Citation Context

...ress computation as a graph with the goal of explicitly separating communication from computation at the programmer interface: neither system addresses problems of heterogeneity or OS support. Sponge =-=[38]-=- is a compilation framework for GPUs using synchronous dataflow (SDF) streaming languages, addressing problems of portability for GPU-side code across different generations of GPUs and CPUs, as well a...

Citation Context

...tion is not flattened. Instead the different levels are mapped to the hierarchy of the GPU, including thread blocks and threads. Several other approaches target GPUs from high-level languages. Sponge =-=[15]-=- generates portable CUDA C code from programs written in StreamIt [23]. Dubach et al. [12] developed a compiler for Lime, a Java-compatible language targeting heterogeneous systems consisting of CPUs ...

Citation Context

...e’s buffer objects encapsulate disjoint memory spaces in a way that is similar to the memory-space transparency provided by Dandelion’s GPU dataflow engine. Flextream [40] is compilation framework for synchronous dataflow models that dynamically adapts applications to FPGA, GPU, or CPU target architectures, and Flextream 12 applications are represented as a graph. All of these works to are essentially complementary to Dandelion, which could use any of these systems as the local GPU dataflow engine; unlike these systems Dandelion does not directly expose the graph-based execution model. Sponge [41] is an SDF language compiler framework that addresses portability challenges for GPUside code across different generations of GPUs and CPUs, as well abstracting hardware details such as memory hierarchy and threading models. Like SDF languages in general [55, 34, 18], scheduling for Sponge is static, while Dandelion cross-compiles for a GPU target on demand and schedules work across available resources dynamically. The Dandelion compiler does not yet focus on optimization of GPU code; the runtime could benefit from implementation of the optimizations from Sponge. Front-end programming models M...

Citation Context

...s W times of the minimum, leading to a throughput a number of factors lower than the peak of GPU [2, 5, 26]. They have been the focused target of some recent studies. However, most prior explorations =-=[2, 12, 18, 20, 25]-=- concentrate on static irregularities, where the memory access patterns are known at compilation time. The type of irregularity in our focus is dynamic: For instance, the content of the indexing array...

Citation Context

...ted work partition between GPU and CPU through compiler and runtime support [17]. In addition, recent years have seen many efforts in optimizing GPU executions through either software (e.g. [8], [9], =-=[15]-=-, [19], [27]–[29]) or hardware (e.g. [13], [18], [23]) approaches. We are not aware of previous studies on optimizing the treatment to implicit synchronizations. X. CONCLUSION In this paper, we presen...

Citation Context

...c thread remapping policy to reduce the impact on performance of branch divergence and non-coalesced accesses. High level program tuning for GPUs has been the target of extensive research too. Sponge =-=[12]-=- is a compiler toolchain for the optimization of streaming programs for GPUs. Dubach et al. [10] propose a set compiler optimizations to improve performance of OpenCL programs generated by the Lime Ja...

Citation Context

.... Unlike our approach, they do not consider the problem of selecting the most suitable device from the host CPU and the GPU to run the code. Programming Frameworks for GPUs Several programming models =-=[10, 11, 15]-=- have been proposed for GPU programming. These approaches provide APIs to develop GPU applications. All these approaches implicitly assume the GPU execution gives the best performance. Optimizing GPU ...

Citation Context

...studies on enhancing GPU memory performance through compiler-assisted techniques. Hormati et al. present a compiler framework to support stream programming on GPU, including some memory optimizations =-=[10]-=-. Zhang et al. propose an online adaptive scheme to enable runtime thread-data remapping to minimize uncoalesced memory accesses [31]. Lee et al. [16] describe some techniques to optimize memory refer...