Download Jacket v2.3

MATLAB® is a registered trademark of MathWorks Inc. (More info)

Release Notes

From Jacket Wiki

Jump to: navigation, search
Back to Documentation

Contents

Version 2.3 (10/22/2012)

  • Requires: CUDA 5.0 drivers
  • Supports NVIDIA K20 GPUs and CUDA 5.0

Installation & Activation

Windows and Linux installers only.

  • Simplified process for installation; get up and running quickly
  • Installer log makes it easier to debug issues during installation

Enhancements

  • PINV added GFOR support and performance enhanced 10x (switched from SVD to QR approach, see more)
  • Jacket HPC supports all operating systems (Windows,Linux,Mac) and no longer requires jacketd
  • Linux installer size decreased (faster Jacket JMC startup)
  • Performance of FFT improved
  • MIN and MAX indexed output is double precision if input double precision
  • GINFO now returns both (a) number of GPUS present and (b) number of GPUS licensed, and avoids printing addons if none present
  • gselect('unused') returns least utilized GPU for coordinating Jacket MGL

Bug Fixes

  • SUM, MIN, MAX, ANY and ALL inside GFOR for scalar inputs
  • licensing error for MRDIVIDE
  • POWER inside GFOR for scalar exponents.
  • avoid unecessarily GPU initialization in multi-GPU scenario (save memory) (SPT-719)
  • fixed memory leak when pushing MATLAB data to device (SPT-679)
  • Jacket JMC allows Jacket DLA, Jacket MGL, Jacket SLA, etc. (SPT-719)
  • IMERODE memory leak (SPT-521)
  • Jacket HPC network licensing correctly uses license files instead of requiring environment variables (SPT-658)
  • detect and warn if arithmetic array larger than 2 GB (unsupported: JKT-2538)

Version 2.2 (6/16/2012)

  • Requires: CUDA 4.2 drivers
  • Kepler Support
    • This version of Jacket is designed for compatibility with NVIDIA's latest line of GPUs, such as the Geforce GTX 600 series, that utilize the efficient Kepler architecture.
  • New Features:
    • Jacket now allows multiple versions to exist.
    • New Installer path: C:\Program Files\AccelerEyes\Jacket\2.2\
  • Known Issues
    • Due to a bug related to the NVIDIA driver, the first Jacket command may take more time than usual to load. AccelerEyes is working with NVIDIA closely to ensure that this issue is fixed.
    • This issue is slated to be fixed with NVIDIA driver version 304 or higher.

Version 2.1 (3/16/2012)

A volume rendering of random data generated by VOLUME
A visualization of stereo reconstruction data by SCATTERCOLORS
  • Requires: CUDA 4.1 drivers
  • New features
    • CIRCSHIFT now supports shift argument as a GPU type
    • Sparse matrix subscripting
    • GFOR
    • Optimization Library Functions (unconstrained minimization)
      • MINGD minimizes an objective utilizing gradient-descent coupled with an Armijo line search
      • MINBFGS minimizes an objective utilizing the BFGS method coupled with line search
      • Utilizes GFOR to compute Jacobian and Hessian when possible (and when not explicitly provided)
    • Graphics
    • NULL added
  • Bugs
    • IMHIST Now works on bin count > 4000.
    • MESHGRID bug fixed
    • SORT bug fixed, correct results on larger sizes
    • Subscript referencing out of bounds produces NAN (or zero/false for integer/logical). Assignment out of bounds has no effect, i.e. assignment is ignored.
    • CAT now works on GINT32 and GUINT32 data types
  • Examples
    • All examples have been simplified for easier understanding and faster performance

Version 2.0 (16/11/2011)

A GPU Mandelbrot Set, generated using ARRAYFUN
Genomics - Multidimensional Scaling for analysis of Crop Gene Expression Patterns
Graphics library: simulating shallow water equations with reflective boundary conditions
Graphics library: image registration visualizing all flow gradients as scatter plot
  • Additional examples in jacket/examples/:
    • JacketDemo - runs a subset of the included examples as side-by-side CPU/GPU comparison
    • black_scholes_example and black_scholes_mgl - Black-Scholes option pricing using all available GPUs
    • brainseg - maximum-likelihood image segmentation of an MRI
    • genomics_mds - High-Throughput Dimensional Scaling (HiT-MDS) to analyze genomic data
    • heart - real-time volume filtering of a heart MRI
    • mandelbrot - Real-time, large domain Mandelbrot image construction
    • music_visualizer - Signal analysis and music beat visualization (see more)
    • volume - real-time volume rendering of the brain

Version 1.8

  • Requires CUDA 4.0 drivers:
  • Built against the CUDA 4.0 Toolkit. We package the CUDA libraries, so users are not required to install this toolkit.

09/29/2011 - v1.8.2

  • New features
    • GINFO sorts devices by theoretical FLOPS (forums)
    • GSELECT between devices any time (not just at startup)
  • Improvements
    • IFFT, IFFT2, IFFTN up to 10 percent faster
    • More specific licensing error messages for various cases.
    • Switch GRAND,GRANDN to use NVIDIA's CURAND library (3x or more speedup)
  • Changes
    • CUMSUM bug fixed inside GFOR (forums)
    • Memory leak inside Reductions fixed (SUM, PROD, MIN, MAX)
    • BWDIST now works properly at larger sizes (forums)
    • MPOWER bug for A^k where k=13
    • Fixed bug for too many file handles being opened
    • Fixed complex imprecision issue for mrdivide and rdivide
    • Fixed an "Invalid MEX File" Error occurring with Jacket 1.8.1 (forums)

08/22/2011 - v1.8.1

  • New features
    • BWLABEL
    • REGIONPROPS
    • New and Improved "Jacket Demo" (see examples/JacketDemo)
    • New Music Visualizer example (described here)
    • New Optical Flow Tracking example (see examples/oflow_example)
    • New Black-Scholes example using GCOMPILE examples/black_scholes_example. ~35X faster than CPU.
  • Improvements
    • Issue with Jacket over Remote Desktop reported on the forums fixed
      • The issue also was preventing GSELECT from functioning properly on some Windows machines.
    • Jacket now gives LM_LICENSE_FILE and LM_LICENSE_PATH precedence over engine/jlicense.dat (When both exist).
    • TIMEIT accepts new calling syntaxes
    • MLDIVIDE is faster for small over-determined systems inside GFOR. (Number of unknowns <= 32).
    • IMHIST generally faster for compute capability >= 1.2
  • API Changes
    • MTIMES returns GPU scalars even if the inputs are vectors.

07/18/2011 - v1.8

  • New Subscripting Features(SUBSREF,SUBSASGN)
    • Double-precision indexing (forum posts 1, 2)
    • Removing items: A(index)=[] (discrete, sorted idx)
    • 5x faster A(boolean_mask)=scalar
  • Mac OSX only supports 64-bit Snow Leopard 10.6.3 or later (discontinued Leopard 10.5 and 32-bit)
  • New functions:
  • Improvements:
  • Examples have been overhauled:
    • Updated so CPU and GPU use the same code (see "Write once, run everywhere")
    • New mandelbrot_example: Using ARRAYFUN to generate a Mandelbrot set - gives typically a 1000x speedup
    • New brainseg example: Segmentation of cranium scan
    • New genomics_mds_example: Multidimensional Scaling for Genomic analysis
    • For a complete description of the current examples: Examples
  • API Changes:
    • SUM, PROD, MIN, MAX, ANY, ALL on vector inputs now return GPU scalars instead of CPU scalars.
  • Graphics:
    • PLOT, PLOT3 now accept linestyle string parameter to set color
    • axis rescaling and panning now corrected so that ticks always visible
    • Added GCOLORMAP, GTITLE, GDRAWNOW
    • Mac OSX support
    • Improved Windows Performance
  • DISPLAY, DISP view contents of each GFOR tile instead of only the first
  • Sparse matrix support for BICGSTAB; support for older versions of MATLAB
  • In many cases, pre-allocation before subscript assignment is no longer needed

Example:

% v1.7
A = gones(n,n);
gfor ii = 1:n
  A(ii,:) = gones(1,n);
gend
 
% v1.8 (no need to preallocate A)
gfor ii = 1:n
  A(ii,:) = gones(1,n);
gend

Version 1.7

New graphics library: image registration using classic Horn-Schunck, Gauss-Newton
New graphics library: debugging image registration code
New graphics library: Simple Example
New graphics library: Tweaked FDTD
  • Requires CUDA 3.2 drivers:
  • Built against the CUDA 3.2 Toolkit. We package the CUDA libraries, so users are not required to install this toolkit.

03/21/2011 - v1.7.1

Enhancements:

  • CUMSUM handles matrices up to 4 dimensions.
  • Cleaned up several examples to use new CLASS syntax to avoid separate CPU and GPU code (see blog)
    • pi, fft, blas, fdtd
  • GFOR-enabled MPOWER

Bug fixes:

  • CULA initialization error on Quadro cards
  • Mac load error for libkublas
  • GRANDN single-precision was producing NaN values
  • Issues with IMHIST for large matrix sizes
  • MTIMES has been fixed for sparse matrix * dense matrix multiplication
  • INTERP1 now works properly for yi = interp1(x,Y,xi,method)*

02/26/2011 - v1.7

New graphics library visualizing FDTD simulation
  • Enhanced convolutions: CONV, CONV2, CONVN
    • No kernel size limitation
    • Supporting double precision and complex numbers
    • Full support with GFOR
  • Graphics Library Refresh
    • Optimized for minimum compute thread impact both in MATLAB and in standalone C++ apps
    • Axis tick marking, automatic scaling of plot data, new slick look
    • Supported primitives: SURF, PLOT, IMAGESC, SCATTER3
      • sophisticated syntaxes on the way (e.g. plot linestyle, marker shape, marker size)
    • New (updated) functions: GHOLD, GSUBPLOT, GCLF, GDRAWNOW (GFIGURE, GCLOSE)
    • IMAGESC performance greatly improved
    • Jacket OpenGL context creation only on demand (compatibility with applications which require their own GL contexts)
    • Drastically reduced disk, memory, and processor footprint
    • CUDA independent graphics mode
  • Improvements
    • ONES, ZEROS, RAND, RANDN, INF, NAN, etc. now available via new usage of CLASS
    • New functions GINF and GNAN added
    • TIMEIT provides robust estimates of both CPU and GPU code snippets (see demos/blas_example/)
    • Support for Remote Desktop (see blog post)
    • New GCOMPILE features:
      • BREAK, RETURN, and CONTINUE
      • Additional EPS syntax: eps('single') and eps('double')
      • Expanded FOR loop support - Instead of only hard-coded literal values, any COLON expression may be used.
  • Bug Fixes
    • GPROFILE
      • now deals correctly with GRANDN, non .m files, mixed precision, NAN
    • Memory leaks fixed as posted on the (forums)
    • FFT plan caching fix which was affecting some combinations
    • GMEX updates for CUDA 3.2, fixes
    • GCACHE corrections to avoid instability caused by presence of older Jacket cache files and now portable across 32bit and 64bit architectures.
  • Limitations
    • The graphics library is not supported for MAC systems in this release.

Versions from 2010 and earlier

For older versions, visit Release Notes 2010 and Earlier.

Views
Personal tools