Bug (?) in StationaryDist_FHorz_Iteration_SemiExo_nProbs_raw

aledinola · March 9, 2026, 12:12pm

Running the function StationaryDist_FHorz_Case1, I got this error message:

Error using gpuArray/reshape
Sparse gpuArray matrices are not supported for this function.

Error in StationaryDist_FHorz_Iteration_SemiExo_nProbs_raw (line 39)
    StationaryDist_jj=reshape(Gammatranspose*StationaryDist_jj,[N_a*N_semiz,N_z]);
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Error in StationaryDist_FHorz_SemiExo (line 100)
        StationaryDist=StationaryDist_FHorz_Iteration_SemiExo_nProbs_raw(jequaloneDist,AgeWeightParamNames,Policy_dsemiexo,Policy_aprime,PolicyProbs,2,N_a,N_semiz,N_z,N_j,pi_semiz_J,pi_z_J,Parameters);
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Error in StationaryDist_FHorz_Case1 (line 183)
    StationaryDist=StationaryDist_FHorz_SemiExo(jequaloneDist,AgeWeightParamNames,Policy,n_d,n_a,simoptions.n_semiz,n_z,N_j,simoptions.pi_semiz_J,pi_z_J,Parameters,simoptions);
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

I think it might be related to me using R2024b instead of R2025b. @robertdkirkby I remember that in a previous post you said that the Tan improvement was running on CPU, not GPU, so I am a bit surprised.

aledinola · March 9, 2026, 12:34pm

I ran the code on R2025b and it runs without error messages.
I realized that SimPanelValues_FHorz_Case1 starts a parallel pool, is this really used?
Thanks

robertdkirkby · March 9, 2026, 7:49pm

Panel data simulations are being done with parfor on CPU (at least they should be).

The SimPanelValues commands call SimPanelIndexes commands, and these contain parfor-loops that call the SimTimeSeriesIndexes commands. For example, you can see this on lines 211-217 of

github.com/vfitoolkit/VFIToolkit-matlab

SimulateTimeSeries/InfHorz/SimPanelIndexes/SimPanelIndexes_InfHorz.m

master

function SimPanel=SimPanelIndexes_InfHorz(InitialDist,Policy,n_d,n_a,n_z,pi_z, simoptions, CondlProbOfSurvival, Parameters)
% Input must already be on CPU
% Simulates a panel based on PolicyIndexes of 'numbersims' agents of length 'simperiods' beginning from randomly drawn InitialDist. 
%
% CondlProbOfSurvival is an optional input. Only needed when using: simoptions.exitinpanel=1, there there is exit, either exog, endog or mix of both.
% Parameters is an optional input. Only needed when you have mixed (endogenous and exogenous) exit.
%
% simoptions are already set, this is an internal use only command
% (SimPanelIndexes is called by SimPanelValues)

%%
simoptions.numbersims=gather(simoptions.numbersims); % This is just to deal with weird error that matlab decided simoptions.numbersims was on gpu and so couldn't be an input to rand()
simoptions.simperiods=gather(simoptions.simperiods);
simoptions.burnin=gather(simoptions.burnin);

simoptions.simperiodsfinal=simoptions.simperiods;
simoptions.simperiods=simoptions.simperiods+simoptions.burnin; % just simulate this many periods, then burnin is removed later

if ~isfield(simoptions,'exitinpanel')
    simoptions.exitinpanel=0;

This file has been truncated. show original

aledinola · March 9, 2026, 8:09pm

Thanks for the clarification, makes sense to parallelize over individuals/observations since they are independent.

Regarding the Matlab version: my understanding is that then the toolkit command for the distribution requires R2025, correct?

robertdkirkby · March 9, 2026, 8:30pm

StationaryDist_FHorz_Iteration_SemiExo_nProbs_raw

should now be on the CPU. I added a gather(). If it still errors tell me (hard to debug as I am on R2025a so it just runs).

aledinola · March 9, 2026, 10:52pm

I can confirm that it runs on Matlab R2024b.