
+    path = os.environ.get("CLAWHDF5_H5DUMP18") or os.path.expanduser(
+        "~/.cache/hdf5-1.8.23/bin/h5dump"
+    )
+    try:
+        out = subprocess.run([path, "--version"], capture_output=True, text=True)
+    except OSError:
+        return None
+    return path if "1.8." in out.stdout else None
+
+
+def write_sample(path, libver):
+    with clawhdf5.File(path, "w", libver=libver) as f:
+        f.create_dataset("x", data=np.arange(10, dtype=np.float64))
+        f.create_dataset(
+            "chunked",
+            data=np.arange(100, dtype=np.int32),
+            chunks=(30,),
+            compression="gzip",
+        )
+        f.create_dataset("z", data=np.array([1 + 2j, -3j], dtype=np.complex128))
+        g = f.create_group("g")
+        g.create_dataset("y", data=np.ones(3, dtype=np.float32))
+        f.attrs["version"] = 1
+
+
+@pytest.mark.parametrize(
+    "libver, sb",
+    [
+        (None, 3),
+        ("v108", 2),
+        (("v108", "v108"), 2),
+        (("v108", "latest"), 2),
+        ("v110", 3),
+        ("v112", 3),
+        ("v114", 3),
+        ("v200", 3),
+        ("latest", 3),
+        (("v110", "v200"), 3),
+    ],
+)
+def test_libver_bounds(h5py, tmp_path, libver, sb):
+    path = str(tmp_path / "libver.h5")
+    write_sample(path, libver)
+    # v108 low bound: HDF5 1.8's version-2 superblock; 1.10 and later: 3.
+    assert superblock_version(path) == sb
+    with h5py.File(path, "r") as f:
+        np.testing.assert_array_equal(f["x"][:], np.arange(10.0))
+        np.testing.assert_array_equal(f["chunked"][:], np.arange(100))
+        np.testing.assert_array_equal(f["z"][:], [1 + 2j, -3j])
+        np.testing.assert_array_equal(f["g/y"][:], np.ones(3))
+        assert f.attrs["version"] == 1
+    with clawhdf5.File(path, "r") as f:
+        np.testing.assert_array_equal(f["chunked"][:], np.arange(100))
+
+
+def test_libver_v108_opens_in_hdf5_1_8(tmp_path):
+    h5dump = h5dump18()
+    if h5dump is None:
+        pytest.skip("no HDF5 1.8 h5dump (CLAWHDF5_H5DUMP18)")
+    path = str(tmp_path / "v108.h5")
+    write_sample(path, "v108")
+    out = subprocess.run([h5dump, path], capture_output=True, text=True)
+    assert out.returncode == 0, out.stderr
+    assert "h5dump error" not in out.stderr
+    assert 'DATASET "y"' in out.stdout
+    # The chunked, deflated dataset (a version-1 B-tree) reads in full.
+    out = subprocess.run(
+        [h5dump, "-w", "0", "-d", "/chunked", path], capture_output=True, text=True
+    )
+    assert out.returncode == 0, out.stderr
+    assert "(0): " + ", ".join(str(k) for k in range(100)) + "\n" in out.stdout
+    # The default (1.10 format) does not open in 1.8: the bound matters.
+    path = str(tmp_path / "default.h5")
+    write_sample(path, None)
+    out = subprocess.run([h5dump, path], capture_output=True, text=True)
+    assert out.returncode != 0
+
+
+def test_libver_earliest_writes_v108_with_a_warning(h5py, tmp_path):
+    path = str(tmp_path / "earliest.h5")
+    with pytest.warns(UserWarning, match="pre-1.8"):
+        write_sample(path, "earliest")
+    assert superblock_version(path) == 2
+    with h5py.File(path, "r") as f:
+        np.testing.assert_array_equal(f["x"][:], np.arange(10.0))
+
+
+@pytest.mark.parametrize(
+    "libver, match",
+    [
+        (("v108", "earliest"), "high bound 'earliest'"),
+        ("v109", "unknown libver 'v109'"),
+        (("latest", "v108"), "newer than the high bound"),
+        (("v108",), "tuple"),
+        (3, "tuple"),
+    ],
+)
+def test_libver_errors(tmp_path, libver, match):
+    with pytest.raises(ValueError, match=match):
+        clawhdf5.File(str(tmp_path / "bad.h5"), "w", libver=libver)
+
+
+def test_libver_v108_high_bound_writes_everything(tmp_path):
+    # Native complex numbers need HDF5 2.0; the Python writer stores complex
+    # as h5py's {r, i} compound, which 1.8 reads, so the 1.8 high bound is
+    # fine for everything clawhdf5.File writes.
+    path = str(tmp_path / "v18only.h5")
+    write_sample(path, ("v108", "v108"))
+    assert superblock_version(path) == 2
+
+
+def test_libver_not_for_editing(tmp_path):
+    path = str(tmp_path / "e.h5")
+    write_sample(path, None)
+    with pytest.raises(NotImplementedError, match="libver"):
+        clawhdf5.File(path, "r+", libver="latest")
+    # Reading ignores it, as the bounds only affect what is written.
+    with clawhdf5.File(path, "r", libver="v108") as f:
+        assert f["x"].shape == (10,)

crates/clawhdf5-tools/src/check.rs

@@ -873,9 +873,9 @@ impl Checker<'_> {
             self.counts.chunk_index_checksummed += 1;
         }
         let filtered = matches!(&info.filters, Ok(Some(p)) if !p.filters.is_empty());
-        let chunk_bytes = cdims.iter().try_fold(u64::from(dt.type_size()), |a, &d| {
-            a.checked_mul(u64::from(d))
-        });
+        let chunk_bytes = cdims
+            .iter()
+            .try_fold(u64::from(dt.type_size()), |a, &d| a.checked_mul(d));
         let max = ds.max_dimensions.clone();
         let mut seen: HashSet<Vec<u64>> = HashSet::with_capacity(chunks.len().min(1 << 20));
         let mut reported = 0usize;
@@ -893,7 +893,7 @@ impl Checker<'_> {
                 ));
             } else {
                 for (i, (&o, &cd)) in c.offsets.iter().zip(cdims).enumerate() {
-                    if o % u64::from(cd) != 0 {
+                    if o % cd != 0 {
                         bad.push(format!(
                             "offset {o} in dimension {i} is not a multiple of the chunk size {cd}"
                         ));

crates/clawhdf5-tools/src/diff.rs

@@ -603,7 +603,22 @@ impl Diff {
                 let dif = match (int_of(&va), int_of(&vb), number(&va), number(&vb)) {
                     (Some(x), Some(y), ..) => x.abs_diff(y).to_string(),
                     (_, _, Some(x), Some(y)) => value::fmt_float((x - y).abs(), 64),
+                    // Each part's absolute difference, as h5diff 2.x prints
+                    // it (`1+0i`).
+                    _ => match (&va, &vb) {
+                        (
+                            Value::Complex { re: ar, im: ai, .. },
+                            Value::Complex { re: br, im: bi, .. },
+                        ) => match (number(ar), number(ai), number(br), number(bi)) {
+                            (Some(ar), Some(ai), Some(br), Some(bi)) => format!(
+                                "{}+{}i",
+                                value::fmt_float((ar - br).abs(), 64),
+                                value::fmt_float((ai - bi).abs(), 64)
+                            ),
                             _ => String::new(),
+                        },
+                        _ => String::new(),
+                    },
                 };
                 rows.push(format!("{pos:<24}{ta:<24}{tb:<24}{dif}"));
             }
@@ -713,6 +728,9 @@ impl Diff {
             (Value::Array(p), Value::Array(q)) | (Value::Seq(p), Value::Seq(q)) => {
                 p.len() == q.len() && p.iter().zip(q).all(|(u, v)| self.equal(a, b, u, v))
             }
+            (Value::Complex { re: pr, im: pi, .. }, Value::Complex { re: qr, im: qi, .. }) => {
+                self.equal(a, b, pr, qr) && self.equal(a, b, pi, qi)
+            }
             (Value::Ref(None), Value::Ref(None)) => true,
             (Value::Ref(Some(p)), Value::Ref(Some(q))) => {
                 // Addresses mean nothing across files: compare the paths the
@@ -846,7 +864,10 @@ fn types_comparable(a: &Datatype, b: &Datatype) -> bool {
         (
             Datatype::Enumeration { base_type: x, .. },
             Datatype::Enumeration { base_type: y, .. },
-        ) => types_comparable(x, y),
+        )
+        | (Datatype::Complex { base_type: x, .. }, Datatype::Complex { base_type: y, .. }) => {
+            types_comparable(x, y)
+        }
         _ => true,
     }
 }

crates/clawhdf5-tools/src/dtype.rs

@@ -135,9 +135,16 @@ pub fn short(dt: &Datatype) -> String {
         return f.short.into();
     }
     match dt {
-        Datatype::Complex { size, base_type } => {
-            short(&Datatype::complex_as_compound(*size, base_type))
-        }
+        // numpy's names (`complex64` is two `float32`) for IEEE parts,
+        // `complex<part>` otherwise.
+        Datatype::Complex { size, base_type } => match base_type.as_ref() {
+            Datatype::FloatingPoint { byte_order, .. } if is_ieee(base_type) => format!(
+                "complex{}{}",
+                u64::from(*size) * 8,
+                if be(byte_order) { "-be" } else { "" }
+            ),
+            _ => format!("complex<{}>", short(base_type)),
+        },
         Datatype::FixedPoint {
             size,
             signed,
@@ -216,8 +223,9 @@ pub fn long(dt: &Datatype) -> String {
         return f.long.into();
     }
     match dt {
-        Datatype::Complex { size, base_type } => {
-            long(&Datatype::complex_as_compound(*size, base_type))
+        // As h5ls 2.x prints a complex type it has no native name for.
+        Datatype::Complex { base_type, .. } => {
+            format!("complex number of\n                   {}", long(base_type))
         }
         Datatype::FixedPoint {
             size,
@@ -361,6 +369,19 @@ fn atomic_ddl(dt: &Datatype) -> Option<String> {
         )),
         // As h5dump 2.x names them (checked against h5dump 2.2.0).
         Datatype::FloatingPoint { .. } => small_float(dt).map(|f| f.ddl.to_string()),
+        // h5dump 2.x's predefined complex names: IEEE binary16/32/64 parts.
+        Datatype::Complex { base_type, .. } => match base_type.as_ref() {
+            Datatype::FloatingPoint {
+                size, byte_order, ..
+            } if is_ieee(base_type) && !matches!(byte_order, DatatypeByteOrder::Vax) => {
+                Some(format!(
+                    "H5T_COMPLEX_IEEE_F{}{}",
+                    u64::from(*size) * 8,
+                    order_suffix(byte_order)
+                ))
+            }
+            _ => None,
+        },
         Datatype::BitField {
             size, byte_order, ..
         } => Some(format!(
@@ -418,6 +439,10 @@ pub fn ddl(dt: &Datatype, ind: usize) -> String {
         Datatype::VariableLength { base_type, .. } => {
             format!("H5T_VLEN {{ {} }}", ddl(base_type, ind))
         }
+        // Parts h5dump has no predefined complex name for.
+        Datatype::Complex { base_type, .. } => {
+            format!("H5T_COMPLEX {{ {} }}", ddl(base_type, ind))
+        }
         Datatype::Opaque { size, tag } => {
             let tag = String::from_utf8_lossy(tag);
             format!(
@@ -498,6 +523,9 @@ fn string_ddl(
 /// hdf5-json type object.
 pub fn json(dt: &Datatype) -> J {
     match dt {
+        // hdf5-json (h5json 2.0.0) has no complex class: a native complex
+        // is written as the `{r, i}` compound h5py uses for complex numbers,
+        // its values as `[re, im]` pairs.
         Datatype::Complex { size, base_type } => {
             json(&Datatype::complex_as_compound(*size, base_type))
         }
@@ -598,7 +626,7 @@ fn pad_json(p: &StringPadding) -> &'static str {
 /// Class name used to decide whether two datatypes can be compared.
 pub fn class(dt: &Datatype) -> &'static str {
     match dt {
-        Datatype::Complex { .. } => "compound",
+        Datatype::Complex { .. } => "complex",
         Datatype::FixedPoint { .. } => "integer",
         Datatype::FloatingPoint { .. } => "float",
         Datatype::Time { .. } => "time",

crates/clawhdf5-tools/src/ls.rs

@@ -292,7 +292,7 @@ impl Ls<'_> {
                     .unwrap_or(0);
                 let bytes = dims
                     .iter()
-                    .try_fold(esize, |a, &d| a.checked_mul(u64::from(d)))
+                    .try_fold(esize, |a, &d| a.checked_mul(d))
                     .map(|b| b.to_string())
                     .unwrap_or_else(|| "?".into());
                 writeln!(

crates/clawhdf5-tools/src/value.rs

@@ -27,6 +27,15 @@ pub enum Value {
     /// An enum member (name, when the value matches one) and its value.
     Enum(Option<String>, i128),
     Compound(Vec<(String, Value)>),
+    /// A native complex number (HDF5 2.0 class 11): real and imaginary
+    /// part. `sign_free` is set when h5dump joins the parts with a bare `+`
+    /// (parts it has no native C type for: binary16, non-IEEE), giving
+    /// `1+-2i`; otherwise the imaginary part carries its sign (`1-2i`).
+    Complex {
+        re: Box<Value>,
+        im: Box<Value>,
+        sign_free: bool,
+    },
     Array(Vec<Value>),
     /// A variable-length sequence.
     Seq(Vec<Value>),
@@ -183,11 +192,18 @@ impl<'a> Decoder<'a> {
             return Value::Error("short element".into());
         };
         match dt {
-            Datatype::Complex { size, base_type } => self.decode(
-                &Datatype::complex_as_compound(*size, base_type),
-                b,
-                depth + 1,
-            ),
+            Datatype::Complex { base_type, .. } => {
+                let part = base_type.type_size() as usize;
+                let (re, im) = b.split_at(part.min(b.len()));
+                Value::Complex {
+                    re: Box::new(self.decode(base_type, re, depth + 1)),
+                    im: Box::new(self.decode(base_type, im, depth + 1)),
+                    // h5dump prints `float`/`double` complex (whatever the
+                    // byte order) with `%g%+gi`, anything else as
+                    // `<re>+<im>i`.
+                    sign_free: !(dtype::is_ieee(base_type) && matches!(part, 4 | 8)),
+                }
+            }
             Datatype::FixedPoint { .. } => match decode_int(dt, b) {
                 Some(v) => Value::Int(v),
                 None => Value::Bytes(b.to_vec()),
@@ -354,6 +370,15 @@ pub fn text(v: &Value, h5paths: &dyn Fn(u64) -> Option<String>) -> String {
                 .collect::<Vec<_>>()
                 .join(", ")
         ),
+        Value::Complex { re, im, sign_free } => {
+            let (re, im) = (text(re, h5paths), text(im, h5paths));
+            let plus = if *sign_free || !im.starts_with('-') {
+                "+"
+            } else {
+                ""
+            };
+            format!("{re}{plus}{im}i")
+        }
         Value::Array(vs) => format!(
             "[ {} ]",
             vs.iter()
@@ -408,6 +433,7 @@ pub fn to_json(v: &Value, h5paths: &dyn Fn(u64) -> Option<String>) -> J {
         Value::Bytes(b) | Value::OtherRef(b) => J::from(hex(b)),
         Value::Enum(_, i) => to_json(&Value::Int(*i), h5paths),
         Value::Compound(ms) => J::Array(ms.iter().map(|(_, v)| to_json(v, h5paths)).collect()),
+        Value::Complex { re, im, .. } => J::Array(vec![to_json(re, h5paths), to_json(im, h5paths)]),
         Value::Array(vs) | Value::Seq(vs) => {
             J::Array(vs.iter().map(|v| to_json(v, h5paths)).collect())
         }

crates/clawhdf5-tools/tests/native_complex_dump.rs

@@ -0,0 +1,202 @@
+//! `h5rs dump` and `h5rs ls` of HDF5 2.0 native complex numbers (datatype
+//! class 11), against the output of h5dump 2.2.0 and h5ls 2.2.0 (the Debian
+//! h5dump CI installs, 1.14.x, predates the type). The h5dump output is
+//! stored next to the fixture; see
+//! `crates/clawhdf5/tests/fixtures/gen_native_complex.py`.
+
+use std::path::PathBuf;
+use std::process::Command;
+
+fn fixture(name: &str) -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+        .join("../clawhdf5/tests/fixtures")
+        .join(name)
+}
+
+fn h5rs(args: &[&str]) -> String {
+    let out = Command::new(env!("CARGO_BIN_EXE_h5rs"))
+        .args(args)
+        .output()
+        .unwrap();
+    assert!(out.status.success(), "h5rs {args:?}: {out:?}");
+    String::from_utf8(out.stdout).unwrap()
+}
+
+/// Split a dump into its lines, with the data lines (`(i): ...`) replaced
+/// by their index, and the data values in order.
+fn split(ddl: &str) -> (Vec<String>, Vec<String>) {
+    let (mut frame, mut data) = (Vec::new(), Vec::new());
+    for line in ddl.lines() {
+        match line.split_once("):") {
+            Some((idx, body)) if idx.trim_start().starts_with('(') => {
+                frame.push(format!("{idx}):"));
+                // Values are split at `, `; array elements come bracketed.
+                data.extend(
+                    body.split(", ")
+                        .map(|s| s.trim().trim_matches(|c| c == '[' || c == ']' || c == ','))
+                        .map(str::trim)
+                        .filter(|s| !s.is_empty() && *s != "{")
+                        .map(String::from),
+                );
+            }
+            _ => {
+                let t = line.trim().trim_end_matches(',');
+                if t.ends_with('i') || t.parse::<f64>().is_ok() {
+                    // A compound member's value on its own line.
+                    data.push(t.to_string());
+                } else {
+                    frame.push(line.to_string());
+                }
+            }
+        }
+    }
+    (frame, data)
+}
+
+/// `a+bi`, `a-bi` or `a+-bi` as its two parts (a real number as one).
+fn parts(v: &str) -> Vec<f64> {
+    let Some(z) = v.strip_suffix('i') else {
+        return vec![v.parse().unwrap()];
+    };
+    let b = z.as_bytes();
+    let at = (1..b.len())
+        .find(|&k| (b[k] == b'+' || b[k] == b'-') && !matches!(b[k - 1], b'e' | b'E' | b'+'))
+        .unwrap_or_else(|| panic!("not a complex value: {v}"));
+    let (re, im) = z.split_at(at);
+    let im = im.strip_prefix('+').unwrap_or(im);
+    vec![re.parse().unwrap(), im.parse().unwrap()]
+}
+
+#[test]
+fn dump_matches_h5dump_2_2() {
+    let path = fixture("native_complex_hdf5_2.h5");
+    let ours = h5rs(&["dump", path.to_str().unwrap()]);
+    let reference = std::fs::read_to_string(fixture("native_complex_hdf5_2.ddl")).unwrap();
+    let (mut our_frame, our_data) = split(&ours);
+    let (mut ref_frame, ref_data) = split(&reference);
+    // The first line names the file as it was given.
+    our_frame.remove(0);
+    ref_frame.remove(0);
+    // Everything but the values is h5dump's byte for byte: the types print
+    // as H5T_COMPLEX_IEEE_F64LE, H5T_ARRAY { [2] H5T_COMPLEX_IEEE_F32LE }, ...
+    assert_eq!(our_frame, ref_frame);
+    assert_eq!(our_data.len(), ref_data.len(), "{our_data:?}\n{ref_data:?}");
+    assert_eq!(our_data.len(), 36);
+    // Values: h5dump prints `%g` (6 digits), inf/nan, and the binary16
+    // parts it has no C type for as `<re>+<im>i` (`1+-2i`); h5rs prints
+    // the shortest round-trip string and Inf/NaN, joined the same way.
+    for (i, (o, r)) in our_data.iter().zip(&ref_data).enumerate() {
+        assert_eq!(o.contains("+-"), r.contains("+-"), "[{i}] {o} vs {r}");
+        let (op, rp) = (parts(o), parts(r));
+        assert_eq!(op.len(), rp.len(), "[{i}] {o} vs {r}");
+        for (x, y) in op.iter().zip(&rp) {
+            let same = if y.is_nan() || y.is_infinite() {
+                x.is_nan() == y.is_nan() && (y.is_nan() || x == y)
+            } else {
+                *x as f32 == *y as f32 || (x - y).abs() <= 5e-6 * y.abs()
+            };
+            assert!(same, "[{i}] h5rs {o}, h5dump {r}");
+            assert_eq!(
+                x.is_sign_negative(),
+                y.is_sign_negative(),
+                "[{i}] {o} vs {r}"
+            );
+        }
+    }
+}
+
+#[test]
+fn ls_describes_complex_like_h5ls_2_2() {
+    let path = fixture("native_complex_hdf5_2.h5");
+    // h5ls 2.2.0 -v, `Type:` of the types it has no native C name for (it
+    // prints `native double _Complex` for the others, as it prints `native
+    // double` where h5rs prints `IEEE 64-bit little-endian float`).
+    for (name, long) in [
+        (
+            "c128be",
+            "complex number of\n                   IEEE 64-bit big-endian float",
+        ),
+        (
+            "c128",
+            "complex number of\n                   IEEE 64-bit little-endian float",
+        ),
+        (
+            "c32",
+            "complex number of\n                   IEEE 16-bit little-endian float",
+        ),
+        (
+            "array",
+            "[2] complex number of\n                   IEEE 32-bit little-endian float",
+        ),
+    ] {
+        let target = format!("{}/{name}", path.display());
+        let verbose = h5rs(&["ls", "-v", &target]);
+        assert!(
+            verbose.contains(&format!("    Type:      {long}\n")),
+            "{name}:\n{verbose}"
+        );
+    }
+    let listing = h5rs(&["ls", path.to_str().unwrap()]);
+    for (name, short) in [
+        ("c64", "complex64"),
+        ("c128", "complex128"),
+        ("c128be", "complex128-be"),
+        ("c32", "complex32"),
+        ("scalar", "complex128"),
+        ("array", "array[2]<complex64>"),
+        ("compound", "compound{z: complex128, k: int64}"),
+    ] {
+        assert!(
+            listing
+                .lines()
+                .any(|l| l.starts_with(&format!("{name} ")) && l.ends_with(&format!(" {short}"))),
+            "{name}:\n{listing}"
+        );
+    }
+}
+
+#[test]
+fn json_dump_uses_the_r_i_compound() {
+    // hdf5-json (h5json 2.0.0) has no complex class: the type is written as
+    // the `{r, i}` compound h5py uses, the values as `[re, im]` pairs.
+    let path = fixture("native_complex_hdf5_2.h5");
+    let j: serde_json::Value =
+        serde_json::from_str(&h5rs(&["dump", "--json", path.to_str().unwrap()])).unwrap();
+    let ds = j["datasets"]
+        .as_object()
+        .unwrap()
+        .values()
+        .find(|d| d["alias"][0] == "/scalar")
+        .unwrap();
+    assert_eq!(ds["type"]["class"], "H5T_COMPOUND");
+    assert_eq!(ds["type"]["fields"][0]["name"], "r");
+    assert_eq!(ds["type"]["fields"][1]["type"]["base"], "H5T_IEEE_F64LE");
+    assert_eq!(ds["value"], serde_json::json!([2.5, -0.5]));
+}
+
+#[test]
+fn diff_compares_complex_values() {
+    let path = fixture("native_complex_hdf5_2.h5");
+    let p = path.to_str().unwrap();
+    // Same values, different byte order: no differences.
+    let out = Command::new(env!("CARGO_BIN_EXE_h5rs"))
+        .args(["diff", p, p, "/c128", "/c128be"])
+        .output()
+        .unwrap();
+    assert!(out.status.success(), "{out:?}");
+    // One element differs in its imaginary part: one difference, exit
+    // status 1, as h5diff 2.2.0 reports it.
+    let out = Command::new(env!("CARGO_BIN_EXE_h5rs"))
+        .args(["diff", "-r", p, p, "/c128", "/c128x"])
+        .output()
+        .unwrap();
+    assert_eq!(out.status.code(), Some(1), "{out:?}");
+    let text = String::from_utf8_lossy(&out.stdout);
+    assert!(text.contains("1 difference(s) found"), "{text}");
+    // h5diff 2.2.0: `[ 1 1 ]  0+1i  1+1i  1+0i`.
+    let row = text.lines().find(|l| l.starts_with("[ 1 1 ]")).unwrap();
+    assert_eq!(
+        row.split_whitespace().collect::<Vec<_>>(),
+        ["[", "1", "1", "]", "0+1i", "1+1i", "1+0i"]
+    );
+}

crates/clawhdf5-wasm/src/core.rs

@@ -312,11 +312,14 @@ impl Reader {
         let base = array_base(dt);
         let is_array = !std::ptr::eq(base, dt);
         Ok(match base {
+            // Read as the part type: an array or complex element is its
+            // parts stored one after another (a complex number's real, then
+            // imaginary part).
             Datatype::FloatingPoint { size, .. } if *size <= 4 => {
-                Data::F32(data_read::read_as_f32(raw, dt).map_err(err)?)
+                Data::F32(data_read::read_as_f32(raw, base).map_err(err)?)
             }
             Datatype::FloatingPoint { .. } => {
-                Data::F64(data_read::read_as_f64(raw, dt).map_err(err)?)
+                Data::F64(data_read::read_as_f64(raw, base).map_err(err)?)
             }
             Datatype::FixedPoint { size, signed, .. } => {
                 let signed_ints = || data_read::read_as_i64(raw, dt).map_err(err);
@@ -393,10 +396,13 @@ fn narrow<S: Copy + std::fmt::Display, T: TryFrom<S>>(v: Vec<S>) -> Result<Vec<T
         .collect()
 }
 
-/// Innermost element type of (possibly nested) array datatypes.
+/// Innermost element type of (possibly nested) array datatypes; for a
+/// native complex type, its part type (each element holds two).
 fn array_base(dt: &Datatype) -> &Datatype {
     match dt {
-        Datatype::Array { base_type, .. } => array_base(base_type),
+        Datatype::Array { base_type, .. } | Datatype::Complex { base_type, .. } => {
+            array_base(base_type)
+        }
         _ => dt,
     }
 }
@@ -411,6 +417,12 @@ fn element_shape(dt: &Datatype) -> Vec<u64> {
             dims.extend(element_shape(base_type));
             dims
         }
+        // `[re, im]`: a complex element reads as its two parts.
+        Datatype::Complex { base_type, .. } => {
+            let mut dims = vec![2];
+            dims.extend(element_shape(base_type));
+            dims
+        }
         _ => Vec::new(),
     }
 }
@@ -487,9 +499,7 @@ pub fn describe(dt: &Datatype) -> String {
         }
     }
     match dt {
-        Datatype::Complex { size, base_type } => {
-            describe(&Datatype::complex_as_compound(*size, base_type))
-        }
+        Datatype::Complex { base_type, .. } => format!("complex<{}>", describe(base_type)),
         Datatype::FixedPoint {
             size,
             signed,
@@ -681,6 +691,74 @@ mod tests {
         assert!(e.contains("not supported"), "{e}");
     }
 
+    #[test]
+    fn native_complex_reads_as_re_im_pairs() {
+        let z = [[1.0f64, -2.0], [0.5, 0.0], [-0.0, 3.25], [f64::MAX, 1e-300]];
+        let mut b = FileBuilder::new();
+        b.create_dataset("z")
+            .with_native_complex_f64_data(&z)
+            .with_shape(&[2, 2]);
+        b.create_dataset("z32")
+            .with_native_complex_f32_data(&[[1.5f32, -2.5]]);
+        let r = Reader::open(b.finish().unwrap()).unwrap();
+        let i = r.info("z").unwrap();
+        assert_eq!(
+            (i.shape, i.dtype, i.element_shape),
+            (vec![2, 2], "complex<f64>".to_string(), vec![2])
+        );
+        let all = r.read("z", None).unwrap();
+        assert_eq!(all.shape, vec![2, 2, 2]);
+        assert_eq!(all.data, Data::F64(z.iter().flatten().copied().collect()));
+        let slab = Hyperslab {
+            start: vec![1, 0],
+            count: vec![1, 2],
+            stride: None,
+            block: None,
+        };
+        let part = r.read("z", Some(&slab)).unwrap();
+        assert_eq!(part.shape, vec![1, 2, 2]);
+        assert_eq!(part.data, Data::F64(vec![-0.0, 3.25, f64::MAX, 1e-300]));
+        let z32 = r.read("z32", None).unwrap();
+        assert_eq!(
+            (z32.shape, z32.data),
+            (vec![1, 2], Data::F32(vec![1.5, -2.5]))
+        );
+    }
+
+    #[test]
+    fn native_complex_written_by_libhdf5_2() {
+        // Written by h5py 3.16 / libhdf5 2.0.0 (gen_native_complex.py).
+        let path = concat!(
+            env!("CARGO_MANIFEST_DIR"),
+            "/../clawhdf5/tests/fixtures/native_complex_hdf5_2.h5"
+        );
+        let r = Reader::open(std::fs::read(path).unwrap()).unwrap();
+        let be = r.read("c128be", None).unwrap();
+        assert_eq!(be.shape, vec![2, 3, 2]);
+        assert_eq!(be.data, r.read("c128", None).unwrap().data);
+        assert_eq!(r.info("c128be").unwrap().dtype, "complex<f64 (big-endian)>");
+        // binary16 parts widen to f32.
+        assert_eq!(
+            r.read("c32", None).unwrap().data,
+            Data::F32(vec![1.0, -2.0, 0.5, 65504.0, 0.0, -0.0])
+        );
+        // An array of complex: array dimensions, then the two parts.
+        let i = r.info("array").unwrap();
+        assert_eq!(
+            (i.dtype.as_str(), i.element_shape),
+            ("array[2]<complex<f32>>", vec![2, 2])
+        );
+        assert_eq!(
+            r.read("array", None).unwrap().data,
+            Data::F32(vec![1.0, 2.0, 3.0, 4.0, -1.0, -1.0, 0.0, 0.0])
+        );
+        let s = r.read("scalar", None).unwrap();
+        assert_eq!((s.shape, s.data), (vec![2], Data::F64(vec![2.5, -0.5])));
+        // A compound holding a complex member is still refused.
+        let e = r.read("compound", None).unwrap_err();
+        assert!(e.contains("compound{z: complex<f64>, k: i64}"), "{e}");
+    }
+
     #[test]
     fn garbage_is_an_error() {
         assert!(Reader::open(vec![0u8; 64]).is_err());

crates/clawhdf5/examples/write_one_chunk.rs

@@ -0,0 +1,46 @@
+//! Write one dataset of `u8` stored as a single chunk, for measuring the
+//! writer's peak memory (`/usr/bin/time -v`):
+//!
+//! ```text
+//! cargo run --release -p clawhdf5 --example write_one_chunk -- BYTES OUT.h5 [slice|owned] [deflate|lz4|zstd]
+//! ```
+//!
+//! `slice` passes the data by reference (`with_u8_data`, which copies it),
+//! `owned` hands the builder the vector (`with_u8_data_owned`). The data is
+//! `i % 251`, one byte per element; with no filter the chunk is stored raw.
+fn main() {
+    let args: Vec<String> = std::env::args().collect();
+    let n: u64 = args[1].parse().expect("BYTES");
+    let out = &args[2];
+    let mode = args.get(3).map_or("slice", String::as_str);
+    let filter = args.get(4).map(String::as_str);
+    let data: Vec<u8> = (0..n).map(|i| (i % 251) as u8).collect();
+    let mut b = clawhdf5::FileBuilder::new();
+    let ds = b.create_dataset("x");
+    match mode {
+        "slice" => {
+            ds.with_u8_data(&data);
+        }
+        "owned" => {
+            ds.with_u8_data_owned(data);
+        }
+        m => panic!("mode {m}"),
+    }
+    ds.with_chunks(&[n]);
+    match filter {
+        None => {}
+        Some("deflate") => {
+            ds.with_deflate(1);
+        }
+        #[cfg(feature = "lz4")]
+        Some("lz4") => {
+            ds.with_lz4();
+        }
+        #[cfg(feature = "zstd")]
+        Some("zstd") => {
+            ds.with_zstd(1);
+        }
+        Some(f) => panic!("filter {f} (not built in?)"),
+    }
+    b.write(out).unwrap();
+}

crates/clawhdf5/src/edit/mod.rs

@@ -408,10 +408,7 @@ impl<'t> ChunkedEdit<'t> {
                 "chunk rank differs from the dataspace".into(),
             ));
         }
-        let cd: Vec<u64> = chunk_dimensions[..rank]
-            .iter()
-            .map(|&c| u64::from(c))
-            .collect();
+        let cd: Vec<u64> = chunk_dimensions[..rank].to_vec();
         // Chunks of 4 GiB or more (HDF5 2.0 writes them with layout message
         // version 5) are read, but not rewritten: the editor holds a chunk
         // it rewrites in memory, compressed and not. Writing values, and a

crates/clawhdf5/src/reader.rs

@@ -1342,7 +1342,10 @@ impl<'f> Dataset<'f> {
     ) -> Result<(Vec<T>, Vec<T>), Error> {
         let dt = self.datatype()?;
         let is_complex = match &dt {
-            // Class 11 parses to this same `{r, i}` compound.
+            Datatype::Complex { size, base_type } => {
+                matches!(base_type.as_ref(), Datatype::FloatingPoint { .. })
+                    && *size == 2 * base_type.type_size()
+            }
             Datatype::Compound { size, members } => {
                 matches!(members.as_slice(), [r, i]
                     if r.name == "r" && i.name == "i"

crates/clawhdf5/src/types.rs

@@ -41,6 +41,13 @@ pub enum DType {
     Array(Box<DType>, Vec<u32>),
     /// Variable-length UTF-8 string stored via a global heap reference.
     VariableLengthString,
+    /// HDF5 2.0 native complex number (datatype class 11) over the given
+    /// floating-point part type: `Complex(F32)` is numpy `complex64`,
+    /// `Complex(F64)` `complex128`, and a half-precision part is
+    /// `Complex(Other("float16"))`. h5py's default complex encoding, the
+    /// compound `{r, i}`, stays a [`DType::Compound`]; `read_complex_f64`
+    /// and `read_complex_f32` read both.
+    Complex(Box<DType>),
     /// Catch-all for HDF5 datatypes that do not map to a specific variant.
     Other(String),
 }
@@ -60,6 +67,7 @@ impl fmt::Display for DType {
             DType::U64 => write!(f, "u64"),
             DType::String => write!(f, "string"),
             DType::VariableLengthString => write!(f, "vlen_string"),
+            DType::Complex(base) => write!(f, "complex<{base}>"),
             DType::Compound(fields) => {
                 write!(f, "compound{{")?;
                 for (i, (name, dt)) in fields.iter().enumerate() {
@@ -148,6 +156,9 @@ pub(crate) fn classify_datatype(dt: &clawhdf5_format::datatype::Datatype) -> DTy
             base_type,
             dimensions,
         } => DType::Array(Box::new(classify_datatype(base_type)), dimensions.clone()),
+        Datatype::Complex { base_type, .. } => {
+            DType::Complex(Box::new(classify_datatype(base_type)))
+        }
         _ => DType::Other(format!("{dt:?}")),
     }
 }

crates/clawhdf5/src/writer.rs

@@ -137,10 +137,15 @@ impl FileBuilder {
         Ok(self.writer.finish()?)
     }
 
-    /// Serialize and write the file to the given path.
+    /// Serialize and write the file to the given path. The file is written
+    /// as it is serialized, not assembled in memory first: unfiltered
+    /// chunks and contiguous data go straight from the datasets' data.
     pub fn write<P: AsRef<std::path::Path>>(self, path: P) -> Result<(), Error> {
-        let bytes = self.finish()?;
-        write_file_atomically(path.as_ref(), &bytes).map_err(Error::Io)
+        use std::io::Write;
+        write_file_atomically_with(path.as_ref(), |f| {
+            self.writer
+                .finish_with(|bytes| f.write_all(bytes).map_err(Error::Io))
+        })
     }
 }
 
@@ -259,8 +264,19 @@ pub fn create_datasets_parallel(specs: Vec<DatasetSpec>) -> Result<Vec<u8>, Erro
 /// file or the complete new one — never a truncated mix. `std::fs::write`
 /// truncates the destination first, so dying mid-write used to destroy the
 /// existing file.
+#[cfg(test)]
 fn write_file_atomically(path: &std::path::Path, bytes: &[u8]) -> std::io::Result<()> {
     use std::io::Write;
+    write_file_atomically_with(path, |f| f.write_all(bytes))
+}
+
+/// [`write_file_atomically`] with the contents written by `write` (into a
+/// buffered writer over the temporary file).
+fn write_file_atomically_with<E: From<std::io::Error>>(
+    path: &std::path::Path,
+    write: impl FnOnce(&mut std::io::BufWriter<std::fs::File>) -> Result<(), E>,
+) -> Result<(), E> {
+    use std::io::Write;
 
     // Same directory as the target, so the rename stays on one filesystem.
     let mut tmp_name = path
@@ -272,11 +288,18 @@ fn write_file_atomically(path: &std::path::Path, bytes: &[u8]) -> std::io::Resul
     tmp_name.push(format!(".tmp-{}", std::process::id()));
     let tmp_path = path.with_file_name(tmp_name);
 
-    let result = (|| {
-        let mut f = std::fs::File::create(&tmp_path)?;
-        f.write_all(bytes)?;
-        f.sync_all()?;
-        std::fs::rename(&tmp_path, path)
+    let result = (|| -> Result<(), E> {
+        // 64 KiB, below glibc's 128 KiB mmap threshold: a 1 MiB buffer was
+        // mapped afresh by each write once the threshold had not risen, and
+        // its page faults made a 1 MiB chunked write 1.6x slower (criterion
+        // `write_2d_chunked/512x512` after the smaller cases, 2026-09-29).
+        // Writes larger than the buffer go straight to the file.
+        let mut f = std::io::BufWriter::with_capacity(64 << 10, std::fs::File::create(&tmp_path)?);
+        write(&mut f)?;
+        f.flush()?;
+        f.get_ref().sync_all()?;
+        std::fs::rename(&tmp_path, path)?;
+        Ok(())
     })();
     if result.is_err() {
         let _ = std::fs::remove_file(&tmp_path);

crates/clawhdf5/tests/fixtures/gen_huge_chunks.py

@@ -2,6 +2,12 @@
 
     python gen_huge_chunks.py filtered OUT.h5     # huge_chunks_filtered.h5
     python gen_huge_chunks.py unfiltered OUT.h5   # generated at test time
+    python gen_huge_chunks.py dims OUT.h5         # huge_chunk_dims.h5
+    python gen_huge_chunks.py dims-unfiltered OUT.h5   # generated at test time
+
+`dims` and `dims-unfiltered` (see `make_dims`) hold `u1` datasets whose
+chunk dimension is N8 = 2**32 + 7: a chunk dimension of 2**32 or more,
+which a layout message of version 5 stores in 5 bytes.
 
 libhdf5 2.x writes a chunk of more than 0xFFFFFFFF bytes with layout message
 version 5 (`H5D__chunk_construct`: "chunk size > 4GB requires
@@ -38,7 +44,8 @@ allocated early (see `make`).
 libhdf5 holds a whole filtered chunk in memory while it writes it, so the
 `filtered` run needs about 4 GiB of memory and 100 s (tank).
 
-Generated 2026-09-28 with h5py 3.16.0 (libhdf5 2.0.0).
+huge_chunks_filtered.h5 was generated 2026-09-28 and huge_chunk_dims.h5
+2026-09-29, both with h5py 3.16.0 (libhdf5 2.0.0).
 """
 import sys
 
@@ -92,12 +99,60 @@ def make(fid, name, index, filtered):
     dsid.close()
 
 
+N8 = 2**32 + 7
+
+
+def make_dims(fid, name, index, filtered):
+    """A `u1` dataset whose chunks are N8 elements (a chunk dimension of
+    2**32 or more). `d[0:10] = 0..9` and `d[N8-10:N8] = 100..109`; where
+    there is a second chunk (`farray`, `earray`: shape N8 + 10),
+    `d[N8:N8+10] = 200..209`.
+
+    filtered (`dims`, committed): deflate level 9 twice, fill value 7;
+    `single` (Single Chunk) and `earray` (Extensible Array, maxshape None).
+    Writing it holds a 4 GiB chunk: 4.1 GiB peak resident memory and about
+    a minute (tank, 2026-09-29).
+    unfiltered (`dims-unfiltered`): fill time "never", sparse; `single`
+    (allocated early, as in `make`) and `farray` (Fixed Array).
+    """
+    dcpl = h5py.h5p.create(h5py.h5p.DATASET_CREATE)
+    if index == "single":
+        shape, maxshape = (N8,), (N8,)
+    elif index == "farray":
+        shape, maxshape = (N8 + 10,), (N8 + 10,)
+    elif index == "earray":
+        shape, maxshape = (N8 + 10,), (UNLIM,)
+    dcpl.set_chunk((N8,))
+    if filtered:
+        dcpl.set_deflate(9)
+        dcpl.set_deflate(9)
+        dcpl.set_fill_value(np.array(7, dtype="u1"))
+    else:
+        dcpl.set_fill_time(h5py.h5d.FILL_TIME_NEVER)
+        if index == "single":
+            dcpl.set_alloc_time(h5py.h5d.ALLOC_TIME_EARLY)
+    space = h5py.h5s.create_simple(shape, maxshape)
+    dsid = h5py.h5d.create(fid, name.encode(), h5py.h5t.STD_U8LE, space, dcpl=dcpl)
+    ds = h5py.Dataset(dsid)
+    ds[0:10] = np.arange(10, dtype="u1")
+    ds[N8 - 10 : N8] = np.arange(100, 110, dtype="u1")
+    if shape[0] > N8:
+        ds[N8 : N8 + 10] = np.arange(200, 210, dtype="u1")
+    dsid.close()
+
+
 def main():
     mode, out = sys.argv[1], sys.argv[2]
-    filtered = {"filtered": True, "unfiltered": False}[mode]
     fapl = h5py.h5p.create(h5py.h5p.FILE_ACCESS)
     fapl.set_libver_bounds(h5py.h5f.LIBVER_EARLIEST, h5py.h5f.LIBVER_V200)
     fid = h5py.h5f.create(out.encode(), h5py.h5f.ACC_TRUNC, fapl=fapl)
+    if mode in ("dims", "dims-unfiltered"):
+        filtered = mode == "dims"
+        for index in ["single", "earray"] if filtered else ["single", "farray"]:
+            make_dims(fid, index, index, filtered)
+        fid.close()
+        return
+    filtered = {"filtered": True, "unfiltered": False}[mode]
     indexes = ["single", "farray", "earray", "btree2"]
     if not filtered:
         indexes.insert(1, "implicit")

crates/clawhdf5/tests/fixtures/gen_native_complex.py

@@ -0,0 +1,75 @@
+"""Generate native_complex_hdf5_2.h5: HDF5 2.0 native complex (class 11).
+
+Datasets (every one class 11, or a type holding it):
+
+    c64       H5T_COMPLEX_IEEE_F32LE, 4 values incl. -0, inf, nan, subnormal
+    c128      H5T_COMPLEX_IEEE_F64LE, 2 x 3
+    c128be    H5T_COMPLEX_IEEE_F64BE, the same values
+    c128x     H5T_COMPLEX_IEEE_F64LE, c128 with element (1,1) changed
+    c32       H5T_COMPLEX_IEEE_F16LE, 3 values
+    scalar    H5T_COMPLEX_IEEE_F64LE, scalar dataspace
+    compound  {z: H5T_COMPLEX_IEEE_F64LE @0, k: H5T_STD_I64LE @16}
+    array     H5T_ARRAY [2] of H5T_COMPLEX_IEEE_F32LE, 2 elements
+
+plus the root attribute `zattr` (H5T_COMPLEX_IEEE_F64LE, 2 values).
+
+Written with h5py 3.16 (libhdf5 2.0.0) through its low-level API, the file
+type as the memory type (no conversion). Re-run only if the fixture ever
+needs regenerating (generated 2026-09-29):
+
+    python gen_native_complex.py native_complex_hdf5_2.h5
+
+The h5dump / h5ls 2.2.0 reference output the tests compare with was made
+with the tools of the libhdf5 2.2.0 build described in gen_mx_floats.py:
+
+    h5dump native_complex_hdf5_2.h5 > native_complex_hdf5_2.ddl
+    h5ls -r -v native_complex_hdf5_2.h5 > native_complex_hdf5_2.ls
+"""
+import sys
+
+import h5py
+import numpy as np
+from h5py import h5a, h5d, h5s, h5t
+
+out = sys.argv[1]
+f = h5py.File(out, "w", libver="latest")
+
+
+def ds(name, tid, arr, shape=None):
+    shape = arr.shape if shape is None else shape
+    sp = h5s.create_simple(shape) if shape != () else h5s.create(h5s.SCALAR)
+    d = h5d.create(f.id, name.encode(), tid, sp)
+    d.write(h5s.ALL, h5s.ALL, np.ascontiguousarray(arr), mtype=tid)
+
+
+nan, inf = float("nan"), float("inf")
+c64 = np.array(
+    [1.5 - 2j, complex(0.0, -0.0), complex(inf, nan), complex(-inf, 1e-40)],
+    dtype=np.complex64,
+)
+ds("c64", h5t.COMPLEX_IEEE_F32LE, c64)
+c128 = np.array(
+    [[1 + 2j, -3.5 + 4e300j, 0.1 + 0.2j], [5e-324 - 0j, 1j, -1]], dtype=np.complex128
+)
+ds("c128", h5t.COMPLEX_IEEE_F64LE, c128)
+ds("c128be", h5t.COMPLEX_IEEE_F64BE, c128.astype(">c16"))
+c128x = c128.copy()
+c128x[1, 1] = 1 + 1j
+ds("c128x", h5t.COMPLEX_IEEE_F64LE, c128x)
+half = np.array([1.0, -2.0, 0.5, 65504.0, 0.0, -0.0], dtype=np.float16)
+ds("c32", h5t.COMPLEX_IEEE_F16LE, half, shape=(3,))
+ds("scalar", h5t.COMPLEX_IEEE_F64LE, np.array(2.5 - 0.5j), shape=())
+
+ct = h5t.create(h5t.COMPOUND, 24)
+ct.insert(b"z", 0, h5t.COMPLEX_IEEE_F64LE)
+ct.insert(b"k", 16, h5t.STD_I64LE)
+cd = np.array([(1 + 1j, 7), (-2.5 + 0j, -8)], dtype=[("z", "<c16"), ("k", "<i8")])
+ds("compound", ct, cd)
+
+at = h5t.array_create(h5t.COMPLEX_IEEE_F32LE, (2,))
+ad = np.array([[1 + 2j, 3 + 4j], [-1 - 1j, 0j]], dtype=np.complex64)
+ds("array", at, ad, shape=(2,))
+
+a = h5a.create(f.id, b"zattr", h5t.COMPLEX_IEEE_F64LE, h5s.create_simple((2,)))
+a.write(np.array([0.5 - 1.5j, 2j]), mtype=h5t.COMPLEX_IEEE_F64LE)
+f.close()

crates/clawhdf5/tests/fixtures/native_complex_hdf5_2.ddl

@@ -0,0 +1,80 @@
+HDF5 "native_complex_hdf5_2.h5" {
+GROUP "/" {
+   ATTRIBUTE "zattr" {
+      DATATYPE  H5T_COMPLEX_IEEE_F64LE
+      DATASPACE  SIMPLE { ( 2 ) / ( 2 ) }
+      DATA {
+      (0): 0.5-1.5i, 0+2i
+      }
+   }
+   DATASET "array" {
+      DATATYPE  H5T_ARRAY { [2] H5T_COMPLEX_IEEE_F32LE }
+      DATASPACE  SIMPLE { ( 2 ) / ( 2 ) }
+      DATA {
+      (0): [ 1+2i, 3+4i ], [ -1-1i, 0+0i ]
+      }
+   }
+   DATASET "c128" {
+      DATATYPE  H5T_COMPLEX_IEEE_F64LE
+      DATASPACE  SIMPLE { ( 2, 3 ) / ( 2, 3 ) }
+      DATA {
+      (0,0): 1+2i, -3.5+4e+300i, 0.1+0.2i,
+      (1,0): 4.94066e-324-0i, 0+1i, -1+0i
+      }
+   }
+   DATASET "c128be" {
+      DATATYPE  H5T_COMPLEX_IEEE_F64BE
+      DATASPACE  SIMPLE { ( 2, 3 ) / ( 2, 3 ) }
+      DATA {
+      (0,0): 1+2i, -3.5+4e+300i, 0.1+0.2i,
+      (1,0): 4.94066e-324-0i, 0+1i, -1+0i
+      }
+   }
+   DATASET "c128x" {
+      DATATYPE  H5T_COMPLEX_IEEE_F64LE
+      DATASPACE  SIMPLE { ( 2, 3 ) / ( 2, 3 ) }
+      DATA {
+      (0,0): 1+2i, -3.5+4e+300i, 0.1+0.2i,
+      (1,0): 4.94066e-324-0i, 1+1i, -1+0i
+      }
+   }
+   DATASET "c32" {
+      DATATYPE  H5T_COMPLEX_IEEE_F16LE
+      DATASPACE  SIMPLE { ( 3 ) / ( 3 ) }
+      DATA {
+      (0): 1+-2i, 0.5+65504i, 0+-0i
+      }
+   }
+   DATASET "c64" {
+      DATATYPE  H5T_COMPLEX_IEEE_F32LE
+      DATASPACE  SIMPLE { ( 4 ) / ( 4 ) }
+      DATA {
+      (0): 1.5-2i, 0-0i, inf+nani, -inf+9.99995e-41i
+      }
+   }
+   DATASET "compound" {
+      DATATYPE  H5T_COMPOUND {
+         H5T_COMPLEX_IEEE_F64LE "z";
+         H5T_STD_I64LE "k";
+      }
+      DATASPACE  SIMPLE { ( 2 ) / ( 2 ) }
+      DATA {
+      (0): {
+            1+1i,
+            7
+         },
+      (1): {
+            -2.5+0i,
+            -8
+         }
+      }
+   }
+   DATASET "scalar" {
+      DATATYPE  H5T_COMPLEX_IEEE_F64LE
+      DATASPACE  SCALAR
+      DATA {
+      (0): 2.5-0.5i
+      }
+   }
+}
+}

crates/clawhdf5/tests/huge_chunks_interop.rs

@@ -81,8 +81,15 @@ fn layout_of(
     let lm = msg(MessageType::DataLayout);
     let layout = DataLayout::parse(&lm.data, sb.offset_size, sb.length_size).unwrap();
     let space = Dataspace::parse(&msg(MessageType::Dataspace).data, sb.length_size).unwrap();
+    // The element size is the layout's last dimension.
+    let elem = match &layout {
+        DataLayout::Chunked {
+            chunk_dimensions, ..
+        } => *chunk_dimensions.last().unwrap() as usize,
+        _ => 8,
+    };
     let (chunks, _) =
-        list_chunks(data, &layout, &space, 8, sb.offset_size, sb.length_size).unwrap();
+        list_chunks(data, &layout, &space, elem, sb.offset_size, sb.length_size).unwrap();
     (lm.data[0], layout, space, chunks)
 }
 
@@ -516,3 +523,343 @@ print("ok")
         }
     }
 }
+
+// ---- Chunk dimensions of 2^32 or more ----
+
+/// Elements per chunk of the `u8` datasets below: a chunk dimension of
+/// 2^32 or more (4 GiB + 7 bytes), which a version-5 layout message stores
+/// in 5 bytes.
+const N8: u64 = (1 << 32) + 7;
+
+fn dims_fixture() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/fixtures/huge_chunk_dims.h5")
+}
+
+/// `u8` values of `sel` in dataset `name` of `file`.
+fn sel_u8(file: &File, name: &str, start: u64, count: u64) -> Vec<u8> {
+    let ds = file.dataset(name).unwrap();
+    let s = Selection::Hyperslab {
+        start: vec![start],
+        stride: vec![1],
+        count: vec![count],
+        block: vec![1],
+    };
+    ds.read_selection(&s).unwrap()
+}
+
+fn u8s(range: std::ops::Range<u8>) -> Vec<u8> {
+    range.collect()
+}
+
+/// `fixtures/huge_chunk_dims.h5` (libhdf5 2.0.0 through h5py 3.16,
+/// `gen_huge_chunks.py dims`): `u8` datasets chunked by N8 = 2^32 + 7, in a
+/// Single Chunk and an Extensible Array index. Their layouts parse to the
+/// whole dimension (it was refused: chunk dimensions were `u32`) and their
+/// chunks list at the right origins.
+#[test]
+fn huge_chunk_dims_list() {
+    let data = std::fs::read(dims_fixture()).unwrap();
+    for (name, index, shape, origins) in [
+        ("single", 1u8, N8, &[0u64][..]),
+        ("earray", 4, N8 + 10, &[0, N8]),
+    ] {
+        let (version, layout, space, mut chunks) = layout_of(&data, name);
+        assert_eq!(version, 5, "{name}");
+        let DataLayout::Chunked {
+            chunk_dimensions,
+            chunk_index_type,
+            ..
+        } = &layout
+        else {
+            panic!("{name}: {layout:?}");
+        };
+        assert_eq!(chunk_dimensions, &[N8, 1], "{name}");
+        assert_eq!(*chunk_index_type, Some(index), "{name}");
+        assert_eq!(space.dimensions, [shape], "{name}");
+        chunks.sort_by_key(|c| c.offsets[0]);
+        let got: Vec<u64> = chunks.iter().map(|c| c.offsets[0]).collect();
+        assert_eq!(got, origins, "{name}");
+    }
+}
+
+/// Decoding the fixture's chunks (4 GiB each, one at a time): written
+/// values, the fill value (7) around them, and the second chunk.
+#[test]
+fn huge_chunk_dims_read() {
+    if !heavy() {
+        return;
+    }
+    let file = File::open(dims_fixture()).unwrap();
+    let mut first = u8s(0..10);
+    first.extend([7, 7]);
+    let mut last = vec![7, 7];
+    last.extend(u8s(100..110));
+    for name in ["single", "earray"] {
+        assert_eq!(sel_u8(&file, name, 0, 12), first, "{name}");
+        assert_eq!(sel_u8(&file, name, N8 - 12, 12), last, "{name}");
+    }
+    let mut across = u8s(108..110);
+    across.extend(u8s(200..210));
+    assert_eq!(sel_u8(&file, "earray", N8 - 2, 12), across);
+}
+
+/// Unfiltered chunks of N8 bytes that libhdf5 2.0.0 wrote (a sparse file,
+/// `gen_huge_chunks.py dims-unfiltered`), in a Single Chunk and a Fixed
+/// Array index: read through a memory map, and through positioned reads
+/// that fetch only the rows selected.
+#[test]
+fn huge_chunk_dims_unfiltered_read() {
+    if !heavy() {
+        return;
+    }
+    let dir = scratch();
+    let Some(path) = generate(dir.path(), "dims-unfiltered") else {
+        return;
+    };
+    let check = |file: &File| {
+        let mut first = u8s(0..10);
+        first.extend([0, 0]);
+        let mut last = vec![0, 0];
+        last.extend(u8s(100..110));
+        for name in ["single", "farray"] {
+            assert_eq!(sel_u8(file, name, 0, 12), first, "{name}");
+            assert_eq!(sel_u8(file, name, N8 - 12, 12), last, "{name}");
+        }
+        let mut across = u8s(108..110);
+        across.extend(u8s(200..210));
+        assert_eq!(sel_u8(file, "farray", N8 - 2, 12), across);
+    };
+    check(&File::open(&path).unwrap());
+    let file = std::fs::File::open(&path).unwrap();
+    let len = file.metadata().unwrap().len();
+    let storage = std::sync::Arc::new(Counting {
+        file,
+        len,
+        read: 0.into(),
+    });
+    check(&File::open_storage(storage.clone()).unwrap());
+    let read = storage.read.load(std::sync::atomic::Ordering::Relaxed);
+    assert!(read < 1 << 20, "{read} bytes read");
+}
+
+/// Run `script` under h5py with `path` as `sys.argv[1]`; skipped without
+/// h5py (a failure under `CLAWHDF5_REQUIRE_INTEROP=1`).
+fn h5py_check(script: &str, path: &Path) {
+    if !python_available() {
+        assert!(!interop_required(), "h5py not available");
+        eprintln!("SKIP: python3 with h5py not available");
+        return;
+    }
+    let out = Command::new(python())
+        .arg("-c")
+        .arg(script)
+        .arg(path)
+        .output()
+        .unwrap();
+    assert!(
+        out.status.success(),
+        "h5py:\n{}",
+        String::from_utf8_lossy(&out.stderr)
+    );
+}
+
+/// `h5dump` 2.x's output for `args` on `path`, checked to succeed without
+/// errors; `None` without `CLAWHDF5_H5DUMP2`.
+fn h5dump2_text(args: &[&str], path: &Path) -> Option<String> {
+    let h5dump = h5dump2()?;
+    let out = Command::new(&h5dump).args(args).arg(path).output().unwrap();
+    let text = format!(
+        "{}{}",
+        String::from_utf8_lossy(&out.stdout),
+        String::from_utf8_lossy(&out.stderr)
+    );
+    assert!(out.status.success(), "h5dump {args:?}:\n{text}");
+    assert!(!text.contains("rror"), "h5dump {args:?}:\n{text}");
+    Some(text)
+}
+
+/// clawhdf5 writes deflated `u8` chunks of N8 elements (a chunk dimension
+/// of 2^32 or more) in a Single Chunk and an Extensible Array index; it,
+/// h5py (libhdf5 2.0.0) and h5dump 2.2.0 (layouts only: it cannot inflate
+/// chunks over 4 GiB, see known-issues) read them.
+#[test]
+fn writer_huge_chunk_dims_round_trip() {
+    if !heavy() {
+        return;
+    }
+    let dir = scratch();
+    let path = dir.path().join("dims.h5");
+    let values = u8s(0..10);
+    let mut b = clawhdf5::FileBuilder::new();
+    for (name, max) in [("single", N8), ("earray", u64::MAX)] {
+        b.create_dataset(name)
+            .with_u8_data(&values)
+            .with_maxshape(&[max])
+            .with_chunks(&[N8])
+            .with_deflate(1);
+    }
+    b.write(&path).unwrap();
+
+    let data = std::fs::read(&path).unwrap();
+    // Deflate level 1 leaves about 40 MiB of each 4 GiB chunk of zeros.
+    assert!(data.len() < 256 << 20, "{} bytes", data.len());
+    for (name, index) in [("single", 1), ("earray", 4)] {
+        let (version, layout, _, chunks) = layout_of(&data, name);
+        assert_eq!(version, 5, "{name}");
+        assert!(
+            matches!(&layout, DataLayout::Chunked { chunk_dimensions, chunk_index_type: Some(t), .. }
+                if *t == index && chunk_dimensions == &[N8, 1]),
+            "{name}: {layout:?}"
+        );
+        assert_eq!(chunks.len(), 1, "{name}");
+    }
+    let file = File::open(&path).unwrap();
+    for name in ["single", "earray"] {
+        assert_eq!(sel_u8(&file, name, 0, 10), values, "{name}");
+    }
+    drop(file);
+
+    h5py_check(
+        r#"
+import sys, h5py
+with h5py.File(sys.argv[1], "r") as f:
+    for name in ("single", "earray"):
+        d = f[name]
+        assert d.chunks == (2**32 + 7,), (name, d.chunks)
+        assert list(d[:]) == list(range(10)), (name, d[:])
+"#,
+        &path,
+    );
+    for name in ["single", "earray"] {
+        if let Some(text) = h5dump2_text(&["-H", "-p", "-d", name], &path) {
+            assert!(text.contains("CHUNKED ( 4294967303 )"), "{name}:\n{text}");
+        }
+    }
+}
+
+/// An unfiltered chunk of N8 bytes (a chunk dimension of 2^32 or more)
+/// written by clawhdf5 from data it was handed (`with_u8_data_owned`):
+/// `FileBuilder::write` streams it to the file without copying it, so the
+/// write needs about the data's own 4 GiB. Read back by clawhdf5, h5py
+/// (a few elements) and h5dump 2.2.0.
+#[test]
+fn writer_unfiltered_huge_chunk_round_trip() {
+    if !heavy() {
+        return;
+    }
+    let dir = scratch();
+    let path = dir.path().join("raw.h5");
+    let value = |i: u64| (i % 251) as u8;
+    let mut b = clawhdf5::FileBuilder::new();
+    b.create_dataset("single")
+        .with_u8_data_owned((0..N8).map(value).collect())
+        .with_chunks(&[N8]);
+    b.write(&path).unwrap();
+
+    let file = File::open(&path).unwrap();
+    let data = file.as_bytes();
+    let (version, layout, space, chunks) = layout_of(data, "single");
+    assert_eq!(version, 5);
+    assert!(
+        matches!(&layout, DataLayout::Chunked { chunk_dimensions, chunk_index_type: Some(1), .. }
+            if chunk_dimensions == &[N8, 1]),
+        "{layout:?}"
+    );
+    assert_eq!(space.dimensions, [N8]);
+    assert_eq!(chunks.len(), 1);
+    assert_eq!(chunks[0].chunk_size, N8);
+    for start in [0, (1 << 32) - 3, N8 - 6] {
+        let want: Vec<u8> = (start..start + 6).map(value).collect();
+        assert_eq!(sel_u8(&file, "single", start, 6), want, "at {start}");
+    }
+    drop(file);
+
+    h5py_check(
+        r#"
+import sys, h5py, numpy as np
+n = 2**32 + 7
+with h5py.File(sys.argv[1], "r") as f:
+    d = f["single"]
+    assert d.chunks == (n,) and d.shape == (n,) and d.compression is None
+    for start in (0, 2**32 - 3, n - 6):
+        want = [i % 251 for i in range(start, start + 6)]
+        assert list(d[start:start + 6]) == want, (start, d[start:start + 6])
+"#,
+        &path,
+    );
+    let start = (1u64 << 32) - 3;
+    if let Some(text) = h5dump2_text(
+        &["-d", "single", "-s", &start.to_string(), "-c", "6"],
+        &path,
+    ) {
+        let want: Vec<String> = (start..start + 6).map(|i| value(i).to_string()).collect();
+        assert!(text.contains(&want.join(", ")), "{text}");
+    }
+}
+
+/// LZ4 and Zstandard chunks of 4 GiB or more: written by clawhdf5, read
+/// back by it and by h5py (through `hdf5plugin`'s LZ4 and Zstandard
+/// filters when installed).
+#[cfg(any(feature = "lz4", feature = "zstd"))]
+#[test]
+fn writer_huge_chunks_lz4_zstd_round_trip() {
+    if !heavy() {
+        return;
+    }
+    let dir = scratch();
+    let path = dir.path().join("codecs.h5");
+    let values = u8s(0..10);
+    let mut b = clawhdf5::FileBuilder::new();
+    let mut names = Vec::new();
+    #[cfg(feature = "lz4")]
+    {
+        b.create_dataset("lz4")
+            .with_u8_data(&values)
+            .with_maxshape(&[N8])
+            .with_chunks(&[N8])
+            .with_lz4();
+        names.push("lz4");
+    }
+    #[cfg(feature = "zstd")]
+    {
+        b.create_dataset("zstd")
+            .with_u8_data(&values)
+            .with_maxshape(&[N8])
+            .with_chunks(&[N8])
+            .with_zstd(3);
+        names.push("zstd");
+    }
+    b.write(&path).unwrap();
+    let data = std::fs::read(&path).unwrap();
+    assert!(data.len() < 256 << 20, "{} bytes", data.len());
+    for name in &names {
+        let (version, _, _, chunks) = layout_of(&data, name);
+        assert_eq!(version, 5, "{name}");
+        assert_eq!(chunks.len(), 1, "{name}");
+    }
+    drop(data);
+    let file = File::open(&path).unwrap();
+    for name in &names {
+        assert_eq!(sel_u8(&file, name, 0, 10), values, "{name}");
+    }
+    drop(file);
+    h5py_check(
+        &format!(
+            r#"
+import sys, h5py
+try:
+    import hdf5plugin
+except ImportError:
+    print("SKIP: no hdf5plugin")
+    sys.exit(0)
+with h5py.File(sys.argv[1], "r") as f:
+    for name in {names:?}:
+        d = f[name]
+        assert d.chunks == (2**32 + 7,), (name, d.chunks)
+        assert list(d[0:10]) == list(range(10)), (name, d[0:10])
+"#,
+            names = names
+        ),
+        &path,
+    );
+}

crates/clawhdf5/tests/integration_tests.rs

@@ -1064,12 +1064,26 @@ fn complex_datasets_and_attributes_round_trip() {
         let ds = file.dataset(name).unwrap();
         assert_eq!(ds.read_complex_f32().unwrap(), z64, "{name}");
         assert_eq!(ds.shape().unwrap(), vec![3]);
-        // Both encodings read as the same {r, i} compound.
-        assert_eq!(
-            ds.dtype().unwrap(),
+        // h5py's encoding is a compound, the native one a complex type.
+        let want = if name == "native64" {
+            DType::Complex(Box::new(DType::F32))
+        } else {
             DType::Compound(vec![("r".into(), DType::F32), ("i".into(), DType::F32)])
-        );
+        };
+        assert_eq!(ds.dtype().unwrap(), want, "{name}");
     }
+    assert_eq!(
+        file.dataset("native128").unwrap().dtype().unwrap(),
+        DType::Complex(Box::new(DType::F64))
+    );
+    assert_eq!(
+        file.dataset("native128").unwrap().raw_datatype().unwrap(),
+        clawhdf5::make_native_complex_f64_type()
+    );
+    assert_eq!(
+        DType::Complex(Box::new(DType::F64)).to_string(),
+        "complex<f64>"
+    );
     for name in ["compound128", "native128"] {
         same(
             file.dataset(name).unwrap().read_complex_f64().unwrap(),
@@ -1094,7 +1108,7 @@ fn complex_datasets_and_attributes_round_trip() {
             data,
         }) => {
             assert!(shape.is_empty());
-            assert_eq!(datatype.type_size(), 16);
+            assert_eq!(datatype, clawhdf5::make_native_complex_f64_type());
             let fields =
                 clawhdf5_format::data_read::read_compound_fields(&data, &datatype).unwrap();
             let part = |k: usize| {

docs/known-issues.md

@@ -17,6 +17,11 @@ Checked against `main` at `9b5803f` on 2026-09-28.
 |---|---|---|
 | [NetCDF-4: differences from netCDF-C](#netcdf-4-differences-from-netcdf-c) | deliberate (floats of other than 4 or 8 bytes, axes netCDF-C leaves without a dimension), and 9 corpus files (external links, values the HDF5 reader refuses) | 2026-09-29 |
 | [Chunks of 4 GiB or more: limits](#chunks-of-4-gib-or-more-limits) | refused (chunk dimensions of 2^32 or more, five filters, editor rewrites), memory (a decoded chunk is held whole) | 2026-09-28 |
+
+
+| [NetCDF-4: variables' dimensions are guessed from sizes](#netcdf-4-variables-dimensions-are-guessed-from-sizes) | **wrong metadata** (`Variable::dimensions`, pure dimension scales listed as variables; values and dimension sizes are right) | 2026-09-28 |
+| [Chunks of 4 GiB or more: limits](#chunks-of-4-gib-or-more-limits) | refused (five filters, editor rewrites), memory (a decoded chunk is held whole) | 2026-09-28 |
+| [NetCDF-4: an unlimited dimension reports size 0](#netcdf-4-an-unlimited-dimension-reports-size-0) | **wrong metadata** (dimension size; variable shapes and values are right) | 2026-09-28 |
 | [Small floats decode as libhdf5 does, not as the OCP MX specification](#small-floats-decode-as-libhdf5-does-not-as-the-ocp-mx-specification) | deliberate: libhdf5's values (FP4/FP6/FP8 E4M3 all-ones exponent is inf/NaN) | 2026-09-28 |
 | [In-place modification (`FileEditor`) limits](#in-place-modification-fileeditor-limits) | refused edits (`Error::Unsupported`), space reuse per editor, no journal | 2026-09-26 |
 | [Python in-place editing limits](#python-in-place-editing-clawhdf5filepath-r-limits) | refused writes (`NotImplementedError`), deliberate conversion differences | 2026-09-27 |
@@ -241,7 +246,10 @@ Values are correct in every case; this is cost only. Selections other than
 
 ## Chunks of 4 GiB or more: limits
 
-**Status:** open (documented 2026-09-28). HDF5 2.0 writes a chunk of more
+**Status:** open (documented 2026-09-28; updated 2026-09-29, when
+[chunk dimensions of 2^32 or more](#chunk-dimensions-of-232-or-more-were-refused)
+and [unfiltered writes](#writing-a-4-gib-unfiltered-chunk-held-several-copies-of-it)
+were fixed). HDF5 2.0 writes a chunk of more
 than 0xFFFFFFFF bytes with layout message version 5 (`H5D__chunk_construct`
 in libhdf5 2.2.0: "chunk size > 4GB requires H5F_LIBVER_V200"), which
 also makes a filtered chunk index element store the chunk's size in "size
@@ -253,34 +261,40 @@ The [fix](#chunks-of-4-gib-or-more-could-not-be-read) is tested by
 `crates/clawhdf5/tests/huge_chunks_interop.rs`; its decoding tests are
 opt-in (`CLAWHDF5_HUGE_CHUNKS=1`, one at a time: `--test-threads=1`).
 What remains:
-- **Chunk dimensions of 2^32 or more** (which libhdf5 2.x also allows under
-  `H5F_LIBVER_V200`) are refused, when read (`InvalidChunkDimensions`)
-  and when written: `DataLayout::Chunked::chunk_dimensions` is `Vec<u32>`.
-  A 4 GiB chunk of 1-byte elements needs one.
 - **Filters the writer refuses for such a chunk** (`FilterError`, before
   anything is written): LZF, bitshuffle, bzip2 and Blosc (their HDF5
   filters record sizes or block lengths in 32 bits, or cannot take such a
   buffer) and pcodec. Deflate, shuffle, Fletcher-32, LZ4 and Zstd are
-  written; only shuffle + deflate is tested end to end (LZ4's framing and
-  the 256 MiB limit it had are unit-tested; Zstd is not tested at this
-  size).
-- **Unfiltered chunks this size are written but not tested end to end:**
-  `FileBuilder` assembles the whole file in memory, several copies of the
-  chunk (well over the 12 GiB the tests may use). Their layout messages
-  are unit-tested (`chunked_write::tests::huge_chunk_layout_messages_and_index_elements`).
+  written and tested end to end (`writer_huge_chunks_round_trip`,
+  `writer_huge_chunk_dims_round_trip`,
+  `writer_huge_chunks_lz4_zstd_round_trip`: read back by clawhdf5 and
+  h5py 3.16, LZ4 and Zstd through hdf5plugin 7.1.0).
 - **`FileEditor`** refuses to rewrite such chunks (see
   [its limits](#in-place-modification-fileeditor-limits)).
 - **Memory:** decoding a filtered chunk holds all of it (4 GiB and more),
   twice when it is shuffled (the inflated and the unshuffled copy, as
-  libhdf5's filters do too); writing one holds the chunk and its shuffled
-  copy. A selection decodes only the chunks it touches; one of an
-  unfiltered chunk reads only the rows it selects, through a memory map or
+  libhdf5's filters do too). Writing a filtered chunk holds its compressed
+  copy and, where it is not a contiguous run of the dataset's data (a
+  chunk past the dataset's edge is zero-padded), the raw chunk; shuffled,
+  the shuffled copy too. An unfiltered chunk is not copied when it is such
+  a run (a dataset stored as one chunk of its own shape): writing one
+  unfiltered chunk of 2^32 + 7 bytes with `with_u8_data_owned` and
+  `FileBuilder::write` peaked at 4.00 GiB resident, the data itself
+  (`write_one_chunk 4294967303 OUT owned`, tank, 2026-09-29, commit
+  `0065c6b`; see [the fix](#writing-a-4-gib-unfiltered-chunk-held-several-copies-of-it)).
+  `FileWriter::finish` (a `Vec`) holds the data and the file.
+  A selection decodes only the chunks it touches; one of an unfiltered
+  chunk reads only the rows it selects, through a memory map or
   positioned reads (`File::open_storage`): every selection of
   `unfiltered_huge_chunks_read` together read under 1 MiB. Peak resident
-  memory of `cargo test --release -p clawhdf5 --test huge_chunks_interop
-  -- --test-threads=1` with `CLAWHDF5_HUGE_CHUNKS=1` (all six tests, h5py
-  included) was 8.06 GiB (tank, 2026-09-28, commit `9143737`); selection
-  reads of the double-deflated fixture alone, 4.0 GiB.
+  memory of `cargo test --release -p clawhdf5 --features lz4,zstd --test
+  huge_chunks_interop -- --test-threads=1` with `CLAWHDF5_HUGE_CHUNKS=1`
+  (all twelve tests, h5py and h5dump included) was 10.04 GiB (tank,
+  2026-09-29, commit `0065c6b`), in `writer_huge_chunks_lz4_zstd_round_trip`
+  (the other tests, 8.2 GiB or less; the first run of the suite, six
+  tests, 8.06 GiB on 2026-09-28). Which step of that test holds 10 GiB
+  (clawhdf5's LZ4 or Zstd write or read, or hdf5plugin's) was not
+  measured separately.
 - **32-bit targets (wasm32):** such a chunk cannot be held in memory:
   reading or writing one is `FormatError::Overflow` ("exceeds the
   addressable size" / "address space"); the file still opens and lists
@@ -315,15 +329,12 @@ wrong data.
   decoded, and external references are an error (object references
   decode). A multi-dimensional numeric attribute is returned as a flat
   array (its shape is not reported; `AttrValue::Raw` carries the shape).
-  HDF5 2.0's native complex type (class 11) is read as the equivalent
-  `{r, i}` compound: values are right (`Dataset::read_complex_f64`, and
-  numpy complex in Python), but `raw_datatype()`, `dtype()`, `h5rs
-  ls`/`dump` and the browser reader show a compound where h5dump 2.x
-  prints `H5T_COMPLEX_IEEE_F64LE`, so `h5rs dump` of such a file does not
-  match h5dump 2.x (h5dump 1.14 cannot read it at all). Writing class 11
-  (added 2026-09-28) is opt-in (`with_native_complex_f64_data`,
-  `make_native_complex_f64_type`); the Python bindings write complex
-  arrays as h5py's compound only.
+  HDF5 2.0's native complex type (class 11) is read as its own type since
+  2026-09-29 ([fixed](#hdf5-20-native-complex-numbers-read-as-a-r-i-compound));
+  writing it is opt-in (`with_native_complex_f64_data`,
+  `make_native_complex_f64_type`), and the Python bindings write complex
+  arrays as h5py's compound only. `h5rs dump --json` writes it as the
+  `{r, i}` compound: hdf5-json (h5json 2.0.0) has no complex class.
 - **Metadata cache images** (read since 2026-09-26) differ from libhdf5 in
   that: libhdf5 fails only the first metadata read of an image it cannot
   load and then reads the file's own (possibly stale) metadata, where we
@@ -533,8 +544,10 @@ browser (`clawhdf5-wasm`'s `openUrl`) open URLs since 2026-09-27.
     again retries it; what was fetched stays cached).
   - Tested under Node 22 and headless Chromium (Playwright's build) against
     a local server, cross-origin included; not in Firefox or Safari.
-- Compound, reference, opaque, bitfield, time and VL-sequence datasets are
-  refused with an error naming the type; attributes of those types come back
+- Compound (h5py's complex numbers, a compound `{r, i}`, included),
+  reference, opaque, bitfield, time and VL-sequence datasets are
+  refused with an error naming the type (HDF5 2.0 native complex datasets
+  read, as `[re, im]` pairs); attributes of those types come back
   as `value: null` with their `dtype`. (VL strings read, with h5py's
   values, through the same `VlResolver` as `File` and `h5rs`.)
 - No Zstd or SZIP (both link C): such datasets fail with
@@ -616,7 +629,7 @@ given.
 
 ## NetCDF-4: files without dimension scales, types and order differed from netCDF-C
 
-**Status:** fixed 2026-09-29 (branch `fix/netcdf-phony-dims`). Affected
+**Status:** fixed 2026-09-29 (#30). Affected
 every release (v2.1.0 to v2.7.0) and `main` after #27. Wrong metadata only
 (names, order and types of dimensions and variables); values were read
 right. Users: the dimensions of a file without dimension scales are now
@@ -669,6 +682,81 @@ netCDF4-python-written files with netCDF-C itself, and
 `corpus_vs_netcdf_c` the corpus (420 of 429 match; the rest are in
 [NetCDF-4: differences from netCDF-C](#netcdf-4-differences-from-netcdf-c)).
 
+## HDF5 2.0 native complex numbers read as a `{r, i}` compound
+
+**Status:** fixed 2026-09-29 (#31). Affected v2.2.0 to v2.7.0 (class 11 has parsed since v2.2.0)
+and `main` until then. Listed until now under
+[HDF5 features still unsupported](#hdf5-features-still-unsupported).
+
+A class-11 datatype (`H5T_COMPLEX_IEEE_F64LE`, ...) parsed into the
+equivalent compound `{r, i}`. Values were right (`read_complex_f64`,
+numpy complex in Python), but `raw_datatype()`, `dtype()`, `h5rs ls`/`dump`
+and the browser reader showed a compound, so `h5rs dump` did not match
+h5dump 2.x, the browser refused the dataset, and a parsed type written
+back out became a compound. `Datatype::parse` now returns
+`Datatype::Complex` (in compounds, arrays and VL types too),
+`Dataset::dtype()` is `DType::Complex(..)`, `h5rs` prints what h5dump,
+h5ls and h5diff 2.2.0 print (checked on tank, 2026-09-29, against a file
+written by h5py 3.16 / libhdf5 2.0.0:
+`cargo test -p clawhdf5-tools --test native_complex_dump`), and the
+browser reads `[re, im]` pairs.
+
+**What users must do:** code that matched a native complex type as a
+`Datatype::Compound` (or a `DType::Compound` of `r`/`i`) must match
+`Datatype::Complex` / `DType::Complex`, or take the compound view with
+`Datatype::complex_as_compound`; exhaustive `match`es over `DType` need a
+`Complex` arm. Files are unchanged; nothing needs rewriting.
+
+## Chunk dimensions of 2^32 or more were refused
+
+**Status:** fixed 2026-09-29 (#32); affected every release (v2.1.0 to v2.7.0). An error, never wrong
+data. Nothing for users to do but upgrade; code that matches on
+`DataLayout::Chunked::chunk_dimensions` gets `u64`s now.
+
+libhdf5 2.x writes a chunk larger than 4 GiB with layout message version
+5, whose dimensions take up to 8 bytes each, so a chunk dimension can be
+2^32 or more (a 4 GiB chunk of 1-byte elements needs one).
+`DataLayout::Chunked::chunk_dimensions` was `Vec<u32>`: such a file was
+refused when opened for reading (`InvalidChunkDimensions`, "larger than
+2^32 - 1"), and the writer refused such a chunk. Chunk dimensions are
+`u64` throughout (format crate, facade, editor, `h5rs`, Python bindings);
+a version-3 layout, whose dimensions are 4 bytes, is never written for
+them (a chunk that large always takes version 5, as in libhdf5). Tests:
+`huge_chunks_interop` (`huge_chunk_dims_list` always, over
+`fixtures/huge_chunk_dims.h5`, which libhdf5 2.0.0 wrote: `u8` chunks of
+2^32 + 7 in a Single Chunk and an Extensible Array index; with
+`CLAWHDF5_HUGE_CHUNKS=1`, reads of it and of an unfiltered sparse file
+h5py writes, and clawhdf5's own such chunks read back by h5py 3.16 and
+h5dump 2.2.0), unit tests of the 5-byte encoding, and the wasm package
+test (the file lists; reading such a chunk on wasm32 is
+`FormatError::Overflow`).
+
+## Writing a 4 GiB unfiltered chunk held several copies of it
+
+**Status:** fixed 2026-09-29 (#32); affected every release (v2.1.0 to v2.7.0). Memory only. To write
+large data with the least memory, hand the builder the data
+(`DatasetBuilder::with_u8_data_owned`) and write with
+`FileBuilder::write` (or `FileWriter::finish_with`).
+
+`FileWriter::finish` copied every chunk into a per-chunk buffer, laid the
+chunks out into one buffer per pass (two passes, the first kept), then
+copied everything into the file's buffer, and `FileBuilder::write` wrote
+that buffer out: about five times a dataset's size for one unfiltered
+chunk, six with `with_u8_data` (which copies its argument). Measured
+before and after on tank, 2026-09-29, peak resident memory of
+`write_one_chunk 1073741824 OUT <mode>` (`crates/clawhdf5/examples`, one
+unfiltered 1 GiB chunk of `u8`): `owned` 5.0 GiB before (the writer at
+`4260af4`), 1.0 GiB after (commit `0065c6b`); `slice` 6.0 GiB before,
+2.0 GiB after. With 4 GiB + 7 bytes and `owned`: 4.00 GiB after. The
+files written are byte for byte the same. Now a chunk that is a contiguous
+run of the dataset's data is borrowed (and a filtered one compressed
+straight from it), the layout refers to chunks instead of copying them,
+and `FileBuilder::write` streams the file to disk
+(`FileWriter::finish_with`); contiguous datasets are not copied either.
+Test: `writer_unfiltered_huge_chunk_round_trip` (with
+`CLAWHDF5_HUGE_CHUNKS=1`: 4 GiB + 7 bytes written and read back by
+clawhdf5, h5py 3.16 and h5dump 2.2.0; the test process peaked at 4.00 GiB).
+
 ## NetCDF-4: variables' dimensions are guessed from sizes
 
 

examples/wasm-viewer/README.md

@@ -123,8 +123,12 @@ else throws an `Error` naming the type.
   `readHyperslab`). Files of 4 GiB or more are refused at open (wasm32).
   Every response body is cut off past the length asked for. More in
   `docs/known-issues.md`.
-- Compound, reference, opaque and variable-length-sequence datasets are
-  refused with an error. Attributes of those types are listed with
+- HDF5 2.0 native complex datasets read as `[re, im]` pairs: `dtype`
+  `complex<f64>`, `elementShape` ending in `2`, the parts interleaved in
+  the typed array.
+- Compound (h5py's complex numbers, a compound `{r, i}`, included),
+  reference, opaque and variable-length-sequence datasets are refused with
+  an error. Attributes of those types are listed with
   `value: null` and their `dtype`.
 - No Zstd or SZIP filters (they link C): such a dataset fails with
   `unsupported filter`. Deflate, shuffle, Fletcher-32, LZ4, N-Bit and

examples/wasm-viewer/test/browser.sh

@@ -93,6 +93,12 @@ expect $H5 "/vlen_str" '<td>"двa"</td>' '<dd>vlen string</dd>'
 expect $H5 "/pairs" '<td>[2, 3]</td>' '<dd>array[2]&lt;i32&gt;</dd>'
 # 3-D: leading dimension held at 0, window over the last two.
 expect $H5 "/cube" '<dd>(2, 5, 6)</dd>' '<td>29</td>' 'dim 0'
+# HDF5 2.0 native complex (written when h5py's libhdf5 is 2.0 or later):
+# each cell is the [re, im] pair.
+if "$PY" -c 'import sys, h5py; sys.exit(not getattr(h5py.get_config(), "has_native_complex", False))'; then
+    expect $H5 "/native_c128" '<dd>complex&lt;f64&gt;</dd>' '<dd>(3, 2)</dd>' '<td>[1, -1.5]</td>' \
+        '<td>[5, -7.5]</td>'
+fi
 # Unsupported type: an error, not values.
 expect $H5 "/table" 'class="error"' 'reading compound{x: f64, n: i32} datasets is not supported'
 # Cross-origin: the page is on 127.0.0.1, the file on localhost. CORS that

examples/wasm-viewer/test/make_fixture.py

@@ -80,6 +80,15 @@ with h5py.File(h5, "w") as f:
                          **hdf5plugin.Zstd())
     comp = np.zeros(2, dtype=[("x", "<f8"), ("n", "<i4")])
     f.create_dataset("table", data=comp)
+    if getattr(h5py.get_config(), "has_native_complex", False):
+        # HDF5 2.0 native complex (class 11): read as [re, im] pairs.
+        from h5py import h5d, h5s, h5t
+
+        for name, t, dt in [(b"native_c128", h5t.COMPLEX_IEEE_F64LE, "<c16"),
+                            (b"native_c64_be", h5t.COMPLEX_IEEE_F32BE, ">c8")]:
+            z = (np.arange(6) - 1.5j * np.arange(6)).reshape(3, 2).astype(dt)
+            d = h5d.create(f.id, name, t, h5s.create_simple(z.shape))
+            d.write(h5s.ALL, h5s.ALL, z, mtype=t)
     g = f.create_group("sensors")
     g.attrs["location"] = "lab"
     g.create_dataset("temp", data=np.array([21.5, 22.0, 22.25], dtype="<f4"))
@@ -105,6 +114,8 @@ def kind(dt):
         return "strings"
     if dt.subdtype is not None:
         return kind(dt.subdtype[0])
+    if dt.kind == "c":
+        return "f64" if dt.itemsize == 16 else "f32"
     if dt.kind == "f":
         return "f64" if dt.itemsize == 8 else "f32"
     if dt.kind in "iu":
@@ -120,20 +131,25 @@ def flat(a, k):
         return [x.decode() if isinstance(x, bytes) else str(x) for x in a.ravel()]
     if k.startswith(("i", "u")):
         return [str(int(x)) for x in a.ravel()]
+    if a.dtype.kind == "c":
+        # [re, im] pairs, in order.
+        a = np.stack([a.real, a.imag], axis=-1)
     return [float(x) for x in a.ravel()]
 
 
 def entry(ds, slab=None):
     k = kind(ds.dtype)
     data = ds[()]
-    e = {"kind": k, "shape": list(np.shape(data)), "values": flat(data, k)}
+    # A complex element reads as its two parts: one more dimension.
+    pair = [2] if ds.dtype.kind == "c" else []
+    e = {"kind": k, "shape": list(np.shape(data)) + pair, "values": flat(data, k)}
     if slab:
         start, count, stride = slab
         idx = tuple(slice(s, s + (c - 1) * st + 1, st)
                     for s, c, st in zip(start, count, stride))
         part = ds[idx]
         e["slab"] = {"start": start, "count": count, "stride": stride,
-                     "shape": list(part.shape), "values": flat(part, k)}
+                     "shape": list(part.shape) + pair, "values": flat(part, k)}
     return e
 
 
@@ -176,7 +192,10 @@ def describe(path):
                                           "datasets": sorted(sets)}
                 for n, o in members.items():
                     walk(key.rstrip("/") + "/" + n, o)
-            elif obj.dtype.names:
+            elif obj.dtype.names or (
+                    obj.dtype.kind == "c"
+                    and obj.id.get_type().get_class() != getattr(h5py.h5t, "COMPLEX", None)):
+                # h5py's own complex numbers are a compound {r, i}: refused.
                 expected["errors"][key] = "compound"
             else:
                 expected["datasets"][key] = entry(obj, slab_for(obj))

examples/wasm-viewer/test/test.mjs

@@ -399,7 +399,7 @@ async function remoteTests() {
   await fails(async () => {
     const f = await pkg.openUrl(`${base}/fix/fixture.h5`, {
       blockSize: 512,
-      fetch: tamper(async (r) => withHeaders(r, { "Content-Range": "bytes 0-511/25752" })),
+      fetch: tamper(async (r) => withHeaders(r, { "Content-Range": `bytes 0-511/${statSync(join(fixDir, "fixture.h5")).size}` })),
     });
     await f.read("/grid");
   }, /the server sent 0-511/, "wrong range");
@@ -514,6 +514,15 @@ async function limitTests() {
       `huge chunk: ${name}`);
   }
   hc.free();
+  // Likewise chunks whose dimension is 2^32 or more (u8 chunks of 2^32 + 7).
+  const dims = pkg.open(new Uint8Array(readFileSync(join(import.meta.dirname, "..", "..", "..",
+    "crates/clawhdf5/tests/fixtures/huge_chunk_dims.h5"))));
+  eq(dims.list("/").map((e) => e.name), ["earray", "single"], "huge chunk dims: list");
+  for (const name of ["single", "earray"]) {
+    await fails(() => dims.readHyperslab(`/${name}`, [0], [4]), /exceeds this platform.s address space/,
+      `huge chunk dims: ${name}`);
+  }
+  dims.free();
 }
 
 // A body of `total` bytes in 64 KiB pieces, made as they are read; `pulled()`
@@ -550,7 +559,7 @@ async function floodTests() {
       const range = new Headers(init.headers).get("Range");
       if (range === "bytes=0-511") return fetch(url, init);
       const m = /^bytes=(\d+)-(\d+)$/.exec(range);
-      return new Response(f.body, { status: 206, headers: { "Content-Range": `bytes ${m[1]}-${m[2]}/25752` } });
+      return new Response(f.body, { status: 206, headers: { "Content-Range": `bytes ${m[1]}-${m[2]}/${statSync(join(fixDir, "fixture.h5")).size}` } });
     },
   }).catch((e) => e);
   // The open itself may need a second range: flooded either way.