Exploring the tradeoffs between programmability and efficiency in data-parallel accelerators
Proceedings of the 38th annual international symposium on Computer architecture
SIMD divergence optimization through intra-warp compaction
Proceedings of the 40th Annual International Symposium on Computer Architecture
Journal of Parallel and Distributed Computing
Exploring the Tradeoffs between Programmability and Efficiency in Data-Parallel Accelerators
ACM Transactions on Computer Systems (TOCS)
A shared matrix unit for a chip multi-core processor
Journal of Parallel and Distributed Computing
Hi-index | 0.00 |
This thesis explores a new approach to building data-parallel accelerators that is based on simplifying the instruction set, microarchitecture, and programming methodology for a vector-thread architecture. The thesis begins by categorizing regular and irregular data-level parallelism (DLP), before presenting several architectural design patterns for data-parallel accelerators including the multiple-instruction multiple-data (MIMD) pattern, the vector single-instruction multiple-data (vector-SIMD) pattern, the single-instruction multiple-thread (SIMT) pattern, and the vector-thread (VT) pattern. Our recently proposed VT pattern includes many control threads that each manage their own array of microthreads. The control thread uses vector memory instructions to efficiently move data and vector fetch instructions to broadcast scalar instructions to all microthreads. These vector mechanisms are complemented by the ability for each microthread to direct its own control flow. In this thesis, I introduce various techniques for building simplified instances of the VT pattern. I propose unifying the VT control-thread and microthread scalar instruction sets to simplify the microarchitecture and programming methodology. I propose a new single-lane VT microarchitecture based on minimal changes to the vector-SIMD pattern. Single-lane cores are simpler to implement than multi-lane cores and can achieve similar energy efficiency. This new microarchitecture uses control processor embedding to mitigate the area overhead of single-lane cores, and uses vector fragments to more efficiently handle both regular and irregular DLP as compared to previous VT architectures. I also propose an explicitly data-parallel VT programming methodology that is based on a slightly modified scalar compiler. This methodology is easier to use than assembly programming, yet simpler to implement than an automatically vectorizing compiler. To evaluate these ideas, we have begun implementing the Maven data-parallel accelerator. This thesis compares a simplified Maven VT core to MIMD, vector-SIMD, and SIMT cores. We have implemented these cores with an ASIC methodology, and I use the resulting gate-level models to evaluate the area, performance, and energy of several compiled microbenchmarks. This work is the first detailed quantitative comparison of the VT pattern to other patterns. My results suggest that future data-parallel accelerators based on simplified VT architectures should be able to combine the energy efficiency of vector-SIMD accelerators with the flexibility of MIMD accelerators. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)