Pre-Processing Software

Auto-Parallelization & Auto-Vectorization Tools

More Features:

• Full Loop Nest Analysis. Loops are analyzed in simple and complicated loop nests; loops containing the largest amount of work are parallelized. Loops do not have to be tightly nested.
• Extended Parallel Regions. VAST/Parallel extends parallel regions to include multiple parallel loops and intervening scalar code. This cuts down on parallel overhead.
• Threshold testing. All parallel systems have some overhead. When VAST/Parallel finds a parallel region, if the amount of work in the region is not clear at compile time, then VAST/Parallel creates a run-time test. Through this run-time test, the parallel region will only be executed if there is enough work; otherwise, the original serial version is executed.
• Dependence Analysis. VAST/Parallel has very sophisticated data dependency analysis capabilities that allow it to optimize complicated situations. All loop nests are examined to see if they can be executed in parallel safely. VAST/Parallel can resolve ambiguous subscripting by examining variable assignments outside of loops, and restructure the use of variables to avoid certain other dependencies.
• Potential Dependence Testing. When dependencies are unclear at compile time, sometimes VAST/Parallel can generate run-time tests to allow parallelism to proceed.
• Special Reduction Optimization. Summations and other reductions are parallelized through the use of locks or critical regions.
• Shared/Private Determination. All variables in a parallel loop are categorized as shared (seen by all threads) or private (copy in each thread). VAST/Parallel can detect and create private arrays.
• Interprocedural Analysis for Parallel Calls. VAST/Parallel can examine call chains to determine their dependencies, and then parallelize loops containing calls or groups of calls outside loops.
• Automatic recognition of parallel cases. When sections of code deal with disjoint operations, VAST/Parallel can process each section in a separate parallel case.
• Superscalar optimizations. VAST/Parallel includes scalar optimizations to boost performance even in a single thread. Parallel optimizations can be done to outer loops while inner loops are optimized for efficient execution on one thread.
• Array Syntax. VAST/Parallel can in general parallelize and optimize multi-dimensional array syntax just as efficiently as loop nests.
• Choice of static or dynamic partitioning of loop iterations. Load balancing can tradeoff with loop overhead. Use dynamic partitioning when you need more load balancing, static partioning when you are concerned about overhead.
• Number of threads can be set with an environment variable. This allows degree of parallelism to be changed from run to run. When the system is busy you can run with two threads, when it is empty you can run with eight threads, without recompiling your program.
• Choice of thread waiting strategy. You can select either busy waiting or sleep waiting for threads, so that the parallel program can adapt to loaded or dedicated workloads on the target system. Use busy waiting on a lightly loaded system, and sleep waiting when another job might need the cycles.
  
  

More Features:

• Optimization of entire loop nests, not just inner loops. Critical optimizations include loop fusion (squeezing multiple loops into one loop), outer loop unrolling (unrolling an outer loop inside an inner loop), loop collapse (making one long loop from a multiple dimension loop), and loop interchange (changing the order of the loops in a loop nest to get more efficient memory access).
• Unrolled vector loops. Unrolling vectorized loops is very important in making sure that the vector instructions are overlapped the the maximum extent possible.
• Vectorization of reduction loops. Includes array summations, dot products, minimum and maximum element of an array, product of array elements, etc. These operations take a large fraction of the CPU time for many programs.
• Vectorization of conditional loops. "if" statements and conditional operators are vectorized.
• Non-aligned vectors can be vectorized efficiently. VAST introduces "permute" operations to align vectors "on the fly" prior to computation.
• 32-bit float and 8, 16 and 32-bit integer vectorization. Integers can be signed and unsigned. Also, VAST can vectorize loops that contain mixed data sizes.
• ALIGNED pragma so that the user can inform VAST-C about arrays that are aligned on 16-byte boundaries. Also the -Valigned command line switch.
• -Vmessages switch to get vectorization messages for all loops in the program. Find out what constructs are inhibiting vectorization of your important loops.
• DISJOINT, NODEPCHK pragmas for disambiguating data dependencies. Especially useful if the target program uses lots of pointers rather than array notation.
• -L parameter for assertion levels to allow vectorization in the presence of pointer arguments. Can be very useful if the program is written to pass most of the data as pointer arguments.
• Vector load lifting. Move all loads to the top of the loop, as far as they will go (safely). Allows the compiler to do a better job of instruction scheduling.
• Vectorization of complex data type. Uses the permute instructions to reorder interleaved complex data so that it can be operated on with the vector unit.
• Testing for stride one on loops with variable stride. Inserts a run-time test to see if variable array strides are all one; executes a vector version of the loop if the strides are one, otherwise executes the original scalar loop.
• Partial vectorization of loops with strided or gather/scatter vectors.
• Vectorization of "table lookup" loops. Loops that have a branch out of the loop can be vectorized in certain cases.

Performance Gains on a Single CPU system:

VAST/Parallel's superscalar optimization technology can enhance the performance of certain types of code on standard, single CPU systems. If your programs spend large amounts of time in nested loops or operating on large arrays, a performance improvement of over 35% may be possible. On other types of code, VAST/Parallel may have little impact.

Performance Gains on Dual CPU System:

VAST/Parallel can automatically parallelize your code and also provides full OpenMP support to enable user-directed parallelization. VAST/Parallel contains sophisticated data dependency analysis technology to detect when optimized execution will be safe, has very advanced in-lining capabilities, and uses interprocedural analysis to optimize across procedure boundaries.

VAST/Parallel fully supports the OpenMP standard. For calculations where you know exactly what you want parallelized, OpenMp provides a portable way to specify this. VAST/Parallel supports all OpenMP directives/pragmas and functions, and provides diagnostics on incorrect use of the directives.

      Special Features:

• Thread private common (choice of methods)
• Orphan directives
• Nested parallelism
• Reduction optimizations
• Environment variables
• Efficient library implementation

The driver(s) that comes with VAST Vector (AltiVec) combines VAST and the compiler(s) in a transparent way, so that (for example) compilation can be as easy as replacing gcc with vcc or f90 with v90 in your makefiles.

There are several ways to use VAST. If your program spends most of its time in clean loops, then VAST may be able to vectorize your program automatically. Often with C programs, depending on the programming style they are written it, a potential "data dependency" between pointers and arrays may prevent some vectorization, and some simple assertions from the user can improve the amount of vectorization. VAST can provide messages that help you understand what parts of your program have been successfully optimized and what parts have not been optimized.

Advanced users may choose to write clean loops for new applications and have VAST automatically generate AltiVec code, rather than doing AltiVec coding by hand. Very advanced users may wish to modify the VAST intermediate C code and change the order or nature of vector operations that VAST generates.