Vectorize a series of interpn calculations with GPU

Let's say I have a cell array with equal sized (but different) 4D datasets:

nPoints = 5;
dataCell = cell(1,nPoints);

 for i = 1:nPoints
     dataCell{i} = rand(3, 2, 4, 3);
 end

For the data arays in each cell, I would like to carry out a 4D interpn. Let's say the interpolation sample arrays and targets for each dimension are as follows:

dimOneArray = [0.1 0.2 0.3];
dimTwoArray = [2 4];
dimThreeArray = [1 2 3 4];
dimFourArray = [0.3 0.6 0.9];

dimOneTarget = 0.15;
dimTwoTarget = 3;
dimThreeTarget = 2.5;
dimFourTarget = 0.4;

Then I would like to carry out nPoints interpn calculations in a vectorized manner i.e. I would like to produce the same result as below (interpArray) without a for loop:

interpArray = nan(nPoints, 1);

for i = 1:nPoints
    interpArray(i) = interpn(dimOneArray, dimTwoArray, dimThreeArray, dimFourArray, cell2mat(dataCell(i)), dimOneTarget, dimTwoTarget, dimThreeTarget, dimFourTarget);
end

How would I use the GPU to carry out this calculation?

Answer 1

[Meet]

Hi,

To perform interpolation of the given data on a GPU, you can perform the following steps:

1. Transfer your “dataCell” array to the GPU using the “gpuArray” function.
2. Initialize an empty array on the GPU to store the final interpolation results.
3. Define a function handle that will be used to perform the interpolation.
4. Use “arrayfun” function to apply the interpolation to each element of the “dataGPU” array.

% Convert dataCell array to gpuArray
dataGPU = cellfun(@gpuArray, dataCell, 'UniformOutput', false);

% Initialize result array which stores the result of the computation
interpArrayGPU = gpuArray.nan(nPoints, 1);

% Defining a function handle for interpolation
interpolateFunction = @(data) interpn(dimOneArray, dimTwoArray, dimThreeArray, dimFourArray, data, dimOneTarget, dimTwoTarget, dimThreeTarget, dimFourTarget);

% Using arrayfun to apply the interpolation function to each element of
% dataGPU array
interpArrayGPU = arrayfun(@(i) interpolateFunction(dataGPU{i}), 1:nPoints);

You can refer to the resources below for more information:

https://www.mathworks.com/help/releases/R2021a/parallel-computing/gpuarray.arrayfun.html

https://www.mathworks.com/help/releases/R2021a/parallel-computing/gpuarray.html

https://www.mathworks.com/help/releases/R2021a/parallel-computing/run-matlab-functions-on-a-gpu.html