Bigarray question

Bigarray question

[Thomas Fischbacher]

Hello everybody,

as far as I can tell, it is easy (but not explicitly allowed in the
documentation) to allocate a Caml bigarray and change the data
pointer inside that bigarray afterwards.

What I am into right now is writing a function which allows me to
walk a caml function over the entries of a sparse matrix operator
managed by a low-level sparse matrix linear algebra C library.
Therefore, it would be rather elegant to allocate bigarrays for
index and data entries only once and pass them on to the ML
callback for every row of the matrix.

There is no problem in implementing this, technically speaking.
But this would require a change to the ML documentation stating
that it is explicitly considered as permissible to change the
data pointer in a bigarray. Can I get that, please?

-- 
best regards,
Thomas Fischbacher
t.fischbacher@soton.ac.uk

Re: [Caml-list] Bigarray question

[Xavier Leroy]

> as far as I can tell, it is easy (but not explicitly allowed in the
> documentation) to allocate a Caml bigarray and change the data
> pointer inside that bigarray afterwards.
> There is no problem in implementing this, technically speaking.
> But this would require a change to the ML documentation stating
> that it is explicitly considered as permissible to change the
> data pointer in a bigarray. Can I get that, please?

I see no reason why this would cause problems, as long as the data
pointer points to C data of the right shape and dimensions.  (Of
course, you could update the dimensions of the bigarray as
appropriate, if needed.)  I don't think I would document this, as I
wouldn't quite know how to word it in the docs, but you have my
encouragements to try and report problems if any.

- Xavier Leroy

Re: [Caml-list] Bigarray question

[Thomas Fischbacher]

Xavier,

>>as far as I can tell, it is easy (but not explicitly allowed in the
>>documentation) to allocate a Caml bigarray and change the data
>>pointer inside that bigarray afterwards.
>>There is no problem in implementing this, technically speaking.
>>But this would require a change to the ML documentation stating
>>that it is explicitly considered as permissible to change the
>>data pointer in a bigarray. Can I get that, please?
> 
> I see no reason why this would cause problems, as long as the data
> pointer points to C data of the right shape and dimensions.  (Of
> course, you could update the dimensions of the bigarray as
> appropriate, if needed.)  I don't think I would document this, as I
> wouldn't quite know how to word it in the docs, but you have my
> encouragements to try and report problems if any.

Just adding a line to the documentation which states that
"it is permitted to dynamically change the data and dims
fields of the bigarray data structure from within C" would be
sufficient.

-- 
best regards,
Thomas Fischbacher
t.fischbacher@soton.ac.uk

Re: [Caml-list] Bigarray question

[Thomas Fischbacher]

Xavier Leroy wrote:

(on whether it is permissible to dynamically change the dim and data
pointers in a ML bigarray):

>>as far as I can tell, it is easy (but not explicitly allowed in the
>>documentation) to allocate a Caml bigarray and change the data
>>pointer inside that bigarray afterwards.
>>There is no problem in implementing this, technically speaking.
>>But this would require a change to the ML documentation stating
>>that it is explicitly considered as permissible to change the
>>data pointer in a bigarray. Can I get that, please?
> 
> 
> I see no reason why this would cause problems, as long as the data
> pointer points to C data of the right shape and dimensions.  (Of
> course, you could update the dimensions of the bigarray as
> appropriate, if needed.)  I don't think I would document this, as I
> wouldn't quite know how to word it in the docs, but you have my
> encouragements to try and report problems if any.


As a matter of fact, considering the two C functions below which I use
in our "nsim" abstract field theory/multiphysics simulation engine to
speed up some internal computation, the first variant would be desirable
but crashes. The second works. What this actually is about is to
systematically go through the rows of a sparse matrix, handled by the
PETSc library. (So, basically, I am using a sparse matrix as sort-of a
hash of row index to entries here.)

CAMLprim value caml_petsc_matrix_call_on_rows_raw(value ml_mx,
						  value ml_fun,
						  value ml_start_row,
						  value ml_end_row
						  )
{
   CAMLparam4(ml_mx,ml_fun,ml_start_row,ml_end_row);
   Mat mx;
   int start_row,end_row,start_own,end_own;
   int i,ncols=0;
   const int *cols=0;
   const PetscScalar *vals=0;
   CAMLlocal2(ba_indices,ba_vals);

   petsc_checkinit();
   petsc_check_mat(ml_mx);

   mx=(Mat)Field(ml_mx,1);

   MatGetOwnershipRange(mx,&start_own,&end_own);

   start_row=Int_val(ml_start_row);
   end_row=Int_val(ml_end_row);

   if(start_row == -1)start_row=start_own;
   if(end_row == -1)end_row=end_own;

   /* We would like to do things this way, but while one may guess that 
it should work,
      this (1) is not covered by what is permitted according to the 
OCaml documentation,
      and (2) indeed produces crashes.
   */
   ba_indices=alloc_bigarray_dims(BIGARRAY_NATIVE_INT | BIGARRAY_C_LAYOUT,
				 1, cols, ncols);

   ba_vals=alloc_bigarray_dims(BIGARRAY_FLOAT64 | BIGARRAY_C_LAYOUT,
			      1, vals, ncols);


   for(i=start_row;i<end_row;i++)
     {
       MatGetRow(mx,i,&ncols,&cols,&vals);

       Bigarray_val(ba_indices)->dim[0]=ncols;
       Bigarray_val(ba_vals)->dim[0]=ncols;

       Bigarray_val(ba_indices)->data=(void*)cols;
       Bigarray_val(ba_vals)->data=(void*)vals;

       callback3(ml_fun,Val_int(i),ba_indices,ba_vals);

       MatRestoreRow(mx,i,&ncols,&cols,&vals);
     }

   CAMLreturn(Val_unit);
}



CAMLprim value caml_petsc_matrix_call_on_rows_raw_defensive(value ml_mx,
							    value ml_fun,
							    value ml_start_row,
							    value ml_end_row
							    )
{
   CAMLparam4(ml_mx,ml_fun,ml_start_row,ml_end_row);
   Mat mx;
   int start_row,end_row,start_own,end_own;
   int i,ncols=0;
   const int *cols=0;
   const PetscScalar *vals=0;
   CAMLlocal2(ba_indices,ba_vals);

   petsc_checkinit();
   petsc_check_mat(ml_mx);

   mx=(Mat)Field(ml_mx,1);

   MatGetOwnershipRange(mx,&start_own,&end_own);

   start_row=Int_val(ml_start_row);
   end_row=Int_val(ml_end_row);

   if(start_row == -1)start_row=start_own;
   if(end_row == -1)end_row=end_own;


   for(i=start_row;i<end_row;i++)
     {
       MatGetRow(mx,i,&ncols,&cols,&vals);

       ba_indices=alloc_bigarray_dims(BIGARRAY_NATIVE_INT | 
BIGARRAY_C_LAYOUT,
				     1, cols, ncols);

       ba_vals=alloc_bigarray_dims(BIGARRAY_FLOAT64 | BIGARRAY_C_LAYOUT,
				  1, vals, ncols);

       callback3(ml_fun,Val_int(i),ba_indices,ba_vals);

       MatRestoreRow(mx,i,&ncols,&cols,&vals);
     }

   CAMLreturn(Val_unit);
}


-- 
best regards,
Thomas Fischbacher
tf@functionality.de

Bigarray question

[Thomas Fischbacher]

As far as I can see from the documentation, it is not explicitly permitted 
to change a once allocated bigarray data structure in such a way that one 
replaces the data pointer by some other data pointer (while retaining 
size and type). This would, however, be quite useful in some situations.

Could this be changed?

-- 
regards,               tf@cip.physik.uni-muenchen.de              (o_
 Thomas Fischbacher -  http://www.cip.physik.uni-muenchen.de/~tf  //\
(lambda (n) ((lambda (p q r) (p p q r)) (lambda (g x y)           V_/_
(if (= x 0) y (g g (- x 1) (* x y)))) n 1))                  (Debian GNU)