Add option to choose between NSYS and NCU profilers (#28)

* Add option to give nvcc extra arguments

* Add test for nvcc options that changes c++ dialect from c++17 to c++14

* Add make and the english language pack to devcontainer to be able to build the documentation

* Update documentation config to automatically import the current version of the package

* Document new --compiler-args argument

* Improve tests coverage by testing for bad arguments and the error output during a failed compilation

* Add IPython to docs requirements to allow the __version__ import for readthedocs env

* Change devcontainer base image to have the latest CUDA toolkit

* Mock the nsight compute tool with a bash script

* Add test to compile with opencv

* Add new page to documentation that contains a new notebook that explains compiling with external libraries

* Add autodocstring vscode extension to devcontainer

* Add function that modifies the default profiler/compiler arguments to allow reusing them in multiple magic command calls

* Update pylint exceptions

* Update contributing instructions

* Change version from 1.0.3 to 1.1.0 due to adding features in a backward-compatible manner

* Install latest CUDA toolkit on the test runner to pass the OpenCV compilation test

* Install opencv in test runner and update code coverage install

* Add CUDA bin to PATH in test and coverage runners

* Add cuda bin to path variable in .bashrc

* Update way to set environment variable PATH in github action

* Change devcontainer base image back to ubuntu:22.04 to match the environment from the test runner

* Add option to choose between NSYS and NCU profilers

* Add tests for choosing the profiler

* Add isort config to help it find local modules so they are not considered 3rd party libraries

* Replace experimental-string-processing black formatter config with enable-unstable-feature as it was removed in version 24.1.0

* Search for profiling tools executable paths when they are required

* Install dev dependencies in editable mode

* Add documentation for using Nsight Systems instead of the default Nsight Compute profiling tool

* Fix cuda typo

* Mention Nsight Systems in README.md

.devcontainer/post_create.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
 # install developer dependencies
-pip install .[dev]
+pip install -e .[dev]
 
 # make sure the developer uses pre-commit hooks
 pre-commit install

README.md

@@ -45,7 +45,7 @@ to own a GPU yourself.
 Here are just a few of the things that nvcc4jupyter does well:
 
   - [Easily run CUDA C++ code](https://nvcc4jupyter.readthedocs.io/en/latest/usage.html#hello-world)
-  - [Profile your code with NVIDIA Nsight Compute](https://nvcc4jupyter.readthedocs.io/en/latest/usage.html#profiling)
+  - [Profile your code with NVIDIA Nsight Compute or Nsight Systems](https://nvcc4jupyter.readthedocs.io/en/latest/usage.html#profiling)
   - [Compile your code with external libraries (e.g. OpenCV)](https://nvcc4jupyter.readthedocs.io/en/latest/notebooks.html#compiling-with-external-libraries)
   - [Share code between different programs in the same notebook / split your code into multiple files for improved readability](https://nvcc4jupyter.readthedocs.io/en/latest/usage.html#groups)
 
@@ -96,7 +96,7 @@ If not using the devcontainer you need to install the package with the
 development dependencies and install the pre-commit hook before commiting any
 changes:
 ```bash
-pip install .[dev]
+pip install -e .[dev]
 pre-commit install
 ```
 

docs/source/magics.rst

@@ -21,7 +21,7 @@ Usage
    - ``%%cuda``: Compile and run this cell.
    - ``%%cuda -p``: Also runs the Nsight Compute profiler.
    - ``%%cuda -p -a "<SPACE SEPARATED PROFILER ARGS>"``: Also runs the Nsight Compute profiler.
-   - ``%%cude -c "<SPACE SEPARATED COMPILER ARGS"``: Passes additional arguments to "nvcc".
+   - ``%%cuda -c "<SPACE SEPARATED COMPILER ARGS"``: Passes additional arguments to "nvcc".
    - ``%%cuda -t``: Outputs the "timeit" built-in magic results.
 
 Options
@@ -36,15 +36,35 @@ Options
 .. _profile:
 
 -p, --profile
-   Boolean. If set, runs the NVIDIA Nsight Compute profiler whose
+   Boolean. If set, runs the NVIDIA Nsight Compute (or NVIDIA Nsight Systems
-   output is appended to standard output.
+   if changed via the \-\-profiler option) profiler whose output is appended to
+   standard output.
+
+.. _profiler:
+
+-l, --profiler
+   String. Can either be "ncu" (the default) to use NVIDIA Nsight Compute
+   profiling tool, or "nsys" to use NVIDIA Nsight Systems profiling tool.
+
+.. _profiler_args:
 
 .. _profiler_args:
 
 -a, --profiler-args
    String. Optional profiler arguments that can be space separated
-   by wrapping them in double quotes. See all options here:
+   by wrapping them in double quotes. Will be passed to the profiler selected
-   `Nsight Compute CLI <https://docs.nvidia.com/nsight-compute/NsightComputeCli/index.html#command-line-options>`_
+   by the \-\-profiler option.. See profiler options here:
+   `Nsight Compute <https://docs.nvidia.com/nsight-compute/NsightComputeCli/index.html#command-line-options>`_
+   or `Nsight Systems <https://docs.nvidia.com/nsight-systems/UserGuide/index.html#command-line-options>`_.
+
+.. _compiler_args:
+
+-c, --compiler-args
+   String. Optional compiler arguments that can be space separated
+   by wrapping them in double quotes. They will be passed to "nvcc".
+   See all options here:
+   `NVCC Options <https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#nvcc-command-options>`_
+
 
 .. _compiler_args:
 

docs/source/usage.rst

@@ -225,10 +225,11 @@ Profiling
 ---------
 
 Another important feature of nvcc4jupyter is its integration with the NVIDIA
-Nsight Compute profiler, which you need to make sure is installed and its
+Nsight Compute / NVIDIA Nsight Systems profilers, which you need to make sure
-executable can be found in a directory in your PATH environment variable.
+are installed and the executables can be found in a directory in your PATH
+environment variable.
 
-In order to use it and provide the profiler with custom arguments, simply run:
+To profile using Nsight Compute with custom arguments:
 
 .. code-block:: c++
 
@@ -256,6 +257,27 @@ Running the cell above will compile and execute the vector addition code in the
     Compute (SM) Throughput             %         1.19
     ----------------------- ------------- ------------
 
+To profile using Nsight Systems with custom arguments:
+
+.. code-block:: c++
+
+    %cuda_group_run --group "vector_add" --profiler nsys --profile --profiler-args "profile --stats=true"
+
+Running the cell above will compile and execute the vector addition code in the
+"vector_add" group and profile it with Nsight Systems. The output will contain
+multiple tables, one of which will look similar to this:
+
+.. code-block::
+
+    [5/8] Executing 'cuda_api_sum' stats report
+
+    Time (%)  Total Time (ns)  Num Calls    Avg (ns)       Med (ns)      Min (ns)     Max (ns)    StdDev (ns)           Name
+    --------  ---------------  ---------  -------------  -------------  -----------  -----------  -----------  ----------------------
+        77.3      200,844,276          1  200,844,276.0  200,844,276.0  200,844,276  200,844,276          0.0  cudaMalloc
+        22.6       58,594,762          2   29,297,381.0   29,297,381.0   29,153,999   29,440,763    202,772.8  cudaMemcpy
+         0.1          305,450          1      305,450.0      305,450.0      305,450      305,450          0.0  cudaLaunchKernel
+         0.0            1,970          1        1,970.0        1,970.0        1,970        1,970          0.0  cuModuleGetLoadingMode
+
 Compiler arguments
 ------------------
 

nvcc4jupyter/__init__.py

@@ -2,7 +2,7 @@
 nvcc4jupyter: CUDA C++ plugin for Jupyter Notebook
 """
 
-from .parsers import set_defaults  # noqa: F401
+from .parsers import Profiler, set_defaults  # noqa: F401
 from .plugin import NVCCPlugin, load_ipython_extension  # noqa: F401
 
 __version__ = "1.1.0"

nvcc4jupyter/parsers.py

@@ -3,14 +3,28 @@ Parsers for the CUDA magic commands.
 """
 
 import argparse
-from typing import Callable, Optional
+from enum import Enum
+from typing import Callable, Optional, Type, TypeVar
 
+
+class Profiler(Enum):
+    """Choice between Nsight Compute and Nsight Systems profilers."""
+
+    NCU = "ncu"
+    NSYS = "nsys"
+
+
+_default_profiler: Profiler = Profiler.NCU
 _default_profiler_args: str = ""
 _default_compiler_args: str = ""
 
+T = TypeVar("T")
+
 
 def set_defaults(
-    compiler_args: Optional[str] = None, profiler_args: Optional[str] = None
+    profiler: Optional[Profiler] = None,
+    compiler_args: Optional[str] = None,
+    profiler_args: Optional[str] = None,
 ) -> None:
     """
     Set the default values for various arguments of the magic commands. These
@@ -18,17 +32,22 @@ def set_defaults(
     to override this behaviour on a cell by cell basis.
 
     Args:
+        profiler: If not None, this value becomes the new default profiler.
+            Defaults to None.
         compiler_args: If not None, this value becomes the new default compiler
-            config. Defaults to "".
+            config. Defaults to None.
         profiler_args: If not None, this value becomes the new default profiler
-            config. Defaults to "".
+            config. Defaults to None.
     """
 
     # pylint: disable=global-statement
+    global _default_profiler
+    if profiler is not None:
+        _default_profiler = profiler
     global _default_compiler_args
-    global _default_profiler_args
     if compiler_args is not None:
         _default_compiler_args = compiler_args
+    global _default_profiler_args
     if profiler_args is not None:
         _default_profiler_args = profiler_args
 
@@ -38,6 +57,11 @@ def str_to_lambda(arg: str) -> Callable[[], str]:
     return lambda: arg
 
 
+def class_to_lambda(arg: str, cls: Type[T]) -> Callable[[], T]:
+    """Convert string value to class and then to lambda"""
+    return lambda: cls(arg)
+
+
 def get_parser_cuda() -> argparse.ArgumentParser:
     """
     %%cuda magic command parser.
@@ -52,8 +76,14 @@ def get_parser_cuda() -> argparse.ArgumentParser:
     parser.add_argument("-t", "--timeit", action="store_true")
     parser.add_argument("-p", "--profile", action="store_true")
 
-    # --profiler-args and --compiler-args values are lambda functions to allow
+    # the type of the following arguments is a lambda lambda function to allow
     # changing the default value at runtime
+    parser.add_argument(
+        "-l",
+        "--profiler",
+        type=lambda arg: class_to_lambda(arg, cls=Profiler),
+        default=lambda: _default_profiler,
+    )
     parser.add_argument(
         "-a",
         "--profiler-args",

nvcc4jupyter/path_utils.py

@@ -0,0 +1,61 @@
+"""
+Helper functions relating to file paths.
+"""
+
+import os
+from glob import glob
+from typing import List, Optional
+
+CUDA_SEARCH_PATHS: List[str] = [
+    "/opt/nvidia/nsight-compute",
+    "/usr/local/cuda",
+    "/opt",
+    "/usr",
+]
+
+
+def is_executable(fpath: str) -> bool:
+    """Check if file exists and is executable"""
+    return os.path.isfile(fpath) and os.access(fpath, os.X_OK)
+
+
+def which(name: str) -> Optional[str]:
+    """Find an executable by name by searching the PATH directories"""
+    for path_dir in os.environ.get("PATH", "").split(os.pathsep):
+        exec_path = os.path.join(path_dir, name)
+        if is_executable(exec_path):
+            return exec_path
+    return None
+
+
+def find_executable(
+    name: str, search_paths: Optional[List[str]] = None
+) -> Optional[str]:
+    """
+    Find an executable, either by searching in the directories of the PATH
+    environment variable or, if that did not work, by searching recursively
+    in directories a list given as parameter.
+
+    Args:
+        name: The name of the executable to be found.
+        search_paths: If None, only executables that are available from PATH
+            will be found. Otherwise, will recursively search these
+            directories. Defaults to None.
+
+    Returns:
+        The path to the executable if it is found, and None otherwise.
+    """
+    if search_paths is None:
+        search_paths = []
+
+    which_path = which(name)
+    if which_path is not None:
+        return which_path
+
+    for search_path in search_paths:
+        search_path = os.path.abspath(search_path)
+        search_path = os.path.join(search_path, f"**/{name}")
+        for exec_path in glob(search_path, recursive=True):
+            return exec_path
+
+    return None

nvcc4jupyter/plugin.py

@@ -9,13 +9,20 @@ import shutil
 import subprocess
 import tempfile
 import uuid
-from typing import List, Optional
+from typing import Dict, List, Optional
 
 # pylint: disable=import-error
 from IPython.core.interactiveshell import InteractiveShell
 from IPython.core.magic import Magics, cell_magic, line_magic, magics_class
 
-from . import parsers
+from .parsers import (
+    Profiler,
+    get_parser_cuda,
+    get_parser_cuda_group_delete,
+    get_parser_cuda_group_run,
+    get_parser_cuda_group_save,
+)
+from .path_utils import CUDA_SEARCH_PATHS, find_executable
 
 DEFAULT_EXEC_FNAME = "cuda_exec.out"
 SHARED_GROUP_NAME = "shared"
@@ -37,14 +44,19 @@ class NVCCPlugin(Magics):
         super().__init__(shell)
         self.shell: InteractiveShell  # type hint not provided by parent class
 
-        self.parser_cuda = parsers.get_parser_cuda()
+        self.parser_cuda = get_parser_cuda()
-        self.parser_cuda_group_save = parsers.get_parser_cuda_group_save()
+        self.parser_cuda_group_save = get_parser_cuda_group_save()
-        self.parser_cuda_group_delete = parsers.get_parser_cuda_group_delete()
+        self.parser_cuda_group_delete = get_parser_cuda_group_delete()
-        self.parser_cuda_group_run = parsers.get_parser_cuda_group_run()
+        self.parser_cuda_group_run = get_parser_cuda_group_run()
 
         self.workdir = tempfile.mkdtemp()
         print(f'Source files will be saved in "{self.workdir}".')
 
+        self.profiler_paths: Dict[Profiler, Optional[str]] = {
+            Profiler.NCU: None,
+            Profiler.NSYS: None,
+        }
+
     def _save_source(
         self, source_name: str, source_code: str, group_name: str
     ) -> None:
@@ -135,11 +147,42 @@ class NVCCPlugin(Magics):
 
         return executable_fpath
 
-    def _run(
+    def _get_profiler_path(self, profiler: Profiler) -> str:
+        """
+        Get the path of the executable of a given profiling tool. Searches
+        the directories of the PATH environment variable and some extra
+        directories where CUDA is usually installed.
+
+        Args:
+            profiler: The profiler whose executable should be found.
+
+        Raises:
+            RuntimeError: If the profiler executable could not be found.
+
+        Returns:
+            The file path of the executable.
+        """
+        profiler_path = self.profiler_paths[profiler]
+        if profiler_path is not None:
+            return profiler_path
+
+        profiler_path = find_executable(profiler.value, CUDA_SEARCH_PATHS)
+        if profiler_path is None:
+            raise RuntimeError(
+                f'Could not find the "{profiler.value}" profiling tool.'
+                " Consider searching for where it is installed and adding its"
+                " directory to the PATH environment variable."
+            )
+
+        self.profiler_paths[profiler] = profiler_path
+        return profiler_path
+
+    def _run(  # pylint: disable=too-many-arguments
         self,
         exec_fpath: str,
         timeit: bool = False,
         profile: bool = False,
+        profiler: Profiler = Profiler.NCU,
         profiler_args: str = "",
     ) -> str:
         """
@@ -150,8 +193,9 @@ class NVCCPlugin(Magics):
             timeit: If True, returns the result of the "timeit" magic instead
                 of the standard output of the CUDA process. Defaults to False.
             profile: If True, the executable is profiled with NVIDIA Nsight
-                Compute profiling tool and its output is added to stdout.
+                Compute or NVIDIA Nsight Systems and the profiling output is
-                Defaults to False.
+                added to stdout. Defaults to False.
+            profiler: The profiling tool to use.
             profiler_args: The profiler arguments used to customize the
                 information gathered by it and its overall behaviour. Defaults
                 to an empty string.
@@ -173,7 +217,8 @@ class NVCCPlugin(Magics):
         else:
             run_args = []
             if profile:
-                run_args.extend(["ncu"] + profiler_args.split())
+                profiler_path = self._get_profiler_path(profiler)
+                run_args.extend([profiler_path] + profiler_args.split())
             run_args.append(exec_fpath)
             output = subprocess.check_output(
                 run_args, stderr=subprocess.STDOUT
@@ -194,6 +239,7 @@ class NVCCPlugin(Magics):
                 exec_fpath=exec_fpath,
                 timeit=args.timeit,
                 profile=args.profile,
+                profiler=args.profiler(),
                 profiler_args=args.profiler_args(),
             )
         except subprocess.CalledProcessError as e:

pyproject.toml

@@ -69,6 +69,7 @@ exclude_lines = [
 
 [tool.isort]
 profile = "black"
+src_paths = ["nvcc4jupyter"] # tells isort where to find local modules to not consider them 3rd party libraries
 
 [tool.bandit]
 exclude_dirs = ["build","dist","tests","scripts"]
@@ -82,7 +83,8 @@ skips = ["B101", "B311", "B404", "B603"]
 [tool.black]
 line-length = 79
 fast = true
-experimental-string-processing = true
+preview = true
+enable-unstable-feature = ["string_processing"]
 
 [tool.coverage.run]
 branch = true

tests/fixtures/fixtures.py

@@ -1,9 +1,11 @@
+import argparse
 import glob
 import os
 
 import pytest
 from IPython.core.interactiveshell import InteractiveShell
 
+from nvcc4jupyter.parsers import Profiler
 from nvcc4jupyter.plugin import NVCCPlugin
 
 
@@ -70,3 +72,14 @@ def multiple_source_fpaths(fixtures_path: str):
     pattern_h = os.path.join(fixtures_path, "multiple_files", "*.h")
     pattern_cu = os.path.join(fixtures_path, "multiple_files", "*.cu")
     return list(glob.glob(pattern_h)) + list(glob.glob(pattern_cu))
+
+
+@pytest.fixture(scope="session")
+def default_args():
+    return argparse.Namespace(
+        timeit=False,
+        profile=True,
+        profiler=lambda: Profiler.NCU,
+        profiler_args=lambda: "",
+        compiler_args=lambda: "",
+    )

tests/fixtures/scripts/ncu

@@ -1,7 +1,7 @@
 #!/bin/bash
 
+echo "[NCU]"
+
 # this is a mock of nsight compute cli tool that just executes the program
 # given as the last argument
 "${@: -1}"
-
-echo "==WARNING== No kernels were profiled"

tests/fixtures/scripts/nsys

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+echo "[NSYS]"
+
+# this is a mock of nsight systems cli tool that just executes the program
+# given as the last argument
+"${@: -1}"

tests/fixtures/scripts/searchforme

@@ -0,0 +1,3 @@
+#!/bin/bash
+
+echo "This is just used to test the path_utils.find_executable function"

tests/test_path_utils.py

@@ -0,0 +1,16 @@
+import os
+
+from nvcc4jupyter.path_utils import find_executable
+
+
+def test_which():
+    assert find_executable("ls") == "/usr/bin/ls"
+
+
+def test_find_executable(fixtures_path: str):
+    exec_path = find_executable("searchforme", [fixtures_path])
+    assert exec_path is not None
+
+    exec_dir, exec_fname = os.path.split(exec_path)
+    assert exec_fname == "searchforme"
+    assert os.path.basename(exec_dir) == "scripts"

tests/test_plugin.py

@@ -1,29 +1,26 @@
-import argparse
 import math
 import os
 import re
 import shutil
 import subprocess
+from argparse import ArgumentParser, Namespace
+from copy import deepcopy
 from typing import List
 
 import pytest
 
-from nvcc4jupyter.parsers import get_parser_cuda, set_defaults
+from nvcc4jupyter.parsers import Profiler, get_parser_cuda, set_defaults
 from nvcc4jupyter.plugin import NVCCPlugin
 
 
-def check_profiler_output(output: str):
+def check_profiler_output(output: str, profiler: str = "[NCU]"):
-    # the profiler output will be a line of "Hello World!" along with some
+    # the output from the profiler will first be a line containing only
-    # warning lines which start with "==WARNING=="
+    # "[NCU]" or "[NSYS]" depending on what profiler was used and another
+    # line containing the string "Hello World!"
     lines = output.strip().split("\n")
-    warn_count = 0
+    assert len(lines) >= 2
-    for line in lines:
+    assert lines[0] == profiler
-        if not line.startswith("==WARNING=="):
+    assert lines[1] == "Hello World!"
-            assert line == "Hello World!"
-        else:
-            warn_count += 1
-    assert warn_count >= 1
-    assert warn_count == len(lines) - 1
 
 
 def copy_source_to_group(
@@ -46,7 +43,7 @@ def before_all(scripts_path: str):
 @pytest.fixture(autouse=True, scope="function")
 def before_each(plugin: NVCCPlugin):
     # BEFORE TESTS
-    set_defaults(compiler_args="", profiler_args="")
+    set_defaults(profiler=Profiler.NCU, compiler_args="", profiler_args="")
     shutil.rmtree(plugin.workdir, ignore_errors=True)
     yield
     # AFTER TESTS
@@ -101,6 +98,7 @@ def test_compile(
 def test_compile_args(
     plugin: NVCCPlugin,
     compiler_cpp_17_fpath: str,
+    default_args: Namespace,
 ):
     gname = "test_compile_args"
     copy_source_to_group(compiler_cpp_17_fpath, gname, plugin.workdir)
@@ -112,21 +110,16 @@ def test_compile_args(
     with pytest.raises(subprocess.CalledProcessError):
         exec_fpath = plugin._compile(gname, compiler_args="--std c++14")
 
-    output = plugin._compile_and_run(
+    args = deepcopy(default_args)
-        group_name=gname,
+    args.compiler_args = lambda: "--std c++14"
-        args=argparse.Namespace(
+    output = plugin._compile_and_run(group_name=gname, args=args)
-            timeit=False,
-            profile=True,
-            profiler_args=lambda: "",
-            compiler_args=lambda: "--std c++14",
-        ),
-    )
     assert "errors detected in the compilation of" in output
 
 
 def test_compile_opencv(
     plugin: NVCCPlugin,
     compiler_opencv_fpath: str,
+    default_args: Namespace,
 ):
     gname = "test_compile_opencv"
     copy_source_to_group(compiler_opencv_fpath, gname, plugin.workdir)
@@ -134,23 +127,14 @@ def test_compile_opencv(
     # check that "pkg-config" exists
     assert subprocess.check_call(["which", "pkg-config"]) == 0
 
+    pkg_config_args = ["pkg-config", "--cflags", "--libs", "opencv4"]
     opencv_compile_options = (
-        subprocess.check_output(
+        subprocess.check_output(args=pkg_config_args).decode().strip()
-            args=["pkg-config", "--cflags", "--libs", "opencv4"]
-        )
-        .decode()
-        .strip()
     )
 
-    output = plugin._compile_and_run(
+    args = deepcopy(default_args)
-        group_name=gname,
+    args.compiler_args = lambda: opencv_compile_options
-        args=argparse.Namespace(
+    output = plugin._compile_and_run(group_name=gname, args=args)
-            timeit=False,
-            profile=True,
-            profiler_args=lambda: "",
-            compiler_args=lambda: opencv_compile_options,
-        ),
-    )
     assert "General configuration for OpenCV" in output
 
 
@@ -199,7 +183,9 @@ def test_run_profile(plugin: NVCCPlugin, sample_cuda_fpath: str):
 
 
 def test_compile_and_run_multiple_files(
-    plugin: NVCCPlugin, multiple_source_fpaths: List[str]
+    plugin: NVCCPlugin,
+    multiple_source_fpaths: List[str],
+    default_args: Namespace,
 ):
     """
     Compiles and executes 3 cuda source files from
@@ -208,20 +194,14 @@ def test_compile_and_run_multiple_files(
     gname = "test_compile_and_run_multiple_files"
     for fpath in multiple_source_fpaths:
         copy_source_to_group(fpath, gname, plugin.workdir)
-    output = plugin._compile_and_run(
+    output = plugin._compile_and_run(group_name=gname, args=default_args)
-        group_name=gname,
-        args=argparse.Namespace(
-            timeit=False,
-            profile=True,
-            profiler_args=lambda: "",
-            compiler_args=lambda: "",
-        ),
-    )
     check_profiler_output(output)
 
 
 def test_compile_and_run_multiple_files_shared(
-    plugin: NVCCPlugin, multiple_source_fpaths: List[str]
+    plugin: NVCCPlugin,
+    multiple_source_fpaths: List[str],
+    default_args: Namespace,
 ):
     """
     Compiles and executes 3 cuda source files from
@@ -236,20 +216,12 @@ def test_compile_and_run_multiple_files_shared(
             copy_source_to_group(fpath, gname, plugin.workdir)
         else:
             copy_source_to_group(fpath, "shared", plugin.workdir)
-    output = plugin._compile_and_run(
+    output = plugin._compile_and_run(group_name=gname, args=default_args)
-        group_name=gname,
-        args=argparse.Namespace(
-            timeit=False,
-            profile=True,
-            profiler_args=lambda: "",
-            compiler_args=lambda: "",
-        ),
-    )
     check_profiler_output(output)
 
 
 def test_read_args(plugin: NVCCPlugin):
-    parser = argparse.ArgumentParser()
+    parser = ArgumentParser()
     parser.add_argument("-a", type=str, required=True)
     parser.add_argument("-b", type=float, required=True)
     args = plugin._read_args(
@@ -292,6 +264,18 @@ def test_magic_cuda(
     check_profiler_output(capsys.readouterr().out)
 
 
+def test_magic_cuda_set_default_profiler(
+    capsys,
+    plugin: NVCCPlugin,
+    sample_cuda_code: str,
+    sample_magic_cu_line: str,
+):
+    # set the default profiler to Nsight Systems
+    set_defaults(profiler=Profiler.NSYS)
+    plugin.cuda(sample_magic_cu_line, sample_cuda_code)
+    check_profiler_output(capsys.readouterr().out, profiler="[NSYS]")
+
+
 def test_magic_cuda_bad_args(
     capsys,
     plugin: NVCCPlugin,