B³D compression library

B³D is a fast (~1 GB/s), GPU based image compression method, developed for fluorescence microscopy applications. Alongside lossless compression, it offers a noise dependent lossy compression mode, where the loss can be tuned as a proportion of the inherent image noise (accounting for photon shot noise and camera read noise). It not only allows for fast compression during imaging, but can achieve compression ratios up to 100 fold.

Corresponding publication: http://www.biorxiv.org/content/early/2017/07/21/164624

Requirements

Microsoft Windows 7 or newer, 64-bit
HDF5 1.8.15 64-bit or later installed. Obtain the latest version here: https://support.hdfgroup.org/HDF5/release/obtain518.html
Microsoft Visual C++ 2013 Redistributable (x64) (installer included in vcredist folder)
CUDA capable graphics card, minimum Compute Capability 3.0
CUDA 8 compatible graphics driver (for GeForce products 378.66 or later)

HDF5 1.10 should also be compatible, but was not tested.

Installation

All necessary binary files can be found in the win64 folder. In order for HDF5 to recognize the filter, the path of this folder has to be stored in the %HDF5_PLUGIN_PATH% environment variable. This can be done automatically by running the install_b3d.bat batch file. It will set the environment variable for the current user, or if it already exists (e.g. other filters are already in use), it will copy the necessary files there.

Reading compressed files

If all the requiriements are present, and installation was successfull, any HDF5 application should be able to automatically use the filter. This functionality can be tested by opening the provided example file B3D_demo_drosophila_compressed.h5 in HDFView for example.

Reading files in Fiji

Fiji comes with its own distribution of HDF5 in the form of jhdf5-14.12.5.jar in the Fiji.app/jars folder. On Windows machines this is not compatible with the filter because a different version of Visual Studio was used to compile the native libraries. To circumvent this issue, we provide a recompiled version of the same file, included in the win64 folder. Copy this file to the Fiji.app/jars folder, overwriting the original if necessary.

To open HDF5 files, you need to enable the HDF5 update site (Help > Update, then click on "Manage update sites" and select "HDF5" from the list). During the update process you might get a warning message: "There are locally modified files". As this is a consequence of the previous step, it's important not to let Fiji overwrite the locally modified files.

Writing compressed files

To write compressed files, the compression has to be set by calling H5Pset_filter on the dataset creation property list (dcpl_id).

H5Pset_filter(dcpl_id, 32016, H5Z_FLAG_OPTIONAL, cd_nelmts, cd_values);

where 32016 is the filter id for B³D compression, and cd_nelmts is the number of elements in cd_values. cd_values is an array comprising of the following elements:

cd_values[0] Quantization step*. Relative to sigma (see paper for details; default = 0). Set to 0 for lossless compression.
cd_values[1] Compression mode. 1: Mode1, 2: Mode2 (see paper for details; default = 1)
cd_values[2] Camera conversion parameter* in DN/e- (default=1000)
cd_values[3] Camera background offset parameter in DN (default = 0)
cd_values[4] Camera read noise parameter* in e- (default = 0)
cd_values[5] Tile size, to optimize parallel execution (default=24)

* Since cd_values can only be integers, for floating point parameters the value*1000 rounded to integers has to be stored in cd_values. For lossless compression all elements can be omitted.

For more information on how to use HDF5 with compression, see the official HDF5 website.

Performance considerations

Filters can only be applied to datasets if chunking is also enabled. For optimal performance, we recommend choosing a chunk size ~2 MB (1M elements with 16 bit data).

Supported data types

Datasets of 2 or 3 dimensions
signed or unsigned 8 bit or 16 bit integers

Usage examples

Example scripts are located in the sample_scripts folder. Currently this includes a Python example (tested with Python 2.7.10 and 3.5.3)

Contact

For any questions / comments about this software, please send an email to Balint Balazs.

B³D Copyright and Software License

B³D source code, binaries and examples are licensed under the MIT license (see LICENSE file) with the exception of the subfolders source/src/B3D_cudaCompress/reduce and source/src/cudaCompress/scan, which contain code adapted from CUDPP (http://cudpp.github.io) and come with a separate license (see file license.txt in the subfolders).

B³D also includes a modified version of cudaCompress (https://github.com/m0bl0/cudaCompress), also licensed under the MIT license.

The binary distribution in the win64 folder also includes the redistributable CUDA runtime library, which comes with it's own license. For details see the LICENSE file and CUDA_EULA.pdf in that folder.

ome/B3D