GPU+CPU	CPU-only

A stream processing framework for high-throughput applications.

Example pipelines can be found in the testbench/ directory. For example, here's a snippet
that reads data from a binary file, copies it to the GPU, runs an FFT, then writes the output back to disk:

# Get a list of binary data files
filenames   = glob.glob('testdata/*.bin')

# Setup pipeline
b_read      = BinaryFileReadBlock(filenames, window_len, 1, 'cf32', core=0)
b_copy      = CopyBlock(b_read, space='cuda', core=1, gpu=0)
b_fft       = FftBlock(b_copy, axes=1, core=2, gpu=0)
b_out       = CopyBlock(b_fft, space='system', core=3)
b_write     = BinaryFileWriteBlock(b_out, core=4)

# Run pipeline
pipeline = bfp.get_default_pipeline()
print pipeline.dot_graph()
pipeline.run()

And here's an example that reads a WAV audio file, generates a spectrum, and writes the output to a filterbank file:

import bifrost as bf

data = bf.blocks.read_wav(['file1.wav', 'file2.wav'], gulp_nframe=4096)
data = bf.blocks.copy(data, space='cuda')
data = bf.views.split_axis(data, 'time', 256, label='fine_time')
data = bf.blocks.fft(data, axes='fine_time', axis_labels='freq')
data = bf.blocks.detect(data, mode='jones')
data = bf.blocks.accumulate(data, 2)
data = bf.blocks.transpose(data, ['time', 'pol', 'freq'])
data = bf.blocks.copy(data, space='cuda_host')
data = bf.blocks.quantize(data, 'i8')
bf.blocks.write_sigproc(data)

pipeline = bf.get_default_pipeline()
pipeline.shutdown_on_signals()
pipeline.run()

• Designed for sustained high-throughput stream processing

• Python and C++ APIs wrap fast C++/CUDA backend

• Native support for both system (CPU) and CUDA (GPU) memory spaces and computation

• Main modules

• Ring buffer: Flexible and thread safe, supports CPU and GPU memory spaces
• Transpose: Arbitrary transpose function for ND arrays

• Experimental modules

• UDP: Fast data capture with memory reordering and unpacking
• Radio astronomy: High-performance signal processing operations

Install dependencies:

$ sudo apt-get install exuberant-ctags

Install dependencies:

• PyCLibrary fork
• Numpy
• matplotlib
• contextlib2
• pint

$ sudo pip install numpy matplotlib contextlib2 pint git+https://github.com/MatthieuDartiailh/pyclibrary.git@dff70337fd904a43de2bdd05dc548620160354d0

Edit user.mk to suit your system, then run:

$ make -j
$ sudo make install 

which will install the library and headers into /usr/local/lib and /usr/local/include respectively.

You can call the following for a local Python installation:

$ sudo make install PYINSTALLFLAGS="--prefix=$HOME/usr/local"

Note that the bifrost module's use of PyCLibrary means it must have access to both the bifrost shared library and the bifrost headers at import time. The LD_LIBRARY_PATH and BIFROST_INCLUDE_PATH environment variables can be used to add search paths for these dependencies respectively.

Install dependencies:

• Docker Engine
• NVIDIA Docker

Build Docker image:

$ make docker

Launch container:

$ nvidia-docker run --rm -it ledatelescope/bifrost

For CPU-only builds:

$ make docker-cpu
$ docker run --rm -it ledatelescope/bifrost

To quickly build the docs using Docker, ensure that you have built a Bifrost container as ledatelescope/bifrost. Then, inside the docs folder, execute ./docker_build_docs.sh, which will create a container called bifrost_docs, then run it, and have it complete the docs-building process for you, outputting the entire html documentation inside docs/html on your machine.

Install sphinx and breathe using pip, and also install Doxygen.

Doxygen documentation can be generated by running:

$ make doc

This documentation can then be used in a Sphinx build by running

$ make html

inside the /docs directory.

• Ben Barsdell
• Daniel Price
• Miles Cranmer
• Hugh Garsden
• Jayce Dowell