Hello,
I've just started trying to use some of sympy's codegen features in an
attempt to create fast routines to output numpy arrays. It seems that not
all of the functionality is there at the moment and I have a few things on
my wishlist:
1. working fcode (and ccode/jscode) for Sum(expr, Tuple())
a. a better method for determining output arguments than (somewhat
opaquely) scraping off a top-level Equality() class.
2. re-working of the tensor class (which does some great things) to put it
in-line with numpy's array class:
a. removal of tensor contraction routines in favor of the Sum() class
b. combining Indexed and IndexedBase classes (I have a test version for
this)
c. changing semantics of tensor indices (Idx class) to simply use (start,
stop, step) to specify selected strides for each tensor dimension (note
that 12*A[2:4] dies with an error because Tuple cannot be ().expand()-ed)
-- these last changes would remove a lot of the tensor functionality in
codegen in favor of some sophistication in translating "Sum"-s. However,
the code generation mechanism seems difficult to expand to me because I
want it to work with arrays ( i.e. numpy.array([Symbol('x')]) ), and
because it doesn't have a facility for injecting new variables to hold
intermediate steps in the computation. A lot more sophistication should go
into creating an internal representation before fcode() and friends get
called, I.e. creating a call graph from cse and optimizing the computation
order of the intermediate variables.
This seems a bit far off, so I'm calling it quits for now and sending a
few helpful patches I made along the way.
~ David M. Rogers
Postdoc, Susan B. Rempe Lab
Sandia National Labs
--
You received this message because you are subscribed to the Google Groups
"sympy-patches" group.
To post to this group, send email to sympy-patches@googlegroups.com.
To unsubscribe from this group, send email to
sympy-patches+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msg/sympy-patches/-/Tdjw8loLZUsJ.
For more options, visit https://groups.google.com/groups/opt_out.
