How to use GPU to speed up computation?

% Define the paths to the three video files
videoPaths = {
    'F:\LIVE Video Quality Challenge (VQC) Database\LIVE Video Quality Challenge (VQC) Database\Video\A002.mp4',
    'F:\LIVE Video Quality Challenge (VQC) Database\LIVE Video Quality Challenge (VQC) Database\Video\A003.mp4',  
    'F:\LIVE Video Quality Challenge (VQC) Database\LIVE Video Quality Challenge (VQC) Database\Video\A004.mp4'  
};

% Define the folder to save the Excel files
outputFolder = 'F:\FEUsingFriquee_csvFiles\';

% Loop over the selected video files
for videoIdx = 1:numel(videoPaths)
    try
        % Open the current video file
        videoObj = VideoReader(videoPaths{videoIdx});
        numFrames = videoObj.NumFrames;

        % Initialize variables to store features
        allFriqueeFeats = [];

        % Create a parallel pool with 6 workers
        if isempty(gcp('nocreate'))
            parpool(6);
        end

        parfor frameIdx = 1:numFrames
            % Read frame
            frame = read(videoObj, frameIdx);

            % Extract features for the current frame
            friqueeFeats = extractFRIQUEEFeatures(frame);

            % Check if the size of features matches the number of frames processed
            if size(friqueeFeats.friqueeALL, 1) ~= 1
                continue; % Skip frames with inconsistent feature sizes
            end

            % Store features for this frame
            allFriqueeFeats = [allFriqueeFeats; friqueeFeats.friqueeALL];
        end

        % Organize features for output
        videoFeatures.allFriqueeFeats = allFriqueeFeats;

        % Define the filename for the current video
        [~, videoName, ~] = fileparts(videoPaths{videoIdx});
        filename = [videoName, '.xlsx'];

        % Full file path for the current video
        fullFilePath = fullfile(outputFolder, filename);

        % Write the feature matrix to Excel file
        writematrix(allFriqueeFeats, fullFilePath);

        % Delete the parallel pool
        delete(gcp);
    catch ME
        fprintf('Error processing video %d: %s\n', videoIdx, ME.message);
        continue; % Continue with the next video file
    end
end

I have to extract features for all frames of each video. the number of frames are 300 and features are 560. The time taken for each video is 8 hours or more and i have to extract feaatures of 500 videos. I have tried to create a pool and use parfor loop to utilize the GPU in order to reduce computation time. However, in task manager, GPU is not being utilized. I have NVIDIA GEFORCE 3050 RTX. How can i use GPU to reduce computation time ?

[Jonas]

parfor uses more than one core of the cpu. to use gpu, you convert your data to gpuArray first

[MashalS]

I have also tried to send the frames and feature extraction to GPU using 'gpuArray function' and then used 'gather' to send results back to CPU,but, since feature extraction is a custom function, the error says it is not GPU compatible. Feature extraction function, which is called in this code is dependent on various other functions. It is very difficult to make changes in all those functions.

[Jonas]

did you profile your function and looked which line did take much time? i could imagine, that reading repeatedly from the videos is the actual problem. use "profile on" before running the code, the start the code, abort if necessary after some loops (in doubt using ctrl+c in the command window) and the type "profile viewer"

[MashalS]

I have exceuted the code for only one video with one frame only. The profiler screenshot is attached above.

[MashalS]

Is it necessary to use inbuilt GPU compatible functions to utilize GPU? is there anyother way?

[Jonas]

do i see that correctly that processing 1 frame only, needs 180s, meaning 3min?

starting with this profile, you may have a look into those function with a big dark band. if you need assistance, you may post those function and explain, what they should do, with is input / output etc.

Answer 1

[Catalytic]

Using the GPU in conjunction with parfor will not be productive unless you have a multi-GPU system and assign each parpool worker a different GPU. Otherwise, the workers will just compete for access to the same GPU bus.

[MashalS]

thankyou for answering. I have attached my device specs. If you can guide accordingly. Regards