cmdstan eigen error vs rstan informative error

[spinkney]

Having a udf with a 0 index compiles fine in both rstan and cmdstan. At run time they produce different errors and I believe cmdstanr should produce a better error.

cmdstanr error

> mod_test$sample(data = list(N = 100, y = rnorm(100)),
+                chains = 2,
+                parallel_chains = 2
+ )
Running MCMC with 2 parallel chains...

Chain 1 test: stan/lib/stan_math/lib/eigen_3.3.7/Eigen/src/Core/DenseCoeffsBase.h:162: Eigen::DenseCoeffsBase<Derived, 0>::CoeffReturnType Eigen::DenseCoeffsBase<Derived, 0>::operator[](Eigen::Index) const [with Derived = Eigen::Matrix<double, -1, 1>; Eigen::DenseCoeffsBase<Derived, 0>::CoeffReturnType = const double&; Eigen::Index = long int]: Assertion `index >= 0 && index < size()' failed.

rstan error

> out_test <- sampling(mod_test_rstan,
+                 data = list(N = 100, y = rnorm(100)),
+                 chains = 2,
+                 iter = 1000)

SAMPLING FOR MODEL 'test' NOW (CHAIN 1).
Chain 1: Unrecoverable error evaluating the log probability at the initial value.
Chain 1: Exception: Exception: []: accessing element out of range. index 0 out of range; expecting index to be between 1 and 100; index position = 1x  (in 'model5a82748ab0ae_test' at line 7)
  (in 'model5a82748ab0ae_test' at line 19)

test model

functions {
  real does_nothing (vector x){
    return x[0];
  }
}
data {
  int<lower=0> N;
  vector[N] y;
}
parameters {
  real mu;
  real<lower=0> sigma;
}
transformed parameters {
  real z = does_nothing(y);
}
model {
  y ~ normal(mu, sigma);
}

versions

cmdstan_version()
[1] "2.24.0"

stan_version()
[1] "2.21.0"

[bbbales2]

This error seems to be triggered specifically in function calls with certain variables.

For instance this gives me a reasonable error (no function):

parameters {
  real a;
}

transformed parameters {
  vector[1] b;
  
  b[0] = 1.0;
}

model {
  a ~ std_normal();
}

This also gives me a reasonable error (matrix in function):

functions {
  real whatever(matrix a) {
    return a[0, 0];
  }
}

parameters {
  real a;
}

transformed parameters {
  matrix[1, 1] b;
  
  b[1, 1] = whatever(b);
}

model {
  a ~ std_normal();
}

And this gives me an Eigen error (row_vector does too):

functions {
  real whatever(vector a) {
    return a[0];
  }
}

parameters {
  real a;
}

transformed parameters {
  vector[1] b;
  
  b[1] = whatever(b);
}

model {
  a ~ std_normal();
}

[rok-cesnovar]

There was one attempt at fixing this: #521

But that proved to cause a performance regression. We should give this another go once the expression rework is done in /stan.

[rok-cesnovar]

The issue covering this is also still open: #489

We can merge the two issues I guess.