AltiVec

Fortran/C/C++ Compilers & Debugging Tools

Resources

      Buy now, contact Sales,
      Questions about your order,
      Resellers, Partners,
     Partial customer list

      Compilers, Debuggers,
      Tools and Libraries,
      Cluster Software

      Commercial, Academic,
      Cluster Software, Upgrades
      Multi-platform Bundles

Contact, FAQs, User Forum Downloads, Registration

Product Information

IMSL math/stat libraries
VAST-F/Parallel
VAST-F/Vector
Fx & Fx2 debuggers
Cluster products
Documentation
Companion Products

AltiVec

AltiVec (also called Velocity Engine by Apple) is the vector processing unit integrated into the G4/G5 processor. In many instances AltiVec operations can dramatically increase the speed of executed code. There are two ways to take advantage of AltiVec without rewriting all your code:

Use a compiler and libraries which automatically take
advantage of AltiVec.
Use tools(optional) which can restructure standard code to include AltiVec directives.

Absoft Pro Fortran offers opportunities to easily improve your code using both these methods. No recoding is necessary and AltiVec directives can be set from the Pro Fortran IDE.

Automatic AltiVec Support

The F95 compiler included with Pro Fortran for OSX and PPC/Linux includes an F90 intrinsic function library optimized for AltiVec. If your code uses these routines you may automatically see a performance increase.
Absoft provides AltiVec optimized versions of BLAS (Basic Linear Algebra Subroutines) on the distribution CD.

Optional AltiVec support

IMSL (optional) can take advantage of the AltiVec optimized routines in the G4/G5 version of the IMSL libraries. IMSL Math and Stat libraries are a collection of 1000 special function math and stat routines.
VAST-F/Vector (optional) will automatically restructure code to take advantage of possible AltiVec performance increase and include AltiVec instructions. VAST-F/Vector is available for Macintosh OSX and PPC/Linux Pro Fortran users.

Examples of user experience with AltiVec:

Craig Hunter of Lockheed developed a comparision of G4/AltiVec vs Intel based systems.
South Hampton College’s High Performance Computing group did comparisons as well.

Detailed AltiVec Support in Pro Fortran and included libraries Below

Velocity Engine Speed Increases

The table below shows the average increase in speed for BLAS routines optimized for Velocity Engine vs. the same routines created using standard optimizations. Both were created using Absoft Pro Fortran. Multiple runs were made using array sizes ranging from 2000 to 200000 elements.

VE performance benefits are only available to G4 users.

BLAS Routine	Average Execution Speed Increase (%)
SDOT	93.57%
SASUM	140.62%
ISAMAX	291.11%
SCOPY	69.44%
SSWAP	50.56%
SSCAL	37.41%
SAXPY	32.73%
SROT	104.42%

Pro Fortran for Power Macintosh and Power/PC Linux is optimized for Velocity Engine™ (AltiVec) in the following manner:

The following F90/95 intrinsic functions have been optimized for Velocity Engine™ (AltiVec) execution.
NOTE: The AltiVec execution unit included in the G4/G5 CPU supports single precision operations only.
These F90/95 intrinsic routines currently only support 4-way parallelism for 32-bit floating point numbers with no optional parameters.

MAXLOC

MAXVAL

MINLOC

MINVAL

The following F90/95 intrinsic routines currently support either 4-way parallelism for 32-bit floating-point numbers, or 2-way parallelism for COMPLEX*8 imaginary numbers with no optional parameters.

DOT_PRODUCT

MATMUL

SUM

The following routines have been optimized for Velocity Engine™ (AltiVec) and provide the foundation for both the Velocity Engine optimized BLAS, LAPACK90, and IMSL Math and Stat Libraries. These routines support either 4-way parallelism for 32-bit floating-point numbers, or 2-way parallelism for COMPLEX*8 imaginary numbers where appropriate. These routines are compatible with both F90/95 and F77 compilers.

CAXPY	CGEMV	ICAMIN	SAXPY	SGEMV
CCOPY	CSCAL	ISAMAX	SCASUM	SROT
CDOTC	CSSCAL	ISAMIN	SCOPY	SSCAL
CDOTU	CSWAP	ISMAX	SDOT	SSUM
CGEMM	ICAMAX	ISMIN	SGEMM	SSWAP