Hi Sai, sorry of the disturbance, I was reading your post and I tried to compile the theano.test() with icpc and the flags that you mentioned -O2 -mkl but I was always getting errors and warning about compiler optimization being disabled. Did you got any warnings or errors when you compiled with icc/icpc? I tried the above on a OS X.
Thanks! On Monday, January 19, 2015 at 5:12:13 PM UTC, sai rajeshwar wrote: > > Hi Fred, Pascal, > > I was testing the PR you 've created, here are the observations. > > I was successfully able to compile with intel compilers and bit of > improvement in the training speed, with flags -O2, -mkl. > However I could not offload it onto MIC(xeon phi) as i expected, tried > many things but somehow couldnot. But it should be possible to > automatically offload python code onto MIC. > > The only thing I can see now is that there is a lower bound on the > matrix size before offloading happens. So if our matrices are too small, > MKL will not offload because the cost of moving the data is less than the > gains from performance. This bound is something like a matrix size of 4096. > How to see the sizes of various matrices as the code runs in theano? what > all parameters could increase it? like larger datasets..or batchsize etc..? > > any suggestions also would be helpful.. thanks > > > *with regards..* > > *M. Sai Rajeswar* > > > *IIT Delhi----------------------------------Cogito Ergo Sum---------* > > On Thu, Jan 15, 2015 at 2:37 AM, Pascal Lamblin <[email protected] > <javascript:>> wrote: > >> On Wed, Jan 07, 2015, Frédéric Bastien wrote: >> > icc/icpc isn't always faster then gcc. I didn't time it. So I don't know >> > what this will do. It will depend of the mode. If the model spend most >> of >> > its time in the BLAS library, it won't make a difference. >> >> I'm not sure that is correct in that case, as I expect icc to be able to >> link with the MIC version of blas, which gcc may not be able to do. >> >> -- >> Pascal >> >> -- >> >> --- >> You received this message because you are subscribed to a topic in the >> Google Groups "theano-users" group. >> To unsubscribe from this topic, visit >> https://groups.google.com/d/topic/theano-users/NpImVMwVESo/unsubscribe. >> To unsubscribe from this group and all its topics, send an email to >> [email protected] <javascript:>. >> For more options, visit https://groups.google.com/d/optout. >> > > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
