[NumPy-Tickets] [NumPy] #1723: Mainly performance improvement in tri and related functions

[NumPy Trac]

#1723: Mainly performance improvement in tri and related functions
----------------------------------------------+-----------------------------
 Reporter:  eat                               |       Owner:  somebody
     Type:  enhancement                       |      Status:  new     
 Priority:  normal                            |   Milestone:  1.5.2   
Component:  Other                             |     Version:  1.5.0   
 Keywords:  tri*, tri*_indices, mask_indices  |  
----------------------------------------------+-----------------------------
 First of all this is enhancement request for functions: tri, tril, triu
 and mask_indices. If these are implemented as demonstrated in the attached
 code, more than 1.5 speedup (old exec time/ new exec time) may be achieved
 (and as far I'm aware the speedup is always >1). Thus this is observed on
 np.__version__== '1.5.1' and sys.version '2.7.1 (r271:86832, Nov 27 2010,
 18:30:46) [MSC v.1500 32 bit (Intel)]'.

 The speedup is achieved by avoiding unnecessary copying of arrays in tri
 and changing the default dtype (when tri is called) from int to bool in
 tril, triu and mask_indices.

 Secondly the documentation of tril states that it will return "of same
 shape and data-type" as input. This is not the case with current
 implementation (at least with dtype bool). My proposal fixes this
 trivially.

 Thirdly, naming of rows and cols. In tri (N, M) while in rest related (M,
 N). I propose consistency with (M, N) as fixed in attachment.

 My attachment is based on version 1.5.1 twodim_base.py. Basically just
 replace "_f" with "" and remove rows after and including "if __name__ ..."
 should give you a working version.

 Additionally all related *_indices functions in twodim_base will gain
 performance.

-- 
Ticket URL: <http://projects.scipy.org/numpy/ticket/1723>
NumPy <http://projects.scipy.org/numpy>
My example project