Calibrate GPU dispatch thresholds for this machine

Runs micro-benchmarks for each (op, backend, size) combination and derives the minimum matrix element count at which each GPU backend reliably outperforms CPU. Results are stored in the current session and optionally persisted to disk for reuse in future sessions.

Usage

amatrix_calibrate(
  backend = NULL,
  ops = c("gemm", "gemv", "crossprod", "rowSums", "colSums", "qr", "chol", "solve",
    "svd"),
  sizes = list(c(64L, 32L), c(128L, 64L), c(256L, 128L), c(512L, 256L), c(1024L, 512L)),
  sparse_densities = c(0.01, 0.05, 0.2),
  n_reps = 10L,
  margin = 0.1,
  persist = TRUE,
  quiet = FALSE
)

Arguments

backend: Character vector of backend names to benchmark. Defaults to all registered non-CPU backends that report available = TRUE.
ops: Character vector of operations to benchmark. Supported values: "matmul" (alias for "gemm"), "gemm", "gemv", "spmv", "spmm", "crossprod", "rowSums", "colSums", "qr", "chol", "solve", "svd".
sizes: List of integer vectors of length 2 giving c(nrow, ncol) test matrix dimensions.
sparse_densities: Numeric vector of target fill densities used when benchmarking "spmv" and "spmm".
n_reps: Positive integer. Number of timed repetitions per benchmark cell, after 2 warm-up repetitions.
margin: Non-negative numeric less than 1. Fraction by which GPU median time must beat CPU to count as a GPU win (default 0.10 means GPU must be at least 10% faster).
persist: Logical. If TRUE (default), save calibration to the user cache directory so future sessions load it automatically.
quiet: Logical. Suppress progress messages.

Value

Invisibly, a list with elements version, calibrated_at (POSIXct), thresholds (nested list keyed by backend then op), and results (data.frame of all benchmark measurements).

Usage

Arguments

Value

See also