On Wed, 25 Jun 2003, Sarath Devarajan wrote:
> #2 : The array that I'm passing say, is of the type,
> array[0..9][0..9][0..9]
>
> I'm already using this array[1000] logic for another array which is
> basically a logical array (denotes whether a particular lattice point is
> an obstacle or a pore)..but since a lot of manipulation needs to be done
> on these 3D arrays, I prefer to retain its structure... Would that be a
> heavy loss in computational time? or buffer space???
No.
> And as to the contiguous sending of the y,z, I think I am doing the
> same..can u check if this is what u meant, Jeff?
>
> for array[0..8][0..9][0..9], i want to send
> subarray[0..8][0..2][0..9] to each process.
> i.e two rows for each of five process..
You can do this in a single send if you use an MPI datatype, or, as you
have in your code below, multiple sends spanning the last 20 elements in
each plane (i.e., the last 2 rows of the plane).
> code part:
>
> for(i=0;i<9;i++)
> {
> if (nid==0)
> {
> MPI_Send(&array[i][0][0],20,MPI_FLOAT,1,tag,MPI_COMM);
> }
> }
>From what you described. this should be fine.
> Inside each processor, I'm creating an array called "node" (runtime) to
> accept the values sent by the "for" loop of the type above. (This is in
> C++, so I am just giving "float node[][][]" of appropriate dimensions.
>
> Does MPI treat the array node as a pointer type array whose values are
> accessed by *node[][][]?
It's not MPI that treats it this way, it's C. i.e., it's the language of
C itself, not any special treatment that MPI uses.
So yes -- node is a pointer. See below.
> Because after working with this "node" and printing out the results, out
> came hexadecimal gibberlish to fill the screen...I'm presuming that the
> addresses were printed out...Does this mean that all changes must be
> made like
>
> *node[0][i][j]=*node[0][i][j]+1;
> and cout<<*node[][][];
Keep in mind that "float *node[][][]" is different than "float
node[][][]". The former is effectively a 4D array, and the latter is
effectively a 3D array (I'm not 100% sure of the C++ syntax; I believe the
latter is not quite right because you would need to specify dimensions in
that case...?). So I'm not quite sure what you're trying to show here;
I'm guessing that you meant to put in indices in the "cout" statement that
you showed...? Otherwise you'll just be showing some kind of pointer
value, but since you've got "*node", I think you're using an extra
dimension that you're not intending to use.
So it all depends on how you setup the "node" array; from your
description, it *sounds* like you really only want 3 dimensions. So you
might be able to do something as simple as (typing off the top of my head
-- pardon any errors...):
// Receive the 3 dimensions from the master
int dims[3];
MPI_Recv(dims, 3, MPI_INT, master, tag, comm, MPI_STATUS_IGNORE);
// Make an array of the right size
float ***node = new float[dims[0]][dims[1]][dims[2]];
// Fill the lower 20 values in several of the planes
for (i = 0; i < 9; ++i)
MPI_Recv(&array[i][0][0], 20, MPI_FLOAT, master, tag,
comm, MPI_STATUS_IGNORE);
Something along those lines.
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
