HDF5

Usage

use HDF5;

or

import HDF5;

Submodules

• C_HDF5
• IOusingMPI

HDF5 bindings for Chapel

This module implements the C-API for HDF5 version 1.10.1, as well as some functionality for reading and writing HDF5 files built on top of the C-API.

Compiling with HDF5

In order to compile a Chapel program that uses this module, the HDF5 library must be installed on the system. The paths to the hdf5_hl.h header file and HDF5 library must be known to the compiler, either by finding them in a default search path, or by using the -I and -L compiler arguments.

The compilation command should look something like this:

chpl -I$PATH_TO_HDF5_DIR \
     -L$PATH_TO_HDF5_LIBS -lhdf5 source.chpl

On Cray systems with the cray-hdf5 module loaded, no compiler flags are necessary to use the HDF5 module.

This module currently requires using a version of the Chapel compiler that is linked with LLVM. See LLVM Support.

Chapel HDF5 API

This module provides some higher-level functions using the HDF5 API. It provides for parallel reads and writes to HDF5 files using multiple Chapel locales.

The native HDF5 functions can be called directly by calling into the C_HDF5 submodule.

proc readAllHDF5Files(locs: [] locale, dirName: string, dsetName: string, filenameStart: string, type eltType, param rank, preprocessor: borrowed nilable HDF5Preprocessor =)

Read the dataset named dsetName from all HDF5 files in the directory dirName with filenames that begin with filenameStart. This will read the files in parallel with one task per locale in the locs array. Specifying the same locale multiple times in the locs array will cause undefined behavior.

Returns a distributed array of ArrayWrapper records containing the arrays that are read. Each instance will reside on the locale where the corresponding data was read.

proc readAllNamedHDF5Files(locs: [] locale, filenames: [] string, dsetName: string, type eltType, param rank, preprocessor: borrowed nilable HDF5Preprocessor =)

Read all HDF5 files named in the filenames array into arrays

proc readNamedHDF5FilesInto1DArrayInt(filenames: [] string, fnCols: int, fnRows: int, dsetName: string, preprocessor: borrowed nilable HDF5Preprocessor =)

Read the datasets named dsetName from the files named in filenames into a 1D array. This function assumes the data elements are all int(64)s. fnCols and fnRows refer to the columns and rows that filenames should have when converting it back to 2-D.

proc readHDF5Dataset(file_id, dsetName: string, data)

Read the dataset named dsetName from the open file that file_id refers to. Store the dataset into the array data. Can read data of type int/uint (size 8, 16, 32, 64), real (size 32, 64), and c_string.

proc getHDF5Type(type eltType)

Return the HDF5 type equivalent to the Chapel type eltType

enum Hdf5OpenMode { Truncate, Append }

Enum indicating the way to open a file for writing. Truncate will clear the contents of an existing file. Append will add to a file that already exists. Both will open a new empty file.

proc writeArraysToHDF5Files(dirName: string, dsetNames: [] string, filenames: [] string, type eltType, param rank: int, data: [] eltTypeArrayWrapperrank, mode: Hdf5OpenMode) throws

Write the arrays from the ArrayWrapper records stored in the data argument to the corresponding filename in the filenames array. The dataset name is taken from the corresponding position in the dsetNames array.

The data argument should be a Block distributed array with dataParTasksPerLocale==1.

Either truncate or append to the file depending on the mode argument.

It would be preferable to find the eltType and rank values directly from the data argument instead of needing explicit arguments for them, but it isn’t obvious how to do that currently. If the generic fields in the ArrayWrapper could be queried that would be a nice replacement for these arguments. e.g. data: [] ArrayWrapper(?eltType, ?rank).

iter hdf5ReadChunks(filename: string, dset: string, chunkShape: domain, type eltType, preprocessor: borrowed nilable HDF5Preprocessor =)

Read the dataset named dset from the HDF5 file named filename. The dataset consists of elements of type eltType. The file will be read in chunks matching the chunkShape domain, and yielded as arrays of that size until the end of the dataset is reached.

This allows operating on a very large dataset in smaller sections, for example when it is too big to fit into the system memory, or to allow each section to fit within cache.

Currently, the chunkShape domain describing the yielded array shape and the shape of the data in the file must both have the same rank. For example, if the data in the file is 2D, chunkShape must be two-dimensional as well. It is expected that this restriction can be relaxed in the future.

class HDF5Preprocessor

A class to preprocess arrays returned by HDF5 file reading procedures. Procedures in this module that take an HDF5Preprocessor argument can accept a subclass of this class with the preprocess method overridden to do preprocessing as desired before returning the data read.

proc preprocess(A: [])

record ArrayWrapper

A record that stores a rectangular array. An array of ArrayWrapper records can store multiple differently sized arrays.

type eltType

param rank: int

var D: domain(rank)

var A: [D] eltType