Git tag: PTB_Beta-2013-04-20_V3.0.10

This release contains a couple of new features and improvements, some rather experimental by now, but also the usual amount of small improvements and bug fixes.

All systems:

• Basic GPGPU compute support:

  A first sketchy implementation of (G)eneral (P)urpose computation on (G)raphics (P)rocessing (U)nits support. We already supported (ab)use of GPUs for other tasks than rendering since the beginning of 2007, with our built-in imaging pipeline and various plugins to accelerate common image processing and stimulus post-processing operations. However, so far all functionality was implemented by reformulating computation tasks as graphics rendering tasks and executing them as a mixture of fixed-function OpenGL rendering and programmable vertex- and fragment-shading via the GLSL OpenGL shading language. This works very well for many visual processing problems, but can become awkward or inefficient for non-graphics compute problems. Technology progressed and nowadays most modern GPUs support GPGPU in a more straightforward and efficient way by use of NVidia’s CUDA API (for NVidia GPUs only) and OpenCL for pretty much all GPUs on the market, including NVidia’s, so it’s time to look into what can be done better with those APIs for stimulus creation. This beta release contains basic support for utilizing NVidia CUDA capable GPU’s under 64-Bit Matlab by use of a high-level open-source toolbox called “GPUmat”, which can be downloaded for Linux and Windows from here: GPUmat toolbox download link

  GPUmat allows to implement computational methods in familiar Matlab language, as it extends Matlab with new GPU data types GPUsingle and GPUdouble. If vectors and arrays are created with this datatype, then the data is stored on the GPU and processed in parallel by the GPU. Psychtoolbox initial implementation provides builtin functionality to detect if GPUmat is installed and to initialize it, if so. It also provides an interface for efficient data transfer and synchronization between GPUmat and Screen/OpenGL. Screen’s floating point framebuffers, offscreen windows and textures can be efficiently passed to GPUmat as GPUsingle matrices to do computations on them and resulting matrices can be converted back into PTB’s representation for post-processing, rendering and display. In this initial release, I focused on getting the high-level setup code right to hopefully allow future painless extensions, and to optimize data transfer between PTB/OpenGL and GPUmat/CUDA.

  Benchmarking showed that the current data transfer works with > 90% of the theoretical peak performance of the tested GPU under Linux, with about 30-40% efficiency on Windows, and IIRC in the 1-2% range on OS X. Profiling showed that the OS X implementation essentially performs data exchange with the speed of a slow software fall-back path and unsurprisingly the deficiency seems to be in the OS X graphics subsystem - another area where the most advanced operating system in the world falls way short of its competitors. This means that if you like to use this functionality, you can gain an almost 100x performance boost for some tasks by running you Mac under Linux instead of OS X.

  There are currently three demos in the PsychDemos/GPGPUDemos subfolder, all showing fast application of a 2D-FFT and inverse FFT to GPU accelerated image filtering by replacing convolution with a point-wise multiply in frequency space.

  Current restrictions: Only 64-Bit Matlab supported, as GPUmat needs object oriented programming support v2 (“classdef” OOP) which is not yet available in stable releases of GNU/Octave. However, octave’s classdef support has made quite a bit of progress, so I’m optimistic this will work for stable octave in the not too far future. Only 64-Bit because I’m too lazy to maintain this for 32-Bit as well. There are not technical reasons, just that people should move to 64-Bit if they want to do this style of computing efficiently. Only NVidia CUDA capable GPU’s for now, as GPUmat is based on CUDA. The long-term goal would be to primarily support OpenCL, as it is cross-platform, cross-vendor and also supports non-GPU compute accelerators. Only Linux and Windows as of today, because GPUmat is not yet available for OS X. I ported it to 64-Bit OS X, but haven’t submitted it upstream for inclusion into the official distro yet, that’s a todo.

  Setup instructions under help PsychGPGPU, demos under help GPGPUDemos.

• RenderDemo shows a new way to use our imaging pipeline and new plugins to perform GPU accelerated color space conversion and calibration. New functions of the imaging pipeline allow to convert from the XYZ tristimulus color space to calibrated RGB output, taking the calibration data from our color calibration routines into account.

  Wikipedia’s definition of XYZ color space

  The advantage is that all the color calibration can now get easily integrated into our stimulus post processing pipeline, and of course speed. The demo also shows a new plugin to convert from xyY color space to XYZ color space. It should be noted that this is only lightly tested. While light testing of a few sample values showed proper behaviour, David Brainard wants to write more exhaustive unit tests at some point in the future. Of course patches for tests are welcome.

• CRS Bits# support: The display subsystem of the Cambridge Research Systems Bits# device is now fully integrated for convenient and efficient use. The I/O system is todo. Demos and tests have been updated to take advantage of new Bits# functionality like improved diagnostic and troubleshooting.

• PR-705 photometer support contributed by Zachary Lindbloom-Brown.

• The ColorCal2 driver now also works with GNU/Octave, thanks to the donation of a ColorCal-II by CRS. The driver seems to work very well on Linux, is a hit and miss on OS X, apparently due to funny new OS X bugs or incompatibilities, and was a no-go for me so far on Windows. The current driver uses USB-HID protocol to communicate with the device, which seems to be flaky on non-Linux. It might make sense to switch it over to good old Serial-over-USB protocol which is probably more trouble-free on the other OS’es.

• OptiCal support, contributed by Andreas Widmann, although it could be this

• Various smaller fixes and improvements, e.g., updates to demos, DrawFormattedText now accepts arbitrary ’rect’angles for layout of text, not only the bounding rectangle of the target window or texture, OS specific fixes, …

Linux:

• Linux Google Nexus-7 / ARM support:

  Included in this release is experimental support for the Google Nexus-7 tablet when run under Ubuntu Linux 13.04 “Raring Ringtail”, 32-Bit ARM edition. This won’t work under the normal Android OS on the tablet, but you will need a tablet edition of Ubuntu desktop Linux, either as only operating system, or as dual-boot setup in parallel to your Android OS. Psychtoolbox seems to work reasonably well on the tablet, with a certain amount of tweaking.

  What works? GetSecs/WaitSecs with usual precision, real time scheduling, sound (PsychPortAudio and 3D OpenAL), keyboard and mouse I/O — touch input as mouse input, IOPort and PsychHID stuff is so far untested, but expected to work normally if you have an OTG adapter cable, Screen() and MOGL 3D OpenGL mostly works, i.e., all the basics work. What doesn’t work is functionality that requires programmable shading support, as we currently only support OpenGL-ES1.1 fixed function rendering, not OpenGL-ES2.0 programmable shading.

  Simple installation instructions for Ubuntu Linux on the Nexus-7 under this link.

  This would wipe your Android OS and data from your device, so no more “Cut the rope” or “Angry Birds” for you, but all the fun of drifting gabor patches and movie playback.

  I chose a dual-boot setup, which preserves Android and its goodness, but requires serious tinkering, and averaging installation instructions over multiple partially incomplete or mutually contradictory howtos, e.g., this howto.

  Anyway, this is mostly a test setup for now to see how much interest there is in Psychtoolbox for mobile devices and what the challenges are. I chose the Nexus-7 because it is the reference device for porting and testing Ubuntu Linux and it is also rather good price/performance at 200$/Euros.

• Linux highly experimental multiple display and rendering backend support: Not really interesting for now, but this is to test next generation display system support for the Linux PTB, ie., non-X11 display setups, EGL, Wayland, raw framebuffer, Android and OpenGL-ES / OpenGL-3/4/…

OS X:

• More fixes for OS X Retina and LCD timestamping brokeness and other OS X bugs. I now recommend to always install the PsychtoolboxKernelDriver if you use OS X 10.7 or later and want precise stimulus onset timing or timestamping. The level of OS X brokenness has reached the point where our own implementation is almost always better than what Apples OS has to offer. E.g., OS X beamposition queries are only supported anymore on NVidia GPU’s and scheduled by Apple for complete removal in a future OS X release. Even on NVidia the mechanism is slightly broken and inaccurate on all modern GPUs and totally broken as of OS X 10.8 for any kind of digital flat panel / LCD / DVI-D etc. - it only sort of works for classic analog VGA connected CRT monitors.