Strange output data with double about cudnn-python-wrappers

[kashif]

@Felix-neko as far as I can see... it seems to be cudnn bug... I cannot see anything wrong that we are doing in calling the APIs... I will keep looking though...

[Felix-neko]

Oy vey!

[crying]

[kashif]

another thing to check might be the workSpaceSizeInBytes... perhaps you can try to explicitly set it and try if that helps?

[kashif]

@Felix-neko also note that when beta = 0 the output is not read and can contain any uninitialized data (including Nan numbers)...

[Felix-neko]

@kashif

But AFAIK, beta=0 means that the old data, whatever it was, should be totally replaced with new data. Isn't it?

AFAIK, what you said is alpha=0.

[kashif]

@Felix-neko true yes you are right... 💣

[kashif]

yes @Felix-neko so then can you kindly try to create a cudnnGetConvolutionForwardAlgorithm() with the type: CUDNN_CONVOLUTION_FWD_NO_WORKSPACE and use that?

[Felix-neko]

I'll try to do it when get to work. Alas, I don't have all the software installed at home.

[kashif]

@Felix-neko I just tried it... but no luck... same behaviour 👎

[philvdm]

I think that alpha and beta are not properly handled in libcudnn_rc2.py
alpha and beta must be passed per reference but that reference MUST be of proper type (e.g double if double type is chosen). From what I see, alpha,beta are always passed as float pointers

[kashif]

@philvdm yes I thought so too, so I explicitly passed double for the double case but it did not help... the API takes just a pointer to the alphas and betas and the docs do not mention their types explicitly....

I have also emailed nvidia with this behaviour as well...

[philvdm]

I am from nvidia ;-) and I can confirm that alpha/beta needs to be of the same type than the tensor.

As a test, just pass Alpha=0 and beta=1, your tensor should be unchanged.

[kashif]

@philvdm ah cool thanks for the info! 🍰 I will try it out...

[kashif]

@philvdm 🎉 cool it works! I will fix up the bindings to set the appropriate ctype depending on the datatype of the tensor/filter...

[hannes-brt]

Thank you very much @kashif, @philvdm, and @Felix-neko for solving this. I just checked again and it works for me too now.

[Felix-neko]

@hannes-brt
Can you also update your cudnn2 branch?

[hannes-brt]

Just use the master branch now, please. All cudnn R1 is backed up on the cudnn1 branch now, so master will have the bleeding edge R2 code.

[Felix-neko]

@hannes-brt
Great! Thank you, comrade!
^__^