Skip to contents

Parallel log(x + eps) / log(base) on dense numeric input.

Source: R/fast_log.R
fast_log.Rd

Mirrors the fast-path that % Log takes inside the query evaluator, but callable directly on an in-memory numeric matrix or vector without wrapping the input in a memory_daf(). Intended for downstream packages whose hot paths already have the matrix materialised in R (e.g. a session-level cache) and want to avoid the single-threaded log(x + eps, base) the query fallback would otherwise incur.

Usage

fast_log(
  x,
  eps = 0,
  base = exp(1),
  threshold = dafr_opt("dafr.kernel_threshold")
)

Arguments

x

Dense numeric matrix or numeric vector. Integers are silently promoted to double.

eps

Non-negative scalar added inside the log. Default 0.

base

Positive scalar. Default exp(1).

threshold

Parallel-for threshold: enable OpenMP only when length(x) >= threshold. Defaults to dafr_opt("dafr.kernel_threshold").

Value

A numeric matrix or vector matching the shape of x. Names and dimnames are preserved.

Details

base values of 2, 10, and exp(1) dispatch to std::log2, std::log10, and std::log respectively — matching R's preference for the more accurate form over log(x) / log(base). Result is equal to log(x + eps, base) to machine precision (verified by the query-julia-compat parity suite).

See also

top_k_per_col() for a similarly scoped OpenMP primitive.

Examples

v <- c(1, 10, 100, 1000)
fast_log(v, base = 10)
#> [1] 0 1 2 3

m <- matrix(runif(12, 0.1, 5), nrow = 3, ncol = 4)
fast_log(m, eps = 1e-5, base = 2)
#>           [,1]       [,2]      [,3]        [,4]
#> [1,] -1.012503 -0.2003633 1.3443281  1.95805029
#> [2,]  2.066345 -2.8754891 0.6039475  2.13275885
#> [3,]  1.605844  1.2541938 1.8840632 -0.06307895