h5pp¶
h5pp
is a C++17 wrapper for HDF5 with focus on simplicity.
In just a few lines of code, h5pp
lets users read and write to HDF5 files, a portable binary format.
h5pp
supports common data types and common containers, such as std::vector.
In particular, h5pp
makes it easy to read and write Eigen matrices and tensors.
Introduction¶
HDF5 is a popular format for portable binary storage of large datasets. With bindings to languages such as Python, Julia, Matlab and many others, it is straightforward to export, import and analyze data in a collaborative setting.
In C/C++ using HDF5 directly is not straightforward. Beginners are met with a steep learning curve to the vast API of HDF5. There are many C/C++ libraries already that simplify the user experience, but as a matter of opinion, things could be even simpler.
The goal of h5pp
is to make HDF5 simple to use in the following sense:
Users should be able to read/write common C++ data-types in a single line of code.
Users should not need prior knowledge of HDF5 for simple tasks.
Sensible defaults should let simple tasks stay simple, e.g., specifying storage layout, chunk dimensions or compression.
Advanced tasks should stay possible, e.g. MPI parallelism.
Logs and error messages should be meaningful to beginners.
Installation should be simple and scalable with opt-in automation.
Features¶
Header-only C++17 template library
High-level front-end to the C API of HDF5
Support for common data types:
short
,int
,long
,long long
(+ unsigned versions),float
,double
,long double
any of the above in C-style arrays
any of the above in
std::complex<>
formany of the above in POD-structs with x,y or x,y,z data members. In
h5pp
these go by the nameScalar2
andScalar3
. These work well together with types such asdouble2
orfloat3
found in CUDA.
std::string
andchar
arrays.Any container such as std::vector with
.data()
member for accessing a contiguous buffer (without conversion to/from row major).Eigen
types such asMatrix
,Array
andTensor
, with automatic conversion to/from row major storage layout.Support for user-defined compound HDF5 types
Support for creating HDF5 tables from user-defined compound HDF5 types.
Modern CMake installation providing targets for simple linking to your projects.
Installation with package managers: conan, conda (and apt using .deb installation file)
(Opt-in) Automatically find or download dependencies using either the Conan package manager or “CMake-only” methods.
Multi-platform: Linux, Windows, OSX. (Developed under Linux)
Usage¶
Using h5pp
is intended to be simple. After initializing a file,
most of the work can be achieved using just two member functions .writeDataset(...)
and .readDataset(...)
.
Example: Writing an std::vector
¶
#include <h5pp/h5pp.h>
int main() {
// Initialize a file
h5pp::File file("myDir/someFile.h5");
// Initialize a vector with 10 doubles
std::vector<double> v (10, 3.14);
// Write the vector to file.
// Inside the file, the data will be stored in a dataset named "myStdVector"
file.writeDataset(v, "myStdVector");
return 0;
}
Find more code examples in the examples directory or in the Wiki.
File permissions¶
h5pp
offers more flags for file access permissions than HDF5. The new flags are primarily intended to
prevent accidental loss of data, but also to clarify intent and avoid mutually exclusive options.
The flags are listed in the order of increasing “danger” that they pose to previously existing files.
| Flag | File exists | No file exists | Comment |
| —- | —- | —- | —- |
| READONLY
| Open with read-only permission | Throw error | Never writes to disk, fails if the file is not found |
| COLLISION_FAIL
| Throw error | Create new file | Never deletes existing files and fails if it already exists |
| RENAME
(default) | Create renamed file | Create new file | Never deletes existing files. Invents a new filename to avoid collision by appending “-#” (#=1,2,3…) to the stem of the filename |
| READWRITE
| Open with read-write permission | Create new file | Never deletes existing files, but is allowed to open/modify |
| BACKUP
| Rename existing file and create new | Create new file | Avoids collision by backing up the existing file, appending “.bak_#” (#=1,2,3…) to the filename |
| REPLACE
| Truncate (overwrite) | Create new file | Deletes the existing file and creates a new one in place |
When a new file is created, the intermediate directories are always created automatically.
When a new file is created,
READWRITE
permission to it is implied.
To give a concrete example, the syntax works as follows
h5pp::File file("myDir/someFile.h5", h5pp::FilePermission::REPLACE);
Storage Layout¶
Unless specified, h5pp
will automatically decide the best storage layout for each dataset. The possible layouts are
H5D_COMPACT
: For scalar or small datasets which can fit in the metadata header. Default on datasets smaller than 32 KB.H5D_CONTIGUOUS
: For medium size datasets. Default on datasets smaller than 512 KB.H5D_CHUNKED
: For large datasets. Default on datasets larger than 512 KB. This layout has some additional features:Chunking, portioning of the data to improve IO performance by caching more efficiently. Chunk dimensions are calculated by
h5pp
if not given specifically.Compression, disabled by default, and only available if HDF5 was built with zlib enabled.
Overwrite with different size (note that the file size never decreases, for instance after overwriting with a smaller dataset).
To specify the layout, pass it as a third argument when writing a new dataset, for instance:
file.writeDataset(myData, "science/myChunkedData", H5D_CHUNKED); // Creates a chunked dataset
Compression¶
Extendable (or chunked) datasets can also be compressed if HDF5 was built with zlib support. Use these functions to set or check the compression level:
file.setCompressionLevel(9); // 0 to 9: 0 to disable compression, 9 for maximum compression.
file.getCompressionLevel(); // Gets the current compression level
h5pp::checkIfCompressionIsAvailable(); // True if your installation of HDF5 has zlib support
or pass a temporary compression level as the fifth argument when writing a dataset:
file.writeDataset(myData, "science/myCompressedData", H5D_CHUNKED, std::nullopt, 8); // Creates a chunked dataset with compression level 8.
Debug and logging¶
Spdlog can be used to emit debugging information efficiently.
The amount of console output (verbosity) can be set to any level between 0
and 5
:
0: trace
(highest verbosity)1: debug
2: info
(default)3: warn
4: error
5: critical
(lowest verbosity)
Set the level when constructing a h5pp::File or by calling the function .setLogLevel(int)
:
int logLevel = 0; // Highest verbosity
// This way...
h5pp::File file("myDir/someFile.h5", h5pp::FilePermission::REPLACE, logLevel);
// or this way
file.setLogLevel(logLevel);
NOTE: Logging works the same with or without Spdlog enabled. When Spdlog is not found, a hand-crafted logger is used in its place to give identical output but without any performance considerations (implemented with STL lists, strings and streams).
Load data into Python¶
HDF5 data is easy to load into Python. Loading integer and floating point data is straightforward. compound data is almost as simple.
HDF5 does not support complex types specifically, but h5pp
enables this through compound HDF5 types. Here is a python example which uses h5py
to load 1D arrays from an HDF5 file generated with h5pp
:
import h5py
import numpy as np
file = h5py.File('myFile.h5', 'r')
# Originally written as std::vector<double> in h5pp
myDoubleArray = np.asarray(file['double-array-dataset'])
# Originally written as std::vector<std::complex<double>> in h5pp
myComplexArray = np.asarray(file['complex-double-array-dataset']).view(dtype=np.complex128)
Notice the cast to dtype=np.complex128
which interprets each element of the array as two doubles
, i.e. the real and imaginary parts are 2 * 64 = 128
bits.
Installation¶
There are currently 4 ways to obtain h5pp
:
git clone https://github.com/DavidAce/h5pp.git
and install (see below)From conda:
conda install -c davidace h5pp
From conan bintray repo
(Debian only) Download the latest release and install with apt:
sudo apt install ./h5pp_<version>_amd64.deb
Requirements¶
C++17 capable compiler. GCC version >= 7 or Clang version >= 7.0
CMake version >= 3.12
HDF5 library, version >= 1.8
Optional dependencies:¶
Eigen: Write Eigen matrices and tensors directly. Tested with version >= 3.3.4
spdlog: Enables logging for debug purposes. Tested with version >= 1.3.1
ghc::filesystem: This drop-in replacement for
std::filesystem
is downloaded and installed automatically when needed, but only ifH5PP_DOWNLOAD_METHOD=<fetch|conan>.
Install methods¶
For full working examples see the directory quickstart
. Find a summary below.
Option 1: Copy the headers¶
Copy the files under h5pp/source/include
and add #include<h5pp/h5pp.h>
.
Make sure to compile with -std=c++17 -lstdc++fs
and link the dependencies hdf5
, Eigen3
and spdlog
. The actual linking
is a non-trivial step, see linking below.
Option 2: Install with CMake¶
Build the library just as any CMake project. For instance, from the project’s root in command-line:
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=<install-dir> ../
make
make install
Headers will be installed under <install-dir>/include
and config files under <install-dir>/share/h5pp/cmake
.
These config files allow you to usefind_package(h5pp)
in your own projects, which in turn defines the target h5pp::h5pp
with everything you need to link h5pp
correctly (including dependencies, if you so choose).
If not set, CMAKE_INSTALL_PREFIX
defaults to ${CMAKE_BINARY_DIR}/install
, where ${CMAKE_BINARY_DIR}
is the directory you are building from.
Option 3: Install with Conan¶
Make sure to install and configure Conan first. Then, either use the cmake-conan integration by passing
-DH5PP_DOWNLOAD_METHOD=conan
as an argument to CMake (see below) or use Conan directly, for instance by running the following command:
$ conan install h5pp/1.8.0@davidace/stable --profile default
This is by far the simplest method and will also make sure to install dependencies HDF5, Eigen3 and spdlog.
Opt-in automatic dependency installation with CMake¶
The CMake flag H5PP_DOWNLOAD_METHOD
controls the automated behavior for finding or installing dependencies. It can take one of three valid strings:
| Option | Description |
| —- | —- |
| none
(default) | No handling of dependencies and linking is left to the user |
| find
| Use CMake’s find_package
to find dependencies pre-installed on your system |
| fetch
(!) | Use CMake-only features to download and install dependencies automatically. Disregards pre-installed dependencies on your system |
| native
| Deprecated. Use fetch
|
| find-or-fetch
| Start with find
and then go to fetch
if not found |
| conan
(!!) | Use the Conan package manager to download and install dependencies automatically. Disregards pre-installed dependencies on your system |
There are several variables you can pass to CMake to guide find_package
calls, see CMake build options below.
(!) Dependencies are installed into CMAKE_INSTALL_PREFIX
. Pass the CMake variable H5PP_DEPS_IN_SUBDIR
to install into separate directories under CMAKE_INSTALL_PREFIX/<libname>
.
(!!) Conan is guided by conanfile.txt
found in this project’s root directory. This method requires conan to be installed prior (for instance through pip
, conda
, apt
, etc). To let CMake find conan you have three options:
Add Conan install (or bin) directory to the environment variable
PATH
.Export Conan install (or bin) directory in the environment variable
CONAN_PREFIX
, i.e. from command line:export CONAN_PREFIX=<path-to-conan>
Give the variable
CONAN_PREFIX
directly to CMake, i.e. from command line:cmake -DCONAN_PREFIX:PATH=<path-to-conan> ...
CMake build options¶
The cmake
step above takes several options, cmake [-DOPTIONS=var] ../
:
| Var | Default | Description |
| —- | —- | —- |
| CMAKE_INSTALL_PREFIX
| ${CMAKE_BINARY_DIR}/install
| Specify h5pp
install directory |
| BUILD_SHARED_LIBS
| OFF
| Link dependencies with static or shared libraries |
| H5PP_ENABLE_TESTS
| OFF
| Build tests (recommended!) |
| H5PP_BUILD_EXAMPLES
| OFF
| Build example programs |
| H5PP_DOWNLOAD_METHOD
| none
| Download method for dependencies, select none
, find
, fetch
, find-or-fetch
or conan
. Fetch
downloads and builds from sources |
| H5PP_PRINT_INFO
| OFF
| Use h5pp with add_subdirectory() |
| H5PP_IS_SUBPROJECT
| OFF
| Print extra CMake info about the host and generated targets during configure |
| H5PP_ENABLE_EIGEN3
| OFF
| Enables Eigen3 linear algebra library support |
| H5PP_ENABLE_SPDLOG
| OFF
| Enables Spdlog support for logging h5pp
internal info to stdout |
| H5PP_DEPS_IN_SUBDIR
| OFF
| Appends <libname>
to install location of dependencies, i.e. CMAKE_INSTALL_PREFIX/<libname>
. This allows simple removal |
| H5PP_PREFER_CONDA_LIBS
| OFF
| Prioritize finding dependencies hdf5
, Eigen3
and spdlog
installed through conda. No effect when H5PP_DOWNLOAD_METHOD=conan
|
The following variables can be set to help guide CMake’s find_package
to your pre-installed software (no defaults):
| Var | Path to |
| —- | —- |
| Eigen3_DIR
| Eigen3Config.cmake |
| Eigen3_ROOT
| Eigen3 install directory |
| EIGEN3_INCLUDE_DIR
| Eigen3 include directory |
| spdlog_DIR
| spdlogConfig.cmake |
| spdlog_ROOT
| Spdlog install directory |
| HDF5_DIR
| HDF5Config.cmake |
| HDF5_ROOT
| HDF5 install directory |
| CONAN_PREFIX
| conan install directory |
Linking¶
Link using CMake targets (easy)¶
h5pp
is easily imported into your project using CMake’s find_package
. Just point it to the h5pp
install directory.
When found, targets are made available to compile and link to dependencies correctly.
A minimal CMakeLists.txt
to use h5pp
would look like:
cmake_minimum_required(VERSION 3.12)
project(myProject)
add_executable(myExecutable main.cpp)
find_package(h5pp PATHS <path-to-h5pp-install-dir> REQUIRED) # If h5pp is installed through conda the path may be $ENV{CONDA_PREFIX}
target_link_libraries(myExecutable PRIVATE h5pp::h5pp)
Targets explained¶
h5pp::h5pp
is the main target including “everything” and should normally be the only target that you need – headers,flags and (if enabled) the found/downloaded dependencies.h5pp::headers
links theh5pp
headers only.h5pp::deps
collects library targets to link all the dependencies that were found/downloaded whenh5pp
was built. These can of course be used independently.If
H5PP_DOWNLOAD_METHOD==find|find-or-fetch|fetch
the targets areEigen3::Eigen
,spdlog::spdlog
andhdf5::hdf5
,If
H5PP_DOWNLOAD_METHOD==conan
the targets areCONAN_PKG::Eigen3
,CONAN_PKG::spdlog
andCONAN_PKG::HDF5
.If
H5PP_DOWNLOAD_METHOD==none
thenh5pp::deps
is empty.
h5pp::flags
sets compile and linker flags to enable C++17 and std::filesystem library, i.e.-std=c++17
and-lstdc++fs
. OnMSVC
it sets/permissive-
to enable logicaland
/or
in C++.
Link manually (not as easy)¶
From the command-line you can of course link using linker flags such as -std=c++17 -lstdc++fs -leigen3 -lspdlog -lhdf5_hl -lhdf5
provided these flags make sense on your system.
You could also use CMake’s find_package(...)
mechanism. A minimal CMakeLists.txt
could be:
cmake_minimum_required(VERSION 3.12)
project(myProject)
add_executable(myExecutable main.cpp)
target_include_directories(myExecutable PRIVATE <path-to-h5pp-headers>)
# Setup h5pp
target_compile_features(myExecutable PRIVATE cxx_std_17)
target_link_libraries(myExecutable PRIVATE stdc++fs)
# Possibly use find_package() here
# Link dependencies (this is the tricky part)
target_include_directories(myExecutable PRIVATE <path-to-Eigen3-include-dir>)
target_include_directories(myExecutable PRIVATE <path-to-spdlog-include-dir>)
target_include_directories(myExecutable PRIVATE <path-to-hdf5-include-dir>)
# Link dependencies (this is the difficult part). Note that you only need the C libs for HDF5.
target_link_libraries(myExecutable PRIVATE hdf5_hl hdf5 rt dl m z pthread) # Possibly more libs, such as aec, dependending on your HDF5 installation
The difficult part is linking to HDF5 libraries and its dependencies.
Use the custom FindHDF5.cmake bundled with h5pp
¶
When installing h5pp
, finding HDF5 and setting up the CMake target hdf5::hdf5
for linking is handled by a custom module for finding HDF5, defined in cmake/FindHDF5.cmake
.
This module wraps the default FindHDF5.cmake
which comes with CMake and uses the same call signature, but fixes some annoyances with naming conventions in different versions of CMake and HDF5 executables.
It reads hints passed through CMake flags to find HDF5 somewhere on your system (e.g. installed via conda
,apt
, brew
, Easybuild
,etc) and defines a CMake target hdf5::hdf5
with everything you need to link correctly.
Most importantly, it avoids injecting shared versions of libraries (dl, zlib, szip, aec) during static builds on older platforms.
You can use the custom module too. Add the path pointing to FindHDF5.cmake
to the variable CMAKE_MODULE_PATH
from within your own project, e.g.:
list(APPEND CMAKE_MODULE_PATH path/to/h5pp/cmake/FindHDF5.cmake)
find_package(HDF5 1.10 COMPONENTS C HL REQUIRED)
if(TARGET hdf5::hdf5)
target_link_libraries(myExecutable PRIVATE hdf5::hdf5)
endif()
These are variables that can be used to guide the custom FindHDF5.cmake
module:
| Var | Where | Description |
| —- | —- | —- |
| CMAKE_MODULE_PATH
| CMake | List of directories where CMake
should search for find-modules |
| CMAKE_PREFIX_PATH
| CMake | List of directories where find_package
should look for dependencies|
| HDF5_ROOT
| CMake/ENV | Path to HDF5 root install directory |
| HDF5_FIND_VERBOSE
| CMake | Prints more information about the search for HDF5. See also HDF5_FIND_DEBUG
in the original module |
| EBROOTHDF5
| ENV | Variable defined by Easybuild with module load HDF5
|
Uninstall¶
The target uninstall
is defined by h5pp
which removes installed headers and dependencies using their respective install manifests.
From the build directory, run the following in the command-line to uninstall:
cmake --build . --target uninstall
To-do¶
In no particular order
Expand documentation. Perhaps a doxygen/sphinx webpage
Expand testing for more edge-cases in
filesystem permissions
user-defined types
tables
Expose more of the C-API:
Support for packed user-defined types. Read more: H5TPack
True support for parallel read/write with MPI
Support row-major <-> col-major transformation for types other than Eigen3 matrices and tensors. For instance, when raw pointers are passed together with dimension initializer list {x,y,z..}. (Although, this can be done by wrapping the data in an Eigen Map object).