新型異構(gòu)并行計(jì)算機(jī)上的數(shù)據(jù)傳輸與程序設(shè)計(jì)陳一峯北京大學(xué)-資料下載頁
2024-10-24 15:42本頁面
【導(dǎo)讀】Tsubame：3GPU/2CPUs. :6GPUs/2CPUs. Dimensionsinatree. memcpy(b+i*8+j*2,a+i*2+j*8,float*a,*b;#insertDataTransfer(a,A,float*a,*b;INIT_GPU($tid$);#insertDataTransfer(a,A,]WIDTH[A]WIDTH[A10]WIDTH[B]WIDTH[B10]WIDTH[C]WIDTH[C10. ]32[A]32/WIDTH[A]16[A]16/WIDTH[A11100100. ]256[B]256/WIDTH[B]32[B]32/WIDTH[B11100100. ]256[C]256/WIDTH[C]16[C]16/WIDTH[C11100100. ]32[T]4[T]256/WIDTH[T]16/WIDTH[T11100100
【正文】 0?????? ???????????? ?????? ]2[C]128[C]256/W I D T H[ C]4[C]4[C]16/W I D T H[ C 1111101001101000???????????? ]32[T]4[T]256/W ID T H[ T]16/W ID T H[ T 11100100620Gflops on Fermi C1060 Large FFT(ICS 10, PPoPP 12) Direct Simulation of Turbulent Flows ? Scale ? 12 distributed arrays 128TB ? Entire Tianhe1A with 7168 GPUs ? Progress ? 4096 3D pleted ? 8192 3D halfway ? and 14336 3D tested for performance. ? Software Technologies ? PARRAY (ACM PPoPP’ 12) code only 300 lines. Discussions ?Performance transparency: macros are piled out. ?Completeness: any index expressions using add/mul/mod/div/fp ?Regular structures from applications and target manycore hardware ?Irregular structures allowed but better supported by other tools ?Typical training = 3 days ?Release in parallelarray
教學(xué)課件相關(guān)推薦
鄂ICP備17016276號-1