starpu-max/doc/doxygen/chapters/basic_examples.doxy

/*
 * This file is part of the StarPU Handbook.
 * Copyright (C) 2009--2011  Universit@'e de Bordeaux 1
 * Copyright (C) 2010, 2011, 2012, 2013  Centre National de la Recherche Scientifique
 * Copyright (C) 2011, 2012 Institut National de Recherche en Informatique et Automatique
 * See the file version.doxy for copying conditions.
 */

/*! \page BasicExamples Basic Examples

\section HelloWorldUsingTheCExtension Hello World Using The C Extension

This section shows how to implement a simple program that submits a task
to StarPU using the StarPU C extension (\ref cExtensions). The complete example, and additional examples,
is available in the directory <c>gcc-plugin/examples</c> of the StarPU
distribution. A similar example showing how to directly use the StarPU's API is shown
in \ref HelloWorldUsingStarPUAPI.

GCC from version 4.5 permit to use the StarPU GCC plug-in (\ref cExtensions). This makes writing a task both simpler and less error-prone.
In a nutshell, all it takes is to declare a task, declare and define its
implementations (for CPU, OpenCL, and/or CUDA), and invoke the task like
a regular C function.  The example below defines <c>my_task</c> which
has a single implementation for CPU:

\snippet hello_pragma.c To be included

The code can then be compiled and linked with GCC and the flag <c>-fplugin</c>:

\verbatim
$ gcc `pkg-config starpu-1.2 --cflags` hello-starpu.c \
    -fplugin=`pkg-config starpu-1.2 --variable=gccplugin` \
    `pkg-config starpu-1.2 --libs`
\endverbatim

The code can also be compiled without the StarPU C extension and will
behave as a normal sequential code.

\verbatim
$ gcc hello-starpu.c
hello-starpu.c:33:1: warning: ‘task’ attribute directive ignored [-Wattributes]
$ ./a.out
Hello, world! With x = 42
\endverbatim

As can be seen above, the C extensions allows programmers to
use StarPU tasks by essentially annotating ``regular'' C code.

\section HelloWorldUsingStarPUAPI Hello World Using StarPU's API

This section shows how to achieve the same result as in the previous
section using StarPU's standard C API.

\subsection RequiredHeaders Required Headers

The header starpu.h should be included in any code using StarPU.

\code{.c}
#include <starpu.h>
\endcode

\subsection DefiningACodelet Defining A Codelet

\code{.c}
struct params
{
    int i;
    float f;
};
void cpu_func(void *buffers[], void *cl_arg)
{
    struct params *params = cl_arg;

    printf("Hello world (params = {%i, %f} )\n", params->i, params->f);
}

struct starpu_codelet cl =
{
    .where = STARPU_CPU,
    .cpu_funcs = { cpu_func, NULL },
    .cpu_funcs_name = { "cpu_func", NULL },
    .nbuffers = 0
};
\endcode

A codelet is a structure that represents a computational kernel. Such a codelet
may contain an implementation of the same kernel on different architectures
(e.g. CUDA, x86, ...). For compatibility, make sure that the whole
structure is properly initialized to zero, either by using the
function starpu_codelet_init(), or by letting the
compiler implicitly do it as examplified above.

The field starpu_codelet::nbuffers specifies the number of data buffers that are
manipulated by the codelet: here the codelet does not access or modify any data
that is controlled by our data management library. Note that the argument
passed to the codelet (the parameter <c>cl_arg</c> of the function
<c>cpu_func</c>) does not count as a buffer since it is not managed by
our data management library, but just contain trivial parameters.

\internal
TODO need a crossref to the proper description of "where" see bla for more ...
\endinternal

We create a codelet which may only be executed on the CPUs. The field
starpu_codelet::where is a bitmask that defines where the codelet may
be executed. Here, the value ::STARPU_CPU means that only CPUs can
execute this codelet. Note that field starpu_codelet::where is
optional, when unset its value is automatically set based on the
availability of the different fields <c>XXX_funcs</c>.
When a CPU core executes a codelet, it calls the function
<c>cpu_func</c>, which \em must have the following prototype:

\code{.c}
void (*cpu_func)(void *buffers[], void *cl_arg);
\endcode

In this example, we can ignore the first argument of this function which gives a
description of the input and output buffers (e.g. the size and the location of
the matrices) since there is none.
The second argument is a pointer to a buffer passed as an
argument to the codelet by the means of the field starpu_task::cl_arg.

\internal
TODO rewrite so that it is a little clearer ?
\endinternal

Be aware that this may be a pointer to a
\em copy of the actual buffer, and not the pointer given by the programmer:
if the codelet modifies this buffer, there is no guarantee that the initial
buffer will be modified as well: this for instance implies that the buffer
cannot be used as a synchronization medium. If synchronization is needed, data
has to be registered to StarPU, see \ref VectorScalingUsingStarPUAPI.

\subsection SubmittingATask Submitting A Task

\code{.c}
void callback_func(void *callback_arg)
{
    printf("Callback function (arg %x)\n", callback_arg);
}

int main(int argc, char **argv)
{
    /* initialize StarPU */
    starpu_init(NULL);

    struct starpu_task *task = starpu_task_create();

    task->cl = &cl; /* Pointer to the codelet defined above */

    struct params params = { 1, 2.0f };
    task->cl_arg = &params;
    task->cl_arg_size = sizeof(params);

    task->callback_func = callback_func;
    task->callback_arg = 0x42;

    /* starpu_task_submit will be a blocking call */
    task->synchronous = 1;

    /* submit the task to StarPU */
    starpu_task_submit(task);

    /* terminate StarPU */
    starpu_shutdown();

    return 0;
}
\endcode

Before submitting any tasks to StarPU, starpu_init() must be called. The
<c>NULL</c> argument specifies that we use the default configuration. Tasks cannot
be submitted after the termination of StarPU by a call to
starpu_shutdown().

In the example above, a task structure is allocated by a call to
starpu_task_create(). This function only allocates and fills the
corresponding structure with the default settings, but it does not
submit the task to StarPU.

\internal
not really clear ;)
\endinternal

The field starpu_task::cl is a pointer to the codelet which the task will
execute: in other words, the codelet structure describes which computational
kernel should be offloaded on the different architectures, and the task
structure is a wrapper containing a codelet and the piece of data on which the
codelet should operate.

The optional field starpu_task::cl_arg field is a pointer to a buffer
(of size starpu_task::cl_arg_size) with some parameters for the kernel
described by the codelet. For instance, if a codelet implements a
computational kernel that multiplies its input vector by a constant,
the constant could be specified by the means of this buffer, instead
of registering it as a StarPU data. It must however be noted that
StarPU avoids making copy whenever possible and rather passes the
pointer as such, so the buffer which is pointed at must be kept allocated
until the task terminates, and if several tasks are submitted with
various parameters, each of them must be given a pointer to their
own buffer.

Once a task has been executed, an optional callback function
starpu_task::callback_func is called when defined.
While the computational kernel could be offloaded on various architectures, the
callback function is always executed on a CPU. The pointer
starpu_task::callback_arg is passed as an argument of the callback
function. The prototype of a callback function must be:

\code{.c}
void (*callback_function)(void *);
\endcode

If the field starpu_task::synchronous is non-zero, task submission
will be synchronous: the function starpu_task_submit() will not return
until the task has been executed. Note that the function starpu_shutdown()
does not guarantee that asynchronous tasks have been executed before
it returns, starpu_task_wait_for_all() can be used to that effect, or
data can be unregistered (starpu_data_unregister()), which will
implicitly wait for all the tasks scheduled to work on it, unless
explicitly disabled thanks to
starpu_data_set_default_sequential_consistency_flag() or
starpu_data_set_sequential_consistency_flag().

\subsection ExecutionOfHelloWorld Execution Of Hello World

\verbatim
$ make hello_world
cc $(pkg-config --cflags starpu-1.2)  $(pkg-config --libs starpu-1.2) hello_world.c -o hello_world
$ ./hello_world
Hello world (params = {1, 2.000000} )
Callback function (arg 42)
\endverbatim

\section VectorScalingUsingTheCExtension Vector Scaling Using the C Extension

The previous example has shown how to submit tasks. In this section,
we show how StarPU tasks can manipulate data.

We will first show how to use the C language extensions provided by
the GCC plug-in (\ref cExtensions). The complete example, and
additional examples, is available in the directory <c>gcc-plugin/examples</c>
of the StarPU distribution. These extensions map directly
to StarPU's main concepts: tasks, task implementations for CPU,
OpenCL, or CUDA, and registered data buffers. The standard C version
that uses StarPU's standard C programming interface is given in \ref
VectorScalingUsingStarPUAPI.

First of all, the vector-scaling task and its simple CPU implementation
has to be defined:

\code{.c}
/* Declare the `vector_scal' task.  */
static void vector_scal (unsigned size, float vector[size],
                         float factor)
  __attribute__ ((task));

/* Define the standard CPU implementation.  */
static void
vector_scal (unsigned size, float vector[size], float factor)
{
  unsigned i;
  for (i = 0; i < size; i++)
    vector[i] *= factor;
}
\endcode

Next, the body of the program, which uses the task defined above, can be
implemented:

\snippet hello_pragma2.c To be included

The function <c>main</c> above does several things:

<ul>
<li>
It initializes StarPU.
</li>
<li>
It allocates <c>vector</c> in the heap; it will automatically be freed
when its scope is left.  Alternatively, good old <c>malloc</c> and
<c>free</c> could have been used, but they are more error-prone and
require more typing.
</li>
<li>
It registers the memory pointed to by <c>vector</c>.  Eventually,
when OpenCL or CUDA task implementations are added, this will allow
StarPU to transfer that memory region between GPUs and the main memory.
Removing this <c>pragma</c> is an error.
</li>
<li>
It invokes the task <c>vector_scal</c>.  The invocation looks the same
as a standard C function call.  However, it is an asynchronous
invocation, meaning that the actual call is performed in parallel with
the caller's continuation.
</li>
<li>
It waits for the termination of the asynchronous call <c>vector_scal</c>.
</li>
<li>
Finally, StarPU is shut down.
</li>
</ul>

The program can be compiled and linked with GCC and the flag <c>-fplugin</c>:

\verbatim
$ gcc `pkg-config starpu-1.2 --cflags` vector_scal.c \
    -fplugin=`pkg-config starpu-1.2 --variable=gccplugin` \
    `pkg-config starpu-1.2 --libs`
\endverbatim

And voilà!

\subsection AddingAnOpenCLTaskImplementation Adding an OpenCL Task Implementation

Now, this is all fine and great, but you certainly want to take
advantage of these newfangled GPUs that your lab just bought, don't you?

So, let's add an OpenCL implementation of the task <c>vector_scal</c>.
We assume that the OpenCL kernel is available in a file,
<c>vector_scal_opencl_kernel.cl</c>, not shown here.  The OpenCL task
implementation is similar to that used with the standard C API
(\ref DefinitionOfTheOpenCLKernel).  It is declared and defined
in our C file like this:

\code{.c}
/* The OpenCL programs, loaded from 'main' (see below). */
static struct starpu_opencl_program cl_programs;

static void vector_scal_opencl (unsigned size, float vector[size],
                                float factor)
  __attribute__ ((task_implementation ("opencl", vector_scal)));

static void
vector_scal_opencl (unsigned size, float vector[size], float factor)
{
  int id, devid, err;
  cl_kernel kernel;
  cl_command_queue queue;
  cl_event event;

  /* VECTOR is GPU memory pointer, not a main memory pointer. */
  cl_mem val = (cl_mem) vector;

  id = starpu_worker_get_id ();
  devid = starpu_worker_get_devid (id);

  /* Prepare to invoke the kernel.  In the future, this will be largely automated.  */
  err = starpu_opencl_load_kernel (&kernel, &queue, &cl_programs,
                                   "vector_mult_opencl", devid);
  if (err != CL_SUCCESS)
    STARPU_OPENCL_REPORT_ERROR (err);

  err = clSetKernelArg (kernel, 0, sizeof (size), &size);
  err |= clSetKernelArg (kernel, 1, sizeof (val), &val);
  err |= clSetKernelArg (kernel, 2, sizeof (factor), &factor);
  if (err)
    STARPU_OPENCL_REPORT_ERROR (err);

  size_t global = 1, local = 1;
  err = clEnqueueNDRangeKernel (queue, kernel, 1, NULL, &global,
                                &local, 0, NULL, &event);
  if (err != CL_SUCCESS)
    STARPU_OPENCL_REPORT_ERROR (err);

  clFinish (queue);
  starpu_opencl_collect_stats (event);
  clReleaseEvent (event);

  /* Done with KERNEL. */
  starpu_opencl_release_kernel (kernel);
}
\endcode

The OpenCL kernel itself must be loaded from <c>main</c>, sometime after
the pragma <c>initialize</c>:

\code{.c}
starpu_opencl_load_opencl_from_file ("vector_scal_opencl_kernel.cl",
                                       &cl_programs, "");
\endcode

And that's it.  The task <c>vector_scal</c> now has an additional
implementation, for OpenCL, which StarPU's scheduler may choose to use
at run-time.  Unfortunately, the <c>vector_scal_opencl</c> above still
has to go through the common OpenCL boilerplate; in the future,
additional extensions will automate most of it.

\subsection AddingACUDATaskImplementation Adding a CUDA Task Implementation

Adding a CUDA implementation of the task is very similar, except that
the implementation itself is typically written in CUDA, and compiled
with <c>nvcc</c>.  Thus, the C file only needs to contain an external
declaration for the task implementation:

\code{.c}
extern void vector_scal_cuda (unsigned size, float vector[size],
                              float factor)
  __attribute__ ((task_implementation ("cuda", vector_scal)));
\endcode

The actual implementation of the CUDA task goes into a separate
compilation unit, in a <c>.cu</c> file.  It is very close to the
implementation when using StarPU's standard C API (\ref DefinitionOfTheCUDAKernel).

\snippet scal_pragma.cu To be included

The complete source code, in the directory <c>gcc-plugin/examples/vector_scal</c>
of the StarPU distribution, also shows how an SSE-specialized
CPU task implementation can be added.

For more details on the C extensions provided by StarPU's GCC plug-in, see
\ref cExtensions.

\section VectorScalingUsingStarPUAPI Vector Scaling Using StarPU's API

This section shows how to achieve the same result as explained in the
previous section using StarPU's standard C API.

The full source code for
this example is given in \ref FullSourceCodeVectorScal.

\subsection SourceCodeOfVectorScaling Source Code of Vector Scaling

Programmers can describe the data layout of their application so that StarPU is
responsible for enforcing data coherency and availability across the machine.
Instead of handling complex (and non-portable) mechanisms to perform data
movements, programmers only declare which piece of data is accessed and/or
modified by a task, and StarPU makes sure that when a computational kernel
starts somewhere (e.g. on a GPU), its data are available locally.

Before submitting those tasks, the programmer first needs to declare the
different pieces of data to StarPU using the functions
<c>starpu_*_data_register</c>. To ease the development of applications
for StarPU, it is possible to describe multiple types of data layout.
A type of data layout is called an <b>interface</b>. There are
different predefined interfaces available in StarPU: here we will
consider the <b>vector interface</b>.

The following lines show how to declare an array of <c>NX</c> elements of type
<c>float</c> using the vector interface:

\code{.c}
float vector[NX];

starpu_data_handle_t vector_handle;
starpu_vector_data_register(&vector_handle, 0, (uintptr_t)vector, NX,
                            sizeof(vector[0]));
\endcode

The first argument, called the <b>data handle</b>, is an opaque pointer which
designates the array in StarPU. This is also the structure which is used to
describe which data is used by a task. The second argument is the node number
where the data originally resides. Here it is 0 since the array <c>vector</c> is in
the main memory. Then comes the pointer <c>vector</c> where the data can be found in main memory,
the number of elements in the vector and the size of each element.
The following shows how to construct a StarPU task that will manipulate the
vector and a constant factor.

\code{.c}
float factor = 3.14;
struct starpu_task *task = starpu_task_create();

task->cl = &cl;                      /* Pointer to the codelet defined below */
task->handles[0] = vector_handle;    /* First parameter of the codelet */
task->cl_arg = &factor;
task->cl_arg_size = sizeof(factor);
task->synchronous = 1;

starpu_task_submit(task);
\endcode

Since the factor is a mere constant float value parameter,
it does not need a preliminary registration, and
can just be passed through the pointer starpu_task::cl_arg like in the previous
example.  The vector parameter is described by its handle.
starpu_task::handles should be set with the handles of the data, the
access modes for the data are defined in the field
starpu_codelet::modes (::STARPU_R for read-only, ::STARPU_W for
write-only and ::STARPU_RW for read and write access).

The definition of the codelet can be written as follows:

\code{.c}
void scal_cpu_func(void *buffers[], void *cl_arg)
{
    unsigned i;
    float *factor = cl_arg;

    /* length of the vector */
    unsigned n = STARPU_VECTOR_GET_NX(buffers[0]);
    /* CPU copy of the vector pointer */
    float *val = (float *)STARPU_VECTOR_GET_PTR(buffers[0]);

    for (i = 0; i < n; i++)
        val[i] *= *factor;
}

struct starpu_codelet cl =
{
    .cpu_funcs = { scal_cpu_func, NULL },
    .cpu_funcs_name = { "scal_cpu_func", NULL },
    .nbuffers = 1,
    .modes = { STARPU_RW }
};
\endcode

The first argument is an array that gives
a description of all the buffers passed in the array starpu_task::handles. The
size of this array is given by the field starpu_codelet::nbuffers. For
the sake of genericity, this array contains pointers to the different
interfaces describing each buffer.  In the case of the <b>vector
interface</b>, the location of the vector (resp. its length) is
accessible in the starpu_vector_interface::ptr (resp.
starpu_vector_interface::nx) of this interface. Since the vector is
accessed in a read-write fashion, any modification will automatically
affect future accesses to this vector made by other tasks.

The second argument of the function <c>scal_cpu_func</c> contains a
pointer to the parameters of the codelet (given in
starpu_task::cl_arg), so that we read the constant factor from this
pointer.

\subsection ExecutionOfVectorScaling Execution of Vector Scaling

\verbatim
$ make vector_scal
cc $(pkg-config --cflags starpu-1.2)  $(pkg-config --libs starpu-1.2)  vector_scal.c   -o vector_scal
$ ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
\endverbatim

\section VectorScalingOnAnHybridCPUGPUMachine Vector Scaling on an Hybrid CPU/GPU Machine

Contrary to the previous examples, the task submitted in this example may not
only be executed by the CPUs, but also by a CUDA device.

\subsection DefinitionOfTheCUDAKernel Definition of the CUDA Kernel

The CUDA implementation can be written as follows. It needs to be compiled with
a CUDA compiler such as nvcc, the NVIDIA CUDA compiler driver. It must be noted
that the vector pointer returned by ::STARPU_VECTOR_GET_PTR is here a
pointer in GPU memory, so that it can be passed as such to the
kernel call <c>vector_mult_cuda</c>.

\snippet vector_scal_cuda.cu To be included

\subsection DefinitionOfTheOpenCLKernel Definition of the OpenCL Kernel

The OpenCL implementation can be written as follows. StarPU provides
tools to compile a OpenCL kernel stored in a file.

\code{.c}
__kernel void vector_mult_opencl(int nx, __global float* val, float factor)
{
        const int i = get_global_id(0);
        if (i < nx) {
                val[i] *= factor;
        }
}
\endcode

Contrary to CUDA and CPU, ::STARPU_VECTOR_GET_DEV_HANDLE has to be used,
which returns a <c>cl_mem</c> (which is not a device pointer, but an OpenCL
handle), which can be passed as such to the OpenCL kernel. The difference is
important when using partitioning, see \ref PartitioningData.

\snippet vector_scal_opencl.c To be included

\subsection DefinitionOfTheMainCode Definition of the Main Code

The CPU implementation is the same as in the previous section.

Here is the source of the main application. You can notice that the fields
starpu_codelet::cuda_funcs and starpu_codelet::opencl_funcs are set to
define the pointers to the CUDA and OpenCL implementations of the
task.

\snippet vector_scal_c.c To be included

\subsection ExecutionOfHybridVectorScaling Execution of Hybrid Vector Scaling

The Makefile given at the beginning of the section must be extended to
give the rules to compile the CUDA source code. Note that the source
file of the OpenCL kernel does not need to be compiled now, it will
be compiled at run-time when calling the function
starpu_opencl_load_opencl_from_file().

\verbatim
CFLAGS  += $(shell pkg-config --cflags starpu-1.2)
LDFLAGS += $(shell pkg-config --libs starpu-1.2)
CC       = gcc

vector_scal: vector_scal.o vector_scal_cpu.o vector_scal_cuda.o vector_scal_opencl.o

%.o: %.cu
       nvcc $(CFLAGS) $< -c $@

clean:
       rm -f vector_scal *.o
\endverbatim

\verbatim
$ make
\endverbatim

and to execute it, with the default configuration:

\verbatim
$ ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
\endverbatim

or for example, by disabling CPU devices:

\verbatim
$ STARPU_NCPU=0 ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
\endverbatim

or by disabling CUDA devices (which may permit to enable the use of OpenCL,
see \ref EnablingOpenCL) :

\verbatim
$ STARPU_NCUDA=0 ./vector_scal
0.000000 3.000000 6.000000 9.000000 12.000000
\endverbatim

*/