CNPreferences

Function reference for CNPreferences. Preferences write LocalPreferences.toml and generally require a fresh Julia process.

Out of the box (no LocalPreferences.toml changes):

Setting	Default
Binary / build mode	JLL prebuilt binaries
Broadcast fusion	on
FUSE_BROADCAST_MIN_OPS	2 (single native operations stay unfused)
Task scope names	off
MIN_SOLVE_MATRIX_SIZE	2048 rows
MIN_SOLVE_TILE_SIZE	512
MIN_CHOLESKY_MATRIX_SIZE	8192 rows
MIN_CHOLESKY_TILE_SIZE	2048
MIN_QR_MATRIX_SIZE	1048576 elements
QR_TILE_SIZE	128
MAX_CHOLESKY_TILES_PER_PROC	4

Build-mode setup (JLL / developer) is documented under Build Modes.

Linear algebra

set_linalg! accepts the constant names above as keywords. All values must be positive integers. Unspecified settings are unchanged. Restart Julia to load the new constants. cuNumeric.versioninfo() shows the effective values.

CNPreferences.set_linalg!(; MIN_SOLVE_MATRIX_SIZE=4096, MIN_SOLVE_TILE_SIZE=512)

See Tuning for selection rules and tuning details.

CNPreferences.set_linalg! Function

set_linalg!(; export_prefs=false, force=true, settings...)

Set linear algebra tuning preferences using their constant names as keywords: MIN_SOLVE_MATRIX_SIZE (2048), MIN_SOLVE_TILE_SIZE (512), MIN_CHOLESKY_MATRIX_SIZE (8192), MIN_CHOLESKY_TILE_SIZE (2048), MIN_QR_MATRIX_SIZE (1048576 elements), QR_TILE_SIZE (128), and MAX_CHOLESKY_TILES_PER_PROC (4). All values must be positive integers. Unspecified settings are unchanged. Restart Julia after changing preferences.

CNPreferences.set_linalg!(; MIN_SOLVE_MATRIX_SIZE=4096, MIN_SOLVE_TILE_SIZE=512)

source

Build mode

CNPreferences.use_jll_binary()      # use the prebuilt wrapper
CNPreferences.use_developer_mode()  # build the wrapper against prebuilt dependencies

Restart Julia after selecting the mode. For developer mode, rebuild the wrapper with Pkg.build("cuNumeric"). See Developer Mode.

Broadcast fusion

Defaults: fusion on, FUSE_BROADCAST_MIN_OPS == 2.

using CNPreferences

CNPreferences.enable_broadcast_fusion!()           # default
CNPreferences.disable_broadcast_fusion!()
CNPreferences.set_broadcast_fusion_min_ops!(2)     # default
CNPreferences.set_broadcast_fusion_min_ops!(1)     # also fuse single-ops

set_broadcast_fusion_min_ops! counts Broadcasted nodes (ops) in the tree:

• 2 (default): fuse multi-op trees such as y .= @. a * b + c. Single native operations like y .= cos.(x) use the C-API path. Functions without a native implementation are fused on the GPU when their broadcast arguments are eligible.

• 1: fuse every eligible expression, including single-ops.

Set the preference in one Julia process, then start a fresh process to use it.

CNPreferences.set_broadcast_fusion! Function

set_broadcast_fusion!(enabled::Bool; export_prefs=false, force=true)

Enable or disable broadcast fusion. When enabled (the default), eligible nested broadcast expressions compile to a single CUDA PTX kernel.

Restart Julia after changing this preference.

source

CNPreferences.enable_broadcast_fusion! Function

enable_broadcast_fusion!(; export_prefs=false, force=true)

Enable broadcast fusion. This is the default.

source

CNPreferences.disable_broadcast_fusion! Function

disable_broadcast_fusion!(; export_prefs=false, force=true)

Disable broadcast fusion so each broadcast node runs as a separate cuNumeric op.

source

CNPreferences.set_broadcast_fusion_min_ops! Function

set_broadcast_fusion_min_ops!(n::Integer; export_prefs=false, force=true)

Fuse only when a broadcast tree has at least n ops. Default is 2, so single ops such as y .= cos.(x) stay on the unfused path. Set n = 1 to fuse those too. Values must be >= 1.

Restart Julia after changing this preference.

source

Task scope names

Default: off. Optional Legate task-scope naming for debugging. When on, cuNumeric wraps many ops in Legate.with_scope so provenance strings (for example matmul, zeros, or fused broadcast.<expr>) appear in Legate logs and profiles. Pair this with --logging legate=debug --log-to-file (or --profile) in LEGATE_CONFIG. See Debugging.

using CNPreferences
CNPreferences.enable_task_scope_names!()
CNPreferences.disable_task_scope_names!()  # default

Restart Julia after changing this preference (it is compile-time in cuNumeric.jl).

CNPreferences.set_task_scope_names! Function

set_task_scope_names!(enabled::Bool; export_prefs=false, force=true)

Enable or disable named Legate task scopes for debugging. Default is off.

When enabled, cuNumeric wraps ops in Legate.with_scope so provenance labels appear in Legate logs/profiles. Pair with LEGATE_CONFIG flags such as --logging legate=debug --log-to-file (set before Julia starts). Requires a fresh Julia process after changing the preference.

source

CNPreferences.enable_task_scope_names! Function

enable_task_scope_names!(; export_prefs=false, force=true)

Enable named Legate task scopes for debugging. See set_task_scope_names!.

source

CNPreferences.disable_task_scope_names! Function

disable_task_scope_names!(; export_prefs=false, force=true)

Disable named Legate task scopes. This is the default.

source