Hi wonghang, Sorry to pester you with emails, but I have some interesting
timing information.
I ran a process using different processors and ways of computing Cholesky()
The results are surprising.
GpuMagmaCholesky()9.0 sec
slinalg.Cholesky(uses cusolver) 2.9 sec
CPU
Hi all,
I found that the cholesky factorization unit test no longer works.
The value returned are completely wrong. It looks like a memory error.
I checked if I skip tril call, the value returned by cuSOLVER is correct.
There should be something wrong in libgpuarray
Hi Wong Hang,
Yes, that's what I saw, the errors started near the end of the matrix.
After that, the numbers appeared random.
I'll try the older version and let you know what I find,
Paul
On Friday, February 7, 2020 at 3:18:23 PM UTC+1, Wong Hang wrote:
>
> I suddenly get the HEAD version of
I suddenly get the HEAD version of libgpuarray works
I found that if I increase the size of the matrix, the error will appear.
The first few rows of the matrix are correct, and then there will be errors
for the remaining rows.
I guess there is a synchronization or memory bug.
$ python3 cho.py
row
I am quite sure I once get correct result even when the matrix is of size
>1000.
Let me do more research and test later and get back to you.
Wong Hang 於 2020年2月7日 週五 下午11:18寫道:
> I suddenly get the HEAD version of libgpuarray works
> I found that if I increase the size of the matrix, the error
Hi Paul,
I think I fixed the issue. Please check the PR
https://github.com/Theano/libgpuarray/pull/589
and you can try to use my branch of libgpuarray to see if it works for you.
For your implementation of MagmaCholesky, you can add profile = True in
~/.theano.rc to see what is the bottleneck of