TY - GEN
T1 - An HSAIL ISA conformed GPU platform
AU - Chen, Heng Yi
AU - Chen, Chung-Ho
AU - Huang, Yun Chi
AU - Hsu, Kuan Chieh
AU - Wang, Chen Chieh
PY - 2016/1/1
Y1 - 2016/1/1
N2 - This paper presents a GPU platform based on the HSA (Heterogeneous System Architecture). The platform contains the OpenCL programming interface, a cycle-level HSAIL GPU simulator, and a finalizer that is able to translate the HSAIL code to our custom binary ISA. The system platform is able to run OpenCL applications, and the results are verified with the real GPU hardware. To improve the scheduling efficiency, we explore warp scheduling policies on our platform including Fair Round- Robin, Loose Round- Robin, Two-Level, and Greedy-Then-Oldest. We found that these algorithms are not optimal in using the load/store units. We propose a Memory Access First (MAF) mechanism which can be adopted to the Two- Level and Greedy-Then-Oldest scheduling policies to reduce the idle time of the load/store units. With the proposed mechanism, the warp scheduler can remove 23 to 34 percent of load/store unit idle time and improve the GPU performance by about 16%.
AB - This paper presents a GPU platform based on the HSA (Heterogeneous System Architecture). The platform contains the OpenCL programming interface, a cycle-level HSAIL GPU simulator, and a finalizer that is able to translate the HSAIL code to our custom binary ISA. The system platform is able to run OpenCL applications, and the results are verified with the real GPU hardware. To improve the scheduling efficiency, we explore warp scheduling policies on our platform including Fair Round- Robin, Loose Round- Robin, Two-Level, and Greedy-Then-Oldest. We found that these algorithms are not optimal in using the load/store units. We propose a Memory Access First (MAF) mechanism which can be adopted to the Two- Level and Greedy-Then-Oldest scheduling policies to reduce the idle time of the load/store units. With the proposed mechanism, the warp scheduler can remove 23 to 34 percent of load/store unit idle time and improve the GPU performance by about 16%.
UR - http://www.scopus.com/inward/record.url?scp=85016782946&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85016782946&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:85016782946
SN - 9781138028937
T3 - Applied System Innovation - Proceedings of the International Conference on Applied System Innovation, ICASI 2015
SP - 243
EP - 247
BT - Applied System Innovation - Proceedings of the International Conference on Applied System Innovation, ICASI 2015
A2 - Lam, Artde Donald Kin-Tak
A2 - Prior, Stephen D.
A2 - Meen, Teen-Hang
PB - CRC Press/Balkema
T2 - International Conference on Applied System Innovation, ICASI 2015
Y2 - 22 May 2015 through 27 May 2015
ER -