N. McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, May 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....software controlled data prefetching to improve memory performance by tolerating cache miss latency. The goal of prefetching is to bring data into the cache before the demand access to that data. Prior research shows that software controlled prefetching is effective in array based Fortran programs [8, 19, 5, 16] We describe a new data flow analysis to identify loop induction variables, and a method to schedule prefetches for array references that contain induction variables in the index expression. We rely on a simplified form of common subexpression elimination to remove redundant prefetches. Our new ....

....effectively. Bernstein, Cohen, Freund, and Maydan implement a variation of Mowry et al. s algorithm for the PowerPC architecture [5] The only loop transformation they use is unrolling. McIntosh extends Mowry et al. s work by focusing on the compiler support necessary for software prefetching [16]. He develops several new compiler techniques to eliminate useless prefetches and to improve prefetch scheduling for array based codes. Our prefetching algorithm is effective without locality analysis and loop transformations. We focus on Java arrays that contain features that make code and data ....

N. McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, May 1998.

....of prefetching linked data structures. C flat also works on array based codes, including indirection arrays. Most of the benefits that Selvidge reports are due to prefetching array elements. McIntosh extends Mowry s work by focusing on the compiler support necessary for software prefetching [72]. He develops several new compiler techniques to eliminate useless prefetches and to improve prefetch scheduling for array based codes. McIntosh develops a new technique for detecting cross loop reuse that provides useful information for improving software prefetching. Cross loop reuse summarizes ....

Nathaniel McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, May 1998.

....developer had already started adding the same feature, preventing that duplication of e ort. 3 Related Work 3.1 Cache Prefetching Cache prefetching as a means of performance improvement is a familiar subject in the elds of computer architecture, operating systems, and compilers. McIntosh[10] describes methods for a compiler to insert instructions to prefetch data that will be needed later in the code. The algorithm used to choose data to fetch is based on an analysis of what data will be needed next during execution of the code. Of course, all the code to be executed in the future ....

Nathaniel McIntosh. Compiler support for software prefetching. Technical Report TR98-303, Rice University, 1998.

....overhead without actually tolerating any latencies. The small number of inner loop iterations in the interchanged postlude suggests the use of outer loop prefetching: applying software pipelining at an outer loop and scheduling prefetches ahead by a certain number of outer loop iterations [McI98] In general, outer loop prefetching algorithms can be ineffective because of increased conflict and capacity misses, requiring code reorganization through strip mining to facilitate a reasonable prefetch distance and reduce the likelihood of early prefetches. However, the short inner loops in ....

Nathaniel McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, May 1998.

....1. Introduction Software controlled data prefetching improves memory performance by hiding memory latency. Its goal is to bring data into the cache before the demand access to that data. Existing research shows the benefits of software prefetching techniques for array based, scientific programs [6, 21, 4, 19]. Given an array, the size of each element, and a regular access pattern, a compiler can compute the address of any element in the array and prefetch it. Prefetching in pointer based codes is difficult because separate dynamically allocated objects are disjoint, and the access patterns are thus ....

N. McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, May 1998.

.... detect coherence misses, especially for compiler parallelized Subroutines Achieved BW BW Utilization compute rhs 252MB s 84 x solve 266MB s 89 y solve 197MB s 66 z solve 262MB s 87 lhsx 321MB s 107 1 lhsy 279MB s 93 lhsz 96MB s 32 Table 2: Memory bandwidth utilization of SP code[8]. 3.3 Implementation Status We are in the process of implementing the bandwidth based performance tool into a whole program compiler. In addition to performance tuning and prediction, the compiler performs global and dynamic computation and data transformations described in [1, 2, 3] In fact, ....

Nathaniel McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, Houston, TX, July 1997.

....improving memory performance by tolerating latency in the memory hierarchy. Compilers statically analyze programs and insert prefetch instructions to load data into the cache prior to use. Previous research shows the benefits of software prefetching techniques in array based scientific programs [4, 14, 2, 13]. Prefetching in array based codes is simpler than in pointer based codes. Given an array, the size of each element and a regular access pattern, the compiler can compute the address of any element in the array and schedule prefetches to elements in a loop that will be accessed in future ....

N. McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, May 1998.

....through cache reuse analysis[18, 17] because they do not incur additional memory cache transfer. The cache conflicts due to the limited associativity of a cache can be estimated by the approximation method used in [8] Finally, the estimation of cache misses considers the cross loop cache reuse [7, 16]. Examining source code is not enough since the actual amount of data transfer can be reduced by datareuse optimizations such as data blocking tiling and loop fusion. The performance tool can get hold of these high level transformations if it has access to compiler decisions. Most data ....

....collect such situations from applications. Coherence misses need also be measured if they are a significant factor to performance. The data analysis techniques in the bandwidth based performance tool have good knowledge to detect coherence misses, especially for compiler parallelized code [16]. 3.3 Related Work The bandwidth based approach is different from previous techniques in its emphasis on monitoring and predicting memory data transfer time for memory constrained applications. Analyses and estimations of bandwidth requirement or constraint is far easier and simpler than those ....

Nathaniel McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, Houston, TX, July 1997.

No context found.

N. McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, May 1998.

No context found.

Nathaniel McIntosh. Compiler Support for Software Prefetching. PhD thesis, Rice University, Houston, TX, July 1997.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC