Determining the best set of optimizations to apply to a kernel to be executed on the graphics processing unit (GPU) is a challenging problem. There are large sets of possible optimization configurations that can be applied, and many applications have multiple kernels. Each kernel may require a specific configuration to achieve the best performance, and moving an application to new hardware often requires a new optimization configuration for each kernel.In this work, we apply optimizations to GPU code using HMPP, a high-level directive-based language and source-tosource compiler that can generate CUDA / OpenCL code. However, programming with high-level languages may mean a loss of performance compared to using low-level languages. Our work shows that it is possible to improve the performance of a high-level language by using auto-tuning. We perform auto-tuning on a large optimization space on GPU kernels, focusing on loop permutation, loop unrolling, tiling, and specifying which loop(s) to parallelize, and show results on convolution kernels, codes in the PolyBench suite, and an implementation of belief propagation for stereo vision. The results show that our auto-tuned HMPP-generated implementations are significantly faster than the default HMPP implementation and can meet or exceed the performance of manually coded CUDA / OpenCL implementations.

No abstract

a b s t r a c tWe present a mathematical framework for constructing and analyzing parallel algorithms for lattice kinetic Monte Carlo (KMC) simulations. The resulting algorithms have the capacity to simulate a wide range of spatio-temporal scales in spatially distributed, non-equilibrium physiochemical processes with complex chemistry and transport micro-mechanisms. Rather than focusing on constructing exactly the stochastic trajectories, our approach relies on approximating the evolution of observables, such as density, coverage, correlations and so on. More specifically, we develop a spatial domain decomposition of the Markov operator (generator) that describes the evolution of all observables according to the kinetic Monte Carlo algorithm. This domain decomposition corresponds to a decomposition of the Markov generator into a hierarchy of operators and can be tailored to specific hierarchical parallel architectures such as multi-core processors or clusters of Graphical Processing Units (GPUs). Based on this operator decomposition, we formulate parallel Fractional step kinetic Monte Carlo algorithms by employing the Trotter Theorem and its randomized variants; these schemes, (a) are partially asynchronous on each fractional step time-window, and (b) are characterized by their communication schedule between processors.The proposed mathematical framework allows us to rigorously justify the numerical and statistical consistency of the proposed algorithms, showing the convergence of our approximating schemes to the original serial KMC. The approach also provides a systematic evaluation of different processor communicating schedules. We carry out a detailed benchmarking of the parallel KMC schemes using available exact solutions, for example, in Isingtype systems and we demonstrate the capabilities of the method to simulate complex spatially distributed reactions at very large scales on GPUs. Finally, we discuss work load balancing between processors and propose a re-balancing scheme based on probabilistic mass transport methods.

As computation becomes increasingly limited by data movement and energy consumption, exploiting locality throughout the memory hierarchy becomes critical to continued performance scaling. Moving computation closer to memory presents an opportunity to reduce both energy and data movement overheads. We explore the use of 3D die stacking to move memory-intensive computations closer to memory. This approach to processing in memory addresses some drawbacks of prior research on in-memory computing and is commercially viable in the foreseeable future.Because 3D stacking provides increased bandwidth, we study throughput-oriented computing using programmable GPU compute units across a broad range of benchmarks, including graph and HPC applications. We also introduce a methodology for rapid design space exploration by analytically predicting performance and energy of in-memory processors based on metrics obtained from execution on today's GPU hardware. Our results show that, on average, viable PIM configurations show moderate performance losses (27%) in return for significant energy efficiency improvements (76% reduction in EDP) relative to a representative mainstream GPU at 22nm technology. At 16nm technology, on average, viable PIM configurations are performance competitive with a representative mainstream GPU (7% speedup) and provide even greater energy efficiency improvements (85% reduction in EDP).