V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....but with the help of modulo operations (as discussed at the end of Section 4) which makes the code more complex. Also, they cannot handle fusion preventing arcs, which is fundamental to be able to consider general codes. Note that folding the memory with modulo operations was also used in [12] but for a given code (no optimization, the order of computations is given) 6 Summary and Future Work In this report, we contributed to answer the questions that remained open after the early work on array contraction by Gao and al. 7] We first proved two NP completeness results: when ....

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.

....on a parameterized version of the input program, avoiding the need to expand a graph for varying parameters and problem sizes, and it can often reduce to a linear program for flexible and e#cient optimization. Polyhedral representations have also been utilized for powerful storage optimizations [22, 28, 31, 20]. In this paper, we aim to bridge the gap and employ the polyhedral representations of the scientific community to analyze the synchronous dataflow graphs of the DSP community. Towards this end, we present a translation from a dataflow graph to a System of A#ne Recurrence Equations (SARE) which ....

....optimization problems that are already the focus of the DSP community. The polyhedral model is appealing because it provides a linear algebra framework that is simple, flexible, and e#cient. 7.3. 1 Bu#er Minimization Storage optimization is one area in which both the scientific community [22, 28, 31, 20] and the DSP community [25, 12, 24] have invested a great deal of energy. Both communities have invented schemes for detecting live ranges, collapsing arrays across dead locations, and sharing bu#ers arrays between di#erent 14 statements. It will be an interesting avenue for future work to use ....

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3-- 4):649--671, May 1998.

....on a parameterized version of the input program, avoiding the need to expand a graph for varying parameters and problem sizes, and it can often reduce to a linear program for exible and ecient optimization. Polyhedral representations have also been utilized for powerful storage optimizations [22, 28, 31, 20]. In this paper, we aim to bridge the gap and employ the polyhedral representations of the scienti c community to analyze the synchronous data ow graphs of the DSP community. Towards this end, we present a translation from a data ow graph to a System of Ane Recurrence Equations (SARE) which are ....

....optimization problems that are already the focus of the DSP community. The polyhedral model is appealing because it provides a linear algebra framework that is simple, exible, and ecient. 7.3. 1 Bu er Minimization Storage optimization is one area in which both the scienti c community [22, 28, 31, 20] and the DSP community [25, 12, 24] have invested a great deal of energy. Both communities have invented schemes for detecting live ranges, collapsing arrays across dead locations, and sharing bu ers arrays between di erent 14 statements. It will be an interesting avenue for future work to use ....

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3{ 4):649-671, May 1998.

....there are AOV s shorter than anyUOV since the AOV must be valid for a smaller range of schedules. Finally, our frameworkgoesbeyond AOV s to unify the notion of occupancy vectors with known ane scheduling techniques. Another related approach to storage management for parallel programs is that of [3, 2, 11]. Given an ane schedule, 11] optimizes storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended in [3, 2] to consider storage mappings for a set of schedules, towards the end of capturing the tradeo between parallelism ....

....since the AOV must be valid for a smaller range of schedules. Finally, our frameworkgoesbeyond AOV s to unify the notion of occupancy vectors with known ane scheduling techniques. Another related approach to storage management for parallel programs is that of [3, 2, 11] Given an ane schedule, [11] optimizes storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended in [3, 2] to consider storage mappings for a set of schedules, towards the end of capturing the tradeo between parallelism and storage. However, these ....

[Article contains additional citation context not shown here]

V. Lefebvre and P.Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3-4):649-671, May 1998.

....such that some storage optimization can precede the nal choice of schedules. Integrating our approach with their framework could provide an interesting avenue for future research. Another related approach to storage management for parallel programs is that of Lefebvre, Feautrier, and Cohen [3, 2, 13]. Given an ane schedule, Lefebvre and Feautrier [13] optimize storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended by Cohen and Lefebvre [3, 2] to consider storage mappings for a set of schedules, towards the end of ....

....nal choice of schedules. Integrating our approach with their framework could provide an interesting avenue for future research. Another related approach to storage management for parallel programs is that of Lefebvre, Feautrier, and Cohen [3, 2, 13] Given an ane schedule, Lefebvre and Feautrier [13] optimize storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended by Cohen and Lefebvre [3, 2] to consider storage mappings for a set of schedules, towards the end of capturing the tradeo between parallelism and ....

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3-4):649-671, May 1998.

....assignment [12] All these techniques allow partial removal of memory based dependences, but may extract less parallelism than conversion to single assignment form. ffl Applying storage mapping optimization techniques [4] Some of these are either schedule independent [16] or schedule dependent [13] yielding better optimizations whether they require former computation of a parallel execution order (scheduling, tiling, etc. or not. Trying to get the best of both directions is the goal of this paper. Our contribution is to show how these two directions can be combined into a unified ....

....Mapping Optimization This shows the need for a memory usage optimization technique. Storage mapping optimization (SMO) 4] consists in reducing memory usage as much as possible as soon as a parallel execution order has been crafted. The technique presented in [4] extending previous results [16, 13] to general loop nests allows automatic generation of a more memory economical program, see Figure 2.b. A single two dimensional array can be used, while keeping the two outer loops parallel. Run time computation of function OE with array Last seems very cheap at first glance; But execution of ....

[Article contains additional citation context not shown here]

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.

....are AOV s shorter than any UOV since the AOV must be valid for a smaller range of schedules. Finally, our framework goes beyond AOV s to unify the notion of occupancy vectors with known ane scheduling techniques. Another related approach to storage management for parallel programs is that of [3, 2, 11]. Given an ane schedule, 11] optimizes storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended in [3, 2] to consider storage mappings for a set of schedules, towards the end of capturing the tradeo between parallelism ....

....since the AOV must be valid for a smaller range of schedules. Finally, our framework goes beyond AOV s to unify the notion of occupancy vectors with known ane scheduling techniques. Another related approach to storage management for parallel programs is that of [3, 2, 11] Given an ane schedule, [11] optimizes storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended in [3, 2] to consider storage mappings for a set of schedules, towards the end of capturing the tradeo between parallelism and storage. However, these ....

[Article contains additional citation context not shown here]

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3-4):649-671, May 1998.

....nested loop, our method applies to general ane dependences across statements and loop nests. Moreover, our framework goes beyond AOV s to unify the notion of occupancy vectors with known ane scheduling techniques. Another related approach to storage management for parallel programs is that of [3, 2, 11]. Given an ane schedule, 11] optimizes storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended in [3, 2] to consider storage mappings for a set of schedules, towards the end of capturing the tradeo between parallelism ....

....to general ane dependences across statements and loop nests. Moreover, our framework goes beyond AOV s to unify the notion of occupancy vectors with known ane scheduling techniques. Another related approach to storage management for parallel programs is that of [3, 2, 11] Given an ane schedule, [11] optimizes storage rst by restricting the size of each array dimension and then by combining distinct arrays via renaming. This work is extended in [3, 2] to consider storage mappings for a set of schedules, towards the end of capturing the tradeo between parallelism and storage. However, these ....

[Article contains additional citation context not shown here]

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3-4):649-671, May 1998.

....assignment [12] All these techniques allow partial removal of memory based dependences, but may extract less parallelism than conversion to single assignment form. ffl Applying storage mapping optimization techniques [4] Some of these are either schedule independent [16] or schedule dependent [13] yielding better optimizations whether they require former computation of a parallel execution order (scheduling, tiling, etc. or not. Trying to get the best of both directions is the goal of this paper. Our contribution is to show how these two directions can be combined into a unified ....

....Mapping Optimization This shows the need for a memory usage optimization technique. Storage mapping optimization (SMO) 4] consists in reducing memory usage as much as possible as soon as a parallel execution order has been crafted. The technique presented in [4] extending previous results [16, 13] to general loop nests allows automatic generation of a more memory economical program, see Figure 2.b. A single two dimensional array can be used, while keeping the two outer loops parallel. Run time computation of function OE with array Last seems very cheap at first glance; But execution of ....

[Article contains additional citation context not shown here]

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.

No context found.

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.

....incoming control paths [7] Parallelization via memory expansion thus requires both moderation in the expansion degree, and efficiency in the run time computation of OE functions, especially for non scalar data structures distributed across processors. Along the lines of [20] and especially [16], we present a general storage mapping optimization framework for expansion of imperative programs, applicable to most loop nest parallelization techniques. The paper is organized as follows: Section 2 studies a motivating example showing what we want to achieve. Section 3 introduces the general ....

....optimized storage mapping, in the sense of [20] On many programs, however, a more memoryeconomical technique consists in computing a legal storage mapping according to a given parallel execution order, instead of finding a universal storage compatible with any execution order. This is done in [16] for affine (a.k.a. static control) loop nests only. The contributions of this paper are the following: ffl Present a robust algorithm for partial expansion. It is adapted from the algorithm in [5] with reduced run time overhead, and generalization to unrestricted array subscripts. Conversion ....

[Article contains additional citation context not shown here]

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.

No context found.

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.

No context found.

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3--4):649--671, May 1998.

No context found.

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24:649--671, 1998.

No context found.

Vincent Lefebvre and Paul Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24:649--671, 1998.

No context found.

Vincent Lefebvre and Paul Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3-4):649--671, May 1998. 33

No context found.

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3--4):649--671, May 1998.

No context found.

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Parallel Computing, 24(3--4):649--671, May 1998.

No context found.

V. Lefebvre and P. Feautrier. Automatic storage management for parallel programs. Journal on Parallel Computing, 24:649--671, 1998.

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