Search code examples
arrays3dmpiopenmpi

MPI Scattering 3D array to tetragonal prismas

Hello everyone, I am trying to scatter 3d array in tetragonal prisma format. I hope this picture will clarify what I mean.

shape

Think of the big cube is 3d array(lets say 4x4x4 dimensions) and P0..3 is processors. (P0 will take [0..1][0..1][0..3] part of big array). I am using datatypes in order to do this, I created 2 datatypes;

MPI_Datatype dtype1, surface,dtype2,tetragonal;

MPI_Type_vector((N)/sqrt(size),
            (N),                               
           (N),         
           MPI_FLOAT,       
           &dtype1);       

 MPI_Type_commit(&dtype1);
 MPI_Type_create_resized(dtype1, 0, 1*sizeof(float), &surface);

 MPI_Type_commit(&surface);





 MPI_Type_vector(N/sqrt(size),    
           1,                  
           sqrt(size),         
           surface,       
           &dtype2);       

 MPI_Type_commit(&dtype2);
 MPI_Type_create_resized(dtype2, 0, 1*sizeof(float), &tetragonal);
 MPI_Type_commit(&tetragonal);

However, I am getting very strange results, I am new on MPI environment so I need some help on this issue to understand how datatypes actually work. Thank you.

(Assume processor size is squared number and N is evenly divisible by sqrt(size) )

Solution

  • You cannot use MPI_Scatter() with this derived datatype, but you can use MPI_Scatterv() (all the counts are equal, but you need to manually calculate the displacements).