Thread: Dynamically Allocated arrays and MPI

Hi all,

I am looking to find some resources online that show how you might send/recv portions of a 3D array in MPI using C.

Basically, I want to be able to send each 2D layer of the 3D matrix to other processors, do some computes, get that 2D matrix back, and re-assemble it into the whole. I can't seem to find any resources that show how best to send the contiguous block of memory in MPI. In psuedo code, i imagine it being something like this

Code:

Declare, allocate, initialize 3DGrid[NX][NY][NZ]
Loop over Z dimension
   
   Send all elements 3DGrid[*][*][NZ] to other processor from main processor
   Recieve elements 3DGrid[*][*][NZ] on the compute processor
   Do some computes on 3DGrid[*][*][NZ] on the compute processor
   Send modified elements 3DGrid[*][*][NZ] back to main processor from Compute processor
   Receieve 3DGrid[*][*][NZ] on main processor
   Reassemble 3DGrid[*][*][NZ] back into the main array on main processor

end Z loop

Has anyone ever done this before?

All there is to it is using the APIs provided by the MPI software, for ex. MPI_Send() and MPI_Recv().

Please refer this site , you may get a clear picture
http://www.ats.ucla.edu/clusters/com...xamples/send.c

I discovered that my problem was that I did not have my arrays defined contiguously. That being said, if I want to loop in this order:
for z =...
for x =...
for y =...

and access everything in the typical way 3darray[x][y][z] it will still be contiguous in memory. How do you do this? every definition I find about allocating the entire space, then space for the pointers doesnt seem to work