GPU-Accelerated Wiener Filter/STC Matrix Extraction

This program makes use of your GPU to extract the Wiener Filter or STC matrix (1,2) from a set of inputs and responses. This can speed up computations >100x compared to commonly used CPU-based implementations. This repository contains a MATLAB wrapper and an example C++ application.

Our paper documenting this algorithm is:

Graphics Processing Unit-Accelerated Code for Computing Second-Order Wiener Kernels and Spike-Triggered Covariance by Omer Mano and Damon A. Clark

Compatibility

This program should be compatible with most modern systems and has been sucessfully tested on the following systems:

• OS X El Capitan with Intel Iris 6100 graphics
• macOS Sierra with Intel Iris 6100 graphics
• Windows 7 with Nvidia 1060 graphics
• Windows 7 with Nvidia 560 graphics
• Windows 8 with AMD Fury X graphics
• Windows 10 with Intel HD 510 graphics
• Windows 10 with Nvidia 970m graphics

We have documented problems running on Macs using Intel HD 5000-series integrated graphics (2014 models and earlier). We believe these are due to non-conforming OpenCL drivers (a known issue on Macs).

Install

To use the MATLAB wrapper, you will need the appropriate .mex file for your system. This can be downloaded from this repository's releases page. This is the best strategy in most cases. Follow the steps in the release's included readme file. If you want to compile the code from scratch, use the steps below.

Compile

Note that this should not be necessary for most users. Instead, use our latest release

1. Make sure you have a compiler installed such as Visual Studio on Windows or XCode on macOS
2. On Windows, download and install the OpenCL SDK from your GPU vendor. MacOS has these libraries built in.

• For Nvidia, use the CUDA SDK
• For AMD, use the AMD APP SDK

3. In MATLAB, navigate to the mex folder in this repository, and type edit compile.m
4. Uncomment the lines pertaining to your OS/GPU vendor, and run the script

• On macOS, you may need to follow the instructions here to enable mex compilation

5. Copy all the files from the mex folder except compile.m and extract2ndOrderKernelGPU.cpp to a convenient location- we'll call this the "install directory".
6. Copy the files in the clFiles directory into your MATLAB path. These files must be placed in the same directory as getClFilePath.m
7. Add your install directory to your MATLAB path: navigate to the install directory in MATLAB and type addpath(pwd) in the terminal.

Usage

To recreate figures from our upcoming paper, follow these steps:

1. Make sure your install directory is in your MATLAB path (see step 7 in "Compile" or the readme.txt in the Release directory)
2. In MATLAB, navigate to the MatlabDemos folder
3. Run generateFigure1B.m

• The first time you run this script, test data will be generated, which will take a few seconds

4. Run generateFigure1CDEF.m Note that this can take several hours to run, depending on your CPU

• As an alternative, run simpleBenchmark.m to get a sense of the speedup on your machine.

For more information regarding the MATLAB wrapper function extract2ndOrderKernelGPU, type help extract2ndOrderKernelGPU in MATLAB. View and run exploreMatrixElements to gain intuition about the output of extract2ndOrderKernelGPU.

For an example of how to use the function in C++ (also useful for creating your own wrapper), see exampleUsage.cpp. Note that the output array will contain numTau * numSpatial * numTau * numSpatial * numROIs elements. These will be organized in the standard STC/Wiener kernel format, such that the response of cell c to a product of stimulus dimension i with stimulus dimension j at timelags t1 and t2, respectively, is stored in the element

t1 +
i *numTau + 
t2*numTau*numSpatial + 
j *numTau*numTau*numSpatial +
c *numTau*numTau*numSpatial*numSpatial

Troubleshooting

Error message	Solution
Invalid MEX-file: The specificied module could not be found	This most likely means OpenCL could not be found on your system. Try reinstalling your graphics drivers.
Undefined function or variable extract2ndOrderKernelGPU	Add the folder containing the correct mex file to your MATLAB path (see step 7 in "Compile" or readme.txt in the Release directory
Attempt to execute SCRIPT extract2ndOrderKernelGPU as function	extract2ndOrderKernel.m is in your MATLAB path, but the mex file is not. Download the mex file from the releases page or compile using the steps above.