计算机系统结构03(向量处理_互联网络_阵列机)(北邮课件).pdf

计算机系统结构 （向量处理＋互联网络＋阵列机） 北京邮电大学计算机科学与技术学院 邝坚 2009年3月 Vector process principles Vector – a set of scalar data items, all of same type, stored in memory. A (a , a , a ,..., a ) 1 2 3 n Vector processor – an ensemble of HW resource, including vector registers, functional pipelines, processing elements, and register counters, for performing vector operations. Vectorization – the conversion from scalar code to vector code. Vectorization Y = a x X LD F0,a ;F0 ← a LV V1,M(X) ;V1 ← X Unrolled-loop-level MULV V2,F0,V1 ;V2 ← a * X parallelism SV M(Y),V2 ;Y ← V2 用标量指令集如何实现？ Vectorization 计算：ci ai+5 +bi ;i 10,11,12,...,1000 FORTRAN(标量) DO 40 I = 10,1000 40 C(I) = A(I+5) + B(I) FORTRAN(向量) C(10:1000) = A(10+5:1000+5) + B(10:1000) Vector instruction types Vector-vector instructions V →V ;V sin(V ) 1 2 2 1 V op V →V ;V V +V j k i 3 1 2 Vector-scalar instructions s op V →V ;V s ×V k i 2 1 Vector-memory instructions M →V ;Vector load V →M ;Vector store Vector instruction types Vector-reduction instructions V →s ;max, min, sum, mean value i j n V op V →s ;dot product, s ∑a ×b j k i