Re: Dot product full-sparse

[Jaroslav Hajek]

On Thu, Aug 12, 2010 at 10:14 AM, Marco Caliari <address@hidden> wrote:
> On Thu, 12 Aug 2010, Jaroslav Hajek wrote:
>
>> On Wed, Aug 11, 2010 at 8:20 AM, Marco Caliari <address@hidden>
>> wrote:
>>>
>>> Dear maintainers,
>>>
>>> if I run
>>>
>>> A = toeplitz ([-2, 1, 1, 1, zeros(1, 856)]);
>>> A = reshape (A, 18490, 40);
>>> SA = sparse (A);
>>> B = rand (size (A));
>>> tic, A .* B;,toc
>>> tic, SA .* B;, toc
>>> tic, SA .* sparse (B);, toc
>>>
>>> I get
>>>
>>> Elapsed time is 0.00745105 seconds.
>>> Elapsed time is 0.229055 seconds.
>>> Elapsed time is 0.0203849 seconds.
>>>
>>> (3.3.51 with Suitesparse 3.5.0, similar results with 3.2.4 with
>>> Suitesparse
>>> 3.4.0). Why SA .* sparse (B) is so faster than SA .* B? What is done by
>>> default when a full-sparse dot product is required?
>>>
>>> Best regards,
>>>
>>> Marco
>>>
>>
>> Apparently this operation is just poorly optimized.
>>
>> Before I start doing anything here, I'd like to know what the Matlab
>> behavior is w.r.t. NaNs. Namely, if
>> A = sparse ([2,3,1,2], [1,1,3,3], [0.1, 0.2, 0.3, 0.4]);
>> B = ones (3); B(1,1) = NaN;
>> C = sparse ([1,2,3,1,2], [1,1,1,3,3], [NaN, 0.1, 0.2, 0.3, 0.4]);
>>
>> what is the output of A.*B, B.*A, A ./ B, A .*C?
>> Of course, newer version is better.
>
> Here are the results with Matlab 7.6.0 (R2008a)
>
>>> A.*B
>
> ans =
>   (1,1)          NaN
>   (2,1)       0.1000
>   (3,1)       0.2000
>   (1,3)       0.3000
>   (2,3)       0.4000
>>>
>>>  B.*A
>
> ans =
>   (1,1)          NaN
>   (2,1)       0.1000
>   (3,1)       0.2000
>   (1,3)       0.3000
>   (2,3)       0.4000
>>>
>>> A ./ B
>
> ans =
>   (1,1)          NaN
>   (2,1)       0.1000
>   (3,1)       0.2000
>   (1,3)       0.3000
>   (2,3)       0.4000
>>>
>>> A .* C
>
> ans =
>   (1,1)          NaN
>   (2,1)       0.0100
>   (3,1)       0.0400
>   (1,3)       0.0900
>   (2,3)       0.1600
>
> Marco
>

Thanks, Marco. Two more questions, please: What happens for:

Inf * speye (3)
speye(3) / NaN

?

-- 
RNDr. Jaroslav Hajek, PhD
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz