Usage

Drop-in replacement

Using cuNumeric as a replacement for NumPy is easy. Users only need to replace:

import numpy as np

with

import cunumeric as np

Running cuNumeric programs

Using legate

The preferred way to run cuNumeric programs is with the legate driver:

legate --cpus 2 --gpus 2 script.py <script options>

All legate command line configuration options must come before the script that is be executed. Any options after the script are passed on to the script.

It is also possible to pass the legate command line configuration options using the LEGATE_CONFIG environment variable:

LEGATE_CONFIG="--cpus 2 --gpus 2" legate script.py <script options>

See the Configuration section Legate driver for more information.

Using standard Python

It is also possible to run cuNumeric programs using the standard Python interpreter, with some limitations:

LEGATE_CONFIG="--cpus 2 --gpus 2" python script.py <script options>

When running programs with this method, configuration options may only be passed via the LEGATE_CONFIG environment variable as shown above.

Additionally, several legate command line configuration options are not available when running programs this way. See the output of legate --help for more details.

Multi-node execution

Using legate

Cunumeric programs can be run in parallel by using the --nodes option to the legate driver, followed by the number of nodes to be used. When running on 2+ nodes, a task launcher must be specified.

Legate currently supports using mpirun, srun, and jsrun as task launchers for multi-node execution via the --launcher command like arguments:

legate --launcher srun --nodes 2 script.py <script options>

See the Configuration section Multi-node execution for more configuration options.

Using a manual task manager

mpirun -np N legate script.py <script options>

It is also possible to use “standard python” in place of the legate driver.

Zero code-change patching

The lgpatch script (in the same location as the legate executable) can help facilitate quick demonstrations of cunumeric on existing codebases that make use of numpy.

To use this tool, invoke it as shown below, with the name of the program to patch:

lgpatch <program> -patch numpy

For example, here is a small test.py program that imports and uses various numpy funtions:

# test.py

import numpy as np
input = np.eye(10, dtype=np.float32)
np.linalg.cholesky(input)

You can invoke lgpatch to run test.py using cunumeric functions instead, without any changes to the original source code. Any standard cunumeric runtime options (e.g. for Measuring API coverage) may also be used:

$ CUNUMERIC_REPORT_COVERAGE=1 LEGATE_CONFIG="--cpus 4"  lgpatch test.py -patch numpy
cuNumeric API coverage: 4/4 (100.0%)