Crash bug

[kasparn]

From Rasmus to Kaspar:

Fejlen opstår, hvis man først kører en simulering med intet threshold, altså med dense matricer. Hvis man så bagefter starter en ny simulering med threshold, altså med sparse matricer, så crasher koden efter endt simulering under ”cudaDestroy()”. Man får simpelthen en memory access fejl.

Hvis man kører med sparse matricer i begge simuleringer (eller med dense), så er der ingen problemer.

Det er en sjov fejl, for man er tilbage i Matlab mellem de to simuleringer, så det burde ikke kunne ske.

Du kan køre den vedhæftede fil med hhv. crash_and_burn = 0 og = 1 for at teste det.

Matlab code to run: clearvars close all

%--- Use either field 1 or field 2 from the NIST example NIST_field = 1; resolution = [136,19,1]; use_CUDA = true; crash_and_burn = 1;

figure1= figure('PaperType','A4','Visible','on','PaperPositionMode', 'auto'); fig1 = axes('Parent',figure1,'Layer','top','FontSize',16); hold on; grid on; box on mu0 = 4pi1e-7;

addpath('../MagTense/matlab/MEX_files'); addpath('../MagTense/matlab/examples/util');

%% Setup the problem for the initial configuration %takes the size of the grid as arguments (nx,ny,nz) and a function handle %that produces the desired field (if not present zero applied field is inferred) tic problem = DefaultMicroMagProblem(resolution(1),resolution(2),resolution(3));

problem = problem.setUseCuda( use_CUDA );

if (crash_and_burn == 1) problem.dem_appr = getMicroMagDemagApproximation('none'); else problem.dem_appr = getMicroMagDemagApproximation('threshold'); end problem.dem_thres = 1e-4; problem.alpha = 4.42e3; problem.gamma = 0;

%initial magnetization problem.m0(:) = 1/sqrt(3);

%time grid on which to solve the problem problem = problem.setTime( linspace(0,100e-9,200) ); problem.setTimeDis = int32(10); HystDir = 1/mu0*[1,1,1] ;

%time-dependent applied field HextFct = @(t) (1e-9-t)' . HystDir . (t<1e-9)';

problem = problem.setHext( HextFct );

solution = struct(); %convert the class obj to a struct so it can be loaded into fortran prob_struct = struct(problem);

solution = MagTenseLandauLifshitzSolver_mex( prob_struct, solution ); toc figure; M_end = squeeze(solution.M(end,:,:)); quiver(solution.pts(:,1),solution.pts(:,2),M_end(:,1),M_end(:,2)); axis equal; title('Starting state - Fortran')

%% Setup problem for the time-dependent solver problem = DefaultMicroMagProblem(resolution(1),resolution(2),resolution(3)); problem.alpha = 4.42e3 ; problem.gamma = 2.21e5 ; % problem.dem_thres = 1e-2; problem.dem_thres = 1e-4; problem.dem_appr = getMicroMagDemagApproximation('threshold'); problem = problem.setTime( linspace(0,1e-9,200) ); % problem.setTimeDis = int32(10);

    problem = problem.setUseCuda( true );

    if (NIST_field == 1)
        %field 1
        HystDir = 1/mu0*[-24.6,4.3,0]/1000 ;
    end
    if (NIST_field == 2)
        %field 2
        HystDir = 1/mu0*[-35.5,-6.3,0]/1000 ;
    end

    HextFct = @(t) (t>-1)' .*HystDir;
    problem = problem.setHext( HextFct );

    problem.m0(:) = solution.M(end,:,:);

    %convert the class obj to a struct so it can be loaded into fortran
    solution_t = struct();
    prob_struct = struct(problem);

solution_t = MagTenseLandauLifshitzSolver_mex( prob_struct, solution_t ); toc

plot(fig1,problem.t,mean(solution_t.M(:,:,1),2),'rx'); plot(fig1,problem.t,mean(solution_t.M(:,:,2),2),'gx'); plot(fig1,problem.t,mean(solution_t.M(:,:,3),2),'bx'); % figure; hold all; for i=2:4; plot(problem.Hext(:,1),problem.Hext(:,i),'.'); end;

%% Run the Matlab version of the micromagnetism code % addpath('....\micromagnetism') % tic % Matlab_model_params.nGrid = resolution'; % Matlab_model_params.Field_dir = HystDir; % Script_3D_Std_Problem_4(fig1,Matlab_model_params); % toc % % %% Compare with published solutions available from NIST webpage % t=linspace(0,1,1000); % data = load(['Published_solutions_field' num2str(NIST_field)]); % mean_avg=data.mean_avg; err=data.err; % colours = [[1 0 0];[0 1 0];[0 0 1]]; % weak_colours = colours + ~colours0.75; % fill_ts=[t,fliplr(t)];
% for j=1:3 % std_errors{j}(1:2,:)=[mean_avg(1,:,j)+err(1,:,j);mean_avg(1,:,j)-err(1,:,j)]; % interval = [std_errors{j}(1,:),fliplr(std_errors{j}(2,:))]; % fill(fig1,1e-9fill_ts,interval,weak_colours(j,:),'linestyle','none') % plot(fig1,1e-9*t,mean_avg(1,:,j),'color',colours(j,:)) % end % % legend(fig1,'Fortran Mx','Fortran My','Fortran Mz','Matlab Mx','Matlab My','Matlab Mz','NIST \sigma{}(Mx)','NIST ','NIST \sigma{}(My)','NIST ','NIST \sigma{}(Mz)','NIST '); % ylabel(fig1,'/M_s') % xlabel(fig1,'Time [ns]') % % xlim(fig1,[0 1e-9]) % figure(figure1) % %

[kasparn]

I think this bug is due to bad memory management in the cuda C code from my side. It's very likely "just" memory that isn't freed correctly.

[rasmusbj]

I agree, it must be due to memory management. However you @kasparn coded that part, so maybe it is easiest if you debug it. It may be related to the allocation and deallocation of the demag tensor, as that is what changes between dense and sparse

[kasparn]

Fixed. Some pointers to arrays in the C-code where not set to NULL after being freed. They could then not be initialized upon subsequent re-call of the function (as it checks on NULL). That caused the crashes.