For your codegen.py example, it is easy to reuse it manually in a Theano Op. A good example is to look at the GpuEye op (and just pretend the the GPU kernel is a CPU one). I don't know of a simpler CPU example that demonstrate that:
https://github.com/Theano/Theano/blob/master/theano/sandbox/cuda/basic_ops.py#L3421 You put the c function in c_support_code() function and call it with code returned by the c_code() fct. If we want to automate that, we need a way to know the order of the parameter passed to the c function if there is more then 1 parameter. Is that possible? Ideally, we need to separate the include from the c function, but I think it work even if we don't. About the Theano speed up, the problem is located in the parsing of the Theano fct inputs. This is done in pure python code. I think we can always use the trust_input=True argument. In the case there is an error, you will get an error message that is harder to understands, as it is an op that will complain that it have bad input. But you won't be easily able to know to witch input it correspond. It could also happen that you don't pass a good input and there is no error. For example if it is used in an op with only python code. But in that case, the computation is probably done correctly, but with the input object received, not the one specified when compiling the theano function. I can't certify this always happen, but I would be surprised to see something else being done. So if you want, just always add trust_input=True, but be ready to have uglier error message and remove the flag to have a better error message in case you have a problem. Ideally, Theano should move that to C code, but I don't think it will happen shortly. On Thu, Aug 8, 2013 at 7:27 PM, Matthew Rocklin <[email protected]> wrote: > @Fred, what easy things can we do on the Theano side to support performant > scalar codes? Your two tricks above were impressive. How can we make this > behavior default? Should this happen on the SymPy side or the Theano side > (Theano side seems better if there is a place for it). > > > On Thu, Aug 8, 2013 at 4:05 PM, Aaron Meurer <[email protected]> wrote: > >> >> >> On Thursday, August 8, 2013, Matthew Rocklin wrote: >> >>> >>> It should be simple to translate SymPy.Piecewise to a recursive >>>>> Theano.switch (after translating SymPy.LT to theano.lt, etc.) I'll >>>>> get on this soon. Does this sound reasonable to you Fred? >>>>> >>>> >>>> It sound reasonable and is the first thing I suggest to try. >>>> >>> >>> Working on this now. >>> >>> > SymPy C Codegen and Theano >>>>> >>>>> @Fred, how hard would it be to leverage SymPy's C codegen in Theano? >>>>> This might be a lot cleaner than wrapping raw SymPy operations and might >>>>> substantially extend Theano's support of scalar expressions. Do you have >>>>> a >>>>> performant Bessel function op? I'll bet SymPy could be made to do this >>>>> quite well. >>>>> >>>>> @Aaron / @Ondrej, if you're reading this thread could you point us to >>>>> the best place to start looking at C codegen in SymPy? Alternatively can >>>>> you point to an active community member who would be able to do so? >>>>> >>>>> >>>> @Matt, you already did a new Theano op with C code. I think it is the >>>> only "easy" way to wrap other people c code in Theano. If the person >>>> already know this C code AND a little of Python AND NumPy C-API, it isn't >>>> very hard to a new Theano op with C code. Otherwise, doing the first such >>>> op ask to learn a few think and could ask a few days. You already did this, >>>> so you have a good idea of the work it need. >>>> >>>> Now the questions is how is done the SymPy code gen? Is just just >>>> string template that is filled with dtype and other stuff? If we can just >>>> call one SymPy function with the information of what we want and it return >>>> a string with the C code it could be relatively easy. The only questions is >>>> about how to handle the variable name to pass the information around. At >>>> worst, we wrap the sympy c code in a c function, then make a small wrapper >>>> c code that take the Theano c variable name and call this function. So not >>>> very hard as Theano provide what is needed. >>>> >>> >>> It looks like codegen is the relevant high-level api call >>> >>> In [1]: from sympy.utilities.codegen import >>> >>> In [2]: expr = sin(x)**2 >>> >>> In [3]: [(c_name, c_code), (h_name, c_header)] = codegen(("f", expr), >>> 'C', 'test', header=False) >>> >>> In [4]: print c_code >>> #include "test.h" >>> #include <math.h> >>> >>> double f(double x) { >>> >>> return pow(sin(x), 2); >>> >>> } >>> >>> The work I did rarely dealt with making and using functions. I'll go >>> over past work and see what I can do. Expect some calls for help though! >>> >> >> If you figure out how to do something that's under-documented, please >> document it. Our code generation needs a lot more documentation, especially >> high-level narrative documentation. >> >> Aaron Meurer >> >> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "sympy" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to [email protected]. >>> To post to this group, send email to [email protected]. >>> Visit this group at http://groups.google.com/group/sympy. >>> For more options, visit https://groups.google.com/groups/opt_out. >>> >>> >>> >> -- >> You received this message because you are subscribed to the Google Groups >> "sympy" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected]. >> To post to this group, send email to [email protected]. >> Visit this group at http://groups.google.com/group/sympy. >> For more options, visit https://groups.google.com/groups/opt_out. >> > > -- > You received this message because you are subscribed to the Google Groups > "sympy" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To post to this group, send email to [email protected]. > Visit this group at http://groups.google.com/group/sympy. > For more options, visit https://groups.google.com/groups/opt_out. > > > -- You received this message because you are subscribed to the Google Groups "sympy" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/sympy. For more options, visit https://groups.google.com/groups/opt_out.
