different result between gpuArray / matrix

2 次查看（过去 30 天）

emar 2017-5-2

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/338388-different-result-between-gpuarray-matrix

评论： Joss Knight 2017-5-13

Hello, I tried to compare the result between two codes two inputs with the same function:

function local_sum_A = local_sum(A,m,n)
B = padarray(A,[m n]);
s = cumsum(B,1);
c = s(1+m:end-1,:)-s(1:end-m-1,:);
s = cumsum(c,2);
local_sum_A = s(:,1+n:end-1)-s(:,1:end-n-1)

It's a function that i found on Matlab. When i compile :

A=rand(20);
s=local_sum(A,21,22);
gpuDevice;
A2=gpuArray(A);
s2=local_sum(A2,21,22);

I have two different results ( difference by a factor 1e-13), but I do a lot of calculations after, so the difference become important (1e-02).

Could you tell me why there is such difference ?

Thank you in advance

2 个评论
显示无隐藏无

Adam 2017-5-2

gpu does calculations differently to the cpu so there will be different levels of accuracy. 1e-13 is insignificant, but the more calculations you do certainly the more the error will inevitably accumulate.

Joss Knight 2017-5-13

To be clear, the error here is between the computed result and the theoretically correct result, not between the CPU result and the GPU result. The CPU result is no more correct than the GPU one. All floating point computation has error, and this error always accumulates the more operations you do.

请先登录，再进行评论。

请先登录，再回答此问题。