Hello NumPy-ers,<div><br></div><div>After some performance analysis, I've designed and implemented a new iterator designed to speed up ufuncs and allow for easier multi-dimensional iteration. The new code is fairly large, but works quite well already. If some people could read the NEP and give some feedback, that would be great! Here's a link:</div>
<div><br></div><div><meta http-equiv="content-type" content="text/html; charset=utf-8"><a href="https://github.com/m-paradox/numpy/blob/mw_neps/doc/neps/new-iterator-ufunc.rst">https://github.com/m-paradox/numpy/blob/mw_neps/doc/neps/new-iterator-ufunc.rst</a></div>
<div><br></div><div>I would also love it if someone could try building the code and play around with it a bit. The github branch is here:</div><div><br></div><div><meta http-equiv="content-type" content="text/html; charset=utf-8"><a href="https://github.com/m-paradox/numpy/tree/new_iterator">https://github.com/m-paradox/numpy/tree/new_iterator</a></div>
<div><br></div><div>To give a taste of the iterator's functionality, below is an example from the NEP for how to implement a "Lambda UFunc." With just a few lines of code, it's possible to replicate something similar to the numexpr library (numexpr still gets a bigger speedup, though). In the example expression I chose, execution time went from 138ms to 61ms.</div>
<div><meta http-equiv="content-type" content="text/html; charset=utf-8"><div><br class="Apple-interchange-newline">Hopefully this is a good Christmas present for NumPy. :)</div><div><br></div><div>Cheers,</div><div>Mark</div>
</div><div><br></div><div><div>Here is the definition of the ``luf`` function.::</div><div><br></div><div><font class="Apple-style-span" face="'courier new', monospace"> def luf(lamdaexpr, *args, **kwargs):</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> """Lambda UFunc</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> </font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> e.g.</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> c = luf(lambda i,j:i+j, a, b, order='K',</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> casting='safe', buffersize=8192)</font></div><div><font class="Apple-style-span" face="'courier new', monospace"><br>
</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> c = np.empty(...)</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> luf(lambda i,j:i+j, a, b, out=c, order='K',</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> casting='safe', buffersize=8192)</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> """</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"><br></font></div><div><font class="Apple-style-span" face="'courier new', monospace"> nargs = len(args)</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> op = args + (kwargs.get('out',None),)</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> it = np.newiter(op, ['buffered','no_inner_iteration'],</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> [['readonly','nbo_aligned']]*nargs +</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> [['writeonly','allocate','no_broadcast']],</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> order=kwargs.get('order','K'),</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> casting=kwargs.get('casting','safe'),</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> buffersize=kwargs.get('buffersize',0))</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> while not it.finished:</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> it[-1] = lamdaexpr(*it[:-1])</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> it.iternext()</font></div><div><font class="Apple-style-span" face="'courier new', monospace"><br></font></div><div><font class="Apple-style-span" face="'courier new', monospace"> return it.operands[-1]</font></div>
<div><br></div><div>Then, by using ``luf`` instead of straight Python expressions, we</div><div>can gain some performance from better cache behavior.::</div><div><br></div><div><font class="Apple-style-span" face="'courier new', monospace"> In [2]: a = np.random.random((50,50,50,10))</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> In [3]: b = np.random.random((50,50,1,10))</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> In [4]: c = np.random.random((50,50,50,1))</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"><br></font></div><div><font class="Apple-style-span" face="'courier new', monospace"> In [5]: timeit 3*a+b-(a/c)</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> 1 loops, best of 3: 138 ms per loop</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"><br></font></div><div><font class="Apple-style-span" face="'courier new', monospace"> In [6]: timeit luf(lambda a,b,c:3*a+b-(a/c), a, b, c)</font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> 10 loops, best of 3: 60.9 ms per loop</font></div><div><font class="Apple-style-span" face="'courier new', monospace"><br></font></div>
<div><font class="Apple-style-span" face="'courier new', monospace"> In [7]: np.all(3*a+b-(a/c) == luf(lambda a,b,c:3*a+b-(a/c), a, b, c))</font></div><div><font class="Apple-style-span" face="'courier new', monospace"> Out[7]: True</font></div>
</div><div><br></div>