Citation Context

...on representative inputs, dynamically observe the memory access patterns, then use the observed access patterns to suggest potential parallelizations that do not violate the observed data dependences =-=[14, 37, 42]-=-. The dynamic analysis may be augmented with a static analysis to recognize parallel patterns such as reductions. These potential parallelizations are typically presented to the developer for approval...

Citation Context

...pared Kremlin’s plans against those parallelized by humans in the SPEC OMP2001 versions. For art and ammp, SPEC OMP versions benefit from serial optimizations compared to their SPEC 2000 counterparts =-=[38]-=-. To exclude the effect of serial optimizations, we applied those optimizations on the SPEC 2000 code before running Kremlin. Our evaluation included only third-party benchmarks that have preexisting ...

Citation Context

... sequences of optimization passes or polyhedral transformations [59, 78]. We started combining Milepost technology with machine-learning based autoparallelization and predictive scheduling techniques =-=[60,51,76]-=-. We have also started investigating staged compilation techniques to balance between static and dynamic optimizations using machine learning in LLVM or Milepost GCC4CIL connected to Mono virtual mach...

Citation Context

...nce all memory addresses are resolved in runtime. Data-dependence profiling has already been used in parallelization efforts like speculative multithreading [4, 7, 17, 23, 29] and finding parallelism =-=[15, 22, 25, 27, 31]-=-. It is also being employed in the commercial tools we mentioned. However, the current algorithm for data-dependence profiling incurs significant costs of time and memory overhead. Surprisingly, altho...

Citation Context

...mization, dynamic data partitioning and predictive scheduling, empirical iterative compilation, statistical analysis, machine learning and decision trees together with program and 25dataset features =-=[50, 62, 56, 51, 47, 60, 45]-=-. More information about collaborative UNIDAPT R&D is available at [30]. 5 Usage Scenarios 5.1 Manual sharing of optimization cases Collective tuning infrastructure provides multiple ways to optimize ...

Citation Context

...zation of loops, where shadow locations are used to track dynamic dependences across loop iterations [22]. Several other approaches of similar nature have been investigated in more recent work (e.g., =-=[2, 19, 30, 31, 35, 39]-=-). The efficient collection of dynamic data-dependence information has also been explored by previous work. Tallam et al. use runtime control flow information to reconstruct significant portions of th...

Citation Context

...plex induction variables (such as iterators through a linked list) are difficult to detect automatically. Second, even if dependences are precisely identified (e.g., using dynamic dependence analysis =-=[43]-=-, program annotations [45], or speculative parallelization [11, 15, 21, 27, 33, 39, 44]), there remain many programs in which memory dependences are accidental artifacts of the implementation and shou...

Citation Context

...on representative inputs, dynamically observe the memory access patterns, then use the observed access patterns to suggest potential parallelizations that do not violate the observed data dependences =-=[34, 30, 10]-=-. These potential parallelizations are then typically presented to the developer for approval. It is possible to use QuickStep in a similar way, to explore potential parallelizations that are then cer...

Citation Context

...sion, the program is then passed to the second stage which is responsible for mapping tasks to both GPUs and CPUs. The number of tasks to be mapped to GPUs and CPUs is determined by the predictor. In =-=[24]-=-, the authors proposed a two-staged parallelization approach combining profiling-driven parallelism detection and ML-based mapping to generate OpenMP annotated parallel programs. In this method, first...