[Numpy-discussion] matrix multiplication

[Keith Goodman]

On Fri, Jun 5, 2009 at 5:19 PM, Christopher Barker
<Chris.Barker at noaa.gov> wrote:
> Robert Kern wrote:
>>>>> x = np.array([1,2,3])
>>>>> timeit x.sum()
>>> 100000 loops, best of 3: 3.01 µs per loop
>>>>> from numpy import sum
>>>>> timeit sum(x)
>>> 100000 loops, best of 3: 4.84 µs per loop
>
> that is a VERY short array, so one extra function call overhead could
> make the difference. Is it really your use case to have such tiny sums
> inside a big loop, and is there no way to vectorize that?

I was trying to make the timeit difference large. It is the overhead
that I was interested in. But it is still noticable when x is a
"typical" size:

>> x = np.arange(1000)
>> timeit x.sum()
100000 loops, best of 3: 5.46 µs per loop
>> from numpy import sum
>> timeit sum(x)
100000 loops, best of 3: 7.31 µs per loop

>> x = np.random.rand(1000)
>> timeit x.sum()
100000 loops, best of 3: 6.81 µs per loop
>> timeit sum(x)
100000 loops, best of 3: 8.36 µs per loop

That reminds me of a big difference between arrays and matrices.
Matrices have the overhead of going through python code (the matrix
class) to get to the core C array code. I coverted an interative
optimization function from matrices to arrays and got a speed up of a
factor of 3.5. Array size was around (500, 10) and (500,)