Profile a fit by solver phase — tulpa

Times a fit one solver phase at a time – the inner Newton solve's eta, scatter (the Hessian and gradient assembly), factorize (Cholesky), line search and final pass, the nested-Laplace outer grid's cell solves, and the NUTS sampler's iterations and gradient evaluations – and returns the breakdown as a data frame. The accumulator aggregates across the parallel outer-grid worker threads and across concurrently run chains, so the reported times cover the whole fit rather than only the calling thread (and can therefore exceed its wall time).

Usage

tulpa_profile(expr, sort = TRUE)

Arguments

expr: An expression that runs a fit (for example a call to tulpa_laplace() or tulpa()). Evaluated once, after the profile counters are reset and timing is switched on.
sort: Logical; order rows by descending time. Default TRUE.

Value

A data frame with one row per phase and columns phase, seconds, calls, ms_per_call (mean wall time per phase call), and share (fraction of the total timed seconds of the leaf phases; NA for an enclosing phase). The fit result is attached as the "value" attribute.

Details

Use it to settle where a fit spends its time, e.g. whether a slow joint occu_cover() fit is bound by the assembly scatter or the Cholesky factorize:


  p <- tulpa_profile(
    tulpa_nested_laplace_joint(..., control = list(integration = "ccd"))
  )
  print(p)            # rows ordered by time; scatter vs factorize at top
  fit <- attr(p, "value")

Timing is off outside tulpa_profile(): an ordinary fit pays one flag read per phase scope and nothing else.

Which fits are timed

The single-response Laplace solve behind tulpa_laplace(), tulpa(mode = "laplace") and every per-cell inner solve of tulpa_nested_laplace() / tulpa(mode = "nested_laplace"): phases eta, scatter, factorize, line_search, log_det, log_lik_prior, hessian_extract and inner_diagnostics.
The joint solver behind tulpa_nested_laplace_joint(), on both its dense and its sparse path (the sparse one adds pattern_build and prep).
The nested-Laplace outer grid, single-response and joint: outer_grid_cell, one call per solved cell.
The samplers: nuts_warmup and nuts_sampling, one call per NUTS iteration per chain, and gradient, one call per log-density gradient evaluation (NUTS leapfrog steps, static HMC, the step-size search and the other samplers that evaluate the gradient through the engine).

An expression that reaches none of these – a Gibbs or quadrature backend that runs neither an engine Newton solve nor an engine gradient, or no fit at all – records nothing, and tulpa_profile() then warns rather than returning an all-zero table as if it were a measurement.

Enclosing phases

outer_grid_cell, nuts_warmup and nuts_sampling ENCLOSE the leaf phases timed inside them (a cell's inner Newton phases, an iteration's gradient evaluations), so their seconds overlap the leaves' rather than adding to them. Their share is NA; the shares of the leaf phases sum to one.

Examples

# \donttest{
set.seed(1)
n <- 200L; X <- cbind(1, rnorm(n))
y <- rbinom(n, 1, plogis(X %*% c(0, 0.5)))
tulpa_profile(tulpa_laplace(y, rep(1L, n), X, family = "binomial"))
# }