==================
Usage and Examples
==================


This library is broken up into three main parts, as well as a certain
compilation and linking framework:

#. :ref:`Core Examples`
#. :ref:`Array Examples`
#. :ref:`BLAS Examples`
#. :ref:`Compilation and Linking`
#. :ref:`Notes`

The ``Core.h`` header contains the necessary macros, flags and objects for interfacing with
basic kernel launching and the CUDA Runtime API. The ``Array.h`` header contains the ``CudaTools::Array``
class which provides a device compatible Array-like class with easy memory management. Lastly,
the ``BLAS.h`` header provides functions BLAS functions through the the cuBLAS library for the GPU,
and Eigen for the CPU. Lastly, a templated Makefile is provided which can be used
for your own project, after following a few rules.

The usage of this libary will be illustrated through examples, and further details
can be found in the other sections. The examples are given in the `samples <https://git.acem.ece.illinois.edu/kjao/CudaTools/src/branch/main/samples>`__ folder.
Throughout this documentation, there are a few common terms that may appear. First,we refer to the CPU as the host, and the GPU as the device. So, a host function refers
to a function runnable on the CPU, and a device function refers to a function that is runnable
on a device. A kernel is a specific function that the host can call to be run on the device.

Core Examples
=============
This file mainly introduces compiler macros and a few classes that are used to improve the
syntax between host and device code. To define and call a kernel, there are a few
macros provided. For example,

.. code-block:: cpp

    KERNEL(add, int x, int y) {
        printf("Kernel: %i\n", x + y);
    }

    int main() {
        CudaTools::Kernel::launch(add, CudaTools::Kernel::basic(1), 1, 1); // Prints 2.
        return 0;
    }

The ``DEFINE_KERNEL(name, ...)`` macro takes in the function name and its arguments.
The second argument in the ``KERNEL()`` macro is are the launch parameters for
kernel. The launch parameters have several items, but for 'embarassingly parallel'
cases, we can simply generate the settings with the number of threads. More detail with
creating launch parameters can be found :ref:`here <CudaTools::Kernel::Settings>`. In the above example,
there is only one thread. The rest of the arguments are just the kernel arguments. For more detail,
see :ref:`here <Macro Functions>`.

.. warning::
   These kernel definitions must be in a file that will be compiled by ``nvcc``. Also,
   for header files, there is an additional macro ``DECLARE_KERNEL(name, ...)`` to declare it
   and make it available to other files.

Since many applications used classes, a macro is provided to 'convert' a class into
being device-compatible. We follow the previous example in a similar fashion.

.. code-block:: cpp

    class intPair {
        DEVICE_CLASS(intPair)
        public:
            int x, y;

            intPair(const int x_, const int y_) : x(x_), y(y_) {
                allocateDevice(); // Allocates memory for this intPair on the device.
                updateDevice().wait(); // Copies the memory on the host to the device and waits until finished.
            };

            ~intPair() { CudaTools::free(that()); };

            HD void swap() {
                int swap = x;
                x = y;
                y = swap;
            };
    };

    KERNEL(swap, intPair* const pair) { pair->swap(); }

    int main() {
        intPair pair(1, 2);
        printf("Before: %u, %u\n", pair.x, pair.y); // Prints 1, 2.

        CudaTools::Kernel::launch(swap, CudaTools::Kernel::basic(1), pair.that()).wait();
        pair.updateHost().wait(); // Copies the memory from the device back to the host and waits until finished.

        printf("After: %u, %u\n", pair.x, pair.y); // Prints 2, 1.
        return 0;
    }

In this example, we create a class called ``intPair``, which is then made available on the device through
the ``DEVICE_CLASS(name)`` macro. Specifically, that macro introduces a few functions, like
``allocateDevice()``, ``updateDevice()``, ``updateHost()``, and ``that()``. The ``that()`` function
returns a pointer to the copy on the device. As a result, the programmer **must** define a destructor
that frees the pointer using ``CudaTools::free(that)``. For more details, see :ref:`here <Device Class>`.

.. warning::
   The ``updateDevice()`` and ``updateHost()`` in most cases will need to be explicitly called
   to push the data on the host to the device, and vice-versa. It is the programmers job to maintain
   where the 'most recent' copy is. If these are not called, various memory errors can occur. Note that,
   when passing a pointer to the kernel, it must be the *device* pointer. Otherwise, an illegal memory
   access would occur.

The kernel argument list should **must** consist of pointers to objects, or a non-reference object.
Otherwise, compilation will fail. In general this is safer, as it forces the programmer to
acknowledge that the device copy is being passed. For the latter case of a non-reference object,
you should only do this if there is no issue in creating a copy of the original object. In the above
example, we could have done this, but for more complicated classes it may result in unwanted behavior.

Lastly, since the point of classes is usually to have some member functions, to have them
available on the device, you must mark them with the compiler macro ``HD`` in front.

We also introduce the ``wait()`` function, which waits for the command to complete before
continuing. Most calls that involve the device are asynchronous, so without proper blocking,
operations dependent on a previous command are not guaranteed to run correctly. If the code is
compiled for CPU, then everything will run synchronously, as per usual.

.. note::
   Almost all functions that are asynchronous provide an optional 'stream' argument,
   where you can give the name of the stream you wish to use. Different streams run
   asynchronous, but operations on the same stream are FIFO. To define a stream to use
   later, you must call ``CudaTools::Manager::get()->addStream("myStream")`` at some point
   before you use it. For more details, see :ref:`here <CudaTools::Manager>`.


Array Examples
==============
The ``Array.h`` file introduces the ``Array`` class, which is a class that provides automatic
memory management between device and host. In particular, it provides functionality on
both the host and device while handling proper memory destruction, with many nice
features. In particular it supports mimics many features of the Python package NumPy.`
We can demonstrate a few here.

.. code-block:: cpp

    KERNEL(times2, const CudaTools::Array<int> arr) {
        CudaTools::Array<int> flat = arr.flattened();
        BASIC_LOOP(arr.shape().items()) { flat[iThread] *= 2; }
    }

    KERNEL(times2double, const CudaTools::Array<double> arr) {
        CudaTools::Array<double> flat = arr.flattened();
        BASIC_LOOP(arr.shape().items()) { flat[iThread] *= 2; }
    }

    int main() {
        CudaTools::Array<int> arrRange = CudaTools::Array<int>::range(0, 10);
        CudaTools::Array<int> arrConst = CudaTools::Array<int>::constant({10}, 1);
        CudaTools::Array<double> arrLinspace = CudaTools::Array<double>::linspace(0, 5, 10);
        CudaTools::Array<int> arrComma({2, 2}); // 2x2 array.
        arrComma << 1, 2, 3, 4;                 // Comma initializer if needed.

        arrRange.updateDevice();
        arrConst.updateDevice();
        arrLinspace.updateDevice();
        arrComma.updateDevice().wait();

        std::cout << "Before Kernel:\n";
        std::cout << arrRange << "\n" << arrConst << "\n" << arrLinspace << "\n" << arrComma << "\n";

        // Call the kernel multiple times asynchronously. Note: since they share same
        // stream, they are not run in parallel, just queued on the device.
        // NOTE: Notice that a view is passed into the kernel, not the Array itself.
        CudaTools::Kernel::launch(times2, CudaTools::Kernel::basic(arrRange.shape().items()), arrRange.view());
        CudaTools::Kernel::launch(times2, CudaTools::Kernel::basic(arrConst.shape().items()), arrConst.view());
        CudaTools::Kernel::launch(times2double, CudaTools::Kernel::basic(arrLinspace.shape().items()), arrLinspace.view());
        CudaTools::Kernel::launch(times2, CudaTools::Kernel::basic(arrComma.shape().items()), arrComma.view()).wait();
        arrRange.updateHost();
        arrConst.updateHost();
        arrLinspace.updateHost();
        arrComma.updateHost().wait(); // Same stream, so you should wait for the last call.

        std::cout << "After Kernel:\n";
        std::cout << arrRange << "\n" << arrConst << "\n" << arrLinspace << "\n" << arrComma << "\n";
        return 0;
    }

In this example, we show a few ways to initialize an ``Array`` through some static functions.
It is templated, so it can (theoretically) support any type. Additionally, you can initialize an
empty ``Array`` by providing its ``Shape`` with an initializer list (ex: ``{2, 2}``). Many of these
array functions and initializers have view-returning and self-assigning versions. For instance,
``.flattened()`` returns a flattened view of an Array, and does not modify the original. For more details,
see :ref:`here <CudaTools::Array<T>>`.

We also note the use of ``BASIC_LOOP(N)``, which is a macro for generating the loop automatically
on the kernel given the number of threads. It is intended to be used only for "embarassingly parallel"
situations and with the ``CudaTools::Kernel::basic()`` launch parameters. If compiling for CPU, it will
mark the loop with ``#pragma parallel for`` and attempt to use OpenMP for parallelism.

.. warning::
   Notice that a view must be passed to the kernel, and not the original object. This

The Array also supports other helpful functions, such as multi-dimensional indexing, slicing, and
a few other functions.

.. code-block:: cpp

    int main() {
        CudaTools::Array<int> arr = CudaTools::Array<int>::constant({100}, 0);
        arr.reshape({4, 5, 5}); // Creates a three dimensional array.

        arr[0][0][0] = 1;     // Axis by axis indexing.
        arr[{1, 0, 0}] = 100; // Specific 'coordinate' indexing.
        std::cout << arr << "\n";

        CudaTools::Array<int> arrRange = CudaTools::Array<int>::range(0, 18);
        auto arrSlice = arr.slice({{1, 3}, {1, 4}, {1, 4}}); // Takes a slice of the center.
        std::cout << "Before Copy:\n" << arrSlice << "\n";
        arrSlice = arrRange; // Copies arrRange into arrSlice. (Does NOT replace!)
        std::cout << "After Copy:\n" << arrSlice << "\n";

        std::cout << "Modified: \n"
                  << arr << "\n"; // The original array is modified, since a slice does not copy.

        CudaTools::Array<int> newArr = arr.copy();                 // Copies the original Array.
        for (auto it = newArr.begin(); it != newArr.end(); ++it) { // Iterate through the array.
            *it = 1;
        }
        std::cout << "Modified New Array:\n" << newArr << "\n";
        std::cout << "Old Array:\n" << arr << "\n"; // The original array was not modified after a copy.
        return 0;
    }

In this example, we demonstrate some of the functionality of the Array. We can do
multi-dimensional indexing, take slices of the Array, and iterate through the Array through an
iterator, in C++ fashion. Particularly, we need to introduce the concept of a "view" of an Array.
An Array either "owns" its data or is a "view" of another Array. You can create a
view manually with the ``.view()`` function.

.. warning::
   When using the assignment operator, if a view is on the left-hand side, it will
   perform a copy of the internal data. However, if the Array is an owner, then it will replace
   the entire Array, and **free the old memory**. This means any view of that previous
   array will now point to invalid places in memory. It is responsibility of the
   programmer to manage this.


Graph Examples
==============
Additionally, there is support for CUDA Graphs, a way of defining a series of kernel
launches and executing later, potentially reducing overhead and timing, as well as
control the specific parallel workflow between CPU and GPU. The following
snippet illustrates this

.. code-block:: cpp

    void myGraph(CudaTools::GraphManager* gm, const CudaTools::Array<uint32_t> A,
                 const CudaTools::Array<uint32_t> B) {
        A.updateDevice("graphStream");
        gm->makeBranch("graphStream", "graphStreamBranch");
        B.updateDevice("graphStreamBranch");
        for (uint32_t iTimes = 0; iTimes < 30; ++iTimes) {
            CudaTools::Kernel::launch(
                collatz, CudaTools::Kernel::basic(A.shape().items(), "graphStream"), A.view());
            CudaTools::Kernel::launch(
                plusOne, CudaTools::Kernel::basic(A.shape().items(), "graphStreamBranch"), B.view());
        }

        gm->joinBranch("graphStream", "graphStreamBranch");
        CudaTools::Kernel::launch(addArray, CudaTools::Kernel::basic(A.shape().items(), "graphStream"),
                                  A.view(), B.view());
        A.updateHost("graphStream");
        B.updateHost("graphStream");
        gm->launchHostFunction("graphStream", addNum, A.view(), 5);
    }

    int main() {
        CudaTools::Manager::get()->addStream("graphStream");
        CudaTools::Manager::get()->addStream("graphStreamBranch");

        CudaTools::Array<uint32_t> A = CudaTools::Array<uint32_t>::constant({100}, 50);
        CudaTools::Array<uint32_t> B = CudaTools::Array<uint32_t>::constant({100}, 0);

        CudaTools::GraphManager gm;
        CudaTools::Graph graph("graphStream", myGraph, &gm, A.view(), B.view());
        TIME(graph.execute().wait(), ExecuteGraph);

        std::cout << A.slice({{0, 10}}) << "\n";
        return 0;
    }

We first create two new streams to be used in the graph, which define the different parallel
streams used. To use CUDA Graphs in CudaTools, we expect the graph to be created from a function, which
should be written as if it will be executed. Note that we do not need to use ``.wait()`` here, since the function
is intended to be captured into the graph. The capture process is done on the creation of the graph, with
the name of the origin stream, the function name, and the arguments of the function. Afterwards,
simply run ``graph.execute()`` to execute the captured graph. On CPU, it will simply run the function.

To access the other functionality like graph branching an capturing host functions, it is
necessary to use the ``CudaTools::GraphManager`` class, which stores a variety of necessary variables
that need to be kept during the lifetime of the graph execution. **Currently, launching host functions sometimes alters the correct blocking of the stream, in particular with copying. It is not yet known if this is an issue with the library or a technicality within CUDA Graphs itself that needs some special care to resolve.** To read more about the syntax, see :ref:`here <CudaTools::GraphManager>`.

.. warning::

   A graph capture essentially 'freezes' the variables used in the capture, like
   function arguments. As a result, the programmer must take care that the variables
   are well-defined. This is especially relevant to variables on the heap, where you need
   to make sure the variable is not a copy. Potentially, always using pointers could work,
   but is not always necessary. Likely ``.view()`` should always be used when dealing with
   ``CudaTools::Array`` objects.


BLAS Examples
=============


Compilation and Linking
=======================
To compile with this library, there are only a few things necessary. First, this library depends on
`Eigen 3.4.0+ <https://eigen.tuxfamily.org/index.php?title=Main_Page>`__, and must be
compiled with C++17. Next, it is recommended you use the provided template ``Makefile``, which can be
easily modified to suit your project needs. It already default handles the compilation
and linking with ``nvcc``, so long as you fulfill a few requirements.

#. Use the compiler flag ``CUDA`` to mark where GPU-specific code is, if necessary.
#. Any files that use ``CUDA`` functionality, (i.e defining a kernel), should have the file
   extension ``.cu.cpp``.
#. When including the ``Core.h`` header file, in only **one** file, you must define the
   macro ``CUDATOOLS_IMPLEMENTATION``. That file must also compile for CUDA, so its
   extension must be ``.cu.cpp``. It's recommended to put this with your kernel definitions.

Afterwards the ``Makefile`` will have two targets ``cpu`` and ``gpu``, which compile
the CPU and GPU compatible binaries respectively. As an example, we can look at the whole
file for the first example:

.. code-block:: cpp

    // main.cu.cpp
    #define CUDATOOLS_IMPLEMENTATION
    #include <Core.h>

    DEFINE_KERNEL(add, int x, int y) {
        printf("Kernel: %i\n", x + y);
    }

    int main() {
        KERNEL(add, CudaTools::Kernel::basic(1), 1, 1); // Prints 2.
        return 0;
    }

.. code-block::

    // Makefile

    CC := g++-10
    NVCC := nvcc
    CFLAGS := -Wall -std=c++17 -fopenmp -MMD
    NVCC_FLAGS := -MMD -std=c++17 -w -Xcompiler

    INCLUDE := ../../
    LIBS_DIR :=
    LIBS_DIR_GPU := /usr/local/cuda/lib64
    LIBS :=
    LIBS_GPU := cuda cudart cublas

    TARGET = coreKernel
    SRC_DIR = .
    BUILD_DIR = build

The lines above are the first few lines of the ``Makefile``, which are the only
lines you should need to modify, consisting of libraries and flags, as well as
the name of the target.

Notes
=====

Complex Numbers
---------------
Dealing with complex numbers is slightly complicated, trying to enforce compatability between
two systems and several different libraries which many not have the right support. We
create a simple barebones host and device compatible complex number class following
the same as ``cuComplex.h``, but with proper C++ operator overloading and class structure. However,
while the underlying data structure is identical to all other complex number structures, there
is a lot of type-casting done underneath the hood to get cuBLAS and Eigen to work well
together, while maintaining one 'unified' complex type.

As a result, there could be some issues and lack of functionality with this at the moment.
For now, it's recommended to use the given ``complex64`` and ``complex128`` types which
should properly adapt and work.