I am trying to gain an understanding why rstan models take much longer to run than the same model built and compiled with cmdstan (through cmdstanr) in R. One model takes 0.1-0.2 seconds for cmdstan but 2-3 seconds for rstan for the same numbers of iterations. For context, this is in an R package where an MCMC step is included in model fitting. Currently it uses cmdstanr for that step but I wanted to compile the Stan program as part of the package so looked at rstan, ideally incorporating the Stan program as part of the broader C++ program making use of rstan’s stan_fit class (although I haven’t got this work yet). An example of the Stan program is at the bottom. I have tried compiling the R package with rstan using compiler options like -fno-math-errno -O3, but it comes nowhere close to cmdstanr’s run times. Is there any way of getting rstan to run comparably to cmdstanr?
An example of the Stan program:
functions {
real partial_sum1_lpdf(array[] real y, int start, int end){
return std_normal_lpdf(y[start:end]);
}
real partial_sum2_lpdf(array[] real y,int start, int end, vector mu,vector sigma){
return normal_lpdf(y[start:end]|mu[start:end],sigma[start:end]);
}
}
data {
int N;
int Q;
vector[N] Xb;
matrix[N,Q] Z;
array[N] real y;
vector[N] sigma;
}
parameters {
array[Q] real gamma;
}
model {
int grainsize = 1;
target += reduce_sum(partial_sum1_lpdf,gamma,grainsize);
target += reduce_sum(partial_sum2_lpdf,y,grainsize,Xb + Z*to_vector(gamma),sqrt(sigma));
}