Preallocating buffer

techniques
performance
#1

I am estimating a multilevel model where for each group, a set of 2-element vectors \alpha, \beta, w is used, with the constraints

0 < \alpha, \beta < 1 \qquad \text{and} \qquad 0 < w

I chose impose a multivariate normal on \text{logit}(\alpha), \text{logit}(\beta), \log(w) (out of convenience).

I implement it in Stan using

  // logit transformation and log Jacobian correction
  real logit_lp(real x) {
    real y;
    y = log(x);
    target += -y - log1m(x);
    return y;
  }

  /* Cross sectional distribution of group-specific parameters.

     Transformed parameters ~ multi_normal_cholesky(cs_mean, cs_L)

     Tws > 0, 0 < alphas, betas < 1
   */
  void cross_sectional_distribution_lp(vector Tws, vector alphas, vector betas,
                                       vector cs_mean, matrix cs_L) {
    vector[6] z;
    z[1:2] = log(Tws);
    target += -sum(z[1:2]);  // log Jacobian adjustment for y = log(x)
    z[3] = logit_lp(alphas[1]);
    z[4] = logit_lp(alphas[2]);
    z[5] = logit_lp(betas[1]);
    z[6] = logit_lp(betas[2]);
    z ~ multi_normal_cholesky(cs_mean, cs_L);
  }

There are many groups. I found that it is significantly (about 4x) faster if I pre-allocate a buffer for z outside the function call, and pass it to the functions.

I am wondering if I could run into any trouble though, eg if Stan implicitly parallelizes code and two such functions could access it concurrently, or something like that.

#2

No

#3

I just realized the question is irrelevant, since I get

SYNTAX ERROR, MESSAGE(S) FROM PARSER:

Cannot assign to function argument variables.
Use local variables instead.

anyway. A typo in the original made me think that I can pre-allocate buffers, but in fact I cannot.