Question

我需要尽快计算矩阵查找（见下文）。

totquadpoint和nbval.Wepast是每个大约200个固定标量（向量flnewcumyieldpos和flneweta1pos的维度）。 nexttsteppenal是一个3D矩阵（尺寸至少为80 * 80 * 80）

由于我需要多次执行此操作（此双循环本身处于三重嵌套循环中），我需要快速完成。有没有办法让它比下面的内容更快？如果在Matlab中不可能，如果我尝试在Java / C中编写那段代码，我可以期待什么样的增益？

--------代码：-----------------

for k=1:totquadpoint
    nexttsteppenalYA1dd=nexttsteppenal(flnewcumyieldpos(k),flneweta1pos(k),:);
    nexttsteppenalYA1du=nexttsteppenal(upflnewcumyieldpos(k),flneweta1pos(k),:);
    nexttsteppenalYA1ud=nexttsteppenal(flnewcumyieldpos(k),upflneweta1pos(k),:);
    nexttsteppenalYA1uu=nexttsteppenal(upflnewcumyieldpos(k),upflneweta1pos(k),:);

    flnewportvalposvec=flnewportvalpos(k,:);
    upflnewportvalposvec=upflnewportvalpos(k,:);

    for we=1:nbval.Wepast                                                
        penddd(k,we)=nexttsteppenalYA1dd(flnewportvalposvec(we));
        penddu(k,we)=nexttsteppenalYA1dd(upflnewportvalposvec(we));
        pendud(k,we)=nexttsteppenalYA1du(flnewportvalposvec(we));
        penduu(k,we)=nexttsteppenalYA1du(upflnewportvalposvec(we));

        penudd(k,we)=nexttsteppenalYA1ud(flnewportvalposvec(we));
        penudu(k,we)=nexttsteppenalYA1ud(upflnewportvalposvec(we));
        penuud(k,we)=nexttsteppenalYA1uu(flnewportvalposvec(we));
        penuuu(k,we)=nexttsteppenalYA1uu(upflnewportvalposvec(we));

    end
end

Answer 1

此：

for w=1:n
    p(k,w)=d(f(w));
end

应该是：

p(k,1:n)=d(f(1:n));

甚至更好，当n是f的长度时：

p(k,:)=d(f);

对于外部循环，您可能需要sub2ind()。

Answer 2

以上是上述代码的可读版本，其中包含表格的模拟值。我的问题是如何在最后优化两个嵌套for循环的速度（我需要运行这个循环可能一百万次）。

quadpoints=800;
nbWe=100;

penddd=zeros(quadpoints,nbWe);
penddu=zeros(quadpoints,nbWe);
pendud=zeros(quadpoints,nbWe);
penduu=zeros(quadpoints,nbWe);
penudd=zeros(quadpoints,nbWe);
penudu=zeros(quadpoints,nbWe);
penuud=zeros(quadpoints,nbWe);
penuuu=zeros(quadpoints,nbWe);

nsp=rand(100,50,100);

ypos=min(sort(poissrnd(50,quadpoints,1),'ascend'),100);
uypos=min(ypos,100);
etapos=min(sort(poissrnd(25,quadpoints,1),'ascend'),50);
uetapos=min(etapos,50);
Vpos=min(sort(poissrnd(50,quadpoints,nbWe),'ascend'),100);
uVpos=min(Vpos,100);

for k=1:quadpoints
     nspdd=nsp(ypos(k),etapos(k),:);
     nspdu=nsp(uypos(k),etapos(k),:);
     nspud=nsp(ypos(k),uetapos(k),:);
     nspuu=nsp(uypos(k),uetapos(k),:);
     dVvec=Vpos(k,:);
     uVvec=uVpos(k,:);

      for we=1:nbWe
     penddd(k,we)=nspdd(dVvec(we));
         penddu(k,we)=nspdd(uVvec(we));
         pendud(k,we)=nspdu(dVvec(we));
         penduu(k,we)=nspdu(uVvec(we));
         penudd(k,we)=nspud(dVvec(we));
         penudu(k,we)=nspud(uVvec(we));
         penuud(k,we)=nspuu(dVvec(we));
         penuuu(k,we)=nspuu(uVvec(we));

      end
end

快速矩阵查找

2 个答案: