Adding just one line in my code slows the GPU.

hello, my first code was

for i = 1:1000
    r=0.8*r+0.2*tanh(A*r+W_in*xc(i+T0)+(0.3)*bias);  
end

In the code above,the gpu is faster than cpu but I have to stack 'r' which is (nx1) matrix, so I just adding 'R' on previous code makes the gpu slower than cpu.

R=zeros(n,1000);
for i = 1:1000
    r=0.8*r+0.2*tanh(A*r+W_in*xc(i+T0)+(0.3)*bias);
    R(:,i)=r;
end

I want to make my code runs fast on gpu than cpu. What should I do??

Answer 1

[Matt J]

Pre-allocate on the GPU. Also, pre-compute things on the GPU that are easily vectorized and don't depend on r.

R=gpuArray.zeros(n,1000);
r=gpuArray(r);
A-gpuArray(A);
increments = W_in*gpuArray( xc((1:1000)+T0)  ) + 0.3*bias;

for i = 1:1000
    r=0.8*r+0.2*tanh(A*r+ increments(i));
    R(:,i)=r;
end

[TAEYOON KIM]

Sorry, I didn't write about pre-allocate. I pre-allocate all elements already. My prblem is when R(:,i)=r part makes for loop slow.

[Matt J]

Still, my other recommendations should help...

[TAEYOON KIM]

Ok thanks a lot!!