Toru Kisuki, Peter M. W. Knijnenburg, and Michael F. P. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In IEEE PACT, pages 237-- 248, 2000.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... with different orders of applying code improvement phases in an attempt to reduce code size [9, 10] Iterative compilation techniques have been used to determine good phase orderings for specific programs [11] and values for optimization parameters such as loop unroll factors and blocking sizcs [17]. In contrast, vista allows a uscr to intcractivcly specify the order and scope in which code improvement phases are applied. 3 VISTA S OPTIMIZATION ENGINE vista s optimization engine is based on vpo, the Very Portable Optimizer [3, 5] vpo has several properties that make it an ideal starting ....

T. Kisuki, E Knijnenburg, and M. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In Proceedings of the 2000.

.... with different orders of applying code improvement phases in an attempt to reduce code size [9, 10] Iterative compilation techniques have been used to determine good phase orderings for specific programs [18] and values for optimization parameters such as loop unroll factors and blocking sizes [17]. In contrast, vista allows a user to interactively specify the order and scope in which code improvement phases are applied. 3 VISTA S OPTIMIZATION ENGINE vista s optimization engine is based on vpo, the Very Portable Optimizer [3, 5] vpo has several properties that make it an ideal starting ....

T. Kisuki, P. Knijnenburg, and M. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In Proceedings of the 2000.

....system that does not expose parameters for tuning is necessarily suboptimal, because no compiler that takes finite time can ahvays guess the best parmneters for all progrmns. Projects that use parameter searching include PHiPAC [1] ATLAS I291, Sparsity [101 [111, FFTW I7] and OCEANS [14] [15] [21] 16] PHiPAC and ATLAS automatically generate nuinerous variants of matrix multiply or other kernels in an attempt to select the best one for a particular task on a pro titular lnachine. PHiPAC and ATLAS use hand crafted templates for handling edge cases, copying data, preferchina, ....

T. Kisuki, P. M. W. Kuijnenbm'g, and M. F. P. O'Boyle. Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation. Technical Report 2000-07, LIACS, Leiden University, 2000.

....system that does not expose parameters for tuning is necessarily suboptimal, because no compiler that takes finite time can always guess the best parameters for all programs. Projects that use parameter searching include PHiPAC [1] ATLAS [29] Sparsity [10] 11] FFTW [7] and OCEANS [14] [15] [21] 16] PHiPAC and ATLAS automatically generate numerous variants of matrix multiply or other kernels in an attempt to select the best one for a particular task on a particular machine. PHiPAC and ATLAS use hand crafted templates for handling edge cases, copying data, prefetching, selecting ....

T. Kisuki, P. M. W. Knijnenburg, and M. F. P. O'Boyle. Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation. Technical Report 2000-07, LIACS, Leiden University, 2000.

....because no compiler that takes finite time can always guess the best parameters for all programs. Projects that use parameter searching include PHiPAC [7] ATLAS [36] Sparsity (Im [17] Im and Yelick [18] BeBOP [6] 16 FFTW ( 12] and iterative compilation (Kisuki et al. 21] Kisuki et al. [22]; O Boyle et al. 31] PHiPAC and ATLAS automatically generate numerous variants of matrix multiply or other kernels in an attempt to select the best one for a particular task on a particular machine. These systems are instructive for both their absolute success and their relative success. The ....

T. Kisuki, P. M. W. Knijnenburg, and M. F. P. O'Boyle. Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation. Technical Report

No context found.

T. Kisuki, P.M.W. Knijnenburg and M.F.P. O'Boyle, Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation, PACT, 2000.

No context found.

T. Kisuki, P.M.W. Knijnenburg and M.F.P. O'Boyle, Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation, PACT, 2000.

No context found.

T. Kisuki, P. Knijnenburg and M. O'Boyle. Combined selection of tile sizes and unroll factors Using iterative compilation. The 2000.

No context found.

T. Kisuki, P.M.W. Knijnenburg and M.F.P. O'Boyle, Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation, Proc. PACT 2000.

No context found.

T. Kisuki, P.M.W. Knijnenburg, and M.F.P. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In IEEE Press, editor, Proceedings of PACT'2000.

No context found.

T. Kisuki, P.M.W. Knijnenburg and M.F.P. O'Boyle, Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation, PACT, 2000.

....so complex an organization that this approach will likely not deliver optimal code. In order to solve this problem, we have proposed iterative compilation where many variants of the source program are generated and the best one is selected by actually pro ling these variants on the target hardware [7]. This framework is essentially target neutral since it consists of a driver module that navigates through the optimization space and a source to source restructurer that allows the speci cation of the transformations it employs. The native compiler is used as back end compiler and it is treated ....

....consists of a driver module that navigates through the optimization space and a source to source restructurer that allows the speci cation of the transformations it employs. The native compiler is used as back end compiler and it is treated together with the platform as a black box. We have shown [7] that this approach outperform existing static approaches signi cantly, albeit at the price of being extremely time consuming. In this paper, we propose to use static models that cover part of the behavior of the target platform, to estimate the e ect of transformations and to exclude ....

[Article contains additional citation context not shown here]

T. Kisuki, P.M.W. Knijnenburg, and M.F.P. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In Proc. PACT, pages 237-246, 2000.

....space considered must be highly restricted. Furthermore, due to architectural evolution, any optimisation strategy is swiftly made redundant. This problem is becoming more severe with each new generation of processor. This paper examines another approach based on iterative compilation [7, 9]. Different transformations are applied, corresponding to points in the transformation space, and their worth evaluated by executing the program. Several evaluations, based on a compiler search strategy, are made upto a certain pre defined limit with the compiler selecting the best one. Although ....

T. Kisuki, P.M.W. Knijnenburg and M.F.P.O'Boyle, Combined selection of tile sizes and unroll factors using iterative compilation. Proc. PACT2000, pages 237-246, 2000.

.... given application and platform [1, 9, 8, 18] In particular, the present authors have proposed an optimization framework called iterative compilation in which many variants of the source program are generated and the best one is selected by actually profiling these variants on the target hardware [11]. This framework is essentially target neutral since it consists of a driver module that navigates through the optimization space and a source to source restructurer that allows the specification of the transformations it employs. The native compiler is used as back end compiler and it is treated ....

....since it consists of a driver module that navigates through the optimization space and a source to source restructurer that allows the specification of the transformations it employs. The native compiler is used as back end compiler and it is treated together with the platform as a black box. In [11] we have shown that this approach can obtain high levels of optimization, outperforming existing approaches significantly. In some cases, improvements of several hundreds of percents are reached. However, we need many profiles for this purpose and compilation time, that includes searching, program ....

[Article contains additional citation context not shown here]

T. Kisuki, P.M.W. Knijnenburg, and M.F.P.O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In Proc. PACT, 2000. To appear. www.liacs.nl/peterk.

....is the entire space. We take a number of samples and order them in a priority queue. Around the best points we define a smaller window. Random search We randomly generate parameters. We conducted a number of experiments to establish good values for the different parameters in the algorithms [10]. 3.2. Benchmarks and Platforms In order to assess the efficiency of iterative compilation for selecting tile sizes and unroll factors, we use many small kernel benchmarks from multimedia applications that exhibit a wide variety of memory access behavior. In this way, we are able to give a ....

....on every platform. Instead, we have considered a total of 82 different experiments. 4. Results In this section we show how much speedup we obtain as a function of the number of iterations, where we show the best speedup found so far. Due to space limitations we only show a few examples. See [10] for a more complete discussion. 4.1. Rectangular Tile Sizes In this section we discuss rectangular tile sizes together with unroll factors. In this case, the search space consists of 20 100 100 = 200; 000 points. We let the search algorithm run for 2000 iterations. The results are given in ....

[Article contains additional citation context not shown here]

T. Kisuki, P.M.W. Knijnenburg, and M.F.P. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. Technical Report 2000-07, LIACS, Leiden University, 2000. Available through www.liacs.nl/peterk.

....to a new architecture while achieving a high level of optimization. What is required is a method where the compiler can receive dynamic feedback regarding its performance and modify its behavior accordingly. To achieve this goal we have introduced a new compilation approach, iterative compilation [13]. In this paper, we incorporate cache models into the iterative compilation system to achieve a high level of optimization of both compiler efficiency and code performance. In this approach, many versions of program with different tile sizes and unrolling factors are generated and their value is ....

....times as well as model information to decide upon the optimal unrolling Driver List of Transformations MT1 Compiler TDL Files F77 Transformed Program Execution Time SSL File Target Platform Cache Model Figure 1: The Compilation Process factor and loop tiling sizes. In previous work [13] we have shown that iterative compilation is powerful approach capable of outperforming existing tile size selection algorithms and reaching a high level of optimization with reasonable compilation time. We have compared the obtained speedups relative to static tile size selection algorithms: the ....

[Article contains additional citation context not shown here]

T. Kisuki, P.M.W. Knijnenburg, and M.F.P. O'Boyle. Combined Selection of Tile Sizes and Unroll Factors Using Iterative Compilation. To appeare in Proc. PACT2000.

.... II 0 20 40 60 80 100 0 5 10 15 20 0 5 10 15 20 25 30 Tile Size Unroll Time UltraSparc Figure 1: Execution Time MxM for Unrolling and Tiling Iterative compilation, based on the selection of high level transformations, has been shown to work across a range of architectures [13] and although preliminary work has shown this approach to be highly e ective [14] the number of executions needed to nd a good program may be prohibitively expensive. We therefore also consider how additional information may be used to guide the search strategy and its e ect on both compiler ....

....evaluated so far and decides which transformations have to be applied next using a search algorithm to steer through the optimization space. We have implemented several search algorithms, including a Genetic Algorithm, Simulated Annealing, Pyramid Search, Window Search and Random Search [13]. In this paper we have included cache models in the driver that are used to decide whether or not to execute the 5 Driver List of Transformations MT1 Compiler TDL Files F77 Transformed Program Execution Time SSL File Target Platform Cache Model Figure 2: The Compilation Process ....

[Article contains additional citation context not shown here]

T. Kisuki, P.M.W. Knijnenburg, and M.F.P. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. Technical Report 2000-07, LIACS, Leiden University, 2000. Submitted to PACT2000.

No context found.

Toru Kisuki, Peter M. W. Knijnenburg, and Michael F. P. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In IEEE PACT, pages 237-- 248, 2000.

No context found.

Kisuki T, Knijnenburg PMW,O'Boyle MFP. Combined selection of tile sizes and unroll factors using iterative compilation. Proceedings of PACT'2000, Parallel Architectures and Compiler Technology. IEEE Press: Philadelphia, PA, 2000; 237--246.

No context found.

T. Kisuki, P. Knijnenburg, and M. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In Parallel Architectures and Compilation Techniques, pages 237--246, 2000.

No context found.

Toru Kisuki, Peter M. W. Knijnenburg, and Michael F. P. O'Boyle. Combined selection of tile sizes and unroll factors using iterative compilation. In IEEE PACT, pages 237-- 248, 2000.

No context found.

P. M. W. Knijnenburg, T. Kisuki, and M. F. P. O'Boyle, "Combined selection of tile sizes and unroll factors using iterative compilation," J. of Supercomputing, vol. 24, no. 1, pp. 43--67, 2003.

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