A methodology pruning the search space of six compiler transformations by addressing them together as one problem and by exploiting the hardware architecture details

Abstract

Today's compilers have a plethora of optimizations-transformations to choose from, and the correct choice, order as well parameters of transformations have a significant/large impact on performance; choosing the correct order and parameters of optimizations has been a long standing problem in compilation research, which until now remains unsolved; the separate subproblems optimization gives a different schedule/binary for each sub-problem and these schedules cannot coexist, as by refining one degrades the other. Researchers try to solve this problem by using iterative compilation techniques but the search space is so big that it cannot be searched even by using modern supercomputers. Moreover, compiler transformations do not take into account the hardware architecture details and data reuse in an efficient way.In this paper, a new iterative compilation methodology is presented which reduces the search space of six compiler transformations by addressing the above problems; the search space is reduced by many orders of magnitude and thus an efficient solution is now capable to be found. The transformations are the following: loop tiling (including the number of the levels of tiling), loop unroll, register allocation, scalar replacement, loop interchange and data array layouts. The search space is reduced a) by addressing the aforementioned transformations together as one problem and not separately, b) by taking into account the custom hardware architecture details (e.g., cache size and associativity) and algorithm characteristics (e.g., data reuse).The proposed methodology has been evaluated over iterative compilation and gcc/icc compilers, on both embedded and general purpose processors; it achieves significant performance gains at many orders of magnitude lower compilation time.

Exploiting Historical Data: Pruning Autotuning Spaces and Estimating the Number of Tuning Steps

“…As a result of the parameters being designed by expert programmers, most of these search spaces are relatively small compared to some other autotuning works 12,13 -the number of tuning parameters is not needlessly excessive, and most parameters only have a few rationally-selected values to choose from. While such spaces are hard to search using methods that partially rely on local search (such as simulated annealing), because most of their dimensions are very small and discontinuous, they also contain a higher proportion of relatively fast configurations, which makes random search a viable strategy.…”

A Survey on Compiler Autotuning using Machine Learning

“…GNU compiler collection (GCC) is the GNU compiler and toolchain project which supports various high-level languages, e.g., C, C++, etc.. "The Free Software Foundation (FSF) distributes GCC under the GNU General Public License (GNU GPL). GCC has played an important [3, 11, 15, 21, 28, 39, 40, 50, 61, 79, 83, 91, 92, 99, 101, 105, 107, 108, 112, 116, 122, 132, 135, 153, 155, 157, 160, 166, 169, 183, 189, 195-198, 200, 202, 205, 206, 208-210, 221, 230, 235, 240, 248, 253, 259, 270, 271] LLVM [14, 16, 18, 19, 21, 28, 29, 41, 72, 111, 112, 170, 171, 184-187, 199, 202, 212] Intel-ICC [44,61,91,112,135,136,160,198,200,202,209,221,240,270] JIT Compiler [52-55, 123, 126, 151, 152, 202, 213, 217, 221, 235] Java Compiler [52-55, 123, 126, 151, 213] Polyhedral Model [43, 44, 199, 202, 208-210, 249, 258, 270, 271] Others [3, 6, 10, 11, 16, 28, 36, 51, 58, 67, 69, 70, 78, 91, 96, 98-101, 105, 106, 142, 146, 151, 152, 162, 162, 163, 170, 171, 178, 201, 202, 213, 216-219, 235-238, 249, 250, 259]…”

Exploiting historical data: Pruning autotuning spaces and estimating the number of tuning steps

“…As a result of the parameters being designed by expert programmers, most of these search spaces are relatively small compared with some other autotuning works 12,13 —the number of tuning parameters is not needlessly excessive, and most parameters only have a few rationally selected values to choose from. While such spaces are hard to search using methods that partially rely on local search (such as simulated annealing), because most of their dimensions are very small and discontinuous, they also contain a higher proportion of relatively fast configurations, which makes random search a viable strategy.…”

“…Several works in the field of compiler autotuning have successfully tackled both the problem of optimization prediction 23 and the problem of pruning tuning spaces 13 . However, these compiler‐oriented works focus on an entirely different aspect of performance optimization than our code‐level autotuning framework, and use different methods, such as optimization clustering, to reduce their search spaces—it is therefore impossible to compare the results of these works with our own results.…”