Transfer distributed array or gpuArray to local workspace



X = gather(A) can operate on the following array data:

  • On a gpuArray: Transfers the elements of A from the GPU to the local workspace and assigns them to X.

  • On a distributed array, outside an spmd statement: Gathers together the elements of A from the multiple workers to the local workspace and assigns them to X.

  • On a codistributed array, inside an spmd statement or communicating job: Gathers together the elements of A and replicates them into X on every worker.

You can call gather on other data types, such as tall arrays (See gather (tall)). If the data type does not support gathering, then gather has no effect.

Gathering GPU arrays or distributed arrays can be costly and is generally not necessary unless you need to use your result with functions that do not support these types of arrays. For more information on function support, see Run MATLAB Functions on a GPU or Run MATLAB Functions with Distributed Arrays.

X = gather(gpuArray(X)), X = gather(distributed(X)), or X = gather(codistributed(X)) return the original array X.

[X1,X2,...,Xn] = gather(A1,A2,...,,Xn) gathers multiple arrays A1,A2,...,,An into the corresponding outputs X1,X2,...,Xn. The number of input arguments and output arguments must match.


X = gather(C,lab) converts a codistributed array C to a variant array X, such that all of the elements are contained on worker lab, and X is a 0-by-0 empty double on all other workers.

[X1,X2,...,Xn] = gather(C1,C2,...,Cn,lab) gathers codistributed arrays C1,C2,...,Cn into corresponding outputs X1,X2,...,Xn, with all elements on worker lab. The number of input arguments and output arguments must match.


collapse all

Gather the results of a GPU operation to the MATLAB® workspace.

G = gpuArray(rand(1024,1));
F = sqrt(G);   % Input and output are both gpuArray
W = gather(G); % Return array to workspace
Name         Size       Bytes  Class

F         1024x1          108  gpuArray
G         1024x1          108  gpuArray
W         1024x1         8192  double

Gather all of the elements from a distributed array D onto the client.

n = 10;
D = distributed(magic(n)); % Distribute array to workers
M = gather(D)              % Return array to client

Distribute a magic square across your workers, then gather the whole matrix onto every worker and then onto the client. This code results in the equivalent of M = magic(n) on all workers and the client.

n = 10;
  C = codistributed(magic(n));
  M = gather(C) % Gather all elements to all workers
S = gather(C)   % Gather elements to client

Gather all of the elements of C onto worker 1, for operations that cannot be performed across distributed arrays.

n = 10;
  C = codistributed(magic(n));
  out = gather(C,1);
  if labindex == 1
    % Characteristic sum for this magic square:
    characteristicSum = sum(1:n^2)/n;
    % Ensure that the diagonal sums are equal to the 
    % characteristic sum:
    areDiagonalsEqual = isequal ...
Lab 1:
  areDiagonalsEqual =

Input Arguments

collapse all

Array to gather, specified as a gpuArray, distributed array, or codistributed array.


Note that gather assembles the codistributed or distributed array in the workspaces of all the workers on which it executes, or on the MATLAB client, respectively, but not both. If you are using gather within an spmd statement, the gathered array is accessible on the client via its corresponding Composite object; see Access Worker Variables with Composites. If you are running gather in a communicating job, you can return the gathered array to the client as an output argument from the task.

As the gather function requires communication between all the workers, you cannot gather data from all the workers onto a single worker by placing the function inside a conditional statement such as if labindex == 1.

Introduced in R2006b