HDF5 for Rust.
The hdf5-metno
crate (also known under the pre-fork name of hdf5
) provides thread-safe Rust bindings and
high-level wrappers for the HDF5 library API. Some of the features include:
- Thread-safety with non-threadsafe libhdf5 builds guaranteed via reentrant mutexes.
- Native representation of most HDF5 types, including variable-length strings and arrays.
- Derive-macro for automatic mapping of user structs and enums to HDF5 types.
- Multi-dimensional array reading/writing interface via
ndarray
.
Direct low-level bindings are also available and are provided in the hdf5-sys
crate.
Requires HDF5 library of version 1.8.4 or later.
This is a fork of https://github.com/aldanor/hdf5-rust used for publishing newer versions of the crate. This includes a change in the crate name to allow publishing on crates.io. For backwards compatibility consider using the following in your Cargo.toml
:
hdf5 = { package = "hdf5-metno", version = "0.9.0" }
#[cfg(feature = "blosc")]
use hdf5::filters::blosc_set_nthreads;
use hdf5::{File, H5Type, Result};
use ndarray::{arr2, s};
#[derive(H5Type, Clone, PartialEq, Debug)] // register with HDF5
#[repr(u8)]
pub enum Color {
R = 1,
G = 2,
B = 3,
}
#[derive(H5Type, Clone, PartialEq, Debug)] // register with HDF5
#[repr(C)]
pub struct Pixel {
xy: (i64, i64),
color: Color,
}
impl Pixel {
pub fn new(x: i64, y: i64, color: Color) -> Self {
Self { xy: (x, y), color }
}
}
fn write_hdf5() -> Result<()> {
use Color::*;
let file = File::create("pixels.h5")?; // open for writing
let group = file.create_group("dir")?; // create a group
#[cfg(feature = "blosc")]
blosc_set_nthreads(2); // set number of blosc threads
let builder = group.new_dataset_builder();
#[cfg(feature = "blosc")]
let builder = builder.blosc_zstd(9, true); // zstd + shuffle
let ds = builder
.with_data(&arr2(&[
// write a 2-D array of data
[Pixel::new(1, 2, R), Pixel::new(2, 3, B)],
[Pixel::new(3, 4, G), Pixel::new(4, 5, R)],
[Pixel::new(5, 6, B), Pixel::new(6, 7, G)],
]))
// finalize and write the dataset
.create("pixels")?;
// create an attr with fixed shape but don't write the data
let attr = ds.new_attr::<Color>().shape([3]).create("colors")?;
// write the attr data
attr.write(&[R, G, B])?;
Ok(())
}
fn read_hdf5() -> Result<()> {
use Color::*;
let file = File::open("pixels.h5")?; // open for reading
let ds = file.dataset("dir/pixels")?; // open the dataset
assert_eq!(
// read a slice of the 2-D dataset and verify it
ds.read_slice::<Pixel, _, _>(s![1.., ..])?,
arr2(&[
[Pixel::new(3, 4, G), Pixel::new(4, 5, R)],
[Pixel::new(5, 6, B), Pixel::new(6, 7, G)],
])
);
let attr = ds.attr("colors")?; // open the attribute
assert_eq!(attr.read_1d::<Color>()?.as_slice().unwrap(), &[R, G, B]);
Ok(())
}
fn main() -> Result<()> {
write_hdf5()?;
read_hdf5()?;
Ok(())
}
hdf5
crate is known to run on these platforms: Linux, macOS, Windows (tested on:
Ubuntu 16.04, 18.04, and 20.04; Windows Server 2019 with both MSVC and GNU
toolchains; macOS Catalina).
hdf5
crate is tested continuously for all three official release channels, and
requires a reasonably recent Rust compiler (e.g. of version 1.80 or newer).
Required HDF5 version is 1.8.4 or newer. The library doesn't have to be built with threadsafe option enabled in order to make the user code threadsafe.
Various HDF5 installation options are supported and tested: via package managers like homebrew and apt; system-wide installations on Windows; conda installations from both the official channels and conda-forge. On Linux and macOS, both OpenMPI and MPICH parallel builds are supported and tested.
The HDF5 C library can also be built from source and linked in statically by
enabling hdf5-sys/static
feature (CMake required).
Build scripts for both hdf5-sys
and hdf5
crates check the actual version of the
HDF5 library that they are being linked against, and some functionality may be conditionally
enabled or disabled at compile time. While this allows supporting multiple versions of HDF5
in a single codebase, this is something the library user should be aware of in case they
choose to use the low level FFI bindings.
If HDF5_DIR
is set, the build script will look there (and nowhere else) for HDF5
headers and binaries (i.e., it will look for headers under $HDF5_DIR/include
).
If HDF5_VERSION
is set, the build script will check that the library version matches
the specified version string; in some cases it may also be used by the build script to
help locating the library (e.g. when both 1.8 and 1.10 are installed via Homebrew on macOS).
It is possible to link against hdf5
conda package; a few notes and tips:
- Point
HDF5_DIR
to conda environment root. - The build script knows about conda environment layout specifics and will adjust
paths accordingly (e.g.
Library
subfolder in Windows environments). - On Windows, environment's
bin
folder must be inPATH
(or the environment can be activated prior to running cargo). - On Linux / macOS, it is recommended to set rpath, e.g. by setting
RUSTFLAGS="-C link-args=-Wl,-rpath,$HDF5_DIR/lib"
. - For old versions of HDF5 conda packages on macOS, it may also be necessary to set
DYLD_FALLBACK_LIBRARY_PATH="$HDF5_DIR/lib"
.
The build script will attempt to use pkg-config first, which will likely work out without further tweaking for the more recent versions of HDF5. The build script will then also look in some standard locations where HDF5 can be found after being apt-installed on Ubuntu.
On macOS, the build script will attempt to locate HDF5 via Homebrew if it's available.
If both 1.8 and 1.10 are installed and available, the default (1.10) will be used
unless HDF5_VERSION
is set.
hdf5
crate fully supports MSVC toolchain, which allows using the
official releases of
HDF5 and is generally the recommended way to go. That being said, previous experiments have
shown that all tests pass on the gnu
target as well, one just needs to be careful with
building the HDF5 binary itself and configuring the build environment.
Few things to note when building on Windows:
hdf5.dll
should be available in the search path at build time and runtime (bothgnu
andmsvc
). This normally requires adding thebin
folder of HDF5 installation toPATH
. If using an official HDF5 release (msvc
only), this will typically be done automatically by the installer.msvc
: installed Visual Studio version should match the HDF5 binary (2013 or 2015). Note that it is not necessary to runvcvars
scripts; Rust build system will take care of that.- When building for either target, make sure that there are no conflicts in the search path (e.g., some binaries from MinGW toolchain may shadow MSVS executables or vice versa).
- The recommended platform for
gnu
target is TDM distribution of MinGW-GCC as it contains bintools for both 32-bit and 64-bit. - The recommended setup for
msvc
target is VS2015 x64 since that matches CI build configuration, however VS2013 and x86 should work equally well.
hdf5
crate is primarily distributed under the terms of both the MIT license and the
Apache License (Version 2.0). See LICENSE-APACHE and
LICENSE-MIT for details.