Changes between Version 6 and Version 7 of PossibleOptimizationAreas

Timestamp:

02/27/06 20:54:17 (7 years ago)

Author:

sasha

Comment:

moved ndarray.fill discussion to a separate page

PossibleOptimizationAreas

@@ -6 +6 @@
  * Default (digits=0) case of around is 10x slower than {{{(x+0.5).astype(int).astype(float)}}}. Resolved: changeset:2151 implements fast rint function and adds a round method to ndarray that is about as fast for digits=0 case.
  
- * As of changeset:2173 {{{x.fill(1)}}} is 2x slower than {{{x += 1}}}.  One should be able to set memory to a constant value faster than autoincrement.
-
-{{{
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'b')" "x.fill(1)"
-10000 loops, best of 3: 69.5 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'h')" "x.fill(1)"
-10000 loops, best of 3: 66.1 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'i')" "x.fill(1)"
-10000 loops, best of 3: 66.3 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'d')" "x.fill(1)"
-10000 loops, best of 3: 73.2 usec per loop
-}}}
-{{{
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'b')" "x += 1"
-10000 loops, best of 3: 58 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'h')" "x += 1"
-10000 loops, best of 3: 33.7 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'i')" "x += 1"
-10000 loops, best of 3: 33.6 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'d')" "x += 1"
-10000 loops, best of 3: 36.9 usec per loop
-}}}
-The [attachment:fast-fill-patch.txt attached patch] results in the following timings:
-{{{
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'b')" "x.fill(1)"
-100000 loops, best of 3: 4.55 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'h')" "x.fill(1)"
-100000 loops, best of 3: 12 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'i')" "x.fill(1)"
-100000 loops, best of 3: 12.4 usec per loop
-> python -m timeit -s "from numpy import zeros; x = zeros(10000,'d')" "x.fill(1)"
-100000 loops, best of 3: 13 usec per loop
-}}}
-Note the more than 10x improvement in the 'b' case.  
+ * As of changeset:2173 {{{x.fill(1)}}} is 2x slower than {{{x += 1}}}.  One should be able to set memory to a constant value faster than autoincrement. [wiki:PossibleOptimizationAreas/FillDiscussion Discussion].