Merge branch 'master' into feature/profiler-tool-choice

* 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

.devcontainer/Dockerfile

@@ -1,15 +1,29 @@
-FROM ubuntu
+FROM ubuntu:22.04
 
 ARG VENV_PATH=/opt/dev-venv
 ENV VENV_ACTIVATE=${VENV_PATH}/bin/activate
+ENV DEBIAN_FRONTEND="noninteractive"
 
+# install the latest CUDA toolkit (https://developer.nvidia.com/cuda-downloads)
 RUN apt update
-RUN apt install -y python3.10-venv nvidia-cuda-toolkit gcc vim git
+RUN apt install -y wget
+RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+RUN dpkg -i cuda-keyring_1.1-1_all.deb
+RUN apt update
+RUN apt -y install cuda-toolkit-12-3
+RUN echo "PATH=\"\$PATH:/usr/local/cuda/bin\"" >> ~/.bashrc
 
-# the mkdir command bypasses a profiler error, which allows us to run it with
+# install OpenCV to test compilation with external libraries
-# host code only to at least check that the profiler parameters are correctly
+RUN apt install -y libopencv-dev pkg-config
-# provided; without this line, some tests will fail
+
-RUN mkdir -p /usr/lib/x86_64-linux-gnu/nsight-compute/sections
+# make & language-pack-en are for documentation
+RUN apt install -y \
+    gcc \
+    git \
+    language-pack-en \
+    make \
+    python3.10-venv \
+    vim
 
 # we create the virtualenv here so that the devcontainer.json setting
 # python.defaultInterpreterPath can be used to find it; if we do it in the

.devcontainer/devcontainer.json

@@ -16,10 +16,12 @@
                 "ms-python.isort",
                 "ms-python.flake8",
                 "ms-python.black-formatter",
-                "ryanluker.vscode-coverage-gutters"
+                "ryanluker.vscode-coverage-gutters",
+                "njpwerner.autodocstring"
             ],
             "settings": {
-                "python.defaultInterpreterPath": "/opt/dev-venv/bin/python"
+                "python.defaultInterpreterPath": "/opt/dev-venv/bin/python",
+                "autoDocstring.docstringFormat": "google-notypes"
             }
         }
     }

.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

.github/workflows/test.yml

@@ -27,14 +27,19 @@ jobs:
         with:
           python-version: ${{ matrix.python-version }}
 
-      # the mkdir command bypasses a profiler error, which allows us to run it
+      - name: Install CUDA toolkit
-      # with host code only to at least check that the profiler parameters are
-      # correctly provided
-      - name: Install CUDA tools
         run: |
           sudo apt update
-          sudo apt install nvidia-cuda-toolkit
+          sudo apt install -y wget
-          sudo mkdir -p /usr/lib/x86_64-linux-gnu/nsight-compute/sections
+          wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+          sudo dpkg -i cuda-keyring_1.1-1_all.deb
+          sudo apt update
+          sudo apt -y install cuda-toolkit-12-3
+          echo "PATH=$PATH:/usr/local/cuda/bin" >> $GITHUB_ENV
+
+      - name: Install OpenCV
+        run: |
+          sudo apt install -y libopencv-dev pkg-config
 
       - name: Install Python dependencies
         run: |
@@ -65,11 +70,19 @@ jobs:
         with:
           python-version: "3.10"
 
-      - name: Install CUDA tools
+      - name: Install CUDA toolkit
         run: |
           sudo apt update
-          sudo apt install nvidia-cuda-toolkit
+          sudo apt install -y wget
-          sudo mkdir -p /usr/lib/x86_64-linux-gnu/nsight-compute/sections
+          wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+          sudo dpkg -i cuda-keyring_1.1-1_all.deb
+          sudo apt update
+          sudo apt -y install cuda-toolkit-12-3
+          echo "PATH=$PATH:/usr/local/cuda/bin" >> $GITHUB_ENV
+
+      - name: Install OpenCV
+        run: |
+          sudo apt install -y libopencv-dev pkg-config
 
       - name: Install Python dependencies
         run: |

README.md

@@ -45,7 +45,8 @@ 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)
 
 ## Install
@@ -88,13 +89,14 @@ The official documentation is hosted on [readthedocs](https://nvcc4jupyter.readt
 
 ## Contributing
 
-Install the package with the development dependencies:
+The recommended setup for development is using the devcontainer in GitHub
-```bash
+Codespaces or locally in VSCode.
-pip install .[dev]
-```
 
-As a developer, make sure you install the pre-commit hook before commiting any changes:
+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 -e .[dev]
 pre-commit install
 ```
 

docs/requirements.txt

@@ -1,2 +1,3 @@
 sphinx==7.1.2
 sphinx-rtd-theme==1.3.0rc1
+IPython>=8.19.0

docs/source/conf.py

@@ -6,11 +6,18 @@
 # -- Project information -----------------------------------------------------
 # https://www.sphinx-doc.org/en/master/usage/configuration.html#project-information
 
+import os
+import sys
+
+sys.path.append(os.path.join("..", ".."))
+from nvcc4jupyter.__init__ import __version__  # noqa: E402
+
 project = "nvcc4jupyter"
 copyright = "2024, Andrei Nechaev & Cosmin Stefan Ciocan"
 author = "Andrei Nechaev & Cosmin Stefan Ciocan"
-release = "1.0.1"
+release = __version__
-version = "1.0.1"
+version = __version__
+
 
 # -- General configuration ---------------------------------------------------
 # https://www.sphinx-doc.org/en/master/usage/configuration.html#general-configuration

docs/source/index.rst

@@ -10,4 +10,5 @@ which provides CUDA capable GPUs with the CUDA toolkit already installed.
    :caption: Contents:
 
    usage
+   notebooks
    magics

docs/source/magics.rst

@@ -21,23 +21,59 @@ 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.
+   - ``%%cuda -c "<SPACE SEPARATED COMPILER ARGS"``: Passes additional arguments to "nvcc".
    - ``%%cuda -t``: Outputs the "timeit" built-in magic results.
 
 Options
 -------
 
+.. _timeit:
+
 -t, --timeit
    Boolean. If set, returns the output of the "timeit" built-in
    ipython magic instead of stdout.
 
+.. _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:
+
+-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>`_
+
 
 .. note::
    If both "\-\-profile" and "\-\-timeit" are used then no profiling is
@@ -47,10 +83,11 @@ Examples
 --------
 ::
 
-   # compile, run, and profile the code in the cell with the Nsight
+   # compile, run, and profile the code in the cell with the Nsight compute
-   # compute profiler while collecting only metrics from the
+   # profiler while collecting only metrics from the "MemoryWorkloadAnalysis"
-   # "MemoryWorkloadAnalysis" section.
+   # section; also provides the "--optimize 3" option to "nvcc" during
-   %%cuda --profile --profiler-args "--section MemoryWorkloadAnalysis"
+   # compilation to optimize host code
+   %%cuda -p -a "--section MemoryWorkloadAnalysis" -c "--optimize 3"
 
 ------
 

docs/source/notebooks.rst

@@ -0,0 +1,34 @@
+*********
+Notebooks
+*********
+
+This page provides a list of useful Jupyter notebooks written with the
+**nvcc4jupyter** library.
+
+.. note::
+   These notebooks are written for Google's Colab, but you may run them in
+   other environments by installing all expected dependencies. If running in
+   Colab, make sure to set the runtime type to a GPU instance (at the time of
+   writing this, T4 is the GPU offered for free by Colab).
+
+------
+
+.. _compiling_with_external_libraries:
+
+Compiling with external libraries
+=================================
+
+[`NOTEBOOK <https://colab.research.google.com/drive/1iuY46DCwv4hy3SqDhJgFeO8kgpHnzjTh?usp=sharing>`_]
+
+If you need to compile CUDA C++ code that uses external libraries in the host
+code (e.g. OpenCV for reading and writing images to disk) then this section is
+for you.
+
+To achieve this, use the :ref:`compiler-args <compiler_args>` option of the
+:ref:`cuda <cuda_magic>` magic command to pass the correct compiler options
+of the OpenCV library to **nvcc** for it to link the OpenCV code with the
+code in your Jupyter cell. Those compiler options can be provided by the
+`pkg-config <https://www.freedesktop.org/wiki/Software/pkg-config/>`_ tool.
+
+In the notebook we show how to use OpenCV to load an image, blur it with a CUDA
+kernel, and then save it back to disk using OpenCV again.

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++
 
@@ -255,3 +256,68 @@ Running the cell above will compile and execute the vector addition code in the
     SM Active Cycles                cycle       383.65
     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
+------------------
+
+In the same way profiler arguments can be passed to the profiling tool,
+compiling arguments can be passed to **nvcc**:
+
+.. code-block:: c++
+
+    %cuda_group_run --group "vector_add" --compiler-args "--optimize 3"
+
+Running the cell above will compile and execute the vector addition code in the
+"vector_add" group. During compilation, **nvcc** receives the "\-\-optimize"
+option which specifies the optimization level for host code.
+
+Set default arguments
+---------------------
+
+In the case where you execute multiple magic commands with the same compiler or
+profiler arguments you can avoid writing them every time by setting the default
+arguments:
+
+.. code-block:: python
+
+    from nvcc4jupyter import set_defaults
+    set_defaults(compiler_args="--optimize 3", profiler_args="--section SpeedOfLight")
+
+The same effect can be achieved by running "set_defaults" once for each config
+due to the fact that the default value is not changed if an a value is not
+given to the "set_defaults" function.
+
+.. code-block:: python
+
+    from nvcc4jupyter import set_defaults
+    set_defaults(compiler_args="--optimize 3")
+    set_defaults(profiler_args="--section SpeedOfLight")
+
+
+Now we can run the following cell without specifying the compiler and profiler
+arguments once again.
+
+.. code-block:: c++
+
+    %cuda_group_run --group "vector_add" --profile

nvcc4jupyter/__init__.py

@@ -2,6 +2,7 @@
 nvcc4jupyter: CUDA C++ plugin for Jupyter Notebook
 """
 
+from .parsers import Profiler, set_defaults  # noqa: F401
 from .plugin import NVCCPlugin, load_ipython_extension  # noqa: F401
 
-__version__ = "1.0.3"
+__version__ = "1.1.0"

nvcc4jupyter/parsers.py

@@ -3,6 +3,63 @@ Parsers for the CUDA magic commands.
 """
 
 import argparse
+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(
+    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
+    values will be used if the user does not explicitly provide those arguments
+    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 None.
+        profiler_args: If not None, this value becomes the new default profiler
+            config. Defaults to None.
+    """
+
+    # pylint: disable=global-statement
+    global _default_profiler
+    if profiler is not None:
+        _default_profiler = profiler
+    global _default_compiler_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
+
+
+def str_to_lambda(arg: str) -> Callable[[], str]:
+    """Convert argparse string to lambda"""
+    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:
@@ -18,7 +75,28 @@ def get_parser_cuda() -> argparse.ArgumentParser:
     )
     parser.add_argument("-t", "--timeit", action="store_true")
     parser.add_argument("-p", "--profile", action="store_true")
-    parser.add_argument("-a", "--profiler-args", type=str, default="")
+
+    # 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",
+        type=str_to_lambda,
+        default=lambda: _default_profiler_args,
+    )
+    parser.add_argument(
+        "-c",
+        "--compiler-args",
+        type=str_to_lambda,
+        default=lambda: _default_compiler_args,
+    )
+
     return parser
 
 

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:
@@ -87,7 +99,10 @@ class NVCCPlugin(Magics):
             shutil.rmtree(group_dirpath)
 
     def _compile(
-        self, group_name: str, executable_fname: str = DEFAULT_EXEC_FNAME
+        self,
+        group_name: str,
+        executable_fname: str = DEFAULT_EXEC_FNAME,
+        compiler_args: str = "",
     ) -> str:
         """
         Compiles all source files in a given group together with all source
@@ -97,6 +112,7 @@ class NVCCPlugin(Magics):
             group_name: The name of the source file group to be compiled.
             executable_fname: The output executable file name. Defaults to
                 "cuda_exec.out".
+            compiler_args: The optional "nvcc" compiler arguments.
 
         Raises:
             RuntimeError: If the group does not exist or if does not have any
@@ -121,27 +137,52 @@ class NVCCPlugin(Magics):
 
         executable_fpath = os.path.join(group_dirpath, executable_fname)
 
-        args = [
+        args = ["nvcc"]
-            "nvcc",
+        args.extend(compiler_args.split())
-            "-I" + shared_dirpath + "," + group_dirpath,
+        args.append("-I" + shared_dirpath + "," + group_dirpath)
-        ]
         args.extend(source_files)
-        args.extend(
+        args.extend(["-o", executable_fpath, "-Wno-deprecated-gpu-targets"])
-            [
+
-                "-o",
-                executable_fpath,
-                "-Wno-deprecated-gpu-targets",
-            ]
-        )
         subprocess.check_output(args, stderr=subprocess.STDOUT)
 
         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:
         """
@@ -152,8 +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.
@@ -175,7 +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
@@ -188,12 +231,16 @@ class NVCCPlugin(Magics):
         self, group_name: str, args: argparse.Namespace
     ) -> str:
         try:
-            exec_fpath = self._compile(group_name)
+            exec_fpath = self._compile(
+                group_name=group_name,
+                compiler_args=args.compiler_args(),
+            )
             output = self._run(
                 exec_fpath=exec_fpath,
                 timeit=args.timeit,
                 profile=args.profile,
-                profiler_args=args.profiler_args,
+                profiler=args.profiler(),
+                profiler_args=args.profiler_args(),
             )
         except subprocess.CalledProcessError as e:
             output = e.output.decode("utf8")

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
@@ -286,6 +288,6 @@ deprecated-modules="optparse,tkinter.tix"
 
 [tool.pylint.'EXCEPTIONS']
 overgeneral-exceptions= [
-    "BaseException",
+    "builtins.BaseException",
-    "Exception"
+    "builtins.Exception"
 ]

tests/fixtures/compiler/cpp_17.cu

@@ -0,0 +1,47 @@
+#include <cstdlib>
+#include <iostream>
+#include <set>
+#include <string>
+#include <iterator>
+
+#include <tuple>
+
+struct S {
+    int n;
+    std::string s;
+    float d;
+    bool operator<(const S& rhs) const
+    {
+        // compares n to rhs.n,
+        // then s to rhs.s,
+        // then d to rhs.d
+        return std::tie(n, s, d) < std::tie(rhs.n, rhs.s, rhs.d);
+    }
+};
+
+int main()
+{
+    std::set<S> mySet;
+
+    // pre C++17:
+    {
+	    S value{42, "Test", 3.14};
+	    std::set<S>::iterator iter;
+	    bool inserted;
+
+	    // unpacks the return val of insert into iter and inserted
+	    std::tie(iter, inserted) = mySet.insert(value);
+
+	    if (inserted)
+		    std::cout << "Value was inserted\n";
+    }
+
+	// with C++17:
+    {
+        S value{100, "abc", 100.0};
+        const auto [iter, inserted] = mySet.insert(value);
+
+        if (inserted)
+		    std::cout << "Value(" << iter->n << ", " << iter->s << ", ...) was inserted" << "\n";
+    }
+}

tests/fixtures/compiler/opencv.cu

@@ -0,0 +1,8 @@
+#include <opencv2/core.hpp>
+#include <iostream>
+
+int main(int argc, char** argv)
+{
+    std::cout << cv::getBuildInformation() << std::endl;
+    return 0;
+}

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
 
 
@@ -27,10 +29,25 @@ def fixtures_path(tests_path):
     return os.path.join(tests_path, "fixtures")
 
 
+@pytest.fixture(scope="session")
+def scripts_path(fixtures_path: str):
+    return os.path.join(fixtures_path, "scripts")
+
+
+@pytest.fixture(scope="session")
+def compiler_cpp_17_fpath(fixtures_path: str):
+    return os.path.join(fixtures_path, "compiler", "cpp_17.cu")
+
+
+@pytest.fixture(scope="session")
+def compiler_opencv_fpath(fixtures_path: str):
+    return os.path.join(fixtures_path, "compiler", "opencv.cu")
+
+
 @pytest.fixture(scope="session")
 def sample_magic_cu_line():
     # fmt: off
-    return '--profile --profiler-args "--metrics l1tex__t_sectors_pipe_lsu_mem_global_op_ld.sum"'  # noqa: E501
+    return '--profile --profiler-args "--metrics l1tex__t_sectors_pipe_lsu_mem_global_op_ld.sum" --compiler-args "--optimize 3"'  # noqa: E501
     # fmt: on
 
 
@@ -55,3 +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

@@ -0,0 +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}"

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

@@ -1,27 +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 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(
@@ -36,11 +35,19 @@ def copy_source_to_group(
     return destination_fpath
 
 
+@pytest.fixture(autouse=True, scope="session")
+def before_all(scripts_path: str):
+    os.environ["PATH"] = scripts_path + os.pathsep + os.environ["PATH"]
+
+
 @pytest.fixture(autouse=True, scope="function")
 def before_each(plugin: NVCCPlugin):
-    shutil.rmtree(plugin.workdir, ignore_errors=True)  # before test
+    # BEFORE TESTS
+    set_defaults(profiler=Profiler.NCU, compiler_args="", profiler_args="")
+    shutil.rmtree(plugin.workdir, ignore_errors=True)
     yield
-    pass  # after test
+    # AFTER TESTS
+    pass
 
 
 def test_save_source(plugin: NVCCPlugin, sample_cuda_code: str) -> None:
@@ -88,6 +95,49 @@ def test_compile(
         plugin._compile(gname)
 
 
+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)
+
+    exec_fpath = plugin._compile(gname, compiler_args="--std c++17")
+    assert os.path.exists(exec_fpath)
+
+    # should fail due to the source file having c++ 17 features
+    with pytest.raises(subprocess.CalledProcessError):
+        exec_fpath = plugin._compile(gname, compiler_args="--std c++14")
+
+    args = deepcopy(default_args)
+    args.compiler_args = lambda: "--std c++14"
+    output = plugin._compile_and_run(group_name=gname, args=args)
+    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)
+
+    # 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(args=pkg_config_args).decode().strip()
+    )
+
+    args = deepcopy(default_args)
+    args.compiler_args = lambda: opencv_compile_options
+    output = plugin._compile_and_run(group_name=gname, args=args)
+    assert "General configuration for OpenCV" in output
+
+
 def test_run(
     plugin: NVCCPlugin,
     sample_cuda_fpath: str,
@@ -133,7 +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
@@ -142,14 +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)
-        gname, argparse.Namespace(timeit=False, profile=True, profiler_args="")
-    )
     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
@@ -164,14 +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)
-        gname, argparse.Namespace(timeit=False, profile=True, profiler_args="")
-    )
     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(
@@ -181,6 +231,29 @@ def test_read_args(plugin: NVCCPlugin):
     assert math.isclose(args.b, 0.75)
 
 
+def test_set_defaults():
+    parser = get_parser_cuda()
+    args = parser.parse_args([])
+    assert args.profiler_args() == ""
+    assert args.compiler_args() == ""
+    set_defaults(profiler_args="123")
+    args = parser.parse_args([])
+    assert args.profiler_args() == "123"
+    assert args.compiler_args() == ""
+    set_defaults(compiler_args="456")
+    args = parser.parse_args([])
+    assert args.profiler_args() == "123"
+    assert args.compiler_args() == "456"
+    set_defaults(profiler_args="")
+    args = parser.parse_args([])
+    assert args.profiler_args() == ""
+    assert args.compiler_args() == "456"
+    set_defaults(profiler_args="123")
+    args = parser.parse_args(["--profiler-args", "789"])
+    assert args.profiler_args() == "789"
+    assert args.compiler_args() == "456"
+
+
 def test_magic_cuda(
     capsys,
     plugin: NVCCPlugin,
@@ -191,6 +264,28 @@ 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,
+    sample_cuda_code: str,
+):
+    plugin.cuda("--this-is-an-unrecognized-argument", sample_cuda_code)
+    output = capsys.readouterr().out
+    assert output.startswith("usage: ")
+
+
 def test_magic_cuda_group_save(plugin: NVCCPlugin, sample_cuda_code: str):
     gname = "test_save_source"
     sname = "sample.cu"