Mathematica 9 is now available
Services & Resources / Wolfram Forums / MathGroup Archive
-----

MathGroup Archive 2010

[Date Index] [Thread Index] [Author Index]

Search the Archive

Re: Can I speed up a matrix * tensor calculation?

  • To: mathgroup at smc.vnet.net
  • Subject: [mg111364] Re: Can I speed up a matrix * tensor calculation?
  • From: David Bailey <dave at removedbailey.co.uk>
  • Date: Thu, 29 Jul 2010 06:42:25 -0400 (EDT)
  • References: <i2m4je$o1h$1@smc.vnet.net>

On 27/07/10 09:15, Jagra wrote:
> I have a tensor and a matrix. Both pretty large.  Both have real
> numbers with machine precision:
>
>      Dimensions[matrix]
>      {10000, 5000}
>
>      Dimensions[tensor]
>      {5000, 7, 10000}
>
> I need to do the following:
>
>      Total[matrix]* Transpose[tensor {2, 3, 1}]];
>
> but this runs very slowly.  I've thought I needed to transpose the
> tensor to multiply it properly.
> The result should give me a vector of 5000 elements.
>
> Any suggestions that can speed this up?
>
> Thanks.
>
> J
>
The square brackets are not matched in your expression, so it is hard to 
know exactly what calculation you need to do. Also, is the '*' meant to 
be a matrix multiplication, and there seems to be a camma missing inside 
the Transpose operation! It would be best if you posted some slow but 
working code, with some idea of the speedup you are hoping to achieve.

Nevertheless, there are a few general points to make here:

1)  If your tensors are sparse, you may be better off to use SparseArray 
objects.

2)  Check if your arrays are stored as packed arrays, testing with 
Developer`PackedArrayQ and converting if necessary. Note that real 
arrays containing the odd integer (such as 0 !) will not pack, and you 
may want to apply N to such an array prior to packing with 
Developer`ToPackedArray.

The difference between working with packed and unpacked arrays of this 
size will be substantial.

David Bailey

http://www.dbaileyconsultancy.co.uk


  • Prev by Date: Re: Substiuting variables within an integral
  • Next by Date: Re: Writing to same file from several kernels?
  • Previous by thread: Can I speed up a matrix * tensor calculation?
  • Next by thread: Parallel search?