Transactional Memory and Language Support for General Purpose Graphics Processors

最近的技术变革给加拿大的软件开发者带来了迫在眉睫的挑战。不断增加的功耗阻碍了计算机核心CPU时钟频率的大幅提高。虽然以前满足通常未声明的软件性能目标只是等待下一代CPU进入市场的问题，但软件开发人员现在越来越多地被迫为更节能的计算设备编写软件，这些设备可以明确地并行处理任务。这些设备以“多核”图形处理单元（gpu）为例，这些图形处理单元在今天销售的大多数个人电脑中都能找到。从根本上说，gpu可以提供比cpu（甚至是“多核”cpu）效率高一个数量级的计算，因为它们将更多的硅资源用于实际的计算工作，而不是发现哪些任务可能是并行的。因此，世界各地的大公司正在安装数量迅速增长的“GPU计算”集群，GPU制造商（如加拿大的ATI/AMD）已经开始提供编程接口，使编写软件更容易使用GPU进行非图形计算。然而，为gpu编写软件仍然具有挑战性，因为软件开发人员需要表达并行性并根据硬件的复杂行为调整他们的代码。我们的研究从三个方面解决了这些问题：通过开发硬件和软件，为软件开发人员提供更可预测的性能；通过直接在GPU硬件中支持称为事务性内存的编程模型；并且通过开发编程语言支持来简化重构代码的过程，或者完全自动化重构过程。这些变化将使gpu在现有和未来的应用领域中更容易使用。我们的研究将直接使加拿大GPU制造商和软件公司受益，这些公司面临着向并行计算系统转变所带来的挑战。

Recent shifts in technologyare creating a looming challenge for Canadian software developers. Increasing power dissipation is preventing substantial increases in clock frequencyfor the CPU at the heart of computers. While previouslymeeting often unstated software performance objectives was simplya matter of waiting for the next CPU generation to reach the market, software developers are now increasinglybeing forced to write software for the class of more power-efficient computing devices which process tasks explicitlyin parallel. These devices are exemplified bythe "manycore" Graphics Processing Units (GPUs) found in most personal computers sold today. Fundamentally, GPUs can offer an order of magnitude more efficient computation than CPUs (even "multicore" CPUs) because theydevote more silicon resources to the actual job of computation rather than to discovering what tasks might be parallel. Consequently, a rapidlygrowing number of "GPU compute" clusters are now being installed in large companies around the world, and GPU manufacturers (such as ATI/AMD in Canada) have begun providing programming interfaces to make it easier to write software to use GPUs for non-graphics computations. However, writing software for GPUs remains challenging due to the need for software developers to express parallelism and tune their code for the complex behavior of the hardware. Our research addresses these problems in three ways: bydeveloping hardware and software for delivering more predictable performance to software developers; bysupporting a programming model called transactional memorydirectlyin GPU hardware; and bydeveloping programming language support to ease the process of restructuring code, or automating the restructuring altogether. Together these changes will make GPUs easier to use in existing and future application domains. Our research will directlybenefit Canadian GPU manufacturers and software companies faced with meeting the challenges resulting from the shift towardsparallel computing systems.