|
|
Blitz Support : |
From: Julian C. Cummings (cummings_at_[hidden])
Date: 2004-06-15 01:18:41
After thinking about this a bit more tonight, I believe that the ordering
computation for expressions involving a reduction needs to be corrected. We
need to add some checks that (a) the ordering value is less than the reduced
Array rank, and (b) the ordering value has not already appeared earlier in
the ordering array. If either condition is violated, we take the ordering
value for the next highest dimension. These fixes would eliminate the
mismatches in ordering values evident in Patrik's examples.
Regards, Julian C.
Dr. Julian C. Cummings
Staff Scientist, CACR/Caltech
(626) 395-2543
cummings_at_[hidden]
> -----Original Message-----
> From: blitz-support-bounces_at_[hidden]
> [mailto:blitz-support-bounces_at_[hidden]] On Behalf Of
> Julian C. Cummings
> Sent: Monday, June 14, 2004 9:04 PM
> To: 'Support list for Blitz++'
> Subject: RE: [Blitz-support] tensor expression weirdness
>
>
> Hi Patrik,
>
> The ordering array specifies the order in which a blitz array
> is stored in memory. Thus, ordering[0] gives the rank of the
> array with unit stride in memory and ordering[N_rank-1] gives
> the rank with the largest stride. What is happening is that
> you are mixing two arrays with different storage orders,
> which is not allowed when using tensor notation.
>
> I think what is wrong here is that the ordering is not
> correctly computed for an array subexpression containing a
> reduction. According to the code in <blitz/array/reduce.h>,
> the ordering for a reduction is the same as the ordering for
> the initial expression being reduced, with the rank reduced
> by one. It is not obvious to me why this should be the case.
> The behavior of the array ordering is also somewhat
> convoluted for arrays that are indexed with tensor notation.
> This is handled by the code in <blitz/array/map.h>, which
> indicates that the ordering for a given dimension d is equal
> to the ordering of the underlying array for the dimension r
> that has been mapped to d, as long as r is less than the rank
> of the underlying array.
>
> A couple of simple examples may help. By default, blitz
> arrays have C-style ordering, so the ordering array will
> contain [N_rank-1, N_rank-2,...,1,0]. The ordering is
> unchanged if we index an array with tensor indices in the
> normal ascending ordering. Thus, the ordering of A and
> A(tensor::i,tensor::j) are the same for a 2d array A. If we
> reverse the tensor indices, the ordering will be reversed.
> If we use a tensor index beyond the rank of the array, the
> ordering will be set to a flag value INT_MIN for each index
> that is skipped over. So if B is a 3d array, the ordering of
> B(i,j,l) is [2,1,INT_MIN,INT_MIN]. When combining arrays in
> an expression, the ordering of each array must match for each
> rank, but the INT_MIN flag value is ignored.
>
> In your first expression, the ordering for the summed
> expression is [2,1,1,0], which then reduces to [2,1,1] after
> the summation. The ordering for the denominator will be
> [INT_MIN,INT_MIN,0], so it does not match in the final
> dimension. In your second expression, the ordering for the
> summed expression is [1,0], which reduces to [1] after
> summation. This obviously doesn't match up with the ordering
> of the denominator, which must be [0]. In neither case does
> the ordering of the denominator match the ordering of the numerator.
>
> So why does your first expression work and your second
> expression fail? The first one uses the assignment operator
> while the second one uses an array constructor. The array
> constructor sets the ordering of the new array to that of the
> expression, so it has to compute this ordering, and that is
> where the mismatch is detected. The assignment operator
> simply assumes that the array on the left-hand side is filled
> in its own default ordering scheme, so the ordering of the
> right-hand side is not checked.
>
> I'm not sure if the behavior of the ordering is incorrect for
> summations or not. If you change your second expression to
> use assignment, it should work OK but I would check the
> results carefully. When in doubt, store the summation in an
> array, and then perform the division separately and make sure
> your results do not change.
>
> Hope this helps,
> Julian C.
>
> Dr. Julian C. Cummings
> Staff Scientist, CACR/Caltech
> (626) 395-2543
> cummings_at_[hidden]
>
>
> > -----Original Message-----
> > From: blitz-support-bounces_at_[hidden]
> > [mailto:blitz-support-bounces_at_[hidden]] On Behalf Of Patrik
> > Sent: Sunday, June 13, 2004 1:48 AM
> > To: Support list for Blitz++
> > Subject: [Blitz-support] tensor expression weirdness
> >
> >
> > hi all,
> > i'm using tensor expressions to do some reductions:
> >
> > broadband_flux =
> > (sum (image (tensor::i, tensor::j, tensor::l)*
> > passbands (tensor::k, tensor::l)*delta_lambda (tensor::l),
> > tensor::l)/ewidth (tensor::k));
> >
> > Array<double,1> broadband_L_integrated
> > (sum (L_lambda_integrated (tensor::j)*
> > passbands_integrated (tensor::i, tensor::j)*
> > delta_lambda_integrated (tensor::j), tensor::j)/
> > ewidth_integrated(tensor::i));
> >
> > The first one works, while the second one gives the following
> > error (it's
> > the division that's the offending part, the sum by itself works):
> >
> > /u7/patrik/include/blitz/array/expr.h:276 Two array operands
> > have different
> > orders: for rank 0, the orders are 1 and 0
> > Assertion failed: 0, file
> > /u7/patrik/include/blitz/array/expr.h, line 279
> >
> > I don't even know what the error message means... The two
> > expressions are
> > very similar (the only difference is that the first one has a
> > 3d array and
> > the second one has a 1d one, so I don't understand why one
> > works but not
> > the other. Any hints?
> >
> > Regards,
> >
> > /Patrik
> >
> > _______________________________________________
> > Blitz-support mailing list
> > Blitz-support_at_[hidden]
> > http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
> >
>
>
> _______________________________________________
> Blitz-support mailing list
> Blitz-support_at_[hidden]
> http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
>