PyCUDA and PyOpenCL backends for ASSET joint prob. matrix calculation

[dizcza]

Benchmarks (compared to the original Python implementation):

• PyCUDA: x1000 and more
• PyOpenCL: x100 and more

Changes to _JSFUniformOrderStat3D class:

1. Changed the default precision from double to float - I find floats perform reasonably well as doubles with my built-in Intel graphics card and yet floats are x4 faster than doubles for both backends. In either case, the users can easily change the precision manually at any time.
2. Added pycuda() backend that allows copying arrays from RAM (virtual Python memory) directly to CUDA global memory.
3. Added pyopencl() backend - suited for all laptops with built-in Intel GPU card. As in the PyCUDA backend, it allows copying directly to (Intel) GPU global memory.
4. Renamed cuda() backend to _cuda(). This backend was the breakthrough in accelerating ASSET with CUDA, but it suffers from disk I/O operations. The _cuda() backend and joint_pmat_old.cu file should be removed once the PyCUDA backend becomes no longer an experimental feature.
5. Added a watchdog that barks when the computed values of a joint prob. matrix are outside the valid [0, 1] interval. For this reason, the tolerance parameter is added in the joint_probability_matrix function.
6. Added description of how to install CUDA and OpenCL support in Elephant installation documentation. In particular, if PyOpenCL backend is used, the users need to disable GPU Hangcheck (described in install.rst) to prevent kernel resets when computation takes a long time to finish. Doing so, of course, makes the system unresponsive until the compute program terminates.

Changes to _pmat_neighbors function:

1. Accelerated with CUDA and OpenCL. Rewrote to _PMatNeighbors class to facilitate different backends. It became thousands of times faster, the exact speedup doesn't matter - it runs in the blink of an eye.

Changes to other Python functions:

1. Optimized du = np.diff(u) in both memory and speed (see the compute function).
2. x5 less memory footprint of the cluster_matrix_entries function with chunking. The chunk size is controlled by the working_memory parameter. With working_memory set to 100, the peak allocation memory (of cluster_matrix_entries and, therefore, ASSET itself since it was the most memory consuming part of ASSET after the bug in pmat is fixed in Memory efficient and faster implementation of ASSET pmat analytical #399) is reduced to, compared to the master branch:

mmat.shape	No chunking, Mb	Chunked, Mb
(150, 150)	4300	815
(170, 170)	7600	1440
(200, 200)	MemErr	3000
(250, 250)	MemErr	8000
(270, 270)	MemErr	11200

---

It's also possible to install pyopencl with pip if you've installed Intel GPU driver manually. You need to make sure that pyopencl sees your system-wide libOpenCL.so*. First, find the location of libOpenCL.so and then provide its directory path as LD_LIBRARY_PATH environment flag. Here are the commands I used:

$ ldconfig -p | grep OpenCL
        libOpenCL.so.1 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libOpenCL.so.1
$ LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu pip install pyopencl

I didn't put it into the install.rst because it's advanced stuff and conda-forge works reliably.

---

CUDA part is not tested because Travis does not provide GPUs. Here are the steps to test it manually on Google Colab:

1. Runtime -> Change runtime type -> GPU
2. pip install -e git+https://github.com/INM-6/elephant#egg=elephant[extras,cuda]
3. Restart the kernel.
4. python /content/src/elephant/elephant/test/test_asset.py

[coveralls]

Coverage decreased (-0.9%) to 88.749% when pulling abe146a on INM-6:cuda/asset into 5e95f77 on NeuralEnsemble:master.

[pep8speaks]

Hello @dizcza! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-02-22 13:48:36 UTC