GPU并行计算研讨.ppt

计算π的例子 Open Computing Language Open Computing Language 2 GFOR循环 GFOR/GEND循环体内语句无先后顺序，各自独立地在GPU的流处理器上执行。 尽量将M代码向量或矩阵化。 GFOR Supported Functions：/wiki/index.php/Category:GFOR Clear; n=1000000; x=grand(n,5); a=grand(1,5); d=gzeros(n,1); tic; gfor ii=1:n d(ii,1)=sum(abs(a-x(ii,:)).*abs(a-x(ii,:))); gend; Toc Elapsed time is 0.069556 seconds. clear; n=1000000; x=rand(n,5); a=rand(1,5); d=zeros(n,1); tic; for ii=1:n d(ii,1)=sum(abs(a-x(ii,:)).*abs(a-x(ii,:))); end; toc Elapsed time is 2.115929 seconds. 加速比= 2.115929 /0.069556 =30.4205 Open Computing Language Open Computing Language Open Computing Language Open Computing Language 3 GCOMPILE 通过GCOMPILE可定义一个GPU的kernel函数。 clc; clear all; square1=gcompile(sumsquare.m); n=10; input1 = grand(n,1); input2=gones(n,1); result =c(input1,input2); result = arrayfun(@square1(), input1,input2); clc; clear all; n=10; input1 = grand(n); input2=gones(n); result =arrayfun(@sumsquare,input1,input2) clc; clear; tic; sums = gcompile(verbatim); %{ function s= sumsquare(x,y) x = x.^2 y = y.^ 2; s=x+y; %} A=gones(10000); B=grand(10000); AA=sums(A,B); toc function s= sumsquare(x,y) x = x.^2 y = y.^ 2; s=x+y; s=sqrt(s); Jaket DLA Jacket DLA NORM CHOL LU SVD QR MLDIVIDE DET MPOWER INV EIG MPOWER LINSOLVE Jacket SLA MTIMES TRANSPOSE, CTRANSPOSE BICG, BICGSTAB, BICGSTABL CGS GMRES LSQR QMR TFQMR 4 Jacket函数库 单双精度线性代数函数库，图形处理、信号处理、统计函数库定义一个GPU的kernel函数。 Open Computing Language Open Computing Language 三、JACKET实例 clear all;clc; X=load(Iris-virginica.txt); X=[X;X;X;X;X];X=[X;X;X;X;X]; Num=size(X,1); rate=0.65; RandInteger=randperm(Num); index_trainX=RandInteger(1:fix(Num*rate)); index_trainX=index_trainX; trainX=X(index_trainX,1:end-1); trainclassY=X(index_trainX,end); index_testZ=RandInteger(fix(Num*rate)+1:end); index_testZ=index_testZ; testZ=X(index_testZ,1:end-1); testZY=X(index_testZ,end); N=size(testZY,1); K=30; type=2norm; result=gzeros(N,1,K); trainX=gsingle(trainX); testZ=gsingle(testZ); tic; gfor k=1:K result(: