Attachment added by trac user oscar.bristol on 2010-01-31: test.py

Attachment added by trac user jpeel on 2011-01-12: 0001-BF-poor-performance-indexing-scalar-record-array.patch

trac user oscar.bristol wrote on 2010-01-31

(Second attempt)
Assigning to a field in a scalar record array like this:

B['a'][:] = A['a'][:]

takes around 50 times as long as assigning like this:

B['a'][:] = A['a']

and 100 times as long as assigning to a size 1 record array. I've attached a script (test.py) that compares scalars and size 1 arrays and the two different methods of indexing above. On my machine, the results are:

$ python test.py 100000 
array,  B['a'][:] = A['a'] 0.25 seconds
array,  B['a']    = A['a'] 0.213 seconds
scalar, B['a'][:] = A['a'] 0.395 seconds
scalar, B['a']    = A['a'] 22.0 seconds

trac user jpeel wrote on 2011-01-12

The problem was that the scalar void type's setfield method makes a copy of the input array even if it is already an array. I made a fast track for when the right hand side value is an array. The fast track uses the getfield method (which is used in the B['a'][:] scalar example above) to get a Numpy array for B['a']. I then just use PyArray_CopyObject to do the transfer. Now, this last method of assignment is right with the others in time (actually faster than scalar B['a'][:] because a slice is not done).

Milestone changed to NumPy 2.0 by trac user jpeel on 2011-01-12

Now it's only 10x as fast ;) @seberg @juliantaylor Either of you interested in taking a look at this?