shard_crossprod: Parallel crossprod() using shard views + output buffers

In shard: Deterministic, Zero-Copy Parallel Execution for R

Parallel crossprod() using shard views + output buffers

Description

Computes crossprod(X, Y) (i.e. t(X) %*% Y) using:

• shared/mmap-backed inputs (one copy),

• block views (no slice materialization),

• BLAS-3 dgemm in each tile,

• an explicit shared output buffer (no gather/bind spikes).

Usage

shard_crossprod(
  X,
  Y,
  workers = NULL,
  block_x = "auto",
  block_y = "auto",
  backing = c("mmap", "shm"),
  materialize = c("auto", "never", "always"),
  materialize_max_bytes = 512 * 1024^2,
  diagnostics = TRUE
)

Arguments

X, Y	Double matrices with the same number of rows.
workers	Number of worker processes.
block_x, block_y	Tile sizes over ncol(X) and ncol(Y). Use "auto" (default) to autotune on the current machine.
backing	Backing for shared inputs and output buffer ("mmap" or "shm").
materialize	Whether to return the result as a standard R matrix: "never" (return buffer handle), "always", or "auto" (materialize if estimated output size is below materialize_max_bytes).
materialize_max_bytes	Threshold for "auto" materialization.
diagnostics	Whether to collect shard_map diagnostics.

Details

This is intended as an ergonomic entry point for the "wow" path: users shouldn't have to manually call share(), view_block(), buffer(), tiles2d(), and shard_map() for common patterns.

Value

A list with:

• buffer: shard_buffer for the result (p x v)

• value: materialized matrix if requested, otherwise NULL

• run: the underlying shard_result from shard_map

• tile: chosen tile sizes

Examples

X <- matrix(rnorm(2000), 100, 20)
Y <- matrix(rnorm(2000), 100, 20)
res <- shard_crossprod(X, Y, block_x = 50, block_y = 10, workers = 2)
pool_stop()
res$value

shard documentation built on April 3, 2026, 9:08 a.m.